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,
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.
459 else if (CC == CallingConv::Fast && Subtarget->hasSSE2())
460 Reg = X86::XMM0; // FP values in X86-32 with fastcc go in XMM0.
462 Reg = X86::ST0; // FP values in X86-32 go in ST0.
465 assert(MVT::isVector(ArgVT) && "Unknown return value type!");
466 Reg = X86::XMM0; // Int/FP vector result -> XMM0.
472 /// LowerRET - Lower an ISD::RET node.
473 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
474 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
476 // Support up returning up to two registers.
477 MVT::ValueType VTs[2];
478 unsigned DestRegs[2];
479 unsigned NumRegs = Op.getNumOperands() / 2;
480 assert(NumRegs <= 2 && "Can only return up to two regs!");
482 for (unsigned i = 0; i != NumRegs; ++i)
483 VTs[i] = Op.getOperand(i*2+1).getValueType();
485 // Determine which register each value should be copied into.
486 GetRetValueLocs(VTs, NumRegs, DestRegs, Subtarget,
487 DAG.getMachineFunction().getFunction()->getCallingConv());
489 // If this is the first return lowered for this function, add the regs to the
490 // liveout set for the function.
491 if (DAG.getMachineFunction().liveout_empty()) {
492 for (unsigned i = 0; i != NumRegs; ++i)
493 DAG.getMachineFunction().addLiveOut(DestRegs[i]);
496 SDOperand Chain = Op.getOperand(0);
499 // Copy the result values into the output registers.
500 if (NumRegs != 1 || DestRegs[0] != X86::ST0) {
501 for (unsigned i = 0; i != NumRegs; ++i) {
502 Chain = DAG.getCopyToReg(Chain, DestRegs[i], Op.getOperand(i*2+1), Flag);
503 Flag = Chain.getValue(1);
506 // We need to handle a destination of ST0 specially, because it isn't really
508 SDOperand Value = Op.getOperand(1);
510 // If this is an FP return with ScalarSSE, we need to move the value from
511 // an XMM register onto the fp-stack.
515 // If this is a load into a scalarsse value, don't store the loaded value
516 // back to the stack, only to reload it: just replace the scalar-sse load.
517 if (ISD::isNON_EXTLoad(Value.Val) &&
518 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
519 Chain = Value.getOperand(0);
520 MemLoc = Value.getOperand(1);
522 // Spill the value to memory and reload it into top of stack.
523 unsigned Size = MVT::getSizeInBits(VTs[0])/8;
524 MachineFunction &MF = DAG.getMachineFunction();
525 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
526 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
527 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
529 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
530 SDOperand Ops[] = { Chain, MemLoc, DAG.getValueType(VTs[0]) };
531 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
532 Chain = Value.getValue(1);
535 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
536 SDOperand Ops[] = { Chain, Value };
537 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
538 Flag = Chain.getValue(1);
541 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
543 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
545 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
549 /// LowerCallResult - Lower the result values of an ISD::CALL into the
550 /// appropriate copies out of appropriate physical registers. This assumes that
551 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
552 /// being lowered. The returns a SDNode with the same number of values as the
554 SDNode *X86TargetLowering::
555 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
556 unsigned CallingConv, SelectionDAG &DAG) {
557 SmallVector<SDOperand, 8> ResultVals;
559 // We support returning up to two registers.
560 MVT::ValueType VTs[2];
561 unsigned DestRegs[2];
562 unsigned NumRegs = TheCall->getNumValues() - 1;
563 assert(NumRegs <= 2 && "Can only return up to two regs!");
565 for (unsigned i = 0; i != NumRegs; ++i)
566 VTs[i] = TheCall->getValueType(i);
568 // Determine which register each value should be copied into.
569 GetRetValueLocs(VTs, NumRegs, DestRegs, Subtarget, CallingConv);
571 // Copy all of the result registers out of their specified physreg.
572 if (NumRegs != 1 || DestRegs[0] != X86::ST0) {
573 for (unsigned i = 0; i != NumRegs; ++i) {
574 Chain = DAG.getCopyFromReg(Chain, DestRegs[i], VTs[i],
576 InFlag = Chain.getValue(2);
577 ResultVals.push_back(Chain.getValue(0));
580 // Copies from the FP stack are special, as ST0 isn't a valid register
581 // before the fp stackifier runs.
583 // Copy ST0 into an RFP register with FP_GET_RESULT.
584 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
585 SDOperand GROps[] = { Chain, InFlag };
586 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
587 Chain = RetVal.getValue(1);
588 InFlag = RetVal.getValue(2);
590 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
593 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
594 // shouldn't be necessary except that RFP cannot be live across
595 // multiple blocks. When stackifier is fixed, they can be uncoupled.
596 MachineFunction &MF = DAG.getMachineFunction();
597 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
598 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
600 Chain, RetVal, StackSlot, DAG.getValueType(VTs[0]), InFlag
602 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
603 RetVal = DAG.getLoad(VTs[0], Chain, StackSlot, NULL, 0);
604 Chain = RetVal.getValue(1);
607 if (VTs[0] == MVT::f32 && !X86ScalarSSE)
608 // FIXME: we would really like to remember that this FP_ROUND
609 // operation is okay to eliminate if we allow excess FP precision.
610 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
611 ResultVals.push_back(RetVal);
614 // Merge everything together with a MERGE_VALUES node.
615 ResultVals.push_back(Chain);
616 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
617 &ResultVals[0], ResultVals.size()).Val;
621 //===----------------------------------------------------------------------===//
622 // C & StdCall Calling Convention implementation
623 //===----------------------------------------------------------------------===//
624 // StdCall calling convention seems to be standard for many Windows' API
625 // routines and around. It differs from C calling convention just a little:
626 // callee should clean up the stack, not caller. Symbols should be also
627 // decorated in some fancy way :) It doesn't support any vector arguments.
629 /// AddLiveIn - This helper function adds the specified physical register to the
630 /// MachineFunction as a live in value. It also creates a corresponding virtual
632 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
633 const TargetRegisterClass *RC) {
634 assert(RC->contains(PReg) && "Not the correct regclass!");
635 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
636 MF.addLiveIn(PReg, VReg);
640 /// HowToPassArgument - Returns how an formal argument of the specified type
641 /// should be passed. If it is through stack, returns the size of the stack
642 /// slot; if it is through integer or XMM register, returns the number of
643 /// integer or XMM registers are needed.
645 HowToPassCallArgument(MVT::ValueType ObjectVT,
647 unsigned NumIntRegs, unsigned NumXMMRegs,
648 unsigned MaxNumIntRegs,
649 unsigned &ObjSize, unsigned &ObjIntRegs,
650 unsigned &ObjXMMRegs) {
655 if (MaxNumIntRegs>3) {
656 // We don't have too much registers on ia32! :)
661 default: assert(0 && "Unhandled argument type!");
663 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
669 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
675 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
681 if (ArgInReg && (NumIntRegs+2 <= MaxNumIntRegs)) {
683 } else if (ArgInReg && (NumIntRegs+1 <= MaxNumIntRegs)) {
708 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
710 unsigned NumArgs = Op.Val->getNumValues() - 1;
711 MachineFunction &MF = DAG.getMachineFunction();
712 MachineFrameInfo *MFI = MF.getFrameInfo();
713 SDOperand Root = Op.getOperand(0);
714 SmallVector<SDOperand, 8> ArgValues;
715 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
717 // Add DAG nodes to load the arguments... On entry to a function on the X86,
718 // the stack frame looks like this:
720 // [ESP] -- return address
721 // [ESP + 4] -- first argument (leftmost lexically)
722 // [ESP + 8] -- second argument, if first argument is <= 4 bytes in size
725 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
726 unsigned NumSRetBytes= 0; // How much bytes on stack used for struct return
727 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
728 unsigned NumIntRegs = 0; // Integer regs used for parameter passing
730 static const unsigned XMMArgRegs[] = {
731 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
733 static const unsigned GPRArgRegs[][3] = {
734 { X86::AL, X86::DL, X86::CL },
735 { X86::AX, X86::DX, X86::CX },
736 { X86::EAX, X86::EDX, X86::ECX }
738 static const TargetRegisterClass* GPRClasses[3] = {
739 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
742 // Handle regparm attribute
743 SmallVector<bool, 8> ArgInRegs(NumArgs, false);
744 SmallVector<bool, 8> SRetArgs(NumArgs, false);
746 for (unsigned i = 0; i<NumArgs; ++i) {
747 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3+i))->getValue();
748 ArgInRegs[i] = (Flags >> 1) & 1;
749 SRetArgs[i] = (Flags >> 2) & 1;
753 for (unsigned i = 0; i < NumArgs; ++i) {
754 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
755 unsigned ArgIncrement = 4;
756 unsigned ObjSize = 0;
757 unsigned ObjXMMRegs = 0;
758 unsigned ObjIntRegs = 0;
762 HowToPassCallArgument(ObjectVT,
764 NumIntRegs, NumXMMRegs, 3,
765 ObjSize, ObjIntRegs, ObjXMMRegs);
768 ArgIncrement = ObjSize;
770 if (ObjIntRegs || ObjXMMRegs) {
772 default: assert(0 && "Unhandled argument type!");
776 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][NumIntRegs];
777 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
778 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
787 assert(!isStdCall && "Unhandled argument type!");
788 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
789 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
792 NumIntRegs += ObjIntRegs;
793 NumXMMRegs += ObjXMMRegs;
796 // XMM arguments have to be aligned on 16-byte boundary.
798 ArgOffset = ((ArgOffset + 15) / 16) * 16;
799 // Create the SelectionDAG nodes corresponding to a load from this
801 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
802 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
803 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
805 ArgOffset += ArgIncrement; // Move on to the next argument.
807 NumSRetBytes += ArgIncrement;
810 ArgValues.push_back(ArgValue);
813 ArgValues.push_back(Root);
815 // If the function takes variable number of arguments, make a frame index for
816 // the start of the first vararg value... for expansion of llvm.va_start.
818 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
820 if (isStdCall && !isVarArg) {
821 BytesToPopOnReturn = ArgOffset; // Callee pops everything..
822 BytesCallerReserves = 0;
824 BytesToPopOnReturn = NumSRetBytes; // Callee pops hidden struct pointer.
825 BytesCallerReserves = ArgOffset;
828 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
829 ReturnAddrIndex = 0; // No return address slot generated yet.
832 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
834 // Return the new list of results.
835 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
836 &ArgValues[0], ArgValues.size());
839 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
841 SDOperand Chain = Op.getOperand(0);
842 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
843 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
844 SDOperand Callee = Op.getOperand(4);
845 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
847 static const unsigned XMMArgRegs[] = {
848 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
850 static const unsigned GPR32ArgRegs[] = {
851 X86::EAX, X86::EDX, X86::ECX
854 // Count how many bytes are to be pushed on the stack.
855 unsigned NumBytes = 0;
856 // Keep track of the number of integer regs passed so far.
857 unsigned NumIntRegs = 0;
858 // Keep track of the number of XMM regs passed so far.
859 unsigned NumXMMRegs = 0;
860 // How much bytes on stack used for struct return
861 unsigned NumSRetBytes= 0;
863 // Handle regparm attribute
864 SmallVector<bool, 8> ArgInRegs(NumOps, false);
865 SmallVector<bool, 8> SRetArgs(NumOps, false);
866 for (unsigned i = 0; i<NumOps; ++i) {
868 dyn_cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
869 ArgInRegs[i] = (Flags >> 1) & 1;
870 SRetArgs[i] = (Flags >> 2) & 1;
873 // Calculate stack frame size
874 for (unsigned i = 0; i != NumOps; ++i) {
875 SDOperand Arg = Op.getOperand(5+2*i);
876 unsigned ArgIncrement = 4;
877 unsigned ObjSize = 0;
878 unsigned ObjIntRegs = 0;
879 unsigned ObjXMMRegs = 0;
881 HowToPassCallArgument(Arg.getValueType(),
883 NumIntRegs, NumXMMRegs, 3,
884 ObjSize, ObjIntRegs, ObjXMMRegs);
886 ArgIncrement = ObjSize;
888 NumIntRegs += ObjIntRegs;
889 NumXMMRegs += ObjXMMRegs;
891 // XMM arguments have to be aligned on 16-byte boundary.
893 NumBytes = ((NumBytes + 15) / 16) * 16;
894 NumBytes += ArgIncrement;
898 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
900 // Arguments go on the stack in reverse order, as specified by the ABI.
901 unsigned ArgOffset = 0;
904 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
905 SmallVector<SDOperand, 8> MemOpChains;
906 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
907 for (unsigned i = 0; i != NumOps; ++i) {
908 SDOperand Arg = Op.getOperand(5+2*i);
909 unsigned ArgIncrement = 4;
910 unsigned ObjSize = 0;
911 unsigned ObjIntRegs = 0;
912 unsigned ObjXMMRegs = 0;
914 HowToPassCallArgument(Arg.getValueType(),
916 NumIntRegs, NumXMMRegs, 3,
917 ObjSize, ObjIntRegs, ObjXMMRegs);
920 ArgIncrement = ObjSize;
922 if (Arg.getValueType() == MVT::i8 || Arg.getValueType() == MVT::i16) {
923 // Promote the integer to 32 bits. If the input type is signed use a
924 // sign extend, otherwise use a zero extend.
925 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
927 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
928 Arg = DAG.getNode(ExtOp, MVT::i32, Arg);
931 if (ObjIntRegs || ObjXMMRegs) {
932 switch (Arg.getValueType()) {
933 default: assert(0 && "Unhandled argument type!");
935 RegsToPass.push_back(std::make_pair(GPR32ArgRegs[NumIntRegs], Arg));
943 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
947 NumIntRegs += ObjIntRegs;
948 NumXMMRegs += ObjXMMRegs;
951 // XMM arguments have to be aligned on 16-byte boundary.
953 ArgOffset = ((ArgOffset + 15) / 16) * 16;
955 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
956 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
957 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
959 ArgOffset += ArgIncrement; // Move on to the next argument.
961 NumSRetBytes += ArgIncrement;
965 // Sanity check: we haven't seen NumSRetBytes > 4
966 assert((NumSRetBytes<=4) &&
967 "Too much space for struct-return pointer requested");
969 if (!MemOpChains.empty())
970 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
971 &MemOpChains[0], MemOpChains.size());
973 // Build a sequence of copy-to-reg nodes chained together with token chain
974 // and flag operands which copy the outgoing args into registers.
976 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
977 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
979 InFlag = Chain.getValue(1);
982 // ELF / PIC requires GOT in the EBX register before function calls via PLT
984 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
985 Subtarget->isPICStyleGOT()) {
986 Chain = DAG.getCopyToReg(Chain, X86::EBX,
987 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
989 InFlag = Chain.getValue(1);
992 // If the callee is a GlobalAddress node (quite common, every direct call is)
993 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
994 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
995 // We should use extra load for direct calls to dllimported functions in
997 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
998 getTargetMachine(), true))
999 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1000 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1001 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1003 // Returns a chain & a flag for retval copy to use.
1004 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1005 SmallVector<SDOperand, 8> Ops;
1006 Ops.push_back(Chain);
1007 Ops.push_back(Callee);
1009 // Add argument registers to the end of the list so that they are known live
1011 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1012 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1013 RegsToPass[i].second.getValueType()));
1015 // Add an implicit use GOT pointer in EBX.
1016 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1017 Subtarget->isPICStyleGOT())
1018 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1021 Ops.push_back(InFlag);
1023 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1024 NodeTys, &Ops[0], Ops.size());
1025 InFlag = Chain.getValue(1);
1027 // Create the CALLSEQ_END node.
1028 unsigned NumBytesForCalleeToPush = 0;
1030 if (CC == CallingConv::X86_StdCall) {
1032 NumBytesForCalleeToPush = NumSRetBytes;
1034 NumBytesForCalleeToPush = NumBytes;
1036 // If this is is a call to a struct-return function, the callee
1037 // pops the hidden struct pointer, so we have to push it back.
1038 // This is common for Darwin/X86, Linux & Mingw32 targets.
1039 NumBytesForCalleeToPush = NumSRetBytes;
1042 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1044 Ops.push_back(Chain);
1045 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1046 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
1047 Ops.push_back(InFlag);
1048 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1049 InFlag = Chain.getValue(1);
1051 // Handle result values, copying them out of physregs into vregs that we
1053 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1057 //===----------------------------------------------------------------------===//
1058 // X86-64 C Calling Convention implementation
1059 //===----------------------------------------------------------------------===//
1061 /// HowToPassX86_64CCCArgument - Returns how an formal argument of the specified
1062 /// type should be passed. If it is through stack, returns the size of the stack
1063 /// slot; if it is through integer or XMM register, returns the number of
1064 /// integer or XMM registers are needed.
1066 HowToPassX86_64CCCArgument(MVT::ValueType ObjectVT,
1067 unsigned NumIntRegs, unsigned NumXMMRegs,
1068 unsigned &ObjSize, unsigned &ObjIntRegs,
1069 unsigned &ObjXMMRegs) {
1075 default: assert(0 && "Unhandled argument type!");
1085 case MVT::i8: ObjSize = 1; break;
1086 case MVT::i16: ObjSize = 2; break;
1087 case MVT::i32: ObjSize = 4; break;
1088 case MVT::i64: ObjSize = 8; break;
1105 case MVT::f32: ObjSize = 4; break;
1106 case MVT::f64: ObjSize = 8; break;
1112 case MVT::v2f64: ObjSize = 16; break;
1120 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1121 unsigned NumArgs = Op.Val->getNumValues() - 1;
1122 MachineFunction &MF = DAG.getMachineFunction();
1123 MachineFrameInfo *MFI = MF.getFrameInfo();
1124 SDOperand Root = Op.getOperand(0);
1125 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1126 SmallVector<SDOperand, 8> ArgValues;
1128 // Add DAG nodes to load the arguments... On entry to a function on the X86,
1129 // the stack frame looks like this:
1131 // [RSP] -- return address
1132 // [RSP + 8] -- first nonreg argument (leftmost lexically)
1133 // [RSP +16] -- second nonreg argument, if 1st argument is <= 8 bytes in size
1136 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1137 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1138 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1140 static const unsigned GPR8ArgRegs[] = {
1141 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1143 static const unsigned GPR16ArgRegs[] = {
1144 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1146 static const unsigned GPR32ArgRegs[] = {
1147 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1149 static const unsigned GPR64ArgRegs[] = {
1150 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1152 static const unsigned XMMArgRegs[] = {
1153 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1154 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1157 for (unsigned i = 0; i < NumArgs; ++i) {
1158 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1159 unsigned ArgIncrement = 8;
1160 unsigned ObjSize = 0;
1161 unsigned ObjIntRegs = 0;
1162 unsigned ObjXMMRegs = 0;
1164 // FIXME: __int128 and long double support?
1165 HowToPassX86_64CCCArgument(ObjectVT, NumIntRegs, NumXMMRegs,
1166 ObjSize, ObjIntRegs, ObjXMMRegs);
1168 ArgIncrement = ObjSize;
1172 if (ObjIntRegs || ObjXMMRegs) {
1174 default: assert(0 && "Unhandled argument type!");
1179 TargetRegisterClass *RC = NULL;
1183 RC = X86::GR8RegisterClass;
1184 Reg = GPR8ArgRegs[NumIntRegs];
1187 RC = X86::GR16RegisterClass;
1188 Reg = GPR16ArgRegs[NumIntRegs];
1191 RC = X86::GR32RegisterClass;
1192 Reg = GPR32ArgRegs[NumIntRegs];
1195 RC = X86::GR64RegisterClass;
1196 Reg = GPR64ArgRegs[NumIntRegs];
1199 Reg = AddLiveIn(MF, Reg, RC);
1200 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1211 TargetRegisterClass *RC= (ObjectVT == MVT::f32) ?
1212 X86::FR32RegisterClass : ((ObjectVT == MVT::f64) ?
1213 X86::FR64RegisterClass : X86::VR128RegisterClass);
1214 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], RC);
1215 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1219 NumIntRegs += ObjIntRegs;
1220 NumXMMRegs += ObjXMMRegs;
1221 } else if (ObjSize) {
1222 // XMM arguments have to be aligned on 16-byte boundary.
1224 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1225 // Create the SelectionDAG nodes corresponding to a load from this
1227 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1228 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1229 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1230 ArgOffset += ArgIncrement; // Move on to the next argument.
1233 ArgValues.push_back(ArgValue);
1236 // If the function takes variable number of arguments, make a frame index for
1237 // the start of the first vararg value... for expansion of llvm.va_start.
1239 // For X86-64, if there are vararg parameters that are passed via
1240 // registers, then we must store them to their spots on the stack so they
1241 // may be loaded by deferencing the result of va_next.
1242 VarArgsGPOffset = NumIntRegs * 8;
1243 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1244 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
1245 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1247 // Store the integer parameter registers.
1248 SmallVector<SDOperand, 8> MemOps;
1249 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1250 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1251 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1252 for (; NumIntRegs != 6; ++NumIntRegs) {
1253 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1254 X86::GR64RegisterClass);
1255 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1256 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1257 MemOps.push_back(Store);
1258 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1259 DAG.getConstant(8, getPointerTy()));
1262 // Now store the XMM (fp + vector) parameter registers.
1263 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1264 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1265 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1266 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1267 X86::VR128RegisterClass);
1268 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1269 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1270 MemOps.push_back(Store);
1271 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1272 DAG.getConstant(16, getPointerTy()));
1274 if (!MemOps.empty())
1275 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1276 &MemOps[0], MemOps.size());
1279 ArgValues.push_back(Root);
1281 ReturnAddrIndex = 0; // No return address slot generated yet.
1282 BytesToPopOnReturn = 0; // Callee pops nothing.
1283 BytesCallerReserves = ArgOffset;
1285 // Return the new list of results.
1286 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1287 &ArgValues[0], ArgValues.size());
1291 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1293 SDOperand Chain = Op.getOperand(0);
1294 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1295 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1296 SDOperand Callee = Op.getOperand(4);
1297 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1299 // Count how many bytes are to be pushed on the stack.
1300 unsigned NumBytes = 0;
1301 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1302 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1304 static const unsigned GPR8ArgRegs[] = {
1305 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1307 static const unsigned GPR16ArgRegs[] = {
1308 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1310 static const unsigned GPR32ArgRegs[] = {
1311 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1313 static const unsigned GPR64ArgRegs[] = {
1314 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1316 static const unsigned XMMArgRegs[] = {
1317 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1318 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1321 for (unsigned i = 0; i != NumOps; ++i) {
1322 SDOperand Arg = Op.getOperand(5+2*i);
1323 MVT::ValueType ArgVT = Arg.getValueType();
1326 default: assert(0 && "Unknown value type!");
1346 else if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1349 // XMM arguments have to be aligned on 16-byte boundary.
1350 NumBytes = ((NumBytes + 15) / 16) * 16;
1357 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1359 // Arguments go on the stack in reverse order, as specified by the ABI.
1360 unsigned ArgOffset = 0;
1363 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1364 SmallVector<SDOperand, 8> MemOpChains;
1365 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1366 for (unsigned i = 0; i != NumOps; ++i) {
1367 SDOperand Arg = Op.getOperand(5+2*i);
1368 MVT::ValueType ArgVT = Arg.getValueType();
1371 default: assert(0 && "Unexpected ValueType for argument!");
1376 if (NumIntRegs < 6) {
1380 case MVT::i8: Reg = GPR8ArgRegs[NumIntRegs]; break;
1381 case MVT::i16: Reg = GPR16ArgRegs[NumIntRegs]; break;
1382 case MVT::i32: Reg = GPR32ArgRegs[NumIntRegs]; break;
1383 case MVT::i64: Reg = GPR64ArgRegs[NumIntRegs]; break;
1385 RegsToPass.push_back(std::make_pair(Reg, Arg));
1388 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1389 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1390 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1402 if (NumXMMRegs < 8) {
1403 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1406 if (ArgVT != MVT::f32 && ArgVT != MVT::f64) {
1407 // XMM arguments have to be aligned on 16-byte boundary.
1408 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1410 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1411 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1412 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1413 if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1421 if (!MemOpChains.empty())
1422 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1423 &MemOpChains[0], MemOpChains.size());
1425 // Build a sequence of copy-to-reg nodes chained together with token chain
1426 // and flag operands which copy the outgoing args into registers.
1428 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1429 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1431 InFlag = Chain.getValue(1);
1435 // From AMD64 ABI document:
1436 // For calls that may call functions that use varargs or stdargs
1437 // (prototype-less calls or calls to functions containing ellipsis (...) in
1438 // the declaration) %al is used as hidden argument to specify the number
1439 // of SSE registers used. The contents of %al do not need to match exactly
1440 // the number of registers, but must be an ubound on the number of SSE
1441 // registers used and is in the range 0 - 8 inclusive.
1442 Chain = DAG.getCopyToReg(Chain, X86::AL,
1443 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1444 InFlag = Chain.getValue(1);
1447 // If the callee is a GlobalAddress node (quite common, every direct call is)
1448 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1449 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1450 // We should use extra load for direct calls to dllimported functions in
1452 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1453 getTargetMachine(), true))
1454 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1455 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1456 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1458 // Returns a chain & a flag for retval copy to use.
1459 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1460 SmallVector<SDOperand, 8> Ops;
1461 Ops.push_back(Chain);
1462 Ops.push_back(Callee);
1464 // Add argument registers to the end of the list so that they are known live
1466 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1467 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1468 RegsToPass[i].second.getValueType()));
1471 Ops.push_back(InFlag);
1473 // FIXME: Do not generate X86ISD::TAILCALL for now.
1474 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1475 NodeTys, &Ops[0], Ops.size());
1476 InFlag = Chain.getValue(1);
1478 // Returns a flag for retval copy to use.
1479 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1481 Ops.push_back(Chain);
1482 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1483 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1484 Ops.push_back(InFlag);
1485 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1486 InFlag = Chain.getValue(1);
1488 // Handle result values, copying them out of physregs into vregs that we
1490 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1493 //===----------------------------------------------------------------------===//
1494 // Fast & FastCall Calling Convention implementation
1495 //===----------------------------------------------------------------------===//
1497 // The X86 'fast' calling convention passes up to two integer arguments in
1498 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
1499 // and requires that the callee pop its arguments off the stack (allowing proper
1500 // tail calls), and has the same return value conventions as C calling convs.
1502 // This calling convention always arranges for the callee pop value to be 8n+4
1503 // bytes, which is needed for tail recursion elimination and stack alignment
1506 // Note that this can be enhanced in the future to pass fp vals in registers
1507 // (when we have a global fp allocator) and do other tricks.
1509 //===----------------------------------------------------------------------===//
1510 // The X86 'fastcall' calling convention passes up to two integer arguments in
1511 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1512 // and requires that the callee pop its arguments off the stack (allowing proper
1513 // tail calls), and has the same return value conventions as C calling convs.
1515 // This calling convention always arranges for the callee pop value to be 8n+4
1516 // bytes, which is needed for tail recursion elimination and stack alignment
1521 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG,
1523 unsigned NumArgs = Op.Val->getNumValues()-1;
1524 MachineFunction &MF = DAG.getMachineFunction();
1525 MachineFrameInfo *MFI = MF.getFrameInfo();
1526 SDOperand Root = Op.getOperand(0);
1527 SmallVector<SDOperand, 8> ArgValues;
1529 // Add DAG nodes to load the arguments... On entry to a function the stack
1530 // frame looks like this:
1532 // [ESP] -- return address
1533 // [ESP + 4] -- first nonreg argument (leftmost lexically)
1534 // [ESP + 8] -- second nonreg argument, if 1st argument is <= 4 bytes in size
1536 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1538 // Keep track of the number of integer regs passed so far. This can be either
1539 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1541 unsigned NumIntRegs = 0;
1542 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1544 static const unsigned XMMArgRegs[] = {
1545 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1548 static const unsigned GPRArgRegs[][2][2] = {
1549 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1550 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1551 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1554 static const TargetRegisterClass* GPRClasses[3] = {
1555 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
1558 unsigned GPRInd = (isFastCall ? 1 : 0);
1559 for (unsigned i = 0; i < NumArgs; ++i) {
1560 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1561 unsigned ArgIncrement = 4;
1562 unsigned ObjSize = 0;
1563 unsigned ObjXMMRegs = 0;
1564 unsigned ObjIntRegs = 0;
1568 HowToPassCallArgument(ObjectVT,
1569 true, // Use as much registers as possible
1570 NumIntRegs, NumXMMRegs,
1571 (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS),
1572 ObjSize, ObjIntRegs, ObjXMMRegs);
1575 ArgIncrement = ObjSize;
1577 if (ObjIntRegs || ObjXMMRegs) {
1579 default: assert(0 && "Unhandled argument type!");
1583 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][GPRInd][NumIntRegs];
1584 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
1585 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1594 assert(!isFastCall && "Unhandled argument type!");
1595 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
1596 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1600 NumIntRegs += ObjIntRegs;
1601 NumXMMRegs += ObjXMMRegs;
1604 // XMM arguments have to be aligned on 16-byte boundary.
1606 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1607 // Create the SelectionDAG nodes corresponding to a load from this
1609 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1610 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1611 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1613 ArgOffset += ArgIncrement; // Move on to the next argument.
1616 ArgValues.push_back(ArgValue);
1619 ArgValues.push_back(Root);
1621 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1622 // arguments and the arguments after the retaddr has been pushed are aligned.
1623 if ((ArgOffset & 7) == 0)
1626 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1627 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1628 ReturnAddrIndex = 0; // No return address slot generated yet.
1629 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
1630 BytesCallerReserves = 0;
1632 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
1634 // Finally, inform the code generator which regs we return values in.
1635 switch (getValueType(MF.getFunction()->getReturnType())) {
1636 default: assert(0 && "Unknown type!");
1637 case MVT::isVoid: break;
1642 MF.addLiveOut(X86::EAX);
1645 MF.addLiveOut(X86::EAX);
1646 MF.addLiveOut(X86::EDX);
1650 MF.addLiveOut(X86::ST0);
1658 assert(!isFastCall && "Unknown result type");
1659 MF.addLiveOut(X86::XMM0);
1663 // Return the new list of results.
1664 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1665 &ArgValues[0], ArgValues.size());
1668 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1670 SDOperand Chain = Op.getOperand(0);
1671 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1672 SDOperand Callee = Op.getOperand(4);
1673 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1675 // Count how many bytes are to be pushed on the stack.
1676 unsigned NumBytes = 0;
1678 // Keep track of the number of integer regs passed so far. This can be either
1679 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1681 unsigned NumIntRegs = 0;
1682 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1684 static const unsigned GPRArgRegs[][2][2] = {
1685 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1686 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1687 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1689 static const unsigned XMMArgRegs[] = {
1690 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1693 bool isFastCall = CC == CallingConv::X86_FastCall;
1694 unsigned GPRInd = isFastCall ? 1 : 0;
1695 for (unsigned i = 0; i != NumOps; ++i) {
1696 SDOperand Arg = Op.getOperand(5+2*i);
1698 switch (Arg.getValueType()) {
1699 default: assert(0 && "Unknown value type!");
1703 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1704 if (NumIntRegs < MaxNumIntRegs) {
1721 assert(!isFastCall && "Unknown value type!");
1725 // XMM arguments have to be aligned on 16-byte boundary.
1726 NumBytes = ((NumBytes + 15) / 16) * 16;
1733 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1734 // arguments and the arguments after the retaddr has been pushed are aligned.
1735 if ((NumBytes & 7) == 0)
1738 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1740 // Arguments go on the stack in reverse order, as specified by the ABI.
1741 unsigned ArgOffset = 0;
1743 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1744 SmallVector<SDOperand, 8> MemOpChains;
1745 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1746 for (unsigned i = 0; i != NumOps; ++i) {
1747 SDOperand Arg = Op.getOperand(5+2*i);
1749 switch (Arg.getValueType()) {
1750 default: assert(0 && "Unexpected ValueType for argument!");
1754 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1755 if (NumIntRegs < MaxNumIntRegs) {
1757 GPRArgRegs[Arg.getValueType()-MVT::i8][GPRInd][NumIntRegs];
1758 RegsToPass.push_back(std::make_pair(RegToUse, Arg));
1764 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1765 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1766 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1771 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1772 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1773 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1783 assert(!isFastCall && "Unexpected ValueType for argument!");
1784 if (NumXMMRegs < 4) {
1785 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1788 // XMM arguments have to be aligned on 16-byte boundary.
1789 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1790 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1791 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1792 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1799 if (!MemOpChains.empty())
1800 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1801 &MemOpChains[0], MemOpChains.size());
1803 // Build a sequence of copy-to-reg nodes chained together with token chain
1804 // and flag operands which copy the outgoing args into registers.
1806 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1807 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1809 InFlag = Chain.getValue(1);
1812 // If the callee is a GlobalAddress node (quite common, every direct call is)
1813 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1814 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1815 // We should use extra load for direct calls to dllimported functions in
1817 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1818 getTargetMachine(), true))
1819 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1820 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1821 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1823 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1825 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1826 Subtarget->isPICStyleGOT()) {
1827 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1828 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1830 InFlag = Chain.getValue(1);
1833 // Returns a chain & a flag for retval copy to use.
1834 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1835 SmallVector<SDOperand, 8> Ops;
1836 Ops.push_back(Chain);
1837 Ops.push_back(Callee);
1839 // Add argument registers to the end of the list so that they are known live
1841 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1842 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1843 RegsToPass[i].second.getValueType()));
1845 // Add an implicit use GOT pointer in EBX.
1846 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1847 Subtarget->isPICStyleGOT())
1848 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1851 Ops.push_back(InFlag);
1853 // FIXME: Do not generate X86ISD::TAILCALL for now.
1854 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1855 NodeTys, &Ops[0], Ops.size());
1856 InFlag = Chain.getValue(1);
1858 // Returns a flag for retval copy to use.
1859 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1861 Ops.push_back(Chain);
1862 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1863 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1864 Ops.push_back(InFlag);
1865 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1866 InFlag = Chain.getValue(1);
1868 // Handle result values, copying them out of physregs into vregs that we
1870 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1873 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1874 if (ReturnAddrIndex == 0) {
1875 // Set up a frame object for the return address.
1876 MachineFunction &MF = DAG.getMachineFunction();
1877 if (Subtarget->is64Bit())
1878 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1880 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1883 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1888 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1889 /// specific condition code. It returns a false if it cannot do a direct
1890 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1892 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1893 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1894 SelectionDAG &DAG) {
1895 X86CC = X86::COND_INVALID;
1897 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1898 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1899 // X > -1 -> X == 0, jump !sign.
1900 RHS = DAG.getConstant(0, RHS.getValueType());
1901 X86CC = X86::COND_NS;
1903 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1904 // X < 0 -> X == 0, jump on sign.
1905 X86CC = X86::COND_S;
1910 switch (SetCCOpcode) {
1912 case ISD::SETEQ: X86CC = X86::COND_E; break;
1913 case ISD::SETGT: X86CC = X86::COND_G; break;
1914 case ISD::SETGE: X86CC = X86::COND_GE; break;
1915 case ISD::SETLT: X86CC = X86::COND_L; break;
1916 case ISD::SETLE: X86CC = X86::COND_LE; break;
1917 case ISD::SETNE: X86CC = X86::COND_NE; break;
1918 case ISD::SETULT: X86CC = X86::COND_B; break;
1919 case ISD::SETUGT: X86CC = X86::COND_A; break;
1920 case ISD::SETULE: X86CC = X86::COND_BE; break;
1921 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1924 // On a floating point condition, the flags are set as follows:
1926 // 0 | 0 | 0 | X > Y
1927 // 0 | 0 | 1 | X < Y
1928 // 1 | 0 | 0 | X == Y
1929 // 1 | 1 | 1 | unordered
1931 switch (SetCCOpcode) {
1934 case ISD::SETEQ: X86CC = X86::COND_E; break;
1935 case ISD::SETOLT: Flip = true; // Fallthrough
1937 case ISD::SETGT: X86CC = X86::COND_A; break;
1938 case ISD::SETOLE: Flip = true; // Fallthrough
1940 case ISD::SETGE: X86CC = X86::COND_AE; break;
1941 case ISD::SETUGT: Flip = true; // Fallthrough
1943 case ISD::SETLT: X86CC = X86::COND_B; break;
1944 case ISD::SETUGE: Flip = true; // Fallthrough
1946 case ISD::SETLE: X86CC = X86::COND_BE; break;
1948 case ISD::SETNE: X86CC = X86::COND_NE; break;
1949 case ISD::SETUO: X86CC = X86::COND_P; break;
1950 case ISD::SETO: X86CC = X86::COND_NP; break;
1953 std::swap(LHS, RHS);
1956 return X86CC != X86::COND_INVALID;
1959 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1960 /// code. Current x86 isa includes the following FP cmov instructions:
1961 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1962 static bool hasFPCMov(unsigned X86CC) {
1978 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1979 /// true if Op is undef or if its value falls within the specified range (L, H].
1980 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1981 if (Op.getOpcode() == ISD::UNDEF)
1984 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1985 return (Val >= Low && Val < Hi);
1988 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1989 /// true if Op is undef or if its value equal to the specified value.
1990 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1991 if (Op.getOpcode() == ISD::UNDEF)
1993 return cast<ConstantSDNode>(Op)->getValue() == Val;
1996 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1997 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1998 bool X86::isPSHUFDMask(SDNode *N) {
1999 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2001 if (N->getNumOperands() != 4)
2004 // Check if the value doesn't reference the second vector.
2005 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2006 SDOperand Arg = N->getOperand(i);
2007 if (Arg.getOpcode() == ISD::UNDEF) continue;
2008 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2009 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
2016 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2017 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2018 bool X86::isPSHUFHWMask(SDNode *N) {
2019 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2021 if (N->getNumOperands() != 8)
2024 // Lower quadword copied in order.
2025 for (unsigned i = 0; i != 4; ++i) {
2026 SDOperand Arg = N->getOperand(i);
2027 if (Arg.getOpcode() == ISD::UNDEF) continue;
2028 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2029 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2033 // Upper quadword shuffled.
2034 for (unsigned i = 4; i != 8; ++i) {
2035 SDOperand Arg = N->getOperand(i);
2036 if (Arg.getOpcode() == ISD::UNDEF) continue;
2037 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2038 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2039 if (Val < 4 || Val > 7)
2046 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2047 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2048 bool X86::isPSHUFLWMask(SDNode *N) {
2049 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2051 if (N->getNumOperands() != 8)
2054 // Upper quadword copied in order.
2055 for (unsigned i = 4; i != 8; ++i)
2056 if (!isUndefOrEqual(N->getOperand(i), i))
2059 // Lower quadword shuffled.
2060 for (unsigned i = 0; i != 4; ++i)
2061 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2067 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2068 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2069 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2070 if (NumElems != 2 && NumElems != 4) return false;
2072 unsigned Half = NumElems / 2;
2073 for (unsigned i = 0; i < Half; ++i)
2074 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2076 for (unsigned i = Half; i < NumElems; ++i)
2077 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2083 bool X86::isSHUFPMask(SDNode *N) {
2084 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2085 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2088 /// isCommutedSHUFP - Returns true if the shuffle mask is except
2089 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2090 /// half elements to come from vector 1 (which would equal the dest.) and
2091 /// the upper half to come from vector 2.
2092 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2093 if (NumOps != 2 && NumOps != 4) return false;
2095 unsigned Half = NumOps / 2;
2096 for (unsigned i = 0; i < Half; ++i)
2097 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2099 for (unsigned i = Half; i < NumOps; ++i)
2100 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2105 static bool isCommutedSHUFP(SDNode *N) {
2106 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2107 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2110 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2111 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2112 bool X86::isMOVHLPSMask(SDNode *N) {
2113 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2115 if (N->getNumOperands() != 4)
2118 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2119 return isUndefOrEqual(N->getOperand(0), 6) &&
2120 isUndefOrEqual(N->getOperand(1), 7) &&
2121 isUndefOrEqual(N->getOperand(2), 2) &&
2122 isUndefOrEqual(N->getOperand(3), 3);
2125 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2126 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2128 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2129 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2131 if (N->getNumOperands() != 4)
2134 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2135 return isUndefOrEqual(N->getOperand(0), 2) &&
2136 isUndefOrEqual(N->getOperand(1), 3) &&
2137 isUndefOrEqual(N->getOperand(2), 2) &&
2138 isUndefOrEqual(N->getOperand(3), 3);
2141 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2142 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2143 bool X86::isMOVLPMask(SDNode *N) {
2144 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2146 unsigned NumElems = N->getNumOperands();
2147 if (NumElems != 2 && NumElems != 4)
2150 for (unsigned i = 0; i < NumElems/2; ++i)
2151 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2154 for (unsigned i = NumElems/2; i < NumElems; ++i)
2155 if (!isUndefOrEqual(N->getOperand(i), i))
2161 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2162 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2164 bool X86::isMOVHPMask(SDNode *N) {
2165 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2167 unsigned NumElems = N->getNumOperands();
2168 if (NumElems != 2 && NumElems != 4)
2171 for (unsigned i = 0; i < NumElems/2; ++i)
2172 if (!isUndefOrEqual(N->getOperand(i), i))
2175 for (unsigned i = 0; i < NumElems/2; ++i) {
2176 SDOperand Arg = N->getOperand(i + NumElems/2);
2177 if (!isUndefOrEqual(Arg, i + NumElems))
2184 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2185 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2186 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2187 bool V2IsSplat = false) {
2188 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2191 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2192 SDOperand BitI = Elts[i];
2193 SDOperand BitI1 = Elts[i+1];
2194 if (!isUndefOrEqual(BitI, j))
2197 if (isUndefOrEqual(BitI1, NumElts))
2200 if (!isUndefOrEqual(BitI1, j + NumElts))
2208 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2209 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2210 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2213 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2214 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2215 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2216 bool V2IsSplat = false) {
2217 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2220 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2221 SDOperand BitI = Elts[i];
2222 SDOperand BitI1 = Elts[i+1];
2223 if (!isUndefOrEqual(BitI, j + NumElts/2))
2226 if (isUndefOrEqual(BitI1, NumElts))
2229 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2237 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2238 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2239 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2242 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2243 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2245 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2246 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2248 unsigned NumElems = N->getNumOperands();
2249 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
2252 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2253 SDOperand BitI = N->getOperand(i);
2254 SDOperand BitI1 = N->getOperand(i+1);
2256 if (!isUndefOrEqual(BitI, j))
2258 if (!isUndefOrEqual(BitI1, j))
2265 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2266 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2267 /// MOVSD, and MOVD, i.e. setting the lowest element.
2268 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2269 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2272 if (!isUndefOrEqual(Elts[0], NumElts))
2275 for (unsigned i = 1; i < NumElts; ++i) {
2276 if (!isUndefOrEqual(Elts[i], i))
2283 bool X86::isMOVLMask(SDNode *N) {
2284 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2285 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2288 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2289 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2290 /// element of vector 2 and the other elements to come from vector 1 in order.
2291 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2292 bool V2IsSplat = false,
2293 bool V2IsUndef = false) {
2294 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2297 if (!isUndefOrEqual(Ops[0], 0))
2300 for (unsigned i = 1; i < NumOps; ++i) {
2301 SDOperand Arg = Ops[i];
2302 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2303 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2304 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2311 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2312 bool V2IsUndef = false) {
2313 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2314 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2315 V2IsSplat, V2IsUndef);
2318 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2319 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2320 bool X86::isMOVSHDUPMask(SDNode *N) {
2321 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2323 if (N->getNumOperands() != 4)
2326 // Expect 1, 1, 3, 3
2327 for (unsigned i = 0; i < 2; ++i) {
2328 SDOperand Arg = N->getOperand(i);
2329 if (Arg.getOpcode() == ISD::UNDEF) continue;
2330 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2331 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2332 if (Val != 1) return false;
2336 for (unsigned i = 2; i < 4; ++i) {
2337 SDOperand Arg = N->getOperand(i);
2338 if (Arg.getOpcode() == ISD::UNDEF) continue;
2339 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2340 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2341 if (Val != 3) return false;
2345 // Don't use movshdup if it can be done with a shufps.
2349 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2350 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2351 bool X86::isMOVSLDUPMask(SDNode *N) {
2352 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2354 if (N->getNumOperands() != 4)
2357 // Expect 0, 0, 2, 2
2358 for (unsigned i = 0; i < 2; ++i) {
2359 SDOperand Arg = N->getOperand(i);
2360 if (Arg.getOpcode() == ISD::UNDEF) continue;
2361 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2362 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2363 if (Val != 0) return false;
2367 for (unsigned i = 2; i < 4; ++i) {
2368 SDOperand Arg = N->getOperand(i);
2369 if (Arg.getOpcode() == ISD::UNDEF) continue;
2370 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2371 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2372 if (Val != 2) return false;
2376 // Don't use movshdup if it can be done with a shufps.
2380 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2381 /// a splat of a single element.
2382 static bool isSplatMask(SDNode *N) {
2383 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2385 // This is a splat operation if each element of the permute is the same, and
2386 // if the value doesn't reference the second vector.
2387 unsigned NumElems = N->getNumOperands();
2388 SDOperand ElementBase;
2390 for (; i != NumElems; ++i) {
2391 SDOperand Elt = N->getOperand(i);
2392 if (isa<ConstantSDNode>(Elt)) {
2398 if (!ElementBase.Val)
2401 for (; i != NumElems; ++i) {
2402 SDOperand Arg = N->getOperand(i);
2403 if (Arg.getOpcode() == ISD::UNDEF) continue;
2404 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2405 if (Arg != ElementBase) return false;
2408 // Make sure it is a splat of the first vector operand.
2409 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2412 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2413 /// a splat of a single element and it's a 2 or 4 element mask.
2414 bool X86::isSplatMask(SDNode *N) {
2415 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2417 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2418 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2420 return ::isSplatMask(N);
2423 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2424 /// specifies a splat of zero element.
2425 bool X86::isSplatLoMask(SDNode *N) {
2426 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2428 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2429 if (!isUndefOrEqual(N->getOperand(i), 0))
2434 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2435 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2437 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2438 unsigned NumOperands = N->getNumOperands();
2439 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2441 for (unsigned i = 0; i < NumOperands; ++i) {
2443 SDOperand Arg = N->getOperand(NumOperands-i-1);
2444 if (Arg.getOpcode() != ISD::UNDEF)
2445 Val = cast<ConstantSDNode>(Arg)->getValue();
2446 if (Val >= NumOperands) Val -= NumOperands;
2448 if (i != NumOperands - 1)
2455 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2456 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2458 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2460 // 8 nodes, but we only care about the last 4.
2461 for (unsigned i = 7; i >= 4; --i) {
2463 SDOperand Arg = N->getOperand(i);
2464 if (Arg.getOpcode() != ISD::UNDEF)
2465 Val = cast<ConstantSDNode>(Arg)->getValue();
2474 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2475 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2477 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2479 // 8 nodes, but we only care about the first 4.
2480 for (int i = 3; i >= 0; --i) {
2482 SDOperand Arg = N->getOperand(i);
2483 if (Arg.getOpcode() != ISD::UNDEF)
2484 Val = cast<ConstantSDNode>(Arg)->getValue();
2493 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2494 /// specifies a 8 element shuffle that can be broken into a pair of
2495 /// PSHUFHW and PSHUFLW.
2496 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2497 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2499 if (N->getNumOperands() != 8)
2502 // Lower quadword shuffled.
2503 for (unsigned i = 0; i != 4; ++i) {
2504 SDOperand Arg = N->getOperand(i);
2505 if (Arg.getOpcode() == ISD::UNDEF) continue;
2506 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2507 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2512 // Upper quadword shuffled.
2513 for (unsigned i = 4; i != 8; ++i) {
2514 SDOperand Arg = N->getOperand(i);
2515 if (Arg.getOpcode() == ISD::UNDEF) continue;
2516 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2517 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2518 if (Val < 4 || Val > 7)
2525 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2526 /// values in ther permute mask.
2527 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2528 SDOperand &V2, SDOperand &Mask,
2529 SelectionDAG &DAG) {
2530 MVT::ValueType VT = Op.getValueType();
2531 MVT::ValueType MaskVT = Mask.getValueType();
2532 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2533 unsigned NumElems = Mask.getNumOperands();
2534 SmallVector<SDOperand, 8> MaskVec;
2536 for (unsigned i = 0; i != NumElems; ++i) {
2537 SDOperand Arg = Mask.getOperand(i);
2538 if (Arg.getOpcode() == ISD::UNDEF) {
2539 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2542 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2543 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2545 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2547 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2551 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2552 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2555 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2556 /// match movhlps. The lower half elements should come from upper half of
2557 /// V1 (and in order), and the upper half elements should come from the upper
2558 /// half of V2 (and in order).
2559 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2560 unsigned NumElems = Mask->getNumOperands();
2563 for (unsigned i = 0, e = 2; i != e; ++i)
2564 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2566 for (unsigned i = 2; i != 4; ++i)
2567 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2572 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2573 /// is promoted to a vector.
2574 static inline bool isScalarLoadToVector(SDNode *N) {
2575 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2576 N = N->getOperand(0).Val;
2577 return ISD::isNON_EXTLoad(N);
2582 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2583 /// match movlp{s|d}. The lower half elements should come from lower half of
2584 /// V1 (and in order), and the upper half elements should come from the upper
2585 /// half of V2 (and in order). And since V1 will become the source of the
2586 /// MOVLP, it must be either a vector load or a scalar load to vector.
2587 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2588 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2590 // Is V2 is a vector load, don't do this transformation. We will try to use
2591 // load folding shufps op.
2592 if (ISD::isNON_EXTLoad(V2))
2595 unsigned NumElems = Mask->getNumOperands();
2596 if (NumElems != 2 && NumElems != 4)
2598 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2599 if (!isUndefOrEqual(Mask->getOperand(i), i))
2601 for (unsigned i = NumElems/2; i != NumElems; ++i)
2602 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2607 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2609 static bool isSplatVector(SDNode *N) {
2610 if (N->getOpcode() != ISD::BUILD_VECTOR)
2613 SDOperand SplatValue = N->getOperand(0);
2614 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2615 if (N->getOperand(i) != SplatValue)
2620 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2622 static bool isUndefShuffle(SDNode *N) {
2623 if (N->getOpcode() != ISD::BUILD_VECTOR)
2626 SDOperand V1 = N->getOperand(0);
2627 SDOperand V2 = N->getOperand(1);
2628 SDOperand Mask = N->getOperand(2);
2629 unsigned NumElems = Mask.getNumOperands();
2630 for (unsigned i = 0; i != NumElems; ++i) {
2631 SDOperand Arg = Mask.getOperand(i);
2632 if (Arg.getOpcode() != ISD::UNDEF) {
2633 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2634 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2636 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2643 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2644 /// that point to V2 points to its first element.
2645 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2646 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2648 bool Changed = false;
2649 SmallVector<SDOperand, 8> MaskVec;
2650 unsigned NumElems = Mask.getNumOperands();
2651 for (unsigned i = 0; i != NumElems; ++i) {
2652 SDOperand Arg = Mask.getOperand(i);
2653 if (Arg.getOpcode() != ISD::UNDEF) {
2654 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2655 if (Val > NumElems) {
2656 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2660 MaskVec.push_back(Arg);
2664 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2665 &MaskVec[0], MaskVec.size());
2669 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2670 /// operation of specified width.
2671 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2672 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2673 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2675 SmallVector<SDOperand, 8> MaskVec;
2676 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2677 for (unsigned i = 1; i != NumElems; ++i)
2678 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2679 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2682 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2683 /// of specified width.
2684 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2685 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2686 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2687 SmallVector<SDOperand, 8> MaskVec;
2688 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2689 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2690 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2692 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2695 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2696 /// of specified width.
2697 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2698 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2699 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2700 unsigned Half = NumElems/2;
2701 SmallVector<SDOperand, 8> MaskVec;
2702 for (unsigned i = 0; i != Half; ++i) {
2703 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2704 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2706 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2709 /// getZeroVector - Returns a vector of specified type with all zero elements.
2711 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2712 assert(MVT::isVector(VT) && "Expected a vector type");
2713 unsigned NumElems = getVectorNumElements(VT);
2714 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2715 bool isFP = MVT::isFloatingPoint(EVT);
2716 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2717 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2718 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2721 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2723 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2724 SDOperand V1 = Op.getOperand(0);
2725 SDOperand Mask = Op.getOperand(2);
2726 MVT::ValueType VT = Op.getValueType();
2727 unsigned NumElems = Mask.getNumOperands();
2728 Mask = getUnpacklMask(NumElems, DAG);
2729 while (NumElems != 4) {
2730 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2733 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2735 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2736 Mask = getZeroVector(MaskVT, DAG);
2737 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2738 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2739 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2742 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2744 static inline bool isZeroNode(SDOperand Elt) {
2745 return ((isa<ConstantSDNode>(Elt) &&
2746 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2747 (isa<ConstantFPSDNode>(Elt) &&
2748 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2751 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2752 /// vector and zero or undef vector.
2753 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2754 unsigned NumElems, unsigned Idx,
2755 bool isZero, SelectionDAG &DAG) {
2756 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2757 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2758 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2759 SDOperand Zero = DAG.getConstant(0, EVT);
2760 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2761 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2762 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2763 &MaskVec[0], MaskVec.size());
2764 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2767 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2769 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2770 unsigned NumNonZero, unsigned NumZero,
2771 SelectionDAG &DAG, TargetLowering &TLI) {
2777 for (unsigned i = 0; i < 16; ++i) {
2778 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2779 if (ThisIsNonZero && First) {
2781 V = getZeroVector(MVT::v8i16, DAG);
2783 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2788 SDOperand ThisElt(0, 0), LastElt(0, 0);
2789 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2790 if (LastIsNonZero) {
2791 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2793 if (ThisIsNonZero) {
2794 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2795 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2796 ThisElt, DAG.getConstant(8, MVT::i8));
2798 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2803 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2804 DAG.getConstant(i/2, TLI.getPointerTy()));
2808 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2811 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2813 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2814 unsigned NumNonZero, unsigned NumZero,
2815 SelectionDAG &DAG, TargetLowering &TLI) {
2821 for (unsigned i = 0; i < 8; ++i) {
2822 bool isNonZero = (NonZeros & (1 << i)) != 0;
2826 V = getZeroVector(MVT::v8i16, DAG);
2828 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2831 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2832 DAG.getConstant(i, TLI.getPointerTy()));
2840 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2841 // All zero's are handled with pxor.
2842 if (ISD::isBuildVectorAllZeros(Op.Val))
2845 // All one's are handled with pcmpeqd.
2846 if (ISD::isBuildVectorAllOnes(Op.Val))
2849 MVT::ValueType VT = Op.getValueType();
2850 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2851 unsigned EVTBits = MVT::getSizeInBits(EVT);
2853 unsigned NumElems = Op.getNumOperands();
2854 unsigned NumZero = 0;
2855 unsigned NumNonZero = 0;
2856 unsigned NonZeros = 0;
2857 std::set<SDOperand> Values;
2858 for (unsigned i = 0; i < NumElems; ++i) {
2859 SDOperand Elt = Op.getOperand(i);
2860 if (Elt.getOpcode() != ISD::UNDEF) {
2862 if (isZeroNode(Elt))
2865 NonZeros |= (1 << i);
2871 if (NumNonZero == 0)
2872 // Must be a mix of zero and undef. Return a zero vector.
2873 return getZeroVector(VT, DAG);
2875 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2876 if (Values.size() == 1)
2879 // Special case for single non-zero element.
2880 if (NumNonZero == 1) {
2881 unsigned Idx = CountTrailingZeros_32(NonZeros);
2882 SDOperand Item = Op.getOperand(Idx);
2883 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2885 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2886 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2889 if (EVTBits == 32) {
2890 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2891 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2893 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2894 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2895 SmallVector<SDOperand, 8> MaskVec;
2896 for (unsigned i = 0; i < NumElems; i++)
2897 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2898 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2899 &MaskVec[0], MaskVec.size());
2900 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2901 DAG.getNode(ISD::UNDEF, VT), Mask);
2905 // Let legalizer expand 2-wide build_vector's.
2909 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2911 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2913 if (V.Val) return V;
2916 if (EVTBits == 16) {
2917 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2919 if (V.Val) return V;
2922 // If element VT is == 32 bits, turn it into a number of shuffles.
2923 SmallVector<SDOperand, 8> V;
2925 if (NumElems == 4 && NumZero > 0) {
2926 for (unsigned i = 0; i < 4; ++i) {
2927 bool isZero = !(NonZeros & (1 << i));
2929 V[i] = getZeroVector(VT, DAG);
2931 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2934 for (unsigned i = 0; i < 2; ++i) {
2935 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2938 V[i] = V[i*2]; // Must be a zero vector.
2941 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2942 getMOVLMask(NumElems, DAG));
2945 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2946 getMOVLMask(NumElems, DAG));
2949 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2950 getUnpacklMask(NumElems, DAG));
2955 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2956 // clears the upper bits.
2957 // FIXME: we can do the same for v4f32 case when we know both parts of
2958 // the lower half come from scalar_to_vector (loadf32). We should do
2959 // that in post legalizer dag combiner with target specific hooks.
2960 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2962 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2963 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2964 SmallVector<SDOperand, 8> MaskVec;
2965 bool Reverse = (NonZeros & 0x3) == 2;
2966 for (unsigned i = 0; i < 2; ++i)
2968 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2970 MaskVec.push_back(DAG.getConstant(i, EVT));
2971 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2972 for (unsigned i = 0; i < 2; ++i)
2974 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2976 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2977 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2978 &MaskVec[0], MaskVec.size());
2979 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2982 if (Values.size() > 2) {
2983 // Expand into a number of unpckl*.
2985 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2986 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2987 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2988 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2989 for (unsigned i = 0; i < NumElems; ++i)
2990 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2992 while (NumElems != 0) {
2993 for (unsigned i = 0; i < NumElems; ++i)
2994 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3005 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3006 SDOperand V1 = Op.getOperand(0);
3007 SDOperand V2 = Op.getOperand(1);
3008 SDOperand PermMask = Op.getOperand(2);
3009 MVT::ValueType VT = Op.getValueType();
3010 unsigned NumElems = PermMask.getNumOperands();
3011 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3012 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3013 bool V1IsSplat = false;
3014 bool V2IsSplat = false;
3016 if (isUndefShuffle(Op.Val))
3017 return DAG.getNode(ISD::UNDEF, VT);
3019 if (isSplatMask(PermMask.Val)) {
3020 if (NumElems <= 4) return Op;
3021 // Promote it to a v4i32 splat.
3022 return PromoteSplat(Op, DAG);
3025 if (X86::isMOVLMask(PermMask.Val))
3026 return (V1IsUndef) ? V2 : Op;
3028 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3029 X86::isMOVSLDUPMask(PermMask.Val) ||
3030 X86::isMOVHLPSMask(PermMask.Val) ||
3031 X86::isMOVHPMask(PermMask.Val) ||
3032 X86::isMOVLPMask(PermMask.Val))
3035 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3036 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3037 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3039 bool Commuted = false;
3040 V1IsSplat = isSplatVector(V1.Val);
3041 V2IsSplat = isSplatVector(V2.Val);
3042 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3043 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3044 std::swap(V1IsSplat, V2IsSplat);
3045 std::swap(V1IsUndef, V2IsUndef);
3049 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3050 if (V2IsUndef) return V1;
3051 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3053 // V2 is a splat, so the mask may be malformed. That is, it may point
3054 // to any V2 element. The instruction selectior won't like this. Get
3055 // a corrected mask and commute to form a proper MOVS{S|D}.
3056 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3057 if (NewMask.Val != PermMask.Val)
3058 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3063 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3064 X86::isUNPCKLMask(PermMask.Val) ||
3065 X86::isUNPCKHMask(PermMask.Val))
3069 // Normalize mask so all entries that point to V2 points to its first
3070 // element then try to match unpck{h|l} again. If match, return a
3071 // new vector_shuffle with the corrected mask.
3072 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3073 if (NewMask.Val != PermMask.Val) {
3074 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3075 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3076 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3077 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3078 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3079 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3084 // Normalize the node to match x86 shuffle ops if needed
3085 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3086 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3089 // Commute is back and try unpck* again.
3090 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3091 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3092 X86::isUNPCKLMask(PermMask.Val) ||
3093 X86::isUNPCKHMask(PermMask.Val))
3097 // If VT is integer, try PSHUF* first, then SHUFP*.
3098 if (MVT::isInteger(VT)) {
3099 if (X86::isPSHUFDMask(PermMask.Val) ||
3100 X86::isPSHUFHWMask(PermMask.Val) ||
3101 X86::isPSHUFLWMask(PermMask.Val)) {
3102 if (V2.getOpcode() != ISD::UNDEF)
3103 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3104 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3108 if (X86::isSHUFPMask(PermMask.Val))
3111 // Handle v8i16 shuffle high / low shuffle node pair.
3112 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3113 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3114 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3115 SmallVector<SDOperand, 8> MaskVec;
3116 for (unsigned i = 0; i != 4; ++i)
3117 MaskVec.push_back(PermMask.getOperand(i));
3118 for (unsigned i = 4; i != 8; ++i)
3119 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3120 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3121 &MaskVec[0], MaskVec.size());
3122 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3124 for (unsigned i = 0; i != 4; ++i)
3125 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3126 for (unsigned i = 4; i != 8; ++i)
3127 MaskVec.push_back(PermMask.getOperand(i));
3128 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3129 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3132 // Floating point cases in the other order.
3133 if (X86::isSHUFPMask(PermMask.Val))
3135 if (X86::isPSHUFDMask(PermMask.Val) ||
3136 X86::isPSHUFHWMask(PermMask.Val) ||
3137 X86::isPSHUFLWMask(PermMask.Val)) {
3138 if (V2.getOpcode() != ISD::UNDEF)
3139 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3140 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3145 if (NumElems == 4) {
3146 MVT::ValueType MaskVT = PermMask.getValueType();
3147 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
3148 SmallVector<std::pair<int, int>, 8> Locs;
3149 Locs.reserve(NumElems);
3150 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3151 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3154 // If no more than two elements come from either vector. This can be
3155 // implemented with two shuffles. First shuffle gather the elements.
3156 // The second shuffle, which takes the first shuffle as both of its
3157 // vector operands, put the elements into the right order.
3158 for (unsigned i = 0; i != NumElems; ++i) {
3159 SDOperand Elt = PermMask.getOperand(i);
3160 if (Elt.getOpcode() == ISD::UNDEF) {
3161 Locs[i] = std::make_pair(-1, -1);
3163 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3164 if (Val < NumElems) {
3165 Locs[i] = std::make_pair(0, NumLo);
3169 Locs[i] = std::make_pair(1, NumHi);
3170 if (2+NumHi < NumElems)
3171 Mask1[2+NumHi] = Elt;
3176 if (NumLo <= 2 && NumHi <= 2) {
3177 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3178 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3179 &Mask1[0], Mask1.size()));
3180 for (unsigned i = 0; i != NumElems; ++i) {
3181 if (Locs[i].first == -1)
3184 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3185 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3186 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3190 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3191 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3192 &Mask2[0], Mask2.size()));
3195 // Break it into (shuffle shuffle_hi, shuffle_lo).
3197 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3198 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3199 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3200 unsigned MaskIdx = 0;
3202 unsigned HiIdx = NumElems/2;
3203 for (unsigned i = 0; i != NumElems; ++i) {
3204 if (i == NumElems/2) {
3210 SDOperand Elt = PermMask.getOperand(i);
3211 if (Elt.getOpcode() == ISD::UNDEF) {
3212 Locs[i] = std::make_pair(-1, -1);
3213 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3214 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3215 (*MaskPtr)[LoIdx] = Elt;
3218 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3219 (*MaskPtr)[HiIdx] = Elt;
3224 SDOperand LoShuffle =
3225 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3226 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3227 &LoMask[0], LoMask.size()));
3228 SDOperand HiShuffle =
3229 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3230 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3231 &HiMask[0], HiMask.size()));
3232 SmallVector<SDOperand, 8> MaskOps;
3233 for (unsigned i = 0; i != NumElems; ++i) {
3234 if (Locs[i].first == -1) {
3235 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3237 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3238 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3241 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3242 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3243 &MaskOps[0], MaskOps.size()));
3250 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3251 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3254 MVT::ValueType VT = Op.getValueType();
3255 // TODO: handle v16i8.
3256 if (MVT::getSizeInBits(VT) == 16) {
3257 // Transform it so it match pextrw which produces a 32-bit result.
3258 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3259 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3260 Op.getOperand(0), Op.getOperand(1));
3261 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3262 DAG.getValueType(VT));
3263 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3264 } else if (MVT::getSizeInBits(VT) == 32) {
3265 SDOperand Vec = Op.getOperand(0);
3266 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3269 // SHUFPS the element to the lowest double word, then movss.
3270 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3271 SmallVector<SDOperand, 8> IdxVec;
3272 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
3273 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3274 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3275 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3276 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3277 &IdxVec[0], IdxVec.size());
3278 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3279 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3280 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3281 DAG.getConstant(0, getPointerTy()));
3282 } else if (MVT::getSizeInBits(VT) == 64) {
3283 SDOperand Vec = Op.getOperand(0);
3284 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3288 // UNPCKHPD the element to the lowest double word, then movsd.
3289 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3290 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3291 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3292 SmallVector<SDOperand, 8> IdxVec;
3293 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
3294 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3295 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3296 &IdxVec[0], IdxVec.size());
3297 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3298 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3299 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3300 DAG.getConstant(0, getPointerTy()));
3307 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3308 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3309 // as its second argument.
3310 MVT::ValueType VT = Op.getValueType();
3311 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
3312 SDOperand N0 = Op.getOperand(0);
3313 SDOperand N1 = Op.getOperand(1);
3314 SDOperand N2 = Op.getOperand(2);
3315 if (MVT::getSizeInBits(BaseVT) == 16) {
3316 if (N1.getValueType() != MVT::i32)
3317 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3318 if (N2.getValueType() != MVT::i32)
3319 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
3320 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3321 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3322 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3325 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3326 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3327 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3328 SmallVector<SDOperand, 8> MaskVec;
3329 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3330 for (unsigned i = 1; i <= 3; ++i)
3331 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3332 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3333 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3334 &MaskVec[0], MaskVec.size()));
3336 // Use two pinsrw instructions to insert a 32 bit value.
3338 if (MVT::isFloatingPoint(N1.getValueType())) {
3339 if (ISD::isNON_EXTLoad(N1.Val)) {
3340 // Just load directly from f32mem to GR32.
3341 LoadSDNode *LD = cast<LoadSDNode>(N1);
3342 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
3343 LD->getSrcValue(), LD->getSrcValueOffset());
3345 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3346 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3347 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3348 DAG.getConstant(0, getPointerTy()));
3351 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3352 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3353 DAG.getConstant(Idx, getPointerTy()));
3354 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3355 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3356 DAG.getConstant(Idx+1, getPointerTy()));
3357 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3365 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3366 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3367 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3370 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3371 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3372 // one of the above mentioned nodes. It has to be wrapped because otherwise
3373 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3374 // be used to form addressing mode. These wrapped nodes will be selected
3377 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3378 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3379 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3381 CP->getAlignment());
3382 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3383 // With PIC, the address is actually $g + Offset.
3384 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3385 !Subtarget->isPICStyleRIPRel()) {
3386 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3387 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3395 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3396 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3397 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3398 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3399 // With PIC, the address is actually $g + Offset.
3400 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3401 !Subtarget->isPICStyleRIPRel()) {
3402 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3403 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3407 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3408 // load the value at address GV, not the value of GV itself. This means that
3409 // the GlobalAddress must be in the base or index register of the address, not
3410 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3411 // The same applies for external symbols during PIC codegen
3412 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3413 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3419 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3420 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3421 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3422 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3423 // With PIC, the address is actually $g + Offset.
3424 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3425 !Subtarget->isPICStyleRIPRel()) {
3426 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3427 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3434 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3435 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3436 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3437 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3438 // With PIC, the address is actually $g + Offset.
3439 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3440 !Subtarget->isPICStyleRIPRel()) {
3441 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3442 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3449 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3450 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3451 "Not an i64 shift!");
3452 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3453 SDOperand ShOpLo = Op.getOperand(0);
3454 SDOperand ShOpHi = Op.getOperand(1);
3455 SDOperand ShAmt = Op.getOperand(2);
3456 SDOperand Tmp1 = isSRA ?
3457 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3458 DAG.getConstant(0, MVT::i32);
3460 SDOperand Tmp2, Tmp3;
3461 if (Op.getOpcode() == ISD::SHL_PARTS) {
3462 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3463 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3465 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3466 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3469 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3470 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3471 DAG.getConstant(32, MVT::i8));
3472 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3473 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3476 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3478 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3479 SmallVector<SDOperand, 4> Ops;
3480 if (Op.getOpcode() == ISD::SHL_PARTS) {
3481 Ops.push_back(Tmp2);
3482 Ops.push_back(Tmp3);
3484 Ops.push_back(InFlag);
3485 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3486 InFlag = Hi.getValue(1);
3489 Ops.push_back(Tmp3);
3490 Ops.push_back(Tmp1);
3492 Ops.push_back(InFlag);
3493 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3495 Ops.push_back(Tmp2);
3496 Ops.push_back(Tmp3);
3498 Ops.push_back(InFlag);
3499 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3500 InFlag = Lo.getValue(1);
3503 Ops.push_back(Tmp3);
3504 Ops.push_back(Tmp1);
3506 Ops.push_back(InFlag);
3507 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3510 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3514 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3517 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3518 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3519 Op.getOperand(0).getValueType() >= MVT::i16 &&
3520 "Unknown SINT_TO_FP to lower!");
3523 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3524 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3525 MachineFunction &MF = DAG.getMachineFunction();
3526 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3527 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3528 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3529 StackSlot, NULL, 0);
3534 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3536 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3537 SmallVector<SDOperand, 8> Ops;
3538 Ops.push_back(Chain);
3539 Ops.push_back(StackSlot);
3540 Ops.push_back(DAG.getValueType(SrcVT));
3541 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3542 Tys, &Ops[0], Ops.size());
3545 Chain = Result.getValue(1);
3546 SDOperand InFlag = Result.getValue(2);
3548 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3549 // shouldn't be necessary except that RFP cannot be live across
3550 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3551 MachineFunction &MF = DAG.getMachineFunction();
3552 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3553 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3554 Tys = DAG.getVTList(MVT::Other);
3555 SmallVector<SDOperand, 8> Ops;
3556 Ops.push_back(Chain);
3557 Ops.push_back(Result);
3558 Ops.push_back(StackSlot);
3559 Ops.push_back(DAG.getValueType(Op.getValueType()));
3560 Ops.push_back(InFlag);
3561 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3562 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3568 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3569 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3570 "Unknown FP_TO_SINT to lower!");
3571 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3573 MachineFunction &MF = DAG.getMachineFunction();
3574 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3575 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3576 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3579 switch (Op.getValueType()) {
3580 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3581 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3582 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3583 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3586 SDOperand Chain = DAG.getEntryNode();
3587 SDOperand Value = Op.getOperand(0);
3589 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3590 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3591 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3593 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3595 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3596 Chain = Value.getValue(1);
3597 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3598 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3601 // Build the FP_TO_INT*_IN_MEM
3602 SDOperand Ops[] = { Chain, Value, StackSlot };
3603 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3606 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3609 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3610 MVT::ValueType VT = Op.getValueType();
3611 const Type *OpNTy = MVT::getTypeForValueType(VT);
3612 std::vector<Constant*> CV;
3613 if (VT == MVT::f64) {
3614 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3615 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3617 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3618 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3619 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3620 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3622 Constant *CS = ConstantStruct::get(CV);
3623 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3624 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3625 SmallVector<SDOperand, 3> Ops;
3626 Ops.push_back(DAG.getEntryNode());
3627 Ops.push_back(CPIdx);
3628 Ops.push_back(DAG.getSrcValue(NULL));
3629 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3630 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3633 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3634 MVT::ValueType VT = Op.getValueType();
3635 const Type *OpNTy = MVT::getTypeForValueType(VT);
3636 std::vector<Constant*> CV;
3637 if (VT == MVT::f64) {
3638 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3639 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3641 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3642 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3643 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3644 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3646 Constant *CS = ConstantStruct::get(CV);
3647 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3648 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3649 SmallVector<SDOperand, 3> Ops;
3650 Ops.push_back(DAG.getEntryNode());
3651 Ops.push_back(CPIdx);
3652 Ops.push_back(DAG.getSrcValue(NULL));
3653 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3654 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3657 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3658 SDOperand Op0 = Op.getOperand(0);
3659 SDOperand Op1 = Op.getOperand(1);
3660 MVT::ValueType VT = Op.getValueType();
3661 MVT::ValueType SrcVT = Op1.getValueType();
3662 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3664 // If second operand is smaller, extend it first.
3665 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3666 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3670 // First get the sign bit of second operand.
3671 std::vector<Constant*> CV;
3672 if (SrcVT == MVT::f64) {
3673 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3674 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3676 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3677 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3678 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3679 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3681 Constant *CS = ConstantStruct::get(CV);
3682 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3683 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3684 SmallVector<SDOperand, 3> Ops;
3685 Ops.push_back(DAG.getEntryNode());
3686 Ops.push_back(CPIdx);
3687 Ops.push_back(DAG.getSrcValue(NULL));
3688 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3689 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3691 // Shift sign bit right or left if the two operands have different types.
3692 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3693 // Op0 is MVT::f32, Op1 is MVT::f64.
3694 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3695 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3696 DAG.getConstant(32, MVT::i32));
3697 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3698 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3699 DAG.getConstant(0, getPointerTy()));
3702 // Clear first operand sign bit.
3704 if (VT == MVT::f64) {
3705 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3706 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3708 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3709 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3710 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3711 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3713 CS = ConstantStruct::get(CV);
3714 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3715 Tys = DAG.getVTList(VT, MVT::Other);
3717 Ops.push_back(DAG.getEntryNode());
3718 Ops.push_back(CPIdx);
3719 Ops.push_back(DAG.getSrcValue(NULL));
3720 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3721 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3723 // Or the value with the sign bit.
3724 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3727 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3729 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3731 SDOperand Op0 = Op.getOperand(0);
3732 SDOperand Op1 = Op.getOperand(1);
3733 SDOperand CC = Op.getOperand(2);
3734 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3735 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3736 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3737 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3740 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3742 SDOperand Ops1[] = { Chain, Op0, Op1 };
3743 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3744 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3745 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3748 assert(isFP && "Illegal integer SetCC!");
3750 SDOperand COps[] = { Chain, Op0, Op1 };
3751 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3753 switch (SetCCOpcode) {
3754 default: assert(false && "Illegal floating point SetCC!");
3755 case ISD::SETOEQ: { // !PF & ZF
3756 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3757 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3758 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3760 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3761 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3763 case ISD::SETUNE: { // PF | !ZF
3764 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3765 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3766 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3768 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3769 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3774 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3775 bool addTest = true;
3776 SDOperand Chain = DAG.getEntryNode();
3777 SDOperand Cond = Op.getOperand(0);
3779 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3781 if (Cond.getOpcode() == ISD::SETCC)
3782 Cond = LowerSETCC(Cond, DAG, Chain);
3784 if (Cond.getOpcode() == X86ISD::SETCC) {
3785 CC = Cond.getOperand(0);
3787 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3788 // (since flag operand cannot be shared). Use it as the condition setting
3789 // operand in place of the X86ISD::SETCC.
3790 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3791 // to use a test instead of duplicating the X86ISD::CMP (for register
3792 // pressure reason)?
3793 SDOperand Cmp = Cond.getOperand(1);
3794 unsigned Opc = Cmp.getOpcode();
3795 bool IllegalFPCMov = !X86ScalarSSE &&
3796 MVT::isFloatingPoint(Op.getValueType()) &&
3797 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3798 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3800 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3801 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3807 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3808 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3809 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3812 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3813 SmallVector<SDOperand, 4> Ops;
3814 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3815 // condition is true.
3816 Ops.push_back(Op.getOperand(2));
3817 Ops.push_back(Op.getOperand(1));
3819 Ops.push_back(Cond.getValue(1));
3820 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3823 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3824 bool addTest = true;
3825 SDOperand Chain = Op.getOperand(0);
3826 SDOperand Cond = Op.getOperand(1);
3827 SDOperand Dest = Op.getOperand(2);
3829 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3831 if (Cond.getOpcode() == ISD::SETCC)
3832 Cond = LowerSETCC(Cond, DAG, Chain);
3834 if (Cond.getOpcode() == X86ISD::SETCC) {
3835 CC = Cond.getOperand(0);
3837 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3838 // (since flag operand cannot be shared). Use it as the condition setting
3839 // operand in place of the X86ISD::SETCC.
3840 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3841 // to use a test instead of duplicating the X86ISD::CMP (for register
3842 // pressure reason)?
3843 SDOperand Cmp = Cond.getOperand(1);
3844 unsigned Opc = Cmp.getOpcode();
3845 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3846 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3847 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3853 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3854 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3855 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3857 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3858 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3861 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3862 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3864 if (Subtarget->is64Bit())
3865 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3867 switch (CallingConv) {
3869 assert(0 && "Unsupported calling convention");
3870 case CallingConv::Fast:
3872 return LowerFastCCCallTo(Op, DAG, CallingConv);
3874 case CallingConv::C:
3875 case CallingConv::X86_StdCall:
3876 return LowerCCCCallTo(Op, DAG, CallingConv);
3877 case CallingConv::X86_FastCall:
3878 return LowerFastCCCallTo(Op, DAG, CallingConv);
3883 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3884 MachineFunction &MF = DAG.getMachineFunction();
3885 const Function* Fn = MF.getFunction();
3886 if (Fn->hasExternalLinkage() &&
3887 Subtarget->isTargetCygMing() &&
3888 Fn->getName() == "main")
3889 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
3891 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3892 if (Subtarget->is64Bit())
3893 return LowerX86_64CCCArguments(Op, DAG);
3897 assert(0 && "Unsupported calling convention");
3898 case CallingConv::Fast:
3900 return LowerFastCCArguments(Op, DAG);
3903 case CallingConv::C:
3904 return LowerCCCArguments(Op, DAG);
3905 case CallingConv::X86_StdCall:
3906 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
3907 return LowerCCCArguments(Op, DAG, true);
3908 case CallingConv::X86_FastCall:
3909 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
3910 return LowerFastCCArguments(Op, DAG, true);
3914 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3915 SDOperand InFlag(0, 0);
3916 SDOperand Chain = Op.getOperand(0);
3918 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3919 if (Align == 0) Align = 1;
3921 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3922 // If not DWORD aligned, call memset if size is less than the threshold.
3923 // It knows how to align to the right boundary first.
3924 if ((Align & 3) != 0 ||
3925 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3926 MVT::ValueType IntPtr = getPointerTy();
3927 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3928 TargetLowering::ArgListTy Args;
3929 TargetLowering::ArgListEntry Entry;
3930 Entry.Node = Op.getOperand(1);
3931 Entry.Ty = IntPtrTy;
3932 Entry.isSigned = false;
3933 Entry.isInReg = false;
3934 Entry.isSRet = false;
3935 Args.push_back(Entry);
3936 // Extend the unsigned i8 argument to be an int value for the call.
3937 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3938 Entry.Ty = IntPtrTy;
3939 Entry.isSigned = false;
3940 Entry.isInReg = false;
3941 Entry.isSRet = false;
3942 Args.push_back(Entry);
3943 Entry.Node = Op.getOperand(3);
3944 Args.push_back(Entry);
3945 std::pair<SDOperand,SDOperand> CallResult =
3946 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3947 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3948 return CallResult.second;
3953 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3954 unsigned BytesLeft = 0;
3955 bool TwoRepStos = false;
3958 uint64_t Val = ValC->getValue() & 255;
3960 // If the value is a constant, then we can potentially use larger sets.
3961 switch (Align & 3) {
3962 case 2: // WORD aligned
3965 Val = (Val << 8) | Val;
3967 case 0: // DWORD aligned
3970 Val = (Val << 8) | Val;
3971 Val = (Val << 16) | Val;
3972 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3975 Val = (Val << 32) | Val;
3978 default: // Byte aligned
3981 Count = Op.getOperand(3);
3985 if (AVT > MVT::i8) {
3987 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3988 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3989 BytesLeft = I->getValue() % UBytes;
3991 assert(AVT >= MVT::i32 &&
3992 "Do not use rep;stos if not at least DWORD aligned");
3993 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3994 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3999 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4001 InFlag = Chain.getValue(1);
4004 Count = Op.getOperand(3);
4005 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4006 InFlag = Chain.getValue(1);
4009 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4011 InFlag = Chain.getValue(1);
4012 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4013 Op.getOperand(1), InFlag);
4014 InFlag = Chain.getValue(1);
4016 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4017 SmallVector<SDOperand, 8> Ops;
4018 Ops.push_back(Chain);
4019 Ops.push_back(DAG.getValueType(AVT));
4020 Ops.push_back(InFlag);
4021 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4024 InFlag = Chain.getValue(1);
4025 Count = Op.getOperand(3);
4026 MVT::ValueType CVT = Count.getValueType();
4027 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4028 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4029 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4031 InFlag = Chain.getValue(1);
4032 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4034 Ops.push_back(Chain);
4035 Ops.push_back(DAG.getValueType(MVT::i8));
4036 Ops.push_back(InFlag);
4037 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4038 } else if (BytesLeft) {
4039 // Issue stores for the last 1 - 7 bytes.
4041 unsigned Val = ValC->getValue() & 255;
4042 unsigned Offset = I->getValue() - BytesLeft;
4043 SDOperand DstAddr = Op.getOperand(1);
4044 MVT::ValueType AddrVT = DstAddr.getValueType();
4045 if (BytesLeft >= 4) {
4046 Val = (Val << 8) | Val;
4047 Val = (Val << 16) | Val;
4048 Value = DAG.getConstant(Val, MVT::i32);
4049 Chain = DAG.getStore(Chain, Value,
4050 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4051 DAG.getConstant(Offset, AddrVT)),
4056 if (BytesLeft >= 2) {
4057 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4058 Chain = DAG.getStore(Chain, Value,
4059 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4060 DAG.getConstant(Offset, AddrVT)),
4065 if (BytesLeft == 1) {
4066 Value = DAG.getConstant(Val, MVT::i8);
4067 Chain = DAG.getStore(Chain, Value,
4068 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4069 DAG.getConstant(Offset, AddrVT)),
4077 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
4078 SDOperand Chain = Op.getOperand(0);
4080 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4081 if (Align == 0) Align = 1;
4083 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4084 // If not DWORD aligned, call memcpy if size is less than the threshold.
4085 // It knows how to align to the right boundary first.
4086 if ((Align & 3) != 0 ||
4087 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4088 MVT::ValueType IntPtr = getPointerTy();
4089 TargetLowering::ArgListTy Args;
4090 TargetLowering::ArgListEntry Entry;
4091 Entry.Ty = getTargetData()->getIntPtrType();
4092 Entry.isSigned = false;
4093 Entry.isInReg = false;
4094 Entry.isSRet = false;
4095 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
4096 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
4097 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
4098 std::pair<SDOperand,SDOperand> CallResult =
4099 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4100 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
4101 return CallResult.second;
4106 unsigned BytesLeft = 0;
4107 bool TwoRepMovs = false;
4108 switch (Align & 3) {
4109 case 2: // WORD aligned
4112 case 0: // DWORD aligned
4114 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4117 default: // Byte aligned
4119 Count = Op.getOperand(3);
4123 if (AVT > MVT::i8) {
4125 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4126 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4127 BytesLeft = I->getValue() % UBytes;
4129 assert(AVT >= MVT::i32 &&
4130 "Do not use rep;movs if not at least DWORD aligned");
4131 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4132 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4137 SDOperand InFlag(0, 0);
4138 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4140 InFlag = Chain.getValue(1);
4141 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4142 Op.getOperand(1), InFlag);
4143 InFlag = Chain.getValue(1);
4144 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4145 Op.getOperand(2), InFlag);
4146 InFlag = Chain.getValue(1);
4148 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4149 SmallVector<SDOperand, 8> Ops;
4150 Ops.push_back(Chain);
4151 Ops.push_back(DAG.getValueType(AVT));
4152 Ops.push_back(InFlag);
4153 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4156 InFlag = Chain.getValue(1);
4157 Count = Op.getOperand(3);
4158 MVT::ValueType CVT = Count.getValueType();
4159 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4160 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4161 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4163 InFlag = Chain.getValue(1);
4164 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4166 Ops.push_back(Chain);
4167 Ops.push_back(DAG.getValueType(MVT::i8));
4168 Ops.push_back(InFlag);
4169 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4170 } else if (BytesLeft) {
4171 // Issue loads and stores for the last 1 - 7 bytes.
4172 unsigned Offset = I->getValue() - BytesLeft;
4173 SDOperand DstAddr = Op.getOperand(1);
4174 MVT::ValueType DstVT = DstAddr.getValueType();
4175 SDOperand SrcAddr = Op.getOperand(2);
4176 MVT::ValueType SrcVT = SrcAddr.getValueType();
4178 if (BytesLeft >= 4) {
4179 Value = DAG.getLoad(MVT::i32, Chain,
4180 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4181 DAG.getConstant(Offset, SrcVT)),
4183 Chain = Value.getValue(1);
4184 Chain = DAG.getStore(Chain, Value,
4185 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4186 DAG.getConstant(Offset, DstVT)),
4191 if (BytesLeft >= 2) {
4192 Value = DAG.getLoad(MVT::i16, Chain,
4193 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4194 DAG.getConstant(Offset, SrcVT)),
4196 Chain = Value.getValue(1);
4197 Chain = DAG.getStore(Chain, Value,
4198 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4199 DAG.getConstant(Offset, DstVT)),
4205 if (BytesLeft == 1) {
4206 Value = DAG.getLoad(MVT::i8, Chain,
4207 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4208 DAG.getConstant(Offset, SrcVT)),
4210 Chain = Value.getValue(1);
4211 Chain = DAG.getStore(Chain, Value,
4212 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4213 DAG.getConstant(Offset, DstVT)),
4222 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4223 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4224 SDOperand TheOp = Op.getOperand(0);
4225 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4226 if (Subtarget->is64Bit()) {
4227 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4228 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4229 MVT::i64, Copy1.getValue(2));
4230 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4231 DAG.getConstant(32, MVT::i8));
4233 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4236 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4237 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4240 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4241 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4242 MVT::i32, Copy1.getValue(2));
4243 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4244 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4245 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4248 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4249 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4251 if (!Subtarget->is64Bit()) {
4252 // vastart just stores the address of the VarArgsFrameIndex slot into the
4253 // memory location argument.
4254 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4255 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4260 // gp_offset (0 - 6 * 8)
4261 // fp_offset (48 - 48 + 8 * 16)
4262 // overflow_arg_area (point to parameters coming in memory).
4264 SmallVector<SDOperand, 8> MemOps;
4265 SDOperand FIN = Op.getOperand(1);
4267 SDOperand Store = DAG.getStore(Op.getOperand(0),
4268 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4269 FIN, SV->getValue(), SV->getOffset());
4270 MemOps.push_back(Store);
4273 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4274 DAG.getConstant(4, getPointerTy()));
4275 Store = DAG.getStore(Op.getOperand(0),
4276 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4277 FIN, SV->getValue(), SV->getOffset());
4278 MemOps.push_back(Store);
4280 // Store ptr to overflow_arg_area
4281 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4282 DAG.getConstant(4, getPointerTy()));
4283 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4284 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4286 MemOps.push_back(Store);
4288 // Store ptr to reg_save_area.
4289 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4290 DAG.getConstant(8, getPointerTy()));
4291 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4292 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4294 MemOps.push_back(Store);
4295 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4299 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4300 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4302 default: return SDOperand(); // Don't custom lower most intrinsics.
4303 // Comparison intrinsics.
4304 case Intrinsic::x86_sse_comieq_ss:
4305 case Intrinsic::x86_sse_comilt_ss:
4306 case Intrinsic::x86_sse_comile_ss:
4307 case Intrinsic::x86_sse_comigt_ss:
4308 case Intrinsic::x86_sse_comige_ss:
4309 case Intrinsic::x86_sse_comineq_ss:
4310 case Intrinsic::x86_sse_ucomieq_ss:
4311 case Intrinsic::x86_sse_ucomilt_ss:
4312 case Intrinsic::x86_sse_ucomile_ss:
4313 case Intrinsic::x86_sse_ucomigt_ss:
4314 case Intrinsic::x86_sse_ucomige_ss:
4315 case Intrinsic::x86_sse_ucomineq_ss:
4316 case Intrinsic::x86_sse2_comieq_sd:
4317 case Intrinsic::x86_sse2_comilt_sd:
4318 case Intrinsic::x86_sse2_comile_sd:
4319 case Intrinsic::x86_sse2_comigt_sd:
4320 case Intrinsic::x86_sse2_comige_sd:
4321 case Intrinsic::x86_sse2_comineq_sd:
4322 case Intrinsic::x86_sse2_ucomieq_sd:
4323 case Intrinsic::x86_sse2_ucomilt_sd:
4324 case Intrinsic::x86_sse2_ucomile_sd:
4325 case Intrinsic::x86_sse2_ucomigt_sd:
4326 case Intrinsic::x86_sse2_ucomige_sd:
4327 case Intrinsic::x86_sse2_ucomineq_sd: {
4329 ISD::CondCode CC = ISD::SETCC_INVALID;
4332 case Intrinsic::x86_sse_comieq_ss:
4333 case Intrinsic::x86_sse2_comieq_sd:
4337 case Intrinsic::x86_sse_comilt_ss:
4338 case Intrinsic::x86_sse2_comilt_sd:
4342 case Intrinsic::x86_sse_comile_ss:
4343 case Intrinsic::x86_sse2_comile_sd:
4347 case Intrinsic::x86_sse_comigt_ss:
4348 case Intrinsic::x86_sse2_comigt_sd:
4352 case Intrinsic::x86_sse_comige_ss:
4353 case Intrinsic::x86_sse2_comige_sd:
4357 case Intrinsic::x86_sse_comineq_ss:
4358 case Intrinsic::x86_sse2_comineq_sd:
4362 case Intrinsic::x86_sse_ucomieq_ss:
4363 case Intrinsic::x86_sse2_ucomieq_sd:
4364 Opc = X86ISD::UCOMI;
4367 case Intrinsic::x86_sse_ucomilt_ss:
4368 case Intrinsic::x86_sse2_ucomilt_sd:
4369 Opc = X86ISD::UCOMI;
4372 case Intrinsic::x86_sse_ucomile_ss:
4373 case Intrinsic::x86_sse2_ucomile_sd:
4374 Opc = X86ISD::UCOMI;
4377 case Intrinsic::x86_sse_ucomigt_ss:
4378 case Intrinsic::x86_sse2_ucomigt_sd:
4379 Opc = X86ISD::UCOMI;
4382 case Intrinsic::x86_sse_ucomige_ss:
4383 case Intrinsic::x86_sse2_ucomige_sd:
4384 Opc = X86ISD::UCOMI;
4387 case Intrinsic::x86_sse_ucomineq_ss:
4388 case Intrinsic::x86_sse2_ucomineq_sd:
4389 Opc = X86ISD::UCOMI;
4395 SDOperand LHS = Op.getOperand(1);
4396 SDOperand RHS = Op.getOperand(2);
4397 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4399 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4400 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4401 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4402 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4403 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4404 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4405 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4410 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4411 // Depths > 0 not supported yet!
4412 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4415 // Just load the return address
4416 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4417 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4420 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4421 // Depths > 0 not supported yet!
4422 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4425 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4426 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4427 DAG.getConstant(4, getPointerTy()));
4430 /// LowerOperation - Provide custom lowering hooks for some operations.
4432 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4433 switch (Op.getOpcode()) {
4434 default: assert(0 && "Should not custom lower this!");
4435 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4436 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4437 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4438 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4439 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4440 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4441 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4442 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4443 case ISD::SHL_PARTS:
4444 case ISD::SRA_PARTS:
4445 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4446 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4447 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4448 case ISD::FABS: return LowerFABS(Op, DAG);
4449 case ISD::FNEG: return LowerFNEG(Op, DAG);
4450 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4451 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4452 case ISD::SELECT: return LowerSELECT(Op, DAG);
4453 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4454 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4455 case ISD::CALL: return LowerCALL(Op, DAG);
4456 case ISD::RET: return LowerRET(Op, DAG);
4457 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4458 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4459 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4460 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4461 case ISD::VASTART: return LowerVASTART(Op, DAG);
4462 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4463 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4464 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4469 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4471 default: return NULL;
4472 case X86ISD::SHLD: return "X86ISD::SHLD";
4473 case X86ISD::SHRD: return "X86ISD::SHRD";
4474 case X86ISD::FAND: return "X86ISD::FAND";
4475 case X86ISD::FOR: return "X86ISD::FOR";
4476 case X86ISD::FXOR: return "X86ISD::FXOR";
4477 case X86ISD::FSRL: return "X86ISD::FSRL";
4478 case X86ISD::FILD: return "X86ISD::FILD";
4479 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4480 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4481 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4482 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4483 case X86ISD::FLD: return "X86ISD::FLD";
4484 case X86ISD::FST: return "X86ISD::FST";
4485 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4486 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4487 case X86ISD::CALL: return "X86ISD::CALL";
4488 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4489 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4490 case X86ISD::CMP: return "X86ISD::CMP";
4491 case X86ISD::COMI: return "X86ISD::COMI";
4492 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4493 case X86ISD::SETCC: return "X86ISD::SETCC";
4494 case X86ISD::CMOV: return "X86ISD::CMOV";
4495 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4496 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4497 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4498 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4499 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4500 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4501 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4502 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4503 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4504 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4505 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4506 case X86ISD::FMAX: return "X86ISD::FMAX";
4507 case X86ISD::FMIN: return "X86ISD::FMIN";
4511 /// isLegalAddressImmediate - Return true if the integer value or
4512 /// GlobalValue can be used as the offset of the target addressing mode.
4513 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
4514 // X86 allows a sign-extended 32-bit immediate field.
4515 return (V > -(1LL << 32) && V < (1LL << 32)-1);
4518 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
4519 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
4520 // field unless we are in small code model.
4521 if (Subtarget->is64Bit() &&
4522 getTargetMachine().getCodeModel() != CodeModel::Small)
4525 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4528 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4529 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4530 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4531 /// are assumed to be legal.
4533 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4534 // Only do shuffles on 128-bit vector types for now.
4535 if (MVT::getSizeInBits(VT) == 64) return false;
4536 return (Mask.Val->getNumOperands() <= 4 ||
4537 isSplatMask(Mask.Val) ||
4538 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4539 X86::isUNPCKLMask(Mask.Val) ||
4540 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4541 X86::isUNPCKHMask(Mask.Val));
4544 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4546 SelectionDAG &DAG) const {
4547 unsigned NumElts = BVOps.size();
4548 // Only do shuffles on 128-bit vector types for now.
4549 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4550 if (NumElts == 2) return true;
4552 return (isMOVLMask(&BVOps[0], 4) ||
4553 isCommutedMOVL(&BVOps[0], 4, true) ||
4554 isSHUFPMask(&BVOps[0], 4) ||
4555 isCommutedSHUFP(&BVOps[0], 4));
4560 //===----------------------------------------------------------------------===//
4561 // X86 Scheduler Hooks
4562 //===----------------------------------------------------------------------===//
4565 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4566 MachineBasicBlock *BB) {
4567 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4568 switch (MI->getOpcode()) {
4569 default: assert(false && "Unexpected instr type to insert");
4570 case X86::CMOV_FR32:
4571 case X86::CMOV_FR64:
4572 case X86::CMOV_V4F32:
4573 case X86::CMOV_V2F64:
4574 case X86::CMOV_V2I64: {
4575 // To "insert" a SELECT_CC instruction, we actually have to insert the
4576 // diamond control-flow pattern. The incoming instruction knows the
4577 // destination vreg to set, the condition code register to branch on, the
4578 // true/false values to select between, and a branch opcode to use.
4579 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4580 ilist<MachineBasicBlock>::iterator It = BB;
4586 // cmpTY ccX, r1, r2
4588 // fallthrough --> copy0MBB
4589 MachineBasicBlock *thisMBB = BB;
4590 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4591 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4593 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4594 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4595 MachineFunction *F = BB->getParent();
4596 F->getBasicBlockList().insert(It, copy0MBB);
4597 F->getBasicBlockList().insert(It, sinkMBB);
4598 // Update machine-CFG edges by first adding all successors of the current
4599 // block to the new block which will contain the Phi node for the select.
4600 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4601 e = BB->succ_end(); i != e; ++i)
4602 sinkMBB->addSuccessor(*i);
4603 // Next, remove all successors of the current block, and add the true
4604 // and fallthrough blocks as its successors.
4605 while(!BB->succ_empty())
4606 BB->removeSuccessor(BB->succ_begin());
4607 BB->addSuccessor(copy0MBB);
4608 BB->addSuccessor(sinkMBB);
4611 // %FalseValue = ...
4612 // # fallthrough to sinkMBB
4615 // Update machine-CFG edges
4616 BB->addSuccessor(sinkMBB);
4619 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4622 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4623 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4624 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4626 delete MI; // The pseudo instruction is gone now.
4630 case X86::FP_TO_INT16_IN_MEM:
4631 case X86::FP_TO_INT32_IN_MEM:
4632 case X86::FP_TO_INT64_IN_MEM: {
4633 // Change the floating point control register to use "round towards zero"
4634 // mode when truncating to an integer value.
4635 MachineFunction *F = BB->getParent();
4636 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4637 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4639 // Load the old value of the high byte of the control word...
4641 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4642 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4644 // Set the high part to be round to zero...
4645 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4648 // Reload the modified control word now...
4649 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4651 // Restore the memory image of control word to original value
4652 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4655 // Get the X86 opcode to use.
4657 switch (MI->getOpcode()) {
4658 default: assert(0 && "illegal opcode!");
4659 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4660 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4661 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4665 MachineOperand &Op = MI->getOperand(0);
4666 if (Op.isRegister()) {
4667 AM.BaseType = X86AddressMode::RegBase;
4668 AM.Base.Reg = Op.getReg();
4670 AM.BaseType = X86AddressMode::FrameIndexBase;
4671 AM.Base.FrameIndex = Op.getFrameIndex();
4673 Op = MI->getOperand(1);
4674 if (Op.isImmediate())
4675 AM.Scale = Op.getImm();
4676 Op = MI->getOperand(2);
4677 if (Op.isImmediate())
4678 AM.IndexReg = Op.getImm();
4679 Op = MI->getOperand(3);
4680 if (Op.isGlobalAddress()) {
4681 AM.GV = Op.getGlobal();
4683 AM.Disp = Op.getImm();
4685 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4686 .addReg(MI->getOperand(4).getReg());
4688 // Reload the original control word now.
4689 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4691 delete MI; // The pseudo instruction is gone now.
4697 //===----------------------------------------------------------------------===//
4698 // X86 Optimization Hooks
4699 //===----------------------------------------------------------------------===//
4701 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4703 uint64_t &KnownZero,
4705 unsigned Depth) const {
4706 unsigned Opc = Op.getOpcode();
4707 assert((Opc >= ISD::BUILTIN_OP_END ||
4708 Opc == ISD::INTRINSIC_WO_CHAIN ||
4709 Opc == ISD::INTRINSIC_W_CHAIN ||
4710 Opc == ISD::INTRINSIC_VOID) &&
4711 "Should use MaskedValueIsZero if you don't know whether Op"
4712 " is a target node!");
4714 KnownZero = KnownOne = 0; // Don't know anything.
4718 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4723 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4724 /// element of the result of the vector shuffle.
4725 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4726 MVT::ValueType VT = N->getValueType(0);
4727 SDOperand PermMask = N->getOperand(2);
4728 unsigned NumElems = PermMask.getNumOperands();
4729 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4731 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4733 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4734 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4735 SDOperand Idx = PermMask.getOperand(i);
4736 if (Idx.getOpcode() == ISD::UNDEF)
4737 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4738 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4743 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4744 /// node is a GlobalAddress + an offset.
4745 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4746 unsigned Opc = N->getOpcode();
4747 if (Opc == X86ISD::Wrapper) {
4748 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4749 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4752 } else if (Opc == ISD::ADD) {
4753 SDOperand N1 = N->getOperand(0);
4754 SDOperand N2 = N->getOperand(1);
4755 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4756 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4758 Offset += V->getSignExtended();
4761 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4762 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4764 Offset += V->getSignExtended();
4772 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4774 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4775 MachineFrameInfo *MFI) {
4776 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4779 SDOperand Loc = N->getOperand(1);
4780 SDOperand BaseLoc = Base->getOperand(1);
4781 if (Loc.getOpcode() == ISD::FrameIndex) {
4782 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4784 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4785 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4786 int FS = MFI->getObjectSize(FI);
4787 int BFS = MFI->getObjectSize(BFI);
4788 if (FS != BFS || FS != Size) return false;
4789 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4791 GlobalValue *GV1 = NULL;
4792 GlobalValue *GV2 = NULL;
4793 int64_t Offset1 = 0;
4794 int64_t Offset2 = 0;
4795 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4796 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4797 if (isGA1 && isGA2 && GV1 == GV2)
4798 return Offset1 == (Offset2 + Dist*Size);
4804 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4805 const X86Subtarget *Subtarget) {
4808 if (isGAPlusOffset(Base, GV, Offset))
4809 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4811 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4812 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4814 // Fixed objects do not specify alignment, however the offsets are known.
4815 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4816 (MFI->getObjectOffset(BFI) % 16) == 0);
4818 return MFI->getObjectAlignment(BFI) >= 16;
4824 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4825 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4826 /// if the load addresses are consecutive, non-overlapping, and in the right
4828 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4829 const X86Subtarget *Subtarget) {
4830 MachineFunction &MF = DAG.getMachineFunction();
4831 MachineFrameInfo *MFI = MF.getFrameInfo();
4832 MVT::ValueType VT = N->getValueType(0);
4833 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4834 SDOperand PermMask = N->getOperand(2);
4835 int NumElems = (int)PermMask.getNumOperands();
4836 SDNode *Base = NULL;
4837 for (int i = 0; i < NumElems; ++i) {
4838 SDOperand Idx = PermMask.getOperand(i);
4839 if (Idx.getOpcode() == ISD::UNDEF) {
4840 if (!Base) return SDOperand();
4843 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4844 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4848 else if (!isConsecutiveLoad(Arg.Val, Base,
4849 i, MVT::getSizeInBits(EVT)/8,MFI))
4854 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4856 LoadSDNode *LD = cast<LoadSDNode>(Base);
4857 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4858 LD->getSrcValueOffset());
4860 // Just use movups, it's shorter.
4861 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4862 SmallVector<SDOperand, 3> Ops;
4863 Ops.push_back(Base->getOperand(0));
4864 Ops.push_back(Base->getOperand(1));
4865 Ops.push_back(Base->getOperand(2));
4866 return DAG.getNode(ISD::BIT_CONVERT, VT,
4867 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4871 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4872 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4873 const X86Subtarget *Subtarget) {
4874 SDOperand Cond = N->getOperand(0);
4876 // If we have SSE[12] support, try to form min/max nodes.
4877 if (Subtarget->hasSSE2() &&
4878 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4879 if (Cond.getOpcode() == ISD::SETCC) {
4880 // Get the LHS/RHS of the select.
4881 SDOperand LHS = N->getOperand(1);
4882 SDOperand RHS = N->getOperand(2);
4883 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4885 unsigned Opcode = 0;
4886 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4889 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4892 if (!UnsafeFPMath) break;
4894 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4896 Opcode = X86ISD::FMIN;
4899 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4902 if (!UnsafeFPMath) break;
4904 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4906 Opcode = X86ISD::FMAX;
4909 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4912 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4915 if (!UnsafeFPMath) break;
4917 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4919 Opcode = X86ISD::FMIN;
4922 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4925 if (!UnsafeFPMath) break;
4927 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4929 Opcode = X86ISD::FMAX;
4935 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4944 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4945 DAGCombinerInfo &DCI) const {
4946 SelectionDAG &DAG = DCI.DAG;
4947 switch (N->getOpcode()) {
4949 case ISD::VECTOR_SHUFFLE:
4950 return PerformShuffleCombine(N, DAG, Subtarget);
4952 return PerformSELECTCombine(N, DAG, Subtarget);
4958 //===----------------------------------------------------------------------===//
4959 // X86 Inline Assembly Support
4960 //===----------------------------------------------------------------------===//
4962 /// getConstraintType - Given a constraint letter, return the type of
4963 /// constraint it is for this target.
4964 X86TargetLowering::ConstraintType
4965 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
4966 switch (ConstraintLetter) {
4975 return C_RegisterClass;
4976 default: return TargetLowering::getConstraintType(ConstraintLetter);
4980 /// isOperandValidForConstraint - Return the specified operand (possibly
4981 /// modified) if the specified SDOperand is valid for the specified target
4982 /// constraint letter, otherwise return null.
4983 SDOperand X86TargetLowering::
4984 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4985 switch (Constraint) {
4988 // Literal immediates are always ok.
4989 if (isa<ConstantSDNode>(Op)) return Op;
4991 // If we are in non-pic codegen mode, we allow the address of a global to
4992 // be used with 'i'.
4993 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
4994 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4995 return SDOperand(0, 0);
4997 if (GA->getOpcode() != ISD::TargetGlobalAddress)
4998 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5003 // Otherwise, not valid for this mode.
5004 return SDOperand(0, 0);
5006 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
5010 std::vector<unsigned> X86TargetLowering::
5011 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5012 MVT::ValueType VT) const {
5013 if (Constraint.size() == 1) {
5014 // FIXME: not handling fp-stack yet!
5015 // FIXME: not handling MMX registers yet ('y' constraint).
5016 switch (Constraint[0]) { // GCC X86 Constraint Letters
5017 default: break; // Unknown constraint letter
5018 case 'A': // EAX/EDX
5019 if (VT == MVT::i32 || VT == MVT::i64)
5020 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5022 case 'r': // GENERAL_REGS
5023 case 'R': // LEGACY_REGS
5024 if (VT == MVT::i64 && Subtarget->is64Bit())
5025 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
5026 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
5027 X86::R8, X86::R9, X86::R10, X86::R11,
5028 X86::R12, X86::R13, X86::R14, X86::R15, 0);
5030 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5031 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
5032 else if (VT == MVT::i16)
5033 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5034 X86::SI, X86::DI, X86::BP, X86::SP, 0);
5035 else if (VT == MVT::i8)
5036 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5038 case 'l': // INDEX_REGS
5040 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5041 X86::ESI, X86::EDI, X86::EBP, 0);
5042 else if (VT == MVT::i16)
5043 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5044 X86::SI, X86::DI, X86::BP, 0);
5045 else if (VT == MVT::i8)
5046 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5048 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5051 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5052 else if (VT == MVT::i16)
5053 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5054 else if (VT == MVT::i8)
5055 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5057 case 'x': // SSE_REGS if SSE1 allowed
5058 if (Subtarget->hasSSE1())
5059 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5060 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5062 return std::vector<unsigned>();
5063 case 'Y': // SSE_REGS if SSE2 allowed
5064 if (Subtarget->hasSSE2())
5065 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5066 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5068 return std::vector<unsigned>();
5072 return std::vector<unsigned>();
5075 std::pair<unsigned, const TargetRegisterClass*>
5076 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5077 MVT::ValueType VT) const {
5078 // Use the default implementation in TargetLowering to convert the register
5079 // constraint into a member of a register class.
5080 std::pair<unsigned, const TargetRegisterClass*> Res;
5081 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5083 // Not found as a standard register?
5084 if (Res.second == 0) {
5085 // GCC calls "st(0)" just plain "st".
5086 if (StringsEqualNoCase("{st}", Constraint)) {
5087 Res.first = X86::ST0;
5088 Res.second = X86::RSTRegisterClass;
5094 // Otherwise, check to see if this is a register class of the wrong value
5095 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5096 // turn into {ax},{dx}.
5097 if (Res.second->hasType(VT))
5098 return Res; // Correct type already, nothing to do.
5100 // All of the single-register GCC register classes map their values onto
5101 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5102 // really want an 8-bit or 32-bit register, map to the appropriate register
5103 // class and return the appropriate register.
5104 if (Res.second != X86::GR16RegisterClass)
5107 if (VT == MVT::i8) {
5108 unsigned DestReg = 0;
5109 switch (Res.first) {
5111 case X86::AX: DestReg = X86::AL; break;
5112 case X86::DX: DestReg = X86::DL; break;
5113 case X86::CX: DestReg = X86::CL; break;
5114 case X86::BX: DestReg = X86::BL; break;
5117 Res.first = DestReg;
5118 Res.second = Res.second = X86::GR8RegisterClass;
5120 } else if (VT == MVT::i32) {
5121 unsigned DestReg = 0;
5122 switch (Res.first) {
5124 case X86::AX: DestReg = X86::EAX; break;
5125 case X86::DX: DestReg = X86::EDX; break;
5126 case X86::CX: DestReg = X86::ECX; break;
5127 case X86::BX: DestReg = X86::EBX; break;
5128 case X86::SI: DestReg = X86::ESI; break;
5129 case X86::DI: DestReg = X86::EDI; break;
5130 case X86::BP: DestReg = X86::EBP; break;
5131 case X86::SP: DestReg = X86::ESP; break;
5134 Res.first = DestReg;
5135 Res.second = Res.second = X86::GR32RegisterClass;
5137 } else if (VT == MVT::i64) {
5138 unsigned DestReg = 0;
5139 switch (Res.first) {
5141 case X86::AX: DestReg = X86::RAX; break;
5142 case X86::DX: DestReg = X86::RDX; break;
5143 case X86::CX: DestReg = X86::RCX; break;
5144 case X86::BX: DestReg = X86::RBX; break;
5145 case X86::SI: DestReg = X86::RSI; break;
5146 case X86::DI: DestReg = X86::RDI; break;
5147 case X86::BP: DestReg = X86::RBP; break;
5148 case X86::SP: DestReg = X86::RSP; break;
5151 Res.first = DestReg;
5152 Res.second = Res.second = X86::GR64RegisterClass;