1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86TargetMachine.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/GlobalVariable.h"
23 #include "llvm/Function.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/ADT/BitVector.h"
26 #include "llvm/ADT/VectorExtras.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineModuleInfo.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/PseudoSourceValue.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Target/TargetOptions.h"
36 #include "llvm/ADT/SmallSet.h"
37 #include "llvm/ADT/StringExtras.h"
38 #include "llvm/Support/CommandLine.h"
42 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
44 // Forward declarations.
45 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
48 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
49 : TargetLowering(TM) {
50 Subtarget = &TM.getSubtarget<X86Subtarget>();
51 X86ScalarSSEf64 = Subtarget->hasSSE2();
52 X86ScalarSSEf32 = Subtarget->hasSSE1();
53 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
55 RegInfo = TM.getRegisterInfo();
58 // Set up the TargetLowering object.
60 // X86 is weird, it always uses i8 for shift amounts and setcc results.
61 setShiftAmountType(MVT::i8);
62 setBooleanContents(ZeroOrOneBooleanContent);
63 setSchedulingPreference(SchedulingForRegPressure);
64 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
65 setStackPointerRegisterToSaveRestore(X86StackPtr);
67 if (Subtarget->isTargetDarwin()) {
68 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
69 setUseUnderscoreSetJmp(false);
70 setUseUnderscoreLongJmp(false);
71 } else if (Subtarget->isTargetMingw()) {
72 // MS runtime is weird: it exports _setjmp, but longjmp!
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(false);
76 setUseUnderscoreSetJmp(true);
77 setUseUnderscoreLongJmp(true);
80 // Set up the register classes.
81 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
82 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
83 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
84 if (Subtarget->is64Bit())
85 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
87 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
89 // We don't accept any truncstore of integer registers.
90 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
91 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
92 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
93 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
94 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
95 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
97 // SETOEQ and SETUNE require checking two conditions.
98 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
99 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
100 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
101 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
102 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
103 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
105 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
107 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
108 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
109 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
111 if (Subtarget->is64Bit()) {
112 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
113 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
114 } else if (!UseSoftFloat) {
115 if (X86ScalarSSEf64) {
116 // We have an impenetrably clever algorithm for ui64->double only.
117 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
119 // We have an algorithm for SSE2, and we turn this into a 64-bit
120 // FILD for other targets.
121 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
124 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
126 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
127 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
130 // SSE has no i16 to fp conversion, only i32
131 if (X86ScalarSSEf32) {
132 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
133 // f32 and f64 cases are Legal, f80 case is not
134 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
136 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
137 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
140 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
141 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
144 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
145 // are Legal, f80 is custom lowered.
146 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
147 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
149 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
151 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
152 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
154 if (X86ScalarSSEf32) {
155 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
156 // f32 and f64 cases are Legal, f80 case is not
157 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
159 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
160 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
163 // Handle FP_TO_UINT by promoting the destination to a larger signed
165 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
166 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
167 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
169 if (Subtarget->is64Bit()) {
170 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
171 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
172 } else if (!UseSoftFloat) {
173 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
174 // Expand FP_TO_UINT into a select.
175 // FIXME: We would like to use a Custom expander here eventually to do
176 // the optimal thing for SSE vs. the default expansion in the legalizer.
177 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
179 // With SSE3 we can use fisttpll to convert to a signed i64; without
180 // SSE, we're stuck with a fistpll.
181 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
184 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
185 if (!X86ScalarSSEf64) {
186 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
187 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
190 // Scalar integer divide and remainder are lowered to use operations that
191 // produce two results, to match the available instructions. This exposes
192 // the two-result form to trivial CSE, which is able to combine x/y and x%y
193 // into a single instruction.
195 // Scalar integer multiply-high is also lowered to use two-result
196 // operations, to match the available instructions. However, plain multiply
197 // (low) operations are left as Legal, as there are single-result
198 // instructions for this in x86. Using the two-result multiply instructions
199 // when both high and low results are needed must be arranged by dagcombine.
200 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
201 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
202 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
203 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
204 setOperationAction(ISD::SREM , MVT::i8 , Expand);
205 setOperationAction(ISD::UREM , MVT::i8 , Expand);
206 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
207 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
208 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
209 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
210 setOperationAction(ISD::SREM , MVT::i16 , Expand);
211 setOperationAction(ISD::UREM , MVT::i16 , Expand);
212 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
213 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
214 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
215 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
216 setOperationAction(ISD::SREM , MVT::i32 , Expand);
217 setOperationAction(ISD::UREM , MVT::i32 , Expand);
218 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
219 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
220 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
221 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
222 setOperationAction(ISD::SREM , MVT::i64 , Expand);
223 setOperationAction(ISD::UREM , MVT::i64 , Expand);
225 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
226 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
227 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
228 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
229 if (Subtarget->is64Bit())
230 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
231 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
232 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
233 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
234 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
235 setOperationAction(ISD::FREM , MVT::f32 , Expand);
236 setOperationAction(ISD::FREM , MVT::f64 , Expand);
237 setOperationAction(ISD::FREM , MVT::f80 , Expand);
238 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
240 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
241 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
242 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
243 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
244 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
245 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
246 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
247 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
248 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
249 if (Subtarget->is64Bit()) {
250 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
251 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
252 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
255 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
256 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
258 // These should be promoted to a larger select which is supported.
259 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
260 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
261 // X86 wants to expand cmov itself.
262 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
263 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
264 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
265 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
266 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
267 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
268 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
269 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
270 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
271 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
272 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
273 if (Subtarget->is64Bit()) {
274 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
275 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
277 // X86 ret instruction may pop stack.
278 setOperationAction(ISD::RET , MVT::Other, Custom);
279 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
282 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
283 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
284 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
285 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
286 if (Subtarget->is64Bit())
287 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
288 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
289 if (Subtarget->is64Bit()) {
290 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
291 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
292 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
293 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
295 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
296 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
297 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
298 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
299 if (Subtarget->is64Bit()) {
300 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
301 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
302 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
305 if (Subtarget->hasSSE1())
306 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
308 if (!Subtarget->hasSSE2())
309 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
311 // Expand certain atomics
312 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
313 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
314 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
315 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
317 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
318 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
319 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
320 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
322 if (!Subtarget->is64Bit()) {
323 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
324 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
325 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
326 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
327 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
328 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
329 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
332 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
333 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
334 // FIXME - use subtarget debug flags
335 if (!Subtarget->isTargetDarwin() &&
336 !Subtarget->isTargetELF() &&
337 !Subtarget->isTargetCygMing()) {
338 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
339 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
342 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
343 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
344 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
345 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
346 if (Subtarget->is64Bit()) {
347 setExceptionPointerRegister(X86::RAX);
348 setExceptionSelectorRegister(X86::RDX);
350 setExceptionPointerRegister(X86::EAX);
351 setExceptionSelectorRegister(X86::EDX);
353 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
354 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
356 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
358 setOperationAction(ISD::TRAP, MVT::Other, Legal);
360 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
361 setOperationAction(ISD::VASTART , MVT::Other, Custom);
362 setOperationAction(ISD::VAEND , MVT::Other, Expand);
363 if (Subtarget->is64Bit()) {
364 setOperationAction(ISD::VAARG , MVT::Other, Custom);
365 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
367 setOperationAction(ISD::VAARG , MVT::Other, Expand);
368 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
371 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
372 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
373 if (Subtarget->is64Bit())
374 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
375 if (Subtarget->isTargetCygMing())
376 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
378 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
380 if (!UseSoftFloat && X86ScalarSSEf64) {
381 // f32 and f64 use SSE.
382 // Set up the FP register classes.
383 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
384 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
386 // Use ANDPD to simulate FABS.
387 setOperationAction(ISD::FABS , MVT::f64, Custom);
388 setOperationAction(ISD::FABS , MVT::f32, Custom);
390 // Use XORP to simulate FNEG.
391 setOperationAction(ISD::FNEG , MVT::f64, Custom);
392 setOperationAction(ISD::FNEG , MVT::f32, Custom);
394 // Use ANDPD and ORPD to simulate FCOPYSIGN.
395 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
396 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
398 // We don't support sin/cos/fmod
399 setOperationAction(ISD::FSIN , MVT::f64, Expand);
400 setOperationAction(ISD::FCOS , MVT::f64, Expand);
401 setOperationAction(ISD::FSIN , MVT::f32, Expand);
402 setOperationAction(ISD::FCOS , MVT::f32, Expand);
404 // Expand FP immediates into loads from the stack, except for the special
406 addLegalFPImmediate(APFloat(+0.0)); // xorpd
407 addLegalFPImmediate(APFloat(+0.0f)); // xorps
408 } else if (!UseSoftFloat && X86ScalarSSEf32) {
409 // Use SSE for f32, x87 for f64.
410 // Set up the FP register classes.
411 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
412 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
414 // Use ANDPS to simulate FABS.
415 setOperationAction(ISD::FABS , MVT::f32, Custom);
417 // Use XORP to simulate FNEG.
418 setOperationAction(ISD::FNEG , MVT::f32, Custom);
420 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
422 // Use ANDPS and ORPS to simulate FCOPYSIGN.
423 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
424 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
426 // We don't support sin/cos/fmod
427 setOperationAction(ISD::FSIN , MVT::f32, Expand);
428 setOperationAction(ISD::FCOS , MVT::f32, Expand);
430 // Special cases we handle for FP constants.
431 addLegalFPImmediate(APFloat(+0.0f)); // xorps
432 addLegalFPImmediate(APFloat(+0.0)); // FLD0
433 addLegalFPImmediate(APFloat(+1.0)); // FLD1
434 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
435 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
438 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
439 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
441 } else if (!UseSoftFloat) {
442 // f32 and f64 in x87.
443 // Set up the FP register classes.
444 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
445 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
447 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
448 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
449 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
450 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
453 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
454 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
456 addLegalFPImmediate(APFloat(+0.0)); // FLD0
457 addLegalFPImmediate(APFloat(+1.0)); // FLD1
458 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
459 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
460 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
461 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
462 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
463 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
466 // Long double always uses X87.
468 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
469 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
470 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
473 APFloat TmpFlt(+0.0);
474 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
476 addLegalFPImmediate(TmpFlt); // FLD0
478 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
479 APFloat TmpFlt2(+1.0);
480 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
482 addLegalFPImmediate(TmpFlt2); // FLD1
483 TmpFlt2.changeSign();
484 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
488 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
489 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
493 // Always use a library call for pow.
494 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
495 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
496 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
498 setOperationAction(ISD::FLOG, MVT::f80, Expand);
499 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
500 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
501 setOperationAction(ISD::FEXP, MVT::f80, Expand);
502 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
504 // First set operation action for all vector types to either promote
505 // (for widening) or expand (for scalarization). Then we will selectively
506 // turn on ones that can be effectively codegen'd.
507 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
508 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
509 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
510 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
511 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
512 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
513 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
514 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
515 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
516 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
517 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
518 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
519 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
520 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
521 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
522 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
523 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
524 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
525 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
526 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
527 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
528 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
529 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
532 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
559 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
560 // with -msoft-float, disable use of MMX as well.
561 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
562 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
563 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
564 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
565 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
566 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
568 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
569 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
570 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
571 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
573 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
574 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
575 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
576 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
578 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
579 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
581 setOperationAction(ISD::AND, MVT::v8i8, Promote);
582 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
583 setOperationAction(ISD::AND, MVT::v4i16, Promote);
584 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
585 setOperationAction(ISD::AND, MVT::v2i32, Promote);
586 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
587 setOperationAction(ISD::AND, MVT::v1i64, Legal);
589 setOperationAction(ISD::OR, MVT::v8i8, Promote);
590 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
591 setOperationAction(ISD::OR, MVT::v4i16, Promote);
592 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
593 setOperationAction(ISD::OR, MVT::v2i32, Promote);
594 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
595 setOperationAction(ISD::OR, MVT::v1i64, Legal);
597 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
598 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
599 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
600 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
601 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
602 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
603 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
605 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
606 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
607 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
608 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
609 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
610 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
611 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
612 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
613 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
615 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
616 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
617 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
618 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
619 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
621 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
622 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
623 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
624 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
626 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
627 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
628 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
629 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
631 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
633 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
634 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
635 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
636 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
637 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
638 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
641 if (!UseSoftFloat && Subtarget->hasSSE1()) {
642 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
644 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
645 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
646 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
647 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
648 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
649 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
650 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
651 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
652 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
653 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
654 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
655 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
658 if (!UseSoftFloat && Subtarget->hasSSE2()) {
659 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
661 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
662 // registers cannot be used even for integer operations.
663 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
664 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
665 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
666 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
668 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
669 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
670 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
671 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
672 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
673 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
674 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
675 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
676 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
677 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
678 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
679 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
680 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
681 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
682 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
683 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
685 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
686 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
690 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
691 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
692 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
693 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
694 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
696 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
697 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
698 MVT VT = (MVT::SimpleValueType)i;
699 // Do not attempt to custom lower non-power-of-2 vectors
700 if (!isPowerOf2_32(VT.getVectorNumElements()))
702 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
703 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
704 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
707 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
708 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
709 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
711 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
712 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
714 if (Subtarget->is64Bit()) {
715 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
716 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
719 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
720 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
721 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
722 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v2i64);
723 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
724 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v2i64);
725 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
726 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v2i64);
727 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
728 AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v2i64);
729 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
730 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v2i64);
733 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
735 // Custom lower v2i64 and v2f64 selects.
736 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
737 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
738 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
739 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
741 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
742 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
743 if (!DisableMMX && Subtarget->hasMMX()) {
744 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
745 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
749 if (Subtarget->hasSSE41()) {
750 // FIXME: Do we need to handle scalar-to-vector here?
751 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
753 // i8 and i16 vectors are custom , because the source register and source
754 // source memory operand types are not the same width. f32 vectors are
755 // custom since the immediate controlling the insert encodes additional
757 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
758 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
759 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
760 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
762 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
763 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
764 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
765 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
767 if (Subtarget->is64Bit()) {
768 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
769 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
773 if (Subtarget->hasSSE42()) {
774 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
777 // We want to custom lower some of our intrinsics.
778 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
780 // Add/Sub/Mul with overflow operations are custom lowered.
781 setOperationAction(ISD::SADDO, MVT::i32, Custom);
782 setOperationAction(ISD::SADDO, MVT::i64, Custom);
783 setOperationAction(ISD::UADDO, MVT::i32, Custom);
784 setOperationAction(ISD::UADDO, MVT::i64, Custom);
785 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
786 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
787 setOperationAction(ISD::USUBO, MVT::i32, Custom);
788 setOperationAction(ISD::USUBO, MVT::i64, Custom);
789 setOperationAction(ISD::SMULO, MVT::i32, Custom);
790 setOperationAction(ISD::SMULO, MVT::i64, Custom);
791 setOperationAction(ISD::UMULO, MVT::i32, Custom);
792 setOperationAction(ISD::UMULO, MVT::i64, Custom);
794 if (!Subtarget->is64Bit()) {
795 // These libcalls are not available in 32-bit.
796 setLibcallName(RTLIB::SHL_I128, 0);
797 setLibcallName(RTLIB::SRL_I128, 0);
798 setLibcallName(RTLIB::SRA_I128, 0);
801 // We have target-specific dag combine patterns for the following nodes:
802 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
803 setTargetDAGCombine(ISD::BUILD_VECTOR);
804 setTargetDAGCombine(ISD::SELECT);
805 setTargetDAGCombine(ISD::SHL);
806 setTargetDAGCombine(ISD::SRA);
807 setTargetDAGCombine(ISD::SRL);
808 setTargetDAGCombine(ISD::STORE);
809 if (Subtarget->is64Bit())
810 setTargetDAGCombine(ISD::MUL);
812 computeRegisterProperties();
814 // FIXME: These should be based on subtarget info. Plus, the values should
815 // be smaller when we are in optimizing for size mode.
816 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
817 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
818 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
819 allowUnalignedMemoryAccesses = true; // x86 supports it!
820 setPrefLoopAlignment(16);
821 benefitFromCodePlacementOpt = true;
825 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
830 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
831 /// the desired ByVal argument alignment.
832 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
835 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
836 if (VTy->getBitWidth() == 128)
838 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
839 unsigned EltAlign = 0;
840 getMaxByValAlign(ATy->getElementType(), EltAlign);
841 if (EltAlign > MaxAlign)
843 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
844 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
845 unsigned EltAlign = 0;
846 getMaxByValAlign(STy->getElementType(i), EltAlign);
847 if (EltAlign > MaxAlign)
856 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
857 /// function arguments in the caller parameter area. For X86, aggregates
858 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
859 /// are at 4-byte boundaries.
860 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
861 if (Subtarget->is64Bit()) {
862 // Max of 8 and alignment of type.
863 unsigned TyAlign = TD->getABITypeAlignment(Ty);
870 if (Subtarget->hasSSE1())
871 getMaxByValAlign(Ty, Align);
875 /// getOptimalMemOpType - Returns the target specific optimal type for load
876 /// and store operations as a result of memset, memcpy, and memmove
877 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
880 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
881 bool isSrcConst, bool isSrcStr,
882 SelectionDAG &DAG) const {
883 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
884 // linux. This is because the stack realignment code can't handle certain
885 // cases like PR2962. This should be removed when PR2962 is fixed.
886 const Function *F = DAG.getMachineFunction().getFunction();
887 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
888 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
889 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
891 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
894 if (Subtarget->is64Bit() && Size >= 8)
899 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
901 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
902 SelectionDAG &DAG) const {
903 if (usesGlobalOffsetTable())
904 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
905 if (!Subtarget->isPICStyleRIPRel())
906 // This doesn't have DebugLoc associated with it, but is not really the
907 // same as a Register.
908 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
913 //===----------------------------------------------------------------------===//
914 // Return Value Calling Convention Implementation
915 //===----------------------------------------------------------------------===//
917 #include "X86GenCallingConv.inc"
919 /// LowerRET - Lower an ISD::RET node.
920 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
921 DebugLoc dl = Op.getDebugLoc();
922 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
924 SmallVector<CCValAssign, 16> RVLocs;
925 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
926 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
927 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
928 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
930 // If this is the first return lowered for this function, add the regs to the
931 // liveout set for the function.
932 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
933 for (unsigned i = 0; i != RVLocs.size(); ++i)
934 if (RVLocs[i].isRegLoc())
935 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
937 SDValue Chain = Op.getOperand(0);
939 // Handle tail call return.
940 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
941 if (Chain.getOpcode() == X86ISD::TAILCALL) {
942 SDValue TailCall = Chain;
943 SDValue TargetAddress = TailCall.getOperand(1);
944 SDValue StackAdjustment = TailCall.getOperand(2);
945 assert(((TargetAddress.getOpcode() == ISD::Register &&
946 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
947 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R9)) ||
948 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
949 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
950 "Expecting an global address, external symbol, or register");
951 assert(StackAdjustment.getOpcode() == ISD::Constant &&
952 "Expecting a const value");
954 SmallVector<SDValue,8> Operands;
955 Operands.push_back(Chain.getOperand(0));
956 Operands.push_back(TargetAddress);
957 Operands.push_back(StackAdjustment);
958 // Copy registers used by the call. Last operand is a flag so it is not
960 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
961 Operands.push_back(Chain.getOperand(i));
963 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
970 SmallVector<SDValue, 6> RetOps;
971 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
972 // Operand #1 = Bytes To Pop
973 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
975 // Copy the result values into the output registers.
976 for (unsigned i = 0; i != RVLocs.size(); ++i) {
977 CCValAssign &VA = RVLocs[i];
978 assert(VA.isRegLoc() && "Can only return in registers!");
979 SDValue ValToCopy = Op.getOperand(i*2+1);
981 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
982 // the RET instruction and handled by the FP Stackifier.
983 if (VA.getLocReg() == X86::ST0 ||
984 VA.getLocReg() == X86::ST1) {
985 // If this is a copy from an xmm register to ST(0), use an FPExtend to
986 // change the value to the FP stack register class.
987 if (isScalarFPTypeInSSEReg(VA.getValVT()))
988 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
989 RetOps.push_back(ValToCopy);
990 // Don't emit a copytoreg.
994 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
995 // which is returned in RAX / RDX.
996 if (Subtarget->is64Bit()) {
997 MVT ValVT = ValToCopy.getValueType();
998 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
999 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1000 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1001 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1005 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1006 Flag = Chain.getValue(1);
1009 // The x86-64 ABI for returning structs by value requires that we copy
1010 // the sret argument into %rax for the return. We saved the argument into
1011 // a virtual register in the entry block, so now we copy the value out
1013 if (Subtarget->is64Bit() &&
1014 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1015 MachineFunction &MF = DAG.getMachineFunction();
1016 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1017 unsigned Reg = FuncInfo->getSRetReturnReg();
1019 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1020 FuncInfo->setSRetReturnReg(Reg);
1022 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1024 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1025 Flag = Chain.getValue(1);
1028 RetOps[0] = Chain; // Update chain.
1030 // Add the flag if we have it.
1032 RetOps.push_back(Flag);
1034 return DAG.getNode(X86ISD::RET_FLAG, dl,
1035 MVT::Other, &RetOps[0], RetOps.size());
1039 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1040 /// appropriate copies out of appropriate physical registers. This assumes that
1041 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1042 /// being lowered. The returns a SDNode with the same number of values as the
1044 SDNode *X86TargetLowering::
1045 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1046 unsigned CallingConv, SelectionDAG &DAG) {
1048 DebugLoc dl = TheCall->getDebugLoc();
1049 // Assign locations to each value returned by this call.
1050 SmallVector<CCValAssign, 16> RVLocs;
1051 bool isVarArg = TheCall->isVarArg();
1052 bool Is64Bit = Subtarget->is64Bit();
1053 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
1054 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1056 SmallVector<SDValue, 8> ResultVals;
1058 // Copy all of the result registers out of their specified physreg.
1059 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1060 CCValAssign &VA = RVLocs[i];
1061 MVT CopyVT = VA.getValVT();
1063 // If this is x86-64, and we disabled SSE, we can't return FP values
1064 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1065 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1066 cerr << "SSE register return with SSE disabled\n";
1070 // If this is a call to a function that returns an fp value on the floating
1071 // point stack, but where we prefer to use the value in xmm registers, copy
1072 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1073 if ((VA.getLocReg() == X86::ST0 ||
1074 VA.getLocReg() == X86::ST1) &&
1075 isScalarFPTypeInSSEReg(VA.getValVT())) {
1080 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1081 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1082 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1083 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1084 MVT::v2i64, InFlag).getValue(1);
1085 Val = Chain.getValue(0);
1086 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1087 Val, DAG.getConstant(0, MVT::i64));
1089 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1090 MVT::i64, InFlag).getValue(1);
1091 Val = Chain.getValue(0);
1093 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1095 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1096 CopyVT, InFlag).getValue(1);
1097 Val = Chain.getValue(0);
1099 InFlag = Chain.getValue(2);
1101 if (CopyVT != VA.getValVT()) {
1102 // Round the F80 the right size, which also moves to the appropriate xmm
1104 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1105 // This truncation won't change the value.
1106 DAG.getIntPtrConstant(1));
1109 ResultVals.push_back(Val);
1112 // Merge everything together with a MERGE_VALUES node.
1113 ResultVals.push_back(Chain);
1114 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1115 &ResultVals[0], ResultVals.size()).getNode();
1119 //===----------------------------------------------------------------------===//
1120 // C & StdCall & Fast Calling Convention implementation
1121 //===----------------------------------------------------------------------===//
1122 // StdCall calling convention seems to be standard for many Windows' API
1123 // routines and around. It differs from C calling convention just a little:
1124 // callee should clean up the stack, not caller. Symbols should be also
1125 // decorated in some fancy way :) It doesn't support any vector arguments.
1126 // For info on fast calling convention see Fast Calling Convention (tail call)
1127 // implementation LowerX86_32FastCCCallTo.
1129 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1131 static bool CallIsStructReturn(CallSDNode *TheCall) {
1132 unsigned NumOps = TheCall->getNumArgs();
1136 return TheCall->getArgFlags(0).isSRet();
1139 /// ArgsAreStructReturn - Determines whether a FORMAL_ARGUMENTS node uses struct
1140 /// return semantics.
1141 static bool ArgsAreStructReturn(SDValue Op) {
1142 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1146 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1149 /// IsCalleePop - Determines whether a CALL or FORMAL_ARGUMENTS node requires
1150 /// the callee to pop its own arguments. Callee pop is necessary to support tail
1152 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1156 switch (CallingConv) {
1159 case CallingConv::X86_StdCall:
1160 return !Subtarget->is64Bit();
1161 case CallingConv::X86_FastCall:
1162 return !Subtarget->is64Bit();
1163 case CallingConv::Fast:
1164 return PerformTailCallOpt;
1168 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1169 /// given CallingConvention value.
1170 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1171 if (Subtarget->is64Bit()) {
1172 if (Subtarget->isTargetWin64())
1173 return CC_X86_Win64_C;
1174 else if (CC == CallingConv::Fast && PerformTailCallOpt)
1175 return CC_X86_64_TailCall;
1180 if (CC == CallingConv::X86_FastCall)
1181 return CC_X86_32_FastCall;
1182 else if (CC == CallingConv::Fast)
1183 return CC_X86_32_FastCC;
1188 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1189 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1191 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1192 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1193 if (CC == CallingConv::X86_FastCall)
1195 else if (CC == CallingConv::X86_StdCall)
1201 /// CallRequiresGOTInRegister - Check whether the call requires the GOT pointer
1202 /// in a register before calling.
1203 bool X86TargetLowering::CallRequiresGOTPtrInReg(bool Is64Bit, bool IsTailCall) {
1204 return !IsTailCall && !Is64Bit &&
1205 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1206 Subtarget->isPICStyleGOT();
1209 /// CallRequiresFnAddressInReg - Check whether the call requires the function
1210 /// address to be loaded in a register.
1212 X86TargetLowering::CallRequiresFnAddressInReg(bool Is64Bit, bool IsTailCall) {
1213 return !Is64Bit && IsTailCall &&
1214 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1215 Subtarget->isPICStyleGOT();
1218 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1219 /// by "Src" to address "Dst" with size and alignment information specified by
1220 /// the specific parameter attribute. The copy will be passed as a byval
1221 /// function parameter.
1223 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1224 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1226 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1227 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1228 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1231 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1232 const CCValAssign &VA,
1233 MachineFrameInfo *MFI,
1235 SDValue Root, unsigned i) {
1236 // Create the nodes corresponding to a load from this parameter slot.
1237 ISD::ArgFlagsTy Flags =
1238 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1239 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1240 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1242 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1243 // changed with more analysis.
1244 // In case of tail call optimization mark all arguments mutable. Since they
1245 // could be overwritten by lowering of arguments in case of a tail call.
1246 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1247 VA.getLocMemOffset(), isImmutable);
1248 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1249 if (Flags.isByVal())
1251 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1252 PseudoSourceValue::getFixedStack(FI), 0);
1256 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1257 MachineFunction &MF = DAG.getMachineFunction();
1258 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1259 DebugLoc dl = Op.getDebugLoc();
1261 const Function* Fn = MF.getFunction();
1262 if (Fn->hasExternalLinkage() &&
1263 Subtarget->isTargetCygMing() &&
1264 Fn->getName() == "main")
1265 FuncInfo->setForceFramePointer(true);
1267 // Decorate the function name.
1268 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1270 MachineFrameInfo *MFI = MF.getFrameInfo();
1271 SDValue Root = Op.getOperand(0);
1272 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1273 unsigned CC = MF.getFunction()->getCallingConv();
1274 bool Is64Bit = Subtarget->is64Bit();
1275 bool IsWin64 = Subtarget->isTargetWin64();
1277 assert(!(isVarArg && CC == CallingConv::Fast) &&
1278 "Var args not supported with calling convention fastcc");
1280 // Assign locations to all of the incoming arguments.
1281 SmallVector<CCValAssign, 16> ArgLocs;
1282 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1283 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1285 SmallVector<SDValue, 8> ArgValues;
1286 unsigned LastVal = ~0U;
1287 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1288 CCValAssign &VA = ArgLocs[i];
1289 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1291 assert(VA.getValNo() != LastVal &&
1292 "Don't support value assigned to multiple locs yet");
1293 LastVal = VA.getValNo();
1295 if (VA.isRegLoc()) {
1296 MVT RegVT = VA.getLocVT();
1297 TargetRegisterClass *RC = NULL;
1298 if (RegVT == MVT::i32)
1299 RC = X86::GR32RegisterClass;
1300 else if (Is64Bit && RegVT == MVT::i64)
1301 RC = X86::GR64RegisterClass;
1302 else if (RegVT == MVT::f32)
1303 RC = X86::FR32RegisterClass;
1304 else if (RegVT == MVT::f64)
1305 RC = X86::FR64RegisterClass;
1306 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1307 RC = X86::VR128RegisterClass;
1308 else if (RegVT.isVector()) {
1309 assert(RegVT.getSizeInBits() == 64);
1311 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1313 // Darwin calling convention passes MMX values in either GPRs or
1314 // XMMs in x86-64. Other targets pass them in memory.
1315 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1316 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1319 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1324 assert(0 && "Unknown argument type!");
1327 unsigned Reg = DAG.getMachineFunction().addLiveIn(VA.getLocReg(), RC);
1328 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1330 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1331 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1333 if (VA.getLocInfo() == CCValAssign::SExt)
1334 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1335 DAG.getValueType(VA.getValVT()));
1336 else if (VA.getLocInfo() == CCValAssign::ZExt)
1337 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1338 DAG.getValueType(VA.getValVT()));
1340 if (VA.getLocInfo() != CCValAssign::Full)
1341 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1343 // Handle MMX values passed in GPRs.
1344 if (Is64Bit && RegVT != VA.getLocVT()) {
1345 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1346 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1347 else if (RC == X86::VR128RegisterClass) {
1348 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1349 ArgValue, DAG.getConstant(0, MVT::i64));
1350 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1354 ArgValues.push_back(ArgValue);
1356 assert(VA.isMemLoc());
1357 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1361 // The x86-64 ABI for returning structs by value requires that we copy
1362 // the sret argument into %rax for the return. Save the argument into
1363 // a virtual register so that we can access it from the return points.
1364 if (Is64Bit && DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1365 MachineFunction &MF = DAG.getMachineFunction();
1366 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1367 unsigned Reg = FuncInfo->getSRetReturnReg();
1369 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1370 FuncInfo->setSRetReturnReg(Reg);
1372 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1373 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1376 unsigned StackSize = CCInfo.getNextStackOffset();
1377 // align stack specially for tail calls
1378 if (PerformTailCallOpt && CC == CallingConv::Fast)
1379 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1381 // If the function takes variable number of arguments, make a frame index for
1382 // the start of the first vararg value... for expansion of llvm.va_start.
1384 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1385 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1388 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1390 // FIXME: We should really autogenerate these arrays
1391 static const unsigned GPR64ArgRegsWin64[] = {
1392 X86::RCX, X86::RDX, X86::R8, X86::R9
1394 static const unsigned XMMArgRegsWin64[] = {
1395 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1397 static const unsigned GPR64ArgRegs64Bit[] = {
1398 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1400 static const unsigned XMMArgRegs64Bit[] = {
1401 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1402 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1404 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1407 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1408 GPR64ArgRegs = GPR64ArgRegsWin64;
1409 XMMArgRegs = XMMArgRegsWin64;
1411 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1412 GPR64ArgRegs = GPR64ArgRegs64Bit;
1413 XMMArgRegs = XMMArgRegs64Bit;
1415 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1417 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1420 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1421 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1422 "SSE register cannot be used when SSE is disabled!");
1423 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1424 "SSE register cannot be used when SSE is disabled!");
1425 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1426 // Kernel mode asks for SSE to be disabled, so don't push them
1428 TotalNumXMMRegs = 0;
1430 // For X86-64, if there are vararg parameters that are passed via
1431 // registers, then we must store them to their spots on the stack so they
1432 // may be loaded by deferencing the result of va_next.
1433 VarArgsGPOffset = NumIntRegs * 8;
1434 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1435 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1436 TotalNumXMMRegs * 16, 16);
1438 // Store the integer parameter registers.
1439 SmallVector<SDValue, 8> MemOps;
1440 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1441 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1442 DAG.getIntPtrConstant(VarArgsGPOffset));
1443 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1444 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1445 X86::GR64RegisterClass);
1446 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1448 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1449 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1450 MemOps.push_back(Store);
1451 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1452 DAG.getIntPtrConstant(8));
1455 // Now store the XMM (fp + vector) parameter registers.
1456 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1457 DAG.getIntPtrConstant(VarArgsFPOffset));
1458 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1459 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1460 X86::VR128RegisterClass);
1461 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1463 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1464 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1465 MemOps.push_back(Store);
1466 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1467 DAG.getIntPtrConstant(16));
1469 if (!MemOps.empty())
1470 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1471 &MemOps[0], MemOps.size());
1475 ArgValues.push_back(Root);
1477 // Some CCs need callee pop.
1478 if (IsCalleePop(isVarArg, CC)) {
1479 BytesToPopOnReturn = StackSize; // Callee pops everything.
1480 BytesCallerReserves = 0;
1482 BytesToPopOnReturn = 0; // Callee pops nothing.
1483 // If this is an sret function, the return should pop the hidden pointer.
1484 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1485 BytesToPopOnReturn = 4;
1486 BytesCallerReserves = StackSize;
1490 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1491 if (CC == CallingConv::X86_FastCall)
1492 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1495 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1497 // Return the new list of results.
1498 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1499 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1503 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1504 const SDValue &StackPtr,
1505 const CCValAssign &VA,
1507 SDValue Arg, ISD::ArgFlagsTy Flags) {
1508 DebugLoc dl = TheCall->getDebugLoc();
1509 unsigned LocMemOffset = VA.getLocMemOffset();
1510 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1511 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1512 if (Flags.isByVal()) {
1513 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1515 return DAG.getStore(Chain, dl, Arg, PtrOff,
1516 PseudoSourceValue::getStack(), LocMemOffset);
1519 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1520 /// optimization is performed and it is required.
1522 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1523 SDValue &OutRetAddr,
1529 if (!IsTailCall || FPDiff==0) return Chain;
1531 // Adjust the Return address stack slot.
1532 MVT VT = getPointerTy();
1533 OutRetAddr = getReturnAddressFrameIndex(DAG);
1535 // Load the "old" Return address.
1536 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1537 return SDValue(OutRetAddr.getNode(), 1);
1540 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1541 /// optimization is performed and it is required (FPDiff!=0).
1543 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1544 SDValue Chain, SDValue RetAddrFrIdx,
1545 bool Is64Bit, int FPDiff, DebugLoc dl) {
1546 // Store the return address to the appropriate stack slot.
1547 if (!FPDiff) return Chain;
1548 // Calculate the new stack slot for the return address.
1549 int SlotSize = Is64Bit ? 8 : 4;
1550 int NewReturnAddrFI =
1551 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1552 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1553 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1554 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1555 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1559 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1560 MachineFunction &MF = DAG.getMachineFunction();
1561 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1562 SDValue Chain = TheCall->getChain();
1563 unsigned CC = TheCall->getCallingConv();
1564 bool isVarArg = TheCall->isVarArg();
1565 bool IsTailCall = TheCall->isTailCall() &&
1566 CC == CallingConv::Fast && PerformTailCallOpt;
1567 SDValue Callee = TheCall->getCallee();
1568 bool Is64Bit = Subtarget->is64Bit();
1569 bool IsStructRet = CallIsStructReturn(TheCall);
1570 DebugLoc dl = TheCall->getDebugLoc();
1572 assert(!(isVarArg && CC == CallingConv::Fast) &&
1573 "Var args not supported with calling convention fastcc");
1575 // Analyze operands of the call, assigning locations to each operand.
1576 SmallVector<CCValAssign, 16> ArgLocs;
1577 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1578 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1580 // Get a count of how many bytes are to be pushed on the stack.
1581 unsigned NumBytes = CCInfo.getNextStackOffset();
1582 if (PerformTailCallOpt && CC == CallingConv::Fast)
1583 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1587 // Lower arguments at fp - stackoffset + fpdiff.
1588 unsigned NumBytesCallerPushed =
1589 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1590 FPDiff = NumBytesCallerPushed - NumBytes;
1592 // Set the delta of movement of the returnaddr stackslot.
1593 // But only set if delta is greater than previous delta.
1594 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1595 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1598 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1600 SDValue RetAddrFrIdx;
1601 // Load return adress for tail calls.
1602 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1605 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1606 SmallVector<SDValue, 8> MemOpChains;
1609 // Walk the register/memloc assignments, inserting copies/loads. In the case
1610 // of tail call optimization arguments are handle later.
1611 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1612 CCValAssign &VA = ArgLocs[i];
1613 SDValue Arg = TheCall->getArg(i);
1614 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1615 bool isByVal = Flags.isByVal();
1617 // Promote the value if needed.
1618 switch (VA.getLocInfo()) {
1619 default: assert(0 && "Unknown loc info!");
1620 case CCValAssign::Full: break;
1621 case CCValAssign::SExt:
1622 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1624 case CCValAssign::ZExt:
1625 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1627 case CCValAssign::AExt:
1628 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1632 if (VA.isRegLoc()) {
1634 MVT RegVT = VA.getLocVT();
1635 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1636 switch (VA.getLocReg()) {
1639 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1641 // Special case: passing MMX values in GPR registers.
1642 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1645 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1646 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1647 // Special case: passing MMX values in XMM registers.
1648 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1649 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1650 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1655 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1657 if (!IsTailCall || (IsTailCall && isByVal)) {
1658 assert(VA.isMemLoc());
1659 if (StackPtr.getNode() == 0)
1660 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1662 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1663 Chain, Arg, Flags));
1668 if (!MemOpChains.empty())
1669 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1670 &MemOpChains[0], MemOpChains.size());
1672 // Build a sequence of copy-to-reg nodes chained together with token chain
1673 // and flag operands which copy the outgoing args into registers.
1675 // Tail call byval lowering might overwrite argument registers so in case of
1676 // tail call optimization the copies to registers are lowered later.
1678 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1679 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1680 RegsToPass[i].second, InFlag);
1681 InFlag = Chain.getValue(1);
1684 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1686 if (CallRequiresGOTPtrInReg(Is64Bit, IsTailCall)) {
1687 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1688 DAG.getNode(X86ISD::GlobalBaseReg,
1689 DebugLoc::getUnknownLoc(),
1692 InFlag = Chain.getValue(1);
1694 // If we are tail calling and generating PIC/GOT style code load the address
1695 // of the callee into ecx. The value in ecx is used as target of the tail
1696 // jump. This is done to circumvent the ebx/callee-saved problem for tail
1697 // calls on PIC/GOT architectures. Normally we would just put the address of
1698 // GOT into ebx and then call target@PLT. But for tail callss ebx would be
1699 // restored (since ebx is callee saved) before jumping to the target@PLT.
1700 if (CallRequiresFnAddressInReg(Is64Bit, IsTailCall)) {
1701 // Note: The actual moving to ecx is done further down.
1702 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1703 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1704 !G->getGlobal()->hasProtectedVisibility())
1705 Callee = LowerGlobalAddress(Callee, DAG);
1706 else if (isa<ExternalSymbolSDNode>(Callee))
1707 Callee = LowerExternalSymbol(Callee,DAG);
1710 if (Is64Bit && isVarArg) {
1711 // From AMD64 ABI document:
1712 // For calls that may call functions that use varargs or stdargs
1713 // (prototype-less calls or calls to functions containing ellipsis (...) in
1714 // the declaration) %al is used as hidden argument to specify the number
1715 // of SSE registers used. The contents of %al do not need to match exactly
1716 // the number of registers, but must be an ubound on the number of SSE
1717 // registers used and is in the range 0 - 8 inclusive.
1719 // FIXME: Verify this on Win64
1720 // Count the number of XMM registers allocated.
1721 static const unsigned XMMArgRegs[] = {
1722 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1723 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1725 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1726 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1727 && "SSE registers cannot be used when SSE is disabled");
1729 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1730 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1731 InFlag = Chain.getValue(1);
1735 // For tail calls lower the arguments to the 'real' stack slot.
1737 SmallVector<SDValue, 8> MemOpChains2;
1740 // Do not flag preceeding copytoreg stuff together with the following stuff.
1742 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1743 CCValAssign &VA = ArgLocs[i];
1744 if (!VA.isRegLoc()) {
1745 assert(VA.isMemLoc());
1746 SDValue Arg = TheCall->getArg(i);
1747 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1748 // Create frame index.
1749 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1750 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1751 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1752 FIN = DAG.getFrameIndex(FI, getPointerTy());
1754 if (Flags.isByVal()) {
1755 // Copy relative to framepointer.
1756 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1757 if (StackPtr.getNode() == 0)
1758 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1760 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1762 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1765 // Store relative to framepointer.
1766 MemOpChains2.push_back(
1767 DAG.getStore(Chain, dl, Arg, FIN,
1768 PseudoSourceValue::getFixedStack(FI), 0));
1773 if (!MemOpChains2.empty())
1774 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1775 &MemOpChains2[0], MemOpChains2.size());
1777 // Copy arguments to their registers.
1778 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1779 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1780 RegsToPass[i].second, InFlag);
1781 InFlag = Chain.getValue(1);
1785 // Store the return address to the appropriate stack slot.
1786 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1790 // If the callee is a GlobalAddress node (quite common, every direct call is)
1791 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1792 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1793 // We should use extra load for direct calls to dllimported functions in
1795 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1796 getTargetMachine(), true))
1797 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy(),
1799 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1800 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1801 } else if (IsTailCall) {
1802 unsigned Opc = Is64Bit ? X86::R9 : X86::EAX;
1804 Chain = DAG.getCopyToReg(Chain, dl,
1805 DAG.getRegister(Opc, getPointerTy()),
1807 Callee = DAG.getRegister(Opc, getPointerTy());
1808 // Add register as live out.
1809 DAG.getMachineFunction().getRegInfo().addLiveOut(Opc);
1812 // Returns a chain & a flag for retval copy to use.
1813 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1814 SmallVector<SDValue, 8> Ops;
1817 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1818 DAG.getIntPtrConstant(0, true), InFlag);
1819 InFlag = Chain.getValue(1);
1821 // Returns a chain & a flag for retval copy to use.
1822 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1826 Ops.push_back(Chain);
1827 Ops.push_back(Callee);
1830 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1832 // Add argument registers to the end of the list so that they are known live
1834 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1835 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1836 RegsToPass[i].second.getValueType()));
1838 // Add an implicit use GOT pointer in EBX.
1839 if (!IsTailCall && !Is64Bit &&
1840 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1841 Subtarget->isPICStyleGOT())
1842 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1844 // Add an implicit use of AL for x86 vararg functions.
1845 if (Is64Bit && isVarArg)
1846 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1848 if (InFlag.getNode())
1849 Ops.push_back(InFlag);
1852 assert(InFlag.getNode() &&
1853 "Flag must be set. Depend on flag being set in LowerRET");
1854 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1855 TheCall->getVTList(), &Ops[0], Ops.size());
1857 return SDValue(Chain.getNode(), Op.getResNo());
1860 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1861 InFlag = Chain.getValue(1);
1863 // Create the CALLSEQ_END node.
1864 unsigned NumBytesForCalleeToPush;
1865 if (IsCalleePop(isVarArg, CC))
1866 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
1867 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
1868 // If this is is a call to a struct-return function, the callee
1869 // pops the hidden struct pointer, so we have to push it back.
1870 // This is common for Darwin/X86, Linux & Mingw32 targets.
1871 NumBytesForCalleeToPush = 4;
1873 NumBytesForCalleeToPush = 0; // Callee pops nothing.
1875 // Returns a flag for retval copy to use.
1876 Chain = DAG.getCALLSEQ_END(Chain,
1877 DAG.getIntPtrConstant(NumBytes, true),
1878 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
1881 InFlag = Chain.getValue(1);
1883 // Handle result values, copying them out of physregs into vregs that we
1885 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1890 //===----------------------------------------------------------------------===//
1891 // Fast Calling Convention (tail call) implementation
1892 //===----------------------------------------------------------------------===//
1894 // Like std call, callee cleans arguments, convention except that ECX is
1895 // reserved for storing the tail called function address. Only 2 registers are
1896 // free for argument passing (inreg). Tail call optimization is performed
1898 // * tailcallopt is enabled
1899 // * caller/callee are fastcc
1900 // On X86_64 architecture with GOT-style position independent code only local
1901 // (within module) calls are supported at the moment.
1902 // To keep the stack aligned according to platform abi the function
1903 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
1904 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
1905 // If a tail called function callee has more arguments than the caller the
1906 // caller needs to make sure that there is room to move the RETADDR to. This is
1907 // achieved by reserving an area the size of the argument delta right after the
1908 // original REtADDR, but before the saved framepointer or the spilled registers
1909 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
1921 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
1922 /// for a 16 byte align requirement.
1923 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
1924 SelectionDAG& DAG) {
1925 MachineFunction &MF = DAG.getMachineFunction();
1926 const TargetMachine &TM = MF.getTarget();
1927 const TargetFrameInfo &TFI = *TM.getFrameInfo();
1928 unsigned StackAlignment = TFI.getStackAlignment();
1929 uint64_t AlignMask = StackAlignment - 1;
1930 int64_t Offset = StackSize;
1931 uint64_t SlotSize = TD->getPointerSize();
1932 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
1933 // Number smaller than 12 so just add the difference.
1934 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
1936 // Mask out lower bits, add stackalignment once plus the 12 bytes.
1937 Offset = ((~AlignMask) & Offset) + StackAlignment +
1938 (StackAlignment-SlotSize);
1943 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1944 /// following the call is a return. A function is eligible if caller/callee
1945 /// calling conventions match, currently only fastcc supports tail calls, and
1946 /// the function CALL is immediatly followed by a RET.
1947 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1949 SelectionDAG& DAG) const {
1950 if (!PerformTailCallOpt)
1953 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1954 MachineFunction &MF = DAG.getMachineFunction();
1955 unsigned CallerCC = MF.getFunction()->getCallingConv();
1956 unsigned CalleeCC= TheCall->getCallingConv();
1957 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1958 SDValue Callee = TheCall->getCallee();
1959 // On x86/32Bit PIC/GOT tail calls are supported.
1960 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ ||
1961 !Subtarget->isPICStyleGOT()|| !Subtarget->is64Bit())
1964 // Can only do local tail calls (in same module, hidden or protected) on
1965 // x86_64 PIC/GOT at the moment.
1966 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1967 return G->getGlobal()->hasHiddenVisibility()
1968 || G->getGlobal()->hasProtectedVisibility();
1976 X86TargetLowering::createFastISel(MachineFunction &mf,
1977 MachineModuleInfo *mmo,
1979 DenseMap<const Value *, unsigned> &vm,
1980 DenseMap<const BasicBlock *,
1981 MachineBasicBlock *> &bm,
1982 DenseMap<const AllocaInst *, int> &am
1984 , SmallSet<Instruction*, 8> &cil
1987 return X86::createFastISel(mf, mmo, dw, vm, bm, am
1995 //===----------------------------------------------------------------------===//
1996 // Other Lowering Hooks
1997 //===----------------------------------------------------------------------===//
2000 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2001 MachineFunction &MF = DAG.getMachineFunction();
2002 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2003 int ReturnAddrIndex = FuncInfo->getRAIndex();
2005 if (ReturnAddrIndex == 0) {
2006 // Set up a frame object for the return address.
2007 uint64_t SlotSize = TD->getPointerSize();
2008 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2009 FuncInfo->setRAIndex(ReturnAddrIndex);
2012 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2016 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2017 /// specific condition code, returning the condition code and the LHS/RHS of the
2018 /// comparison to make.
2019 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2020 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2022 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2023 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2024 // X > -1 -> X == 0, jump !sign.
2025 RHS = DAG.getConstant(0, RHS.getValueType());
2026 return X86::COND_NS;
2027 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2028 // X < 0 -> X == 0, jump on sign.
2030 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2032 RHS = DAG.getConstant(0, RHS.getValueType());
2033 return X86::COND_LE;
2037 switch (SetCCOpcode) {
2038 default: assert(0 && "Invalid integer condition!");
2039 case ISD::SETEQ: return X86::COND_E;
2040 case ISD::SETGT: return X86::COND_G;
2041 case ISD::SETGE: return X86::COND_GE;
2042 case ISD::SETLT: return X86::COND_L;
2043 case ISD::SETLE: return X86::COND_LE;
2044 case ISD::SETNE: return X86::COND_NE;
2045 case ISD::SETULT: return X86::COND_B;
2046 case ISD::SETUGT: return X86::COND_A;
2047 case ISD::SETULE: return X86::COND_BE;
2048 case ISD::SETUGE: return X86::COND_AE;
2052 // First determine if it is required or is profitable to flip the operands.
2054 // If LHS is a foldable load, but RHS is not, flip the condition.
2055 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2056 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2057 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2058 std::swap(LHS, RHS);
2061 switch (SetCCOpcode) {
2067 std::swap(LHS, RHS);
2071 // On a floating point condition, the flags are set as follows:
2073 // 0 | 0 | 0 | X > Y
2074 // 0 | 0 | 1 | X < Y
2075 // 1 | 0 | 0 | X == Y
2076 // 1 | 1 | 1 | unordered
2077 switch (SetCCOpcode) {
2078 default: assert(0 && "Condcode should be pre-legalized away");
2080 case ISD::SETEQ: return X86::COND_E;
2081 case ISD::SETOLT: // flipped
2083 case ISD::SETGT: return X86::COND_A;
2084 case ISD::SETOLE: // flipped
2086 case ISD::SETGE: return X86::COND_AE;
2087 case ISD::SETUGT: // flipped
2089 case ISD::SETLT: return X86::COND_B;
2090 case ISD::SETUGE: // flipped
2092 case ISD::SETLE: return X86::COND_BE;
2094 case ISD::SETNE: return X86::COND_NE;
2095 case ISD::SETUO: return X86::COND_P;
2096 case ISD::SETO: return X86::COND_NP;
2100 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2101 /// code. Current x86 isa includes the following FP cmov instructions:
2102 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2103 static bool hasFPCMov(unsigned X86CC) {
2119 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2120 /// the specified range (L, H].
2121 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2122 return (Val < 0) || (Val >= Low && Val < Hi);
2125 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2126 /// specified value.
2127 static bool isUndefOrEqual(int Val, int CmpVal) {
2128 if (Val < 0 || Val == CmpVal)
2133 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2134 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2135 /// the second operand.
2136 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2137 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2138 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2139 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2140 return (Mask[0] < 2 && Mask[1] < 2);
2144 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2145 SmallVector<int, 8> M;
2147 return ::isPSHUFDMask(M, N->getValueType(0));
2150 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2151 /// is suitable for input to PSHUFHW.
2152 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2153 if (VT != MVT::v8i16)
2156 // Lower quadword copied in order or undef.
2157 for (int i = 0; i != 4; ++i)
2158 if (Mask[i] >= 0 && Mask[i] != i)
2161 // Upper quadword shuffled.
2162 for (int i = 4; i != 8; ++i)
2163 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2169 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2170 SmallVector<int, 8> M;
2172 return ::isPSHUFHWMask(M, N->getValueType(0));
2175 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2176 /// is suitable for input to PSHUFLW.
2177 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2178 if (VT != MVT::v8i16)
2181 // Upper quadword copied in order.
2182 for (int i = 4; i != 8; ++i)
2183 if (Mask[i] >= 0 && Mask[i] != i)
2186 // Lower quadword shuffled.
2187 for (int i = 0; i != 4; ++i)
2194 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2195 SmallVector<int, 8> M;
2197 return ::isPSHUFLWMask(M, N->getValueType(0));
2200 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2201 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2202 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2203 int NumElems = VT.getVectorNumElements();
2204 if (NumElems != 2 && NumElems != 4)
2207 int Half = NumElems / 2;
2208 for (int i = 0; i < Half; ++i)
2209 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2211 for (int i = Half; i < NumElems; ++i)
2212 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2218 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2219 SmallVector<int, 8> M;
2221 return ::isSHUFPMask(M, N->getValueType(0));
2224 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2225 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2226 /// half elements to come from vector 1 (which would equal the dest.) and
2227 /// the upper half to come from vector 2.
2228 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2229 int NumElems = VT.getVectorNumElements();
2231 if (NumElems != 2 && NumElems != 4)
2234 int Half = NumElems / 2;
2235 for (int i = 0; i < Half; ++i)
2236 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2238 for (int i = Half; i < NumElems; ++i)
2239 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2244 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2245 SmallVector<int, 8> M;
2247 return isCommutedSHUFPMask(M, N->getValueType(0));
2250 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2251 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2252 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2253 if (N->getValueType(0).getVectorNumElements() != 4)
2256 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2257 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2258 isUndefOrEqual(N->getMaskElt(1), 7) &&
2259 isUndefOrEqual(N->getMaskElt(2), 2) &&
2260 isUndefOrEqual(N->getMaskElt(3), 3);
2263 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2264 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2265 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2266 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2268 if (NumElems != 2 && NumElems != 4)
2271 for (unsigned i = 0; i < NumElems/2; ++i)
2272 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2275 for (unsigned i = NumElems/2; i < NumElems; ++i)
2276 if (!isUndefOrEqual(N->getMaskElt(i), i))
2282 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2283 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2285 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2286 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2288 if (NumElems != 2 && NumElems != 4)
2291 for (unsigned i = 0; i < NumElems/2; ++i)
2292 if (!isUndefOrEqual(N->getMaskElt(i), i))
2295 for (unsigned i = 0; i < NumElems/2; ++i)
2296 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2302 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2303 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2305 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2306 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2311 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2312 isUndefOrEqual(N->getMaskElt(1), 3) &&
2313 isUndefOrEqual(N->getMaskElt(2), 2) &&
2314 isUndefOrEqual(N->getMaskElt(3), 3);
2317 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2318 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2319 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2320 bool V2IsSplat = false) {
2321 int NumElts = VT.getVectorNumElements();
2322 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2325 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2327 int BitI1 = Mask[i+1];
2328 if (!isUndefOrEqual(BitI, j))
2331 if (!isUndefOrEqual(BitI1, NumElts))
2334 if (!isUndefOrEqual(BitI1, j + NumElts))
2341 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2342 SmallVector<int, 8> M;
2344 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2347 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2348 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2349 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, MVT VT,
2350 bool V2IsSplat = false) {
2351 int NumElts = VT.getVectorNumElements();
2352 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2355 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2357 int BitI1 = Mask[i+1];
2358 if (!isUndefOrEqual(BitI, j + NumElts/2))
2361 if (isUndefOrEqual(BitI1, NumElts))
2364 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2371 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2372 SmallVector<int, 8> M;
2374 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2377 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2378 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2380 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2381 int NumElems = VT.getVectorNumElements();
2382 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2385 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2387 int BitI1 = Mask[i+1];
2388 if (!isUndefOrEqual(BitI, j))
2390 if (!isUndefOrEqual(BitI1, j))
2396 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2397 SmallVector<int, 8> M;
2399 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2402 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2403 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2405 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2406 int NumElems = VT.getVectorNumElements();
2407 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2410 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2412 int BitI1 = Mask[i+1];
2413 if (!isUndefOrEqual(BitI, j))
2415 if (!isUndefOrEqual(BitI1, j))
2421 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2422 SmallVector<int, 8> M;
2424 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2427 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2428 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2429 /// MOVSD, and MOVD, i.e. setting the lowest element.
2430 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2431 if (VT.getVectorElementType().getSizeInBits() < 32)
2434 int NumElts = VT.getVectorNumElements();
2436 if (!isUndefOrEqual(Mask[0], NumElts))
2439 for (int i = 1; i < NumElts; ++i)
2440 if (!isUndefOrEqual(Mask[i], i))
2446 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2447 SmallVector<int, 8> M;
2449 return ::isMOVLMask(M, N->getValueType(0));
2452 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2453 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2454 /// element of vector 2 and the other elements to come from vector 1 in order.
2455 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2456 bool V2IsSplat = false, bool V2IsUndef = false) {
2457 int NumOps = VT.getVectorNumElements();
2458 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2461 if (!isUndefOrEqual(Mask[0], 0))
2464 for (int i = 1; i < NumOps; ++i)
2465 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2466 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2467 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2473 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2474 bool V2IsUndef = false) {
2475 SmallVector<int, 8> M;
2477 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2480 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2481 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2482 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2483 if (N->getValueType(0).getVectorNumElements() != 4)
2486 // Expect 1, 1, 3, 3
2487 for (unsigned i = 0; i < 2; ++i) {
2488 int Elt = N->getMaskElt(i);
2489 if (Elt >= 0 && Elt != 1)
2494 for (unsigned i = 2; i < 4; ++i) {
2495 int Elt = N->getMaskElt(i);
2496 if (Elt >= 0 && Elt != 3)
2501 // Don't use movshdup if it can be done with a shufps.
2502 // FIXME: verify that matching u, u, 3, 3 is what we want.
2506 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2507 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2508 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2509 if (N->getValueType(0).getVectorNumElements() != 4)
2512 // Expect 0, 0, 2, 2
2513 for (unsigned i = 0; i < 2; ++i)
2514 if (N->getMaskElt(i) > 0)
2518 for (unsigned i = 2; i < 4; ++i) {
2519 int Elt = N->getMaskElt(i);
2520 if (Elt >= 0 && Elt != 2)
2525 // Don't use movsldup if it can be done with a shufps.
2529 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2530 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2531 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2532 int e = N->getValueType(0).getVectorNumElements() / 2;
2534 for (int i = 0; i < e; ++i)
2535 if (!isUndefOrEqual(N->getMaskElt(i), i))
2537 for (int i = 0; i < e; ++i)
2538 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2543 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2544 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2546 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2547 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2548 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2550 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2552 for (int i = 0; i < NumOperands; ++i) {
2553 int Val = SVOp->getMaskElt(NumOperands-i-1);
2554 if (Val < 0) Val = 0;
2555 if (Val >= NumOperands) Val -= NumOperands;
2557 if (i != NumOperands - 1)
2563 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2564 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2566 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2567 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2569 // 8 nodes, but we only care about the last 4.
2570 for (unsigned i = 7; i >= 4; --i) {
2571 int Val = SVOp->getMaskElt(i);
2580 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2581 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2583 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2584 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2586 // 8 nodes, but we only care about the first 4.
2587 for (int i = 3; i >= 0; --i) {
2588 int Val = SVOp->getMaskElt(i);
2597 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2598 /// their permute mask.
2599 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2600 SelectionDAG &DAG) {
2601 MVT VT = SVOp->getValueType(0);
2602 unsigned NumElems = VT.getVectorNumElements();
2603 SmallVector<int, 8> MaskVec;
2605 for (unsigned i = 0; i != NumElems; ++i) {
2606 int idx = SVOp->getMaskElt(i);
2608 MaskVec.push_back(idx);
2609 else if (idx < (int)NumElems)
2610 MaskVec.push_back(idx + NumElems);
2612 MaskVec.push_back(idx - NumElems);
2614 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2615 SVOp->getOperand(0), &MaskVec[0]);
2618 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2619 /// the two vector operands have swapped position.
2620 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
2621 unsigned NumElems = VT.getVectorNumElements();
2622 for (unsigned i = 0; i != NumElems; ++i) {
2626 else if (idx < (int)NumElems)
2627 Mask[i] = idx + NumElems;
2629 Mask[i] = idx - NumElems;
2633 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2634 /// match movhlps. The lower half elements should come from upper half of
2635 /// V1 (and in order), and the upper half elements should come from the upper
2636 /// half of V2 (and in order).
2637 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2638 if (Op->getValueType(0).getVectorNumElements() != 4)
2640 for (unsigned i = 0, e = 2; i != e; ++i)
2641 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2643 for (unsigned i = 2; i != 4; ++i)
2644 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2649 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2650 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2652 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2653 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2655 N = N->getOperand(0).getNode();
2656 if (!ISD::isNON_EXTLoad(N))
2659 *LD = cast<LoadSDNode>(N);
2663 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2664 /// match movlp{s|d}. The lower half elements should come from lower half of
2665 /// V1 (and in order), and the upper half elements should come from the upper
2666 /// half of V2 (and in order). And since V1 will become the source of the
2667 /// MOVLP, it must be either a vector load or a scalar load to vector.
2668 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2669 ShuffleVectorSDNode *Op) {
2670 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2672 // Is V2 is a vector load, don't do this transformation. We will try to use
2673 // load folding shufps op.
2674 if (ISD::isNON_EXTLoad(V2))
2677 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2679 if (NumElems != 2 && NumElems != 4)
2681 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2682 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2684 for (unsigned i = NumElems/2; i != NumElems; ++i)
2685 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2690 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2692 static bool isSplatVector(SDNode *N) {
2693 if (N->getOpcode() != ISD::BUILD_VECTOR)
2696 SDValue SplatValue = N->getOperand(0);
2697 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2698 if (N->getOperand(i) != SplatValue)
2703 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2705 static inline bool isZeroNode(SDValue Elt) {
2706 return ((isa<ConstantSDNode>(Elt) &&
2707 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2708 (isa<ConstantFPSDNode>(Elt) &&
2709 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2712 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2713 /// to an zero vector.
2714 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2715 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2716 SDValue V1 = N->getOperand(0);
2717 SDValue V2 = N->getOperand(1);
2718 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2719 for (unsigned i = 0; i != NumElems; ++i) {
2720 int Idx = N->getMaskElt(i);
2721 if (Idx >= (int)NumElems) {
2722 unsigned Opc = V2.getOpcode();
2723 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2725 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V2.getOperand(Idx-NumElems)))
2727 } else if (Idx >= 0) {
2728 unsigned Opc = V1.getOpcode();
2729 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2731 if (Opc != ISD::BUILD_VECTOR || !isZeroNode(V1.getOperand(Idx)))
2738 /// getZeroVector - Returns a vector of specified type with all zero elements.
2740 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2742 assert(VT.isVector() && "Expected a vector type");
2744 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2745 // type. This ensures they get CSE'd.
2747 if (VT.getSizeInBits() == 64) { // MMX
2748 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2749 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2750 } else if (HasSSE2) { // SSE2
2751 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2752 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2754 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2755 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2757 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2760 /// getOnesVector - Returns a vector of specified type with all bits set.
2762 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2763 assert(VT.isVector() && "Expected a vector type");
2765 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2766 // type. This ensures they get CSE'd.
2767 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2769 if (VT.getSizeInBits() == 64) // MMX
2770 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2772 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2773 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2777 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2778 /// that point to V2 points to its first element.
2779 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2780 MVT VT = SVOp->getValueType(0);
2781 unsigned NumElems = VT.getVectorNumElements();
2783 bool Changed = false;
2784 SmallVector<int, 8> MaskVec;
2785 SVOp->getMask(MaskVec);
2787 for (unsigned i = 0; i != NumElems; ++i) {
2788 if (MaskVec[i] > (int)NumElems) {
2789 MaskVec[i] = NumElems;
2794 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2795 SVOp->getOperand(1), &MaskVec[0]);
2796 return SDValue(SVOp, 0);
2799 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2800 /// operation of specified width.
2801 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2803 unsigned NumElems = VT.getVectorNumElements();
2804 SmallVector<int, 8> Mask;
2805 Mask.push_back(NumElems);
2806 for (unsigned i = 1; i != NumElems; ++i)
2808 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2811 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2812 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2814 unsigned NumElems = VT.getVectorNumElements();
2815 SmallVector<int, 8> Mask;
2816 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2818 Mask.push_back(i + NumElems);
2820 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2823 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2824 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2826 unsigned NumElems = VT.getVectorNumElements();
2827 unsigned Half = NumElems/2;
2828 SmallVector<int, 8> Mask;
2829 for (unsigned i = 0; i != Half; ++i) {
2830 Mask.push_back(i + Half);
2831 Mask.push_back(i + NumElems + Half);
2833 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2836 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2837 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2839 if (SV->getValueType(0).getVectorNumElements() <= 4)
2840 return SDValue(SV, 0);
2842 MVT PVT = MVT::v4f32;
2843 MVT VT = SV->getValueType(0);
2844 DebugLoc dl = SV->getDebugLoc();
2845 SDValue V1 = SV->getOperand(0);
2846 int NumElems = VT.getVectorNumElements();
2847 int EltNo = SV->getSplatIndex();
2849 // unpack elements to the correct location
2850 while (NumElems > 4) {
2851 if (EltNo < NumElems/2) {
2852 V1 = getUnpackl(DAG, dl, VT, V1, V1);
2854 V1 = getUnpackh(DAG, dl, VT, V1, V1);
2855 EltNo -= NumElems/2;
2860 // Perform the splat.
2861 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
2862 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2863 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
2864 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2867 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2868 /// vector of zero or undef vector. This produces a shuffle where the low
2869 /// element of V2 is swizzled into the zero/undef vector, landing at element
2870 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2871 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2872 bool isZero, bool HasSSE2,
2873 SelectionDAG &DAG) {
2874 MVT VT = V2.getValueType();
2876 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
2877 unsigned NumElems = VT.getVectorNumElements();
2878 SmallVector<int, 16> MaskVec;
2879 for (unsigned i = 0; i != NumElems; ++i)
2880 // If this is the insertion idx, put the low elt of V2 here.
2881 MaskVec.push_back(i == Idx ? NumElems : i);
2882 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
2885 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
2886 /// a shuffle that is zero.
2888 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
2889 bool Low, SelectionDAG &DAG) {
2890 unsigned NumZeros = 0;
2891 for (int i = 0; i < NumElems; ++i) {
2892 unsigned Index = Low ? i : NumElems-i-1;
2893 int Idx = SVOp->getMaskElt(Index);
2898 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
2899 if (Elt.getNode() && isZeroNode(Elt))
2907 /// isVectorShift - Returns true if the shuffle can be implemented as a
2908 /// logical left or right shift of a vector.
2909 /// FIXME: split into pslldqi, psrldqi, palignr variants.
2910 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
2911 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
2912 int NumElems = SVOp->getValueType(0).getVectorNumElements();
2915 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
2918 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
2922 bool SeenV1 = false;
2923 bool SeenV2 = false;
2924 for (int i = NumZeros; i < NumElems; ++i) {
2925 int Val = isLeft ? (i - NumZeros) : i;
2926 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
2938 if (SeenV1 && SeenV2)
2941 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
2947 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2949 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
2950 unsigned NumNonZero, unsigned NumZero,
2951 SelectionDAG &DAG, TargetLowering &TLI) {
2955 DebugLoc dl = Op.getDebugLoc();
2958 for (unsigned i = 0; i < 16; ++i) {
2959 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2960 if (ThisIsNonZero && First) {
2962 V = getZeroVector(MVT::v8i16, true, DAG, dl);
2964 V = DAG.getUNDEF(MVT::v8i16);
2969 SDValue ThisElt(0, 0), LastElt(0, 0);
2970 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2971 if (LastIsNonZero) {
2972 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
2973 MVT::i16, Op.getOperand(i-1));
2975 if (ThisIsNonZero) {
2976 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
2977 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
2978 ThisElt, DAG.getConstant(8, MVT::i8));
2980 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
2984 if (ThisElt.getNode())
2985 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
2986 DAG.getIntPtrConstant(i/2));
2990 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
2993 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
2995 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
2996 unsigned NumNonZero, unsigned NumZero,
2997 SelectionDAG &DAG, TargetLowering &TLI) {
3001 DebugLoc dl = Op.getDebugLoc();
3004 for (unsigned i = 0; i < 8; ++i) {
3005 bool isNonZero = (NonZeros & (1 << i)) != 0;
3009 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3011 V = DAG.getUNDEF(MVT::v8i16);
3014 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3015 MVT::v8i16, V, Op.getOperand(i),
3016 DAG.getIntPtrConstant(i));
3023 /// getVShift - Return a vector logical shift node.
3025 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3026 unsigned NumBits, SelectionDAG &DAG,
3027 const TargetLowering &TLI, DebugLoc dl) {
3028 bool isMMX = VT.getSizeInBits() == 64;
3029 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3030 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3031 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3032 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3033 DAG.getNode(Opc, dl, ShVT, SrcOp,
3034 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3038 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3039 DebugLoc dl = Op.getDebugLoc();
3040 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3041 if (ISD::isBuildVectorAllZeros(Op.getNode())
3042 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3043 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3044 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3045 // eliminated on x86-32 hosts.
3046 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3049 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3050 return getOnesVector(Op.getValueType(), DAG, dl);
3051 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3054 MVT VT = Op.getValueType();
3055 MVT EVT = VT.getVectorElementType();
3056 unsigned EVTBits = EVT.getSizeInBits();
3058 unsigned NumElems = Op.getNumOperands();
3059 unsigned NumZero = 0;
3060 unsigned NumNonZero = 0;
3061 unsigned NonZeros = 0;
3062 bool IsAllConstants = true;
3063 SmallSet<SDValue, 8> Values;
3064 for (unsigned i = 0; i < NumElems; ++i) {
3065 SDValue Elt = Op.getOperand(i);
3066 if (Elt.getOpcode() == ISD::UNDEF)
3069 if (Elt.getOpcode() != ISD::Constant &&
3070 Elt.getOpcode() != ISD::ConstantFP)
3071 IsAllConstants = false;
3072 if (isZeroNode(Elt))
3075 NonZeros |= (1 << i);
3080 if (NumNonZero == 0) {
3081 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3082 return DAG.getUNDEF(VT);
3085 // Special case for single non-zero, non-undef, element.
3086 if (NumNonZero == 1) {
3087 unsigned Idx = CountTrailingZeros_32(NonZeros);
3088 SDValue Item = Op.getOperand(Idx);
3090 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3091 // the value are obviously zero, truncate the value to i32 and do the
3092 // insertion that way. Only do this if the value is non-constant or if the
3093 // value is a constant being inserted into element 0. It is cheaper to do
3094 // a constant pool load than it is to do a movd + shuffle.
3095 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3096 (!IsAllConstants || Idx == 0)) {
3097 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3098 // Handle MMX and SSE both.
3099 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3100 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3102 // Truncate the value (which may itself be a constant) to i32, and
3103 // convert it to a vector with movd (S2V+shuffle to zero extend).
3104 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3105 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3106 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3107 Subtarget->hasSSE2(), DAG);
3109 // Now we have our 32-bit value zero extended in the low element of
3110 // a vector. If Idx != 0, swizzle it into place.
3112 SmallVector<int, 4> Mask;
3113 Mask.push_back(Idx);
3114 for (unsigned i = 1; i != VecElts; ++i)
3116 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3117 DAG.getUNDEF(Item.getValueType()),
3120 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3124 // If we have a constant or non-constant insertion into the low element of
3125 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3126 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3127 // depending on what the source datatype is.
3130 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3131 } else if (EVT == MVT::i32 || EVT == MVT::f32 || EVT == MVT::f64 ||
3132 (EVT == MVT::i64 && Subtarget->is64Bit())) {
3133 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3134 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3135 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3137 } else if (EVT == MVT::i16 || EVT == MVT::i8) {
3138 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3139 MVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3140 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3141 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3142 Subtarget->hasSSE2(), DAG);
3143 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3147 // Is it a vector logical left shift?
3148 if (NumElems == 2 && Idx == 1 &&
3149 isZeroNode(Op.getOperand(0)) && !isZeroNode(Op.getOperand(1))) {
3150 unsigned NumBits = VT.getSizeInBits();
3151 return getVShift(true, VT,
3152 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3153 VT, Op.getOperand(1)),
3154 NumBits/2, DAG, *this, dl);
3157 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3160 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3161 // is a non-constant being inserted into an element other than the low one,
3162 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3163 // movd/movss) to move this into the low element, then shuffle it into
3165 if (EVTBits == 32) {
3166 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3168 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3169 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3170 Subtarget->hasSSE2(), DAG);
3171 SmallVector<int, 8> MaskVec;
3172 for (unsigned i = 0; i < NumElems; i++)
3173 MaskVec.push_back(i == Idx ? 0 : 1);
3174 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3178 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3179 if (Values.size() == 1)
3182 // A vector full of immediates; various special cases are already
3183 // handled, so this is best done with a single constant-pool load.
3187 // Let legalizer expand 2-wide build_vectors.
3188 if (EVTBits == 64) {
3189 if (NumNonZero == 1) {
3190 // One half is zero or undef.
3191 unsigned Idx = CountTrailingZeros_32(NonZeros);
3192 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3193 Op.getOperand(Idx));
3194 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3195 Subtarget->hasSSE2(), DAG);
3200 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3201 if (EVTBits == 8 && NumElems == 16) {
3202 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3204 if (V.getNode()) return V;
3207 if (EVTBits == 16 && NumElems == 8) {
3208 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3210 if (V.getNode()) return V;
3213 // If element VT is == 32 bits, turn it into a number of shuffles.
3214 SmallVector<SDValue, 8> V;
3216 if (NumElems == 4 && NumZero > 0) {
3217 for (unsigned i = 0; i < 4; ++i) {
3218 bool isZero = !(NonZeros & (1 << i));
3220 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3222 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3225 for (unsigned i = 0; i < 2; ++i) {
3226 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3229 V[i] = V[i*2]; // Must be a zero vector.
3232 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3235 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3238 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3243 SmallVector<int, 8> MaskVec;
3244 bool Reverse = (NonZeros & 0x3) == 2;
3245 for (unsigned i = 0; i < 2; ++i)
3246 MaskVec.push_back(Reverse ? 1-i : i);
3247 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3248 for (unsigned i = 0; i < 2; ++i)
3249 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3250 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3253 if (Values.size() > 2) {
3254 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3255 // values to be inserted is equal to the number of elements, in which case
3256 // use the unpack code below in the hopes of matching the consecutive elts
3257 // load merge pattern for shuffles.
3258 // FIXME: We could probably just check that here directly.
3259 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3260 getSubtarget()->hasSSE41()) {
3261 V[0] = DAG.getUNDEF(VT);
3262 for (unsigned i = 0; i < NumElems; ++i)
3263 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3264 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3265 Op.getOperand(i), DAG.getIntPtrConstant(i));
3268 // Expand into a number of unpckl*.
3270 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3271 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3272 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3273 for (unsigned i = 0; i < NumElems; ++i)
3274 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3276 while (NumElems != 0) {
3277 for (unsigned i = 0; i < NumElems; ++i)
3278 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3287 // v8i16 shuffles - Prefer shuffles in the following order:
3288 // 1. [all] pshuflw, pshufhw, optional move
3289 // 2. [ssse3] 1 x pshufb
3290 // 3. [ssse3] 2 x pshufb + 1 x por
3291 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3293 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3294 SelectionDAG &DAG, X86TargetLowering &TLI) {
3295 SDValue V1 = SVOp->getOperand(0);
3296 SDValue V2 = SVOp->getOperand(1);
3297 DebugLoc dl = SVOp->getDebugLoc();
3298 SmallVector<int, 8> MaskVals;
3300 // Determine if more than 1 of the words in each of the low and high quadwords
3301 // of the result come from the same quadword of one of the two inputs. Undef
3302 // mask values count as coming from any quadword, for better codegen.
3303 SmallVector<unsigned, 4> LoQuad(4);
3304 SmallVector<unsigned, 4> HiQuad(4);
3305 BitVector InputQuads(4);
3306 for (unsigned i = 0; i < 8; ++i) {
3307 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3308 int EltIdx = SVOp->getMaskElt(i);
3309 MaskVals.push_back(EltIdx);
3318 InputQuads.set(EltIdx / 4);
3321 int BestLoQuad = -1;
3322 unsigned MaxQuad = 1;
3323 for (unsigned i = 0; i < 4; ++i) {
3324 if (LoQuad[i] > MaxQuad) {
3326 MaxQuad = LoQuad[i];
3330 int BestHiQuad = -1;
3332 for (unsigned i = 0; i < 4; ++i) {
3333 if (HiQuad[i] > MaxQuad) {
3335 MaxQuad = HiQuad[i];
3339 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3340 // of the two input vectors, shuffle them into one input vector so only a
3341 // single pshufb instruction is necessary. If There are more than 2 input
3342 // quads, disable the next transformation since it does not help SSSE3.
3343 bool V1Used = InputQuads[0] || InputQuads[1];
3344 bool V2Used = InputQuads[2] || InputQuads[3];
3345 if (TLI.getSubtarget()->hasSSSE3()) {
3346 if (InputQuads.count() == 2 && V1Used && V2Used) {
3347 BestLoQuad = InputQuads.find_first();
3348 BestHiQuad = InputQuads.find_next(BestLoQuad);
3350 if (InputQuads.count() > 2) {
3356 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3357 // the shuffle mask. If a quad is scored as -1, that means that it contains
3358 // words from all 4 input quadwords.
3360 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3361 SmallVector<int, 8> MaskV;
3362 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3363 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3364 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3365 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3366 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3367 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3369 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3370 // source words for the shuffle, to aid later transformations.
3371 bool AllWordsInNewV = true;
3372 bool InOrder[2] = { true, true };
3373 for (unsigned i = 0; i != 8; ++i) {
3374 int idx = MaskVals[i];
3376 InOrder[i/4] = false;
3377 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3379 AllWordsInNewV = false;
3383 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3384 if (AllWordsInNewV) {
3385 for (int i = 0; i != 8; ++i) {
3386 int idx = MaskVals[i];
3389 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3390 if ((idx != i) && idx < 4)
3392 if ((idx != i) && idx > 3)
3401 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3402 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3403 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3404 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3405 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3409 // If we have SSSE3, and all words of the result are from 1 input vector,
3410 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3411 // is present, fall back to case 4.
3412 if (TLI.getSubtarget()->hasSSSE3()) {
3413 SmallVector<SDValue,16> pshufbMask;
3415 // If we have elements from both input vectors, set the high bit of the
3416 // shuffle mask element to zero out elements that come from V2 in the V1
3417 // mask, and elements that come from V1 in the V2 mask, so that the two
3418 // results can be OR'd together.
3419 bool TwoInputs = V1Used && V2Used;
3420 for (unsigned i = 0; i != 8; ++i) {
3421 int EltIdx = MaskVals[i] * 2;
3422 if (TwoInputs && (EltIdx >= 16)) {
3423 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3424 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3427 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3428 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3430 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3431 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3432 DAG.getNode(ISD::BUILD_VECTOR, dl,
3433 MVT::v16i8, &pshufbMask[0], 16));
3435 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3437 // Calculate the shuffle mask for the second input, shuffle it, and
3438 // OR it with the first shuffled input.
3440 for (unsigned i = 0; i != 8; ++i) {
3441 int EltIdx = MaskVals[i] * 2;
3443 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3444 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3447 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3448 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3450 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3451 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3452 DAG.getNode(ISD::BUILD_VECTOR, dl,
3453 MVT::v16i8, &pshufbMask[0], 16));
3454 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3455 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3458 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3459 // and update MaskVals with new element order.
3460 BitVector InOrder(8);
3461 if (BestLoQuad >= 0) {
3462 SmallVector<int, 8> MaskV;
3463 for (int i = 0; i != 4; ++i) {
3464 int idx = MaskVals[i];
3466 MaskV.push_back(-1);
3468 } else if ((idx / 4) == BestLoQuad) {
3469 MaskV.push_back(idx & 3);
3472 MaskV.push_back(-1);
3475 for (unsigned i = 4; i != 8; ++i)
3477 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3481 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3482 // and update MaskVals with the new element order.
3483 if (BestHiQuad >= 0) {
3484 SmallVector<int, 8> MaskV;
3485 for (unsigned i = 0; i != 4; ++i)
3487 for (unsigned i = 4; i != 8; ++i) {
3488 int idx = MaskVals[i];
3490 MaskV.push_back(-1);
3492 } else if ((idx / 4) == BestHiQuad) {
3493 MaskV.push_back((idx & 3) + 4);
3496 MaskV.push_back(-1);
3499 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3503 // In case BestHi & BestLo were both -1, which means each quadword has a word
3504 // from each of the four input quadwords, calculate the InOrder bitvector now
3505 // before falling through to the insert/extract cleanup.
3506 if (BestLoQuad == -1 && BestHiQuad == -1) {
3508 for (int i = 0; i != 8; ++i)
3509 if (MaskVals[i] < 0 || MaskVals[i] == i)
3513 // The other elements are put in the right place using pextrw and pinsrw.
3514 for (unsigned i = 0; i != 8; ++i) {
3517 int EltIdx = MaskVals[i];
3520 SDValue ExtOp = (EltIdx < 8)
3521 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3522 DAG.getIntPtrConstant(EltIdx))
3523 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3524 DAG.getIntPtrConstant(EltIdx - 8));
3525 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3526 DAG.getIntPtrConstant(i));
3531 // v16i8 shuffles - Prefer shuffles in the following order:
3532 // 1. [ssse3] 1 x pshufb
3533 // 2. [ssse3] 2 x pshufb + 1 x por
3534 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3536 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3537 SelectionDAG &DAG, X86TargetLowering &TLI) {
3538 SDValue V1 = SVOp->getOperand(0);
3539 SDValue V2 = SVOp->getOperand(1);
3540 DebugLoc dl = SVOp->getDebugLoc();
3541 SmallVector<int, 16> MaskVals;
3542 SVOp->getMask(MaskVals);
3544 // If we have SSSE3, case 1 is generated when all result bytes come from
3545 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3546 // present, fall back to case 3.
3547 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3550 for (unsigned i = 0; i < 16; ++i) {
3551 int EltIdx = MaskVals[i];
3560 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3561 if (TLI.getSubtarget()->hasSSSE3()) {
3562 SmallVector<SDValue,16> pshufbMask;
3564 // If all result elements are from one input vector, then only translate
3565 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3567 // Otherwise, we have elements from both input vectors, and must zero out
3568 // elements that come from V2 in the first mask, and V1 in the second mask
3569 // so that we can OR them together.
3570 bool TwoInputs = !(V1Only || V2Only);
3571 for (unsigned i = 0; i != 16; ++i) {
3572 int EltIdx = MaskVals[i];
3573 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3574 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3577 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3579 // If all the elements are from V2, assign it to V1 and return after
3580 // building the first pshufb.
3583 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3584 DAG.getNode(ISD::BUILD_VECTOR, dl,
3585 MVT::v16i8, &pshufbMask[0], 16));
3589 // Calculate the shuffle mask for the second input, shuffle it, and
3590 // OR it with the first shuffled input.
3592 for (unsigned i = 0; i != 16; ++i) {
3593 int EltIdx = MaskVals[i];
3595 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3598 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3600 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3601 DAG.getNode(ISD::BUILD_VECTOR, dl,
3602 MVT::v16i8, &pshufbMask[0], 16));
3603 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3606 // No SSSE3 - Calculate in place words and then fix all out of place words
3607 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3608 // the 16 different words that comprise the two doublequadword input vectors.
3609 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3610 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3611 SDValue NewV = V2Only ? V2 : V1;
3612 for (int i = 0; i != 8; ++i) {
3613 int Elt0 = MaskVals[i*2];
3614 int Elt1 = MaskVals[i*2+1];
3616 // This word of the result is all undef, skip it.
3617 if (Elt0 < 0 && Elt1 < 0)
3620 // This word of the result is already in the correct place, skip it.
3621 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3623 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3626 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3627 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3630 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3631 // using a single extract together, load it and store it.
3632 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3633 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3634 DAG.getIntPtrConstant(Elt1 / 2));
3635 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3636 DAG.getIntPtrConstant(i));
3640 // If Elt1 is defined, extract it from the appropriate source. If the
3641 // source byte is not also odd, shift the extracted word left 8 bits
3642 // otherwise clear the bottom 8 bits if we need to do an or.
3644 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3645 DAG.getIntPtrConstant(Elt1 / 2));
3646 if ((Elt1 & 1) == 0)
3647 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3648 DAG.getConstant(8, TLI.getShiftAmountTy()));
3650 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3651 DAG.getConstant(0xFF00, MVT::i16));
3653 // If Elt0 is defined, extract it from the appropriate source. If the
3654 // source byte is not also even, shift the extracted word right 8 bits. If
3655 // Elt1 was also defined, OR the extracted values together before
3656 // inserting them in the result.
3658 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3659 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3660 if ((Elt0 & 1) != 0)
3661 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3662 DAG.getConstant(8, TLI.getShiftAmountTy()));
3664 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3665 DAG.getConstant(0x00FF, MVT::i16));
3666 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3669 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3670 DAG.getIntPtrConstant(i));
3672 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3675 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3676 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3677 /// done when every pair / quad of shuffle mask elements point to elements in
3678 /// the right sequence. e.g.
3679 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3681 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3683 TargetLowering &TLI, DebugLoc dl) {
3684 MVT VT = SVOp->getValueType(0);
3685 SDValue V1 = SVOp->getOperand(0);
3686 SDValue V2 = SVOp->getOperand(1);
3687 unsigned NumElems = VT.getVectorNumElements();
3688 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3689 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3690 MVT MaskEltVT = MaskVT.getVectorElementType();
3692 switch (VT.getSimpleVT()) {
3693 default: assert(false && "Unexpected!");
3694 case MVT::v4f32: NewVT = MVT::v2f64; break;
3695 case MVT::v4i32: NewVT = MVT::v2i64; break;
3696 case MVT::v8i16: NewVT = MVT::v4i32; break;
3697 case MVT::v16i8: NewVT = MVT::v4i32; break;
3700 if (NewWidth == 2) {
3706 int Scale = NumElems / NewWidth;
3707 SmallVector<int, 8> MaskVec;
3708 for (unsigned i = 0; i < NumElems; i += Scale) {
3710 for (int j = 0; j < Scale; ++j) {
3711 int EltIdx = SVOp->getMaskElt(i+j);
3715 StartIdx = EltIdx - (EltIdx % Scale);
3716 if (EltIdx != StartIdx + j)
3720 MaskVec.push_back(-1);
3722 MaskVec.push_back(StartIdx / Scale);
3725 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3726 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3727 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3730 /// getVZextMovL - Return a zero-extending vector move low node.
3732 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3733 SDValue SrcOp, SelectionDAG &DAG,
3734 const X86Subtarget *Subtarget, DebugLoc dl) {
3735 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3736 LoadSDNode *LD = NULL;
3737 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3738 LD = dyn_cast<LoadSDNode>(SrcOp);
3740 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3742 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3743 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3744 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3745 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3746 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3748 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3749 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3750 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3751 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3759 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3760 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3761 DAG.getNode(ISD::BIT_CONVERT, dl,
3765 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3768 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3769 SDValue V1 = SVOp->getOperand(0);
3770 SDValue V2 = SVOp->getOperand(1);
3771 DebugLoc dl = SVOp->getDebugLoc();
3772 MVT VT = SVOp->getValueType(0);
3774 SmallVector<std::pair<int, int>, 8> Locs;
3776 SmallVector<int, 8> Mask1(4U, -1);
3777 SmallVector<int, 8> PermMask;
3778 SVOp->getMask(PermMask);
3782 for (unsigned i = 0; i != 4; ++i) {
3783 int Idx = PermMask[i];
3785 Locs[i] = std::make_pair(-1, -1);
3787 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3789 Locs[i] = std::make_pair(0, NumLo);
3793 Locs[i] = std::make_pair(1, NumHi);
3795 Mask1[2+NumHi] = Idx;
3801 if (NumLo <= 2 && NumHi <= 2) {
3802 // If no more than two elements come from either vector. This can be
3803 // implemented with two shuffles. First shuffle gather the elements.
3804 // The second shuffle, which takes the first shuffle as both of its
3805 // vector operands, put the elements into the right order.
3806 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3808 SmallVector<int, 8> Mask2(4U, -1);
3810 for (unsigned i = 0; i != 4; ++i) {
3811 if (Locs[i].first == -1)
3814 unsigned Idx = (i < 2) ? 0 : 4;
3815 Idx += Locs[i].first * 2 + Locs[i].second;
3820 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3821 } else if (NumLo == 3 || NumHi == 3) {
3822 // Otherwise, we must have three elements from one vector, call it X, and
3823 // one element from the other, call it Y. First, use a shufps to build an
3824 // intermediate vector with the one element from Y and the element from X
3825 // that will be in the same half in the final destination (the indexes don't
3826 // matter). Then, use a shufps to build the final vector, taking the half
3827 // containing the element from Y from the intermediate, and the other half
3830 // Normalize it so the 3 elements come from V1.
3831 CommuteVectorShuffleMask(PermMask, VT);
3835 // Find the element from V2.
3837 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3838 int Val = PermMask[HiIndex];
3845 Mask1[0] = PermMask[HiIndex];
3847 Mask1[2] = PermMask[HiIndex^1];
3849 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3852 Mask1[0] = PermMask[0];
3853 Mask1[1] = PermMask[1];
3854 Mask1[2] = HiIndex & 1 ? 6 : 4;
3855 Mask1[3] = HiIndex & 1 ? 4 : 6;
3856 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3858 Mask1[0] = HiIndex & 1 ? 2 : 0;
3859 Mask1[1] = HiIndex & 1 ? 0 : 2;
3860 Mask1[2] = PermMask[2];
3861 Mask1[3] = PermMask[3];
3866 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
3870 // Break it into (shuffle shuffle_hi, shuffle_lo).
3872 SmallVector<int,8> LoMask(4U, -1);
3873 SmallVector<int,8> HiMask(4U, -1);
3875 SmallVector<int,8> *MaskPtr = &LoMask;
3876 unsigned MaskIdx = 0;
3879 for (unsigned i = 0; i != 4; ++i) {
3886 int Idx = PermMask[i];
3888 Locs[i] = std::make_pair(-1, -1);
3889 } else if (Idx < 4) {
3890 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3891 (*MaskPtr)[LoIdx] = Idx;
3894 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3895 (*MaskPtr)[HiIdx] = Idx;
3900 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
3901 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
3902 SmallVector<int, 8> MaskOps;
3903 for (unsigned i = 0; i != 4; ++i) {
3904 if (Locs[i].first == -1) {
3905 MaskOps.push_back(-1);
3907 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
3908 MaskOps.push_back(Idx);
3911 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
3915 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3916 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
3917 SDValue V1 = Op.getOperand(0);
3918 SDValue V2 = Op.getOperand(1);
3919 MVT VT = Op.getValueType();
3920 DebugLoc dl = Op.getDebugLoc();
3921 unsigned NumElems = VT.getVectorNumElements();
3922 bool isMMX = VT.getSizeInBits() == 64;
3923 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3924 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3925 bool V1IsSplat = false;
3926 bool V2IsSplat = false;
3928 if (isZeroShuffle(SVOp))
3929 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3931 // Promote splats to v4f32.
3932 if (SVOp->isSplat()) {
3933 if (isMMX || NumElems < 4)
3935 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
3938 // If the shuffle can be profitably rewritten as a narrower shuffle, then
3940 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
3941 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
3942 if (NewOp.getNode())
3943 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3944 LowerVECTOR_SHUFFLE(NewOp, DAG));
3945 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
3946 // FIXME: Figure out a cleaner way to do this.
3947 // Try to make use of movq to zero out the top part.
3948 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
3949 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
3950 if (NewOp.getNode()) {
3951 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
3952 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
3953 DAG, Subtarget, dl);
3955 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
3956 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
3957 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
3958 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
3959 DAG, Subtarget, dl);
3963 if (X86::isPSHUFDMask(SVOp))
3966 // Check if this can be converted into a logical shift.
3967 bool isLeft = false;
3970 bool isShift = getSubtarget()->hasSSE2() &&
3971 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
3972 if (isShift && ShVal.hasOneUse()) {
3973 // If the shifted value has multiple uses, it may be cheaper to use
3974 // v_set0 + movlhps or movhlps, etc.
3975 MVT EVT = VT.getVectorElementType();
3976 ShAmt *= EVT.getSizeInBits();
3977 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
3980 if (X86::isMOVLMask(SVOp)) {
3983 if (ISD::isBuildVectorAllZeros(V1.getNode()))
3984 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
3989 // FIXME: fold these into legal mask.
3990 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
3991 X86::isMOVSLDUPMask(SVOp) ||
3992 X86::isMOVHLPSMask(SVOp) ||
3993 X86::isMOVHPMask(SVOp) ||
3994 X86::isMOVLPMask(SVOp)))
3997 if (ShouldXformToMOVHLPS(SVOp) ||
3998 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
3999 return CommuteVectorShuffle(SVOp, DAG);
4002 // No better options. Use a vshl / vsrl.
4003 MVT EVT = VT.getVectorElementType();
4004 ShAmt *= EVT.getSizeInBits();
4005 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4008 bool Commuted = false;
4009 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4010 // 1,1,1,1 -> v8i16 though.
4011 V1IsSplat = isSplatVector(V1.getNode());
4012 V2IsSplat = isSplatVector(V2.getNode());
4014 // Canonicalize the splat or undef, if present, to be on the RHS.
4015 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4016 Op = CommuteVectorShuffle(SVOp, DAG);
4017 SVOp = cast<ShuffleVectorSDNode>(Op);
4018 V1 = SVOp->getOperand(0);
4019 V2 = SVOp->getOperand(1);
4020 std::swap(V1IsSplat, V2IsSplat);
4021 std::swap(V1IsUndef, V2IsUndef);
4025 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4026 // Shuffling low element of v1 into undef, just return v1.
4029 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4030 // the instruction selector will not match, so get a canonical MOVL with
4031 // swapped operands to undo the commute.
4032 return getMOVL(DAG, dl, VT, V2, V1);
4035 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4036 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4037 X86::isUNPCKLMask(SVOp) ||
4038 X86::isUNPCKHMask(SVOp))
4042 // Normalize mask so all entries that point to V2 points to its first
4043 // element then try to match unpck{h|l} again. If match, return a
4044 // new vector_shuffle with the corrected mask.
4045 SDValue NewMask = NormalizeMask(SVOp, DAG);
4046 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4047 if (NSVOp != SVOp) {
4048 if (X86::isUNPCKLMask(NSVOp, true)) {
4050 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4057 // Commute is back and try unpck* again.
4058 // FIXME: this seems wrong.
4059 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4060 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4061 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4062 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4063 X86::isUNPCKLMask(NewSVOp) ||
4064 X86::isUNPCKHMask(NewSVOp))
4068 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4070 // Normalize the node to match x86 shuffle ops if needed
4071 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4072 return CommuteVectorShuffle(SVOp, DAG);
4074 // Check for legal shuffle and return?
4075 SmallVector<int, 16> PermMask;
4076 SVOp->getMask(PermMask);
4077 if (isShuffleMaskLegal(PermMask, VT))
4080 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4081 if (VT == MVT::v8i16) {
4082 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4083 if (NewOp.getNode())
4087 if (VT == MVT::v16i8) {
4088 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4089 if (NewOp.getNode())
4093 // Handle all 4 wide cases with a number of shuffles except for MMX.
4094 if (NumElems == 4 && !isMMX)
4095 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4101 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4102 SelectionDAG &DAG) {
4103 MVT VT = Op.getValueType();
4104 DebugLoc dl = Op.getDebugLoc();
4105 if (VT.getSizeInBits() == 8) {
4106 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4107 Op.getOperand(0), Op.getOperand(1));
4108 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4109 DAG.getValueType(VT));
4110 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4111 } else if (VT.getSizeInBits() == 16) {
4112 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4113 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4115 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4116 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4117 DAG.getNode(ISD::BIT_CONVERT, dl,
4121 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4122 Op.getOperand(0), Op.getOperand(1));
4123 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4124 DAG.getValueType(VT));
4125 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4126 } else if (VT == MVT::f32) {
4127 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4128 // the result back to FR32 register. It's only worth matching if the
4129 // result has a single use which is a store or a bitcast to i32. And in
4130 // the case of a store, it's not worth it if the index is a constant 0,
4131 // because a MOVSSmr can be used instead, which is smaller and faster.
4132 if (!Op.hasOneUse())
4134 SDNode *User = *Op.getNode()->use_begin();
4135 if ((User->getOpcode() != ISD::STORE ||
4136 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4137 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4138 (User->getOpcode() != ISD::BIT_CONVERT ||
4139 User->getValueType(0) != MVT::i32))
4141 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4142 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4145 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4146 } else if (VT == MVT::i32) {
4147 // ExtractPS works with constant index.
4148 if (isa<ConstantSDNode>(Op.getOperand(1)))
4156 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4157 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4160 if (Subtarget->hasSSE41()) {
4161 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4166 MVT VT = Op.getValueType();
4167 DebugLoc dl = Op.getDebugLoc();
4168 // TODO: handle v16i8.
4169 if (VT.getSizeInBits() == 16) {
4170 SDValue Vec = Op.getOperand(0);
4171 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4173 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4174 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4175 DAG.getNode(ISD::BIT_CONVERT, dl,
4178 // Transform it so it match pextrw which produces a 32-bit result.
4179 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4180 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4181 Op.getOperand(0), Op.getOperand(1));
4182 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4183 DAG.getValueType(VT));
4184 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4185 } else if (VT.getSizeInBits() == 32) {
4186 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4190 // SHUFPS the element to the lowest double word, then movss.
4191 int Mask[4] = { Idx, -1, -1, -1 };
4192 MVT VVT = Op.getOperand(0).getValueType();
4193 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4194 DAG.getUNDEF(VVT), Mask);
4195 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4196 DAG.getIntPtrConstant(0));
4197 } else if (VT.getSizeInBits() == 64) {
4198 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4199 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4200 // to match extract_elt for f64.
4201 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4205 // UNPCKHPD the element to the lowest double word, then movsd.
4206 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4207 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4208 int Mask[2] = { 1, -1 };
4209 MVT VVT = Op.getOperand(0).getValueType();
4210 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4211 DAG.getUNDEF(VVT), Mask);
4212 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4213 DAG.getIntPtrConstant(0));
4220 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4221 MVT VT = Op.getValueType();
4222 MVT EVT = VT.getVectorElementType();
4223 DebugLoc dl = Op.getDebugLoc();
4225 SDValue N0 = Op.getOperand(0);
4226 SDValue N1 = Op.getOperand(1);
4227 SDValue N2 = Op.getOperand(2);
4229 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4230 isa<ConstantSDNode>(N2)) {
4231 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4233 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4235 if (N1.getValueType() != MVT::i32)
4236 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4237 if (N2.getValueType() != MVT::i32)
4238 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4239 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4240 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4241 // Bits [7:6] of the constant are the source select. This will always be
4242 // zero here. The DAG Combiner may combine an extract_elt index into these
4243 // bits. For example (insert (extract, 3), 2) could be matched by putting
4244 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4245 // Bits [5:4] of the constant are the destination select. This is the
4246 // value of the incoming immediate.
4247 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4248 // combine either bitwise AND or insert of float 0.0 to set these bits.
4249 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4250 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4251 } else if (EVT == MVT::i32) {
4252 // InsertPS works with constant index.
4253 if (isa<ConstantSDNode>(N2))
4260 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4261 MVT VT = Op.getValueType();
4262 MVT EVT = VT.getVectorElementType();
4264 if (Subtarget->hasSSE41())
4265 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4270 DebugLoc dl = Op.getDebugLoc();
4271 SDValue N0 = Op.getOperand(0);
4272 SDValue N1 = Op.getOperand(1);
4273 SDValue N2 = Op.getOperand(2);
4275 if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4276 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4277 // as its second argument.
4278 if (N1.getValueType() != MVT::i32)
4279 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4280 if (N2.getValueType() != MVT::i32)
4281 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4282 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4288 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4289 DebugLoc dl = Op.getDebugLoc();
4290 if (Op.getValueType() == MVT::v2f32)
4291 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4292 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4293 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4294 Op.getOperand(0))));
4296 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4297 MVT VT = MVT::v2i32;
4298 switch (Op.getValueType().getSimpleVT()) {
4305 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4306 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4309 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4310 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4311 // one of the above mentioned nodes. It has to be wrapped because otherwise
4312 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4313 // be used to form addressing mode. These wrapped nodes will be selected
4316 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4317 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4318 // FIXME there isn't really any debug info here, should come from the parent
4319 DebugLoc dl = CP->getDebugLoc();
4320 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4321 CP->getAlignment());
4322 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4323 // With PIC, the address is actually $g + Offset.
4324 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4325 !Subtarget->isPICStyleRIPRel()) {
4326 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4327 DAG.getNode(X86ISD::GlobalBaseReg,
4328 DebugLoc::getUnknownLoc(),
4337 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4339 SelectionDAG &DAG) const {
4340 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
4341 bool ExtraLoadRequired =
4342 Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false);
4344 // Create the TargetGlobalAddress node, folding in the constant
4345 // offset if it is legal.
4347 if (!IsPic && !ExtraLoadRequired && isInt32(Offset)) {
4348 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4351 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0);
4352 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4354 // With PIC, the address is actually $g + Offset.
4355 if (IsPic && !Subtarget->isPICStyleRIPRel()) {
4356 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4357 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4361 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
4362 // load the value at address GV, not the value of GV itself. This means that
4363 // the GlobalAddress must be in the base or index register of the address, not
4364 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
4365 // The same applies for external symbols during PIC codegen
4366 if (ExtraLoadRequired)
4367 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4368 PseudoSourceValue::getGOT(), 0);
4370 // If there was a non-zero offset that we didn't fold, create an explicit
4373 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4374 DAG.getConstant(Offset, getPointerTy()));
4380 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4381 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4382 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4383 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4387 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4388 SDValue *InFlag, const MVT PtrVT, unsigned ReturnReg) {
4389 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4390 DebugLoc dl = GA->getDebugLoc();
4391 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4392 GA->getValueType(0),
4395 SDValue Ops[] = { Chain, TGA, *InFlag };
4396 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4398 SDValue Ops[] = { Chain, TGA };
4399 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4401 SDValue Flag = Chain.getValue(1);
4402 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4405 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4407 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4410 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4411 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4412 DAG.getNode(X86ISD::GlobalBaseReg,
4413 DebugLoc::getUnknownLoc(),
4415 InFlag = Chain.getValue(1);
4417 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX);
4420 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4422 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4424 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT, X86::RAX);
4427 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4428 // "local exec" model.
4429 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4430 const MVT PtrVT, TLSModel::Model model,
4432 DebugLoc dl = GA->getDebugLoc();
4433 // Get the Thread Pointer
4434 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4435 DebugLoc::getUnknownLoc(), PtrVT,
4436 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4439 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4442 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4444 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4445 GA->getValueType(0),
4447 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
4449 if (model == TLSModel::InitialExec)
4450 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4451 PseudoSourceValue::getGOT(), 0);
4453 // The address of the thread local variable is the add of the thread
4454 // pointer with the offset of the variable.
4455 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4459 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4460 // TODO: implement the "local dynamic" model
4461 // TODO: implement the "initial exec"model for pic executables
4462 assert(Subtarget->isTargetELF() &&
4463 "TLS not implemented for non-ELF targets");
4464 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4465 GlobalValue *GV = GA->getGlobal();
4466 TLSModel::Model model =
4467 getTLSModel (GV, getTargetMachine().getRelocationModel());
4468 if (Subtarget->is64Bit()) {
4470 case TLSModel::GeneralDynamic:
4471 case TLSModel::LocalDynamic: // not implemented
4472 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4474 case TLSModel::InitialExec:
4475 case TLSModel::LocalExec:
4476 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, true);
4480 case TLSModel::GeneralDynamic:
4481 case TLSModel::LocalDynamic: // not implemented
4482 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4484 case TLSModel::InitialExec:
4485 case TLSModel::LocalExec:
4486 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model, false);
4489 assert(0 && "Unreachable");
4494 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4495 // FIXME there isn't really any debug info here
4496 DebugLoc dl = Op.getDebugLoc();
4497 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4498 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
4499 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4500 // With PIC, the address is actually $g + Offset.
4501 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4502 !Subtarget->isPICStyleRIPRel()) {
4503 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4504 DAG.getNode(X86ISD::GlobalBaseReg,
4505 DebugLoc::getUnknownLoc(),
4513 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4514 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4515 // FIXME there isn't really any debug into here
4516 DebugLoc dl = JT->getDebugLoc();
4517 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
4518 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4519 // With PIC, the address is actually $g + Offset.
4520 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4521 !Subtarget->isPICStyleRIPRel()) {
4522 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4523 DAG.getNode(X86ISD::GlobalBaseReg,
4524 DebugLoc::getUnknownLoc(),
4532 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4533 /// take a 2 x i32 value to shift plus a shift amount.
4534 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4535 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4536 MVT VT = Op.getValueType();
4537 unsigned VTBits = VT.getSizeInBits();
4538 DebugLoc dl = Op.getDebugLoc();
4539 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4540 SDValue ShOpLo = Op.getOperand(0);
4541 SDValue ShOpHi = Op.getOperand(1);
4542 SDValue ShAmt = Op.getOperand(2);
4543 SDValue Tmp1 = isSRA ?
4544 DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4545 DAG.getConstant(VTBits - 1, MVT::i8)) :
4546 DAG.getConstant(0, VT);
4549 if (Op.getOpcode() == ISD::SHL_PARTS) {
4550 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4551 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4553 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4554 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4557 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4558 DAG.getConstant(VTBits, MVT::i8));
4559 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4560 AndNode, DAG.getConstant(0, MVT::i8));
4563 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4564 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4565 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4567 if (Op.getOpcode() == ISD::SHL_PARTS) {
4568 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4569 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4571 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4572 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4575 SDValue Ops[2] = { Lo, Hi };
4576 return DAG.getMergeValues(Ops, 2, dl);
4579 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4580 MVT SrcVT = Op.getOperand(0).getValueType();
4582 if (SrcVT.isVector()) {
4583 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4589 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4590 "Unknown SINT_TO_FP to lower!");
4592 // These are really Legal; return the operand so the caller accepts it as
4594 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4596 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4597 Subtarget->is64Bit()) {
4601 DebugLoc dl = Op.getDebugLoc();
4602 unsigned Size = SrcVT.getSizeInBits()/8;
4603 MachineFunction &MF = DAG.getMachineFunction();
4604 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4605 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4606 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4608 PseudoSourceValue::getFixedStack(SSFI), 0);
4609 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4612 SDValue X86TargetLowering::BuildFILD(SDValue Op, MVT SrcVT, SDValue Chain,
4614 SelectionDAG &DAG) {
4616 DebugLoc dl = Op.getDebugLoc();
4618 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4620 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4622 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4623 SmallVector<SDValue, 8> Ops;
4624 Ops.push_back(Chain);
4625 Ops.push_back(StackSlot);
4626 Ops.push_back(DAG.getValueType(SrcVT));
4627 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4628 Tys, &Ops[0], Ops.size());
4631 Chain = Result.getValue(1);
4632 SDValue InFlag = Result.getValue(2);
4634 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4635 // shouldn't be necessary except that RFP cannot be live across
4636 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4637 MachineFunction &MF = DAG.getMachineFunction();
4638 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4639 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4640 Tys = DAG.getVTList(MVT::Other);
4641 SmallVector<SDValue, 8> Ops;
4642 Ops.push_back(Chain);
4643 Ops.push_back(Result);
4644 Ops.push_back(StackSlot);
4645 Ops.push_back(DAG.getValueType(Op.getValueType()));
4646 Ops.push_back(InFlag);
4647 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4648 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4649 PseudoSourceValue::getFixedStack(SSFI), 0);
4655 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4656 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4657 // This algorithm is not obvious. Here it is in C code, more or less:
4659 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4660 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4661 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4663 // Copy ints to xmm registers.
4664 __m128i xh = _mm_cvtsi32_si128( hi );
4665 __m128i xl = _mm_cvtsi32_si128( lo );
4667 // Combine into low half of a single xmm register.
4668 __m128i x = _mm_unpacklo_epi32( xh, xl );
4672 // Merge in appropriate exponents to give the integer bits the right
4674 x = _mm_unpacklo_epi32( x, exp );
4676 // Subtract away the biases to deal with the IEEE-754 double precision
4678 d = _mm_sub_pd( (__m128d) x, bias );
4680 // All conversions up to here are exact. The correctly rounded result is
4681 // calculated using the current rounding mode using the following
4683 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4684 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4685 // store doesn't really need to be here (except
4686 // maybe to zero the other double)
4691 DebugLoc dl = Op.getDebugLoc();
4693 // Build some magic constants.
4694 std::vector<Constant*> CV0;
4695 CV0.push_back(ConstantInt::get(APInt(32, 0x45300000)));
4696 CV0.push_back(ConstantInt::get(APInt(32, 0x43300000)));
4697 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4698 CV0.push_back(ConstantInt::get(APInt(32, 0)));
4699 Constant *C0 = ConstantVector::get(CV0);
4700 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4702 std::vector<Constant*> CV1;
4703 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4530000000000000ULL))));
4704 CV1.push_back(ConstantFP::get(APFloat(APInt(64, 0x4330000000000000ULL))));
4705 Constant *C1 = ConstantVector::get(CV1);
4706 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4708 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4709 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4711 DAG.getIntPtrConstant(1)));
4712 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4713 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4715 DAG.getIntPtrConstant(0)));
4716 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4717 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4718 PseudoSourceValue::getConstantPool(), 0,
4720 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4721 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4722 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4723 PseudoSourceValue::getConstantPool(), 0,
4725 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4727 // Add the halves; easiest way is to swap them into another reg first.
4728 int ShufMask[2] = { 1, -1 };
4729 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4730 DAG.getUNDEF(MVT::v2f64), ShufMask);
4731 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4732 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4733 DAG.getIntPtrConstant(0));
4736 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4737 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4738 DebugLoc dl = Op.getDebugLoc();
4739 // FP constant to bias correct the final result.
4740 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4743 // Load the 32-bit value into an XMM register.
4744 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4745 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4747 DAG.getIntPtrConstant(0)));
4749 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4750 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4751 DAG.getIntPtrConstant(0));
4753 // Or the load with the bias.
4754 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4755 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4756 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4758 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4759 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4760 MVT::v2f64, Bias)));
4761 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4762 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4763 DAG.getIntPtrConstant(0));
4765 // Subtract the bias.
4766 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
4768 // Handle final rounding.
4769 MVT DestVT = Op.getValueType();
4771 if (DestVT.bitsLT(MVT::f64)) {
4772 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
4773 DAG.getIntPtrConstant(0));
4774 } else if (DestVT.bitsGT(MVT::f64)) {
4775 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
4778 // Handle final rounding.
4782 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4783 SDValue N0 = Op.getOperand(0);
4784 DebugLoc dl = Op.getDebugLoc();
4786 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
4787 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
4788 // the optimization here.
4789 if (DAG.SignBitIsZero(N0))
4790 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
4792 MVT SrcVT = N0.getValueType();
4793 if (SrcVT == MVT::i64) {
4794 // We only handle SSE2 f64 target here; caller can expand the rest.
4795 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
4798 return LowerUINT_TO_FP_i64(Op, DAG);
4799 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
4800 return LowerUINT_TO_FP_i32(Op, DAG);
4803 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
4805 // Make a 64-bit buffer, and use it to build an FILD.
4806 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
4807 SDValue WordOff = DAG.getConstant(4, getPointerTy());
4808 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
4809 getPointerTy(), StackSlot, WordOff);
4810 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4811 StackSlot, NULL, 0);
4812 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
4813 OffsetSlot, NULL, 0);
4814 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
4817 std::pair<SDValue,SDValue> X86TargetLowering::
4818 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
4819 DebugLoc dl = Op.getDebugLoc();
4821 MVT DstTy = Op.getValueType();
4824 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
4828 assert(DstTy.getSimpleVT() <= MVT::i64 &&
4829 DstTy.getSimpleVT() >= MVT::i16 &&
4830 "Unknown FP_TO_SINT to lower!");
4832 // These are really Legal.
4833 if (DstTy == MVT::i32 &&
4834 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4835 return std::make_pair(SDValue(), SDValue());
4836 if (Subtarget->is64Bit() &&
4837 DstTy == MVT::i64 &&
4838 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
4839 return std::make_pair(SDValue(), SDValue());
4841 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
4843 MachineFunction &MF = DAG.getMachineFunction();
4844 unsigned MemSize = DstTy.getSizeInBits()/8;
4845 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4846 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4849 switch (DstTy.getSimpleVT()) {
4850 default: assert(0 && "Invalid FP_TO_SINT to lower!");
4851 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
4852 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
4853 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
4856 SDValue Chain = DAG.getEntryNode();
4857 SDValue Value = Op.getOperand(0);
4858 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
4859 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
4860 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
4861 PseudoSourceValue::getFixedStack(SSFI), 0);
4862 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
4864 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
4866 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
4867 Chain = Value.getValue(1);
4868 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
4869 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4872 // Build the FP_TO_INT*_IN_MEM
4873 SDValue Ops[] = { Chain, Value, StackSlot };
4874 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
4876 return std::make_pair(FIST, StackSlot);
4879 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
4880 if (Op.getValueType().isVector()) {
4881 if (Op.getValueType() == MVT::v2i32 &&
4882 Op.getOperand(0).getValueType() == MVT::v2f64) {
4888 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
4889 SDValue FIST = Vals.first, StackSlot = Vals.second;
4890 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
4891 if (FIST.getNode() == 0) return Op;
4894 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
4895 FIST, StackSlot, NULL, 0);
4898 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
4899 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
4900 SDValue FIST = Vals.first, StackSlot = Vals.second;
4901 assert(FIST.getNode() && "Unexpected failure");
4904 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
4905 FIST, StackSlot, NULL, 0);
4908 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
4909 DebugLoc dl = Op.getDebugLoc();
4910 MVT VT = Op.getValueType();
4913 EltVT = VT.getVectorElementType();
4914 std::vector<Constant*> CV;
4915 if (EltVT == MVT::f64) {
4916 Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
4920 Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
4926 Constant *C = ConstantVector::get(CV);
4927 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
4928 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
4929 PseudoSourceValue::getConstantPool(), 0,
4931 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
4934 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
4935 DebugLoc dl = Op.getDebugLoc();
4936 MVT VT = Op.getValueType();
4938 unsigned EltNum = 1;
4939 if (VT.isVector()) {
4940 EltVT = VT.getVectorElementType();
4941 EltNum = VT.getVectorNumElements();
4943 std::vector<Constant*> CV;
4944 if (EltVT == MVT::f64) {
4945 Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
4949 Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
4955 Constant *C = ConstantVector::get(CV);
4956 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
4957 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
4958 PseudoSourceValue::getConstantPool(), 0,
4960 if (VT.isVector()) {
4961 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4962 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
4963 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4965 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
4967 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
4971 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
4972 SDValue Op0 = Op.getOperand(0);
4973 SDValue Op1 = Op.getOperand(1);
4974 DebugLoc dl = Op.getDebugLoc();
4975 MVT VT = Op.getValueType();
4976 MVT SrcVT = Op1.getValueType();
4978 // If second operand is smaller, extend it first.
4979 if (SrcVT.bitsLT(VT)) {
4980 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
4983 // And if it is bigger, shrink it first.
4984 if (SrcVT.bitsGT(VT)) {
4985 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
4989 // At this point the operands and the result should have the same
4990 // type, and that won't be f80 since that is not custom lowered.
4992 // First get the sign bit of second operand.
4993 std::vector<Constant*> CV;
4994 if (SrcVT == MVT::f64) {
4995 CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
4996 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
4998 CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
4999 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5000 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5001 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5003 Constant *C = ConstantVector::get(CV);
5004 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5005 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5006 PseudoSourceValue::getConstantPool(), 0,
5008 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5010 // Shift sign bit right or left if the two operands have different types.
5011 if (SrcVT.bitsGT(VT)) {
5012 // Op0 is MVT::f32, Op1 is MVT::f64.
5013 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5014 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5015 DAG.getConstant(32, MVT::i32));
5016 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5017 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5018 DAG.getIntPtrConstant(0));
5021 // Clear first operand sign bit.
5023 if (VT == MVT::f64) {
5024 CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
5025 CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
5027 CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
5028 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5029 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5030 CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
5032 C = ConstantVector::get(CV);
5033 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5034 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5035 PseudoSourceValue::getConstantPool(), 0,
5037 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5039 // Or the value with the sign bit.
5040 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5043 /// Emit nodes that will be selected as "test Op0,Op0", or something
5045 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5046 SelectionDAG &DAG) {
5047 DebugLoc dl = Op.getDebugLoc();
5049 // CF and OF aren't always set the way we want. Determine which
5050 // of these we need.
5051 bool NeedCF = false;
5052 bool NeedOF = false;
5054 case X86::COND_A: case X86::COND_AE:
5055 case X86::COND_B: case X86::COND_BE:
5058 case X86::COND_G: case X86::COND_GE:
5059 case X86::COND_L: case X86::COND_LE:
5060 case X86::COND_O: case X86::COND_NO:
5066 // See if we can use the EFLAGS value from the operand instead of
5067 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5068 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5069 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5070 unsigned Opcode = 0;
5071 unsigned NumOperands = 0;
5072 switch (Op.getNode()->getOpcode()) {
5074 // Due to an isel shortcoming, be conservative if this add is likely to
5075 // be selected as part of a load-modify-store instruction. When the root
5076 // node in a match is a store, isel doesn't know how to remap non-chain
5077 // non-flag uses of other nodes in the match, such as the ADD in this
5078 // case. This leads to the ADD being left around and reselected, with
5079 // the result being two adds in the output.
5080 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5081 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5082 if (UI->getOpcode() == ISD::STORE)
5084 if (ConstantSDNode *C =
5085 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5086 // An add of one will be selected as an INC.
5087 if (C->getAPIntValue() == 1) {
5088 Opcode = X86ISD::INC;
5092 // An add of negative one (subtract of one) will be selected as a DEC.
5093 if (C->getAPIntValue().isAllOnesValue()) {
5094 Opcode = X86ISD::DEC;
5099 // Otherwise use a regular EFLAGS-setting add.
5100 Opcode = X86ISD::ADD;
5104 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5105 // likely to be selected as part of a load-modify-store instruction.
5106 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5107 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5108 if (UI->getOpcode() == ISD::STORE)
5110 // Otherwise use a regular EFLAGS-setting sub.
5111 Opcode = X86ISD::SUB;
5118 return SDValue(Op.getNode(), 1);
5124 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5125 SmallVector<SDValue, 4> Ops;
5126 for (unsigned i = 0; i != NumOperands; ++i)
5127 Ops.push_back(Op.getOperand(i));
5128 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5129 DAG.ReplaceAllUsesWith(Op, New);
5130 return SDValue(New.getNode(), 1);
5134 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5135 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5136 DAG.getConstant(0, Op.getValueType()));
5139 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5141 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5142 SelectionDAG &DAG) {
5143 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5144 if (C->getAPIntValue() == 0)
5145 return EmitTest(Op0, X86CC, DAG);
5147 DebugLoc dl = Op0.getDebugLoc();
5148 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5151 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5152 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5153 SDValue Op0 = Op.getOperand(0);
5154 SDValue Op1 = Op.getOperand(1);
5155 DebugLoc dl = Op.getDebugLoc();
5156 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5158 // Lower (X & (1 << N)) == 0 to BT(X, N).
5159 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5160 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5161 if (Op0.getOpcode() == ISD::AND &&
5163 Op1.getOpcode() == ISD::Constant &&
5164 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5165 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5167 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5168 if (ConstantSDNode *Op010C =
5169 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5170 if (Op010C->getZExtValue() == 1) {
5171 LHS = Op0.getOperand(0);
5172 RHS = Op0.getOperand(1).getOperand(1);
5174 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5175 if (ConstantSDNode *Op000C =
5176 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5177 if (Op000C->getZExtValue() == 1) {
5178 LHS = Op0.getOperand(1);
5179 RHS = Op0.getOperand(0).getOperand(1);
5181 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5182 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5183 SDValue AndLHS = Op0.getOperand(0);
5184 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5185 LHS = AndLHS.getOperand(0);
5186 RHS = AndLHS.getOperand(1);
5190 if (LHS.getNode()) {
5191 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5192 // instruction. Since the shift amount is in-range-or-undefined, we know
5193 // that doing a bittest on the i16 value is ok. We extend to i32 because
5194 // the encoding for the i16 version is larger than the i32 version.
5195 if (LHS.getValueType() == MVT::i8)
5196 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5198 // If the operand types disagree, extend the shift amount to match. Since
5199 // BT ignores high bits (like shifts) we can use anyextend.
5200 if (LHS.getValueType() != RHS.getValueType())
5201 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5203 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5204 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5205 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5206 DAG.getConstant(Cond, MVT::i8), BT);
5210 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5211 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5213 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5214 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5215 DAG.getConstant(X86CC, MVT::i8), Cond);
5218 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5220 SDValue Op0 = Op.getOperand(0);
5221 SDValue Op1 = Op.getOperand(1);
5222 SDValue CC = Op.getOperand(2);
5223 MVT VT = Op.getValueType();
5224 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5225 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5226 DebugLoc dl = Op.getDebugLoc();
5230 MVT VT0 = Op0.getValueType();
5231 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5232 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5235 switch (SetCCOpcode) {
5238 case ISD::SETEQ: SSECC = 0; break;
5240 case ISD::SETGT: Swap = true; // Fallthrough
5242 case ISD::SETOLT: SSECC = 1; break;
5244 case ISD::SETGE: Swap = true; // Fallthrough
5246 case ISD::SETOLE: SSECC = 2; break;
5247 case ISD::SETUO: SSECC = 3; break;
5249 case ISD::SETNE: SSECC = 4; break;
5250 case ISD::SETULE: Swap = true;
5251 case ISD::SETUGE: SSECC = 5; break;
5252 case ISD::SETULT: Swap = true;
5253 case ISD::SETUGT: SSECC = 6; break;
5254 case ISD::SETO: SSECC = 7; break;
5257 std::swap(Op0, Op1);
5259 // In the two special cases we can't handle, emit two comparisons.
5261 if (SetCCOpcode == ISD::SETUEQ) {
5263 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5264 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5265 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5267 else if (SetCCOpcode == ISD::SETONE) {
5269 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5270 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5271 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5273 assert(0 && "Illegal FP comparison");
5275 // Handle all other FP comparisons here.
5276 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5279 // We are handling one of the integer comparisons here. Since SSE only has
5280 // GT and EQ comparisons for integer, swapping operands and multiple
5281 // operations may be required for some comparisons.
5282 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5283 bool Swap = false, Invert = false, FlipSigns = false;
5285 switch (VT.getSimpleVT()) {
5287 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5288 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5289 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5290 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5293 switch (SetCCOpcode) {
5295 case ISD::SETNE: Invert = true;
5296 case ISD::SETEQ: Opc = EQOpc; break;
5297 case ISD::SETLT: Swap = true;
5298 case ISD::SETGT: Opc = GTOpc; break;
5299 case ISD::SETGE: Swap = true;
5300 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5301 case ISD::SETULT: Swap = true;
5302 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5303 case ISD::SETUGE: Swap = true;
5304 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5307 std::swap(Op0, Op1);
5309 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5310 // bits of the inputs before performing those operations.
5312 MVT EltVT = VT.getVectorElementType();
5313 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5315 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5316 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5318 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5319 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5322 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5324 // If the logical-not of the result is required, perform that now.
5326 Result = DAG.getNOT(dl, Result, VT);
5331 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5332 static bool isX86LogicalCmp(SDValue Op) {
5333 unsigned Opc = Op.getNode()->getOpcode();
5334 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5336 if (Op.getResNo() == 1 &&
5337 (Opc == X86ISD::ADD ||
5338 Opc == X86ISD::SUB ||
5339 Opc == X86ISD::SMUL ||
5340 Opc == X86ISD::UMUL ||
5341 Opc == X86ISD::INC ||
5342 Opc == X86ISD::DEC))
5348 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5349 bool addTest = true;
5350 SDValue Cond = Op.getOperand(0);
5351 DebugLoc dl = Op.getDebugLoc();
5354 if (Cond.getOpcode() == ISD::SETCC)
5355 Cond = LowerSETCC(Cond, DAG);
5357 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5358 // setting operand in place of the X86ISD::SETCC.
5359 if (Cond.getOpcode() == X86ISD::SETCC) {
5360 CC = Cond.getOperand(0);
5362 SDValue Cmp = Cond.getOperand(1);
5363 unsigned Opc = Cmp.getOpcode();
5364 MVT VT = Op.getValueType();
5366 bool IllegalFPCMov = false;
5367 if (VT.isFloatingPoint() && !VT.isVector() &&
5368 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5369 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5371 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5372 Opc == X86ISD::BT) { // FIXME
5379 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5380 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5383 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5384 SmallVector<SDValue, 4> Ops;
5385 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5386 // condition is true.
5387 Ops.push_back(Op.getOperand(2));
5388 Ops.push_back(Op.getOperand(1));
5390 Ops.push_back(Cond);
5391 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5394 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5395 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5396 // from the AND / OR.
5397 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5398 Opc = Op.getOpcode();
5399 if (Opc != ISD::OR && Opc != ISD::AND)
5401 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5402 Op.getOperand(0).hasOneUse() &&
5403 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5404 Op.getOperand(1).hasOneUse());
5407 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5408 // 1 and that the SETCC node has a single use.
5409 static bool isXor1OfSetCC(SDValue Op) {
5410 if (Op.getOpcode() != ISD::XOR)
5412 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5413 if (N1C && N1C->getAPIntValue() == 1) {
5414 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5415 Op.getOperand(0).hasOneUse();
5420 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5421 bool addTest = true;
5422 SDValue Chain = Op.getOperand(0);
5423 SDValue Cond = Op.getOperand(1);
5424 SDValue Dest = Op.getOperand(2);
5425 DebugLoc dl = Op.getDebugLoc();
5428 if (Cond.getOpcode() == ISD::SETCC)
5429 Cond = LowerSETCC(Cond, DAG);
5431 // FIXME: LowerXALUO doesn't handle these!!
5432 else if (Cond.getOpcode() == X86ISD::ADD ||
5433 Cond.getOpcode() == X86ISD::SUB ||
5434 Cond.getOpcode() == X86ISD::SMUL ||
5435 Cond.getOpcode() == X86ISD::UMUL)
5436 Cond = LowerXALUO(Cond, DAG);
5439 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5440 // setting operand in place of the X86ISD::SETCC.
5441 if (Cond.getOpcode() == X86ISD::SETCC) {
5442 CC = Cond.getOperand(0);
5444 SDValue Cmp = Cond.getOperand(1);
5445 unsigned Opc = Cmp.getOpcode();
5446 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5447 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5451 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5455 // These can only come from an arithmetic instruction with overflow,
5456 // e.g. SADDO, UADDO.
5457 Cond = Cond.getNode()->getOperand(1);
5464 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5465 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5466 if (CondOpc == ISD::OR) {
5467 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5468 // two branches instead of an explicit OR instruction with a
5470 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5471 isX86LogicalCmp(Cmp)) {
5472 CC = Cond.getOperand(0).getOperand(0);
5473 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5474 Chain, Dest, CC, Cmp);
5475 CC = Cond.getOperand(1).getOperand(0);
5479 } else { // ISD::AND
5480 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5481 // two branches instead of an explicit AND instruction with a
5482 // separate test. However, we only do this if this block doesn't
5483 // have a fall-through edge, because this requires an explicit
5484 // jmp when the condition is false.
5485 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5486 isX86LogicalCmp(Cmp) &&
5487 Op.getNode()->hasOneUse()) {
5488 X86::CondCode CCode =
5489 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5490 CCode = X86::GetOppositeBranchCondition(CCode);
5491 CC = DAG.getConstant(CCode, MVT::i8);
5492 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5493 // Look for an unconditional branch following this conditional branch.
5494 // We need this because we need to reverse the successors in order
5495 // to implement FCMP_OEQ.
5496 if (User.getOpcode() == ISD::BR) {
5497 SDValue FalseBB = User.getOperand(1);
5499 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5500 assert(NewBR == User);
5503 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5504 Chain, Dest, CC, Cmp);
5505 X86::CondCode CCode =
5506 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5507 CCode = X86::GetOppositeBranchCondition(CCode);
5508 CC = DAG.getConstant(CCode, MVT::i8);
5514 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5515 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5516 // It should be transformed during dag combiner except when the condition
5517 // is set by a arithmetics with overflow node.
5518 X86::CondCode CCode =
5519 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5520 CCode = X86::GetOppositeBranchCondition(CCode);
5521 CC = DAG.getConstant(CCode, MVT::i8);
5522 Cond = Cond.getOperand(0).getOperand(1);
5528 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5529 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5531 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5532 Chain, Dest, CC, Cond);
5536 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5537 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5538 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5539 // that the guard pages used by the OS virtual memory manager are allocated in
5540 // correct sequence.
5542 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5543 SelectionDAG &DAG) {
5544 assert(Subtarget->isTargetCygMing() &&
5545 "This should be used only on Cygwin/Mingw targets");
5546 DebugLoc dl = Op.getDebugLoc();
5549 SDValue Chain = Op.getOperand(0);
5550 SDValue Size = Op.getOperand(1);
5551 // FIXME: Ensure alignment here
5555 MVT IntPtr = getPointerTy();
5556 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5558 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5560 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5561 Flag = Chain.getValue(1);
5563 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5564 SDValue Ops[] = { Chain,
5565 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5566 DAG.getRegister(X86::EAX, IntPtr),
5567 DAG.getRegister(X86StackPtr, SPTy),
5569 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5570 Flag = Chain.getValue(1);
5572 Chain = DAG.getCALLSEQ_END(Chain,
5573 DAG.getIntPtrConstant(0, true),
5574 DAG.getIntPtrConstant(0, true),
5577 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5579 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5580 return DAG.getMergeValues(Ops1, 2, dl);
5584 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5586 SDValue Dst, SDValue Src,
5587 SDValue Size, unsigned Align,
5589 uint64_t DstSVOff) {
5590 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5592 // If not DWORD aligned or size is more than the threshold, call the library.
5593 // The libc version is likely to be faster for these cases. It can use the
5594 // address value and run time information about the CPU.
5595 if ((Align & 3) != 0 ||
5597 ConstantSize->getZExtValue() >
5598 getSubtarget()->getMaxInlineSizeThreshold()) {
5599 SDValue InFlag(0, 0);
5601 // Check to see if there is a specialized entry-point for memory zeroing.
5602 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5604 if (const char *bzeroEntry = V &&
5605 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5606 MVT IntPtr = getPointerTy();
5607 const Type *IntPtrTy = TD->getIntPtrType();
5608 TargetLowering::ArgListTy Args;
5609 TargetLowering::ArgListEntry Entry;
5611 Entry.Ty = IntPtrTy;
5612 Args.push_back(Entry);
5614 Args.push_back(Entry);
5615 std::pair<SDValue,SDValue> CallResult =
5616 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5617 CallingConv::C, false,
5618 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5619 return CallResult.second;
5622 // Otherwise have the target-independent code call memset.
5626 uint64_t SizeVal = ConstantSize->getZExtValue();
5627 SDValue InFlag(0, 0);
5630 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5631 unsigned BytesLeft = 0;
5632 bool TwoRepStos = false;
5635 uint64_t Val = ValC->getZExtValue() & 255;
5637 // If the value is a constant, then we can potentially use larger sets.
5638 switch (Align & 3) {
5639 case 2: // WORD aligned
5642 Val = (Val << 8) | Val;
5644 case 0: // DWORD aligned
5647 Val = (Val << 8) | Val;
5648 Val = (Val << 16) | Val;
5649 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5652 Val = (Val << 32) | Val;
5655 default: // Byte aligned
5658 Count = DAG.getIntPtrConstant(SizeVal);
5662 if (AVT.bitsGT(MVT::i8)) {
5663 unsigned UBytes = AVT.getSizeInBits() / 8;
5664 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5665 BytesLeft = SizeVal % UBytes;
5668 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5670 InFlag = Chain.getValue(1);
5673 Count = DAG.getIntPtrConstant(SizeVal);
5674 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5675 InFlag = Chain.getValue(1);
5678 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5681 InFlag = Chain.getValue(1);
5682 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5685 InFlag = Chain.getValue(1);
5687 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5688 SmallVector<SDValue, 8> Ops;
5689 Ops.push_back(Chain);
5690 Ops.push_back(DAG.getValueType(AVT));
5691 Ops.push_back(InFlag);
5692 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5695 InFlag = Chain.getValue(1);
5697 MVT CVT = Count.getValueType();
5698 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5699 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5700 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5703 InFlag = Chain.getValue(1);
5704 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5706 Ops.push_back(Chain);
5707 Ops.push_back(DAG.getValueType(MVT::i8));
5708 Ops.push_back(InFlag);
5709 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5710 } else if (BytesLeft) {
5711 // Handle the last 1 - 7 bytes.
5712 unsigned Offset = SizeVal - BytesLeft;
5713 MVT AddrVT = Dst.getValueType();
5714 MVT SizeVT = Size.getValueType();
5716 Chain = DAG.getMemset(Chain, dl,
5717 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5718 DAG.getConstant(Offset, AddrVT)),
5720 DAG.getConstant(BytesLeft, SizeVT),
5721 Align, DstSV, DstSVOff + Offset);
5724 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5729 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5730 SDValue Chain, SDValue Dst, SDValue Src,
5731 SDValue Size, unsigned Align,
5733 const Value *DstSV, uint64_t DstSVOff,
5734 const Value *SrcSV, uint64_t SrcSVOff) {
5735 // This requires the copy size to be a constant, preferrably
5736 // within a subtarget-specific limit.
5737 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5740 uint64_t SizeVal = ConstantSize->getZExtValue();
5741 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5744 /// If not DWORD aligned, call the library.
5745 if ((Align & 3) != 0)
5750 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5753 unsigned UBytes = AVT.getSizeInBits() / 8;
5754 unsigned CountVal = SizeVal / UBytes;
5755 SDValue Count = DAG.getIntPtrConstant(CountVal);
5756 unsigned BytesLeft = SizeVal % UBytes;
5758 SDValue InFlag(0, 0);
5759 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5762 InFlag = Chain.getValue(1);
5763 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5766 InFlag = Chain.getValue(1);
5767 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5770 InFlag = Chain.getValue(1);
5772 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5773 SmallVector<SDValue, 8> Ops;
5774 Ops.push_back(Chain);
5775 Ops.push_back(DAG.getValueType(AVT));
5776 Ops.push_back(InFlag);
5777 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
5779 SmallVector<SDValue, 4> Results;
5780 Results.push_back(RepMovs);
5782 // Handle the last 1 - 7 bytes.
5783 unsigned Offset = SizeVal - BytesLeft;
5784 MVT DstVT = Dst.getValueType();
5785 MVT SrcVT = Src.getValueType();
5786 MVT SizeVT = Size.getValueType();
5787 Results.push_back(DAG.getMemcpy(Chain, dl,
5788 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
5789 DAG.getConstant(Offset, DstVT)),
5790 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
5791 DAG.getConstant(Offset, SrcVT)),
5792 DAG.getConstant(BytesLeft, SizeVT),
5793 Align, AlwaysInline,
5794 DstSV, DstSVOff + Offset,
5795 SrcSV, SrcSVOff + Offset));
5798 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5799 &Results[0], Results.size());
5802 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
5803 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
5804 DebugLoc dl = Op.getDebugLoc();
5806 if (!Subtarget->is64Bit()) {
5807 // vastart just stores the address of the VarArgsFrameIndex slot into the
5808 // memory location argument.
5809 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5810 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
5814 // gp_offset (0 - 6 * 8)
5815 // fp_offset (48 - 48 + 8 * 16)
5816 // overflow_arg_area (point to parameters coming in memory).
5818 SmallVector<SDValue, 8> MemOps;
5819 SDValue FIN = Op.getOperand(1);
5821 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
5822 DAG.getConstant(VarArgsGPOffset, MVT::i32),
5824 MemOps.push_back(Store);
5827 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5828 FIN, DAG.getIntPtrConstant(4));
5829 Store = DAG.getStore(Op.getOperand(0), dl,
5830 DAG.getConstant(VarArgsFPOffset, MVT::i32),
5832 MemOps.push_back(Store);
5834 // Store ptr to overflow_arg_area
5835 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5836 FIN, DAG.getIntPtrConstant(4));
5837 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
5838 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
5839 MemOps.push_back(Store);
5841 // Store ptr to reg_save_area.
5842 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5843 FIN, DAG.getIntPtrConstant(8));
5844 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
5845 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
5846 MemOps.push_back(Store);
5847 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5848 &MemOps[0], MemOps.size());
5851 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
5852 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5853 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
5854 SDValue Chain = Op.getOperand(0);
5855 SDValue SrcPtr = Op.getOperand(1);
5856 SDValue SrcSV = Op.getOperand(2);
5858 assert(0 && "VAArgInst is not yet implemented for x86-64!");
5863 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
5864 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
5865 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
5866 SDValue Chain = Op.getOperand(0);
5867 SDValue DstPtr = Op.getOperand(1);
5868 SDValue SrcPtr = Op.getOperand(2);
5869 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
5870 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
5871 DebugLoc dl = Op.getDebugLoc();
5873 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
5874 DAG.getIntPtrConstant(24), 8, false,
5875 DstSV, 0, SrcSV, 0);
5879 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
5880 DebugLoc dl = Op.getDebugLoc();
5881 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5883 default: return SDValue(); // Don't custom lower most intrinsics.
5884 // Comparison intrinsics.
5885 case Intrinsic::x86_sse_comieq_ss:
5886 case Intrinsic::x86_sse_comilt_ss:
5887 case Intrinsic::x86_sse_comile_ss:
5888 case Intrinsic::x86_sse_comigt_ss:
5889 case Intrinsic::x86_sse_comige_ss:
5890 case Intrinsic::x86_sse_comineq_ss:
5891 case Intrinsic::x86_sse_ucomieq_ss:
5892 case Intrinsic::x86_sse_ucomilt_ss:
5893 case Intrinsic::x86_sse_ucomile_ss:
5894 case Intrinsic::x86_sse_ucomigt_ss:
5895 case Intrinsic::x86_sse_ucomige_ss:
5896 case Intrinsic::x86_sse_ucomineq_ss:
5897 case Intrinsic::x86_sse2_comieq_sd:
5898 case Intrinsic::x86_sse2_comilt_sd:
5899 case Intrinsic::x86_sse2_comile_sd:
5900 case Intrinsic::x86_sse2_comigt_sd:
5901 case Intrinsic::x86_sse2_comige_sd:
5902 case Intrinsic::x86_sse2_comineq_sd:
5903 case Intrinsic::x86_sse2_ucomieq_sd:
5904 case Intrinsic::x86_sse2_ucomilt_sd:
5905 case Intrinsic::x86_sse2_ucomile_sd:
5906 case Intrinsic::x86_sse2_ucomigt_sd:
5907 case Intrinsic::x86_sse2_ucomige_sd:
5908 case Intrinsic::x86_sse2_ucomineq_sd: {
5910 ISD::CondCode CC = ISD::SETCC_INVALID;
5913 case Intrinsic::x86_sse_comieq_ss:
5914 case Intrinsic::x86_sse2_comieq_sd:
5918 case Intrinsic::x86_sse_comilt_ss:
5919 case Intrinsic::x86_sse2_comilt_sd:
5923 case Intrinsic::x86_sse_comile_ss:
5924 case Intrinsic::x86_sse2_comile_sd:
5928 case Intrinsic::x86_sse_comigt_ss:
5929 case Intrinsic::x86_sse2_comigt_sd:
5933 case Intrinsic::x86_sse_comige_ss:
5934 case Intrinsic::x86_sse2_comige_sd:
5938 case Intrinsic::x86_sse_comineq_ss:
5939 case Intrinsic::x86_sse2_comineq_sd:
5943 case Intrinsic::x86_sse_ucomieq_ss:
5944 case Intrinsic::x86_sse2_ucomieq_sd:
5945 Opc = X86ISD::UCOMI;
5948 case Intrinsic::x86_sse_ucomilt_ss:
5949 case Intrinsic::x86_sse2_ucomilt_sd:
5950 Opc = X86ISD::UCOMI;
5953 case Intrinsic::x86_sse_ucomile_ss:
5954 case Intrinsic::x86_sse2_ucomile_sd:
5955 Opc = X86ISD::UCOMI;
5958 case Intrinsic::x86_sse_ucomigt_ss:
5959 case Intrinsic::x86_sse2_ucomigt_sd:
5960 Opc = X86ISD::UCOMI;
5963 case Intrinsic::x86_sse_ucomige_ss:
5964 case Intrinsic::x86_sse2_ucomige_sd:
5965 Opc = X86ISD::UCOMI;
5968 case Intrinsic::x86_sse_ucomineq_ss:
5969 case Intrinsic::x86_sse2_ucomineq_sd:
5970 Opc = X86ISD::UCOMI;
5975 SDValue LHS = Op.getOperand(1);
5976 SDValue RHS = Op.getOperand(2);
5977 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
5978 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
5979 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5980 DAG.getConstant(X86CC, MVT::i8), Cond);
5981 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
5984 // Fix vector shift instructions where the last operand is a non-immediate
5986 case Intrinsic::x86_sse2_pslli_w:
5987 case Intrinsic::x86_sse2_pslli_d:
5988 case Intrinsic::x86_sse2_pslli_q:
5989 case Intrinsic::x86_sse2_psrli_w:
5990 case Intrinsic::x86_sse2_psrli_d:
5991 case Intrinsic::x86_sse2_psrli_q:
5992 case Intrinsic::x86_sse2_psrai_w:
5993 case Intrinsic::x86_sse2_psrai_d:
5994 case Intrinsic::x86_mmx_pslli_w:
5995 case Intrinsic::x86_mmx_pslli_d:
5996 case Intrinsic::x86_mmx_pslli_q:
5997 case Intrinsic::x86_mmx_psrli_w:
5998 case Intrinsic::x86_mmx_psrli_d:
5999 case Intrinsic::x86_mmx_psrli_q:
6000 case Intrinsic::x86_mmx_psrai_w:
6001 case Intrinsic::x86_mmx_psrai_d: {
6002 SDValue ShAmt = Op.getOperand(2);
6003 if (isa<ConstantSDNode>(ShAmt))
6006 unsigned NewIntNo = 0;
6007 MVT ShAmtVT = MVT::v4i32;
6009 case Intrinsic::x86_sse2_pslli_w:
6010 NewIntNo = Intrinsic::x86_sse2_psll_w;
6012 case Intrinsic::x86_sse2_pslli_d:
6013 NewIntNo = Intrinsic::x86_sse2_psll_d;
6015 case Intrinsic::x86_sse2_pslli_q:
6016 NewIntNo = Intrinsic::x86_sse2_psll_q;
6018 case Intrinsic::x86_sse2_psrli_w:
6019 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6021 case Intrinsic::x86_sse2_psrli_d:
6022 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6024 case Intrinsic::x86_sse2_psrli_q:
6025 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6027 case Intrinsic::x86_sse2_psrai_w:
6028 NewIntNo = Intrinsic::x86_sse2_psra_w;
6030 case Intrinsic::x86_sse2_psrai_d:
6031 NewIntNo = Intrinsic::x86_sse2_psra_d;
6034 ShAmtVT = MVT::v2i32;
6036 case Intrinsic::x86_mmx_pslli_w:
6037 NewIntNo = Intrinsic::x86_mmx_psll_w;
6039 case Intrinsic::x86_mmx_pslli_d:
6040 NewIntNo = Intrinsic::x86_mmx_psll_d;
6042 case Intrinsic::x86_mmx_pslli_q:
6043 NewIntNo = Intrinsic::x86_mmx_psll_q;
6045 case Intrinsic::x86_mmx_psrli_w:
6046 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6048 case Intrinsic::x86_mmx_psrli_d:
6049 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6051 case Intrinsic::x86_mmx_psrli_q:
6052 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6054 case Intrinsic::x86_mmx_psrai_w:
6055 NewIntNo = Intrinsic::x86_mmx_psra_w;
6057 case Intrinsic::x86_mmx_psrai_d:
6058 NewIntNo = Intrinsic::x86_mmx_psra_d;
6060 default: abort(); // Can't reach here.
6065 MVT VT = Op.getValueType();
6066 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6067 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6068 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6069 DAG.getConstant(NewIntNo, MVT::i32),
6070 Op.getOperand(1), ShAmt);
6075 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6076 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6077 DebugLoc dl = Op.getDebugLoc();
6080 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6082 DAG.getConstant(TD->getPointerSize(),
6083 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6084 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6085 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6090 // Just load the return address.
6091 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6092 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6093 RetAddrFI, NULL, 0);
6096 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6097 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6098 MFI->setFrameAddressIsTaken(true);
6099 MVT VT = Op.getValueType();
6100 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6101 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6102 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6103 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6105 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6109 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6110 SelectionDAG &DAG) {
6111 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6114 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6116 MachineFunction &MF = DAG.getMachineFunction();
6117 SDValue Chain = Op.getOperand(0);
6118 SDValue Offset = Op.getOperand(1);
6119 SDValue Handler = Op.getOperand(2);
6120 DebugLoc dl = Op.getDebugLoc();
6122 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6124 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6126 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6127 DAG.getIntPtrConstant(-TD->getPointerSize()));
6128 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6129 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6130 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6131 MF.getRegInfo().addLiveOut(StoreAddrReg);
6133 return DAG.getNode(X86ISD::EH_RETURN, dl,
6135 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6138 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6139 SelectionDAG &DAG) {
6140 SDValue Root = Op.getOperand(0);
6141 SDValue Trmp = Op.getOperand(1); // trampoline
6142 SDValue FPtr = Op.getOperand(2); // nested function
6143 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6144 DebugLoc dl = Op.getDebugLoc();
6146 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6148 const X86InstrInfo *TII =
6149 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6151 if (Subtarget->is64Bit()) {
6152 SDValue OutChains[6];
6154 // Large code-model.
6156 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6157 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6159 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6160 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6162 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6164 // Load the pointer to the nested function into R11.
6165 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6166 SDValue Addr = Trmp;
6167 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6170 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6171 DAG.getConstant(2, MVT::i64));
6172 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6174 // Load the 'nest' parameter value into R10.
6175 // R10 is specified in X86CallingConv.td
6176 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6177 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6178 DAG.getConstant(10, MVT::i64));
6179 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6180 Addr, TrmpAddr, 10);
6182 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6183 DAG.getConstant(12, MVT::i64));
6184 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6186 // Jump to the nested function.
6187 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6188 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6189 DAG.getConstant(20, MVT::i64));
6190 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6191 Addr, TrmpAddr, 20);
6193 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6194 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6195 DAG.getConstant(22, MVT::i64));
6196 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6200 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6201 return DAG.getMergeValues(Ops, 2, dl);
6203 const Function *Func =
6204 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6205 unsigned CC = Func->getCallingConv();
6210 assert(0 && "Unsupported calling convention");
6211 case CallingConv::C:
6212 case CallingConv::X86_StdCall: {
6213 // Pass 'nest' parameter in ECX.
6214 // Must be kept in sync with X86CallingConv.td
6217 // Check that ECX wasn't needed by an 'inreg' parameter.
6218 const FunctionType *FTy = Func->getFunctionType();
6219 const AttrListPtr &Attrs = Func->getAttributes();
6221 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6222 unsigned InRegCount = 0;
6225 for (FunctionType::param_iterator I = FTy->param_begin(),
6226 E = FTy->param_end(); I != E; ++I, ++Idx)
6227 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6228 // FIXME: should only count parameters that are lowered to integers.
6229 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6231 if (InRegCount > 2) {
6232 cerr << "Nest register in use - reduce number of inreg parameters!\n";
6238 case CallingConv::X86_FastCall:
6239 case CallingConv::Fast:
6240 // Pass 'nest' parameter in EAX.
6241 // Must be kept in sync with X86CallingConv.td
6246 SDValue OutChains[4];
6249 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6250 DAG.getConstant(10, MVT::i32));
6251 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6253 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6254 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6255 OutChains[0] = DAG.getStore(Root, dl,
6256 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6259 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6260 DAG.getConstant(1, MVT::i32));
6261 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6263 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6264 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6265 DAG.getConstant(5, MVT::i32));
6266 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6267 TrmpAddr, 5, false, 1);
6269 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6270 DAG.getConstant(6, MVT::i32));
6271 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6274 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6275 return DAG.getMergeValues(Ops, 2, dl);
6279 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6281 The rounding mode is in bits 11:10 of FPSR, and has the following
6288 FLT_ROUNDS, on the other hand, expects the following:
6295 To perform the conversion, we do:
6296 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6299 MachineFunction &MF = DAG.getMachineFunction();
6300 const TargetMachine &TM = MF.getTarget();
6301 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6302 unsigned StackAlignment = TFI.getStackAlignment();
6303 MVT VT = Op.getValueType();
6304 DebugLoc dl = Op.getDebugLoc();
6306 // Save FP Control Word to stack slot
6307 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6308 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6310 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6311 DAG.getEntryNode(), StackSlot);
6313 // Load FP Control Word from stack slot
6314 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6316 // Transform as necessary
6318 DAG.getNode(ISD::SRL, dl, MVT::i16,
6319 DAG.getNode(ISD::AND, dl, MVT::i16,
6320 CWD, DAG.getConstant(0x800, MVT::i16)),
6321 DAG.getConstant(11, MVT::i8));
6323 DAG.getNode(ISD::SRL, dl, MVT::i16,
6324 DAG.getNode(ISD::AND, dl, MVT::i16,
6325 CWD, DAG.getConstant(0x400, MVT::i16)),
6326 DAG.getConstant(9, MVT::i8));
6329 DAG.getNode(ISD::AND, dl, MVT::i16,
6330 DAG.getNode(ISD::ADD, dl, MVT::i16,
6331 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6332 DAG.getConstant(1, MVT::i16)),
6333 DAG.getConstant(3, MVT::i16));
6336 return DAG.getNode((VT.getSizeInBits() < 16 ?
6337 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6340 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6341 MVT VT = Op.getValueType();
6343 unsigned NumBits = VT.getSizeInBits();
6344 DebugLoc dl = Op.getDebugLoc();
6346 Op = Op.getOperand(0);
6347 if (VT == MVT::i8) {
6348 // Zero extend to i32 since there is not an i8 bsr.
6350 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6353 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6354 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6355 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6357 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6358 SmallVector<SDValue, 4> Ops;
6360 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6361 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6362 Ops.push_back(Op.getValue(1));
6363 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6365 // Finally xor with NumBits-1.
6366 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6369 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6373 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6374 MVT VT = Op.getValueType();
6376 unsigned NumBits = VT.getSizeInBits();
6377 DebugLoc dl = Op.getDebugLoc();
6379 Op = Op.getOperand(0);
6380 if (VT == MVT::i8) {
6382 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6385 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6386 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6387 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6389 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6390 SmallVector<SDValue, 4> Ops;
6392 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6393 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6394 Ops.push_back(Op.getValue(1));
6395 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6398 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6402 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6403 MVT VT = Op.getValueType();
6404 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6405 DebugLoc dl = Op.getDebugLoc();
6407 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6408 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6409 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6410 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6411 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6413 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6414 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6415 // return AloBlo + AloBhi + AhiBlo;
6417 SDValue A = Op.getOperand(0);
6418 SDValue B = Op.getOperand(1);
6420 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6421 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6422 A, DAG.getConstant(32, MVT::i32));
6423 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6424 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6425 B, DAG.getConstant(32, MVT::i32));
6426 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6427 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6429 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6430 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6432 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6433 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6435 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6436 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6437 AloBhi, DAG.getConstant(32, MVT::i32));
6438 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6439 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6440 AhiBlo, DAG.getConstant(32, MVT::i32));
6441 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6442 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6447 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6448 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6449 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6450 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6451 // has only one use.
6452 SDNode *N = Op.getNode();
6453 SDValue LHS = N->getOperand(0);
6454 SDValue RHS = N->getOperand(1);
6455 unsigned BaseOp = 0;
6457 DebugLoc dl = Op.getDebugLoc();
6459 switch (Op.getOpcode()) {
6460 default: assert(0 && "Unknown ovf instruction!");
6462 // A subtract of one will be selected as a INC. Note that INC doesn't
6463 // set CF, so we can't do this for UADDO.
6464 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6465 if (C->getAPIntValue() == 1) {
6466 BaseOp = X86ISD::INC;
6470 BaseOp = X86ISD::ADD;
6474 BaseOp = X86ISD::ADD;
6478 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6479 // set CF, so we can't do this for USUBO.
6480 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6481 if (C->getAPIntValue() == 1) {
6482 BaseOp = X86ISD::DEC;
6486 BaseOp = X86ISD::SUB;
6490 BaseOp = X86ISD::SUB;
6494 BaseOp = X86ISD::SMUL;
6498 BaseOp = X86ISD::UMUL;
6503 // Also sets EFLAGS.
6504 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6505 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6508 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6509 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6511 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6515 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6516 MVT T = Op.getValueType();
6517 DebugLoc dl = Op.getDebugLoc();
6520 switch(T.getSimpleVT()) {
6522 assert(false && "Invalid value type!");
6523 case MVT::i8: Reg = X86::AL; size = 1; break;
6524 case MVT::i16: Reg = X86::AX; size = 2; break;
6525 case MVT::i32: Reg = X86::EAX; size = 4; break;
6527 assert(Subtarget->is64Bit() && "Node not type legal!");
6528 Reg = X86::RAX; size = 8;
6531 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6532 Op.getOperand(2), SDValue());
6533 SDValue Ops[] = { cpIn.getValue(0),
6536 DAG.getTargetConstant(size, MVT::i8),
6538 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6539 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6541 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6545 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6546 SelectionDAG &DAG) {
6547 assert(Subtarget->is64Bit() && "Result not type legalized?");
6548 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6549 SDValue TheChain = Op.getOperand(0);
6550 DebugLoc dl = Op.getDebugLoc();
6551 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6552 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6553 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6555 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6556 DAG.getConstant(32, MVT::i8));
6558 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6561 return DAG.getMergeValues(Ops, 2, dl);
6564 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6565 SDNode *Node = Op.getNode();
6566 DebugLoc dl = Node->getDebugLoc();
6567 MVT T = Node->getValueType(0);
6568 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6569 DAG.getConstant(0, T), Node->getOperand(2));
6570 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6571 cast<AtomicSDNode>(Node)->getMemoryVT(),
6572 Node->getOperand(0),
6573 Node->getOperand(1), negOp,
6574 cast<AtomicSDNode>(Node)->getSrcValue(),
6575 cast<AtomicSDNode>(Node)->getAlignment());
6578 /// LowerOperation - Provide custom lowering hooks for some operations.
6580 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6581 switch (Op.getOpcode()) {
6582 default: assert(0 && "Should not custom lower this!");
6583 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6584 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6585 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6586 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6587 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6588 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6589 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6590 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6591 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6592 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6593 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6594 case ISD::SHL_PARTS:
6595 case ISD::SRA_PARTS:
6596 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6597 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6598 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6599 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6600 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6601 case ISD::FABS: return LowerFABS(Op, DAG);
6602 case ISD::FNEG: return LowerFNEG(Op, DAG);
6603 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6604 case ISD::SETCC: return LowerSETCC(Op, DAG);
6605 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6606 case ISD::SELECT: return LowerSELECT(Op, DAG);
6607 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6608 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6609 case ISD::CALL: return LowerCALL(Op, DAG);
6610 case ISD::RET: return LowerRET(Op, DAG);
6611 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6612 case ISD::VASTART: return LowerVASTART(Op, DAG);
6613 case ISD::VAARG: return LowerVAARG(Op, DAG);
6614 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6615 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6616 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6617 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6618 case ISD::FRAME_TO_ARGS_OFFSET:
6619 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6620 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6621 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6622 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6623 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6624 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6625 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6626 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6632 case ISD::UMULO: return LowerXALUO(Op, DAG);
6633 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6637 void X86TargetLowering::
6638 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6639 SelectionDAG &DAG, unsigned NewOp) {
6640 MVT T = Node->getValueType(0);
6641 DebugLoc dl = Node->getDebugLoc();
6642 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6644 SDValue Chain = Node->getOperand(0);
6645 SDValue In1 = Node->getOperand(1);
6646 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6647 Node->getOperand(2), DAG.getIntPtrConstant(0));
6648 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6649 Node->getOperand(2), DAG.getIntPtrConstant(1));
6650 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6651 // have a MemOperand. Pass the info through as a normal operand.
6652 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6653 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6654 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6655 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6656 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6657 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6658 Results.push_back(Result.getValue(2));
6661 /// ReplaceNodeResults - Replace a node with an illegal result type
6662 /// with a new node built out of custom code.
6663 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6664 SmallVectorImpl<SDValue>&Results,
6665 SelectionDAG &DAG) {
6666 DebugLoc dl = N->getDebugLoc();
6667 switch (N->getOpcode()) {
6669 assert(false && "Do not know how to custom type legalize this operation!");
6671 case ISD::FP_TO_SINT: {
6672 std::pair<SDValue,SDValue> Vals =
6673 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
6674 SDValue FIST = Vals.first, StackSlot = Vals.second;
6675 if (FIST.getNode() != 0) {
6676 MVT VT = N->getValueType(0);
6677 // Return a load from the stack slot.
6678 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6682 case ISD::READCYCLECOUNTER: {
6683 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6684 SDValue TheChain = N->getOperand(0);
6685 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6686 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6688 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6690 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6691 SDValue Ops[] = { eax, edx };
6692 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6693 Results.push_back(edx.getValue(1));
6696 case ISD::ATOMIC_CMP_SWAP: {
6697 MVT T = N->getValueType(0);
6698 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6699 SDValue cpInL, cpInH;
6700 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6701 DAG.getConstant(0, MVT::i32));
6702 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6703 DAG.getConstant(1, MVT::i32));
6704 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6705 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6707 SDValue swapInL, swapInH;
6708 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6709 DAG.getConstant(0, MVT::i32));
6710 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6711 DAG.getConstant(1, MVT::i32));
6712 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6714 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6715 swapInL.getValue(1));
6716 SDValue Ops[] = { swapInH.getValue(0),
6718 swapInH.getValue(1) };
6719 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6720 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6721 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6722 MVT::i32, Result.getValue(1));
6723 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6724 MVT::i32, cpOutL.getValue(2));
6725 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6726 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6727 Results.push_back(cpOutH.getValue(1));
6730 case ISD::ATOMIC_LOAD_ADD:
6731 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6733 case ISD::ATOMIC_LOAD_AND:
6734 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6736 case ISD::ATOMIC_LOAD_NAND:
6737 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6739 case ISD::ATOMIC_LOAD_OR:
6740 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6742 case ISD::ATOMIC_LOAD_SUB:
6743 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6745 case ISD::ATOMIC_LOAD_XOR:
6746 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6748 case ISD::ATOMIC_SWAP:
6749 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6754 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6756 default: return NULL;
6757 case X86ISD::BSF: return "X86ISD::BSF";
6758 case X86ISD::BSR: return "X86ISD::BSR";
6759 case X86ISD::SHLD: return "X86ISD::SHLD";
6760 case X86ISD::SHRD: return "X86ISD::SHRD";
6761 case X86ISD::FAND: return "X86ISD::FAND";
6762 case X86ISD::FOR: return "X86ISD::FOR";
6763 case X86ISD::FXOR: return "X86ISD::FXOR";
6764 case X86ISD::FSRL: return "X86ISD::FSRL";
6765 case X86ISD::FILD: return "X86ISD::FILD";
6766 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6767 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6768 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6769 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6770 case X86ISD::FLD: return "X86ISD::FLD";
6771 case X86ISD::FST: return "X86ISD::FST";
6772 case X86ISD::CALL: return "X86ISD::CALL";
6773 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
6774 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
6775 case X86ISD::BT: return "X86ISD::BT";
6776 case X86ISD::CMP: return "X86ISD::CMP";
6777 case X86ISD::COMI: return "X86ISD::COMI";
6778 case X86ISD::UCOMI: return "X86ISD::UCOMI";
6779 case X86ISD::SETCC: return "X86ISD::SETCC";
6780 case X86ISD::CMOV: return "X86ISD::CMOV";
6781 case X86ISD::BRCOND: return "X86ISD::BRCOND";
6782 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
6783 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
6784 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
6785 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
6786 case X86ISD::Wrapper: return "X86ISD::Wrapper";
6787 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
6788 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
6789 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
6790 case X86ISD::PINSRB: return "X86ISD::PINSRB";
6791 case X86ISD::PINSRW: return "X86ISD::PINSRW";
6792 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
6793 case X86ISD::FMAX: return "X86ISD::FMAX";
6794 case X86ISD::FMIN: return "X86ISD::FMIN";
6795 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
6796 case X86ISD::FRCP: return "X86ISD::FRCP";
6797 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
6798 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
6799 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
6800 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
6801 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
6802 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
6803 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
6804 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
6805 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
6806 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
6807 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
6808 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
6809 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
6810 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
6811 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
6812 case X86ISD::VSHL: return "X86ISD::VSHL";
6813 case X86ISD::VSRL: return "X86ISD::VSRL";
6814 case X86ISD::CMPPD: return "X86ISD::CMPPD";
6815 case X86ISD::CMPPS: return "X86ISD::CMPPS";
6816 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
6817 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
6818 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
6819 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
6820 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
6821 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
6822 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
6823 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
6824 case X86ISD::ADD: return "X86ISD::ADD";
6825 case X86ISD::SUB: return "X86ISD::SUB";
6826 case X86ISD::SMUL: return "X86ISD::SMUL";
6827 case X86ISD::UMUL: return "X86ISD::UMUL";
6828 case X86ISD::INC: return "X86ISD::INC";
6829 case X86ISD::DEC: return "X86ISD::DEC";
6830 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
6834 // isLegalAddressingMode - Return true if the addressing mode represented
6835 // by AM is legal for this target, for a load/store of the specified type.
6836 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6837 const Type *Ty) const {
6838 // X86 supports extremely general addressing modes.
6840 // X86 allows a sign-extended 32-bit immediate field as a displacement.
6841 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
6845 // We can only fold this if we don't need an extra load.
6846 if (Subtarget->GVRequiresExtraLoad(AM.BaseGV, getTargetMachine(), false))
6848 // If BaseGV requires a register, we cannot also have a BaseReg.
6849 if (Subtarget->GVRequiresRegister(AM.BaseGV, getTargetMachine(), false) &&
6853 // X86-64 only supports addr of globals in small code model.
6854 if (Subtarget->is64Bit()) {
6855 if (getTargetMachine().getCodeModel() != CodeModel::Small)
6857 // If lower 4G is not available, then we must use rip-relative addressing.
6858 if (AM.BaseOffs || AM.Scale > 1)
6869 // These scales always work.
6874 // These scales are formed with basereg+scalereg. Only accept if there is
6879 default: // Other stuff never works.
6887 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
6888 if (!Ty1->isInteger() || !Ty2->isInteger())
6890 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6891 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6892 if (NumBits1 <= NumBits2)
6894 return Subtarget->is64Bit() || NumBits1 < 64;
6897 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
6898 if (!VT1.isInteger() || !VT2.isInteger())
6900 unsigned NumBits1 = VT1.getSizeInBits();
6901 unsigned NumBits2 = VT2.getSizeInBits();
6902 if (NumBits1 <= NumBits2)
6904 return Subtarget->is64Bit() || NumBits1 < 64;
6907 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
6908 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
6909 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
6912 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
6913 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
6914 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
6917 bool X86TargetLowering::isNarrowingProfitable(MVT VT1, MVT VT2) const {
6918 // i16 instructions are longer (0x66 prefix) and potentially slower.
6919 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
6922 /// isShuffleMaskLegal - Targets can use this to indicate that they only
6923 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
6924 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
6925 /// are assumed to be legal.
6927 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6929 // Only do shuffles on 128-bit vector types for now.
6930 if (VT.getSizeInBits() == 64)
6933 // FIXME: pshufb, blends, palignr, shifts.
6934 return (VT.getVectorNumElements() == 2 ||
6935 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
6936 isMOVLMask(M, VT) ||
6937 isSHUFPMask(M, VT) ||
6938 isPSHUFDMask(M, VT) ||
6939 isPSHUFHWMask(M, VT) ||
6940 isPSHUFLWMask(M, VT) ||
6941 isUNPCKLMask(M, VT) ||
6942 isUNPCKHMask(M, VT) ||
6943 isUNPCKL_v_undef_Mask(M, VT) ||
6944 isUNPCKH_v_undef_Mask(M, VT));
6948 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
6950 unsigned NumElts = VT.getVectorNumElements();
6951 // FIXME: This collection of masks seems suspect.
6954 if (NumElts == 4 && VT.getSizeInBits() == 128) {
6955 return (isMOVLMask(Mask, VT) ||
6956 isCommutedMOVLMask(Mask, VT, true) ||
6957 isSHUFPMask(Mask, VT) ||
6958 isCommutedSHUFPMask(Mask, VT));
6963 //===----------------------------------------------------------------------===//
6964 // X86 Scheduler Hooks
6965 //===----------------------------------------------------------------------===//
6967 // private utility function
6969 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
6970 MachineBasicBlock *MBB,
6978 TargetRegisterClass *RC,
6979 bool invSrc) const {
6980 // For the atomic bitwise operator, we generate
6983 // ld t1 = [bitinstr.addr]
6984 // op t2 = t1, [bitinstr.val]
6986 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
6988 // fallthrough -->nextMBB
6989 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6990 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
6991 MachineFunction::iterator MBBIter = MBB;
6994 /// First build the CFG
6995 MachineFunction *F = MBB->getParent();
6996 MachineBasicBlock *thisMBB = MBB;
6997 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
6998 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
6999 F->insert(MBBIter, newMBB);
7000 F->insert(MBBIter, nextMBB);
7002 // Move all successors to thisMBB to nextMBB
7003 nextMBB->transferSuccessors(thisMBB);
7005 // Update thisMBB to fall through to newMBB
7006 thisMBB->addSuccessor(newMBB);
7008 // newMBB jumps to itself and fall through to nextMBB
7009 newMBB->addSuccessor(nextMBB);
7010 newMBB->addSuccessor(newMBB);
7012 // Insert instructions into newMBB based on incoming instruction
7013 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7014 "unexpected number of operands");
7015 DebugLoc dl = bInstr->getDebugLoc();
7016 MachineOperand& destOper = bInstr->getOperand(0);
7017 MachineOperand* argOpers[2 + X86AddrNumOperands];
7018 int numArgs = bInstr->getNumOperands() - 1;
7019 for (int i=0; i < numArgs; ++i)
7020 argOpers[i] = &bInstr->getOperand(i+1);
7022 // x86 address has 4 operands: base, index, scale, and displacement
7023 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7024 int valArgIndx = lastAddrIndx + 1;
7026 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7027 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7028 for (int i=0; i <= lastAddrIndx; ++i)
7029 (*MIB).addOperand(*argOpers[i]);
7031 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7033 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7038 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7039 assert((argOpers[valArgIndx]->isReg() ||
7040 argOpers[valArgIndx]->isImm()) &&
7042 if (argOpers[valArgIndx]->isReg())
7043 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7045 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7047 (*MIB).addOperand(*argOpers[valArgIndx]);
7049 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7052 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7053 for (int i=0; i <= lastAddrIndx; ++i)
7054 (*MIB).addOperand(*argOpers[i]);
7056 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7057 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7059 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7063 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7065 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7069 // private utility function: 64 bit atomics on 32 bit host.
7071 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7072 MachineBasicBlock *MBB,
7077 bool invSrc) const {
7078 // For the atomic bitwise operator, we generate
7079 // thisMBB (instructions are in pairs, except cmpxchg8b)
7080 // ld t1,t2 = [bitinstr.addr]
7082 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7083 // op t5, t6 <- out1, out2, [bitinstr.val]
7084 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7085 // mov ECX, EBX <- t5, t6
7086 // mov EAX, EDX <- t1, t2
7087 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7088 // mov t3, t4 <- EAX, EDX
7090 // result in out1, out2
7091 // fallthrough -->nextMBB
7093 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7094 const unsigned LoadOpc = X86::MOV32rm;
7095 const unsigned copyOpc = X86::MOV32rr;
7096 const unsigned NotOpc = X86::NOT32r;
7097 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7098 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7099 MachineFunction::iterator MBBIter = MBB;
7102 /// First build the CFG
7103 MachineFunction *F = MBB->getParent();
7104 MachineBasicBlock *thisMBB = MBB;
7105 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7106 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7107 F->insert(MBBIter, newMBB);
7108 F->insert(MBBIter, nextMBB);
7110 // Move all successors to thisMBB to nextMBB
7111 nextMBB->transferSuccessors(thisMBB);
7113 // Update thisMBB to fall through to newMBB
7114 thisMBB->addSuccessor(newMBB);
7116 // newMBB jumps to itself and fall through to nextMBB
7117 newMBB->addSuccessor(nextMBB);
7118 newMBB->addSuccessor(newMBB);
7120 DebugLoc dl = bInstr->getDebugLoc();
7121 // Insert instructions into newMBB based on incoming instruction
7122 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7123 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7124 "unexpected number of operands");
7125 MachineOperand& dest1Oper = bInstr->getOperand(0);
7126 MachineOperand& dest2Oper = bInstr->getOperand(1);
7127 MachineOperand* argOpers[2 + X86AddrNumOperands];
7128 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7129 argOpers[i] = &bInstr->getOperand(i+2);
7131 // x86 address has 4 operands: base, index, scale, and displacement
7132 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7134 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7135 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7136 for (int i=0; i <= lastAddrIndx; ++i)
7137 (*MIB).addOperand(*argOpers[i]);
7138 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7139 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7140 // add 4 to displacement.
7141 for (int i=0; i <= lastAddrIndx-2; ++i)
7142 (*MIB).addOperand(*argOpers[i]);
7143 MachineOperand newOp3 = *(argOpers[3]);
7145 newOp3.setImm(newOp3.getImm()+4);
7147 newOp3.setOffset(newOp3.getOffset()+4);
7148 (*MIB).addOperand(newOp3);
7149 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7151 // t3/4 are defined later, at the bottom of the loop
7152 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7153 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7154 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7155 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7156 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7157 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7159 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7160 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7162 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7163 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7169 int valArgIndx = lastAddrIndx + 1;
7170 assert((argOpers[valArgIndx]->isReg() ||
7171 argOpers[valArgIndx]->isImm()) &&
7173 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7174 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7175 if (argOpers[valArgIndx]->isReg())
7176 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7178 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7179 if (regOpcL != X86::MOV32rr)
7181 (*MIB).addOperand(*argOpers[valArgIndx]);
7182 assert(argOpers[valArgIndx + 1]->isReg() ==
7183 argOpers[valArgIndx]->isReg());
7184 assert(argOpers[valArgIndx + 1]->isImm() ==
7185 argOpers[valArgIndx]->isImm());
7186 if (argOpers[valArgIndx + 1]->isReg())
7187 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7189 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7190 if (regOpcH != X86::MOV32rr)
7192 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7194 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7196 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7199 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7201 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7204 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7205 for (int i=0; i <= lastAddrIndx; ++i)
7206 (*MIB).addOperand(*argOpers[i]);
7208 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7209 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7211 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7212 MIB.addReg(X86::EAX);
7213 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7214 MIB.addReg(X86::EDX);
7217 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7219 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7223 // private utility function
7225 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7226 MachineBasicBlock *MBB,
7227 unsigned cmovOpc) const {
7228 // For the atomic min/max operator, we generate
7231 // ld t1 = [min/max.addr]
7232 // mov t2 = [min/max.val]
7234 // cmov[cond] t2 = t1
7236 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7238 // fallthrough -->nextMBB
7240 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7241 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7242 MachineFunction::iterator MBBIter = MBB;
7245 /// First build the CFG
7246 MachineFunction *F = MBB->getParent();
7247 MachineBasicBlock *thisMBB = MBB;
7248 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7249 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7250 F->insert(MBBIter, newMBB);
7251 F->insert(MBBIter, nextMBB);
7253 // Move all successors to thisMBB to nextMBB
7254 nextMBB->transferSuccessors(thisMBB);
7256 // Update thisMBB to fall through to newMBB
7257 thisMBB->addSuccessor(newMBB);
7259 // newMBB jumps to newMBB and fall through to nextMBB
7260 newMBB->addSuccessor(nextMBB);
7261 newMBB->addSuccessor(newMBB);
7263 DebugLoc dl = mInstr->getDebugLoc();
7264 // Insert instructions into newMBB based on incoming instruction
7265 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7266 "unexpected number of operands");
7267 MachineOperand& destOper = mInstr->getOperand(0);
7268 MachineOperand* argOpers[2 + X86AddrNumOperands];
7269 int numArgs = mInstr->getNumOperands() - 1;
7270 for (int i=0; i < numArgs; ++i)
7271 argOpers[i] = &mInstr->getOperand(i+1);
7273 // x86 address has 4 operands: base, index, scale, and displacement
7274 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7275 int valArgIndx = lastAddrIndx + 1;
7277 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7278 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7279 for (int i=0; i <= lastAddrIndx; ++i)
7280 (*MIB).addOperand(*argOpers[i]);
7282 // We only support register and immediate values
7283 assert((argOpers[valArgIndx]->isReg() ||
7284 argOpers[valArgIndx]->isImm()) &&
7287 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7288 if (argOpers[valArgIndx]->isReg())
7289 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7291 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7292 (*MIB).addOperand(*argOpers[valArgIndx]);
7294 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7297 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7302 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7303 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7307 // Cmp and exchange if none has modified the memory location
7308 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7309 for (int i=0; i <= lastAddrIndx; ++i)
7310 (*MIB).addOperand(*argOpers[i]);
7312 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7313 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7315 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7316 MIB.addReg(X86::EAX);
7319 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7321 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7327 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7328 MachineBasicBlock *BB) const {
7329 DebugLoc dl = MI->getDebugLoc();
7330 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7331 switch (MI->getOpcode()) {
7332 default: assert(false && "Unexpected instr type to insert");
7333 case X86::CMOV_V1I64:
7334 case X86::CMOV_FR32:
7335 case X86::CMOV_FR64:
7336 case X86::CMOV_V4F32:
7337 case X86::CMOV_V2F64:
7338 case X86::CMOV_V2I64: {
7339 // To "insert" a SELECT_CC instruction, we actually have to insert the
7340 // diamond control-flow pattern. The incoming instruction knows the
7341 // destination vreg to set, the condition code register to branch on, the
7342 // true/false values to select between, and a branch opcode to use.
7343 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7344 MachineFunction::iterator It = BB;
7350 // cmpTY ccX, r1, r2
7352 // fallthrough --> copy0MBB
7353 MachineBasicBlock *thisMBB = BB;
7354 MachineFunction *F = BB->getParent();
7355 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7356 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7358 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7359 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7360 F->insert(It, copy0MBB);
7361 F->insert(It, sinkMBB);
7362 // Update machine-CFG edges by transferring all successors of the current
7363 // block to the new block which will contain the Phi node for the select.
7364 sinkMBB->transferSuccessors(BB);
7366 // Add the true and fallthrough blocks as its successors.
7367 BB->addSuccessor(copy0MBB);
7368 BB->addSuccessor(sinkMBB);
7371 // %FalseValue = ...
7372 // # fallthrough to sinkMBB
7375 // Update machine-CFG edges
7376 BB->addSuccessor(sinkMBB);
7379 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7382 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7383 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7384 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7386 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7390 case X86::FP32_TO_INT16_IN_MEM:
7391 case X86::FP32_TO_INT32_IN_MEM:
7392 case X86::FP32_TO_INT64_IN_MEM:
7393 case X86::FP64_TO_INT16_IN_MEM:
7394 case X86::FP64_TO_INT32_IN_MEM:
7395 case X86::FP64_TO_INT64_IN_MEM:
7396 case X86::FP80_TO_INT16_IN_MEM:
7397 case X86::FP80_TO_INT32_IN_MEM:
7398 case X86::FP80_TO_INT64_IN_MEM: {
7399 // Change the floating point control register to use "round towards zero"
7400 // mode when truncating to an integer value.
7401 MachineFunction *F = BB->getParent();
7402 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7403 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7405 // Load the old value of the high byte of the control word...
7407 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7408 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7411 // Set the high part to be round to zero...
7412 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7415 // Reload the modified control word now...
7416 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7418 // Restore the memory image of control word to original value
7419 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7422 // Get the X86 opcode to use.
7424 switch (MI->getOpcode()) {
7425 default: assert(0 && "illegal opcode!");
7426 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7427 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7428 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7429 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7430 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7431 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7432 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7433 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7434 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7438 MachineOperand &Op = MI->getOperand(0);
7440 AM.BaseType = X86AddressMode::RegBase;
7441 AM.Base.Reg = Op.getReg();
7443 AM.BaseType = X86AddressMode::FrameIndexBase;
7444 AM.Base.FrameIndex = Op.getIndex();
7446 Op = MI->getOperand(1);
7448 AM.Scale = Op.getImm();
7449 Op = MI->getOperand(2);
7451 AM.IndexReg = Op.getImm();
7452 Op = MI->getOperand(3);
7453 if (Op.isGlobal()) {
7454 AM.GV = Op.getGlobal();
7456 AM.Disp = Op.getImm();
7458 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7459 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7461 // Reload the original control word now.
7462 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7464 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7467 case X86::ATOMAND32:
7468 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7469 X86::AND32ri, X86::MOV32rm,
7470 X86::LCMPXCHG32, X86::MOV32rr,
7471 X86::NOT32r, X86::EAX,
7472 X86::GR32RegisterClass);
7474 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7475 X86::OR32ri, X86::MOV32rm,
7476 X86::LCMPXCHG32, X86::MOV32rr,
7477 X86::NOT32r, X86::EAX,
7478 X86::GR32RegisterClass);
7479 case X86::ATOMXOR32:
7480 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7481 X86::XOR32ri, X86::MOV32rm,
7482 X86::LCMPXCHG32, X86::MOV32rr,
7483 X86::NOT32r, X86::EAX,
7484 X86::GR32RegisterClass);
7485 case X86::ATOMNAND32:
7486 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7487 X86::AND32ri, X86::MOV32rm,
7488 X86::LCMPXCHG32, X86::MOV32rr,
7489 X86::NOT32r, X86::EAX,
7490 X86::GR32RegisterClass, true);
7491 case X86::ATOMMIN32:
7492 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7493 case X86::ATOMMAX32:
7494 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7495 case X86::ATOMUMIN32:
7496 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7497 case X86::ATOMUMAX32:
7498 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7500 case X86::ATOMAND16:
7501 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7502 X86::AND16ri, X86::MOV16rm,
7503 X86::LCMPXCHG16, X86::MOV16rr,
7504 X86::NOT16r, X86::AX,
7505 X86::GR16RegisterClass);
7507 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7508 X86::OR16ri, X86::MOV16rm,
7509 X86::LCMPXCHG16, X86::MOV16rr,
7510 X86::NOT16r, X86::AX,
7511 X86::GR16RegisterClass);
7512 case X86::ATOMXOR16:
7513 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7514 X86::XOR16ri, X86::MOV16rm,
7515 X86::LCMPXCHG16, X86::MOV16rr,
7516 X86::NOT16r, X86::AX,
7517 X86::GR16RegisterClass);
7518 case X86::ATOMNAND16:
7519 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7520 X86::AND16ri, X86::MOV16rm,
7521 X86::LCMPXCHG16, X86::MOV16rr,
7522 X86::NOT16r, X86::AX,
7523 X86::GR16RegisterClass, true);
7524 case X86::ATOMMIN16:
7525 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7526 case X86::ATOMMAX16:
7527 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7528 case X86::ATOMUMIN16:
7529 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7530 case X86::ATOMUMAX16:
7531 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7534 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7535 X86::AND8ri, X86::MOV8rm,
7536 X86::LCMPXCHG8, X86::MOV8rr,
7537 X86::NOT8r, X86::AL,
7538 X86::GR8RegisterClass);
7540 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7541 X86::OR8ri, X86::MOV8rm,
7542 X86::LCMPXCHG8, X86::MOV8rr,
7543 X86::NOT8r, X86::AL,
7544 X86::GR8RegisterClass);
7546 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7547 X86::XOR8ri, X86::MOV8rm,
7548 X86::LCMPXCHG8, X86::MOV8rr,
7549 X86::NOT8r, X86::AL,
7550 X86::GR8RegisterClass);
7551 case X86::ATOMNAND8:
7552 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7553 X86::AND8ri, X86::MOV8rm,
7554 X86::LCMPXCHG8, X86::MOV8rr,
7555 X86::NOT8r, X86::AL,
7556 X86::GR8RegisterClass, true);
7557 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7558 // This group is for 64-bit host.
7559 case X86::ATOMAND64:
7560 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7561 X86::AND64ri32, X86::MOV64rm,
7562 X86::LCMPXCHG64, X86::MOV64rr,
7563 X86::NOT64r, X86::RAX,
7564 X86::GR64RegisterClass);
7566 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7567 X86::OR64ri32, X86::MOV64rm,
7568 X86::LCMPXCHG64, X86::MOV64rr,
7569 X86::NOT64r, X86::RAX,
7570 X86::GR64RegisterClass);
7571 case X86::ATOMXOR64:
7572 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7573 X86::XOR64ri32, X86::MOV64rm,
7574 X86::LCMPXCHG64, X86::MOV64rr,
7575 X86::NOT64r, X86::RAX,
7576 X86::GR64RegisterClass);
7577 case X86::ATOMNAND64:
7578 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7579 X86::AND64ri32, X86::MOV64rm,
7580 X86::LCMPXCHG64, X86::MOV64rr,
7581 X86::NOT64r, X86::RAX,
7582 X86::GR64RegisterClass, true);
7583 case X86::ATOMMIN64:
7584 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7585 case X86::ATOMMAX64:
7586 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7587 case X86::ATOMUMIN64:
7588 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7589 case X86::ATOMUMAX64:
7590 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7592 // This group does 64-bit operations on a 32-bit host.
7593 case X86::ATOMAND6432:
7594 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7595 X86::AND32rr, X86::AND32rr,
7596 X86::AND32ri, X86::AND32ri,
7598 case X86::ATOMOR6432:
7599 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7600 X86::OR32rr, X86::OR32rr,
7601 X86::OR32ri, X86::OR32ri,
7603 case X86::ATOMXOR6432:
7604 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7605 X86::XOR32rr, X86::XOR32rr,
7606 X86::XOR32ri, X86::XOR32ri,
7608 case X86::ATOMNAND6432:
7609 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7610 X86::AND32rr, X86::AND32rr,
7611 X86::AND32ri, X86::AND32ri,
7613 case X86::ATOMADD6432:
7614 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7615 X86::ADD32rr, X86::ADC32rr,
7616 X86::ADD32ri, X86::ADC32ri,
7618 case X86::ATOMSUB6432:
7619 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7620 X86::SUB32rr, X86::SBB32rr,
7621 X86::SUB32ri, X86::SBB32ri,
7623 case X86::ATOMSWAP6432:
7624 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7625 X86::MOV32rr, X86::MOV32rr,
7626 X86::MOV32ri, X86::MOV32ri,
7631 //===----------------------------------------------------------------------===//
7632 // X86 Optimization Hooks
7633 //===----------------------------------------------------------------------===//
7635 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7639 const SelectionDAG &DAG,
7640 unsigned Depth) const {
7641 unsigned Opc = Op.getOpcode();
7642 assert((Opc >= ISD::BUILTIN_OP_END ||
7643 Opc == ISD::INTRINSIC_WO_CHAIN ||
7644 Opc == ISD::INTRINSIC_W_CHAIN ||
7645 Opc == ISD::INTRINSIC_VOID) &&
7646 "Should use MaskedValueIsZero if you don't know whether Op"
7647 " is a target node!");
7649 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7658 // These nodes' second result is a boolean.
7659 if (Op.getResNo() == 0)
7663 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7664 Mask.getBitWidth() - 1);
7669 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7670 /// node is a GlobalAddress + offset.
7671 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7672 GlobalValue* &GA, int64_t &Offset) const{
7673 if (N->getOpcode() == X86ISD::Wrapper) {
7674 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7675 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7676 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7680 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7683 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7684 const TargetLowering &TLI) {
7687 if (TLI.isGAPlusOffset(Base, GV, Offset))
7688 return (GV->getAlignment() >= N && (Offset % N) == 0);
7689 // DAG combine handles the stack object case.
7693 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
7694 MVT EVT, LoadSDNode *&LDBase,
7695 unsigned &LastLoadedElt,
7696 SelectionDAG &DAG, MachineFrameInfo *MFI,
7697 const TargetLowering &TLI) {
7700 for (unsigned i = 0; i < NumElems; ++i) {
7701 if (N->getMaskElt(i) < 0) {
7707 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7708 if (!Elt.getNode() ||
7709 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7712 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
7714 LDBase = cast<LoadSDNode>(Elt.getNode());
7718 if (Elt.getOpcode() == ISD::UNDEF)
7721 LoadSDNode *LD = cast<LoadSDNode>(Elt);
7722 if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
7729 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7730 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7731 /// if the load addresses are consecutive, non-overlapping, and in the right
7732 /// order. In the case of v2i64, it will see if it can rewrite the
7733 /// shuffle to be an appropriate build vector so it can take advantage of
7734 // performBuildVectorCombine.
7735 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7736 const TargetLowering &TLI) {
7737 DebugLoc dl = N->getDebugLoc();
7738 MVT VT = N->getValueType(0);
7739 MVT EVT = VT.getVectorElementType();
7740 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7741 unsigned NumElems = VT.getVectorNumElements();
7743 if (VT.getSizeInBits() != 128)
7746 // For x86-32 machines, if we see an insert and then a shuffle in a v2i64
7747 // where the upper half is 0, it is advantageous to rewrite it as a build
7748 // vector of (0, val) so it can use movq.
7749 if (VT == MVT::v2i64) {
7751 In[0] = N->getOperand(0);
7752 In[1] = N->getOperand(1);
7753 int Idx0 = SVN->getMaskElt(0);
7754 int Idx1 = SVN->getMaskElt(1);
7755 // FIXME: can we take advantage of undef index?
7756 if (Idx0 >= 0 && Idx1 >= 0 &&
7757 In[Idx0/2].getOpcode() == ISD::INSERT_VECTOR_ELT &&
7758 In[Idx1/2].getOpcode() == ISD::BUILD_VECTOR) {
7759 ConstantSDNode* InsertVecIdx =
7760 dyn_cast<ConstantSDNode>(In[Idx0/2].getOperand(2));
7762 InsertVecIdx->getZExtValue() == (unsigned)(Idx0 % 2) &&
7763 isZeroNode(In[Idx1/2].getOperand(Idx1 % 2))) {
7764 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
7765 In[Idx0/2].getOperand(1),
7766 In[Idx1/2].getOperand(Idx1 % 2));
7771 // Try to combine a vector_shuffle into a 128-bit load.
7772 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7773 LoadSDNode *LD = NULL;
7774 unsigned LastLoadedElt;
7775 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
7779 if (LastLoadedElt == NumElems - 1) {
7780 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
7781 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7782 LD->getSrcValue(), LD->getSrcValueOffset(),
7784 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7785 LD->getSrcValue(), LD->getSrcValueOffset(),
7786 LD->isVolatile(), LD->getAlignment());
7787 } else if (NumElems == 4 && LastLoadedElt == 1) {
7788 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
7789 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7790 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
7791 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
7796 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
7797 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
7798 const X86Subtarget *Subtarget) {
7799 DebugLoc DL = N->getDebugLoc();
7800 SDValue Cond = N->getOperand(0);
7801 // Get the LHS/RHS of the select.
7802 SDValue LHS = N->getOperand(1);
7803 SDValue RHS = N->getOperand(2);
7805 // If we have SSE[12] support, try to form min/max nodes.
7806 if (Subtarget->hasSSE2() &&
7807 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
7808 Cond.getOpcode() == ISD::SETCC) {
7809 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
7811 unsigned Opcode = 0;
7812 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
7815 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
7818 if (!UnsafeFPMath) break;
7820 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
7822 Opcode = X86ISD::FMIN;
7825 case ISD::SETOGT: // (X > Y) ? X : Y -> max
7828 if (!UnsafeFPMath) break;
7830 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
7832 Opcode = X86ISD::FMAX;
7835 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
7838 case ISD::SETOGT: // (X > Y) ? Y : X -> min
7841 if (!UnsafeFPMath) break;
7843 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
7845 Opcode = X86ISD::FMIN;
7848 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
7851 if (!UnsafeFPMath) break;
7853 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
7855 Opcode = X86ISD::FMAX;
7861 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
7864 // If this is a select between two integer constants, try to do some
7866 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
7867 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
7868 // Don't do this for crazy integer types.
7869 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
7870 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
7871 // so that TrueC (the true value) is larger than FalseC.
7872 bool NeedsCondInvert = false;
7874 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
7875 // Efficiently invertible.
7876 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
7877 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
7878 isa<ConstantSDNode>(Cond.getOperand(1))))) {
7879 NeedsCondInvert = true;
7880 std::swap(TrueC, FalseC);
7883 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
7884 if (FalseC->getAPIntValue() == 0 &&
7885 TrueC->getAPIntValue().isPowerOf2()) {
7886 if (NeedsCondInvert) // Invert the condition if needed.
7887 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7888 DAG.getConstant(1, Cond.getValueType()));
7890 // Zero extend the condition if needed.
7891 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
7893 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
7894 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
7895 DAG.getConstant(ShAmt, MVT::i8));
7898 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
7899 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
7900 if (NeedsCondInvert) // Invert the condition if needed.
7901 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7902 DAG.getConstant(1, Cond.getValueType()));
7904 // Zero extend the condition if needed.
7905 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
7906 FalseC->getValueType(0), Cond);
7907 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
7908 SDValue(FalseC, 0));
7911 // Optimize cases that will turn into an LEA instruction. This requires
7912 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
7913 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
7914 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
7915 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
7917 bool isFastMultiplier = false;
7919 switch ((unsigned char)Diff) {
7921 case 1: // result = add base, cond
7922 case 2: // result = lea base( , cond*2)
7923 case 3: // result = lea base(cond, cond*2)
7924 case 4: // result = lea base( , cond*4)
7925 case 5: // result = lea base(cond, cond*4)
7926 case 8: // result = lea base( , cond*8)
7927 case 9: // result = lea base(cond, cond*8)
7928 isFastMultiplier = true;
7933 if (isFastMultiplier) {
7934 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
7935 if (NeedsCondInvert) // Invert the condition if needed.
7936 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
7937 DAG.getConstant(1, Cond.getValueType()));
7939 // Zero extend the condition if needed.
7940 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
7942 // Scale the condition by the difference.
7944 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
7945 DAG.getConstant(Diff, Cond.getValueType()));
7947 // Add the base if non-zero.
7948 if (FalseC->getAPIntValue() != 0)
7949 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
7950 SDValue(FalseC, 0));
7960 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
7961 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
7962 TargetLowering::DAGCombinerInfo &DCI) {
7963 DebugLoc DL = N->getDebugLoc();
7965 // If the flag operand isn't dead, don't touch this CMOV.
7966 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
7969 // If this is a select between two integer constants, try to do some
7970 // optimizations. Note that the operands are ordered the opposite of SELECT
7972 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
7973 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
7974 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
7975 // larger than FalseC (the false value).
7976 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
7978 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
7979 CC = X86::GetOppositeBranchCondition(CC);
7980 std::swap(TrueC, FalseC);
7983 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
7984 // This is efficient for any integer data type (including i8/i16) and
7986 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
7987 SDValue Cond = N->getOperand(3);
7988 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
7989 DAG.getConstant(CC, MVT::i8), Cond);
7991 // Zero extend the condition if needed.
7992 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
7994 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
7995 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
7996 DAG.getConstant(ShAmt, MVT::i8));
7997 if (N->getNumValues() == 2) // Dead flag value?
7998 return DCI.CombineTo(N, Cond, SDValue());
8002 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8003 // for any integer data type, including i8/i16.
8004 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8005 SDValue Cond = N->getOperand(3);
8006 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8007 DAG.getConstant(CC, MVT::i8), Cond);
8009 // Zero extend the condition if needed.
8010 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8011 FalseC->getValueType(0), Cond);
8012 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8013 SDValue(FalseC, 0));
8015 if (N->getNumValues() == 2) // Dead flag value?
8016 return DCI.CombineTo(N, Cond, SDValue());
8020 // Optimize cases that will turn into an LEA instruction. This requires
8021 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8022 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8023 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8024 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8026 bool isFastMultiplier = false;
8028 switch ((unsigned char)Diff) {
8030 case 1: // result = add base, cond
8031 case 2: // result = lea base( , cond*2)
8032 case 3: // result = lea base(cond, cond*2)
8033 case 4: // result = lea base( , cond*4)
8034 case 5: // result = lea base(cond, cond*4)
8035 case 8: // result = lea base( , cond*8)
8036 case 9: // result = lea base(cond, cond*8)
8037 isFastMultiplier = true;
8042 if (isFastMultiplier) {
8043 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8044 SDValue Cond = N->getOperand(3);
8045 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8046 DAG.getConstant(CC, MVT::i8), Cond);
8047 // Zero extend the condition if needed.
8048 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8050 // Scale the condition by the difference.
8052 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8053 DAG.getConstant(Diff, Cond.getValueType()));
8055 // Add the base if non-zero.
8056 if (FalseC->getAPIntValue() != 0)
8057 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8058 SDValue(FalseC, 0));
8059 if (N->getNumValues() == 2) // Dead flag value?
8060 return DCI.CombineTo(N, Cond, SDValue());
8070 /// PerformMulCombine - Optimize a single multiply with constant into two
8071 /// in order to implement it with two cheaper instructions, e.g.
8072 /// LEA + SHL, LEA + LEA.
8073 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8074 TargetLowering::DAGCombinerInfo &DCI) {
8075 if (DAG.getMachineFunction().
8076 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8079 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8082 MVT VT = N->getValueType(0);
8086 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8089 uint64_t MulAmt = C->getZExtValue();
8090 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8093 uint64_t MulAmt1 = 0;
8094 uint64_t MulAmt2 = 0;
8095 if ((MulAmt % 9) == 0) {
8097 MulAmt2 = MulAmt / 9;
8098 } else if ((MulAmt % 5) == 0) {
8100 MulAmt2 = MulAmt / 5;
8101 } else if ((MulAmt % 3) == 0) {
8103 MulAmt2 = MulAmt / 3;
8106 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8107 DebugLoc DL = N->getDebugLoc();
8109 if (isPowerOf2_64(MulAmt2) &&
8110 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8111 // If second multiplifer is pow2, issue it first. We want the multiply by
8112 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8114 std::swap(MulAmt1, MulAmt2);
8117 if (isPowerOf2_64(MulAmt1))
8118 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8119 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8121 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8122 DAG.getConstant(MulAmt1, VT));
8124 if (isPowerOf2_64(MulAmt2))
8125 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8126 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8128 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8129 DAG.getConstant(MulAmt2, VT));
8131 // Do not add new nodes to DAG combiner worklist.
8132 DCI.CombineTo(N, NewMul, false);
8138 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8140 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8141 const X86Subtarget *Subtarget) {
8142 // On X86 with SSE2 support, we can transform this to a vector shift if
8143 // all elements are shifted by the same amount. We can't do this in legalize
8144 // because the a constant vector is typically transformed to a constant pool
8145 // so we have no knowledge of the shift amount.
8146 if (!Subtarget->hasSSE2())
8149 MVT VT = N->getValueType(0);
8150 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8153 SDValue ShAmtOp = N->getOperand(1);
8154 MVT EltVT = VT.getVectorElementType();
8155 DebugLoc DL = N->getDebugLoc();
8157 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8158 unsigned NumElts = VT.getVectorNumElements();
8160 for (; i != NumElts; ++i) {
8161 SDValue Arg = ShAmtOp.getOperand(i);
8162 if (Arg.getOpcode() == ISD::UNDEF) continue;
8166 for (; i != NumElts; ++i) {
8167 SDValue Arg = ShAmtOp.getOperand(i);
8168 if (Arg.getOpcode() == ISD::UNDEF) continue;
8169 if (Arg != BaseShAmt) {
8173 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8174 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8175 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8176 DAG.getIntPtrConstant(0));
8180 if (EltVT.bitsGT(MVT::i32))
8181 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8182 else if (EltVT.bitsLT(MVT::i32))
8183 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8185 // The shift amount is identical so we can do a vector shift.
8186 SDValue ValOp = N->getOperand(0);
8187 switch (N->getOpcode()) {
8189 assert(0 && "Unknown shift opcode!");
8192 if (VT == MVT::v2i64)
8193 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8194 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8196 if (VT == MVT::v4i32)
8197 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8198 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8200 if (VT == MVT::v8i16)
8201 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8202 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8206 if (VT == MVT::v4i32)
8207 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8208 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8210 if (VT == MVT::v8i16)
8211 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8212 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8216 if (VT == MVT::v2i64)
8217 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8218 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8220 if (VT == MVT::v4i32)
8221 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8222 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8224 if (VT == MVT::v8i16)
8225 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8226 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8233 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8234 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8235 const X86Subtarget *Subtarget) {
8236 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8237 // the FP state in cases where an emms may be missing.
8238 // A preferable solution to the general problem is to figure out the right
8239 // places to insert EMMS. This qualifies as a quick hack.
8241 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8242 StoreSDNode *St = cast<StoreSDNode>(N);
8243 MVT VT = St->getValue().getValueType();
8244 if (VT.getSizeInBits() != 64)
8247 const Function *F = DAG.getMachineFunction().getFunction();
8248 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8249 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8250 && Subtarget->hasSSE2();
8251 if ((VT.isVector() ||
8252 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8253 isa<LoadSDNode>(St->getValue()) &&
8254 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8255 St->getChain().hasOneUse() && !St->isVolatile()) {
8256 SDNode* LdVal = St->getValue().getNode();
8258 int TokenFactorIndex = -1;
8259 SmallVector<SDValue, 8> Ops;
8260 SDNode* ChainVal = St->getChain().getNode();
8261 // Must be a store of a load. We currently handle two cases: the load
8262 // is a direct child, and it's under an intervening TokenFactor. It is
8263 // possible to dig deeper under nested TokenFactors.
8264 if (ChainVal == LdVal)
8265 Ld = cast<LoadSDNode>(St->getChain());
8266 else if (St->getValue().hasOneUse() &&
8267 ChainVal->getOpcode() == ISD::TokenFactor) {
8268 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8269 if (ChainVal->getOperand(i).getNode() == LdVal) {
8270 TokenFactorIndex = i;
8271 Ld = cast<LoadSDNode>(St->getValue());
8273 Ops.push_back(ChainVal->getOperand(i));
8277 if (!Ld || !ISD::isNormalLoad(Ld))
8280 // If this is not the MMX case, i.e. we are just turning i64 load/store
8281 // into f64 load/store, avoid the transformation if there are multiple
8282 // uses of the loaded value.
8283 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8286 DebugLoc LdDL = Ld->getDebugLoc();
8287 DebugLoc StDL = N->getDebugLoc();
8288 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8289 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8291 if (Subtarget->is64Bit() || F64IsLegal) {
8292 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8293 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8294 Ld->getBasePtr(), Ld->getSrcValue(),
8295 Ld->getSrcValueOffset(), Ld->isVolatile(),
8296 Ld->getAlignment());
8297 SDValue NewChain = NewLd.getValue(1);
8298 if (TokenFactorIndex != -1) {
8299 Ops.push_back(NewChain);
8300 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8303 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8304 St->getSrcValue(), St->getSrcValueOffset(),
8305 St->isVolatile(), St->getAlignment());
8308 // Otherwise, lower to two pairs of 32-bit loads / stores.
8309 SDValue LoAddr = Ld->getBasePtr();
8310 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8311 DAG.getConstant(4, MVT::i32));
8313 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8314 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8315 Ld->isVolatile(), Ld->getAlignment());
8316 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8317 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8319 MinAlign(Ld->getAlignment(), 4));
8321 SDValue NewChain = LoLd.getValue(1);
8322 if (TokenFactorIndex != -1) {
8323 Ops.push_back(LoLd);
8324 Ops.push_back(HiLd);
8325 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8329 LoAddr = St->getBasePtr();
8330 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8331 DAG.getConstant(4, MVT::i32));
8333 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8334 St->getSrcValue(), St->getSrcValueOffset(),
8335 St->isVolatile(), St->getAlignment());
8336 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8338 St->getSrcValueOffset() + 4,
8340 MinAlign(St->getAlignment(), 4));
8341 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8346 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8347 /// X86ISD::FXOR nodes.
8348 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8349 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8350 // F[X]OR(0.0, x) -> x
8351 // F[X]OR(x, 0.0) -> x
8352 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8353 if (C->getValueAPF().isPosZero())
8354 return N->getOperand(1);
8355 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8356 if (C->getValueAPF().isPosZero())
8357 return N->getOperand(0);
8361 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8362 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8363 // FAND(0.0, x) -> 0.0
8364 // FAND(x, 0.0) -> 0.0
8365 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8366 if (C->getValueAPF().isPosZero())
8367 return N->getOperand(0);
8368 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8369 if (C->getValueAPF().isPosZero())
8370 return N->getOperand(1);
8374 static SDValue PerformBTCombine(SDNode *N,
8376 TargetLowering::DAGCombinerInfo &DCI) {
8377 // BT ignores high bits in the bit index operand.
8378 SDValue Op1 = N->getOperand(1);
8379 if (Op1.hasOneUse()) {
8380 unsigned BitWidth = Op1.getValueSizeInBits();
8381 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8382 APInt KnownZero, KnownOne;
8383 TargetLowering::TargetLoweringOpt TLO(DAG);
8384 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8385 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8386 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8387 DCI.CommitTargetLoweringOpt(TLO);
8392 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8393 SDValue Op = N->getOperand(0);
8394 if (Op.getOpcode() == ISD::BIT_CONVERT)
8395 Op = Op.getOperand(0);
8396 MVT VT = N->getValueType(0), OpVT = Op.getValueType();
8397 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8398 VT.getVectorElementType().getSizeInBits() ==
8399 OpVT.getVectorElementType().getSizeInBits()) {
8400 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8405 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8406 DAGCombinerInfo &DCI) const {
8407 SelectionDAG &DAG = DCI.DAG;
8408 switch (N->getOpcode()) {
8410 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8411 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8412 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8413 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8416 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8417 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8419 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8420 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8421 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8422 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
8428 //===----------------------------------------------------------------------===//
8429 // X86 Inline Assembly Support
8430 //===----------------------------------------------------------------------===//
8432 /// getConstraintType - Given a constraint letter, return the type of
8433 /// constraint it is for this target.
8434 X86TargetLowering::ConstraintType
8435 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8436 if (Constraint.size() == 1) {
8437 switch (Constraint[0]) {
8449 return C_RegisterClass;
8457 return TargetLowering::getConstraintType(Constraint);
8460 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8461 /// with another that has more specific requirements based on the type of the
8462 /// corresponding operand.
8463 const char *X86TargetLowering::
8464 LowerXConstraint(MVT ConstraintVT) const {
8465 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8466 // 'f' like normal targets.
8467 if (ConstraintVT.isFloatingPoint()) {
8468 if (Subtarget->hasSSE2())
8470 if (Subtarget->hasSSE1())
8474 return TargetLowering::LowerXConstraint(ConstraintVT);
8477 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8478 /// vector. If it is invalid, don't add anything to Ops.
8479 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8482 std::vector<SDValue>&Ops,
8483 SelectionDAG &DAG) const {
8484 SDValue Result(0, 0);
8486 switch (Constraint) {
8489 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8490 if (C->getZExtValue() <= 31) {
8491 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8497 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8498 if (C->getZExtValue() <= 63) {
8499 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8505 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8506 if (C->getZExtValue() <= 255) {
8507 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8513 // 32-bit signed value
8514 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8515 const ConstantInt *CI = C->getConstantIntValue();
8516 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8517 // Widen to 64 bits here to get it sign extended.
8518 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8521 // FIXME gcc accepts some relocatable values here too, but only in certain
8522 // memory models; it's complicated.
8527 // 32-bit unsigned value
8528 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8529 const ConstantInt *CI = C->getConstantIntValue();
8530 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8531 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8535 // FIXME gcc accepts some relocatable values here too, but only in certain
8536 // memory models; it's complicated.
8540 // Literal immediates are always ok.
8541 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8542 // Widen to 64 bits here to get it sign extended.
8543 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8547 // If we are in non-pic codegen mode, we allow the address of a global (with
8548 // an optional displacement) to be used with 'i'.
8549 GlobalAddressSDNode *GA = 0;
8552 // Match either (GA), (GA+C), (GA+C1+C2), etc.
8554 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
8555 Offset += GA->getOffset();
8557 } else if (Op.getOpcode() == ISD::ADD) {
8558 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8559 Offset += C->getZExtValue();
8560 Op = Op.getOperand(0);
8563 } else if (Op.getOpcode() == ISD::SUB) {
8564 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8565 Offset += -C->getZExtValue();
8566 Op = Op.getOperand(0);
8571 // Otherwise, this isn't something we can handle, reject it.
8576 Op = LowerGlobalAddress(GA->getGlobal(), Op.getDebugLoc(), Offset, DAG);
8578 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
8585 if (Result.getNode()) {
8586 Ops.push_back(Result);
8589 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8593 std::vector<unsigned> X86TargetLowering::
8594 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8596 if (Constraint.size() == 1) {
8597 // FIXME: not handling fp-stack yet!
8598 switch (Constraint[0]) { // GCC X86 Constraint Letters
8599 default: break; // Unknown constraint letter
8600 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
8603 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
8604 else if (VT == MVT::i16)
8605 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
8606 else if (VT == MVT::i8)
8607 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
8608 else if (VT == MVT::i64)
8609 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
8614 return std::vector<unsigned>();
8617 std::pair<unsigned, const TargetRegisterClass*>
8618 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8620 // First, see if this is a constraint that directly corresponds to an LLVM
8622 if (Constraint.size() == 1) {
8623 // GCC Constraint Letters
8624 switch (Constraint[0]) {
8626 case 'r': // GENERAL_REGS
8627 case 'R': // LEGACY_REGS
8628 case 'l': // INDEX_REGS
8630 return std::make_pair(0U, X86::GR8RegisterClass);
8632 return std::make_pair(0U, X86::GR16RegisterClass);
8633 if (VT == MVT::i32 || !Subtarget->is64Bit())
8634 return std::make_pair(0U, X86::GR32RegisterClass);
8635 return std::make_pair(0U, X86::GR64RegisterClass);
8636 case 'f': // FP Stack registers.
8637 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
8638 // value to the correct fpstack register class.
8639 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
8640 return std::make_pair(0U, X86::RFP32RegisterClass);
8641 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
8642 return std::make_pair(0U, X86::RFP64RegisterClass);
8643 return std::make_pair(0U, X86::RFP80RegisterClass);
8644 case 'y': // MMX_REGS if MMX allowed.
8645 if (!Subtarget->hasMMX()) break;
8646 return std::make_pair(0U, X86::VR64RegisterClass);
8647 case 'Y': // SSE_REGS if SSE2 allowed
8648 if (!Subtarget->hasSSE2()) break;
8650 case 'x': // SSE_REGS if SSE1 allowed
8651 if (!Subtarget->hasSSE1()) break;
8653 switch (VT.getSimpleVT()) {
8655 // Scalar SSE types.
8658 return std::make_pair(0U, X86::FR32RegisterClass);
8661 return std::make_pair(0U, X86::FR64RegisterClass);
8669 return std::make_pair(0U, X86::VR128RegisterClass);
8675 // Use the default implementation in TargetLowering to convert the register
8676 // constraint into a member of a register class.
8677 std::pair<unsigned, const TargetRegisterClass*> Res;
8678 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8680 // Not found as a standard register?
8681 if (Res.second == 0) {
8682 // GCC calls "st(0)" just plain "st".
8683 if (StringsEqualNoCase("{st}", Constraint)) {
8684 Res.first = X86::ST0;
8685 Res.second = X86::RFP80RegisterClass;
8687 // 'A' means EAX + EDX.
8688 if (Constraint == "A") {
8689 Res.first = X86::EAX;
8690 Res.second = X86::GRADRegisterClass;
8695 // Otherwise, check to see if this is a register class of the wrong value
8696 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
8697 // turn into {ax},{dx}.
8698 if (Res.second->hasType(VT))
8699 return Res; // Correct type already, nothing to do.
8701 // All of the single-register GCC register classes map their values onto
8702 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
8703 // really want an 8-bit or 32-bit register, map to the appropriate register
8704 // class and return the appropriate register.
8705 if (Res.second == X86::GR16RegisterClass) {
8706 if (VT == MVT::i8) {
8707 unsigned DestReg = 0;
8708 switch (Res.first) {
8710 case X86::AX: DestReg = X86::AL; break;
8711 case X86::DX: DestReg = X86::DL; break;
8712 case X86::CX: DestReg = X86::CL; break;
8713 case X86::BX: DestReg = X86::BL; break;
8716 Res.first = DestReg;
8717 Res.second = X86::GR8RegisterClass;
8719 } else if (VT == MVT::i32) {
8720 unsigned DestReg = 0;
8721 switch (Res.first) {
8723 case X86::AX: DestReg = X86::EAX; break;
8724 case X86::DX: DestReg = X86::EDX; break;
8725 case X86::CX: DestReg = X86::ECX; break;
8726 case X86::BX: DestReg = X86::EBX; break;
8727 case X86::SI: DestReg = X86::ESI; break;
8728 case X86::DI: DestReg = X86::EDI; break;
8729 case X86::BP: DestReg = X86::EBP; break;
8730 case X86::SP: DestReg = X86::ESP; break;
8733 Res.first = DestReg;
8734 Res.second = X86::GR32RegisterClass;
8736 } else if (VT == MVT::i64) {
8737 unsigned DestReg = 0;
8738 switch (Res.first) {
8740 case X86::AX: DestReg = X86::RAX; break;
8741 case X86::DX: DestReg = X86::RDX; break;
8742 case X86::CX: DestReg = X86::RCX; break;
8743 case X86::BX: DestReg = X86::RBX; break;
8744 case X86::SI: DestReg = X86::RSI; break;
8745 case X86::DI: DestReg = X86::RDI; break;
8746 case X86::BP: DestReg = X86::RBP; break;
8747 case X86::SP: DestReg = X86::RSP; break;
8750 Res.first = DestReg;
8751 Res.second = X86::GR64RegisterClass;
8754 } else if (Res.second == X86::FR32RegisterClass ||
8755 Res.second == X86::FR64RegisterClass ||
8756 Res.second == X86::VR128RegisterClass) {
8757 // Handle references to XMM physical registers that got mapped into the
8758 // wrong class. This can happen with constraints like {xmm0} where the
8759 // target independent register mapper will just pick the first match it can
8760 // find, ignoring the required type.
8762 Res.second = X86::FR32RegisterClass;
8763 else if (VT == MVT::f64)
8764 Res.second = X86::FR64RegisterClass;
8765 else if (X86::VR128RegisterClass->hasType(VT))
8766 Res.second = X86::VR128RegisterClass;
8772 //===----------------------------------------------------------------------===//
8773 // X86 Widen vector type
8774 //===----------------------------------------------------------------------===//
8776 /// getWidenVectorType: given a vector type, returns the type to widen
8777 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
8778 /// If there is no vector type that we want to widen to, returns MVT::Other
8779 /// When and where to widen is target dependent based on the cost of
8780 /// scalarizing vs using the wider vector type.
8782 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
8783 assert(VT.isVector());
8784 if (isTypeLegal(VT))
8787 // TODO: In computeRegisterProperty, we can compute the list of legal vector
8788 // type based on element type. This would speed up our search (though
8789 // it may not be worth it since the size of the list is relatively
8791 MVT EltVT = VT.getVectorElementType();
8792 unsigned NElts = VT.getVectorNumElements();
8794 // On X86, it make sense to widen any vector wider than 1
8798 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
8799 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
8800 MVT SVT = (MVT::SimpleValueType)nVT;
8802 if (isTypeLegal(SVT) &&
8803 SVT.getVectorElementType() == EltVT &&
8804 SVT.getVectorNumElements() > NElts)