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];
451 case MVT::i8: Reg = X86::AL; break;
452 case MVT::i16: Reg = X86::AX; break;
453 case MVT::i32: Reg = X86::EAX; break;
454 case MVT::i64: Reg = X86::RAX; break;
457 if (Subtarget->is64Bit())
458 Reg = X86::XMM0; // FP values in X86-64 go in XMM0.
460 Reg = X86::ST0; // FP values in X86-32 go in ST0.
463 assert(MVT::isVector(ArgVT) && "Unknown return value type!");
464 Reg = X86::XMM0; // Int/FP vector result -> XMM0.
470 /// LowerCallResult - Lower the result values of an ISD::CALL into the
471 /// appropriate copies out of appropriate physical registers. This assumes that
472 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
473 /// being lowered. The returns a SDNode with the same number of values as the
475 SDNode *X86TargetLowering::
476 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
477 unsigned CallingConv, SelectionDAG &DAG) {
478 SmallVector<SDOperand, 8> ResultVals;
480 // We support returning up to two registers.
481 MVT::ValueType VTs[2];
482 unsigned DestRegs[2];
483 unsigned NumRegs = TheCall->getNumValues() - 1;
484 assert(NumRegs <= 2 && "Can only return up to two regs!");
486 for (unsigned i = 0; i != NumRegs; ++i)
487 VTs[i] = TheCall->getValueType(i);
489 // Determine which register each value should be copied into.
490 GetRetValueLocs(VTs, NumRegs, DestRegs, Subtarget, CallingConv);
492 // Copy all of the result registers out of their specified physreg.
493 if (NumRegs != 1 || DestRegs[0] != X86::ST0) {
494 for (unsigned i = 0; i != NumRegs; ++i) {
495 Chain = DAG.getCopyFromReg(Chain, DestRegs[i], VTs[i],
497 InFlag = Chain.getValue(2);
498 ResultVals.push_back(Chain.getValue(0));
501 // Copies from the FP stack are special, as ST0 isn't a valid register
502 // before the fp stackifier runs.
504 // Copy ST0 into an RFP register with FP_GET_RESULT.
505 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
506 SDOperand GROps[] = { Chain, InFlag };
507 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
508 Chain = RetVal.getValue(1);
509 InFlag = RetVal.getValue(2);
511 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
514 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
515 // shouldn't be necessary except that RFP cannot be live across
516 // multiple blocks. When stackifier is fixed, they can be uncoupled.
517 MachineFunction &MF = DAG.getMachineFunction();
518 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
519 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
521 Chain, RetVal, StackSlot, DAG.getValueType(VTs[0]), InFlag
523 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
524 RetVal = DAG.getLoad(VTs[0], Chain, StackSlot, NULL, 0);
525 Chain = RetVal.getValue(1);
528 if (VTs[0] == MVT::f32 && !X86ScalarSSE)
529 // FIXME: we would really like to remember that this FP_ROUND
530 // operation is okay to eliminate if we allow excess FP precision.
531 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
532 ResultVals.push_back(RetVal);
535 // Merge everything together with a MERGE_VALUES node.
536 ResultVals.push_back(Chain);
537 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
538 &ResultVals[0], ResultVals.size()).Val;
542 //===----------------------------------------------------------------------===//
543 // C & StdCall Calling Convention implementation
544 //===----------------------------------------------------------------------===//
545 // StdCall calling convention seems to be standard for many Windows' API
546 // routines and around. It differs from C calling convention just a little:
547 // callee should clean up the stack, not caller. Symbols should be also
548 // decorated in some fancy way :) It doesn't support any vector arguments.
550 /// AddLiveIn - This helper function adds the specified physical register to the
551 /// MachineFunction as a live in value. It also creates a corresponding virtual
553 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
554 const TargetRegisterClass *RC) {
555 assert(RC->contains(PReg) && "Not the correct regclass!");
556 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
557 MF.addLiveIn(PReg, VReg);
561 /// HowToPassArgument - Returns how an formal argument of the specified type
562 /// should be passed. If it is through stack, returns the size of the stack
563 /// slot; if it is through integer or XMM register, returns the number of
564 /// integer or XMM registers are needed.
566 HowToPassCallArgument(MVT::ValueType ObjectVT,
568 unsigned NumIntRegs, unsigned NumXMMRegs,
569 unsigned MaxNumIntRegs,
570 unsigned &ObjSize, unsigned &ObjIntRegs,
571 unsigned &ObjXMMRegs,
572 bool AllowVectors = true) {
577 if (MaxNumIntRegs>3) {
578 // We don't have too much registers on ia32! :)
583 default: assert(0 && "Unhandled argument type!");
585 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
591 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
597 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
603 if (ArgInReg && (NumIntRegs+2 <= MaxNumIntRegs)) {
605 } else if (ArgInReg && (NumIntRegs+1 <= MaxNumIntRegs)) {
629 assert(0 && "Unhandled argument type [vector]!");
633 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
635 unsigned NumArgs = Op.Val->getNumValues() - 1;
636 MachineFunction &MF = DAG.getMachineFunction();
637 MachineFrameInfo *MFI = MF.getFrameInfo();
638 SDOperand Root = Op.getOperand(0);
639 SmallVector<SDOperand, 8> ArgValues;
640 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
642 // Add DAG nodes to load the arguments... On entry to a function on the X86,
643 // the stack frame looks like this:
645 // [ESP] -- return address
646 // [ESP + 4] -- first argument (leftmost lexically)
647 // [ESP + 8] -- second argument, if first argument is <= 4 bytes in size
650 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
651 unsigned NumSRetBytes= 0; // How much bytes on stack used for struct return
652 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
653 unsigned NumIntRegs = 0; // Integer regs used for parameter passing
655 static const unsigned XMMArgRegs[] = {
656 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
658 static const unsigned GPRArgRegs[][3] = {
659 { X86::AL, X86::DL, X86::CL },
660 { X86::AX, X86::DX, X86::CX },
661 { X86::EAX, X86::EDX, X86::ECX }
663 static const TargetRegisterClass* GPRClasses[3] = {
664 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
667 // Handle regparm attribute
668 SmallVector<bool, 8> ArgInRegs(NumArgs, false);
669 SmallVector<bool, 8> SRetArgs(NumArgs, false);
671 for (unsigned i = 0; i<NumArgs; ++i) {
672 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3+i))->getValue();
673 ArgInRegs[i] = (Flags >> 1) & 1;
674 SRetArgs[i] = (Flags >> 2) & 1;
678 for (unsigned i = 0; i < NumArgs; ++i) {
679 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
680 unsigned ArgIncrement = 4;
681 unsigned ObjSize = 0;
682 unsigned ObjXMMRegs = 0;
683 unsigned ObjIntRegs = 0;
687 HowToPassCallArgument(ObjectVT,
689 NumIntRegs, NumXMMRegs, 3,
690 ObjSize, ObjIntRegs, ObjXMMRegs,
694 ArgIncrement = ObjSize;
696 if (ObjIntRegs || ObjXMMRegs) {
698 default: assert(0 && "Unhandled argument type!");
702 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][NumIntRegs];
703 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
704 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
713 assert(!isStdCall && "Unhandled argument type!");
714 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
715 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
718 NumIntRegs += ObjIntRegs;
719 NumXMMRegs += ObjXMMRegs;
722 // XMM arguments have to be aligned on 16-byte boundary.
724 ArgOffset = ((ArgOffset + 15) / 16) * 16;
725 // Create the SelectionDAG nodes corresponding to a load from this
727 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
728 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
729 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
731 ArgOffset += ArgIncrement; // Move on to the next argument.
733 NumSRetBytes += ArgIncrement;
736 ArgValues.push_back(ArgValue);
739 ArgValues.push_back(Root);
741 // If the function takes variable number of arguments, make a frame index for
742 // the start of the first vararg value... for expansion of llvm.va_start.
744 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
746 if (isStdCall && !isVarArg) {
747 BytesToPopOnReturn = ArgOffset; // Callee pops everything..
748 BytesCallerReserves = 0;
750 BytesToPopOnReturn = NumSRetBytes; // Callee pops hidden struct pointer.
751 BytesCallerReserves = ArgOffset;
754 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
755 ReturnAddrIndex = 0; // No return address slot generated yet.
758 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
760 // Return the new list of results.
761 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
762 &ArgValues[0], ArgValues.size());
765 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
767 SDOperand Chain = Op.getOperand(0);
768 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
769 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
770 SDOperand Callee = Op.getOperand(4);
771 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
773 static const unsigned XMMArgRegs[] = {
774 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
776 static const unsigned GPR32ArgRegs[] = {
777 X86::EAX, X86::EDX, X86::ECX
780 // Count how many bytes are to be pushed on the stack.
781 unsigned NumBytes = 0;
782 // Keep track of the number of integer regs passed so far.
783 unsigned NumIntRegs = 0;
784 // Keep track of the number of XMM regs passed so far.
785 unsigned NumXMMRegs = 0;
786 // How much bytes on stack used for struct return
787 unsigned NumSRetBytes= 0;
789 // Handle regparm attribute
790 SmallVector<bool, 8> ArgInRegs(NumOps, false);
791 SmallVector<bool, 8> SRetArgs(NumOps, false);
792 for (unsigned i = 0; i<NumOps; ++i) {
794 dyn_cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
795 ArgInRegs[i] = (Flags >> 1) & 1;
796 SRetArgs[i] = (Flags >> 2) & 1;
799 // Calculate stack frame size
800 for (unsigned i = 0; i != NumOps; ++i) {
801 SDOperand Arg = Op.getOperand(5+2*i);
802 unsigned ArgIncrement = 4;
803 unsigned ObjSize = 0;
804 unsigned ObjIntRegs = 0;
805 unsigned ObjXMMRegs = 0;
807 HowToPassCallArgument(Arg.getValueType(),
809 NumIntRegs, NumXMMRegs, 3,
810 ObjSize, ObjIntRegs, ObjXMMRegs,
811 CC != CallingConv::X86_StdCall);
813 ArgIncrement = ObjSize;
815 NumIntRegs += ObjIntRegs;
816 NumXMMRegs += ObjXMMRegs;
818 // XMM arguments have to be aligned on 16-byte boundary.
820 NumBytes = ((NumBytes + 15) / 16) * 16;
821 NumBytes += ArgIncrement;
825 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
827 // Arguments go on the stack in reverse order, as specified by the ABI.
828 unsigned ArgOffset = 0;
831 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
832 SmallVector<SDOperand, 8> MemOpChains;
833 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
834 for (unsigned i = 0; i != NumOps; ++i) {
835 SDOperand Arg = Op.getOperand(5+2*i);
836 unsigned ArgIncrement = 4;
837 unsigned ObjSize = 0;
838 unsigned ObjIntRegs = 0;
839 unsigned ObjXMMRegs = 0;
841 HowToPassCallArgument(Arg.getValueType(),
843 NumIntRegs, NumXMMRegs, 3,
844 ObjSize, ObjIntRegs, ObjXMMRegs,
845 CC != CallingConv::X86_StdCall);
848 ArgIncrement = ObjSize;
850 if (Arg.getValueType() == MVT::i8 || Arg.getValueType() == MVT::i16) {
851 // Promote the integer to 32 bits. If the input type is signed use a
852 // sign extend, otherwise use a zero extend.
853 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
855 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
856 Arg = DAG.getNode(ExtOp, MVT::i32, Arg);
859 if (ObjIntRegs || ObjXMMRegs) {
860 switch (Arg.getValueType()) {
861 default: assert(0 && "Unhandled argument type!");
863 RegsToPass.push_back(std::make_pair(GPR32ArgRegs[NumIntRegs], Arg));
871 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
875 NumIntRegs += ObjIntRegs;
876 NumXMMRegs += ObjXMMRegs;
879 // XMM arguments have to be aligned on 16-byte boundary.
881 ArgOffset = ((ArgOffset + 15) / 16) * 16;
883 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
884 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
885 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
887 ArgOffset += ArgIncrement; // Move on to the next argument.
889 NumSRetBytes += ArgIncrement;
893 // Sanity check: we haven't seen NumSRetBytes > 4
894 assert((NumSRetBytes<=4) &&
895 "Too much space for struct-return pointer requested");
897 if (!MemOpChains.empty())
898 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
899 &MemOpChains[0], MemOpChains.size());
901 // Build a sequence of copy-to-reg nodes chained together with token chain
902 // and flag operands which copy the outgoing args into registers.
904 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
905 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
907 InFlag = Chain.getValue(1);
910 // ELF / PIC requires GOT in the EBX register before function calls via PLT
912 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
913 Subtarget->isPICStyleGOT()) {
914 Chain = DAG.getCopyToReg(Chain, X86::EBX,
915 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
917 InFlag = Chain.getValue(1);
920 // If the callee is a GlobalAddress node (quite common, every direct call is)
921 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
922 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
923 // We should use extra load for direct calls to dllimported functions in
925 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
926 getTargetMachine(), true))
927 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
928 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
929 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
931 // Returns a chain & a flag for retval copy to use.
932 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
933 SmallVector<SDOperand, 8> Ops;
934 Ops.push_back(Chain);
935 Ops.push_back(Callee);
937 // Add argument registers to the end of the list so that they are known live
939 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
940 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
941 RegsToPass[i].second.getValueType()));
943 // Add an implicit use GOT pointer in EBX.
944 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
945 Subtarget->isPICStyleGOT())
946 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
949 Ops.push_back(InFlag);
951 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
952 NodeTys, &Ops[0], Ops.size());
953 InFlag = Chain.getValue(1);
955 // Create the CALLSEQ_END node.
956 unsigned NumBytesForCalleeToPush = 0;
958 if (CC == CallingConv::X86_StdCall) {
960 NumBytesForCalleeToPush = NumSRetBytes;
962 NumBytesForCalleeToPush = NumBytes;
964 // If this is is a call to a struct-return function, the callee
965 // pops the hidden struct pointer, so we have to push it back.
966 // This is common for Darwin/X86, Linux & Mingw32 targets.
967 NumBytesForCalleeToPush = NumSRetBytes;
970 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
972 Ops.push_back(Chain);
973 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
974 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
975 Ops.push_back(InFlag);
976 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
977 InFlag = Chain.getValue(1);
979 // Handle result values, copying them out of physregs into vregs that we
981 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
985 //===----------------------------------------------------------------------===//
986 // X86-64 C Calling Convention implementation
987 //===----------------------------------------------------------------------===//
989 /// HowToPassX86_64CCCArgument - Returns how an formal argument of the specified
990 /// type should be passed. If it is through stack, returns the size of the stack
991 /// slot; if it is through integer or XMM register, returns the number of
992 /// integer or XMM registers are needed.
994 HowToPassX86_64CCCArgument(MVT::ValueType ObjectVT,
995 unsigned NumIntRegs, unsigned NumXMMRegs,
996 unsigned &ObjSize, unsigned &ObjIntRegs,
997 unsigned &ObjXMMRegs) {
1003 default: assert(0 && "Unhandled argument type!");
1013 case MVT::i8: ObjSize = 1; break;
1014 case MVT::i16: ObjSize = 2; break;
1015 case MVT::i32: ObjSize = 4; break;
1016 case MVT::i64: ObjSize = 8; break;
1033 case MVT::f32: ObjSize = 4; break;
1034 case MVT::f64: ObjSize = 8; break;
1040 case MVT::v2f64: ObjSize = 16; break;
1048 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1049 unsigned NumArgs = Op.Val->getNumValues() - 1;
1050 MachineFunction &MF = DAG.getMachineFunction();
1051 MachineFrameInfo *MFI = MF.getFrameInfo();
1052 SDOperand Root = Op.getOperand(0);
1053 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1054 SmallVector<SDOperand, 8> ArgValues;
1056 // Add DAG nodes to load the arguments... On entry to a function on the X86,
1057 // the stack frame looks like this:
1059 // [RSP] -- return address
1060 // [RSP + 8] -- first nonreg argument (leftmost lexically)
1061 // [RSP +16] -- second nonreg argument, if 1st argument is <= 8 bytes in size
1064 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1065 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1066 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1068 static const unsigned GPR8ArgRegs[] = {
1069 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1071 static const unsigned GPR16ArgRegs[] = {
1072 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1074 static const unsigned GPR32ArgRegs[] = {
1075 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1077 static const unsigned GPR64ArgRegs[] = {
1078 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1080 static const unsigned XMMArgRegs[] = {
1081 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1082 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1085 for (unsigned i = 0; i < NumArgs; ++i) {
1086 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1087 unsigned ArgIncrement = 8;
1088 unsigned ObjSize = 0;
1089 unsigned ObjIntRegs = 0;
1090 unsigned ObjXMMRegs = 0;
1092 // FIXME: __int128 and long double support?
1093 HowToPassX86_64CCCArgument(ObjectVT, NumIntRegs, NumXMMRegs,
1094 ObjSize, ObjIntRegs, ObjXMMRegs);
1096 ArgIncrement = ObjSize;
1100 if (ObjIntRegs || ObjXMMRegs) {
1102 default: assert(0 && "Unhandled argument type!");
1107 TargetRegisterClass *RC = NULL;
1111 RC = X86::GR8RegisterClass;
1112 Reg = GPR8ArgRegs[NumIntRegs];
1115 RC = X86::GR16RegisterClass;
1116 Reg = GPR16ArgRegs[NumIntRegs];
1119 RC = X86::GR32RegisterClass;
1120 Reg = GPR32ArgRegs[NumIntRegs];
1123 RC = X86::GR64RegisterClass;
1124 Reg = GPR64ArgRegs[NumIntRegs];
1127 Reg = AddLiveIn(MF, Reg, RC);
1128 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1139 TargetRegisterClass *RC= (ObjectVT == MVT::f32) ?
1140 X86::FR32RegisterClass : ((ObjectVT == MVT::f64) ?
1141 X86::FR64RegisterClass : X86::VR128RegisterClass);
1142 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], RC);
1143 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1147 NumIntRegs += ObjIntRegs;
1148 NumXMMRegs += ObjXMMRegs;
1149 } else if (ObjSize) {
1150 // XMM arguments have to be aligned on 16-byte boundary.
1152 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1153 // Create the SelectionDAG nodes corresponding to a load from this
1155 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1156 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1157 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1158 ArgOffset += ArgIncrement; // Move on to the next argument.
1161 ArgValues.push_back(ArgValue);
1164 // If the function takes variable number of arguments, make a frame index for
1165 // the start of the first vararg value... for expansion of llvm.va_start.
1167 // For X86-64, if there are vararg parameters that are passed via
1168 // registers, then we must store them to their spots on the stack so they
1169 // may be loaded by deferencing the result of va_next.
1170 VarArgsGPOffset = NumIntRegs * 8;
1171 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1172 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
1173 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1175 // Store the integer parameter registers.
1176 SmallVector<SDOperand, 8> MemOps;
1177 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1178 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1179 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1180 for (; NumIntRegs != 6; ++NumIntRegs) {
1181 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1182 X86::GR64RegisterClass);
1183 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1184 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1185 MemOps.push_back(Store);
1186 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1187 DAG.getConstant(8, getPointerTy()));
1190 // Now store the XMM (fp + vector) parameter registers.
1191 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1192 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1193 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1194 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1195 X86::VR128RegisterClass);
1196 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1197 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1198 MemOps.push_back(Store);
1199 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1200 DAG.getConstant(16, getPointerTy()));
1202 if (!MemOps.empty())
1203 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1204 &MemOps[0], MemOps.size());
1207 ArgValues.push_back(Root);
1209 ReturnAddrIndex = 0; // No return address slot generated yet.
1210 BytesToPopOnReturn = 0; // Callee pops nothing.
1211 BytesCallerReserves = ArgOffset;
1213 // Return the new list of results.
1214 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1215 &ArgValues[0], ArgValues.size());
1219 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1221 SDOperand Chain = Op.getOperand(0);
1222 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1223 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1224 SDOperand Callee = Op.getOperand(4);
1225 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1227 // Count how many bytes are to be pushed on the stack.
1228 unsigned NumBytes = 0;
1229 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1230 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1232 static const unsigned GPR8ArgRegs[] = {
1233 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1235 static const unsigned GPR16ArgRegs[] = {
1236 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1238 static const unsigned GPR32ArgRegs[] = {
1239 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1241 static const unsigned GPR64ArgRegs[] = {
1242 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1244 static const unsigned XMMArgRegs[] = {
1245 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1246 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1249 for (unsigned i = 0; i != NumOps; ++i) {
1250 SDOperand Arg = Op.getOperand(5+2*i);
1251 MVT::ValueType ArgVT = Arg.getValueType();
1254 default: assert(0 && "Unknown value type!");
1274 else if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1277 // XMM arguments have to be aligned on 16-byte boundary.
1278 NumBytes = ((NumBytes + 15) / 16) * 16;
1285 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1287 // Arguments go on the stack in reverse order, as specified by the ABI.
1288 unsigned ArgOffset = 0;
1291 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1292 SmallVector<SDOperand, 8> MemOpChains;
1293 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1294 for (unsigned i = 0; i != NumOps; ++i) {
1295 SDOperand Arg = Op.getOperand(5+2*i);
1296 MVT::ValueType ArgVT = Arg.getValueType();
1299 default: assert(0 && "Unexpected ValueType for argument!");
1304 if (NumIntRegs < 6) {
1308 case MVT::i8: Reg = GPR8ArgRegs[NumIntRegs]; break;
1309 case MVT::i16: Reg = GPR16ArgRegs[NumIntRegs]; break;
1310 case MVT::i32: Reg = GPR32ArgRegs[NumIntRegs]; break;
1311 case MVT::i64: Reg = GPR64ArgRegs[NumIntRegs]; break;
1313 RegsToPass.push_back(std::make_pair(Reg, Arg));
1316 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1317 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1318 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1330 if (NumXMMRegs < 8) {
1331 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1334 if (ArgVT != MVT::f32 && ArgVT != MVT::f64) {
1335 // XMM arguments have to be aligned on 16-byte boundary.
1336 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1338 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1339 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1340 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1341 if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1349 if (!MemOpChains.empty())
1350 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1351 &MemOpChains[0], MemOpChains.size());
1353 // Build a sequence of copy-to-reg nodes chained together with token chain
1354 // and flag operands which copy the outgoing args into registers.
1356 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1357 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1359 InFlag = Chain.getValue(1);
1363 // From AMD64 ABI document:
1364 // For calls that may call functions that use varargs or stdargs
1365 // (prototype-less calls or calls to functions containing ellipsis (...) in
1366 // the declaration) %al is used as hidden argument to specify the number
1367 // of SSE registers used. The contents of %al do not need to match exactly
1368 // the number of registers, but must be an ubound on the number of SSE
1369 // registers used and is in the range 0 - 8 inclusive.
1370 Chain = DAG.getCopyToReg(Chain, X86::AL,
1371 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1372 InFlag = Chain.getValue(1);
1375 // If the callee is a GlobalAddress node (quite common, every direct call is)
1376 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1377 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1378 // We should use extra load for direct calls to dllimported functions in
1380 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1381 getTargetMachine(), true))
1382 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1383 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1384 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1386 // Returns a chain & a flag for retval copy to use.
1387 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1388 SmallVector<SDOperand, 8> Ops;
1389 Ops.push_back(Chain);
1390 Ops.push_back(Callee);
1392 // Add argument registers to the end of the list so that they are known live
1394 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1395 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1396 RegsToPass[i].second.getValueType()));
1399 Ops.push_back(InFlag);
1401 // FIXME: Do not generate X86ISD::TAILCALL for now.
1402 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1403 NodeTys, &Ops[0], Ops.size());
1404 InFlag = Chain.getValue(1);
1406 // Returns a flag for retval copy to use.
1407 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1409 Ops.push_back(Chain);
1410 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1411 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1412 Ops.push_back(InFlag);
1413 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1414 InFlag = Chain.getValue(1);
1416 // Handle result values, copying them out of physregs into vregs that we
1418 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1421 //===----------------------------------------------------------------------===//
1422 // Fast & FastCall Calling Convention implementation
1423 //===----------------------------------------------------------------------===//
1425 // The X86 'fast' calling convention passes up to two integer arguments in
1426 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
1427 // and requires that the callee pop its arguments off the stack (allowing proper
1428 // tail calls), and has the same return value conventions as C calling convs.
1430 // This calling convention always arranges for the callee pop value to be 8n+4
1431 // bytes, which is needed for tail recursion elimination and stack alignment
1434 // Note that this can be enhanced in the future to pass fp vals in registers
1435 // (when we have a global fp allocator) and do other tricks.
1437 //===----------------------------------------------------------------------===//
1438 // The X86 'fastcall' calling convention passes up to two integer arguments in
1439 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1440 // and requires that the callee pop its arguments off the stack (allowing proper
1441 // tail calls), and has the same return value conventions as C calling convs.
1443 // This calling convention always arranges for the callee pop value to be 8n+4
1444 // bytes, which is needed for tail recursion elimination and stack alignment
1449 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG,
1451 unsigned NumArgs = Op.Val->getNumValues()-1;
1452 MachineFunction &MF = DAG.getMachineFunction();
1453 MachineFrameInfo *MFI = MF.getFrameInfo();
1454 SDOperand Root = Op.getOperand(0);
1455 SmallVector<SDOperand, 8> ArgValues;
1457 // Add DAG nodes to load the arguments... On entry to a function the stack
1458 // frame looks like this:
1460 // [ESP] -- return address
1461 // [ESP + 4] -- first nonreg argument (leftmost lexically)
1462 // [ESP + 8] -- second nonreg argument, if 1st argument is <= 4 bytes in size
1464 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1466 // Keep track of the number of integer regs passed so far. This can be either
1467 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1469 unsigned NumIntRegs = 0;
1470 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1472 static const unsigned XMMArgRegs[] = {
1473 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1476 static const unsigned GPRArgRegs[][2][2] = {
1477 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1478 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1479 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1482 static const TargetRegisterClass* GPRClasses[3] = {
1483 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
1486 unsigned GPRInd = (isFastCall ? 1 : 0);
1487 for (unsigned i = 0; i < NumArgs; ++i) {
1488 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1489 unsigned ArgIncrement = 4;
1490 unsigned ObjSize = 0;
1491 unsigned ObjXMMRegs = 0;
1492 unsigned ObjIntRegs = 0;
1496 HowToPassCallArgument(ObjectVT,
1497 true, // Use as much registers as possible
1498 NumIntRegs, NumXMMRegs,
1499 (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS),
1500 ObjSize, ObjIntRegs, ObjXMMRegs,
1504 ArgIncrement = ObjSize;
1506 if (ObjIntRegs || ObjXMMRegs) {
1508 default: assert(0 && "Unhandled argument type!");
1512 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][GPRInd][NumIntRegs];
1513 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
1514 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1523 assert(!isFastCall && "Unhandled argument type!");
1524 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
1525 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1529 NumIntRegs += ObjIntRegs;
1530 NumXMMRegs += ObjXMMRegs;
1533 // XMM arguments have to be aligned on 16-byte boundary.
1535 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1536 // Create the SelectionDAG nodes corresponding to a load from this
1538 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1539 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1540 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1542 ArgOffset += ArgIncrement; // Move on to the next argument.
1545 ArgValues.push_back(ArgValue);
1548 ArgValues.push_back(Root);
1550 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1551 // arguments and the arguments after the retaddr has been pushed are aligned.
1552 if ((ArgOffset & 7) == 0)
1555 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1556 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1557 ReturnAddrIndex = 0; // No return address slot generated yet.
1558 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
1559 BytesCallerReserves = 0;
1561 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
1563 // Finally, inform the code generator which regs we return values in.
1564 switch (getValueType(MF.getFunction()->getReturnType())) {
1565 default: assert(0 && "Unknown type!");
1566 case MVT::isVoid: break;
1571 MF.addLiveOut(X86::EAX);
1574 MF.addLiveOut(X86::EAX);
1575 MF.addLiveOut(X86::EDX);
1579 MF.addLiveOut(X86::ST0);
1587 assert(!isFastCall && "Unknown result type");
1588 MF.addLiveOut(X86::XMM0);
1592 // Return the new list of results.
1593 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1594 &ArgValues[0], ArgValues.size());
1597 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1599 SDOperand Chain = Op.getOperand(0);
1600 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1601 SDOperand Callee = Op.getOperand(4);
1602 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1604 // Count how many bytes are to be pushed on the stack.
1605 unsigned NumBytes = 0;
1607 // Keep track of the number of integer regs passed so far. This can be either
1608 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1610 unsigned NumIntRegs = 0;
1611 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1613 static const unsigned GPRArgRegs[][2][2] = {
1614 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1615 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1616 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1618 static const unsigned XMMArgRegs[] = {
1619 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1622 bool isFastCall = CC == CallingConv::X86_FastCall;
1623 unsigned GPRInd = isFastCall ? 1 : 0;
1624 for (unsigned i = 0; i != NumOps; ++i) {
1625 SDOperand Arg = Op.getOperand(5+2*i);
1627 switch (Arg.getValueType()) {
1628 default: assert(0 && "Unknown value type!");
1632 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1633 if (NumIntRegs < MaxNumIntRegs) {
1650 assert(!isFastCall && "Unknown value type!");
1654 // XMM arguments have to be aligned on 16-byte boundary.
1655 NumBytes = ((NumBytes + 15) / 16) * 16;
1662 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1663 // arguments and the arguments after the retaddr has been pushed are aligned.
1664 if ((NumBytes & 7) == 0)
1667 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1669 // Arguments go on the stack in reverse order, as specified by the ABI.
1670 unsigned ArgOffset = 0;
1672 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1673 SmallVector<SDOperand, 8> MemOpChains;
1674 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1675 for (unsigned i = 0; i != NumOps; ++i) {
1676 SDOperand Arg = Op.getOperand(5+2*i);
1678 switch (Arg.getValueType()) {
1679 default: assert(0 && "Unexpected ValueType for argument!");
1683 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1684 if (NumIntRegs < MaxNumIntRegs) {
1686 GPRArgRegs[Arg.getValueType()-MVT::i8][GPRInd][NumIntRegs];
1687 RegsToPass.push_back(std::make_pair(RegToUse, Arg));
1693 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1694 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1695 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1700 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1701 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1702 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1712 assert(!isFastCall && "Unexpected ValueType for argument!");
1713 if (NumXMMRegs < 4) {
1714 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1717 // XMM arguments have to be aligned on 16-byte boundary.
1718 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1719 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1720 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1721 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1728 if (!MemOpChains.empty())
1729 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1730 &MemOpChains[0], MemOpChains.size());
1732 // Build a sequence of copy-to-reg nodes chained together with token chain
1733 // and flag operands which copy the outgoing args into registers.
1735 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1736 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1738 InFlag = Chain.getValue(1);
1741 // If the callee is a GlobalAddress node (quite common, every direct call is)
1742 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1743 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1744 // We should use extra load for direct calls to dllimported functions in
1746 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1747 getTargetMachine(), true))
1748 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1749 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1750 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1752 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1754 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1755 Subtarget->isPICStyleGOT()) {
1756 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1757 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1759 InFlag = Chain.getValue(1);
1762 // Returns a chain & a flag for retval copy to use.
1763 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1764 SmallVector<SDOperand, 8> Ops;
1765 Ops.push_back(Chain);
1766 Ops.push_back(Callee);
1768 // Add argument registers to the end of the list so that they are known live
1770 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1771 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1772 RegsToPass[i].second.getValueType()));
1774 // Add an implicit use GOT pointer in EBX.
1775 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1776 Subtarget->isPICStyleGOT())
1777 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1780 Ops.push_back(InFlag);
1782 // FIXME: Do not generate X86ISD::TAILCALL for now.
1783 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1784 NodeTys, &Ops[0], Ops.size());
1785 InFlag = Chain.getValue(1);
1787 // Returns a flag for retval copy to use.
1788 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1790 Ops.push_back(Chain);
1791 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1792 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1793 Ops.push_back(InFlag);
1794 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1795 InFlag = Chain.getValue(1);
1797 // Handle result values, copying them out of physregs into vregs that we
1799 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1802 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1803 if (ReturnAddrIndex == 0) {
1804 // Set up a frame object for the return address.
1805 MachineFunction &MF = DAG.getMachineFunction();
1806 if (Subtarget->is64Bit())
1807 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1809 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1812 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1817 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1818 /// specific condition code. It returns a false if it cannot do a direct
1819 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1821 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1822 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1823 SelectionDAG &DAG) {
1824 X86CC = X86::COND_INVALID;
1826 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1827 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1828 // X > -1 -> X == 0, jump !sign.
1829 RHS = DAG.getConstant(0, RHS.getValueType());
1830 X86CC = X86::COND_NS;
1832 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1833 // X < 0 -> X == 0, jump on sign.
1834 X86CC = X86::COND_S;
1839 switch (SetCCOpcode) {
1841 case ISD::SETEQ: X86CC = X86::COND_E; break;
1842 case ISD::SETGT: X86CC = X86::COND_G; break;
1843 case ISD::SETGE: X86CC = X86::COND_GE; break;
1844 case ISD::SETLT: X86CC = X86::COND_L; break;
1845 case ISD::SETLE: X86CC = X86::COND_LE; break;
1846 case ISD::SETNE: X86CC = X86::COND_NE; break;
1847 case ISD::SETULT: X86CC = X86::COND_B; break;
1848 case ISD::SETUGT: X86CC = X86::COND_A; break;
1849 case ISD::SETULE: X86CC = X86::COND_BE; break;
1850 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1853 // On a floating point condition, the flags are set as follows:
1855 // 0 | 0 | 0 | X > Y
1856 // 0 | 0 | 1 | X < Y
1857 // 1 | 0 | 0 | X == Y
1858 // 1 | 1 | 1 | unordered
1860 switch (SetCCOpcode) {
1863 case ISD::SETEQ: X86CC = X86::COND_E; break;
1864 case ISD::SETOLT: Flip = true; // Fallthrough
1866 case ISD::SETGT: X86CC = X86::COND_A; break;
1867 case ISD::SETOLE: Flip = true; // Fallthrough
1869 case ISD::SETGE: X86CC = X86::COND_AE; break;
1870 case ISD::SETUGT: Flip = true; // Fallthrough
1872 case ISD::SETLT: X86CC = X86::COND_B; break;
1873 case ISD::SETUGE: Flip = true; // Fallthrough
1875 case ISD::SETLE: X86CC = X86::COND_BE; break;
1877 case ISD::SETNE: X86CC = X86::COND_NE; break;
1878 case ISD::SETUO: X86CC = X86::COND_P; break;
1879 case ISD::SETO: X86CC = X86::COND_NP; break;
1882 std::swap(LHS, RHS);
1885 return X86CC != X86::COND_INVALID;
1888 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1889 /// code. Current x86 isa includes the following FP cmov instructions:
1890 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1891 static bool hasFPCMov(unsigned X86CC) {
1907 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1908 /// true if Op is undef or if its value falls within the specified range (L, H].
1909 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1910 if (Op.getOpcode() == ISD::UNDEF)
1913 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1914 return (Val >= Low && Val < Hi);
1917 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1918 /// true if Op is undef or if its value equal to the specified value.
1919 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1920 if (Op.getOpcode() == ISD::UNDEF)
1922 return cast<ConstantSDNode>(Op)->getValue() == Val;
1925 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1926 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1927 bool X86::isPSHUFDMask(SDNode *N) {
1928 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1930 if (N->getNumOperands() != 4)
1933 // Check if the value doesn't reference the second vector.
1934 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1935 SDOperand Arg = N->getOperand(i);
1936 if (Arg.getOpcode() == ISD::UNDEF) continue;
1937 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1938 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1945 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1946 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1947 bool X86::isPSHUFHWMask(SDNode *N) {
1948 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1950 if (N->getNumOperands() != 8)
1953 // Lower quadword copied in order.
1954 for (unsigned i = 0; i != 4; ++i) {
1955 SDOperand Arg = N->getOperand(i);
1956 if (Arg.getOpcode() == ISD::UNDEF) continue;
1957 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1958 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1962 // Upper quadword shuffled.
1963 for (unsigned i = 4; i != 8; ++i) {
1964 SDOperand Arg = N->getOperand(i);
1965 if (Arg.getOpcode() == ISD::UNDEF) continue;
1966 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1967 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1968 if (Val < 4 || Val > 7)
1975 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1976 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1977 bool X86::isPSHUFLWMask(SDNode *N) {
1978 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1980 if (N->getNumOperands() != 8)
1983 // Upper quadword copied in order.
1984 for (unsigned i = 4; i != 8; ++i)
1985 if (!isUndefOrEqual(N->getOperand(i), i))
1988 // Lower quadword shuffled.
1989 for (unsigned i = 0; i != 4; ++i)
1990 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1996 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1997 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1998 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1999 if (NumElems != 2 && NumElems != 4) return false;
2001 unsigned Half = NumElems / 2;
2002 for (unsigned i = 0; i < Half; ++i)
2003 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2005 for (unsigned i = Half; i < NumElems; ++i)
2006 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2012 bool X86::isSHUFPMask(SDNode *N) {
2013 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2014 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2017 /// isCommutedSHUFP - Returns true if the shuffle mask is except
2018 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2019 /// half elements to come from vector 1 (which would equal the dest.) and
2020 /// the upper half to come from vector 2.
2021 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2022 if (NumOps != 2 && NumOps != 4) return false;
2024 unsigned Half = NumOps / 2;
2025 for (unsigned i = 0; i < Half; ++i)
2026 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2028 for (unsigned i = Half; i < NumOps; ++i)
2029 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2034 static bool isCommutedSHUFP(SDNode *N) {
2035 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2036 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2039 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2040 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2041 bool X86::isMOVHLPSMask(SDNode *N) {
2042 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2044 if (N->getNumOperands() != 4)
2047 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2048 return isUndefOrEqual(N->getOperand(0), 6) &&
2049 isUndefOrEqual(N->getOperand(1), 7) &&
2050 isUndefOrEqual(N->getOperand(2), 2) &&
2051 isUndefOrEqual(N->getOperand(3), 3);
2054 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2055 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2057 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2058 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2060 if (N->getNumOperands() != 4)
2063 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2064 return isUndefOrEqual(N->getOperand(0), 2) &&
2065 isUndefOrEqual(N->getOperand(1), 3) &&
2066 isUndefOrEqual(N->getOperand(2), 2) &&
2067 isUndefOrEqual(N->getOperand(3), 3);
2070 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2071 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2072 bool X86::isMOVLPMask(SDNode *N) {
2073 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2075 unsigned NumElems = N->getNumOperands();
2076 if (NumElems != 2 && NumElems != 4)
2079 for (unsigned i = 0; i < NumElems/2; ++i)
2080 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2083 for (unsigned i = NumElems/2; i < NumElems; ++i)
2084 if (!isUndefOrEqual(N->getOperand(i), i))
2090 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2091 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2093 bool X86::isMOVHPMask(SDNode *N) {
2094 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2096 unsigned NumElems = N->getNumOperands();
2097 if (NumElems != 2 && NumElems != 4)
2100 for (unsigned i = 0; i < NumElems/2; ++i)
2101 if (!isUndefOrEqual(N->getOperand(i), i))
2104 for (unsigned i = 0; i < NumElems/2; ++i) {
2105 SDOperand Arg = N->getOperand(i + NumElems/2);
2106 if (!isUndefOrEqual(Arg, i + NumElems))
2113 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2114 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2115 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2116 bool V2IsSplat = false) {
2117 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2120 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2121 SDOperand BitI = Elts[i];
2122 SDOperand BitI1 = Elts[i+1];
2123 if (!isUndefOrEqual(BitI, j))
2126 if (isUndefOrEqual(BitI1, NumElts))
2129 if (!isUndefOrEqual(BitI1, j + NumElts))
2137 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2138 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2139 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2142 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2143 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2144 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2145 bool V2IsSplat = false) {
2146 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2149 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2150 SDOperand BitI = Elts[i];
2151 SDOperand BitI1 = Elts[i+1];
2152 if (!isUndefOrEqual(BitI, j + NumElts/2))
2155 if (isUndefOrEqual(BitI1, NumElts))
2158 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2166 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2167 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2168 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2171 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2172 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2174 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2175 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2177 unsigned NumElems = N->getNumOperands();
2178 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
2181 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2182 SDOperand BitI = N->getOperand(i);
2183 SDOperand BitI1 = N->getOperand(i+1);
2185 if (!isUndefOrEqual(BitI, j))
2187 if (!isUndefOrEqual(BitI1, j))
2194 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2195 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2196 /// MOVSD, and MOVD, i.e. setting the lowest element.
2197 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2198 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2201 if (!isUndefOrEqual(Elts[0], NumElts))
2204 for (unsigned i = 1; i < NumElts; ++i) {
2205 if (!isUndefOrEqual(Elts[i], i))
2212 bool X86::isMOVLMask(SDNode *N) {
2213 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2214 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2217 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2218 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2219 /// element of vector 2 and the other elements to come from vector 1 in order.
2220 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2221 bool V2IsSplat = false,
2222 bool V2IsUndef = false) {
2223 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2226 if (!isUndefOrEqual(Ops[0], 0))
2229 for (unsigned i = 1; i < NumOps; ++i) {
2230 SDOperand Arg = Ops[i];
2231 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2232 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2233 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2240 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2241 bool V2IsUndef = false) {
2242 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2243 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2244 V2IsSplat, V2IsUndef);
2247 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2248 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2249 bool X86::isMOVSHDUPMask(SDNode *N) {
2250 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2252 if (N->getNumOperands() != 4)
2255 // Expect 1, 1, 3, 3
2256 for (unsigned i = 0; i < 2; ++i) {
2257 SDOperand Arg = N->getOperand(i);
2258 if (Arg.getOpcode() == ISD::UNDEF) continue;
2259 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2260 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2261 if (Val != 1) return false;
2265 for (unsigned i = 2; i < 4; ++i) {
2266 SDOperand Arg = N->getOperand(i);
2267 if (Arg.getOpcode() == ISD::UNDEF) continue;
2268 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2269 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2270 if (Val != 3) return false;
2274 // Don't use movshdup if it can be done with a shufps.
2278 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2279 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2280 bool X86::isMOVSLDUPMask(SDNode *N) {
2281 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2283 if (N->getNumOperands() != 4)
2286 // Expect 0, 0, 2, 2
2287 for (unsigned i = 0; i < 2; ++i) {
2288 SDOperand Arg = N->getOperand(i);
2289 if (Arg.getOpcode() == ISD::UNDEF) continue;
2290 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2291 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2292 if (Val != 0) return false;
2296 for (unsigned i = 2; i < 4; ++i) {
2297 SDOperand Arg = N->getOperand(i);
2298 if (Arg.getOpcode() == ISD::UNDEF) continue;
2299 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2300 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2301 if (Val != 2) return false;
2305 // Don't use movshdup if it can be done with a shufps.
2309 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2310 /// a splat of a single element.
2311 static bool isSplatMask(SDNode *N) {
2312 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2314 // This is a splat operation if each element of the permute is the same, and
2315 // if the value doesn't reference the second vector.
2316 unsigned NumElems = N->getNumOperands();
2317 SDOperand ElementBase;
2319 for (; i != NumElems; ++i) {
2320 SDOperand Elt = N->getOperand(i);
2321 if (isa<ConstantSDNode>(Elt)) {
2327 if (!ElementBase.Val)
2330 for (; i != NumElems; ++i) {
2331 SDOperand Arg = N->getOperand(i);
2332 if (Arg.getOpcode() == ISD::UNDEF) continue;
2333 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2334 if (Arg != ElementBase) return false;
2337 // Make sure it is a splat of the first vector operand.
2338 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2341 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2342 /// a splat of a single element and it's a 2 or 4 element mask.
2343 bool X86::isSplatMask(SDNode *N) {
2344 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2346 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2347 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2349 return ::isSplatMask(N);
2352 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2353 /// specifies a splat of zero element.
2354 bool X86::isSplatLoMask(SDNode *N) {
2355 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2357 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2358 if (!isUndefOrEqual(N->getOperand(i), 0))
2363 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2364 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2366 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2367 unsigned NumOperands = N->getNumOperands();
2368 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2370 for (unsigned i = 0; i < NumOperands; ++i) {
2372 SDOperand Arg = N->getOperand(NumOperands-i-1);
2373 if (Arg.getOpcode() != ISD::UNDEF)
2374 Val = cast<ConstantSDNode>(Arg)->getValue();
2375 if (Val >= NumOperands) Val -= NumOperands;
2377 if (i != NumOperands - 1)
2384 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2385 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2387 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2389 // 8 nodes, but we only care about the last 4.
2390 for (unsigned i = 7; i >= 4; --i) {
2392 SDOperand Arg = N->getOperand(i);
2393 if (Arg.getOpcode() != ISD::UNDEF)
2394 Val = cast<ConstantSDNode>(Arg)->getValue();
2403 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2404 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2406 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2408 // 8 nodes, but we only care about the first 4.
2409 for (int i = 3; i >= 0; --i) {
2411 SDOperand Arg = N->getOperand(i);
2412 if (Arg.getOpcode() != ISD::UNDEF)
2413 Val = cast<ConstantSDNode>(Arg)->getValue();
2422 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2423 /// specifies a 8 element shuffle that can be broken into a pair of
2424 /// PSHUFHW and PSHUFLW.
2425 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2426 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2428 if (N->getNumOperands() != 8)
2431 // Lower quadword shuffled.
2432 for (unsigned i = 0; i != 4; ++i) {
2433 SDOperand Arg = N->getOperand(i);
2434 if (Arg.getOpcode() == ISD::UNDEF) continue;
2435 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2436 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2441 // Upper quadword shuffled.
2442 for (unsigned i = 4; i != 8; ++i) {
2443 SDOperand Arg = N->getOperand(i);
2444 if (Arg.getOpcode() == ISD::UNDEF) continue;
2445 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2446 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2447 if (Val < 4 || Val > 7)
2454 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2455 /// values in ther permute mask.
2456 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2457 SDOperand &V2, SDOperand &Mask,
2458 SelectionDAG &DAG) {
2459 MVT::ValueType VT = Op.getValueType();
2460 MVT::ValueType MaskVT = Mask.getValueType();
2461 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2462 unsigned NumElems = Mask.getNumOperands();
2463 SmallVector<SDOperand, 8> MaskVec;
2465 for (unsigned i = 0; i != NumElems; ++i) {
2466 SDOperand Arg = Mask.getOperand(i);
2467 if (Arg.getOpcode() == ISD::UNDEF) {
2468 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2471 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2472 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2474 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2476 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2480 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2481 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2484 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2485 /// match movhlps. The lower half elements should come from upper half of
2486 /// V1 (and in order), and the upper half elements should come from the upper
2487 /// half of V2 (and in order).
2488 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2489 unsigned NumElems = Mask->getNumOperands();
2492 for (unsigned i = 0, e = 2; i != e; ++i)
2493 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2495 for (unsigned i = 2; i != 4; ++i)
2496 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2501 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2502 /// is promoted to a vector.
2503 static inline bool isScalarLoadToVector(SDNode *N) {
2504 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2505 N = N->getOperand(0).Val;
2506 return ISD::isNON_EXTLoad(N);
2511 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2512 /// match movlp{s|d}. The lower half elements should come from lower half of
2513 /// V1 (and in order), and the upper half elements should come from the upper
2514 /// half of V2 (and in order). And since V1 will become the source of the
2515 /// MOVLP, it must be either a vector load or a scalar load to vector.
2516 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2517 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2519 // Is V2 is a vector load, don't do this transformation. We will try to use
2520 // load folding shufps op.
2521 if (ISD::isNON_EXTLoad(V2))
2524 unsigned NumElems = Mask->getNumOperands();
2525 if (NumElems != 2 && NumElems != 4)
2527 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2528 if (!isUndefOrEqual(Mask->getOperand(i), i))
2530 for (unsigned i = NumElems/2; i != NumElems; ++i)
2531 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2536 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2538 static bool isSplatVector(SDNode *N) {
2539 if (N->getOpcode() != ISD::BUILD_VECTOR)
2542 SDOperand SplatValue = N->getOperand(0);
2543 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2544 if (N->getOperand(i) != SplatValue)
2549 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2551 static bool isUndefShuffle(SDNode *N) {
2552 if (N->getOpcode() != ISD::BUILD_VECTOR)
2555 SDOperand V1 = N->getOperand(0);
2556 SDOperand V2 = N->getOperand(1);
2557 SDOperand Mask = N->getOperand(2);
2558 unsigned NumElems = Mask.getNumOperands();
2559 for (unsigned i = 0; i != NumElems; ++i) {
2560 SDOperand Arg = Mask.getOperand(i);
2561 if (Arg.getOpcode() != ISD::UNDEF) {
2562 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2563 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2565 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2572 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2573 /// that point to V2 points to its first element.
2574 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2575 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2577 bool Changed = false;
2578 SmallVector<SDOperand, 8> MaskVec;
2579 unsigned NumElems = Mask.getNumOperands();
2580 for (unsigned i = 0; i != NumElems; ++i) {
2581 SDOperand Arg = Mask.getOperand(i);
2582 if (Arg.getOpcode() != ISD::UNDEF) {
2583 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2584 if (Val > NumElems) {
2585 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2589 MaskVec.push_back(Arg);
2593 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2594 &MaskVec[0], MaskVec.size());
2598 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2599 /// operation of specified width.
2600 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2601 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2602 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2604 SmallVector<SDOperand, 8> MaskVec;
2605 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2606 for (unsigned i = 1; i != NumElems; ++i)
2607 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2608 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2611 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2612 /// of specified width.
2613 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2614 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2615 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2616 SmallVector<SDOperand, 8> MaskVec;
2617 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2618 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2619 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2621 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2624 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2625 /// of specified width.
2626 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2627 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2628 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2629 unsigned Half = NumElems/2;
2630 SmallVector<SDOperand, 8> MaskVec;
2631 for (unsigned i = 0; i != Half; ++i) {
2632 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2633 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2635 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2638 /// getZeroVector - Returns a vector of specified type with all zero elements.
2640 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2641 assert(MVT::isVector(VT) && "Expected a vector type");
2642 unsigned NumElems = getVectorNumElements(VT);
2643 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2644 bool isFP = MVT::isFloatingPoint(EVT);
2645 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2646 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2647 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2650 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2652 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2653 SDOperand V1 = Op.getOperand(0);
2654 SDOperand Mask = Op.getOperand(2);
2655 MVT::ValueType VT = Op.getValueType();
2656 unsigned NumElems = Mask.getNumOperands();
2657 Mask = getUnpacklMask(NumElems, DAG);
2658 while (NumElems != 4) {
2659 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2662 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2664 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2665 Mask = getZeroVector(MaskVT, DAG);
2666 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2667 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2668 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2671 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2673 static inline bool isZeroNode(SDOperand Elt) {
2674 return ((isa<ConstantSDNode>(Elt) &&
2675 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2676 (isa<ConstantFPSDNode>(Elt) &&
2677 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2680 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2681 /// vector and zero or undef vector.
2682 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2683 unsigned NumElems, unsigned Idx,
2684 bool isZero, SelectionDAG &DAG) {
2685 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2686 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2687 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2688 SDOperand Zero = DAG.getConstant(0, EVT);
2689 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2690 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2691 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2692 &MaskVec[0], MaskVec.size());
2693 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2696 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2698 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2699 unsigned NumNonZero, unsigned NumZero,
2700 SelectionDAG &DAG, TargetLowering &TLI) {
2706 for (unsigned i = 0; i < 16; ++i) {
2707 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2708 if (ThisIsNonZero && First) {
2710 V = getZeroVector(MVT::v8i16, DAG);
2712 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2717 SDOperand ThisElt(0, 0), LastElt(0, 0);
2718 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2719 if (LastIsNonZero) {
2720 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2722 if (ThisIsNonZero) {
2723 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2724 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2725 ThisElt, DAG.getConstant(8, MVT::i8));
2727 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2732 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2733 DAG.getConstant(i/2, TLI.getPointerTy()));
2737 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2740 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2742 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2743 unsigned NumNonZero, unsigned NumZero,
2744 SelectionDAG &DAG, TargetLowering &TLI) {
2750 for (unsigned i = 0; i < 8; ++i) {
2751 bool isNonZero = (NonZeros & (1 << i)) != 0;
2755 V = getZeroVector(MVT::v8i16, DAG);
2757 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2760 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2761 DAG.getConstant(i, TLI.getPointerTy()));
2769 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2770 // All zero's are handled with pxor.
2771 if (ISD::isBuildVectorAllZeros(Op.Val))
2774 // All one's are handled with pcmpeqd.
2775 if (ISD::isBuildVectorAllOnes(Op.Val))
2778 MVT::ValueType VT = Op.getValueType();
2779 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2780 unsigned EVTBits = MVT::getSizeInBits(EVT);
2782 unsigned NumElems = Op.getNumOperands();
2783 unsigned NumZero = 0;
2784 unsigned NumNonZero = 0;
2785 unsigned NonZeros = 0;
2786 std::set<SDOperand> Values;
2787 for (unsigned i = 0; i < NumElems; ++i) {
2788 SDOperand Elt = Op.getOperand(i);
2789 if (Elt.getOpcode() != ISD::UNDEF) {
2791 if (isZeroNode(Elt))
2794 NonZeros |= (1 << i);
2800 if (NumNonZero == 0)
2801 // Must be a mix of zero and undef. Return a zero vector.
2802 return getZeroVector(VT, DAG);
2804 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2805 if (Values.size() == 1)
2808 // Special case for single non-zero element.
2809 if (NumNonZero == 1) {
2810 unsigned Idx = CountTrailingZeros_32(NonZeros);
2811 SDOperand Item = Op.getOperand(Idx);
2812 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2814 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2815 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2818 if (EVTBits == 32) {
2819 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2820 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2822 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2823 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2824 SmallVector<SDOperand, 8> MaskVec;
2825 for (unsigned i = 0; i < NumElems; i++)
2826 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2827 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2828 &MaskVec[0], MaskVec.size());
2829 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2830 DAG.getNode(ISD::UNDEF, VT), Mask);
2834 // Let legalizer expand 2-wide build_vector's.
2838 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2840 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2842 if (V.Val) return V;
2845 if (EVTBits == 16) {
2846 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2848 if (V.Val) return V;
2851 // If element VT is == 32 bits, turn it into a number of shuffles.
2852 SmallVector<SDOperand, 8> V;
2854 if (NumElems == 4 && NumZero > 0) {
2855 for (unsigned i = 0; i < 4; ++i) {
2856 bool isZero = !(NonZeros & (1 << i));
2858 V[i] = getZeroVector(VT, DAG);
2860 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2863 for (unsigned i = 0; i < 2; ++i) {
2864 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2867 V[i] = V[i*2]; // Must be a zero vector.
2870 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2871 getMOVLMask(NumElems, DAG));
2874 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2875 getMOVLMask(NumElems, DAG));
2878 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2879 getUnpacklMask(NumElems, DAG));
2884 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2885 // clears the upper bits.
2886 // FIXME: we can do the same for v4f32 case when we know both parts of
2887 // the lower half come from scalar_to_vector (loadf32). We should do
2888 // that in post legalizer dag combiner with target specific hooks.
2889 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2891 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2892 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2893 SmallVector<SDOperand, 8> MaskVec;
2894 bool Reverse = (NonZeros & 0x3) == 2;
2895 for (unsigned i = 0; i < 2; ++i)
2897 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2899 MaskVec.push_back(DAG.getConstant(i, EVT));
2900 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2901 for (unsigned i = 0; i < 2; ++i)
2903 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2905 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2906 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2907 &MaskVec[0], MaskVec.size());
2908 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2911 if (Values.size() > 2) {
2912 // Expand into a number of unpckl*.
2914 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2915 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2916 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2917 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2918 for (unsigned i = 0; i < NumElems; ++i)
2919 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2921 while (NumElems != 0) {
2922 for (unsigned i = 0; i < NumElems; ++i)
2923 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2934 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2935 SDOperand V1 = Op.getOperand(0);
2936 SDOperand V2 = Op.getOperand(1);
2937 SDOperand PermMask = Op.getOperand(2);
2938 MVT::ValueType VT = Op.getValueType();
2939 unsigned NumElems = PermMask.getNumOperands();
2940 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2941 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2942 bool V1IsSplat = false;
2943 bool V2IsSplat = false;
2945 if (isUndefShuffle(Op.Val))
2946 return DAG.getNode(ISD::UNDEF, VT);
2948 if (isSplatMask(PermMask.Val)) {
2949 if (NumElems <= 4) return Op;
2950 // Promote it to a v4i32 splat.
2951 return PromoteSplat(Op, DAG);
2954 if (X86::isMOVLMask(PermMask.Val))
2955 return (V1IsUndef) ? V2 : Op;
2957 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2958 X86::isMOVSLDUPMask(PermMask.Val) ||
2959 X86::isMOVHLPSMask(PermMask.Val) ||
2960 X86::isMOVHPMask(PermMask.Val) ||
2961 X86::isMOVLPMask(PermMask.Val))
2964 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2965 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2966 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2968 bool Commuted = false;
2969 V1IsSplat = isSplatVector(V1.Val);
2970 V2IsSplat = isSplatVector(V2.Val);
2971 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2972 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2973 std::swap(V1IsSplat, V2IsSplat);
2974 std::swap(V1IsUndef, V2IsUndef);
2978 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2979 if (V2IsUndef) return V1;
2980 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2982 // V2 is a splat, so the mask may be malformed. That is, it may point
2983 // to any V2 element. The instruction selectior won't like this. Get
2984 // a corrected mask and commute to form a proper MOVS{S|D}.
2985 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2986 if (NewMask.Val != PermMask.Val)
2987 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2992 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2993 X86::isUNPCKLMask(PermMask.Val) ||
2994 X86::isUNPCKHMask(PermMask.Val))
2998 // Normalize mask so all entries that point to V2 points to its first
2999 // element then try to match unpck{h|l} again. If match, return a
3000 // new vector_shuffle with the corrected mask.
3001 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3002 if (NewMask.Val != PermMask.Val) {
3003 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3004 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3005 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3006 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3007 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3008 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3013 // Normalize the node to match x86 shuffle ops if needed
3014 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3015 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3018 // Commute is back and try unpck* again.
3019 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3020 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3021 X86::isUNPCKLMask(PermMask.Val) ||
3022 X86::isUNPCKHMask(PermMask.Val))
3026 // If VT is integer, try PSHUF* first, then SHUFP*.
3027 if (MVT::isInteger(VT)) {
3028 if (X86::isPSHUFDMask(PermMask.Val) ||
3029 X86::isPSHUFHWMask(PermMask.Val) ||
3030 X86::isPSHUFLWMask(PermMask.Val)) {
3031 if (V2.getOpcode() != ISD::UNDEF)
3032 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3033 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3037 if (X86::isSHUFPMask(PermMask.Val))
3040 // Handle v8i16 shuffle high / low shuffle node pair.
3041 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3042 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3043 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3044 SmallVector<SDOperand, 8> MaskVec;
3045 for (unsigned i = 0; i != 4; ++i)
3046 MaskVec.push_back(PermMask.getOperand(i));
3047 for (unsigned i = 4; i != 8; ++i)
3048 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3049 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3050 &MaskVec[0], MaskVec.size());
3051 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3053 for (unsigned i = 0; i != 4; ++i)
3054 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3055 for (unsigned i = 4; i != 8; ++i)
3056 MaskVec.push_back(PermMask.getOperand(i));
3057 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3058 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3061 // Floating point cases in the other order.
3062 if (X86::isSHUFPMask(PermMask.Val))
3064 if (X86::isPSHUFDMask(PermMask.Val) ||
3065 X86::isPSHUFHWMask(PermMask.Val) ||
3066 X86::isPSHUFLWMask(PermMask.Val)) {
3067 if (V2.getOpcode() != ISD::UNDEF)
3068 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3069 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3074 if (NumElems == 4) {
3075 MVT::ValueType MaskVT = PermMask.getValueType();
3076 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
3077 SmallVector<std::pair<int, int>, 8> Locs;
3078 Locs.reserve(NumElems);
3079 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3080 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3083 // If no more than two elements come from either vector. This can be
3084 // implemented with two shuffles. First shuffle gather the elements.
3085 // The second shuffle, which takes the first shuffle as both of its
3086 // vector operands, put the elements into the right order.
3087 for (unsigned i = 0; i != NumElems; ++i) {
3088 SDOperand Elt = PermMask.getOperand(i);
3089 if (Elt.getOpcode() == ISD::UNDEF) {
3090 Locs[i] = std::make_pair(-1, -1);
3092 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3093 if (Val < NumElems) {
3094 Locs[i] = std::make_pair(0, NumLo);
3098 Locs[i] = std::make_pair(1, NumHi);
3099 if (2+NumHi < NumElems)
3100 Mask1[2+NumHi] = Elt;
3105 if (NumLo <= 2 && NumHi <= 2) {
3106 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3107 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3108 &Mask1[0], Mask1.size()));
3109 for (unsigned i = 0; i != NumElems; ++i) {
3110 if (Locs[i].first == -1)
3113 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3114 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3115 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3119 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3120 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3121 &Mask2[0], Mask2.size()));
3124 // Break it into (shuffle shuffle_hi, shuffle_lo).
3126 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3127 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3128 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3129 unsigned MaskIdx = 0;
3131 unsigned HiIdx = NumElems/2;
3132 for (unsigned i = 0; i != NumElems; ++i) {
3133 if (i == NumElems/2) {
3139 SDOperand Elt = PermMask.getOperand(i);
3140 if (Elt.getOpcode() == ISD::UNDEF) {
3141 Locs[i] = std::make_pair(-1, -1);
3142 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3143 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3144 (*MaskPtr)[LoIdx] = Elt;
3147 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3148 (*MaskPtr)[HiIdx] = Elt;
3153 SDOperand LoShuffle =
3154 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3155 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3156 &LoMask[0], LoMask.size()));
3157 SDOperand HiShuffle =
3158 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3159 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3160 &HiMask[0], HiMask.size()));
3161 SmallVector<SDOperand, 8> MaskOps;
3162 for (unsigned i = 0; i != NumElems; ++i) {
3163 if (Locs[i].first == -1) {
3164 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3166 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3167 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3170 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3171 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3172 &MaskOps[0], MaskOps.size()));
3179 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3180 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3183 MVT::ValueType VT = Op.getValueType();
3184 // TODO: handle v16i8.
3185 if (MVT::getSizeInBits(VT) == 16) {
3186 // Transform it so it match pextrw which produces a 32-bit result.
3187 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3188 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3189 Op.getOperand(0), Op.getOperand(1));
3190 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3191 DAG.getValueType(VT));
3192 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3193 } else if (MVT::getSizeInBits(VT) == 32) {
3194 SDOperand Vec = Op.getOperand(0);
3195 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3198 // SHUFPS the element to the lowest double word, then movss.
3199 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3200 SmallVector<SDOperand, 8> IdxVec;
3201 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
3202 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3203 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3204 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3205 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3206 &IdxVec[0], IdxVec.size());
3207 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3208 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3209 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3210 DAG.getConstant(0, getPointerTy()));
3211 } else if (MVT::getSizeInBits(VT) == 64) {
3212 SDOperand Vec = Op.getOperand(0);
3213 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3217 // UNPCKHPD the element to the lowest double word, then movsd.
3218 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3219 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3220 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3221 SmallVector<SDOperand, 8> IdxVec;
3222 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
3223 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3224 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3225 &IdxVec[0], IdxVec.size());
3226 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3227 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3228 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3229 DAG.getConstant(0, getPointerTy()));
3236 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3237 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3238 // as its second argument.
3239 MVT::ValueType VT = Op.getValueType();
3240 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
3241 SDOperand N0 = Op.getOperand(0);
3242 SDOperand N1 = Op.getOperand(1);
3243 SDOperand N2 = Op.getOperand(2);
3244 if (MVT::getSizeInBits(BaseVT) == 16) {
3245 if (N1.getValueType() != MVT::i32)
3246 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3247 if (N2.getValueType() != MVT::i32)
3248 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
3249 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3250 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3251 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3254 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3255 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3256 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3257 SmallVector<SDOperand, 8> MaskVec;
3258 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3259 for (unsigned i = 1; i <= 3; ++i)
3260 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3261 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3262 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3263 &MaskVec[0], MaskVec.size()));
3265 // Use two pinsrw instructions to insert a 32 bit value.
3267 if (MVT::isFloatingPoint(N1.getValueType())) {
3268 if (ISD::isNON_EXTLoad(N1.Val)) {
3269 // Just load directly from f32mem to GR32.
3270 LoadSDNode *LD = cast<LoadSDNode>(N1);
3271 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
3272 LD->getSrcValue(), LD->getSrcValueOffset());
3274 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3275 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3276 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3277 DAG.getConstant(0, getPointerTy()));
3280 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3281 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3282 DAG.getConstant(Idx, getPointerTy()));
3283 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3284 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3285 DAG.getConstant(Idx+1, getPointerTy()));
3286 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3294 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3295 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3296 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3299 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3300 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3301 // one of the above mentioned nodes. It has to be wrapped because otherwise
3302 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3303 // be used to form addressing mode. These wrapped nodes will be selected
3306 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3307 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3308 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3310 CP->getAlignment());
3311 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3312 // With PIC, the address is actually $g + Offset.
3313 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3314 !Subtarget->isPICStyleRIPRel()) {
3315 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3316 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3324 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3325 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3326 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3327 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3328 // With PIC, the address is actually $g + Offset.
3329 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3330 !Subtarget->isPICStyleRIPRel()) {
3331 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3332 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3336 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3337 // load the value at address GV, not the value of GV itself. This means that
3338 // the GlobalAddress must be in the base or index register of the address, not
3339 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3340 // The same applies for external symbols during PIC codegen
3341 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3342 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3348 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3349 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3350 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3351 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3352 // With PIC, the address is actually $g + Offset.
3353 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3354 !Subtarget->isPICStyleRIPRel()) {
3355 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3356 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3363 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3364 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3365 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3366 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3367 // With PIC, the address is actually $g + Offset.
3368 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3369 !Subtarget->isPICStyleRIPRel()) {
3370 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3371 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3378 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3379 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3380 "Not an i64 shift!");
3381 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3382 SDOperand ShOpLo = Op.getOperand(0);
3383 SDOperand ShOpHi = Op.getOperand(1);
3384 SDOperand ShAmt = Op.getOperand(2);
3385 SDOperand Tmp1 = isSRA ?
3386 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3387 DAG.getConstant(0, MVT::i32);
3389 SDOperand Tmp2, Tmp3;
3390 if (Op.getOpcode() == ISD::SHL_PARTS) {
3391 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3392 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3394 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3395 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3398 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3399 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3400 DAG.getConstant(32, MVT::i8));
3401 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3402 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3405 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3407 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3408 SmallVector<SDOperand, 4> Ops;
3409 if (Op.getOpcode() == ISD::SHL_PARTS) {
3410 Ops.push_back(Tmp2);
3411 Ops.push_back(Tmp3);
3413 Ops.push_back(InFlag);
3414 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3415 InFlag = Hi.getValue(1);
3418 Ops.push_back(Tmp3);
3419 Ops.push_back(Tmp1);
3421 Ops.push_back(InFlag);
3422 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3424 Ops.push_back(Tmp2);
3425 Ops.push_back(Tmp3);
3427 Ops.push_back(InFlag);
3428 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3429 InFlag = Lo.getValue(1);
3432 Ops.push_back(Tmp3);
3433 Ops.push_back(Tmp1);
3435 Ops.push_back(InFlag);
3436 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3439 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3443 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3446 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3447 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3448 Op.getOperand(0).getValueType() >= MVT::i16 &&
3449 "Unknown SINT_TO_FP to lower!");
3452 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3453 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3454 MachineFunction &MF = DAG.getMachineFunction();
3455 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3456 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3457 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3458 StackSlot, NULL, 0);
3463 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3465 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3466 SmallVector<SDOperand, 8> Ops;
3467 Ops.push_back(Chain);
3468 Ops.push_back(StackSlot);
3469 Ops.push_back(DAG.getValueType(SrcVT));
3470 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3471 Tys, &Ops[0], Ops.size());
3474 Chain = Result.getValue(1);
3475 SDOperand InFlag = Result.getValue(2);
3477 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3478 // shouldn't be necessary except that RFP cannot be live across
3479 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3480 MachineFunction &MF = DAG.getMachineFunction();
3481 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3482 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3483 Tys = DAG.getVTList(MVT::Other);
3484 SmallVector<SDOperand, 8> Ops;
3485 Ops.push_back(Chain);
3486 Ops.push_back(Result);
3487 Ops.push_back(StackSlot);
3488 Ops.push_back(DAG.getValueType(Op.getValueType()));
3489 Ops.push_back(InFlag);
3490 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3491 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3497 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3498 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3499 "Unknown FP_TO_SINT to lower!");
3500 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3502 MachineFunction &MF = DAG.getMachineFunction();
3503 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3504 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3505 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3508 switch (Op.getValueType()) {
3509 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3510 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3511 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3512 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3515 SDOperand Chain = DAG.getEntryNode();
3516 SDOperand Value = Op.getOperand(0);
3518 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3519 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3520 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3522 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3524 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3525 Chain = Value.getValue(1);
3526 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3527 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3530 // Build the FP_TO_INT*_IN_MEM
3531 SDOperand Ops[] = { Chain, Value, StackSlot };
3532 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3535 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3538 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3539 MVT::ValueType VT = Op.getValueType();
3540 const Type *OpNTy = MVT::getTypeForValueType(VT);
3541 std::vector<Constant*> CV;
3542 if (VT == MVT::f64) {
3543 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3544 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3546 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3547 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3548 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3549 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3551 Constant *CS = ConstantStruct::get(CV);
3552 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3553 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3554 SmallVector<SDOperand, 3> Ops;
3555 Ops.push_back(DAG.getEntryNode());
3556 Ops.push_back(CPIdx);
3557 Ops.push_back(DAG.getSrcValue(NULL));
3558 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3559 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3562 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3563 MVT::ValueType VT = Op.getValueType();
3564 const Type *OpNTy = MVT::getTypeForValueType(VT);
3565 std::vector<Constant*> CV;
3566 if (VT == MVT::f64) {
3567 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3568 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3570 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3571 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3572 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3573 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3575 Constant *CS = ConstantStruct::get(CV);
3576 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3577 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3578 SmallVector<SDOperand, 3> Ops;
3579 Ops.push_back(DAG.getEntryNode());
3580 Ops.push_back(CPIdx);
3581 Ops.push_back(DAG.getSrcValue(NULL));
3582 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3583 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3586 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3587 SDOperand Op0 = Op.getOperand(0);
3588 SDOperand Op1 = Op.getOperand(1);
3589 MVT::ValueType VT = Op.getValueType();
3590 MVT::ValueType SrcVT = Op1.getValueType();
3591 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3593 // If second operand is smaller, extend it first.
3594 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3595 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3599 // First get the sign bit of second operand.
3600 std::vector<Constant*> CV;
3601 if (SrcVT == MVT::f64) {
3602 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3603 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3605 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3606 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3607 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3608 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3610 Constant *CS = ConstantStruct::get(CV);
3611 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3612 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3613 SmallVector<SDOperand, 3> Ops;
3614 Ops.push_back(DAG.getEntryNode());
3615 Ops.push_back(CPIdx);
3616 Ops.push_back(DAG.getSrcValue(NULL));
3617 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3618 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3620 // Shift sign bit right or left if the two operands have different types.
3621 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3622 // Op0 is MVT::f32, Op1 is MVT::f64.
3623 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3624 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3625 DAG.getConstant(32, MVT::i32));
3626 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3627 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3628 DAG.getConstant(0, getPointerTy()));
3631 // Clear first operand sign bit.
3633 if (VT == MVT::f64) {
3634 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3635 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3637 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3638 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3639 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3640 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3642 CS = ConstantStruct::get(CV);
3643 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3644 Tys = DAG.getVTList(VT, MVT::Other);
3646 Ops.push_back(DAG.getEntryNode());
3647 Ops.push_back(CPIdx);
3648 Ops.push_back(DAG.getSrcValue(NULL));
3649 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3650 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3652 // Or the value with the sign bit.
3653 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3656 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3658 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3660 SDOperand Op0 = Op.getOperand(0);
3661 SDOperand Op1 = Op.getOperand(1);
3662 SDOperand CC = Op.getOperand(2);
3663 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3664 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3665 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3666 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3669 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3671 SDOperand Ops1[] = { Chain, Op0, Op1 };
3672 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3673 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3674 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3677 assert(isFP && "Illegal integer SetCC!");
3679 SDOperand COps[] = { Chain, Op0, Op1 };
3680 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3682 switch (SetCCOpcode) {
3683 default: assert(false && "Illegal floating point SetCC!");
3684 case ISD::SETOEQ: { // !PF & ZF
3685 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3686 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3687 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3689 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3690 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3692 case ISD::SETUNE: { // PF | !ZF
3693 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3694 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3695 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3697 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3698 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3703 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3704 bool addTest = true;
3705 SDOperand Chain = DAG.getEntryNode();
3706 SDOperand Cond = Op.getOperand(0);
3708 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3710 if (Cond.getOpcode() == ISD::SETCC)
3711 Cond = LowerSETCC(Cond, DAG, Chain);
3713 if (Cond.getOpcode() == X86ISD::SETCC) {
3714 CC = Cond.getOperand(0);
3716 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3717 // (since flag operand cannot be shared). Use it as the condition setting
3718 // operand in place of the X86ISD::SETCC.
3719 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3720 // to use a test instead of duplicating the X86ISD::CMP (for register
3721 // pressure reason)?
3722 SDOperand Cmp = Cond.getOperand(1);
3723 unsigned Opc = Cmp.getOpcode();
3724 bool IllegalFPCMov = !X86ScalarSSE &&
3725 MVT::isFloatingPoint(Op.getValueType()) &&
3726 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3727 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3729 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3730 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3736 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3737 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3738 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3741 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3742 SmallVector<SDOperand, 4> Ops;
3743 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3744 // condition is true.
3745 Ops.push_back(Op.getOperand(2));
3746 Ops.push_back(Op.getOperand(1));
3748 Ops.push_back(Cond.getValue(1));
3749 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3752 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3753 bool addTest = true;
3754 SDOperand Chain = Op.getOperand(0);
3755 SDOperand Cond = Op.getOperand(1);
3756 SDOperand Dest = Op.getOperand(2);
3758 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3760 if (Cond.getOpcode() == ISD::SETCC)
3761 Cond = LowerSETCC(Cond, DAG, Chain);
3763 if (Cond.getOpcode() == X86ISD::SETCC) {
3764 CC = Cond.getOperand(0);
3766 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3767 // (since flag operand cannot be shared). Use it as the condition setting
3768 // operand in place of the X86ISD::SETCC.
3769 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3770 // to use a test instead of duplicating the X86ISD::CMP (for register
3771 // pressure reason)?
3772 SDOperand Cmp = Cond.getOperand(1);
3773 unsigned Opc = Cmp.getOpcode();
3774 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3775 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3776 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3782 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3783 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3784 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3786 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3787 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3790 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3791 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3793 if (Subtarget->is64Bit())
3794 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3796 switch (CallingConv) {
3798 assert(0 && "Unsupported calling convention");
3799 case CallingConv::Fast:
3801 return LowerFastCCCallTo(Op, DAG, CallingConv);
3803 case CallingConv::C:
3804 case CallingConv::X86_StdCall:
3805 return LowerCCCCallTo(Op, DAG, CallingConv);
3806 case CallingConv::X86_FastCall:
3807 return LowerFastCCCallTo(Op, DAG, CallingConv);
3811 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
3812 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
3814 // Support up returning up to two registers.
3815 MVT::ValueType VTs[2];
3816 unsigned DestRegs[2];
3817 unsigned NumRegs = Op.getNumOperands() / 2;
3818 assert(NumRegs <= 2 && "Can only return up to two regs!");
3820 for (unsigned i = 0; i != NumRegs; ++i)
3821 VTs[i] = Op.getOperand(i*2+1).getValueType();
3823 // Determine which register each value should be copied into.
3824 GetRetValueLocs(VTs, NumRegs, DestRegs, Subtarget,
3825 DAG.getMachineFunction().getFunction()->getCallingConv());
3827 // If this is the first return lowered for this function, add the regs to the
3828 // liveout set for the function.
3829 if (DAG.getMachineFunction().liveout_empty()) {
3830 for (unsigned i = 0; i != NumRegs; ++i)
3831 DAG.getMachineFunction().addLiveOut(DestRegs[i]);
3834 SDOperand Chain = Op.getOperand(0);
3837 // Copy the result values into the output registers.
3838 if (NumRegs != 1 || DestRegs[0] != X86::ST0) {
3839 for (unsigned i = 0; i != NumRegs; ++i) {
3840 Chain = DAG.getCopyToReg(Chain, DestRegs[i], Op.getOperand(i*2+1), Flag);
3841 Flag = Chain.getValue(1);
3844 // We need to handle a destination of ST0 specially, because it isn't really
3846 SDOperand Value = Op.getOperand(1);
3848 // If this is an FP return with ScalarSSE, we need to move the value from
3849 // an XMM register onto the fp-stack.
3853 // If this is a load into a scalarsse value, don't store the loaded value
3854 // back to the stack, only to reload it: just replace the scalar-sse load.
3855 if (ISD::isNON_EXTLoad(Value.Val) &&
3856 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
3857 Chain = Value.getOperand(0);
3858 MemLoc = Value.getOperand(1);
3860 // Spill the value to memory and reload it into top of stack.
3861 unsigned Size = MVT::getSizeInBits(VTs[0])/8;
3862 MachineFunction &MF = DAG.getMachineFunction();
3863 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3864 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
3865 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
3867 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3868 SDOperand Ops[] = { Chain, MemLoc, DAG.getValueType(VTs[0]) };
3869 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3870 Chain = Value.getValue(1);
3873 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3874 SDOperand Ops[] = { Chain, Value };
3875 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
3876 Flag = Chain.getValue(1);
3879 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
3881 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
3883 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
3887 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3888 MachineFunction &MF = DAG.getMachineFunction();
3889 const Function* Fn = MF.getFunction();
3890 if (Fn->hasExternalLinkage() &&
3891 Subtarget->isTargetCygMing() &&
3892 Fn->getName() == "main")
3893 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
3895 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3896 if (Subtarget->is64Bit())
3897 return LowerX86_64CCCArguments(Op, DAG);
3901 assert(0 && "Unsupported calling convention");
3902 case CallingConv::Fast:
3904 return LowerFastCCArguments(Op, DAG);
3907 case CallingConv::C:
3908 return LowerCCCArguments(Op, DAG);
3909 case CallingConv::X86_StdCall:
3910 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
3911 return LowerCCCArguments(Op, DAG, true);
3912 case CallingConv::X86_FastCall:
3913 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
3914 return LowerFastCCArguments(Op, DAG, true);
3918 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3919 SDOperand InFlag(0, 0);
3920 SDOperand Chain = Op.getOperand(0);
3922 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3923 if (Align == 0) Align = 1;
3925 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3926 // If not DWORD aligned, call memset if size is less than the threshold.
3927 // It knows how to align to the right boundary first.
3928 if ((Align & 3) != 0 ||
3929 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3930 MVT::ValueType IntPtr = getPointerTy();
3931 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3932 TargetLowering::ArgListTy Args;
3933 TargetLowering::ArgListEntry Entry;
3934 Entry.Node = Op.getOperand(1);
3935 Entry.Ty = IntPtrTy;
3936 Entry.isSigned = false;
3937 Entry.isInReg = false;
3938 Entry.isSRet = false;
3939 Args.push_back(Entry);
3940 // Extend the unsigned i8 argument to be an int value for the call.
3941 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3942 Entry.Ty = IntPtrTy;
3943 Entry.isSigned = false;
3944 Entry.isInReg = false;
3945 Entry.isSRet = false;
3946 Args.push_back(Entry);
3947 Entry.Node = Op.getOperand(3);
3948 Args.push_back(Entry);
3949 std::pair<SDOperand,SDOperand> CallResult =
3950 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3951 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3952 return CallResult.second;
3957 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3958 unsigned BytesLeft = 0;
3959 bool TwoRepStos = false;
3962 uint64_t Val = ValC->getValue() & 255;
3964 // If the value is a constant, then we can potentially use larger sets.
3965 switch (Align & 3) {
3966 case 2: // WORD aligned
3969 Val = (Val << 8) | Val;
3971 case 0: // DWORD aligned
3974 Val = (Val << 8) | Val;
3975 Val = (Val << 16) | Val;
3976 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3979 Val = (Val << 32) | Val;
3982 default: // Byte aligned
3985 Count = Op.getOperand(3);
3989 if (AVT > MVT::i8) {
3991 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3992 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3993 BytesLeft = I->getValue() % UBytes;
3995 assert(AVT >= MVT::i32 &&
3996 "Do not use rep;stos if not at least DWORD aligned");
3997 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3998 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4003 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4005 InFlag = Chain.getValue(1);
4008 Count = Op.getOperand(3);
4009 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4010 InFlag = Chain.getValue(1);
4013 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4015 InFlag = Chain.getValue(1);
4016 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4017 Op.getOperand(1), InFlag);
4018 InFlag = Chain.getValue(1);
4020 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4021 SmallVector<SDOperand, 8> Ops;
4022 Ops.push_back(Chain);
4023 Ops.push_back(DAG.getValueType(AVT));
4024 Ops.push_back(InFlag);
4025 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4028 InFlag = Chain.getValue(1);
4029 Count = Op.getOperand(3);
4030 MVT::ValueType CVT = Count.getValueType();
4031 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4032 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4033 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4035 InFlag = Chain.getValue(1);
4036 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4038 Ops.push_back(Chain);
4039 Ops.push_back(DAG.getValueType(MVT::i8));
4040 Ops.push_back(InFlag);
4041 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4042 } else if (BytesLeft) {
4043 // Issue stores for the last 1 - 7 bytes.
4045 unsigned Val = ValC->getValue() & 255;
4046 unsigned Offset = I->getValue() - BytesLeft;
4047 SDOperand DstAddr = Op.getOperand(1);
4048 MVT::ValueType AddrVT = DstAddr.getValueType();
4049 if (BytesLeft >= 4) {
4050 Val = (Val << 8) | Val;
4051 Val = (Val << 16) | Val;
4052 Value = DAG.getConstant(Val, MVT::i32);
4053 Chain = DAG.getStore(Chain, Value,
4054 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4055 DAG.getConstant(Offset, AddrVT)),
4060 if (BytesLeft >= 2) {
4061 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4062 Chain = DAG.getStore(Chain, Value,
4063 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4064 DAG.getConstant(Offset, AddrVT)),
4069 if (BytesLeft == 1) {
4070 Value = DAG.getConstant(Val, MVT::i8);
4071 Chain = DAG.getStore(Chain, Value,
4072 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4073 DAG.getConstant(Offset, AddrVT)),
4081 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
4082 SDOperand Chain = Op.getOperand(0);
4084 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4085 if (Align == 0) Align = 1;
4087 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4088 // If not DWORD aligned, call memcpy if size is less than the threshold.
4089 // It knows how to align to the right boundary first.
4090 if ((Align & 3) != 0 ||
4091 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4092 MVT::ValueType IntPtr = getPointerTy();
4093 TargetLowering::ArgListTy Args;
4094 TargetLowering::ArgListEntry Entry;
4095 Entry.Ty = getTargetData()->getIntPtrType();
4096 Entry.isSigned = false;
4097 Entry.isInReg = false;
4098 Entry.isSRet = false;
4099 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
4100 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
4101 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
4102 std::pair<SDOperand,SDOperand> CallResult =
4103 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4104 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
4105 return CallResult.second;
4110 unsigned BytesLeft = 0;
4111 bool TwoRepMovs = false;
4112 switch (Align & 3) {
4113 case 2: // WORD aligned
4116 case 0: // DWORD aligned
4118 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4121 default: // Byte aligned
4123 Count = Op.getOperand(3);
4127 if (AVT > MVT::i8) {
4129 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4130 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4131 BytesLeft = I->getValue() % UBytes;
4133 assert(AVT >= MVT::i32 &&
4134 "Do not use rep;movs if not at least DWORD aligned");
4135 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4136 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4141 SDOperand InFlag(0, 0);
4142 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4144 InFlag = Chain.getValue(1);
4145 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4146 Op.getOperand(1), InFlag);
4147 InFlag = Chain.getValue(1);
4148 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4149 Op.getOperand(2), InFlag);
4150 InFlag = Chain.getValue(1);
4152 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4153 SmallVector<SDOperand, 8> Ops;
4154 Ops.push_back(Chain);
4155 Ops.push_back(DAG.getValueType(AVT));
4156 Ops.push_back(InFlag);
4157 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4160 InFlag = Chain.getValue(1);
4161 Count = Op.getOperand(3);
4162 MVT::ValueType CVT = Count.getValueType();
4163 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4164 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4165 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4167 InFlag = Chain.getValue(1);
4168 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4170 Ops.push_back(Chain);
4171 Ops.push_back(DAG.getValueType(MVT::i8));
4172 Ops.push_back(InFlag);
4173 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4174 } else if (BytesLeft) {
4175 // Issue loads and stores for the last 1 - 7 bytes.
4176 unsigned Offset = I->getValue() - BytesLeft;
4177 SDOperand DstAddr = Op.getOperand(1);
4178 MVT::ValueType DstVT = DstAddr.getValueType();
4179 SDOperand SrcAddr = Op.getOperand(2);
4180 MVT::ValueType SrcVT = SrcAddr.getValueType();
4182 if (BytesLeft >= 4) {
4183 Value = DAG.getLoad(MVT::i32, Chain,
4184 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4185 DAG.getConstant(Offset, SrcVT)),
4187 Chain = Value.getValue(1);
4188 Chain = DAG.getStore(Chain, Value,
4189 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4190 DAG.getConstant(Offset, DstVT)),
4195 if (BytesLeft >= 2) {
4196 Value = DAG.getLoad(MVT::i16, Chain,
4197 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4198 DAG.getConstant(Offset, SrcVT)),
4200 Chain = Value.getValue(1);
4201 Chain = DAG.getStore(Chain, Value,
4202 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4203 DAG.getConstant(Offset, DstVT)),
4209 if (BytesLeft == 1) {
4210 Value = DAG.getLoad(MVT::i8, Chain,
4211 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4212 DAG.getConstant(Offset, SrcVT)),
4214 Chain = Value.getValue(1);
4215 Chain = DAG.getStore(Chain, Value,
4216 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4217 DAG.getConstant(Offset, DstVT)),
4226 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4227 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4228 SDOperand TheOp = Op.getOperand(0);
4229 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4230 if (Subtarget->is64Bit()) {
4231 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4232 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4233 MVT::i64, Copy1.getValue(2));
4234 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4235 DAG.getConstant(32, MVT::i8));
4237 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4240 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4241 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4244 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4245 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4246 MVT::i32, Copy1.getValue(2));
4247 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4248 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4249 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4252 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4253 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4255 if (!Subtarget->is64Bit()) {
4256 // vastart just stores the address of the VarArgsFrameIndex slot into the
4257 // memory location argument.
4258 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4259 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4264 // gp_offset (0 - 6 * 8)
4265 // fp_offset (48 - 48 + 8 * 16)
4266 // overflow_arg_area (point to parameters coming in memory).
4268 SmallVector<SDOperand, 8> MemOps;
4269 SDOperand FIN = Op.getOperand(1);
4271 SDOperand Store = DAG.getStore(Op.getOperand(0),
4272 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4273 FIN, SV->getValue(), SV->getOffset());
4274 MemOps.push_back(Store);
4277 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4278 DAG.getConstant(4, getPointerTy()));
4279 Store = DAG.getStore(Op.getOperand(0),
4280 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4281 FIN, SV->getValue(), SV->getOffset());
4282 MemOps.push_back(Store);
4284 // Store ptr to overflow_arg_area
4285 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4286 DAG.getConstant(4, getPointerTy()));
4287 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4288 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4290 MemOps.push_back(Store);
4292 // Store ptr to reg_save_area.
4293 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4294 DAG.getConstant(8, getPointerTy()));
4295 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4296 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4298 MemOps.push_back(Store);
4299 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4303 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4304 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4306 default: return SDOperand(); // Don't custom lower most intrinsics.
4307 // Comparison intrinsics.
4308 case Intrinsic::x86_sse_comieq_ss:
4309 case Intrinsic::x86_sse_comilt_ss:
4310 case Intrinsic::x86_sse_comile_ss:
4311 case Intrinsic::x86_sse_comigt_ss:
4312 case Intrinsic::x86_sse_comige_ss:
4313 case Intrinsic::x86_sse_comineq_ss:
4314 case Intrinsic::x86_sse_ucomieq_ss:
4315 case Intrinsic::x86_sse_ucomilt_ss:
4316 case Intrinsic::x86_sse_ucomile_ss:
4317 case Intrinsic::x86_sse_ucomigt_ss:
4318 case Intrinsic::x86_sse_ucomige_ss:
4319 case Intrinsic::x86_sse_ucomineq_ss:
4320 case Intrinsic::x86_sse2_comieq_sd:
4321 case Intrinsic::x86_sse2_comilt_sd:
4322 case Intrinsic::x86_sse2_comile_sd:
4323 case Intrinsic::x86_sse2_comigt_sd:
4324 case Intrinsic::x86_sse2_comige_sd:
4325 case Intrinsic::x86_sse2_comineq_sd:
4326 case Intrinsic::x86_sse2_ucomieq_sd:
4327 case Intrinsic::x86_sse2_ucomilt_sd:
4328 case Intrinsic::x86_sse2_ucomile_sd:
4329 case Intrinsic::x86_sse2_ucomigt_sd:
4330 case Intrinsic::x86_sse2_ucomige_sd:
4331 case Intrinsic::x86_sse2_ucomineq_sd: {
4333 ISD::CondCode CC = ISD::SETCC_INVALID;
4336 case Intrinsic::x86_sse_comieq_ss:
4337 case Intrinsic::x86_sse2_comieq_sd:
4341 case Intrinsic::x86_sse_comilt_ss:
4342 case Intrinsic::x86_sse2_comilt_sd:
4346 case Intrinsic::x86_sse_comile_ss:
4347 case Intrinsic::x86_sse2_comile_sd:
4351 case Intrinsic::x86_sse_comigt_ss:
4352 case Intrinsic::x86_sse2_comigt_sd:
4356 case Intrinsic::x86_sse_comige_ss:
4357 case Intrinsic::x86_sse2_comige_sd:
4361 case Intrinsic::x86_sse_comineq_ss:
4362 case Intrinsic::x86_sse2_comineq_sd:
4366 case Intrinsic::x86_sse_ucomieq_ss:
4367 case Intrinsic::x86_sse2_ucomieq_sd:
4368 Opc = X86ISD::UCOMI;
4371 case Intrinsic::x86_sse_ucomilt_ss:
4372 case Intrinsic::x86_sse2_ucomilt_sd:
4373 Opc = X86ISD::UCOMI;
4376 case Intrinsic::x86_sse_ucomile_ss:
4377 case Intrinsic::x86_sse2_ucomile_sd:
4378 Opc = X86ISD::UCOMI;
4381 case Intrinsic::x86_sse_ucomigt_ss:
4382 case Intrinsic::x86_sse2_ucomigt_sd:
4383 Opc = X86ISD::UCOMI;
4386 case Intrinsic::x86_sse_ucomige_ss:
4387 case Intrinsic::x86_sse2_ucomige_sd:
4388 Opc = X86ISD::UCOMI;
4391 case Intrinsic::x86_sse_ucomineq_ss:
4392 case Intrinsic::x86_sse2_ucomineq_sd:
4393 Opc = X86ISD::UCOMI;
4399 SDOperand LHS = Op.getOperand(1);
4400 SDOperand RHS = Op.getOperand(2);
4401 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4403 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4404 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4405 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4406 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4407 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4408 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4409 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4414 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4415 // Depths > 0 not supported yet!
4416 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4419 // Just load the return address
4420 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4421 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4424 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4425 // Depths > 0 not supported yet!
4426 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4429 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4430 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4431 DAG.getConstant(4, getPointerTy()));
4434 /// LowerOperation - Provide custom lowering hooks for some operations.
4436 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4437 switch (Op.getOpcode()) {
4438 default: assert(0 && "Should not custom lower this!");
4439 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4440 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4441 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4442 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4443 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4444 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4445 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4446 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4447 case ISD::SHL_PARTS:
4448 case ISD::SRA_PARTS:
4449 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4450 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4451 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4452 case ISD::FABS: return LowerFABS(Op, DAG);
4453 case ISD::FNEG: return LowerFNEG(Op, DAG);
4454 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4455 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4456 case ISD::SELECT: return LowerSELECT(Op, DAG);
4457 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4458 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4459 case ISD::CALL: return LowerCALL(Op, DAG);
4460 case ISD::RET: return LowerRET(Op, DAG);
4461 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4462 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4463 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4464 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4465 case ISD::VASTART: return LowerVASTART(Op, DAG);
4466 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4467 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4468 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4473 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4475 default: return NULL;
4476 case X86ISD::SHLD: return "X86ISD::SHLD";
4477 case X86ISD::SHRD: return "X86ISD::SHRD";
4478 case X86ISD::FAND: return "X86ISD::FAND";
4479 case X86ISD::FOR: return "X86ISD::FOR";
4480 case X86ISD::FXOR: return "X86ISD::FXOR";
4481 case X86ISD::FSRL: return "X86ISD::FSRL";
4482 case X86ISD::FILD: return "X86ISD::FILD";
4483 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4484 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4485 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4486 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4487 case X86ISD::FLD: return "X86ISD::FLD";
4488 case X86ISD::FST: return "X86ISD::FST";
4489 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4490 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4491 case X86ISD::CALL: return "X86ISD::CALL";
4492 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4493 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4494 case X86ISD::CMP: return "X86ISD::CMP";
4495 case X86ISD::COMI: return "X86ISD::COMI";
4496 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4497 case X86ISD::SETCC: return "X86ISD::SETCC";
4498 case X86ISD::CMOV: return "X86ISD::CMOV";
4499 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4500 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4501 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4502 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4503 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4504 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4505 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4506 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4507 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4508 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4509 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4510 case X86ISD::FMAX: return "X86ISD::FMAX";
4511 case X86ISD::FMIN: return "X86ISD::FMIN";
4515 /// isLegalAddressImmediate - Return true if the integer value or
4516 /// GlobalValue can be used as the offset of the target addressing mode.
4517 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
4518 // X86 allows a sign-extended 32-bit immediate field.
4519 return (V > -(1LL << 32) && V < (1LL << 32)-1);
4522 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
4523 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
4524 // field unless we are in small code model.
4525 if (Subtarget->is64Bit() &&
4526 getTargetMachine().getCodeModel() != CodeModel::Small)
4529 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4532 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4533 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4534 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4535 /// are assumed to be legal.
4537 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4538 // Only do shuffles on 128-bit vector types for now.
4539 if (MVT::getSizeInBits(VT) == 64) return false;
4540 return (Mask.Val->getNumOperands() <= 4 ||
4541 isSplatMask(Mask.Val) ||
4542 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4543 X86::isUNPCKLMask(Mask.Val) ||
4544 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4545 X86::isUNPCKHMask(Mask.Val));
4548 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4550 SelectionDAG &DAG) const {
4551 unsigned NumElts = BVOps.size();
4552 // Only do shuffles on 128-bit vector types for now.
4553 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4554 if (NumElts == 2) return true;
4556 return (isMOVLMask(&BVOps[0], 4) ||
4557 isCommutedMOVL(&BVOps[0], 4, true) ||
4558 isSHUFPMask(&BVOps[0], 4) ||
4559 isCommutedSHUFP(&BVOps[0], 4));
4564 //===----------------------------------------------------------------------===//
4565 // X86 Scheduler Hooks
4566 //===----------------------------------------------------------------------===//
4569 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4570 MachineBasicBlock *BB) {
4571 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4572 switch (MI->getOpcode()) {
4573 default: assert(false && "Unexpected instr type to insert");
4574 case X86::CMOV_FR32:
4575 case X86::CMOV_FR64:
4576 case X86::CMOV_V4F32:
4577 case X86::CMOV_V2F64:
4578 case X86::CMOV_V2I64: {
4579 // To "insert" a SELECT_CC instruction, we actually have to insert the
4580 // diamond control-flow pattern. The incoming instruction knows the
4581 // destination vreg to set, the condition code register to branch on, the
4582 // true/false values to select between, and a branch opcode to use.
4583 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4584 ilist<MachineBasicBlock>::iterator It = BB;
4590 // cmpTY ccX, r1, r2
4592 // fallthrough --> copy0MBB
4593 MachineBasicBlock *thisMBB = BB;
4594 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4595 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4597 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4598 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4599 MachineFunction *F = BB->getParent();
4600 F->getBasicBlockList().insert(It, copy0MBB);
4601 F->getBasicBlockList().insert(It, sinkMBB);
4602 // Update machine-CFG edges by first adding all successors of the current
4603 // block to the new block which will contain the Phi node for the select.
4604 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4605 e = BB->succ_end(); i != e; ++i)
4606 sinkMBB->addSuccessor(*i);
4607 // Next, remove all successors of the current block, and add the true
4608 // and fallthrough blocks as its successors.
4609 while(!BB->succ_empty())
4610 BB->removeSuccessor(BB->succ_begin());
4611 BB->addSuccessor(copy0MBB);
4612 BB->addSuccessor(sinkMBB);
4615 // %FalseValue = ...
4616 // # fallthrough to sinkMBB
4619 // Update machine-CFG edges
4620 BB->addSuccessor(sinkMBB);
4623 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4626 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4627 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4628 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4630 delete MI; // The pseudo instruction is gone now.
4634 case X86::FP_TO_INT16_IN_MEM:
4635 case X86::FP_TO_INT32_IN_MEM:
4636 case X86::FP_TO_INT64_IN_MEM: {
4637 // Change the floating point control register to use "round towards zero"
4638 // mode when truncating to an integer value.
4639 MachineFunction *F = BB->getParent();
4640 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4641 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4643 // Load the old value of the high byte of the control word...
4645 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4646 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4648 // Set the high part to be round to zero...
4649 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4652 // Reload the modified control word now...
4653 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4655 // Restore the memory image of control word to original value
4656 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4659 // Get the X86 opcode to use.
4661 switch (MI->getOpcode()) {
4662 default: assert(0 && "illegal opcode!");
4663 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4664 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4665 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4669 MachineOperand &Op = MI->getOperand(0);
4670 if (Op.isRegister()) {
4671 AM.BaseType = X86AddressMode::RegBase;
4672 AM.Base.Reg = Op.getReg();
4674 AM.BaseType = X86AddressMode::FrameIndexBase;
4675 AM.Base.FrameIndex = Op.getFrameIndex();
4677 Op = MI->getOperand(1);
4678 if (Op.isImmediate())
4679 AM.Scale = Op.getImm();
4680 Op = MI->getOperand(2);
4681 if (Op.isImmediate())
4682 AM.IndexReg = Op.getImm();
4683 Op = MI->getOperand(3);
4684 if (Op.isGlobalAddress()) {
4685 AM.GV = Op.getGlobal();
4687 AM.Disp = Op.getImm();
4689 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4690 .addReg(MI->getOperand(4).getReg());
4692 // Reload the original control word now.
4693 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4695 delete MI; // The pseudo instruction is gone now.
4701 //===----------------------------------------------------------------------===//
4702 // X86 Optimization Hooks
4703 //===----------------------------------------------------------------------===//
4705 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4707 uint64_t &KnownZero,
4709 unsigned Depth) const {
4710 unsigned Opc = Op.getOpcode();
4711 assert((Opc >= ISD::BUILTIN_OP_END ||
4712 Opc == ISD::INTRINSIC_WO_CHAIN ||
4713 Opc == ISD::INTRINSIC_W_CHAIN ||
4714 Opc == ISD::INTRINSIC_VOID) &&
4715 "Should use MaskedValueIsZero if you don't know whether Op"
4716 " is a target node!");
4718 KnownZero = KnownOne = 0; // Don't know anything.
4722 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4727 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4728 /// element of the result of the vector shuffle.
4729 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4730 MVT::ValueType VT = N->getValueType(0);
4731 SDOperand PermMask = N->getOperand(2);
4732 unsigned NumElems = PermMask.getNumOperands();
4733 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4735 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4737 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4738 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4739 SDOperand Idx = PermMask.getOperand(i);
4740 if (Idx.getOpcode() == ISD::UNDEF)
4741 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4742 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4747 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4748 /// node is a GlobalAddress + an offset.
4749 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4750 unsigned Opc = N->getOpcode();
4751 if (Opc == X86ISD::Wrapper) {
4752 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4753 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4756 } else if (Opc == ISD::ADD) {
4757 SDOperand N1 = N->getOperand(0);
4758 SDOperand N2 = N->getOperand(1);
4759 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4760 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4762 Offset += V->getSignExtended();
4765 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4766 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4768 Offset += V->getSignExtended();
4776 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4778 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4779 MachineFrameInfo *MFI) {
4780 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4783 SDOperand Loc = N->getOperand(1);
4784 SDOperand BaseLoc = Base->getOperand(1);
4785 if (Loc.getOpcode() == ISD::FrameIndex) {
4786 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4788 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4789 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4790 int FS = MFI->getObjectSize(FI);
4791 int BFS = MFI->getObjectSize(BFI);
4792 if (FS != BFS || FS != Size) return false;
4793 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4795 GlobalValue *GV1 = NULL;
4796 GlobalValue *GV2 = NULL;
4797 int64_t Offset1 = 0;
4798 int64_t Offset2 = 0;
4799 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4800 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4801 if (isGA1 && isGA2 && GV1 == GV2)
4802 return Offset1 == (Offset2 + Dist*Size);
4808 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4809 const X86Subtarget *Subtarget) {
4812 if (isGAPlusOffset(Base, GV, Offset))
4813 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4815 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4816 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4818 // Fixed objects do not specify alignment, however the offsets are known.
4819 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4820 (MFI->getObjectOffset(BFI) % 16) == 0);
4822 return MFI->getObjectAlignment(BFI) >= 16;
4828 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4829 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4830 /// if the load addresses are consecutive, non-overlapping, and in the right
4832 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4833 const X86Subtarget *Subtarget) {
4834 MachineFunction &MF = DAG.getMachineFunction();
4835 MachineFrameInfo *MFI = MF.getFrameInfo();
4836 MVT::ValueType VT = N->getValueType(0);
4837 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4838 SDOperand PermMask = N->getOperand(2);
4839 int NumElems = (int)PermMask.getNumOperands();
4840 SDNode *Base = NULL;
4841 for (int i = 0; i < NumElems; ++i) {
4842 SDOperand Idx = PermMask.getOperand(i);
4843 if (Idx.getOpcode() == ISD::UNDEF) {
4844 if (!Base) return SDOperand();
4847 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4848 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4852 else if (!isConsecutiveLoad(Arg.Val, Base,
4853 i, MVT::getSizeInBits(EVT)/8,MFI))
4858 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4860 LoadSDNode *LD = cast<LoadSDNode>(Base);
4861 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4862 LD->getSrcValueOffset());
4864 // Just use movups, it's shorter.
4865 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4866 SmallVector<SDOperand, 3> Ops;
4867 Ops.push_back(Base->getOperand(0));
4868 Ops.push_back(Base->getOperand(1));
4869 Ops.push_back(Base->getOperand(2));
4870 return DAG.getNode(ISD::BIT_CONVERT, VT,
4871 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4875 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4876 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4877 const X86Subtarget *Subtarget) {
4878 SDOperand Cond = N->getOperand(0);
4880 // If we have SSE[12] support, try to form min/max nodes.
4881 if (Subtarget->hasSSE2() &&
4882 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4883 if (Cond.getOpcode() == ISD::SETCC) {
4884 // Get the LHS/RHS of the select.
4885 SDOperand LHS = N->getOperand(1);
4886 SDOperand RHS = N->getOperand(2);
4887 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4889 unsigned Opcode = 0;
4890 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4893 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4896 if (!UnsafeFPMath) break;
4898 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4900 Opcode = X86ISD::FMIN;
4903 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4906 if (!UnsafeFPMath) break;
4908 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4910 Opcode = X86ISD::FMAX;
4913 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4916 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4919 if (!UnsafeFPMath) break;
4921 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4923 Opcode = X86ISD::FMIN;
4926 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4929 if (!UnsafeFPMath) break;
4931 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4933 Opcode = X86ISD::FMAX;
4939 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4948 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4949 DAGCombinerInfo &DCI) const {
4950 SelectionDAG &DAG = DCI.DAG;
4951 switch (N->getOpcode()) {
4953 case ISD::VECTOR_SHUFFLE:
4954 return PerformShuffleCombine(N, DAG, Subtarget);
4956 return PerformSELECTCombine(N, DAG, Subtarget);
4962 //===----------------------------------------------------------------------===//
4963 // X86 Inline Assembly Support
4964 //===----------------------------------------------------------------------===//
4966 /// getConstraintType - Given a constraint letter, return the type of
4967 /// constraint it is for this target.
4968 X86TargetLowering::ConstraintType
4969 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
4970 switch (ConstraintLetter) {
4979 return C_RegisterClass;
4980 default: return TargetLowering::getConstraintType(ConstraintLetter);
4984 /// isOperandValidForConstraint - Return the specified operand (possibly
4985 /// modified) if the specified SDOperand is valid for the specified target
4986 /// constraint letter, otherwise return null.
4987 SDOperand X86TargetLowering::
4988 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4989 switch (Constraint) {
4992 // Literal immediates are always ok.
4993 if (isa<ConstantSDNode>(Op)) return Op;
4995 // If we are in non-pic codegen mode, we allow the address of a global to
4996 // be used with 'i'.
4997 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
4998 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4999 return SDOperand(0, 0);
5001 if (GA->getOpcode() != ISD::TargetGlobalAddress)
5002 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5007 // Otherwise, not valid for this mode.
5008 return SDOperand(0, 0);
5010 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
5014 std::vector<unsigned> X86TargetLowering::
5015 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5016 MVT::ValueType VT) const {
5017 if (Constraint.size() == 1) {
5018 // FIXME: not handling fp-stack yet!
5019 // FIXME: not handling MMX registers yet ('y' constraint).
5020 switch (Constraint[0]) { // GCC X86 Constraint Letters
5021 default: break; // Unknown constraint letter
5022 case 'A': // EAX/EDX
5023 if (VT == MVT::i32 || VT == MVT::i64)
5024 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5026 case 'r': // GENERAL_REGS
5027 case 'R': // LEGACY_REGS
5028 if (VT == MVT::i64 && Subtarget->is64Bit())
5029 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
5030 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
5031 X86::R8, X86::R9, X86::R10, X86::R11,
5032 X86::R12, X86::R13, X86::R14, X86::R15, 0);
5034 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5035 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
5036 else if (VT == MVT::i16)
5037 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5038 X86::SI, X86::DI, X86::BP, X86::SP, 0);
5039 else if (VT == MVT::i8)
5040 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5042 case 'l': // INDEX_REGS
5044 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5045 X86::ESI, X86::EDI, X86::EBP, 0);
5046 else if (VT == MVT::i16)
5047 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5048 X86::SI, X86::DI, X86::BP, 0);
5049 else if (VT == MVT::i8)
5050 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5052 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5055 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5056 else if (VT == MVT::i16)
5057 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5058 else if (VT == MVT::i8)
5059 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5061 case 'x': // SSE_REGS if SSE1 allowed
5062 if (Subtarget->hasSSE1())
5063 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5064 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5066 return std::vector<unsigned>();
5067 case 'Y': // SSE_REGS if SSE2 allowed
5068 if (Subtarget->hasSSE2())
5069 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5070 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5072 return std::vector<unsigned>();
5076 return std::vector<unsigned>();
5079 std::pair<unsigned, const TargetRegisterClass*>
5080 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5081 MVT::ValueType VT) const {
5082 // Use the default implementation in TargetLowering to convert the register
5083 // constraint into a member of a register class.
5084 std::pair<unsigned, const TargetRegisterClass*> Res;
5085 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5087 // Not found as a standard register?
5088 if (Res.second == 0) {
5089 // GCC calls "st(0)" just plain "st".
5090 if (StringsEqualNoCase("{st}", Constraint)) {
5091 Res.first = X86::ST0;
5092 Res.second = X86::RSTRegisterClass;
5098 // Otherwise, check to see if this is a register class of the wrong value
5099 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5100 // turn into {ax},{dx}.
5101 if (Res.second->hasType(VT))
5102 return Res; // Correct type already, nothing to do.
5104 // All of the single-register GCC register classes map their values onto
5105 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5106 // really want an 8-bit or 32-bit register, map to the appropriate register
5107 // class and return the appropriate register.
5108 if (Res.second != X86::GR16RegisterClass)
5111 if (VT == MVT::i8) {
5112 unsigned DestReg = 0;
5113 switch (Res.first) {
5115 case X86::AX: DestReg = X86::AL; break;
5116 case X86::DX: DestReg = X86::DL; break;
5117 case X86::CX: DestReg = X86::CL; break;
5118 case X86::BX: DestReg = X86::BL; break;
5121 Res.first = DestReg;
5122 Res.second = Res.second = X86::GR8RegisterClass;
5124 } else if (VT == MVT::i32) {
5125 unsigned DestReg = 0;
5126 switch (Res.first) {
5128 case X86::AX: DestReg = X86::EAX; break;
5129 case X86::DX: DestReg = X86::EDX; break;
5130 case X86::CX: DestReg = X86::ECX; break;
5131 case X86::BX: DestReg = X86::EBX; break;
5132 case X86::SI: DestReg = X86::ESI; break;
5133 case X86::DI: DestReg = X86::EDI; break;
5134 case X86::BP: DestReg = X86::EBP; break;
5135 case X86::SP: DestReg = X86::ESP; break;
5138 Res.first = DestReg;
5139 Res.second = Res.second = X86::GR32RegisterClass;
5141 } else if (VT == MVT::i64) {
5142 unsigned DestReg = 0;
5143 switch (Res.first) {
5145 case X86::AX: DestReg = X86::RAX; break;
5146 case X86::DX: DestReg = X86::RDX; break;
5147 case X86::CX: DestReg = X86::RCX; break;
5148 case X86::BX: DestReg = X86::RBX; break;
5149 case X86::SI: DestReg = X86::RSI; break;
5150 case X86::DI: DestReg = X86::RDI; break;
5151 case X86::BP: DestReg = X86::RBP; break;
5152 case X86::SP: DestReg = X86::RSP; break;
5155 Res.first = DestReg;
5156 Res.second = Res.second = X86::GR64RegisterClass;