1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source 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 "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/ADT/VectorExtras.h"
26 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/CodeGen/SSARegMap.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/ADT/StringExtras.h"
39 static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
40 cl::desc("Enable fastcc on X86"));
41 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
42 : TargetLowering(TM) {
43 Subtarget = &TM.getSubtarget<X86Subtarget>();
44 X86ScalarSSE = Subtarget->hasSSE2();
45 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
47 // Set up the TargetLowering object.
49 // X86 is weird, it always uses i8 for shift amounts and setcc results.
50 setShiftAmountType(MVT::i8);
51 setSetCCResultType(MVT::i8);
52 setSetCCResultContents(ZeroOrOneSetCCResult);
53 setSchedulingPreference(SchedulingForRegPressure);
54 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
55 setStackPointerRegisterToSaveRestore(X86StackPtr);
57 if (Subtarget->isTargetDarwin()) {
58 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
59 setUseUnderscoreSetJmp(false);
60 setUseUnderscoreLongJmp(false);
61 } else if (Subtarget->isTargetMingw()) {
62 // MS runtime is weird: it exports _setjmp, but longjmp!
63 setUseUnderscoreSetJmp(true);
64 setUseUnderscoreLongJmp(false);
66 setUseUnderscoreSetJmp(true);
67 setUseUnderscoreLongJmp(true);
70 // Add legal addressing mode scale values.
71 addLegalAddressScale(8);
72 addLegalAddressScale(4);
73 addLegalAddressScale(2);
74 // Enter the ones which require both scale + index last. These are more
76 addLegalAddressScale(9);
77 addLegalAddressScale(5);
78 addLegalAddressScale(3);
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 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
89 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
91 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
92 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
93 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
95 if (Subtarget->is64Bit()) {
96 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
97 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
100 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
101 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
103 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
106 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
108 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
109 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
110 // SSE has no i16 to fp conversion, only i32
112 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
114 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
115 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
118 if (!Subtarget->is64Bit()) {
119 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
120 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
121 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
124 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
126 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
127 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
130 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
132 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
133 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
136 // Handle FP_TO_UINT by promoting the destination to a larger signed
138 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
139 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
140 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
142 if (Subtarget->is64Bit()) {
143 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
144 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
146 if (X86ScalarSSE && !Subtarget->hasSSE3())
147 // Expand FP_TO_UINT into a select.
148 // FIXME: We would like to use a Custom expander here eventually to do
149 // the optimal thing for SSE vs. the default expansion in the legalizer.
150 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
152 // With SSE3 we can use fisttpll to convert to a signed i64.
153 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
156 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
158 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
159 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
162 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
163 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
164 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
165 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
166 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
167 if (Subtarget->is64Bit())
168 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
169 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
170 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
171 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
172 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
173 setOperationAction(ISD::FREM , MVT::f64 , Expand);
175 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
176 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
177 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
178 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
179 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
180 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
181 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
182 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
183 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
184 if (Subtarget->is64Bit()) {
185 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
186 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
187 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
190 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
191 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
193 // These should be promoted to a larger select which is supported.
194 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
195 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
196 // X86 wants to expand cmov itself.
197 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
198 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
199 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
200 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
201 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
202 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
203 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
204 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
205 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
206 if (Subtarget->is64Bit()) {
207 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
208 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
210 // X86 ret instruction may pop stack.
211 setOperationAction(ISD::RET , MVT::Other, Custom);
213 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
214 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
215 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
216 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
217 if (Subtarget->is64Bit()) {
218 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
219 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
220 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
221 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
223 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
224 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
225 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
226 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
227 // X86 wants to expand memset / memcpy itself.
228 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
229 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
231 // We don't have line number support yet.
232 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
233 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
234 // FIXME - use subtarget debug flags
235 if (!Subtarget->isTargetDarwin() &&
236 !Subtarget->isTargetELF() &&
237 !Subtarget->isTargetCygMing())
238 setOperationAction(ISD::LABEL, MVT::Other, Expand);
240 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
241 setOperationAction(ISD::VASTART , MVT::Other, Custom);
243 // Use the default implementation.
244 setOperationAction(ISD::VAARG , MVT::Other, Expand);
245 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
246 setOperationAction(ISD::VAEND , MVT::Other, Expand);
247 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
248 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
249 if (Subtarget->is64Bit())
250 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
251 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
254 // Set up the FP register classes.
255 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
256 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
258 // Use ANDPD to simulate FABS.
259 setOperationAction(ISD::FABS , MVT::f64, Custom);
260 setOperationAction(ISD::FABS , MVT::f32, Custom);
262 // Use XORP to simulate FNEG.
263 setOperationAction(ISD::FNEG , MVT::f64, Custom);
264 setOperationAction(ISD::FNEG , MVT::f32, Custom);
266 // Use ANDPD and ORPD to simulate FCOPYSIGN.
267 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
268 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
270 // We don't support sin/cos/fmod
271 setOperationAction(ISD::FSIN , MVT::f64, Expand);
272 setOperationAction(ISD::FCOS , MVT::f64, Expand);
273 setOperationAction(ISD::FREM , MVT::f64, Expand);
274 setOperationAction(ISD::FSIN , MVT::f32, Expand);
275 setOperationAction(ISD::FCOS , MVT::f32, Expand);
276 setOperationAction(ISD::FREM , MVT::f32, Expand);
278 // Expand FP immediates into loads from the stack, except for the special
280 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
281 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
282 addLegalFPImmediate(+0.0); // xorps / xorpd
284 // Set up the FP register classes.
285 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
287 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
288 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
289 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
292 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
293 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
296 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
297 addLegalFPImmediate(+0.0); // FLD0
298 addLegalFPImmediate(+1.0); // FLD1
299 addLegalFPImmediate(-0.0); // FLD0/FCHS
300 addLegalFPImmediate(-1.0); // FLD1/FCHS
303 // First set operation action for all vector types to expand. Then we
304 // will selectively turn on ones that can be effectively codegen'd.
305 for (unsigned VT = (unsigned)MVT::Vector + 1;
306 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
307 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
308 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
309 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
310 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
311 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
312 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
313 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
314 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
315 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
316 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
317 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
318 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
319 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
320 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
321 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
324 if (Subtarget->hasMMX()) {
325 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
326 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
327 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
329 // FIXME: add MMX packed arithmetics
330 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Expand);
331 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Expand);
332 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Expand);
335 if (Subtarget->hasSSE1()) {
336 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
338 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
339 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
340 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
341 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
342 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
343 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
344 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
345 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
346 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
349 if (Subtarget->hasSSE2()) {
350 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
351 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
352 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
353 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
354 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
356 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
357 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
358 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
359 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
360 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
361 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
362 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
363 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
364 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
365 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
366 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
368 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
369 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
370 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
371 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
372 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
373 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
375 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
376 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
377 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
378 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
379 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
381 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
382 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
383 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
384 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
385 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
386 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
388 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
389 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
390 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
391 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
392 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
393 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
394 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
395 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
396 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
397 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
398 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
399 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
402 // Custom lower v2i64 and v2f64 selects.
403 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
404 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
405 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
406 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
409 // We want to custom lower some of our intrinsics.
410 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
412 // We have target-specific dag combine patterns for the following nodes:
413 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
414 setTargetDAGCombine(ISD::SELECT);
416 computeRegisterProperties();
418 // FIXME: These should be based on subtarget info. Plus, the values should
419 // be smaller when we are in optimizing for size mode.
420 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
421 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
422 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
423 allowUnalignedMemoryAccesses = true; // x86 supports it!
427 //===----------------------------------------------------------------------===//
428 // Return Value Calling Convention Implementation
429 //===----------------------------------------------------------------------===//
431 /// GetRetValueLocs - If we are returning a set of values with the specified
432 /// value types, determine the set of registers each one will land in. This
433 /// sets one element of the ResultRegs array for each element in the VTs array.
434 static void GetRetValueLocs(const MVT::ValueType *VTs, unsigned NumVTs,
435 unsigned *ResultRegs,
436 const X86Subtarget *Subtarget,
437 unsigned CallingConv) {
438 if (NumVTs == 0) return;
441 ResultRegs[0] = VTs[0] == MVT::i64 ? X86::RAX : X86::EAX;
442 ResultRegs[1] = VTs[1] == MVT::i64 ? X86::RDX : X86::EDX;
446 // Otherwise, NumVTs is 1.
447 MVT::ValueType ArgVT = VTs[0];
449 if (MVT::isVector(ArgVT)) // Integer or FP vector result -> XMM0.
450 ResultRegs[0] = X86::XMM0;
451 else if (MVT::isFloatingPoint(ArgVT) && Subtarget->is64Bit())
452 // FP values in X86-64 go in XMM0.
453 ResultRegs[0] = X86::XMM0;
454 else if (MVT::isFloatingPoint(ArgVT))
455 // FP values in X86-32 go in ST0.
456 ResultRegs[0] = X86::ST0;
458 assert(MVT::isInteger(ArgVT) && "Unknown return value type!");
460 // Integer result -> EAX / RAX.
461 // The C calling convention guarantees the return value has been
462 // promoted to at least MVT::i32. The X86-64 ABI doesn't require the
463 // value to be promoted MVT::i64. So we don't have to extend it to
465 ResultRegs[0] = (ArgVT == MVT::i64) ? X86::RAX : X86::EAX;
470 //===----------------------------------------------------------------------===//
471 // C & StdCall Calling Convention implementation
472 //===----------------------------------------------------------------------===//
473 // StdCall calling convention seems to be standard for many Windows' API
474 // routines and around. It differs from C calling convention just a little:
475 // callee should clean up the stack, not caller. Symbols should be also
476 // decorated in some fancy way :) It doesn't support any vector arguments.
478 /// AddLiveIn - This helper function adds the specified physical register to the
479 /// MachineFunction as a live in value. It also creates a corresponding virtual
481 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
482 const TargetRegisterClass *RC) {
483 assert(RC->contains(PReg) && "Not the correct regclass!");
484 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
485 MF.addLiveIn(PReg, VReg);
489 /// HowToPassArgument - Returns how an formal argument of the specified type
490 /// should be passed. If it is through stack, returns the size of the stack
491 /// slot; if it is through integer or XMM register, returns the number of
492 /// integer or XMM registers are needed.
494 HowToPassCallArgument(MVT::ValueType ObjectVT,
496 unsigned NumIntRegs, unsigned NumXMMRegs,
497 unsigned MaxNumIntRegs,
498 unsigned &ObjSize, unsigned &ObjIntRegs,
499 unsigned &ObjXMMRegs,
500 bool AllowVectors = true) {
505 if (MaxNumIntRegs>3) {
506 // We don't have too much registers on ia32! :)
511 default: assert(0 && "Unhandled argument type!");
513 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
519 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
525 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
531 if (ArgInReg && (NumIntRegs+2 <= MaxNumIntRegs)) {
533 } else if (ArgInReg && (NumIntRegs+1 <= MaxNumIntRegs)) {
557 assert(0 && "Unhandled argument type [vector]!");
561 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
563 unsigned NumArgs = Op.Val->getNumValues() - 1;
564 MachineFunction &MF = DAG.getMachineFunction();
565 MachineFrameInfo *MFI = MF.getFrameInfo();
566 SDOperand Root = Op.getOperand(0);
567 SmallVector<SDOperand, 8> ArgValues;
568 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
570 // Add DAG nodes to load the arguments... On entry to a function on the X86,
571 // the stack frame looks like this:
573 // [ESP] -- return address
574 // [ESP + 4] -- first argument (leftmost lexically)
575 // [ESP + 8] -- second argument, if first argument is <= 4 bytes in size
578 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
579 unsigned NumSRetBytes= 0; // How much bytes on stack used for struct return
580 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
581 unsigned NumIntRegs = 0; // Integer regs used for parameter passing
583 static const unsigned XMMArgRegs[] = {
584 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
586 static const unsigned GPRArgRegs[][3] = {
587 { X86::AL, X86::DL, X86::CL },
588 { X86::AX, X86::DX, X86::CX },
589 { X86::EAX, X86::EDX, X86::ECX }
591 static const TargetRegisterClass* GPRClasses[3] = {
592 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
595 // Handle regparm attribute
596 SmallVector<bool, 8> ArgInRegs(NumArgs, false);
597 SmallVector<bool, 8> SRetArgs(NumArgs, false);
599 for (unsigned i = 0; i<NumArgs; ++i) {
600 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3+i))->getValue();
601 ArgInRegs[i] = (Flags >> 1) & 1;
602 SRetArgs[i] = (Flags >> 2) & 1;
606 for (unsigned i = 0; i < NumArgs; ++i) {
607 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
608 unsigned ArgIncrement = 4;
609 unsigned ObjSize = 0;
610 unsigned ObjXMMRegs = 0;
611 unsigned ObjIntRegs = 0;
615 HowToPassCallArgument(ObjectVT,
617 NumIntRegs, NumXMMRegs, 3,
618 ObjSize, ObjIntRegs, ObjXMMRegs,
622 ArgIncrement = ObjSize;
624 if (ObjIntRegs || ObjXMMRegs) {
626 default: assert(0 && "Unhandled argument type!");
630 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][NumIntRegs];
631 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
632 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
641 assert(!isStdCall && "Unhandled argument type!");
642 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
643 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
646 NumIntRegs += ObjIntRegs;
647 NumXMMRegs += ObjXMMRegs;
650 // XMM arguments have to be aligned on 16-byte boundary.
652 ArgOffset = ((ArgOffset + 15) / 16) * 16;
653 // Create the SelectionDAG nodes corresponding to a load from this
655 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
656 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
657 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
659 ArgOffset += ArgIncrement; // Move on to the next argument.
661 NumSRetBytes += ArgIncrement;
664 ArgValues.push_back(ArgValue);
667 ArgValues.push_back(Root);
669 // If the function takes variable number of arguments, make a frame index for
670 // the start of the first vararg value... for expansion of llvm.va_start.
672 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
674 if (isStdCall && !isVarArg) {
675 BytesToPopOnReturn = ArgOffset; // Callee pops everything..
676 BytesCallerReserves = 0;
678 BytesToPopOnReturn = NumSRetBytes; // Callee pops hidden struct pointer.
679 BytesCallerReserves = ArgOffset;
682 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
683 ReturnAddrIndex = 0; // No return address slot generated yet.
686 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
688 // Return the new list of results.
689 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
690 &ArgValues[0], ArgValues.size());
693 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
695 SDOperand Chain = Op.getOperand(0);
696 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
697 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
698 SDOperand Callee = Op.getOperand(4);
699 MVT::ValueType RetVT= Op.Val->getValueType(0);
700 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
702 static const unsigned XMMArgRegs[] = {
703 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
705 static const unsigned GPR32ArgRegs[] = {
706 X86::EAX, X86::EDX, X86::ECX
709 // Count how many bytes are to be pushed on the stack.
710 unsigned NumBytes = 0;
711 // Keep track of the number of integer regs passed so far.
712 unsigned NumIntRegs = 0;
713 // Keep track of the number of XMM regs passed so far.
714 unsigned NumXMMRegs = 0;
715 // How much bytes on stack used for struct return
716 unsigned NumSRetBytes= 0;
718 // Handle regparm attribute
719 SmallVector<bool, 8> ArgInRegs(NumOps, false);
720 SmallVector<bool, 8> SRetArgs(NumOps, false);
721 for (unsigned i = 0; i<NumOps; ++i) {
723 dyn_cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
724 ArgInRegs[i] = (Flags >> 1) & 1;
725 SRetArgs[i] = (Flags >> 2) & 1;
728 // Calculate stack frame size
729 for (unsigned i = 0; i != NumOps; ++i) {
730 SDOperand Arg = Op.getOperand(5+2*i);
731 unsigned ArgIncrement = 4;
732 unsigned ObjSize = 0;
733 unsigned ObjIntRegs = 0;
734 unsigned ObjXMMRegs = 0;
736 HowToPassCallArgument(Arg.getValueType(),
738 NumIntRegs, NumXMMRegs, 3,
739 ObjSize, ObjIntRegs, ObjXMMRegs,
742 ArgIncrement = ObjSize;
744 NumIntRegs += ObjIntRegs;
745 NumXMMRegs += ObjXMMRegs;
747 // XMM arguments have to be aligned on 16-byte boundary.
749 NumBytes = ((NumBytes + 15) / 16) * 16;
750 NumBytes += ArgIncrement;
754 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
756 // Arguments go on the stack in reverse order, as specified by the ABI.
757 unsigned ArgOffset = 0;
760 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
761 SmallVector<SDOperand, 8> MemOpChains;
762 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
763 for (unsigned i = 0; i != NumOps; ++i) {
764 SDOperand Arg = Op.getOperand(5+2*i);
765 unsigned ArgIncrement = 4;
766 unsigned ObjSize = 0;
767 unsigned ObjIntRegs = 0;
768 unsigned ObjXMMRegs = 0;
770 HowToPassCallArgument(Arg.getValueType(),
772 NumIntRegs, NumXMMRegs, 3,
773 ObjSize, ObjIntRegs, ObjXMMRegs,
777 ArgIncrement = ObjSize;
779 if (Arg.getValueType() == MVT::i8 || Arg.getValueType() == MVT::i16) {
780 // Promote the integer to 32 bits. If the input type is signed use a
781 // sign extend, otherwise use a zero extend.
782 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
784 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
785 Arg = DAG.getNode(ExtOp, MVT::i32, Arg);
788 if (ObjIntRegs || ObjXMMRegs) {
789 switch (Arg.getValueType()) {
790 default: assert(0 && "Unhandled argument type!");
792 RegsToPass.push_back(std::make_pair(GPR32ArgRegs[NumIntRegs], Arg));
800 assert(!isStdCall && "Unhandled argument type!");
801 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
805 NumIntRegs += ObjIntRegs;
806 NumXMMRegs += ObjXMMRegs;
809 // XMM arguments have to be aligned on 16-byte boundary.
811 ArgOffset = ((ArgOffset + 15) / 16) * 16;
813 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
814 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
815 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
817 ArgOffset += ArgIncrement; // Move on to the next argument.
819 NumSRetBytes += ArgIncrement;
823 // Sanity check: we haven't seen NumSRetBytes > 4
824 assert((NumSRetBytes<=4) &&
825 "Too much space for struct-return pointer requested");
827 if (!MemOpChains.empty())
828 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
829 &MemOpChains[0], MemOpChains.size());
831 // Build a sequence of copy-to-reg nodes chained together with token chain
832 // and flag operands which copy the outgoing args into registers.
834 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
835 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
837 InFlag = Chain.getValue(1);
840 // ELF / PIC requires GOT in the EBX register before function calls via PLT
842 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
843 Subtarget->isPICStyleGOT()) {
844 Chain = DAG.getCopyToReg(Chain, X86::EBX,
845 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
847 InFlag = Chain.getValue(1);
850 // If the callee is a GlobalAddress node (quite common, every direct call is)
851 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
852 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
853 // We should use extra load for direct calls to dllimported functions in
855 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
856 getTargetMachine(), true))
857 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
858 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
859 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
861 // Returns a chain & a flag for retval copy to use.
862 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
863 SmallVector<SDOperand, 8> Ops;
864 Ops.push_back(Chain);
865 Ops.push_back(Callee);
867 // Add argument registers to the end of the list so that they are known live
869 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
870 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
871 RegsToPass[i].second.getValueType()));
873 // Add an implicit use GOT pointer in EBX.
874 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
875 Subtarget->isPICStyleGOT())
876 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
879 Ops.push_back(InFlag);
881 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
882 NodeTys, &Ops[0], Ops.size());
883 InFlag = Chain.getValue(1);
885 // Create the CALLSEQ_END node.
886 unsigned NumBytesForCalleeToPush = 0;
890 NumBytesForCalleeToPush = NumSRetBytes;
892 NumBytesForCalleeToPush = NumBytes;
895 // If this is is a call to a struct-return function, the callee
896 // pops the hidden struct pointer, so we have to push it back.
897 // This is common for Darwin/X86, Linux & Mingw32 targets.
898 NumBytesForCalleeToPush = NumSRetBytes;
901 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
903 Ops.push_back(Chain);
904 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
905 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
906 Ops.push_back(InFlag);
907 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
908 if (RetVT != MVT::Other)
909 InFlag = Chain.getValue(1);
911 SmallVector<SDOperand, 8> ResultVals;
913 default: assert(0 && "Unknown value type to return!");
915 NodeTys = DAG.getVTList(MVT::Other);
918 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
919 ResultVals.push_back(Chain.getValue(0));
920 NodeTys = DAG.getVTList(MVT::i8, MVT::Other);
923 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
924 ResultVals.push_back(Chain.getValue(0));
925 NodeTys = DAG.getVTList(MVT::i16, MVT::Other);
928 if (Op.Val->getValueType(1) == MVT::i32) {
929 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
930 ResultVals.push_back(Chain.getValue(0));
931 Chain = DAG.getCopyFromReg(Chain, X86::EDX, MVT::i32,
932 Chain.getValue(2)).getValue(1);
933 ResultVals.push_back(Chain.getValue(0));
934 NodeTys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
936 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
937 ResultVals.push_back(Chain.getValue(0));
938 NodeTys = DAG.getVTList(MVT::i32, MVT::Other);
947 assert(!isStdCall && "Unknown value type to return!");
948 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
949 ResultVals.push_back(Chain.getValue(0));
950 NodeTys = DAG.getVTList(RetVT, MVT::Other);
954 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
955 SDOperand GROps[] = { Chain, InFlag };
956 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
957 Chain = RetVal.getValue(1);
958 InFlag = RetVal.getValue(2);
960 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
961 // shouldn't be necessary except that RFP cannot be live across
962 // multiple blocks. When stackifier is fixed, they can be uncoupled.
963 MachineFunction &MF = DAG.getMachineFunction();
964 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
965 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
966 Tys = DAG.getVTList(MVT::Other);
968 Chain, RetVal, StackSlot, DAG.getValueType(RetVT), InFlag
970 Chain = DAG.getNode(X86ISD::FST, Tys, Ops, 5);
971 RetVal = DAG.getLoad(RetVT, Chain, StackSlot, NULL, 0);
972 Chain = RetVal.getValue(1);
975 if (RetVT == MVT::f32 && !X86ScalarSSE)
976 // FIXME: we would really like to remember that this FP_ROUND
977 // operation is okay to eliminate if we allow excess FP precision.
978 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
979 ResultVals.push_back(RetVal);
980 NodeTys = DAG.getVTList(RetVT, MVT::Other);
985 // Merge everything together with a MERGE_VALUES node.
986 ResultVals.push_back(Chain);
987 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
988 &ResultVals[0], ResultVals.size());
989 return Res.getValue(Op.ResNo);
993 //===----------------------------------------------------------------------===//
994 // X86-64 C Calling Convention implementation
995 //===----------------------------------------------------------------------===//
997 /// HowToPassX86_64CCCArgument - Returns how an formal argument of the specified
998 /// type should be passed. If it is through stack, returns the size of the stack
999 /// slot; if it is through integer or XMM register, returns the number of
1000 /// integer or XMM registers are needed.
1002 HowToPassX86_64CCCArgument(MVT::ValueType ObjectVT,
1003 unsigned NumIntRegs, unsigned NumXMMRegs,
1004 unsigned &ObjSize, unsigned &ObjIntRegs,
1005 unsigned &ObjXMMRegs) {
1011 default: assert(0 && "Unhandled argument type!");
1021 case MVT::i8: ObjSize = 1; break;
1022 case MVT::i16: ObjSize = 2; break;
1023 case MVT::i32: ObjSize = 4; break;
1024 case MVT::i64: ObjSize = 8; break;
1041 case MVT::f32: ObjSize = 4; break;
1042 case MVT::f64: ObjSize = 8; break;
1048 case MVT::v2f64: ObjSize = 16; break;
1056 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1057 unsigned NumArgs = Op.Val->getNumValues() - 1;
1058 MachineFunction &MF = DAG.getMachineFunction();
1059 MachineFrameInfo *MFI = MF.getFrameInfo();
1060 SDOperand Root = Op.getOperand(0);
1061 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1062 SmallVector<SDOperand, 8> ArgValues;
1064 // Add DAG nodes to load the arguments... On entry to a function on the X86,
1065 // the stack frame looks like this:
1067 // [RSP] -- return address
1068 // [RSP + 8] -- first nonreg argument (leftmost lexically)
1069 // [RSP +16] -- second nonreg argument, if 1st argument is <= 8 bytes in size
1072 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1073 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1074 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1076 static const unsigned GPR8ArgRegs[] = {
1077 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1079 static const unsigned GPR16ArgRegs[] = {
1080 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1082 static const unsigned GPR32ArgRegs[] = {
1083 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1085 static const unsigned GPR64ArgRegs[] = {
1086 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1088 static const unsigned XMMArgRegs[] = {
1089 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1090 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1093 for (unsigned i = 0; i < NumArgs; ++i) {
1094 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1095 unsigned ArgIncrement = 8;
1096 unsigned ObjSize = 0;
1097 unsigned ObjIntRegs = 0;
1098 unsigned ObjXMMRegs = 0;
1100 // FIXME: __int128 and long double support?
1101 HowToPassX86_64CCCArgument(ObjectVT, NumIntRegs, NumXMMRegs,
1102 ObjSize, ObjIntRegs, ObjXMMRegs);
1104 ArgIncrement = ObjSize;
1108 if (ObjIntRegs || ObjXMMRegs) {
1110 default: assert(0 && "Unhandled argument type!");
1115 TargetRegisterClass *RC = NULL;
1119 RC = X86::GR8RegisterClass;
1120 Reg = GPR8ArgRegs[NumIntRegs];
1123 RC = X86::GR16RegisterClass;
1124 Reg = GPR16ArgRegs[NumIntRegs];
1127 RC = X86::GR32RegisterClass;
1128 Reg = GPR32ArgRegs[NumIntRegs];
1131 RC = X86::GR64RegisterClass;
1132 Reg = GPR64ArgRegs[NumIntRegs];
1135 Reg = AddLiveIn(MF, Reg, RC);
1136 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1147 TargetRegisterClass *RC= (ObjectVT == MVT::f32) ?
1148 X86::FR32RegisterClass : ((ObjectVT == MVT::f64) ?
1149 X86::FR64RegisterClass : X86::VR128RegisterClass);
1150 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], RC);
1151 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1155 NumIntRegs += ObjIntRegs;
1156 NumXMMRegs += ObjXMMRegs;
1157 } else if (ObjSize) {
1158 // XMM arguments have to be aligned on 16-byte boundary.
1160 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1161 // Create the SelectionDAG nodes corresponding to a load from this
1163 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1164 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1165 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1166 ArgOffset += ArgIncrement; // Move on to the next argument.
1169 ArgValues.push_back(ArgValue);
1172 // If the function takes variable number of arguments, make a frame index for
1173 // the start of the first vararg value... for expansion of llvm.va_start.
1175 // For X86-64, if there are vararg parameters that are passed via
1176 // registers, then we must store them to their spots on the stack so they
1177 // may be loaded by deferencing the result of va_next.
1178 VarArgsGPOffset = NumIntRegs * 8;
1179 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1180 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
1181 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1183 // Store the integer parameter registers.
1184 SmallVector<SDOperand, 8> MemOps;
1185 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1186 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1187 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1188 for (; NumIntRegs != 6; ++NumIntRegs) {
1189 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1190 X86::GR64RegisterClass);
1191 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1192 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1193 MemOps.push_back(Store);
1194 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1195 DAG.getConstant(8, getPointerTy()));
1198 // Now store the XMM (fp + vector) parameter registers.
1199 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1200 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1201 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1202 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1203 X86::VR128RegisterClass);
1204 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1205 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1206 MemOps.push_back(Store);
1207 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1208 DAG.getConstant(16, getPointerTy()));
1210 if (!MemOps.empty())
1211 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1212 &MemOps[0], MemOps.size());
1215 ArgValues.push_back(Root);
1217 ReturnAddrIndex = 0; // No return address slot generated yet.
1218 BytesToPopOnReturn = 0; // Callee pops nothing.
1219 BytesCallerReserves = ArgOffset;
1221 // Return the new list of results.
1222 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1223 &ArgValues[0], ArgValues.size());
1227 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG) {
1228 SDOperand Chain = Op.getOperand(0);
1229 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1230 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1231 SDOperand Callee = Op.getOperand(4);
1232 MVT::ValueType RetVT= Op.Val->getValueType(0);
1233 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1235 // Count how many bytes are to be pushed on the stack.
1236 unsigned NumBytes = 0;
1237 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1238 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1240 static const unsigned GPR8ArgRegs[] = {
1241 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1243 static const unsigned GPR16ArgRegs[] = {
1244 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1246 static const unsigned GPR32ArgRegs[] = {
1247 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1249 static const unsigned GPR64ArgRegs[] = {
1250 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1252 static const unsigned XMMArgRegs[] = {
1253 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1254 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1257 for (unsigned i = 0; i != NumOps; ++i) {
1258 SDOperand Arg = Op.getOperand(5+2*i);
1259 MVT::ValueType ArgVT = Arg.getValueType();
1262 default: assert(0 && "Unknown value type!");
1282 else if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1285 // XMM arguments have to be aligned on 16-byte boundary.
1286 NumBytes = ((NumBytes + 15) / 16) * 16;
1293 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1295 // Arguments go on the stack in reverse order, as specified by the ABI.
1296 unsigned ArgOffset = 0;
1299 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1300 SmallVector<SDOperand, 8> MemOpChains;
1301 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1302 for (unsigned i = 0; i != NumOps; ++i) {
1303 SDOperand Arg = Op.getOperand(5+2*i);
1304 MVT::ValueType ArgVT = Arg.getValueType();
1307 default: assert(0 && "Unexpected ValueType for argument!");
1312 if (NumIntRegs < 6) {
1316 case MVT::i8: Reg = GPR8ArgRegs[NumIntRegs]; break;
1317 case MVT::i16: Reg = GPR16ArgRegs[NumIntRegs]; break;
1318 case MVT::i32: Reg = GPR32ArgRegs[NumIntRegs]; break;
1319 case MVT::i64: Reg = GPR64ArgRegs[NumIntRegs]; break;
1321 RegsToPass.push_back(std::make_pair(Reg, Arg));
1324 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1325 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1326 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1338 if (NumXMMRegs < 8) {
1339 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1342 if (ArgVT != MVT::f32 && ArgVT != MVT::f64) {
1343 // XMM arguments have to be aligned on 16-byte boundary.
1344 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1346 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1347 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1348 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1349 if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1357 if (!MemOpChains.empty())
1358 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1359 &MemOpChains[0], MemOpChains.size());
1361 // Build a sequence of copy-to-reg nodes chained together with token chain
1362 // and flag operands which copy the outgoing args into registers.
1364 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1365 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1367 InFlag = Chain.getValue(1);
1371 // From AMD64 ABI document:
1372 // For calls that may call functions that use varargs or stdargs
1373 // (prototype-less calls or calls to functions containing ellipsis (...) in
1374 // the declaration) %al is used as hidden argument to specify the number
1375 // of SSE registers used. The contents of %al do not need to match exactly
1376 // the number of registers, but must be an ubound on the number of SSE
1377 // registers used and is in the range 0 - 8 inclusive.
1378 Chain = DAG.getCopyToReg(Chain, X86::AL,
1379 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1380 InFlag = Chain.getValue(1);
1383 // If the callee is a GlobalAddress node (quite common, every direct call is)
1384 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1385 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1386 // We should use extra load for direct calls to dllimported functions in
1388 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1389 getTargetMachine(), true))
1390 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1391 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1392 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1394 // Returns a chain & a flag for retval copy to use.
1395 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1396 SmallVector<SDOperand, 8> Ops;
1397 Ops.push_back(Chain);
1398 Ops.push_back(Callee);
1400 // Add argument registers to the end of the list so that they are known live
1402 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1403 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1404 RegsToPass[i].second.getValueType()));
1407 Ops.push_back(InFlag);
1409 // FIXME: Do not generate X86ISD::TAILCALL for now.
1410 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1411 NodeTys, &Ops[0], Ops.size());
1412 InFlag = Chain.getValue(1);
1414 // Returns a flag for retval copy to use.
1415 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1417 Ops.push_back(Chain);
1418 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1419 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1420 Ops.push_back(InFlag);
1421 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1422 if (RetVT != MVT::Other)
1423 InFlag = Chain.getValue(1);
1425 SmallVector<SDOperand, 8> ResultVals;
1427 default: assert(0 && "Unknown value type to return!");
1429 NodeTys = DAG.getVTList(MVT::Other);
1432 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
1433 ResultVals.push_back(Chain.getValue(0));
1434 NodeTys = DAG.getVTList(MVT::i8, MVT::Other);
1437 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
1438 ResultVals.push_back(Chain.getValue(0));
1439 NodeTys = DAG.getVTList(MVT::i16, MVT::Other);
1442 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1443 ResultVals.push_back(Chain.getValue(0));
1444 NodeTys = DAG.getVTList(MVT::i32, MVT::Other);
1447 if (Op.Val->getValueType(1) == MVT::i64) {
1448 // FIXME: __int128 support?
1449 Chain = DAG.getCopyFromReg(Chain, X86::RAX, MVT::i64, InFlag).getValue(1);
1450 ResultVals.push_back(Chain.getValue(0));
1451 Chain = DAG.getCopyFromReg(Chain, X86::RDX, MVT::i64,
1452 Chain.getValue(2)).getValue(1);
1453 ResultVals.push_back(Chain.getValue(0));
1454 NodeTys = DAG.getVTList(MVT::i64, MVT::i64, MVT::Other);
1456 Chain = DAG.getCopyFromReg(Chain, X86::RAX, MVT::i64, InFlag).getValue(1);
1457 ResultVals.push_back(Chain.getValue(0));
1458 NodeTys = DAG.getVTList(MVT::i64, MVT::Other);
1469 // FIXME: long double support?
1470 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
1471 ResultVals.push_back(Chain.getValue(0));
1472 NodeTys = DAG.getVTList(RetVT, MVT::Other);
1476 // Merge everything together with a MERGE_VALUES node.
1477 ResultVals.push_back(Chain);
1478 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1479 &ResultVals[0], ResultVals.size());
1480 return Res.getValue(Op.ResNo);
1483 //===----------------------------------------------------------------------===//
1484 // Fast & FastCall Calling Convention implementation
1485 //===----------------------------------------------------------------------===//
1487 // The X86 'fast' calling convention passes up to two integer arguments in
1488 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
1489 // and requires that the callee pop its arguments off the stack (allowing proper
1490 // tail calls), and has the same return value conventions as C calling convs.
1492 // This calling convention always arranges for the callee pop value to be 8n+4
1493 // bytes, which is needed for tail recursion elimination and stack alignment
1496 // Note that this can be enhanced in the future to pass fp vals in registers
1497 // (when we have a global fp allocator) and do other tricks.
1499 //===----------------------------------------------------------------------===//
1500 // The X86 'fastcall' calling convention passes up to two integer arguments in
1501 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1502 // and requires that the callee pop its arguments off the stack (allowing proper
1503 // tail calls), and has the same return value conventions as C calling convs.
1505 // This calling convention always arranges for the callee pop value to be 8n+4
1506 // bytes, which is needed for tail recursion elimination and stack alignment
1511 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG,
1513 unsigned NumArgs = Op.Val->getNumValues()-1;
1514 MachineFunction &MF = DAG.getMachineFunction();
1515 MachineFrameInfo *MFI = MF.getFrameInfo();
1516 SDOperand Root = Op.getOperand(0);
1517 SmallVector<SDOperand, 8> ArgValues;
1519 // Add DAG nodes to load the arguments... On entry to a function the stack
1520 // frame looks like this:
1522 // [ESP] -- return address
1523 // [ESP + 4] -- first nonreg argument (leftmost lexically)
1524 // [ESP + 8] -- second nonreg argument, if 1st argument is <= 4 bytes in size
1526 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1528 // Keep track of the number of integer regs passed so far. This can be either
1529 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1531 unsigned NumIntRegs = 0;
1532 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1534 static const unsigned XMMArgRegs[] = {
1535 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1538 static const unsigned GPRArgRegs[][2][2] = {
1539 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1540 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1541 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1544 static const TargetRegisterClass* GPRClasses[3] = {
1545 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
1548 unsigned GPRInd = (isFastCall ? 1 : 0);
1549 for (unsigned i = 0; i < NumArgs; ++i) {
1550 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1551 unsigned ArgIncrement = 4;
1552 unsigned ObjSize = 0;
1553 unsigned ObjXMMRegs = 0;
1554 unsigned ObjIntRegs = 0;
1558 HowToPassCallArgument(ObjectVT,
1559 true, // Use as much registers as possible
1560 NumIntRegs, NumXMMRegs,
1561 (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS),
1562 ObjSize, ObjIntRegs, ObjXMMRegs,
1566 ArgIncrement = ObjSize;
1568 if (ObjIntRegs || ObjXMMRegs) {
1570 default: assert(0 && "Unhandled argument type!");
1574 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][GPRInd][NumIntRegs];
1575 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
1576 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1585 assert(!isFastCall && "Unhandled argument type!");
1586 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
1587 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1591 NumIntRegs += ObjIntRegs;
1592 NumXMMRegs += ObjXMMRegs;
1595 // XMM arguments have to be aligned on 16-byte boundary.
1597 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1598 // Create the SelectionDAG nodes corresponding to a load from this
1600 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1601 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1602 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1604 ArgOffset += ArgIncrement; // Move on to the next argument.
1607 ArgValues.push_back(ArgValue);
1610 ArgValues.push_back(Root);
1612 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1613 // arguments and the arguments after the retaddr has been pushed are aligned.
1614 if ((ArgOffset & 7) == 0)
1617 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1618 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1619 ReturnAddrIndex = 0; // No return address slot generated yet.
1620 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
1621 BytesCallerReserves = 0;
1623 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
1625 // Finally, inform the code generator which regs we return values in.
1626 switch (getValueType(MF.getFunction()->getReturnType())) {
1627 default: assert(0 && "Unknown type!");
1628 case MVT::isVoid: break;
1633 MF.addLiveOut(X86::EAX);
1636 MF.addLiveOut(X86::EAX);
1637 MF.addLiveOut(X86::EDX);
1641 MF.addLiveOut(X86::ST0);
1649 assert(!isFastCall && "Unknown result type");
1650 MF.addLiveOut(X86::XMM0);
1654 // Return the new list of results.
1655 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1656 &ArgValues[0], ArgValues.size());
1659 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1661 SDOperand Chain = Op.getOperand(0);
1662 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1663 SDOperand Callee = Op.getOperand(4);
1664 MVT::ValueType RetVT= Op.Val->getValueType(0);
1665 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1667 // Count how many bytes are to be pushed on the stack.
1668 unsigned NumBytes = 0;
1670 // Keep track of the number of integer regs passed so far. This can be either
1671 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1673 unsigned NumIntRegs = 0;
1674 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1676 static const unsigned GPRArgRegs[][2][2] = {
1677 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1678 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1679 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1681 static const unsigned XMMArgRegs[] = {
1682 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1685 unsigned GPRInd = (isFastCall ? 1 : 0);
1686 for (unsigned i = 0; i != NumOps; ++i) {
1687 SDOperand Arg = Op.getOperand(5+2*i);
1689 switch (Arg.getValueType()) {
1690 default: assert(0 && "Unknown value type!");
1694 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1695 if (NumIntRegs < MaxNumIntRegs) {
1712 assert(!isFastCall && "Unknown value type!");
1716 // XMM arguments have to be aligned on 16-byte boundary.
1717 NumBytes = ((NumBytes + 15) / 16) * 16;
1724 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1725 // arguments and the arguments after the retaddr has been pushed are aligned.
1726 if ((NumBytes & 7) == 0)
1729 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1731 // Arguments go on the stack in reverse order, as specified by the ABI.
1732 unsigned ArgOffset = 0;
1734 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1735 SmallVector<SDOperand, 8> MemOpChains;
1736 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1737 for (unsigned i = 0; i != NumOps; ++i) {
1738 SDOperand Arg = Op.getOperand(5+2*i);
1740 switch (Arg.getValueType()) {
1741 default: assert(0 && "Unexpected ValueType for argument!");
1745 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1746 if (NumIntRegs < MaxNumIntRegs) {
1748 GPRArgRegs[Arg.getValueType()-MVT::i8][GPRInd][NumIntRegs];
1749 RegsToPass.push_back(std::make_pair(RegToUse, Arg));
1755 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1756 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1757 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1762 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1763 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1764 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1774 assert(!isFastCall && "Unexpected ValueType for argument!");
1775 if (NumXMMRegs < 4) {
1776 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1779 // XMM arguments have to be aligned on 16-byte boundary.
1780 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1781 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1782 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1783 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1790 if (!MemOpChains.empty())
1791 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1792 &MemOpChains[0], MemOpChains.size());
1794 // Build a sequence of copy-to-reg nodes chained together with token chain
1795 // and flag operands which copy the outgoing args into registers.
1797 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1798 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1800 InFlag = Chain.getValue(1);
1803 // If the callee is a GlobalAddress node (quite common, every direct call is)
1804 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1805 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1806 // We should use extra load for direct calls to dllimported functions in
1808 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1809 getTargetMachine(), true))
1810 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1811 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1812 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1814 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1816 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1817 Subtarget->isPICStyleGOT()) {
1818 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1819 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1821 InFlag = Chain.getValue(1);
1824 // Returns a chain & a flag for retval copy to use.
1825 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1826 SmallVector<SDOperand, 8> Ops;
1827 Ops.push_back(Chain);
1828 Ops.push_back(Callee);
1830 // Add argument registers to the end of the list so that they are known live
1832 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1833 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1834 RegsToPass[i].second.getValueType()));
1836 // Add an implicit use GOT pointer in EBX.
1837 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1838 Subtarget->isPICStyleGOT())
1839 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1842 Ops.push_back(InFlag);
1844 // FIXME: Do not generate X86ISD::TAILCALL for now.
1845 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1846 NodeTys, &Ops[0], Ops.size());
1847 InFlag = Chain.getValue(1);
1849 // Returns a flag for retval copy to use.
1850 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1852 Ops.push_back(Chain);
1853 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1854 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1855 Ops.push_back(InFlag);
1856 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1857 if (RetVT != MVT::Other)
1858 InFlag = Chain.getValue(1);
1860 SmallVector<SDOperand, 8> ResultVals;
1862 default: assert(0 && "Unknown value type to return!");
1864 NodeTys = DAG.getVTList(MVT::Other);
1867 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
1868 ResultVals.push_back(Chain.getValue(0));
1869 NodeTys = DAG.getVTList(MVT::i8, MVT::Other);
1872 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
1873 ResultVals.push_back(Chain.getValue(0));
1874 NodeTys = DAG.getVTList(MVT::i16, MVT::Other);
1877 if (Op.Val->getValueType(1) == MVT::i32) {
1878 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1879 ResultVals.push_back(Chain.getValue(0));
1880 Chain = DAG.getCopyFromReg(Chain, X86::EDX, MVT::i32,
1881 Chain.getValue(2)).getValue(1);
1882 ResultVals.push_back(Chain.getValue(0));
1883 NodeTys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
1885 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1886 ResultVals.push_back(Chain.getValue(0));
1887 NodeTys = DAG.getVTList(MVT::i32, MVT::Other);
1897 assert(0 && "Unknown value type to return!");
1899 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
1900 ResultVals.push_back(Chain.getValue(0));
1901 NodeTys = DAG.getVTList(RetVT, MVT::Other);
1906 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
1907 SmallVector<SDOperand, 8> Ops;
1908 Ops.push_back(Chain);
1909 Ops.push_back(InFlag);
1910 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys,
1911 &Ops[0], Ops.size());
1912 Chain = RetVal.getValue(1);
1913 InFlag = RetVal.getValue(2);
1915 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
1916 // shouldn't be necessary except that RFP cannot be live across
1917 // multiple blocks. When stackifier is fixed, they can be uncoupled.
1918 MachineFunction &MF = DAG.getMachineFunction();
1919 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
1920 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
1921 Tys = DAG.getVTList(MVT::Other);
1923 Ops.push_back(Chain);
1924 Ops.push_back(RetVal);
1925 Ops.push_back(StackSlot);
1926 Ops.push_back(DAG.getValueType(RetVT));
1927 Ops.push_back(InFlag);
1928 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
1929 RetVal = DAG.getLoad(RetVT, Chain, StackSlot, NULL, 0);
1930 Chain = RetVal.getValue(1);
1933 if (RetVT == MVT::f32 && !X86ScalarSSE)
1934 // FIXME: we would really like to remember that this FP_ROUND
1935 // operation is okay to eliminate if we allow excess FP precision.
1936 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
1937 ResultVals.push_back(RetVal);
1938 NodeTys = DAG.getVTList(RetVT, MVT::Other);
1944 // Merge everything together with a MERGE_VALUES node.
1945 ResultVals.push_back(Chain);
1946 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1947 &ResultVals[0], ResultVals.size());
1948 return Res.getValue(Op.ResNo);
1951 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1952 if (ReturnAddrIndex == 0) {
1953 // Set up a frame object for the return address.
1954 MachineFunction &MF = DAG.getMachineFunction();
1955 if (Subtarget->is64Bit())
1956 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1958 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1961 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1966 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1967 /// specific condition code. It returns a false if it cannot do a direct
1968 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1970 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1971 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1972 SelectionDAG &DAG) {
1973 X86CC = X86::COND_INVALID;
1975 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1976 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1977 // X > -1 -> X == 0, jump !sign.
1978 RHS = DAG.getConstant(0, RHS.getValueType());
1979 X86CC = X86::COND_NS;
1981 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1982 // X < 0 -> X == 0, jump on sign.
1983 X86CC = X86::COND_S;
1988 switch (SetCCOpcode) {
1990 case ISD::SETEQ: X86CC = X86::COND_E; break;
1991 case ISD::SETGT: X86CC = X86::COND_G; break;
1992 case ISD::SETGE: X86CC = X86::COND_GE; break;
1993 case ISD::SETLT: X86CC = X86::COND_L; break;
1994 case ISD::SETLE: X86CC = X86::COND_LE; break;
1995 case ISD::SETNE: X86CC = X86::COND_NE; break;
1996 case ISD::SETULT: X86CC = X86::COND_B; break;
1997 case ISD::SETUGT: X86CC = X86::COND_A; break;
1998 case ISD::SETULE: X86CC = X86::COND_BE; break;
1999 case ISD::SETUGE: X86CC = X86::COND_AE; break;
2002 // On a floating point condition, the flags are set as follows:
2004 // 0 | 0 | 0 | X > Y
2005 // 0 | 0 | 1 | X < Y
2006 // 1 | 0 | 0 | X == Y
2007 // 1 | 1 | 1 | unordered
2009 switch (SetCCOpcode) {
2012 case ISD::SETEQ: X86CC = X86::COND_E; break;
2013 case ISD::SETOLT: Flip = true; // Fallthrough
2015 case ISD::SETGT: X86CC = X86::COND_A; break;
2016 case ISD::SETOLE: Flip = true; // Fallthrough
2018 case ISD::SETGE: X86CC = X86::COND_AE; break;
2019 case ISD::SETUGT: Flip = true; // Fallthrough
2021 case ISD::SETLT: X86CC = X86::COND_B; break;
2022 case ISD::SETUGE: Flip = true; // Fallthrough
2024 case ISD::SETLE: X86CC = X86::COND_BE; break;
2026 case ISD::SETNE: X86CC = X86::COND_NE; break;
2027 case ISD::SETUO: X86CC = X86::COND_P; break;
2028 case ISD::SETO: X86CC = X86::COND_NP; break;
2031 std::swap(LHS, RHS);
2034 return X86CC != X86::COND_INVALID;
2037 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2038 /// code. Current x86 isa includes the following FP cmov instructions:
2039 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2040 static bool hasFPCMov(unsigned X86CC) {
2056 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2057 /// true if Op is undef or if its value falls within the specified range (L, H].
2058 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2059 if (Op.getOpcode() == ISD::UNDEF)
2062 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2063 return (Val >= Low && Val < Hi);
2066 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2067 /// true if Op is undef or if its value equal to the specified value.
2068 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2069 if (Op.getOpcode() == ISD::UNDEF)
2071 return cast<ConstantSDNode>(Op)->getValue() == Val;
2074 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2075 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2076 bool X86::isPSHUFDMask(SDNode *N) {
2077 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2079 if (N->getNumOperands() != 4)
2082 // Check if the value doesn't reference the second vector.
2083 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2084 SDOperand Arg = N->getOperand(i);
2085 if (Arg.getOpcode() == ISD::UNDEF) continue;
2086 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2087 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
2094 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2095 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2096 bool X86::isPSHUFHWMask(SDNode *N) {
2097 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2099 if (N->getNumOperands() != 8)
2102 // Lower quadword copied in order.
2103 for (unsigned i = 0; i != 4; ++i) {
2104 SDOperand Arg = N->getOperand(i);
2105 if (Arg.getOpcode() == ISD::UNDEF) continue;
2106 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2107 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2111 // Upper quadword shuffled.
2112 for (unsigned i = 4; i != 8; ++i) {
2113 SDOperand Arg = N->getOperand(i);
2114 if (Arg.getOpcode() == ISD::UNDEF) continue;
2115 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2116 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2117 if (Val < 4 || Val > 7)
2124 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2125 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2126 bool X86::isPSHUFLWMask(SDNode *N) {
2127 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2129 if (N->getNumOperands() != 8)
2132 // Upper quadword copied in order.
2133 for (unsigned i = 4; i != 8; ++i)
2134 if (!isUndefOrEqual(N->getOperand(i), i))
2137 // Lower quadword shuffled.
2138 for (unsigned i = 0; i != 4; ++i)
2139 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2145 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2146 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2147 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2148 if (NumElems != 2 && NumElems != 4) return false;
2150 unsigned Half = NumElems / 2;
2151 for (unsigned i = 0; i < Half; ++i)
2152 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2154 for (unsigned i = Half; i < NumElems; ++i)
2155 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2161 bool X86::isSHUFPMask(SDNode *N) {
2162 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2163 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2166 /// isCommutedSHUFP - Returns true if the shuffle mask is except
2167 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2168 /// half elements to come from vector 1 (which would equal the dest.) and
2169 /// the upper half to come from vector 2.
2170 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2171 if (NumOps != 2 && NumOps != 4) return false;
2173 unsigned Half = NumOps / 2;
2174 for (unsigned i = 0; i < Half; ++i)
2175 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2177 for (unsigned i = Half; i < NumOps; ++i)
2178 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2183 static bool isCommutedSHUFP(SDNode *N) {
2184 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2185 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2188 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2189 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2190 bool X86::isMOVHLPSMask(SDNode *N) {
2191 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2193 if (N->getNumOperands() != 4)
2196 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2197 return isUndefOrEqual(N->getOperand(0), 6) &&
2198 isUndefOrEqual(N->getOperand(1), 7) &&
2199 isUndefOrEqual(N->getOperand(2), 2) &&
2200 isUndefOrEqual(N->getOperand(3), 3);
2203 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2204 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2206 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2207 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2209 if (N->getNumOperands() != 4)
2212 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2213 return isUndefOrEqual(N->getOperand(0), 2) &&
2214 isUndefOrEqual(N->getOperand(1), 3) &&
2215 isUndefOrEqual(N->getOperand(2), 2) &&
2216 isUndefOrEqual(N->getOperand(3), 3);
2219 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2220 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2221 bool X86::isMOVLPMask(SDNode *N) {
2222 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2224 unsigned NumElems = N->getNumOperands();
2225 if (NumElems != 2 && NumElems != 4)
2228 for (unsigned i = 0; i < NumElems/2; ++i)
2229 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2232 for (unsigned i = NumElems/2; i < NumElems; ++i)
2233 if (!isUndefOrEqual(N->getOperand(i), i))
2239 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2240 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2242 bool X86::isMOVHPMask(SDNode *N) {
2243 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2245 unsigned NumElems = N->getNumOperands();
2246 if (NumElems != 2 && NumElems != 4)
2249 for (unsigned i = 0; i < NumElems/2; ++i)
2250 if (!isUndefOrEqual(N->getOperand(i), i))
2253 for (unsigned i = 0; i < NumElems/2; ++i) {
2254 SDOperand Arg = N->getOperand(i + NumElems/2);
2255 if (!isUndefOrEqual(Arg, i + NumElems))
2262 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2263 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2264 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2265 bool V2IsSplat = false) {
2266 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2269 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2270 SDOperand BitI = Elts[i];
2271 SDOperand BitI1 = Elts[i+1];
2272 if (!isUndefOrEqual(BitI, j))
2275 if (isUndefOrEqual(BitI1, NumElts))
2278 if (!isUndefOrEqual(BitI1, j + NumElts))
2286 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2287 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2288 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2291 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2292 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2293 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2294 bool V2IsSplat = false) {
2295 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2298 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2299 SDOperand BitI = Elts[i];
2300 SDOperand BitI1 = Elts[i+1];
2301 if (!isUndefOrEqual(BitI, j + NumElts/2))
2304 if (isUndefOrEqual(BitI1, NumElts))
2307 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2315 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2316 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2317 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2320 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2321 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2323 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2324 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2326 unsigned NumElems = N->getNumOperands();
2327 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
2330 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2331 SDOperand BitI = N->getOperand(i);
2332 SDOperand BitI1 = N->getOperand(i+1);
2334 if (!isUndefOrEqual(BitI, j))
2336 if (!isUndefOrEqual(BitI1, j))
2343 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2344 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2345 /// MOVSD, and MOVD, i.e. setting the lowest element.
2346 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2347 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2350 if (!isUndefOrEqual(Elts[0], NumElts))
2353 for (unsigned i = 1; i < NumElts; ++i) {
2354 if (!isUndefOrEqual(Elts[i], i))
2361 bool X86::isMOVLMask(SDNode *N) {
2362 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2363 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2366 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2367 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2368 /// element of vector 2 and the other elements to come from vector 1 in order.
2369 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2370 bool V2IsSplat = false,
2371 bool V2IsUndef = false) {
2372 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2375 if (!isUndefOrEqual(Ops[0], 0))
2378 for (unsigned i = 1; i < NumOps; ++i) {
2379 SDOperand Arg = Ops[i];
2380 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2381 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2382 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2389 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2390 bool V2IsUndef = false) {
2391 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2392 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2393 V2IsSplat, V2IsUndef);
2396 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2397 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2398 bool X86::isMOVSHDUPMask(SDNode *N) {
2399 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2401 if (N->getNumOperands() != 4)
2404 // Expect 1, 1, 3, 3
2405 for (unsigned i = 0; i < 2; ++i) {
2406 SDOperand Arg = N->getOperand(i);
2407 if (Arg.getOpcode() == ISD::UNDEF) continue;
2408 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2409 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2410 if (Val != 1) return false;
2414 for (unsigned i = 2; i < 4; ++i) {
2415 SDOperand Arg = N->getOperand(i);
2416 if (Arg.getOpcode() == ISD::UNDEF) continue;
2417 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2418 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2419 if (Val != 3) return false;
2423 // Don't use movshdup if it can be done with a shufps.
2427 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2428 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2429 bool X86::isMOVSLDUPMask(SDNode *N) {
2430 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2432 if (N->getNumOperands() != 4)
2435 // Expect 0, 0, 2, 2
2436 for (unsigned i = 0; i < 2; ++i) {
2437 SDOperand Arg = N->getOperand(i);
2438 if (Arg.getOpcode() == ISD::UNDEF) continue;
2439 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2440 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2441 if (Val != 0) return false;
2445 for (unsigned i = 2; i < 4; ++i) {
2446 SDOperand Arg = N->getOperand(i);
2447 if (Arg.getOpcode() == ISD::UNDEF) continue;
2448 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2449 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2450 if (Val != 2) return false;
2454 // Don't use movshdup if it can be done with a shufps.
2458 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2459 /// a splat of a single element.
2460 static bool isSplatMask(SDNode *N) {
2461 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2463 // This is a splat operation if each element of the permute is the same, and
2464 // if the value doesn't reference the second vector.
2465 unsigned NumElems = N->getNumOperands();
2466 SDOperand ElementBase;
2468 for (; i != NumElems; ++i) {
2469 SDOperand Elt = N->getOperand(i);
2470 if (isa<ConstantSDNode>(Elt)) {
2476 if (!ElementBase.Val)
2479 for (; i != NumElems; ++i) {
2480 SDOperand Arg = N->getOperand(i);
2481 if (Arg.getOpcode() == ISD::UNDEF) continue;
2482 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2483 if (Arg != ElementBase) return false;
2486 // Make sure it is a splat of the first vector operand.
2487 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2490 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2491 /// a splat of a single element and it's a 2 or 4 element mask.
2492 bool X86::isSplatMask(SDNode *N) {
2493 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2495 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2496 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2498 return ::isSplatMask(N);
2501 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2502 /// specifies a splat of zero element.
2503 bool X86::isSplatLoMask(SDNode *N) {
2504 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2506 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2507 if (!isUndefOrEqual(N->getOperand(i), 0))
2512 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2513 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2515 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2516 unsigned NumOperands = N->getNumOperands();
2517 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2519 for (unsigned i = 0; i < NumOperands; ++i) {
2521 SDOperand Arg = N->getOperand(NumOperands-i-1);
2522 if (Arg.getOpcode() != ISD::UNDEF)
2523 Val = cast<ConstantSDNode>(Arg)->getValue();
2524 if (Val >= NumOperands) Val -= NumOperands;
2526 if (i != NumOperands - 1)
2533 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2534 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2536 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2538 // 8 nodes, but we only care about the last 4.
2539 for (unsigned i = 7; i >= 4; --i) {
2541 SDOperand Arg = N->getOperand(i);
2542 if (Arg.getOpcode() != ISD::UNDEF)
2543 Val = cast<ConstantSDNode>(Arg)->getValue();
2552 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2553 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2555 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2557 // 8 nodes, but we only care about the first 4.
2558 for (int i = 3; i >= 0; --i) {
2560 SDOperand Arg = N->getOperand(i);
2561 if (Arg.getOpcode() != ISD::UNDEF)
2562 Val = cast<ConstantSDNode>(Arg)->getValue();
2571 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2572 /// specifies a 8 element shuffle that can be broken into a pair of
2573 /// PSHUFHW and PSHUFLW.
2574 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2575 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2577 if (N->getNumOperands() != 8)
2580 // Lower quadword shuffled.
2581 for (unsigned i = 0; i != 4; ++i) {
2582 SDOperand Arg = N->getOperand(i);
2583 if (Arg.getOpcode() == ISD::UNDEF) continue;
2584 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2585 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2590 // Upper quadword shuffled.
2591 for (unsigned i = 4; i != 8; ++i) {
2592 SDOperand Arg = N->getOperand(i);
2593 if (Arg.getOpcode() == ISD::UNDEF) continue;
2594 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2595 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2596 if (Val < 4 || Val > 7)
2603 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2604 /// values in ther permute mask.
2605 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2606 SDOperand &V2, SDOperand &Mask,
2607 SelectionDAG &DAG) {
2608 MVT::ValueType VT = Op.getValueType();
2609 MVT::ValueType MaskVT = Mask.getValueType();
2610 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2611 unsigned NumElems = Mask.getNumOperands();
2612 SmallVector<SDOperand, 8> MaskVec;
2614 for (unsigned i = 0; i != NumElems; ++i) {
2615 SDOperand Arg = Mask.getOperand(i);
2616 if (Arg.getOpcode() == ISD::UNDEF) {
2617 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2620 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2621 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2623 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2625 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2629 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2630 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
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(SDNode *Mask) {
2638 unsigned NumElems = Mask->getNumOperands();
2641 for (unsigned i = 0, e = 2; i != e; ++i)
2642 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2644 for (unsigned i = 2; i != 4; ++i)
2645 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2650 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2651 /// is promoted to a vector.
2652 static inline bool isScalarLoadToVector(SDNode *N) {
2653 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2654 N = N->getOperand(0).Val;
2655 return ISD::isNON_EXTLoad(N);
2660 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2661 /// match movlp{s|d}. The lower half elements should come from lower half of
2662 /// V1 (and in order), and the upper half elements should come from the upper
2663 /// half of V2 (and in order). And since V1 will become the source of the
2664 /// MOVLP, it must be either a vector load or a scalar load to vector.
2665 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2666 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2668 // Is V2 is a vector load, don't do this transformation. We will try to use
2669 // load folding shufps op.
2670 if (ISD::isNON_EXTLoad(V2))
2673 unsigned NumElems = Mask->getNumOperands();
2674 if (NumElems != 2 && NumElems != 4)
2676 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2677 if (!isUndefOrEqual(Mask->getOperand(i), i))
2679 for (unsigned i = NumElems/2; i != NumElems; ++i)
2680 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2685 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2687 static bool isSplatVector(SDNode *N) {
2688 if (N->getOpcode() != ISD::BUILD_VECTOR)
2691 SDOperand SplatValue = N->getOperand(0);
2692 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2693 if (N->getOperand(i) != SplatValue)
2698 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2700 static bool isUndefShuffle(SDNode *N) {
2701 if (N->getOpcode() != ISD::BUILD_VECTOR)
2704 SDOperand V1 = N->getOperand(0);
2705 SDOperand V2 = N->getOperand(1);
2706 SDOperand Mask = N->getOperand(2);
2707 unsigned NumElems = Mask.getNumOperands();
2708 for (unsigned i = 0; i != NumElems; ++i) {
2709 SDOperand Arg = Mask.getOperand(i);
2710 if (Arg.getOpcode() != ISD::UNDEF) {
2711 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2712 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2714 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2721 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2722 /// that point to V2 points to its first element.
2723 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2724 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2726 bool Changed = false;
2727 SmallVector<SDOperand, 8> MaskVec;
2728 unsigned NumElems = Mask.getNumOperands();
2729 for (unsigned i = 0; i != NumElems; ++i) {
2730 SDOperand Arg = Mask.getOperand(i);
2731 if (Arg.getOpcode() != ISD::UNDEF) {
2732 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2733 if (Val > NumElems) {
2734 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2738 MaskVec.push_back(Arg);
2742 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2743 &MaskVec[0], MaskVec.size());
2747 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2748 /// operation of specified width.
2749 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2750 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2751 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2753 SmallVector<SDOperand, 8> MaskVec;
2754 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2755 for (unsigned i = 1; i != NumElems; ++i)
2756 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2757 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2760 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2761 /// of specified width.
2762 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2763 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2764 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2765 SmallVector<SDOperand, 8> MaskVec;
2766 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2767 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2768 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2770 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2773 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2774 /// of specified width.
2775 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2776 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2777 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2778 unsigned Half = NumElems/2;
2779 SmallVector<SDOperand, 8> MaskVec;
2780 for (unsigned i = 0; i != Half; ++i) {
2781 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2782 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2784 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2787 /// getZeroVector - Returns a vector of specified type with all zero elements.
2789 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2790 assert(MVT::isVector(VT) && "Expected a vector type");
2791 unsigned NumElems = getVectorNumElements(VT);
2792 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2793 bool isFP = MVT::isFloatingPoint(EVT);
2794 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2795 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2796 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2799 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2801 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2802 SDOperand V1 = Op.getOperand(0);
2803 SDOperand Mask = Op.getOperand(2);
2804 MVT::ValueType VT = Op.getValueType();
2805 unsigned NumElems = Mask.getNumOperands();
2806 Mask = getUnpacklMask(NumElems, DAG);
2807 while (NumElems != 4) {
2808 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2811 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2813 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2814 Mask = getZeroVector(MaskVT, DAG);
2815 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2816 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2817 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2820 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2822 static inline bool isZeroNode(SDOperand Elt) {
2823 return ((isa<ConstantSDNode>(Elt) &&
2824 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2825 (isa<ConstantFPSDNode>(Elt) &&
2826 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2829 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2830 /// vector and zero or undef vector.
2831 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2832 unsigned NumElems, unsigned Idx,
2833 bool isZero, SelectionDAG &DAG) {
2834 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2835 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2836 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2837 SDOperand Zero = DAG.getConstant(0, EVT);
2838 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2839 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2840 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2841 &MaskVec[0], MaskVec.size());
2842 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2845 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2847 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2848 unsigned NumNonZero, unsigned NumZero,
2849 SelectionDAG &DAG, TargetLowering &TLI) {
2855 for (unsigned i = 0; i < 16; ++i) {
2856 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2857 if (ThisIsNonZero && First) {
2859 V = getZeroVector(MVT::v8i16, DAG);
2861 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2866 SDOperand ThisElt(0, 0), LastElt(0, 0);
2867 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2868 if (LastIsNonZero) {
2869 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2871 if (ThisIsNonZero) {
2872 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2873 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2874 ThisElt, DAG.getConstant(8, MVT::i8));
2876 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2881 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2882 DAG.getConstant(i/2, TLI.getPointerTy()));
2886 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2889 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2891 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2892 unsigned NumNonZero, unsigned NumZero,
2893 SelectionDAG &DAG, TargetLowering &TLI) {
2899 for (unsigned i = 0; i < 8; ++i) {
2900 bool isNonZero = (NonZeros & (1 << i)) != 0;
2904 V = getZeroVector(MVT::v8i16, DAG);
2906 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2909 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2910 DAG.getConstant(i, TLI.getPointerTy()));
2918 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2919 // All zero's are handled with pxor.
2920 if (ISD::isBuildVectorAllZeros(Op.Val))
2923 // All one's are handled with pcmpeqd.
2924 if (ISD::isBuildVectorAllOnes(Op.Val))
2927 MVT::ValueType VT = Op.getValueType();
2928 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2929 unsigned EVTBits = MVT::getSizeInBits(EVT);
2931 unsigned NumElems = Op.getNumOperands();
2932 unsigned NumZero = 0;
2933 unsigned NumNonZero = 0;
2934 unsigned NonZeros = 0;
2935 std::set<SDOperand> Values;
2936 for (unsigned i = 0; i < NumElems; ++i) {
2937 SDOperand Elt = Op.getOperand(i);
2938 if (Elt.getOpcode() != ISD::UNDEF) {
2940 if (isZeroNode(Elt))
2943 NonZeros |= (1 << i);
2949 if (NumNonZero == 0)
2950 // Must be a mix of zero and undef. Return a zero vector.
2951 return getZeroVector(VT, DAG);
2953 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2954 if (Values.size() == 1)
2957 // Special case for single non-zero element.
2958 if (NumNonZero == 1) {
2959 unsigned Idx = CountTrailingZeros_32(NonZeros);
2960 SDOperand Item = Op.getOperand(Idx);
2961 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2963 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2964 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2967 if (EVTBits == 32) {
2968 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2969 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2971 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2972 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2973 SmallVector<SDOperand, 8> MaskVec;
2974 for (unsigned i = 0; i < NumElems; i++)
2975 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2976 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2977 &MaskVec[0], MaskVec.size());
2978 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2979 DAG.getNode(ISD::UNDEF, VT), Mask);
2983 // Let legalizer expand 2-wide build_vector's.
2987 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2989 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2991 if (V.Val) return V;
2994 if (EVTBits == 16) {
2995 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2997 if (V.Val) return V;
3000 // If element VT is == 32 bits, turn it into a number of shuffles.
3001 SmallVector<SDOperand, 8> V;
3003 if (NumElems == 4 && NumZero > 0) {
3004 for (unsigned i = 0; i < 4; ++i) {
3005 bool isZero = !(NonZeros & (1 << i));
3007 V[i] = getZeroVector(VT, DAG);
3009 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3012 for (unsigned i = 0; i < 2; ++i) {
3013 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3016 V[i] = V[i*2]; // Must be a zero vector.
3019 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3020 getMOVLMask(NumElems, DAG));
3023 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3024 getMOVLMask(NumElems, DAG));
3027 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3028 getUnpacklMask(NumElems, DAG));
3033 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3034 // clears the upper bits.
3035 // FIXME: we can do the same for v4f32 case when we know both parts of
3036 // the lower half come from scalar_to_vector (loadf32). We should do
3037 // that in post legalizer dag combiner with target specific hooks.
3038 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3040 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3041 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
3042 SmallVector<SDOperand, 8> MaskVec;
3043 bool Reverse = (NonZeros & 0x3) == 2;
3044 for (unsigned i = 0; i < 2; ++i)
3046 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3048 MaskVec.push_back(DAG.getConstant(i, EVT));
3049 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3050 for (unsigned i = 0; i < 2; ++i)
3052 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3054 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3055 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3056 &MaskVec[0], MaskVec.size());
3057 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3060 if (Values.size() > 2) {
3061 // Expand into a number of unpckl*.
3063 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3064 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3065 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3066 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3067 for (unsigned i = 0; i < NumElems; ++i)
3068 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3070 while (NumElems != 0) {
3071 for (unsigned i = 0; i < NumElems; ++i)
3072 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3083 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3084 SDOperand V1 = Op.getOperand(0);
3085 SDOperand V2 = Op.getOperand(1);
3086 SDOperand PermMask = Op.getOperand(2);
3087 MVT::ValueType VT = Op.getValueType();
3088 unsigned NumElems = PermMask.getNumOperands();
3089 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3090 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3091 bool V1IsSplat = false;
3092 bool V2IsSplat = false;
3094 if (isUndefShuffle(Op.Val))
3095 return DAG.getNode(ISD::UNDEF, VT);
3097 if (isSplatMask(PermMask.Val)) {
3098 if (NumElems <= 4) return Op;
3099 // Promote it to a v4i32 splat.
3100 return PromoteSplat(Op, DAG);
3103 if (X86::isMOVLMask(PermMask.Val))
3104 return (V1IsUndef) ? V2 : Op;
3106 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3107 X86::isMOVSLDUPMask(PermMask.Val) ||
3108 X86::isMOVHLPSMask(PermMask.Val) ||
3109 X86::isMOVHPMask(PermMask.Val) ||
3110 X86::isMOVLPMask(PermMask.Val))
3113 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3114 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3115 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3117 bool Commuted = false;
3118 V1IsSplat = isSplatVector(V1.Val);
3119 V2IsSplat = isSplatVector(V2.Val);
3120 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3121 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3122 std::swap(V1IsSplat, V2IsSplat);
3123 std::swap(V1IsUndef, V2IsUndef);
3127 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3128 if (V2IsUndef) return V1;
3129 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3131 // V2 is a splat, so the mask may be malformed. That is, it may point
3132 // to any V2 element. The instruction selectior won't like this. Get
3133 // a corrected mask and commute to form a proper MOVS{S|D}.
3134 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3135 if (NewMask.Val != PermMask.Val)
3136 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3141 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3142 X86::isUNPCKLMask(PermMask.Val) ||
3143 X86::isUNPCKHMask(PermMask.Val))
3147 // Normalize mask so all entries that point to V2 points to its first
3148 // element then try to match unpck{h|l} again. If match, return a
3149 // new vector_shuffle with the corrected mask.
3150 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3151 if (NewMask.Val != PermMask.Val) {
3152 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3153 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3154 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3155 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3156 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3157 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3162 // Normalize the node to match x86 shuffle ops if needed
3163 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3164 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3167 // Commute is back and try unpck* again.
3168 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3169 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3170 X86::isUNPCKLMask(PermMask.Val) ||
3171 X86::isUNPCKHMask(PermMask.Val))
3175 // If VT is integer, try PSHUF* first, then SHUFP*.
3176 if (MVT::isInteger(VT)) {
3177 if (X86::isPSHUFDMask(PermMask.Val) ||
3178 X86::isPSHUFHWMask(PermMask.Val) ||
3179 X86::isPSHUFLWMask(PermMask.Val)) {
3180 if (V2.getOpcode() != ISD::UNDEF)
3181 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3182 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3186 if (X86::isSHUFPMask(PermMask.Val))
3189 // Handle v8i16 shuffle high / low shuffle node pair.
3190 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3191 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3192 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3193 SmallVector<SDOperand, 8> MaskVec;
3194 for (unsigned i = 0; i != 4; ++i)
3195 MaskVec.push_back(PermMask.getOperand(i));
3196 for (unsigned i = 4; i != 8; ++i)
3197 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3198 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3199 &MaskVec[0], MaskVec.size());
3200 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3202 for (unsigned i = 0; i != 4; ++i)
3203 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3204 for (unsigned i = 4; i != 8; ++i)
3205 MaskVec.push_back(PermMask.getOperand(i));
3206 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3207 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3210 // Floating point cases in the other order.
3211 if (X86::isSHUFPMask(PermMask.Val))
3213 if (X86::isPSHUFDMask(PermMask.Val) ||
3214 X86::isPSHUFHWMask(PermMask.Val) ||
3215 X86::isPSHUFLWMask(PermMask.Val)) {
3216 if (V2.getOpcode() != ISD::UNDEF)
3217 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3218 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3223 if (NumElems == 4) {
3224 MVT::ValueType MaskVT = PermMask.getValueType();
3225 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
3226 SmallVector<std::pair<int, int>, 8> Locs;
3227 Locs.reserve(NumElems);
3228 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3229 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3232 // If no more than two elements come from either vector. This can be
3233 // implemented with two shuffles. First shuffle gather the elements.
3234 // The second shuffle, which takes the first shuffle as both of its
3235 // vector operands, put the elements into the right order.
3236 for (unsigned i = 0; i != NumElems; ++i) {
3237 SDOperand Elt = PermMask.getOperand(i);
3238 if (Elt.getOpcode() == ISD::UNDEF) {
3239 Locs[i] = std::make_pair(-1, -1);
3241 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3242 if (Val < NumElems) {
3243 Locs[i] = std::make_pair(0, NumLo);
3247 Locs[i] = std::make_pair(1, NumHi);
3248 if (2+NumHi < NumElems)
3249 Mask1[2+NumHi] = Elt;
3254 if (NumLo <= 2 && NumHi <= 2) {
3255 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3256 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3257 &Mask1[0], Mask1.size()));
3258 for (unsigned i = 0; i != NumElems; ++i) {
3259 if (Locs[i].first == -1)
3262 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3263 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3264 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3268 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3269 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3270 &Mask2[0], Mask2.size()));
3273 // Break it into (shuffle shuffle_hi, shuffle_lo).
3275 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3276 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3277 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3278 unsigned MaskIdx = 0;
3280 unsigned HiIdx = NumElems/2;
3281 for (unsigned i = 0; i != NumElems; ++i) {
3282 if (i == NumElems/2) {
3288 SDOperand Elt = PermMask.getOperand(i);
3289 if (Elt.getOpcode() == ISD::UNDEF) {
3290 Locs[i] = std::make_pair(-1, -1);
3291 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3292 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3293 (*MaskPtr)[LoIdx] = Elt;
3296 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3297 (*MaskPtr)[HiIdx] = Elt;
3302 SDOperand LoShuffle =
3303 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3304 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3305 &LoMask[0], LoMask.size()));
3306 SDOperand HiShuffle =
3307 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3308 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3309 &HiMask[0], HiMask.size()));
3310 SmallVector<SDOperand, 8> MaskOps;
3311 for (unsigned i = 0; i != NumElems; ++i) {
3312 if (Locs[i].first == -1) {
3313 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3315 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3316 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3319 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3320 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3321 &MaskOps[0], MaskOps.size()));
3328 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3329 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3332 MVT::ValueType VT = Op.getValueType();
3333 // TODO: handle v16i8.
3334 if (MVT::getSizeInBits(VT) == 16) {
3335 // Transform it so it match pextrw which produces a 32-bit result.
3336 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3337 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3338 Op.getOperand(0), Op.getOperand(1));
3339 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3340 DAG.getValueType(VT));
3341 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3342 } else if (MVT::getSizeInBits(VT) == 32) {
3343 SDOperand Vec = Op.getOperand(0);
3344 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3347 // SHUFPS the element to the lowest double word, then movss.
3348 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3349 SmallVector<SDOperand, 8> IdxVec;
3350 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
3351 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3352 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3353 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3354 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3355 &IdxVec[0], IdxVec.size());
3356 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3357 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3358 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3359 DAG.getConstant(0, getPointerTy()));
3360 } else if (MVT::getSizeInBits(VT) == 64) {
3361 SDOperand Vec = Op.getOperand(0);
3362 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3366 // UNPCKHPD the element to the lowest double word, then movsd.
3367 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3368 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3369 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3370 SmallVector<SDOperand, 8> IdxVec;
3371 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
3372 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3373 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3374 &IdxVec[0], IdxVec.size());
3375 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3376 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3377 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3378 DAG.getConstant(0, getPointerTy()));
3385 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3386 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3387 // as its second argument.
3388 MVT::ValueType VT = Op.getValueType();
3389 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
3390 SDOperand N0 = Op.getOperand(0);
3391 SDOperand N1 = Op.getOperand(1);
3392 SDOperand N2 = Op.getOperand(2);
3393 if (MVT::getSizeInBits(BaseVT) == 16) {
3394 if (N1.getValueType() != MVT::i32)
3395 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3396 if (N2.getValueType() != MVT::i32)
3397 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
3398 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3399 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3400 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3403 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3404 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3405 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3406 SmallVector<SDOperand, 8> MaskVec;
3407 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3408 for (unsigned i = 1; i <= 3; ++i)
3409 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3410 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3411 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3412 &MaskVec[0], MaskVec.size()));
3414 // Use two pinsrw instructions to insert a 32 bit value.
3416 if (MVT::isFloatingPoint(N1.getValueType())) {
3417 if (ISD::isNON_EXTLoad(N1.Val)) {
3418 // Just load directly from f32mem to GR32.
3419 LoadSDNode *LD = cast<LoadSDNode>(N1);
3420 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
3421 LD->getSrcValue(), LD->getSrcValueOffset());
3423 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3424 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3425 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3426 DAG.getConstant(0, getPointerTy()));
3429 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3430 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3431 DAG.getConstant(Idx, getPointerTy()));
3432 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3433 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3434 DAG.getConstant(Idx+1, getPointerTy()));
3435 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3443 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3444 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3445 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3448 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3449 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3450 // one of the above mentioned nodes. It has to be wrapped because otherwise
3451 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3452 // be used to form addressing mode. These wrapped nodes will be selected
3455 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3456 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3457 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3459 CP->getAlignment());
3460 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3461 // With PIC, the address is actually $g + Offset.
3462 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3463 !Subtarget->isPICStyleRIPRel()) {
3464 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3465 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3473 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3474 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3475 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3476 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3477 // With PIC, the address is actually $g + Offset.
3478 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3479 !Subtarget->isPICStyleRIPRel()) {
3480 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3481 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3485 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3486 // load the value at address GV, not the value of GV itself. This means that
3487 // the GlobalAddress must be in the base or index register of the address, not
3488 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3489 // The same applies for external symbols during PIC codegen
3490 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3491 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3497 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3498 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3499 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3500 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3501 // With PIC, the address is actually $g + Offset.
3502 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3503 !Subtarget->isPICStyleRIPRel()) {
3504 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3505 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3512 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3513 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3514 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3515 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3516 // With PIC, the address is actually $g + Offset.
3517 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3518 !Subtarget->isPICStyleRIPRel()) {
3519 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3520 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3527 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3528 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3529 "Not an i64 shift!");
3530 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3531 SDOperand ShOpLo = Op.getOperand(0);
3532 SDOperand ShOpHi = Op.getOperand(1);
3533 SDOperand ShAmt = Op.getOperand(2);
3534 SDOperand Tmp1 = isSRA ?
3535 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3536 DAG.getConstant(0, MVT::i32);
3538 SDOperand Tmp2, Tmp3;
3539 if (Op.getOpcode() == ISD::SHL_PARTS) {
3540 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3541 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3543 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3544 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3547 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3548 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3549 DAG.getConstant(32, MVT::i8));
3550 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3551 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3554 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3556 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3557 SmallVector<SDOperand, 4> Ops;
3558 if (Op.getOpcode() == ISD::SHL_PARTS) {
3559 Ops.push_back(Tmp2);
3560 Ops.push_back(Tmp3);
3562 Ops.push_back(InFlag);
3563 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3564 InFlag = Hi.getValue(1);
3567 Ops.push_back(Tmp3);
3568 Ops.push_back(Tmp1);
3570 Ops.push_back(InFlag);
3571 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3573 Ops.push_back(Tmp2);
3574 Ops.push_back(Tmp3);
3576 Ops.push_back(InFlag);
3577 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3578 InFlag = Lo.getValue(1);
3581 Ops.push_back(Tmp3);
3582 Ops.push_back(Tmp1);
3584 Ops.push_back(InFlag);
3585 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3588 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3592 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3595 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3596 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3597 Op.getOperand(0).getValueType() >= MVT::i16 &&
3598 "Unknown SINT_TO_FP to lower!");
3601 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3602 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3603 MachineFunction &MF = DAG.getMachineFunction();
3604 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3605 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3606 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3607 StackSlot, NULL, 0);
3612 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3614 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3615 SmallVector<SDOperand, 8> Ops;
3616 Ops.push_back(Chain);
3617 Ops.push_back(StackSlot);
3618 Ops.push_back(DAG.getValueType(SrcVT));
3619 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3620 Tys, &Ops[0], Ops.size());
3623 Chain = Result.getValue(1);
3624 SDOperand InFlag = Result.getValue(2);
3626 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3627 // shouldn't be necessary except that RFP cannot be live across
3628 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3629 MachineFunction &MF = DAG.getMachineFunction();
3630 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3631 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3632 Tys = DAG.getVTList(MVT::Other);
3633 SmallVector<SDOperand, 8> Ops;
3634 Ops.push_back(Chain);
3635 Ops.push_back(Result);
3636 Ops.push_back(StackSlot);
3637 Ops.push_back(DAG.getValueType(Op.getValueType()));
3638 Ops.push_back(InFlag);
3639 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3640 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3646 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3647 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3648 "Unknown FP_TO_SINT to lower!");
3649 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3651 MachineFunction &MF = DAG.getMachineFunction();
3652 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3653 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3654 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3657 switch (Op.getValueType()) {
3658 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3659 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3660 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3661 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3664 SDOperand Chain = DAG.getEntryNode();
3665 SDOperand Value = Op.getOperand(0);
3667 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3668 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3669 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3671 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3673 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3674 Chain = Value.getValue(1);
3675 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3676 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3679 // Build the FP_TO_INT*_IN_MEM
3680 SDOperand Ops[] = { Chain, Value, StackSlot };
3681 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3684 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3687 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3688 MVT::ValueType VT = Op.getValueType();
3689 const Type *OpNTy = MVT::getTypeForValueType(VT);
3690 std::vector<Constant*> CV;
3691 if (VT == MVT::f64) {
3692 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3693 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3695 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3696 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3697 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3698 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3700 Constant *CS = ConstantStruct::get(CV);
3701 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3702 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3703 SmallVector<SDOperand, 3> Ops;
3704 Ops.push_back(DAG.getEntryNode());
3705 Ops.push_back(CPIdx);
3706 Ops.push_back(DAG.getSrcValue(NULL));
3707 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3708 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3711 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3712 MVT::ValueType VT = Op.getValueType();
3713 const Type *OpNTy = MVT::getTypeForValueType(VT);
3714 std::vector<Constant*> CV;
3715 if (VT == MVT::f64) {
3716 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3717 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3719 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3720 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3721 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3722 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3724 Constant *CS = ConstantStruct::get(CV);
3725 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3726 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3727 SmallVector<SDOperand, 3> Ops;
3728 Ops.push_back(DAG.getEntryNode());
3729 Ops.push_back(CPIdx);
3730 Ops.push_back(DAG.getSrcValue(NULL));
3731 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3732 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3735 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3736 SDOperand Op0 = Op.getOperand(0);
3737 SDOperand Op1 = Op.getOperand(1);
3738 MVT::ValueType VT = Op.getValueType();
3739 MVT::ValueType SrcVT = Op1.getValueType();
3740 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3742 // If second operand is smaller, extend it first.
3743 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3744 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3748 // First get the sign bit of second operand.
3749 std::vector<Constant*> CV;
3750 if (SrcVT == MVT::f64) {
3751 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3752 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3754 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3755 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3756 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3757 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3759 Constant *CS = ConstantStruct::get(CV);
3760 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3761 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3762 SmallVector<SDOperand, 3> Ops;
3763 Ops.push_back(DAG.getEntryNode());
3764 Ops.push_back(CPIdx);
3765 Ops.push_back(DAG.getSrcValue(NULL));
3766 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3767 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3769 // Shift sign bit right or left if the two operands have different types.
3770 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3771 // Op0 is MVT::f32, Op1 is MVT::f64.
3772 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3773 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3774 DAG.getConstant(32, MVT::i32));
3775 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3776 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3777 DAG.getConstant(0, getPointerTy()));
3780 // Clear first operand sign bit.
3782 if (VT == MVT::f64) {
3783 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3784 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3786 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3787 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3788 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3789 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3791 CS = ConstantStruct::get(CV);
3792 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3793 Tys = DAG.getVTList(VT, MVT::Other);
3795 Ops.push_back(DAG.getEntryNode());
3796 Ops.push_back(CPIdx);
3797 Ops.push_back(DAG.getSrcValue(NULL));
3798 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3799 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3801 // Or the value with the sign bit.
3802 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3805 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3807 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3809 SDOperand Op0 = Op.getOperand(0);
3810 SDOperand Op1 = Op.getOperand(1);
3811 SDOperand CC = Op.getOperand(2);
3812 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3813 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3814 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3815 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3818 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3820 SDOperand Ops1[] = { Chain, Op0, Op1 };
3821 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3822 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3823 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3826 assert(isFP && "Illegal integer SetCC!");
3828 SDOperand COps[] = { Chain, Op0, Op1 };
3829 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3831 switch (SetCCOpcode) {
3832 default: assert(false && "Illegal floating point SetCC!");
3833 case ISD::SETOEQ: { // !PF & ZF
3834 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3835 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3836 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3838 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3839 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3841 case ISD::SETUNE: { // PF | !ZF
3842 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3843 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3844 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3846 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3847 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3852 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3853 bool addTest = true;
3854 SDOperand Chain = DAG.getEntryNode();
3855 SDOperand Cond = Op.getOperand(0);
3857 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3859 if (Cond.getOpcode() == ISD::SETCC)
3860 Cond = LowerSETCC(Cond, DAG, Chain);
3862 if (Cond.getOpcode() == X86ISD::SETCC) {
3863 CC = Cond.getOperand(0);
3865 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3866 // (since flag operand cannot be shared). Use it as the condition setting
3867 // operand in place of the X86ISD::SETCC.
3868 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3869 // to use a test instead of duplicating the X86ISD::CMP (for register
3870 // pressure reason)?
3871 SDOperand Cmp = Cond.getOperand(1);
3872 unsigned Opc = Cmp.getOpcode();
3873 bool IllegalFPCMov = !X86ScalarSSE &&
3874 MVT::isFloatingPoint(Op.getValueType()) &&
3875 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3876 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3878 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3879 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3885 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3886 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3887 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3890 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3891 SmallVector<SDOperand, 4> Ops;
3892 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3893 // condition is true.
3894 Ops.push_back(Op.getOperand(2));
3895 Ops.push_back(Op.getOperand(1));
3897 Ops.push_back(Cond.getValue(1));
3898 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3901 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3902 bool addTest = true;
3903 SDOperand Chain = Op.getOperand(0);
3904 SDOperand Cond = Op.getOperand(1);
3905 SDOperand Dest = Op.getOperand(2);
3907 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3909 if (Cond.getOpcode() == ISD::SETCC)
3910 Cond = LowerSETCC(Cond, DAG, Chain);
3912 if (Cond.getOpcode() == X86ISD::SETCC) {
3913 CC = Cond.getOperand(0);
3915 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3916 // (since flag operand cannot be shared). Use it as the condition setting
3917 // operand in place of the X86ISD::SETCC.
3918 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3919 // to use a test instead of duplicating the X86ISD::CMP (for register
3920 // pressure reason)?
3921 SDOperand Cmp = Cond.getOperand(1);
3922 unsigned Opc = Cmp.getOpcode();
3923 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3924 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3925 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3931 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3932 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3933 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3935 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3936 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3939 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3940 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3942 if (Subtarget->is64Bit())
3943 return LowerX86_64CCCCallTo(Op, DAG);
3945 switch (CallingConv) {
3947 assert(0 && "Unsupported calling convention");
3948 case CallingConv::Fast:
3950 return LowerFastCCCallTo(Op, DAG);
3953 case CallingConv::C:
3954 return LowerCCCCallTo(Op, DAG);
3955 case CallingConv::X86_StdCall:
3956 return LowerCCCCallTo(Op, DAG, true);
3957 case CallingConv::X86_FastCall:
3958 return LowerFastCCCallTo(Op, DAG, true);
3962 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
3963 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
3965 // Support up returning up to two registers.
3966 MVT::ValueType VTs[2];
3967 unsigned DestRegs[2];
3968 unsigned NumRegs = Op.getNumOperands() / 2;
3969 assert(NumRegs <= 2 && "Can only return up to two regs!");
3971 for (unsigned i = 0; i != NumRegs; ++i)
3972 VTs[i] = Op.getOperand(i*2+1).getValueType();
3974 // Determine which register each value should be copied into.
3975 GetRetValueLocs(VTs, NumRegs, DestRegs, Subtarget,
3976 DAG.getMachineFunction().getFunction()->getCallingConv());
3978 // If this is the first return lowered for this function, add the regs to the
3979 // liveout set for the function.
3980 if (DAG.getMachineFunction().liveout_empty()) {
3981 for (unsigned i = 0; i != NumRegs; ++i)
3982 DAG.getMachineFunction().addLiveOut(DestRegs[i]);
3985 SDOperand Chain = Op.getOperand(0);
3988 // Copy the result values into the output registers.
3989 if (NumRegs != 1 || DestRegs[0] != X86::ST0) {
3990 for (unsigned i = 0; i != NumRegs; ++i) {
3991 Chain = DAG.getCopyToReg(Chain, DestRegs[i], Op.getOperand(i*2+1), Flag);
3992 Flag = Chain.getValue(1);
3995 // We need to handle a destination of ST0 specially, because it isn't really
3997 SDOperand Value = Op.getOperand(1);
3999 // If this is an FP return with ScalarSSE, we need to move the value from
4000 // an XMM register onto the fp-stack.
4004 // If this is a load into a scalarsse value, don't store the loaded value
4005 // back to the stack, only to reload it: just replace the scalar-sse load.
4006 if (ISD::isNON_EXTLoad(Value.Val) &&
4007 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
4008 Chain = Value.getOperand(0);
4009 MemLoc = Value.getOperand(1);
4011 // Spill the value to memory and reload it into top of stack.
4012 unsigned Size = MVT::getSizeInBits(VTs[0])/8;
4013 MachineFunction &MF = DAG.getMachineFunction();
4014 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4015 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
4016 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
4018 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
4019 SDOperand Ops[] = { Chain, MemLoc, DAG.getValueType(VTs[0]) };
4020 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4021 Chain = Value.getValue(1);
4024 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4025 SDOperand Ops[] = { Chain, Value };
4026 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
4027 Flag = Chain.getValue(1);
4030 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
4032 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
4034 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
4038 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
4039 MachineFunction &MF = DAG.getMachineFunction();
4040 const Function* Fn = MF.getFunction();
4041 if (Fn->hasExternalLinkage() &&
4042 Subtarget->isTargetCygMing() &&
4043 Fn->getName() == "main")
4044 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
4046 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4047 if (Subtarget->is64Bit())
4048 return LowerX86_64CCCArguments(Op, DAG);
4052 assert(0 && "Unsupported calling convention");
4053 case CallingConv::Fast:
4055 return LowerFastCCArguments(Op, DAG);
4058 case CallingConv::C:
4059 return LowerCCCArguments(Op, DAG);
4060 case CallingConv::X86_StdCall:
4061 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
4062 return LowerCCCArguments(Op, DAG, true);
4063 case CallingConv::X86_FastCall:
4064 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
4065 return LowerFastCCArguments(Op, DAG, true);
4069 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4070 SDOperand InFlag(0, 0);
4071 SDOperand Chain = Op.getOperand(0);
4073 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4074 if (Align == 0) Align = 1;
4076 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4077 // If not DWORD aligned, call memset if size is less than the threshold.
4078 // It knows how to align to the right boundary first.
4079 if ((Align & 3) != 0 ||
4080 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4081 MVT::ValueType IntPtr = getPointerTy();
4082 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4083 TargetLowering::ArgListTy Args;
4084 TargetLowering::ArgListEntry Entry;
4085 Entry.Node = Op.getOperand(1);
4086 Entry.Ty = IntPtrTy;
4087 Entry.isSigned = false;
4088 Entry.isInReg = false;
4089 Entry.isSRet = false;
4090 Args.push_back(Entry);
4091 // Extend the unsigned i8 argument to be an int value for the call.
4092 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4093 Entry.Ty = IntPtrTy;
4094 Entry.isSigned = false;
4095 Entry.isInReg = false;
4096 Entry.isSRet = false;
4097 Args.push_back(Entry);
4098 Entry.Node = Op.getOperand(3);
4099 Args.push_back(Entry);
4100 std::pair<SDOperand,SDOperand> CallResult =
4101 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4102 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4103 return CallResult.second;
4108 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4109 unsigned BytesLeft = 0;
4110 bool TwoRepStos = false;
4113 uint64_t Val = ValC->getValue() & 255;
4115 // If the value is a constant, then we can potentially use larger sets.
4116 switch (Align & 3) {
4117 case 2: // WORD aligned
4120 Val = (Val << 8) | Val;
4122 case 0: // DWORD aligned
4125 Val = (Val << 8) | Val;
4126 Val = (Val << 16) | Val;
4127 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4130 Val = (Val << 32) | Val;
4133 default: // Byte aligned
4136 Count = Op.getOperand(3);
4140 if (AVT > MVT::i8) {
4142 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4143 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4144 BytesLeft = I->getValue() % UBytes;
4146 assert(AVT >= MVT::i32 &&
4147 "Do not use rep;stos if not at least DWORD aligned");
4148 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4149 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4154 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4156 InFlag = Chain.getValue(1);
4159 Count = Op.getOperand(3);
4160 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4161 InFlag = Chain.getValue(1);
4164 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4166 InFlag = Chain.getValue(1);
4167 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4168 Op.getOperand(1), InFlag);
4169 InFlag = Chain.getValue(1);
4171 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4172 SmallVector<SDOperand, 8> Ops;
4173 Ops.push_back(Chain);
4174 Ops.push_back(DAG.getValueType(AVT));
4175 Ops.push_back(InFlag);
4176 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4179 InFlag = Chain.getValue(1);
4180 Count = Op.getOperand(3);
4181 MVT::ValueType CVT = Count.getValueType();
4182 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4183 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4184 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4186 InFlag = Chain.getValue(1);
4187 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4189 Ops.push_back(Chain);
4190 Ops.push_back(DAG.getValueType(MVT::i8));
4191 Ops.push_back(InFlag);
4192 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4193 } else if (BytesLeft) {
4194 // Issue stores for the last 1 - 7 bytes.
4196 unsigned Val = ValC->getValue() & 255;
4197 unsigned Offset = I->getValue() - BytesLeft;
4198 SDOperand DstAddr = Op.getOperand(1);
4199 MVT::ValueType AddrVT = DstAddr.getValueType();
4200 if (BytesLeft >= 4) {
4201 Val = (Val << 8) | Val;
4202 Val = (Val << 16) | Val;
4203 Value = DAG.getConstant(Val, MVT::i32);
4204 Chain = DAG.getStore(Chain, Value,
4205 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4206 DAG.getConstant(Offset, AddrVT)),
4211 if (BytesLeft >= 2) {
4212 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4213 Chain = DAG.getStore(Chain, Value,
4214 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4215 DAG.getConstant(Offset, AddrVT)),
4220 if (BytesLeft == 1) {
4221 Value = DAG.getConstant(Val, MVT::i8);
4222 Chain = DAG.getStore(Chain, Value,
4223 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4224 DAG.getConstant(Offset, AddrVT)),
4232 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
4233 SDOperand Chain = Op.getOperand(0);
4235 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4236 if (Align == 0) Align = 1;
4238 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4239 // If not DWORD aligned, call memcpy if size is less than the threshold.
4240 // It knows how to align to the right boundary first.
4241 if ((Align & 3) != 0 ||
4242 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4243 MVT::ValueType IntPtr = getPointerTy();
4244 TargetLowering::ArgListTy Args;
4245 TargetLowering::ArgListEntry Entry;
4246 Entry.Ty = getTargetData()->getIntPtrType();
4247 Entry.isSigned = false;
4248 Entry.isInReg = false;
4249 Entry.isSRet = false;
4250 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
4251 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
4252 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
4253 std::pair<SDOperand,SDOperand> CallResult =
4254 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4255 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
4256 return CallResult.second;
4261 unsigned BytesLeft = 0;
4262 bool TwoRepMovs = false;
4263 switch (Align & 3) {
4264 case 2: // WORD aligned
4267 case 0: // DWORD aligned
4269 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4272 default: // Byte aligned
4274 Count = Op.getOperand(3);
4278 if (AVT > MVT::i8) {
4280 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4281 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4282 BytesLeft = I->getValue() % UBytes;
4284 assert(AVT >= MVT::i32 &&
4285 "Do not use rep;movs if not at least DWORD aligned");
4286 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4287 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4292 SDOperand InFlag(0, 0);
4293 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4295 InFlag = Chain.getValue(1);
4296 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4297 Op.getOperand(1), InFlag);
4298 InFlag = Chain.getValue(1);
4299 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4300 Op.getOperand(2), InFlag);
4301 InFlag = Chain.getValue(1);
4303 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4304 SmallVector<SDOperand, 8> Ops;
4305 Ops.push_back(Chain);
4306 Ops.push_back(DAG.getValueType(AVT));
4307 Ops.push_back(InFlag);
4308 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4311 InFlag = Chain.getValue(1);
4312 Count = Op.getOperand(3);
4313 MVT::ValueType CVT = Count.getValueType();
4314 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4315 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4316 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4318 InFlag = Chain.getValue(1);
4319 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4321 Ops.push_back(Chain);
4322 Ops.push_back(DAG.getValueType(MVT::i8));
4323 Ops.push_back(InFlag);
4324 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4325 } else if (BytesLeft) {
4326 // Issue loads and stores for the last 1 - 7 bytes.
4327 unsigned Offset = I->getValue() - BytesLeft;
4328 SDOperand DstAddr = Op.getOperand(1);
4329 MVT::ValueType DstVT = DstAddr.getValueType();
4330 SDOperand SrcAddr = Op.getOperand(2);
4331 MVT::ValueType SrcVT = SrcAddr.getValueType();
4333 if (BytesLeft >= 4) {
4334 Value = DAG.getLoad(MVT::i32, Chain,
4335 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4336 DAG.getConstant(Offset, SrcVT)),
4338 Chain = Value.getValue(1);
4339 Chain = DAG.getStore(Chain, Value,
4340 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4341 DAG.getConstant(Offset, DstVT)),
4346 if (BytesLeft >= 2) {
4347 Value = DAG.getLoad(MVT::i16, Chain,
4348 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4349 DAG.getConstant(Offset, SrcVT)),
4351 Chain = Value.getValue(1);
4352 Chain = DAG.getStore(Chain, Value,
4353 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4354 DAG.getConstant(Offset, DstVT)),
4360 if (BytesLeft == 1) {
4361 Value = DAG.getLoad(MVT::i8, Chain,
4362 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4363 DAG.getConstant(Offset, SrcVT)),
4365 Chain = Value.getValue(1);
4366 Chain = DAG.getStore(Chain, Value,
4367 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4368 DAG.getConstant(Offset, DstVT)),
4377 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4378 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4379 SDOperand TheOp = Op.getOperand(0);
4380 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4381 if (Subtarget->is64Bit()) {
4382 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4383 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4384 MVT::i64, Copy1.getValue(2));
4385 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4386 DAG.getConstant(32, MVT::i8));
4388 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4391 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4392 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4395 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4396 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4397 MVT::i32, Copy1.getValue(2));
4398 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4399 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4400 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4403 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4404 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4406 if (!Subtarget->is64Bit()) {
4407 // vastart just stores the address of the VarArgsFrameIndex slot into the
4408 // memory location argument.
4409 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4410 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4415 // gp_offset (0 - 6 * 8)
4416 // fp_offset (48 - 48 + 8 * 16)
4417 // overflow_arg_area (point to parameters coming in memory).
4419 SmallVector<SDOperand, 8> MemOps;
4420 SDOperand FIN = Op.getOperand(1);
4422 SDOperand Store = DAG.getStore(Op.getOperand(0),
4423 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4424 FIN, SV->getValue(), SV->getOffset());
4425 MemOps.push_back(Store);
4428 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4429 DAG.getConstant(4, getPointerTy()));
4430 Store = DAG.getStore(Op.getOperand(0),
4431 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4432 FIN, SV->getValue(), SV->getOffset());
4433 MemOps.push_back(Store);
4435 // Store ptr to overflow_arg_area
4436 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4437 DAG.getConstant(4, getPointerTy()));
4438 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4439 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4441 MemOps.push_back(Store);
4443 // Store ptr to reg_save_area.
4444 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4445 DAG.getConstant(8, getPointerTy()));
4446 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4447 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4449 MemOps.push_back(Store);
4450 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4454 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4455 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4457 default: return SDOperand(); // Don't custom lower most intrinsics.
4458 // Comparison intrinsics.
4459 case Intrinsic::x86_sse_comieq_ss:
4460 case Intrinsic::x86_sse_comilt_ss:
4461 case Intrinsic::x86_sse_comile_ss:
4462 case Intrinsic::x86_sse_comigt_ss:
4463 case Intrinsic::x86_sse_comige_ss:
4464 case Intrinsic::x86_sse_comineq_ss:
4465 case Intrinsic::x86_sse_ucomieq_ss:
4466 case Intrinsic::x86_sse_ucomilt_ss:
4467 case Intrinsic::x86_sse_ucomile_ss:
4468 case Intrinsic::x86_sse_ucomigt_ss:
4469 case Intrinsic::x86_sse_ucomige_ss:
4470 case Intrinsic::x86_sse_ucomineq_ss:
4471 case Intrinsic::x86_sse2_comieq_sd:
4472 case Intrinsic::x86_sse2_comilt_sd:
4473 case Intrinsic::x86_sse2_comile_sd:
4474 case Intrinsic::x86_sse2_comigt_sd:
4475 case Intrinsic::x86_sse2_comige_sd:
4476 case Intrinsic::x86_sse2_comineq_sd:
4477 case Intrinsic::x86_sse2_ucomieq_sd:
4478 case Intrinsic::x86_sse2_ucomilt_sd:
4479 case Intrinsic::x86_sse2_ucomile_sd:
4480 case Intrinsic::x86_sse2_ucomigt_sd:
4481 case Intrinsic::x86_sse2_ucomige_sd:
4482 case Intrinsic::x86_sse2_ucomineq_sd: {
4484 ISD::CondCode CC = ISD::SETCC_INVALID;
4487 case Intrinsic::x86_sse_comieq_ss:
4488 case Intrinsic::x86_sse2_comieq_sd:
4492 case Intrinsic::x86_sse_comilt_ss:
4493 case Intrinsic::x86_sse2_comilt_sd:
4497 case Intrinsic::x86_sse_comile_ss:
4498 case Intrinsic::x86_sse2_comile_sd:
4502 case Intrinsic::x86_sse_comigt_ss:
4503 case Intrinsic::x86_sse2_comigt_sd:
4507 case Intrinsic::x86_sse_comige_ss:
4508 case Intrinsic::x86_sse2_comige_sd:
4512 case Intrinsic::x86_sse_comineq_ss:
4513 case Intrinsic::x86_sse2_comineq_sd:
4517 case Intrinsic::x86_sse_ucomieq_ss:
4518 case Intrinsic::x86_sse2_ucomieq_sd:
4519 Opc = X86ISD::UCOMI;
4522 case Intrinsic::x86_sse_ucomilt_ss:
4523 case Intrinsic::x86_sse2_ucomilt_sd:
4524 Opc = X86ISD::UCOMI;
4527 case Intrinsic::x86_sse_ucomile_ss:
4528 case Intrinsic::x86_sse2_ucomile_sd:
4529 Opc = X86ISD::UCOMI;
4532 case Intrinsic::x86_sse_ucomigt_ss:
4533 case Intrinsic::x86_sse2_ucomigt_sd:
4534 Opc = X86ISD::UCOMI;
4537 case Intrinsic::x86_sse_ucomige_ss:
4538 case Intrinsic::x86_sse2_ucomige_sd:
4539 Opc = X86ISD::UCOMI;
4542 case Intrinsic::x86_sse_ucomineq_ss:
4543 case Intrinsic::x86_sse2_ucomineq_sd:
4544 Opc = X86ISD::UCOMI;
4550 SDOperand LHS = Op.getOperand(1);
4551 SDOperand RHS = Op.getOperand(2);
4552 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4554 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4555 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4556 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4557 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4558 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4559 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4560 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4565 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4566 // Depths > 0 not supported yet!
4567 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4570 // Just load the return address
4571 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4572 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4575 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4576 // Depths > 0 not supported yet!
4577 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4580 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4581 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4582 DAG.getConstant(4, getPointerTy()));
4585 /// LowerOperation - Provide custom lowering hooks for some operations.
4587 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4588 switch (Op.getOpcode()) {
4589 default: assert(0 && "Should not custom lower this!");
4590 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4591 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4592 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4593 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4594 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4595 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4596 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4597 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4598 case ISD::SHL_PARTS:
4599 case ISD::SRA_PARTS:
4600 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4601 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4602 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4603 case ISD::FABS: return LowerFABS(Op, DAG);
4604 case ISD::FNEG: return LowerFNEG(Op, DAG);
4605 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4606 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4607 case ISD::SELECT: return LowerSELECT(Op, DAG);
4608 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4609 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4610 case ISD::CALL: return LowerCALL(Op, DAG);
4611 case ISD::RET: return LowerRET(Op, DAG);
4612 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4613 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4614 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4615 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4616 case ISD::VASTART: return LowerVASTART(Op, DAG);
4617 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4618 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4619 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4624 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4626 default: return NULL;
4627 case X86ISD::SHLD: return "X86ISD::SHLD";
4628 case X86ISD::SHRD: return "X86ISD::SHRD";
4629 case X86ISD::FAND: return "X86ISD::FAND";
4630 case X86ISD::FOR: return "X86ISD::FOR";
4631 case X86ISD::FXOR: return "X86ISD::FXOR";
4632 case X86ISD::FSRL: return "X86ISD::FSRL";
4633 case X86ISD::FILD: return "X86ISD::FILD";
4634 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4635 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4636 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4637 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4638 case X86ISD::FLD: return "X86ISD::FLD";
4639 case X86ISD::FST: return "X86ISD::FST";
4640 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4641 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4642 case X86ISD::CALL: return "X86ISD::CALL";
4643 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4644 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4645 case X86ISD::CMP: return "X86ISD::CMP";
4646 case X86ISD::COMI: return "X86ISD::COMI";
4647 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4648 case X86ISD::SETCC: return "X86ISD::SETCC";
4649 case X86ISD::CMOV: return "X86ISD::CMOV";
4650 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4651 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4652 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4653 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4654 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4655 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4656 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4657 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4658 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4659 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4660 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4661 case X86ISD::FMAX: return "X86ISD::FMAX";
4662 case X86ISD::FMIN: return "X86ISD::FMIN";
4666 /// isLegalAddressImmediate - Return true if the integer value or
4667 /// GlobalValue can be used as the offset of the target addressing mode.
4668 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
4669 // X86 allows a sign-extended 32-bit immediate field.
4670 return (V > -(1LL << 32) && V < (1LL << 32)-1);
4673 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
4674 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
4675 // field unless we are in small code model.
4676 if (Subtarget->is64Bit() &&
4677 getTargetMachine().getCodeModel() != CodeModel::Small)
4680 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4683 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4684 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4685 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4686 /// are assumed to be legal.
4688 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4689 // Only do shuffles on 128-bit vector types for now.
4690 if (MVT::getSizeInBits(VT) == 64) return false;
4691 return (Mask.Val->getNumOperands() <= 4 ||
4692 isSplatMask(Mask.Val) ||
4693 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4694 X86::isUNPCKLMask(Mask.Val) ||
4695 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4696 X86::isUNPCKHMask(Mask.Val));
4699 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4701 SelectionDAG &DAG) const {
4702 unsigned NumElts = BVOps.size();
4703 // Only do shuffles on 128-bit vector types for now.
4704 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4705 if (NumElts == 2) return true;
4707 return (isMOVLMask(&BVOps[0], 4) ||
4708 isCommutedMOVL(&BVOps[0], 4, true) ||
4709 isSHUFPMask(&BVOps[0], 4) ||
4710 isCommutedSHUFP(&BVOps[0], 4));
4715 //===----------------------------------------------------------------------===//
4716 // X86 Scheduler Hooks
4717 //===----------------------------------------------------------------------===//
4720 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4721 MachineBasicBlock *BB) {
4722 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4723 switch (MI->getOpcode()) {
4724 default: assert(false && "Unexpected instr type to insert");
4725 case X86::CMOV_FR32:
4726 case X86::CMOV_FR64:
4727 case X86::CMOV_V4F32:
4728 case X86::CMOV_V2F64:
4729 case X86::CMOV_V2I64: {
4730 // To "insert" a SELECT_CC instruction, we actually have to insert the
4731 // diamond control-flow pattern. The incoming instruction knows the
4732 // destination vreg to set, the condition code register to branch on, the
4733 // true/false values to select between, and a branch opcode to use.
4734 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4735 ilist<MachineBasicBlock>::iterator It = BB;
4741 // cmpTY ccX, r1, r2
4743 // fallthrough --> copy0MBB
4744 MachineBasicBlock *thisMBB = BB;
4745 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4746 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4748 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4749 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4750 MachineFunction *F = BB->getParent();
4751 F->getBasicBlockList().insert(It, copy0MBB);
4752 F->getBasicBlockList().insert(It, sinkMBB);
4753 // Update machine-CFG edges by first adding all successors of the current
4754 // block to the new block which will contain the Phi node for the select.
4755 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4756 e = BB->succ_end(); i != e; ++i)
4757 sinkMBB->addSuccessor(*i);
4758 // Next, remove all successors of the current block, and add the true
4759 // and fallthrough blocks as its successors.
4760 while(!BB->succ_empty())
4761 BB->removeSuccessor(BB->succ_begin());
4762 BB->addSuccessor(copy0MBB);
4763 BB->addSuccessor(sinkMBB);
4766 // %FalseValue = ...
4767 // # fallthrough to sinkMBB
4770 // Update machine-CFG edges
4771 BB->addSuccessor(sinkMBB);
4774 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4777 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4778 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4779 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4781 delete MI; // The pseudo instruction is gone now.
4785 case X86::FP_TO_INT16_IN_MEM:
4786 case X86::FP_TO_INT32_IN_MEM:
4787 case X86::FP_TO_INT64_IN_MEM: {
4788 // Change the floating point control register to use "round towards zero"
4789 // mode when truncating to an integer value.
4790 MachineFunction *F = BB->getParent();
4791 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4792 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4794 // Load the old value of the high byte of the control word...
4796 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4797 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4799 // Set the high part to be round to zero...
4800 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4803 // Reload the modified control word now...
4804 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4806 // Restore the memory image of control word to original value
4807 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4810 // Get the X86 opcode to use.
4812 switch (MI->getOpcode()) {
4813 default: assert(0 && "illegal opcode!");
4814 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4815 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4816 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4820 MachineOperand &Op = MI->getOperand(0);
4821 if (Op.isRegister()) {
4822 AM.BaseType = X86AddressMode::RegBase;
4823 AM.Base.Reg = Op.getReg();
4825 AM.BaseType = X86AddressMode::FrameIndexBase;
4826 AM.Base.FrameIndex = Op.getFrameIndex();
4828 Op = MI->getOperand(1);
4829 if (Op.isImmediate())
4830 AM.Scale = Op.getImm();
4831 Op = MI->getOperand(2);
4832 if (Op.isImmediate())
4833 AM.IndexReg = Op.getImm();
4834 Op = MI->getOperand(3);
4835 if (Op.isGlobalAddress()) {
4836 AM.GV = Op.getGlobal();
4838 AM.Disp = Op.getImm();
4840 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4841 .addReg(MI->getOperand(4).getReg());
4843 // Reload the original control word now.
4844 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4846 delete MI; // The pseudo instruction is gone now.
4852 //===----------------------------------------------------------------------===//
4853 // X86 Optimization Hooks
4854 //===----------------------------------------------------------------------===//
4856 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4858 uint64_t &KnownZero,
4860 unsigned Depth) const {
4861 unsigned Opc = Op.getOpcode();
4862 assert((Opc >= ISD::BUILTIN_OP_END ||
4863 Opc == ISD::INTRINSIC_WO_CHAIN ||
4864 Opc == ISD::INTRINSIC_W_CHAIN ||
4865 Opc == ISD::INTRINSIC_VOID) &&
4866 "Should use MaskedValueIsZero if you don't know whether Op"
4867 " is a target node!");
4869 KnownZero = KnownOne = 0; // Don't know anything.
4873 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4878 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4879 /// element of the result of the vector shuffle.
4880 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4881 MVT::ValueType VT = N->getValueType(0);
4882 SDOperand PermMask = N->getOperand(2);
4883 unsigned NumElems = PermMask.getNumOperands();
4884 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4886 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4888 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4889 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4890 SDOperand Idx = PermMask.getOperand(i);
4891 if (Idx.getOpcode() == ISD::UNDEF)
4892 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4893 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4898 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4899 /// node is a GlobalAddress + an offset.
4900 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4901 unsigned Opc = N->getOpcode();
4902 if (Opc == X86ISD::Wrapper) {
4903 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4904 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4907 } else if (Opc == ISD::ADD) {
4908 SDOperand N1 = N->getOperand(0);
4909 SDOperand N2 = N->getOperand(1);
4910 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4911 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4913 Offset += V->getSignExtended();
4916 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4917 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4919 Offset += V->getSignExtended();
4927 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4929 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4930 MachineFrameInfo *MFI) {
4931 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4934 SDOperand Loc = N->getOperand(1);
4935 SDOperand BaseLoc = Base->getOperand(1);
4936 if (Loc.getOpcode() == ISD::FrameIndex) {
4937 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4939 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4940 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4941 int FS = MFI->getObjectSize(FI);
4942 int BFS = MFI->getObjectSize(BFI);
4943 if (FS != BFS || FS != Size) return false;
4944 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4946 GlobalValue *GV1 = NULL;
4947 GlobalValue *GV2 = NULL;
4948 int64_t Offset1 = 0;
4949 int64_t Offset2 = 0;
4950 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4951 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4952 if (isGA1 && isGA2 && GV1 == GV2)
4953 return Offset1 == (Offset2 + Dist*Size);
4959 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4960 const X86Subtarget *Subtarget) {
4963 if (isGAPlusOffset(Base, GV, Offset))
4964 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4966 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4967 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4969 // Fixed objects do not specify alignment, however the offsets are known.
4970 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4971 (MFI->getObjectOffset(BFI) % 16) == 0);
4973 return MFI->getObjectAlignment(BFI) >= 16;
4979 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4980 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4981 /// if the load addresses are consecutive, non-overlapping, and in the right
4983 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4984 const X86Subtarget *Subtarget) {
4985 MachineFunction &MF = DAG.getMachineFunction();
4986 MachineFrameInfo *MFI = MF.getFrameInfo();
4987 MVT::ValueType VT = N->getValueType(0);
4988 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4989 SDOperand PermMask = N->getOperand(2);
4990 int NumElems = (int)PermMask.getNumOperands();
4991 SDNode *Base = NULL;
4992 for (int i = 0; i < NumElems; ++i) {
4993 SDOperand Idx = PermMask.getOperand(i);
4994 if (Idx.getOpcode() == ISD::UNDEF) {
4995 if (!Base) return SDOperand();
4998 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4999 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5003 else if (!isConsecutiveLoad(Arg.Val, Base,
5004 i, MVT::getSizeInBits(EVT)/8,MFI))
5009 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5011 LoadSDNode *LD = cast<LoadSDNode>(Base);
5012 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5013 LD->getSrcValueOffset());
5015 // Just use movups, it's shorter.
5016 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
5017 SmallVector<SDOperand, 3> Ops;
5018 Ops.push_back(Base->getOperand(0));
5019 Ops.push_back(Base->getOperand(1));
5020 Ops.push_back(Base->getOperand(2));
5021 return DAG.getNode(ISD::BIT_CONVERT, VT,
5022 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
5026 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5027 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5028 const X86Subtarget *Subtarget) {
5029 SDOperand Cond = N->getOperand(0);
5031 // If we have SSE[12] support, try to form min/max nodes.
5032 if (Subtarget->hasSSE2() &&
5033 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5034 if (Cond.getOpcode() == ISD::SETCC) {
5035 // Get the LHS/RHS of the select.
5036 SDOperand LHS = N->getOperand(1);
5037 SDOperand RHS = N->getOperand(2);
5038 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5040 unsigned Opcode = 0;
5041 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5044 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5047 if (!UnsafeFPMath) break;
5049 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5051 Opcode = X86ISD::FMIN;
5054 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5057 if (!UnsafeFPMath) break;
5059 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5061 Opcode = X86ISD::FMAX;
5064 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5067 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5070 if (!UnsafeFPMath) break;
5072 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5074 Opcode = X86ISD::FMIN;
5077 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5080 if (!UnsafeFPMath) break;
5082 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5084 Opcode = X86ISD::FMAX;
5090 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5099 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5100 DAGCombinerInfo &DCI) const {
5101 SelectionDAG &DAG = DCI.DAG;
5102 switch (N->getOpcode()) {
5104 case ISD::VECTOR_SHUFFLE:
5105 return PerformShuffleCombine(N, DAG, Subtarget);
5107 return PerformSELECTCombine(N, DAG, Subtarget);
5113 //===----------------------------------------------------------------------===//
5114 // X86 Inline Assembly Support
5115 //===----------------------------------------------------------------------===//
5117 /// getConstraintType - Given a constraint letter, return the type of
5118 /// constraint it is for this target.
5119 X86TargetLowering::ConstraintType
5120 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
5121 switch (ConstraintLetter) {
5130 return C_RegisterClass;
5131 default: return TargetLowering::getConstraintType(ConstraintLetter);
5135 /// isOperandValidForConstraint - Return the specified operand (possibly
5136 /// modified) if the specified SDOperand is valid for the specified target
5137 /// constraint letter, otherwise return null.
5138 SDOperand X86TargetLowering::
5139 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
5140 switch (Constraint) {
5143 // Literal immediates are always ok.
5144 if (isa<ConstantSDNode>(Op)) return Op;
5146 // If we are in non-pic codegen mode, we allow the address of a global to
5147 // be used with 'i'.
5148 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
5149 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
5150 return SDOperand(0, 0);
5152 if (GA->getOpcode() != ISD::TargetGlobalAddress)
5153 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5158 // Otherwise, not valid for this mode.
5159 return SDOperand(0, 0);
5161 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
5165 std::vector<unsigned> X86TargetLowering::
5166 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5167 MVT::ValueType VT) const {
5168 if (Constraint.size() == 1) {
5169 // FIXME: not handling fp-stack yet!
5170 // FIXME: not handling MMX registers yet ('y' constraint).
5171 switch (Constraint[0]) { // GCC X86 Constraint Letters
5172 default: break; // Unknown constraint letter
5173 case 'A': // EAX/EDX
5174 if (VT == MVT::i32 || VT == MVT::i64)
5175 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5177 case 'r': // GENERAL_REGS
5178 case 'R': // LEGACY_REGS
5179 if (VT == MVT::i64 && Subtarget->is64Bit())
5180 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
5181 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
5182 X86::R8, X86::R9, X86::R10, X86::R11,
5183 X86::R12, X86::R13, X86::R14, X86::R15, 0);
5185 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5186 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
5187 else if (VT == MVT::i16)
5188 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5189 X86::SI, X86::DI, X86::BP, X86::SP, 0);
5190 else if (VT == MVT::i8)
5191 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5193 case 'l': // INDEX_REGS
5195 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5196 X86::ESI, X86::EDI, X86::EBP, 0);
5197 else if (VT == MVT::i16)
5198 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5199 X86::SI, X86::DI, X86::BP, 0);
5200 else if (VT == MVT::i8)
5201 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5203 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5206 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5207 else if (VT == MVT::i16)
5208 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5209 else if (VT == MVT::i8)
5210 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5212 case 'x': // SSE_REGS if SSE1 allowed
5213 if (Subtarget->hasSSE1())
5214 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5215 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5217 return std::vector<unsigned>();
5218 case 'Y': // SSE_REGS if SSE2 allowed
5219 if (Subtarget->hasSSE2())
5220 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5221 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5223 return std::vector<unsigned>();
5227 return std::vector<unsigned>();
5230 std::pair<unsigned, const TargetRegisterClass*>
5231 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5232 MVT::ValueType VT) const {
5233 // Use the default implementation in TargetLowering to convert the register
5234 // constraint into a member of a register class.
5235 std::pair<unsigned, const TargetRegisterClass*> Res;
5236 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5238 // Not found as a standard register?
5239 if (Res.second == 0) {
5240 // GCC calls "st(0)" just plain "st".
5241 if (StringsEqualNoCase("{st}", Constraint)) {
5242 Res.first = X86::ST0;
5243 Res.second = X86::RSTRegisterClass;
5249 // Otherwise, check to see if this is a register class of the wrong value
5250 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5251 // turn into {ax},{dx}.
5252 if (Res.second->hasType(VT))
5253 return Res; // Correct type already, nothing to do.
5255 // All of the single-register GCC register classes map their values onto
5256 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5257 // really want an 8-bit or 32-bit register, map to the appropriate register
5258 // class and return the appropriate register.
5259 if (Res.second != X86::GR16RegisterClass)
5262 if (VT == MVT::i8) {
5263 unsigned DestReg = 0;
5264 switch (Res.first) {
5266 case X86::AX: DestReg = X86::AL; break;
5267 case X86::DX: DestReg = X86::DL; break;
5268 case X86::CX: DestReg = X86::CL; break;
5269 case X86::BX: DestReg = X86::BL; break;
5272 Res.first = DestReg;
5273 Res.second = Res.second = X86::GR8RegisterClass;
5275 } else if (VT == MVT::i32) {
5276 unsigned DestReg = 0;
5277 switch (Res.first) {
5279 case X86::AX: DestReg = X86::EAX; break;
5280 case X86::DX: DestReg = X86::EDX; break;
5281 case X86::CX: DestReg = X86::ECX; break;
5282 case X86::BX: DestReg = X86::EBX; break;
5283 case X86::SI: DestReg = X86::ESI; break;
5284 case X86::DI: DestReg = X86::EDI; break;
5285 case X86::BP: DestReg = X86::EBP; break;
5286 case X86::SP: DestReg = X86::ESP; break;
5289 Res.first = DestReg;
5290 Res.second = Res.second = X86::GR32RegisterClass;
5292 } else if (VT == MVT::i64) {
5293 unsigned DestReg = 0;
5294 switch (Res.first) {
5296 case X86::AX: DestReg = X86::RAX; break;
5297 case X86::DX: DestReg = X86::RDX; break;
5298 case X86::CX: DestReg = X86::RCX; break;
5299 case X86::BX: DestReg = X86::RBX; break;
5300 case X86::SI: DestReg = X86::RSI; break;
5301 case X86::DI: DestReg = X86::RDI; break;
5302 case X86::BP: DestReg = X86::RBP; break;
5303 case X86::SP: DestReg = X86::RSP; break;
5306 Res.first = DestReg;
5307 Res.second = Res.second = X86::GR64RegisterClass;