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()).getValue(Op.ResNo);
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 class CallingConvState {
1062 uint32_t UsedRegs[(X86::NUM_TARGET_REGS+31)/32];
1063 unsigned StackOffset;
1064 const MRegisterInfo &MRI;
1066 CallingConvState(const MRegisterInfo &mri) : MRI(mri) {
1067 // No stack is used.
1070 UsedRegs.resize(MRI.getNumRegs());
1071 // No registers are used.
1072 memset(UsedRegs, 0, sizeof(UsedRegs));
1075 unsigned getNextStackOffset() const { return StackOffset; }
1077 /// isAllocated - Return true if the specified register (or an alias) is
1079 bool isAllocated(unsigned Reg) const {
1080 return UsedRegs[Reg/32] & (1 << (Reg&31));
1083 /// getFirstUnallocated - Return the first unallocated register in the set, or
1084 /// NumRegs if they are all allocated.
1085 unsigned getFirstUnallocated(const unsigned *Regs, unsigned NumRegs) const {
1086 for (unsigned i = 0; i != NumRegs; ++i)
1087 if (!isAllocated(Regs[i]))
1092 /// AllocateReg - Attempt to allocate one of the specified registers. If none
1093 /// are available, return zero. Otherwise, return the first one available,
1094 /// marking it and any aliases as allocated.
1095 unsigned AllocateReg(const unsigned *Regs, unsigned NumRegs) {
1096 unsigned FirstUnalloc = getFirstUnallocated(Regs, NumRegs);
1097 if (FirstUnalloc == NumRegs)
1098 return 0; // Didn't find the reg.
1100 // Mark the register and any aliases as allocated.
1101 unsigned Reg = Regs[FirstUnalloc];
1103 if (const unsigned *RegAliases = MRI.getAliasSet(Reg))
1104 for (; *RegAliases; ++RegAliases)
1105 MarkAllocated(*RegAliases);
1109 /// AllocateStack - Allocate a chunk of stack space with the specified size
1111 unsigned AllocateStack(unsigned Size, unsigned Align) {
1112 assert(Align && ((Align-1) & Align) == 0); // Align is power of 2.
1113 StackOffset = ((StackOffset + Align-1) & ~(Align-1));
1114 unsigned Result = StackOffset;
1115 StackOffset += Size;
1119 void MarkAllocated(unsigned Reg) {
1120 UsedRegs[Reg/32] |= 1 << (Reg&31);
1124 /// X86_64_CCC_AssignArgument - Implement the X86-64 C Calling Convention.
1125 template<typename Client, typename DataTy>
1126 static void X86_64_CCC_AssignArgument(Client &C, CallingConvState &State,
1127 MVT::ValueType ArgVT, unsigned ArgFlags,
1129 MVT::ValueType LocVT = ArgVT;
1130 unsigned ExtendType = ISD::ANY_EXTEND;
1132 // Promote the integer to 32 bits. If the input type is signed use a
1133 // sign extend, otherwise use a zero extend.
1134 if (ArgVT == MVT::i8 || ArgVT == MVT::i16) {
1136 ExtendType = (ArgFlags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1139 // If this is a 32-bit value, assign to a 32-bit register if any are
1141 if (LocVT == MVT::i32) {
1142 static const unsigned GPR32ArgRegs[] = {
1143 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1145 if (unsigned Reg = State.AllocateReg(GPR32ArgRegs, 6)) {
1146 C.AssignToReg(Data, Reg, ArgVT, LocVT, ExtendType);
1151 // If this is a 64-bit value, assign to a 64-bit register if any are
1153 if (LocVT == MVT::i64) {
1154 static const unsigned GPR64ArgRegs[] = {
1155 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1157 if (unsigned Reg = State.AllocateReg(GPR64ArgRegs, 6)) {
1158 C.AssignToReg(Data, Reg, ArgVT, LocVT, ExtendType);
1163 // If this is a FP or vector type, assign to an XMM reg if any are
1165 if (MVT::isVector(LocVT) || MVT::isFloatingPoint(LocVT)) {
1166 static const unsigned XMMArgRegs[] = {
1167 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1168 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1170 if (unsigned Reg = State.AllocateReg(XMMArgRegs, 8)) {
1171 C.AssignToReg(Data, Reg, ArgVT, LocVT, ExtendType);
1176 // Integer/FP values get stored in stack slots that are 8 bytes in size and
1177 // 8-byte aligned if there are no more registers to hold them.
1178 if (LocVT == MVT::i32 || LocVT == MVT::i64 ||
1179 LocVT == MVT::f32 || LocVT == MVT::f64) {
1180 unsigned Offset = State.AllocateStack(8, 8);
1181 C.AssignToStack(Data, Offset, ArgVT, LocVT, ExtendType);
1185 // Vectors get 16-byte stack slots that are 16-byte aligned.
1186 if (MVT::isVector(LocVT)) {
1187 unsigned Offset = State.AllocateStack(16, 16);
1188 C.AssignToStack(Data, Offset, ArgVT, LocVT, ExtendType);
1191 assert(0 && "Unknown argument type!");
1194 class LowerArgumentsClient {
1196 X86TargetLowering &TLI;
1197 SmallVector<SDOperand, 8> &ArgValues;
1200 LowerArgumentsClient(SelectionDAG &dag, X86TargetLowering &tli,
1201 SmallVector<SDOperand, 8> &argvalues,
1203 : DAG(dag), TLI(tli), ArgValues(argvalues), Chain(chain) {
1207 void AssignToReg(SDOperand Arg, unsigned RegNo,
1208 MVT::ValueType ArgVT, MVT::ValueType RegVT,
1209 unsigned ExtendType) {
1210 TargetRegisterClass *RC = NULL;
1211 if (RegVT == MVT::i32)
1212 RC = X86::GR32RegisterClass;
1213 else if (RegVT == MVT::i64)
1214 RC = X86::GR64RegisterClass;
1215 else if (RegVT == MVT::f32)
1216 RC = X86::FR32RegisterClass;
1217 else if (RegVT == MVT::f64)
1218 RC = X86::FR64RegisterClass;
1220 RC = X86::VR128RegisterClass;
1223 SDOperand ArgValue = DAG.getCopyFromReg(Chain, RegNo, RegVT);
1224 AddLiveIn(DAG.getMachineFunction(), RegNo, RC);
1226 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1227 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1229 if (ArgVT < RegVT) {
1230 if (ExtendType == ISD::SIGN_EXTEND) {
1231 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1232 DAG.getValueType(ArgVT));
1233 } else if (ExtendType == ISD::ZERO_EXTEND) {
1234 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1235 DAG.getValueType(ArgVT));
1237 ArgValue = DAG.getNode(ISD::TRUNCATE, ArgVT, ArgValue);
1240 ArgValues.push_back(ArgValue);
1243 void AssignToStack(SDOperand Arg, unsigned Offset,
1244 MVT::ValueType ArgVT, MVT::ValueType DestVT,
1245 unsigned ExtendType) {
1246 // Create the SelectionDAG nodes corresponding to a load from this
1248 MachineFunction &MF = DAG.getMachineFunction();
1249 MachineFrameInfo *MFI = MF.getFrameInfo();
1250 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(ArgVT)/8, Offset);
1251 SDOperand FIN = DAG.getFrameIndex(FI, TLI.getPointerTy());
1252 ArgValues.push_back(DAG.getLoad(ArgVT, Chain, FIN, NULL, 0));
1256 class LowerCallArgumentsClient {
1258 X86TargetLowering &TLI;
1259 SmallVector<std::pair<unsigned, SDOperand>, 8> &RegsToPass;
1260 SmallVector<SDOperand, 8> &MemOpChains;
1264 LowerCallArgumentsClient(SelectionDAG &dag, X86TargetLowering &tli,
1265 SmallVector<std::pair<unsigned, SDOperand>, 8> &rtp,
1266 SmallVector<SDOperand, 8> &moc,
1268 : DAG(dag), TLI(tli), RegsToPass(rtp), MemOpChains(moc), Chain(chain) {
1272 void AssignToReg(SDOperand Arg, unsigned RegNo,
1273 MVT::ValueType ArgVT, MVT::ValueType RegVT,
1274 unsigned ExtendType) {
1275 // If the argument has to be extended somehow before being passed, do so.
1277 Arg = DAG.getNode(ExtendType, RegVT, Arg);
1279 RegsToPass.push_back(std::make_pair(RegNo, Arg));
1282 void AssignToStack(SDOperand Arg, unsigned Offset,
1283 MVT::ValueType ArgVT, MVT::ValueType DestVT,
1284 unsigned ExtendType) {
1285 // If the argument has to be extended somehow before being stored, do so.
1287 Arg = DAG.getNode(ExtendType, DestVT, Arg);
1289 SDOperand SP = getSP();
1290 SDOperand PtrOff = DAG.getConstant(Offset, SP.getValueType());
1291 PtrOff = DAG.getNode(ISD::ADD, SP.getValueType(), SP, PtrOff);
1292 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1296 if (StackPtr.Val == 0) {
1297 MVT::ValueType PtrTy = TLI.getPointerTy();
1298 StackPtr = DAG.getRegister(TLI.getStackPtrReg(), PtrTy);
1304 class EmptyArgumentsClient {
1306 EmptyArgumentsClient() {}
1308 void AssignToReg(SDOperand Arg, unsigned RegNo,
1309 MVT::ValueType ArgVT, MVT::ValueType RegVT,
1310 unsigned ExtendType) {
1313 void AssignToStack(SDOperand Arg, unsigned Offset,
1314 MVT::ValueType ArgVT, MVT::ValueType DestVT,
1315 unsigned ExtendType) {
1321 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1322 unsigned NumArgs = Op.Val->getNumValues() - 1;
1323 MachineFunction &MF = DAG.getMachineFunction();
1324 MachineFrameInfo *MFI = MF.getFrameInfo();
1325 SDOperand Root = Op.getOperand(0);
1326 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1328 static const unsigned GPR64ArgRegs[] = {
1329 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1331 static const unsigned XMMArgRegs[] = {
1332 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1333 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1336 SmallVector<SDOperand, 8> ArgValues;
1339 CallingConvState CCState(*getTargetMachine().getRegisterInfo());
1340 LowerArgumentsClient Client(DAG, *this, ArgValues, Root);
1342 for (unsigned i = 0; i != NumArgs; ++i) {
1343 MVT::ValueType ArgVT = Op.getValue(i).getValueType();
1344 unsigned ArgFlags = cast<ConstantSDNode>(Op.getOperand(3+i))->getValue();
1346 X86_64_CCC_AssignArgument(Client, CCState, ArgVT, ArgFlags, SDOperand());
1349 unsigned StackSize = CCState.getNextStackOffset();
1351 // If the function takes variable number of arguments, make a frame index for
1352 // the start of the first vararg value... for expansion of llvm.va_start.
1354 unsigned NumIntRegs = CCState.getFirstUnallocated(GPR64ArgRegs, 6);
1355 unsigned NumXMMRegs = CCState.getFirstUnallocated(XMMArgRegs, 8);
1357 // For X86-64, if there are vararg parameters that are passed via
1358 // registers, then we must store them to their spots on the stack so they
1359 // may be loaded by deferencing the result of va_next.
1360 VarArgsGPOffset = NumIntRegs * 8;
1361 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1362 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1363 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1365 // Store the integer parameter registers.
1366 SmallVector<SDOperand, 8> MemOps;
1367 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1368 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1369 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1370 for (; NumIntRegs != 6; ++NumIntRegs) {
1371 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1372 X86::GR64RegisterClass);
1373 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1374 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1375 MemOps.push_back(Store);
1376 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1377 DAG.getConstant(8, getPointerTy()));
1380 // Now store the XMM (fp + vector) parameter registers.
1381 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1382 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1383 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1384 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1385 X86::VR128RegisterClass);
1386 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1387 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1388 MemOps.push_back(Store);
1389 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1390 DAG.getConstant(16, getPointerTy()));
1392 if (!MemOps.empty())
1393 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1394 &MemOps[0], MemOps.size());
1397 ArgValues.push_back(Root);
1399 ReturnAddrIndex = 0; // No return address slot generated yet.
1400 BytesToPopOnReturn = 0; // Callee pops nothing.
1401 BytesCallerReserves = StackSize;
1403 // Return the new list of results.
1404 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1405 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1409 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1411 SDOperand Chain = Op.getOperand(0);
1412 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1413 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1414 SDOperand Callee = Op.getOperand(4);
1415 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1417 // Count how many bytes are to be pushed on the stack.
1418 unsigned NumBytes = 0;
1420 CallingConvState CCState(*getTargetMachine().getRegisterInfo());
1421 EmptyArgumentsClient Client;
1423 for (unsigned i = 0; i != NumOps; ++i) {
1424 SDOperand Arg = Op.getOperand(5+2*i);
1425 MVT::ValueType ArgVT = Arg.getValueType();
1427 cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1428 X86_64_CCC_AssignArgument(Client, CCState, ArgVT, ArgFlags, Arg);
1431 NumBytes = CCState.getNextStackOffset();
1435 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1437 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1438 SmallVector<SDOperand, 8> MemOpChains;
1440 CallingConvState CCState(*getTargetMachine().getRegisterInfo());
1441 LowerCallArgumentsClient Client(DAG, *this, RegsToPass, MemOpChains, Chain);
1443 for (unsigned i = 0; i != NumOps; ++i) {
1444 SDOperand Arg = Op.getOperand(5+2*i);
1445 MVT::ValueType ArgVT = Arg.getValueType();
1447 cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1448 X86_64_CCC_AssignArgument(Client, CCState, ArgVT, ArgFlags, Arg);
1451 if (!MemOpChains.empty())
1452 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1453 &MemOpChains[0], MemOpChains.size());
1455 // Build a sequence of copy-to-reg nodes chained together with token chain
1456 // and flag operands which copy the outgoing args into registers.
1458 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1459 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1461 InFlag = Chain.getValue(1);
1465 // From AMD64 ABI document:
1466 // For calls that may call functions that use varargs or stdargs
1467 // (prototype-less calls or calls to functions containing ellipsis (...) in
1468 // the declaration) %al is used as hidden argument to specify the number
1469 // of SSE registers used. The contents of %al do not need to match exactly
1470 // the number of registers, but must be an ubound on the number of SSE
1471 // registers used and is in the range 0 - 8 inclusive.
1473 // Count the number of XMM registers allocated.
1474 static const unsigned XMMArgRegs[] = {
1475 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1476 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1478 unsigned NumXMMRegs = CCState.getFirstUnallocated(XMMArgRegs, 8);
1480 Chain = DAG.getCopyToReg(Chain, X86::AL,
1481 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1482 InFlag = Chain.getValue(1);
1485 // If the callee is a GlobalAddress node (quite common, every direct call is)
1486 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1487 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1488 // We should use extra load for direct calls to dllimported functions in
1490 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1491 getTargetMachine(), true))
1492 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1493 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1494 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1496 // Returns a chain & a flag for retval copy to use.
1497 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1498 SmallVector<SDOperand, 8> Ops;
1499 Ops.push_back(Chain);
1500 Ops.push_back(Callee);
1502 // Add argument registers to the end of the list so that they are known live
1504 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1505 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1506 RegsToPass[i].second.getValueType()));
1509 Ops.push_back(InFlag);
1511 // FIXME: Do not generate X86ISD::TAILCALL for now.
1512 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1513 NodeTys, &Ops[0], Ops.size());
1514 InFlag = Chain.getValue(1);
1516 // Returns a flag for retval copy to use.
1517 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1519 Ops.push_back(Chain);
1520 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1521 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1522 Ops.push_back(InFlag);
1523 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1524 InFlag = Chain.getValue(1);
1526 // Handle result values, copying them out of physregs into vregs that we
1528 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1531 //===----------------------------------------------------------------------===//
1532 // Fast & FastCall Calling Convention implementation
1533 //===----------------------------------------------------------------------===//
1535 // The X86 'fast' calling convention passes up to two integer arguments in
1536 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
1537 // and requires that the callee pop its arguments off the stack (allowing proper
1538 // tail calls), and has the same return value conventions as C calling convs.
1540 // This calling convention always arranges for the callee pop value to be 8n+4
1541 // bytes, which is needed for tail recursion elimination and stack alignment
1544 // Note that this can be enhanced in the future to pass fp vals in registers
1545 // (when we have a global fp allocator) and do other tricks.
1547 //===----------------------------------------------------------------------===//
1548 // The X86 'fastcall' calling convention passes up to two integer arguments in
1549 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1550 // and requires that the callee pop its arguments off the stack (allowing proper
1551 // tail calls), and has the same return value conventions as C calling convs.
1553 // This calling convention always arranges for the callee pop value to be 8n+4
1554 // bytes, which is needed for tail recursion elimination and stack alignment
1557 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG,
1559 unsigned NumArgs = Op.Val->getNumValues()-1;
1560 MachineFunction &MF = DAG.getMachineFunction();
1561 MachineFrameInfo *MFI = MF.getFrameInfo();
1562 SDOperand Root = Op.getOperand(0);
1563 SmallVector<SDOperand, 8> ArgValues;
1565 // Add DAG nodes to load the arguments... On entry to a function the stack
1566 // frame looks like this:
1568 // [ESP] -- return address
1569 // [ESP + 4] -- first nonreg argument (leftmost lexically)
1570 // [ESP + 8] -- second nonreg argument, if 1st argument is <= 4 bytes in size
1572 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1574 // Keep track of the number of integer regs passed so far. This can be either
1575 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1577 unsigned NumIntRegs = 0;
1578 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1580 static const unsigned XMMArgRegs[] = {
1581 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1584 static const unsigned GPRArgRegs[][2][2] = {
1585 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1586 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1587 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1590 static const TargetRegisterClass* GPRClasses[3] = {
1591 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
1594 unsigned GPRInd = (isFastCall ? 1 : 0);
1595 for (unsigned i = 0; i < NumArgs; ++i) {
1596 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1597 unsigned ArgIncrement = 4;
1598 unsigned ObjSize = 0;
1599 unsigned ObjXMMRegs = 0;
1600 unsigned ObjIntRegs = 0;
1604 HowToPassCallArgument(ObjectVT,
1605 true, // Use as much registers as possible
1606 NumIntRegs, NumXMMRegs,
1607 (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS),
1608 ObjSize, ObjIntRegs, ObjXMMRegs);
1611 ArgIncrement = ObjSize;
1613 if (ObjIntRegs || ObjXMMRegs) {
1615 default: assert(0 && "Unhandled argument type!");
1619 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][GPRInd][NumIntRegs];
1620 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
1621 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1630 assert(!isFastCall && "Unhandled argument type!");
1631 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
1632 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1636 NumIntRegs += ObjIntRegs;
1637 NumXMMRegs += ObjXMMRegs;
1640 // XMM arguments have to be aligned on 16-byte boundary.
1642 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1643 // Create the SelectionDAG nodes corresponding to a load from this
1645 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1646 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1647 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1649 ArgOffset += ArgIncrement; // Move on to the next argument.
1652 ArgValues.push_back(ArgValue);
1655 ArgValues.push_back(Root);
1657 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1658 // arguments and the arguments after the retaddr has been pushed are aligned.
1659 if ((ArgOffset & 7) == 0)
1662 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1663 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1664 ReturnAddrIndex = 0; // No return address slot generated yet.
1665 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
1666 BytesCallerReserves = 0;
1668 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
1670 // Finally, inform the code generator which regs we return values in.
1671 switch (getValueType(MF.getFunction()->getReturnType())) {
1672 default: assert(0 && "Unknown type!");
1673 case MVT::isVoid: break;
1678 MF.addLiveOut(X86::EAX);
1681 MF.addLiveOut(X86::EAX);
1682 MF.addLiveOut(X86::EDX);
1686 MF.addLiveOut(X86::ST0);
1694 assert(!isFastCall && "Unknown result type");
1695 MF.addLiveOut(X86::XMM0);
1699 // Return the new list of results.
1700 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1701 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1704 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1706 SDOperand Chain = Op.getOperand(0);
1707 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1708 SDOperand Callee = Op.getOperand(4);
1709 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1711 // Count how many bytes are to be pushed on the stack.
1712 unsigned NumBytes = 0;
1714 // Keep track of the number of integer regs passed so far. This can be either
1715 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1717 unsigned NumIntRegs = 0;
1718 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1720 static const unsigned GPRArgRegs[][2][2] = {
1721 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1722 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1723 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1725 static const unsigned XMMArgRegs[] = {
1726 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1729 bool isFastCall = CC == CallingConv::X86_FastCall;
1730 unsigned GPRInd = isFastCall ? 1 : 0;
1731 for (unsigned i = 0; i != NumOps; ++i) {
1732 SDOperand Arg = Op.getOperand(5+2*i);
1734 switch (Arg.getValueType()) {
1735 default: assert(0 && "Unknown value type!");
1739 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1740 if (NumIntRegs < MaxNumIntRegs) {
1757 assert(!isFastCall && "Unknown value type!");
1761 // XMM arguments have to be aligned on 16-byte boundary.
1762 NumBytes = ((NumBytes + 15) / 16) * 16;
1769 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1770 // arguments and the arguments after the retaddr has been pushed are aligned.
1771 if ((NumBytes & 7) == 0)
1774 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1776 // Arguments go on the stack in reverse order, as specified by the ABI.
1777 unsigned ArgOffset = 0;
1779 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1780 SmallVector<SDOperand, 8> MemOpChains;
1781 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1782 for (unsigned i = 0; i != NumOps; ++i) {
1783 SDOperand Arg = Op.getOperand(5+2*i);
1785 switch (Arg.getValueType()) {
1786 default: assert(0 && "Unexpected ValueType for argument!");
1790 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1791 if (NumIntRegs < MaxNumIntRegs) {
1793 GPRArgRegs[Arg.getValueType()-MVT::i8][GPRInd][NumIntRegs];
1794 RegsToPass.push_back(std::make_pair(RegToUse, Arg));
1800 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1801 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1802 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1807 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1808 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1809 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1819 assert(!isFastCall && "Unexpected ValueType for argument!");
1820 if (NumXMMRegs < 4) {
1821 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1824 // XMM arguments have to be aligned on 16-byte boundary.
1825 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1826 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1827 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1828 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1835 if (!MemOpChains.empty())
1836 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1837 &MemOpChains[0], MemOpChains.size());
1839 // Build a sequence of copy-to-reg nodes chained together with token chain
1840 // and flag operands which copy the outgoing args into registers.
1842 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1843 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1845 InFlag = Chain.getValue(1);
1848 // If the callee is a GlobalAddress node (quite common, every direct call is)
1849 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1850 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1851 // We should use extra load for direct calls to dllimported functions in
1853 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1854 getTargetMachine(), true))
1855 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1856 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1857 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1859 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1861 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1862 Subtarget->isPICStyleGOT()) {
1863 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1864 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1866 InFlag = Chain.getValue(1);
1869 // Returns a chain & a flag for retval copy to use.
1870 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1871 SmallVector<SDOperand, 8> Ops;
1872 Ops.push_back(Chain);
1873 Ops.push_back(Callee);
1875 // Add argument registers to the end of the list so that they are known live
1877 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1878 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1879 RegsToPass[i].second.getValueType()));
1881 // Add an implicit use GOT pointer in EBX.
1882 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1883 Subtarget->isPICStyleGOT())
1884 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1887 Ops.push_back(InFlag);
1889 // FIXME: Do not generate X86ISD::TAILCALL for now.
1890 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1891 NodeTys, &Ops[0], Ops.size());
1892 InFlag = Chain.getValue(1);
1894 // Returns a flag for retval copy to use.
1895 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1897 Ops.push_back(Chain);
1898 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1899 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1900 Ops.push_back(InFlag);
1901 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1902 InFlag = Chain.getValue(1);
1904 // Handle result values, copying them out of physregs into vregs that we
1906 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1909 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1910 if (ReturnAddrIndex == 0) {
1911 // Set up a frame object for the return address.
1912 MachineFunction &MF = DAG.getMachineFunction();
1913 if (Subtarget->is64Bit())
1914 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1916 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1919 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1924 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1925 /// specific condition code. It returns a false if it cannot do a direct
1926 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1928 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1929 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1930 SelectionDAG &DAG) {
1931 X86CC = X86::COND_INVALID;
1933 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1934 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1935 // X > -1 -> X == 0, jump !sign.
1936 RHS = DAG.getConstant(0, RHS.getValueType());
1937 X86CC = X86::COND_NS;
1939 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1940 // X < 0 -> X == 0, jump on sign.
1941 X86CC = X86::COND_S;
1946 switch (SetCCOpcode) {
1948 case ISD::SETEQ: X86CC = X86::COND_E; break;
1949 case ISD::SETGT: X86CC = X86::COND_G; break;
1950 case ISD::SETGE: X86CC = X86::COND_GE; break;
1951 case ISD::SETLT: X86CC = X86::COND_L; break;
1952 case ISD::SETLE: X86CC = X86::COND_LE; break;
1953 case ISD::SETNE: X86CC = X86::COND_NE; break;
1954 case ISD::SETULT: X86CC = X86::COND_B; break;
1955 case ISD::SETUGT: X86CC = X86::COND_A; break;
1956 case ISD::SETULE: X86CC = X86::COND_BE; break;
1957 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1960 // On a floating point condition, the flags are set as follows:
1962 // 0 | 0 | 0 | X > Y
1963 // 0 | 0 | 1 | X < Y
1964 // 1 | 0 | 0 | X == Y
1965 // 1 | 1 | 1 | unordered
1967 switch (SetCCOpcode) {
1970 case ISD::SETEQ: X86CC = X86::COND_E; break;
1971 case ISD::SETOLT: Flip = true; // Fallthrough
1973 case ISD::SETGT: X86CC = X86::COND_A; break;
1974 case ISD::SETOLE: Flip = true; // Fallthrough
1976 case ISD::SETGE: X86CC = X86::COND_AE; break;
1977 case ISD::SETUGT: Flip = true; // Fallthrough
1979 case ISD::SETLT: X86CC = X86::COND_B; break;
1980 case ISD::SETUGE: Flip = true; // Fallthrough
1982 case ISD::SETLE: X86CC = X86::COND_BE; break;
1984 case ISD::SETNE: X86CC = X86::COND_NE; break;
1985 case ISD::SETUO: X86CC = X86::COND_P; break;
1986 case ISD::SETO: X86CC = X86::COND_NP; break;
1989 std::swap(LHS, RHS);
1992 return X86CC != X86::COND_INVALID;
1995 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1996 /// code. Current x86 isa includes the following FP cmov instructions:
1997 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1998 static bool hasFPCMov(unsigned X86CC) {
2014 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2015 /// true if Op is undef or if its value falls within the specified range (L, H].
2016 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2017 if (Op.getOpcode() == ISD::UNDEF)
2020 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2021 return (Val >= Low && Val < Hi);
2024 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2025 /// true if Op is undef or if its value equal to the specified value.
2026 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2027 if (Op.getOpcode() == ISD::UNDEF)
2029 return cast<ConstantSDNode>(Op)->getValue() == Val;
2032 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2033 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2034 bool X86::isPSHUFDMask(SDNode *N) {
2035 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2037 if (N->getNumOperands() != 4)
2040 // Check if the value doesn't reference the second vector.
2041 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2042 SDOperand Arg = N->getOperand(i);
2043 if (Arg.getOpcode() == ISD::UNDEF) continue;
2044 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2045 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
2052 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2053 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2054 bool X86::isPSHUFHWMask(SDNode *N) {
2055 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2057 if (N->getNumOperands() != 8)
2060 // Lower quadword copied in order.
2061 for (unsigned i = 0; i != 4; ++i) {
2062 SDOperand Arg = N->getOperand(i);
2063 if (Arg.getOpcode() == ISD::UNDEF) continue;
2064 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2065 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2069 // Upper quadword shuffled.
2070 for (unsigned i = 4; i != 8; ++i) {
2071 SDOperand Arg = N->getOperand(i);
2072 if (Arg.getOpcode() == ISD::UNDEF) continue;
2073 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2074 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2075 if (Val < 4 || Val > 7)
2082 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2083 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2084 bool X86::isPSHUFLWMask(SDNode *N) {
2085 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2087 if (N->getNumOperands() != 8)
2090 // Upper quadword copied in order.
2091 for (unsigned i = 4; i != 8; ++i)
2092 if (!isUndefOrEqual(N->getOperand(i), i))
2095 // Lower quadword shuffled.
2096 for (unsigned i = 0; i != 4; ++i)
2097 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2103 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2104 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2105 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2106 if (NumElems != 2 && NumElems != 4) return false;
2108 unsigned Half = NumElems / 2;
2109 for (unsigned i = 0; i < Half; ++i)
2110 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2112 for (unsigned i = Half; i < NumElems; ++i)
2113 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2119 bool X86::isSHUFPMask(SDNode *N) {
2120 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2121 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2124 /// isCommutedSHUFP - Returns true if the shuffle mask is except
2125 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2126 /// half elements to come from vector 1 (which would equal the dest.) and
2127 /// the upper half to come from vector 2.
2128 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2129 if (NumOps != 2 && NumOps != 4) return false;
2131 unsigned Half = NumOps / 2;
2132 for (unsigned i = 0; i < Half; ++i)
2133 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2135 for (unsigned i = Half; i < NumOps; ++i)
2136 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2141 static bool isCommutedSHUFP(SDNode *N) {
2142 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2143 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2146 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2147 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2148 bool X86::isMOVHLPSMask(SDNode *N) {
2149 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2151 if (N->getNumOperands() != 4)
2154 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2155 return isUndefOrEqual(N->getOperand(0), 6) &&
2156 isUndefOrEqual(N->getOperand(1), 7) &&
2157 isUndefOrEqual(N->getOperand(2), 2) &&
2158 isUndefOrEqual(N->getOperand(3), 3);
2161 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2162 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2164 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2165 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2167 if (N->getNumOperands() != 4)
2170 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2171 return isUndefOrEqual(N->getOperand(0), 2) &&
2172 isUndefOrEqual(N->getOperand(1), 3) &&
2173 isUndefOrEqual(N->getOperand(2), 2) &&
2174 isUndefOrEqual(N->getOperand(3), 3);
2177 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2178 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2179 bool X86::isMOVLPMask(SDNode *N) {
2180 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2182 unsigned NumElems = N->getNumOperands();
2183 if (NumElems != 2 && NumElems != 4)
2186 for (unsigned i = 0; i < NumElems/2; ++i)
2187 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2190 for (unsigned i = NumElems/2; i < NumElems; ++i)
2191 if (!isUndefOrEqual(N->getOperand(i), i))
2197 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2198 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2200 bool X86::isMOVHPMask(SDNode *N) {
2201 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2203 unsigned NumElems = N->getNumOperands();
2204 if (NumElems != 2 && NumElems != 4)
2207 for (unsigned i = 0; i < NumElems/2; ++i)
2208 if (!isUndefOrEqual(N->getOperand(i), i))
2211 for (unsigned i = 0; i < NumElems/2; ++i) {
2212 SDOperand Arg = N->getOperand(i + NumElems/2);
2213 if (!isUndefOrEqual(Arg, i + NumElems))
2220 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2221 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2222 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2223 bool V2IsSplat = false) {
2224 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2227 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2228 SDOperand BitI = Elts[i];
2229 SDOperand BitI1 = Elts[i+1];
2230 if (!isUndefOrEqual(BitI, j))
2233 if (isUndefOrEqual(BitI1, NumElts))
2236 if (!isUndefOrEqual(BitI1, j + NumElts))
2244 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2245 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2246 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2249 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2250 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2251 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2252 bool V2IsSplat = false) {
2253 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2256 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2257 SDOperand BitI = Elts[i];
2258 SDOperand BitI1 = Elts[i+1];
2259 if (!isUndefOrEqual(BitI, j + NumElts/2))
2262 if (isUndefOrEqual(BitI1, NumElts))
2265 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2273 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2274 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2275 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2278 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2279 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2281 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2282 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2284 unsigned NumElems = N->getNumOperands();
2285 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
2288 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2289 SDOperand BitI = N->getOperand(i);
2290 SDOperand BitI1 = N->getOperand(i+1);
2292 if (!isUndefOrEqual(BitI, j))
2294 if (!isUndefOrEqual(BitI1, j))
2301 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2302 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2303 /// MOVSD, and MOVD, i.e. setting the lowest element.
2304 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2305 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2308 if (!isUndefOrEqual(Elts[0], NumElts))
2311 for (unsigned i = 1; i < NumElts; ++i) {
2312 if (!isUndefOrEqual(Elts[i], i))
2319 bool X86::isMOVLMask(SDNode *N) {
2320 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2321 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2324 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2325 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2326 /// element of vector 2 and the other elements to come from vector 1 in order.
2327 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2328 bool V2IsSplat = false,
2329 bool V2IsUndef = false) {
2330 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2333 if (!isUndefOrEqual(Ops[0], 0))
2336 for (unsigned i = 1; i < NumOps; ++i) {
2337 SDOperand Arg = Ops[i];
2338 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2339 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2340 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2347 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2348 bool V2IsUndef = false) {
2349 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2350 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2351 V2IsSplat, V2IsUndef);
2354 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2355 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2356 bool X86::isMOVSHDUPMask(SDNode *N) {
2357 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2359 if (N->getNumOperands() != 4)
2362 // Expect 1, 1, 3, 3
2363 for (unsigned i = 0; i < 2; ++i) {
2364 SDOperand Arg = N->getOperand(i);
2365 if (Arg.getOpcode() == ISD::UNDEF) continue;
2366 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2367 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2368 if (Val != 1) return false;
2372 for (unsigned i = 2; i < 4; ++i) {
2373 SDOperand Arg = N->getOperand(i);
2374 if (Arg.getOpcode() == ISD::UNDEF) continue;
2375 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2376 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2377 if (Val != 3) return false;
2381 // Don't use movshdup if it can be done with a shufps.
2385 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2386 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2387 bool X86::isMOVSLDUPMask(SDNode *N) {
2388 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2390 if (N->getNumOperands() != 4)
2393 // Expect 0, 0, 2, 2
2394 for (unsigned i = 0; i < 2; ++i) {
2395 SDOperand Arg = N->getOperand(i);
2396 if (Arg.getOpcode() == ISD::UNDEF) continue;
2397 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2398 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2399 if (Val != 0) return false;
2403 for (unsigned i = 2; i < 4; ++i) {
2404 SDOperand Arg = N->getOperand(i);
2405 if (Arg.getOpcode() == ISD::UNDEF) continue;
2406 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2407 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2408 if (Val != 2) return false;
2412 // Don't use movshdup if it can be done with a shufps.
2416 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2417 /// a splat of a single element.
2418 static bool isSplatMask(SDNode *N) {
2419 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2421 // This is a splat operation if each element of the permute is the same, and
2422 // if the value doesn't reference the second vector.
2423 unsigned NumElems = N->getNumOperands();
2424 SDOperand ElementBase;
2426 for (; i != NumElems; ++i) {
2427 SDOperand Elt = N->getOperand(i);
2428 if (isa<ConstantSDNode>(Elt)) {
2434 if (!ElementBase.Val)
2437 for (; i != NumElems; ++i) {
2438 SDOperand Arg = N->getOperand(i);
2439 if (Arg.getOpcode() == ISD::UNDEF) continue;
2440 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2441 if (Arg != ElementBase) return false;
2444 // Make sure it is a splat of the first vector operand.
2445 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2448 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2449 /// a splat of a single element and it's a 2 or 4 element mask.
2450 bool X86::isSplatMask(SDNode *N) {
2451 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2453 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2454 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2456 return ::isSplatMask(N);
2459 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2460 /// specifies a splat of zero element.
2461 bool X86::isSplatLoMask(SDNode *N) {
2462 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2464 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2465 if (!isUndefOrEqual(N->getOperand(i), 0))
2470 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2471 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2473 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2474 unsigned NumOperands = N->getNumOperands();
2475 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2477 for (unsigned i = 0; i < NumOperands; ++i) {
2479 SDOperand Arg = N->getOperand(NumOperands-i-1);
2480 if (Arg.getOpcode() != ISD::UNDEF)
2481 Val = cast<ConstantSDNode>(Arg)->getValue();
2482 if (Val >= NumOperands) Val -= NumOperands;
2484 if (i != NumOperands - 1)
2491 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2492 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2494 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2496 // 8 nodes, but we only care about the last 4.
2497 for (unsigned i = 7; i >= 4; --i) {
2499 SDOperand Arg = N->getOperand(i);
2500 if (Arg.getOpcode() != ISD::UNDEF)
2501 Val = cast<ConstantSDNode>(Arg)->getValue();
2510 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2511 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2513 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2515 // 8 nodes, but we only care about the first 4.
2516 for (int i = 3; i >= 0; --i) {
2518 SDOperand Arg = N->getOperand(i);
2519 if (Arg.getOpcode() != ISD::UNDEF)
2520 Val = cast<ConstantSDNode>(Arg)->getValue();
2529 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2530 /// specifies a 8 element shuffle that can be broken into a pair of
2531 /// PSHUFHW and PSHUFLW.
2532 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2533 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2535 if (N->getNumOperands() != 8)
2538 // Lower quadword shuffled.
2539 for (unsigned i = 0; i != 4; ++i) {
2540 SDOperand Arg = N->getOperand(i);
2541 if (Arg.getOpcode() == ISD::UNDEF) continue;
2542 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2543 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2548 // Upper quadword shuffled.
2549 for (unsigned i = 4; i != 8; ++i) {
2550 SDOperand Arg = N->getOperand(i);
2551 if (Arg.getOpcode() == ISD::UNDEF) continue;
2552 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2553 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2554 if (Val < 4 || Val > 7)
2561 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2562 /// values in ther permute mask.
2563 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2564 SDOperand &V2, SDOperand &Mask,
2565 SelectionDAG &DAG) {
2566 MVT::ValueType VT = Op.getValueType();
2567 MVT::ValueType MaskVT = Mask.getValueType();
2568 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2569 unsigned NumElems = Mask.getNumOperands();
2570 SmallVector<SDOperand, 8> MaskVec;
2572 for (unsigned i = 0; i != NumElems; ++i) {
2573 SDOperand Arg = Mask.getOperand(i);
2574 if (Arg.getOpcode() == ISD::UNDEF) {
2575 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2578 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2579 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2581 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2583 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2587 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2588 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2591 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2592 /// match movhlps. The lower half elements should come from upper half of
2593 /// V1 (and in order), and the upper half elements should come from the upper
2594 /// half of V2 (and in order).
2595 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2596 unsigned NumElems = Mask->getNumOperands();
2599 for (unsigned i = 0, e = 2; i != e; ++i)
2600 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2602 for (unsigned i = 2; i != 4; ++i)
2603 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2608 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2609 /// is promoted to a vector.
2610 static inline bool isScalarLoadToVector(SDNode *N) {
2611 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2612 N = N->getOperand(0).Val;
2613 return ISD::isNON_EXTLoad(N);
2618 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2619 /// match movlp{s|d}. The lower half elements should come from lower half of
2620 /// V1 (and in order), and the upper half elements should come from the upper
2621 /// half of V2 (and in order). And since V1 will become the source of the
2622 /// MOVLP, it must be either a vector load or a scalar load to vector.
2623 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2624 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2626 // Is V2 is a vector load, don't do this transformation. We will try to use
2627 // load folding shufps op.
2628 if (ISD::isNON_EXTLoad(V2))
2631 unsigned NumElems = Mask->getNumOperands();
2632 if (NumElems != 2 && NumElems != 4)
2634 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2635 if (!isUndefOrEqual(Mask->getOperand(i), i))
2637 for (unsigned i = NumElems/2; i != NumElems; ++i)
2638 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2643 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2645 static bool isSplatVector(SDNode *N) {
2646 if (N->getOpcode() != ISD::BUILD_VECTOR)
2649 SDOperand SplatValue = N->getOperand(0);
2650 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2651 if (N->getOperand(i) != SplatValue)
2656 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2658 static bool isUndefShuffle(SDNode *N) {
2659 if (N->getOpcode() != ISD::BUILD_VECTOR)
2662 SDOperand V1 = N->getOperand(0);
2663 SDOperand V2 = N->getOperand(1);
2664 SDOperand Mask = N->getOperand(2);
2665 unsigned NumElems = Mask.getNumOperands();
2666 for (unsigned i = 0; i != NumElems; ++i) {
2667 SDOperand Arg = Mask.getOperand(i);
2668 if (Arg.getOpcode() != ISD::UNDEF) {
2669 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2670 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2672 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2679 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2680 /// that point to V2 points to its first element.
2681 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2682 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2684 bool Changed = false;
2685 SmallVector<SDOperand, 8> MaskVec;
2686 unsigned NumElems = Mask.getNumOperands();
2687 for (unsigned i = 0; i != NumElems; ++i) {
2688 SDOperand Arg = Mask.getOperand(i);
2689 if (Arg.getOpcode() != ISD::UNDEF) {
2690 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2691 if (Val > NumElems) {
2692 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2696 MaskVec.push_back(Arg);
2700 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2701 &MaskVec[0], MaskVec.size());
2705 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2706 /// operation of specified width.
2707 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2708 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2709 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2711 SmallVector<SDOperand, 8> MaskVec;
2712 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2713 for (unsigned i = 1; i != NumElems; ++i)
2714 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2715 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2718 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2719 /// of specified width.
2720 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2721 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2722 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2723 SmallVector<SDOperand, 8> MaskVec;
2724 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2725 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2726 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2728 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2731 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2732 /// of specified width.
2733 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2734 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2735 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2736 unsigned Half = NumElems/2;
2737 SmallVector<SDOperand, 8> MaskVec;
2738 for (unsigned i = 0; i != Half; ++i) {
2739 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2740 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2742 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2745 /// getZeroVector - Returns a vector of specified type with all zero elements.
2747 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2748 assert(MVT::isVector(VT) && "Expected a vector type");
2749 unsigned NumElems = getVectorNumElements(VT);
2750 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2751 bool isFP = MVT::isFloatingPoint(EVT);
2752 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2753 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2754 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2757 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2759 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2760 SDOperand V1 = Op.getOperand(0);
2761 SDOperand Mask = Op.getOperand(2);
2762 MVT::ValueType VT = Op.getValueType();
2763 unsigned NumElems = Mask.getNumOperands();
2764 Mask = getUnpacklMask(NumElems, DAG);
2765 while (NumElems != 4) {
2766 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2769 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2771 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2772 Mask = getZeroVector(MaskVT, DAG);
2773 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2774 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2775 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2778 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2780 static inline bool isZeroNode(SDOperand Elt) {
2781 return ((isa<ConstantSDNode>(Elt) &&
2782 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2783 (isa<ConstantFPSDNode>(Elt) &&
2784 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2787 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2788 /// vector and zero or undef vector.
2789 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2790 unsigned NumElems, unsigned Idx,
2791 bool isZero, SelectionDAG &DAG) {
2792 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2793 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2794 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2795 SDOperand Zero = DAG.getConstant(0, EVT);
2796 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2797 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2798 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2799 &MaskVec[0], MaskVec.size());
2800 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2803 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2805 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2806 unsigned NumNonZero, unsigned NumZero,
2807 SelectionDAG &DAG, TargetLowering &TLI) {
2813 for (unsigned i = 0; i < 16; ++i) {
2814 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2815 if (ThisIsNonZero && First) {
2817 V = getZeroVector(MVT::v8i16, DAG);
2819 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2824 SDOperand ThisElt(0, 0), LastElt(0, 0);
2825 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2826 if (LastIsNonZero) {
2827 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2829 if (ThisIsNonZero) {
2830 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2831 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2832 ThisElt, DAG.getConstant(8, MVT::i8));
2834 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2839 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2840 DAG.getConstant(i/2, TLI.getPointerTy()));
2844 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2847 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2849 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2850 unsigned NumNonZero, unsigned NumZero,
2851 SelectionDAG &DAG, TargetLowering &TLI) {
2857 for (unsigned i = 0; i < 8; ++i) {
2858 bool isNonZero = (NonZeros & (1 << i)) != 0;
2862 V = getZeroVector(MVT::v8i16, DAG);
2864 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2867 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2868 DAG.getConstant(i, TLI.getPointerTy()));
2876 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2877 // All zero's are handled with pxor.
2878 if (ISD::isBuildVectorAllZeros(Op.Val))
2881 // All one's are handled with pcmpeqd.
2882 if (ISD::isBuildVectorAllOnes(Op.Val))
2885 MVT::ValueType VT = Op.getValueType();
2886 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2887 unsigned EVTBits = MVT::getSizeInBits(EVT);
2889 unsigned NumElems = Op.getNumOperands();
2890 unsigned NumZero = 0;
2891 unsigned NumNonZero = 0;
2892 unsigned NonZeros = 0;
2893 std::set<SDOperand> Values;
2894 for (unsigned i = 0; i < NumElems; ++i) {
2895 SDOperand Elt = Op.getOperand(i);
2896 if (Elt.getOpcode() != ISD::UNDEF) {
2898 if (isZeroNode(Elt))
2901 NonZeros |= (1 << i);
2907 if (NumNonZero == 0)
2908 // Must be a mix of zero and undef. Return a zero vector.
2909 return getZeroVector(VT, DAG);
2911 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2912 if (Values.size() == 1)
2915 // Special case for single non-zero element.
2916 if (NumNonZero == 1) {
2917 unsigned Idx = CountTrailingZeros_32(NonZeros);
2918 SDOperand Item = Op.getOperand(Idx);
2919 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2921 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2922 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2925 if (EVTBits == 32) {
2926 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2927 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2929 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2930 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2931 SmallVector<SDOperand, 8> MaskVec;
2932 for (unsigned i = 0; i < NumElems; i++)
2933 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2934 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2935 &MaskVec[0], MaskVec.size());
2936 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2937 DAG.getNode(ISD::UNDEF, VT), Mask);
2941 // Let legalizer expand 2-wide build_vector's.
2945 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2947 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2949 if (V.Val) return V;
2952 if (EVTBits == 16) {
2953 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2955 if (V.Val) return V;
2958 // If element VT is == 32 bits, turn it into a number of shuffles.
2959 SmallVector<SDOperand, 8> V;
2961 if (NumElems == 4 && NumZero > 0) {
2962 for (unsigned i = 0; i < 4; ++i) {
2963 bool isZero = !(NonZeros & (1 << i));
2965 V[i] = getZeroVector(VT, DAG);
2967 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2970 for (unsigned i = 0; i < 2; ++i) {
2971 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2974 V[i] = V[i*2]; // Must be a zero vector.
2977 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2978 getMOVLMask(NumElems, DAG));
2981 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2982 getMOVLMask(NumElems, DAG));
2985 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2986 getUnpacklMask(NumElems, DAG));
2991 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2992 // clears the upper bits.
2993 // FIXME: we can do the same for v4f32 case when we know both parts of
2994 // the lower half come from scalar_to_vector (loadf32). We should do
2995 // that in post legalizer dag combiner with target specific hooks.
2996 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2998 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2999 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
3000 SmallVector<SDOperand, 8> MaskVec;
3001 bool Reverse = (NonZeros & 0x3) == 2;
3002 for (unsigned i = 0; i < 2; ++i)
3004 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3006 MaskVec.push_back(DAG.getConstant(i, EVT));
3007 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3008 for (unsigned i = 0; i < 2; ++i)
3010 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3012 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3013 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3014 &MaskVec[0], MaskVec.size());
3015 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3018 if (Values.size() > 2) {
3019 // Expand into a number of unpckl*.
3021 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3022 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3023 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3024 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3025 for (unsigned i = 0; i < NumElems; ++i)
3026 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3028 while (NumElems != 0) {
3029 for (unsigned i = 0; i < NumElems; ++i)
3030 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3041 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3042 SDOperand V1 = Op.getOperand(0);
3043 SDOperand V2 = Op.getOperand(1);
3044 SDOperand PermMask = Op.getOperand(2);
3045 MVT::ValueType VT = Op.getValueType();
3046 unsigned NumElems = PermMask.getNumOperands();
3047 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3048 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3049 bool V1IsSplat = false;
3050 bool V2IsSplat = false;
3052 if (isUndefShuffle(Op.Val))
3053 return DAG.getNode(ISD::UNDEF, VT);
3055 if (isSplatMask(PermMask.Val)) {
3056 if (NumElems <= 4) return Op;
3057 // Promote it to a v4i32 splat.
3058 return PromoteSplat(Op, DAG);
3061 if (X86::isMOVLMask(PermMask.Val))
3062 return (V1IsUndef) ? V2 : Op;
3064 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3065 X86::isMOVSLDUPMask(PermMask.Val) ||
3066 X86::isMOVHLPSMask(PermMask.Val) ||
3067 X86::isMOVHPMask(PermMask.Val) ||
3068 X86::isMOVLPMask(PermMask.Val))
3071 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3072 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3073 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3075 bool Commuted = false;
3076 V1IsSplat = isSplatVector(V1.Val);
3077 V2IsSplat = isSplatVector(V2.Val);
3078 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3079 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3080 std::swap(V1IsSplat, V2IsSplat);
3081 std::swap(V1IsUndef, V2IsUndef);
3085 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3086 if (V2IsUndef) return V1;
3087 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3089 // V2 is a splat, so the mask may be malformed. That is, it may point
3090 // to any V2 element. The instruction selectior won't like this. Get
3091 // a corrected mask and commute to form a proper MOVS{S|D}.
3092 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3093 if (NewMask.Val != PermMask.Val)
3094 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3099 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3100 X86::isUNPCKLMask(PermMask.Val) ||
3101 X86::isUNPCKHMask(PermMask.Val))
3105 // Normalize mask so all entries that point to V2 points to its first
3106 // element then try to match unpck{h|l} again. If match, return a
3107 // new vector_shuffle with the corrected mask.
3108 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3109 if (NewMask.Val != PermMask.Val) {
3110 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3111 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3112 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3113 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3114 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3115 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3120 // Normalize the node to match x86 shuffle ops if needed
3121 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3122 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3125 // Commute is back and try unpck* again.
3126 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3127 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3128 X86::isUNPCKLMask(PermMask.Val) ||
3129 X86::isUNPCKHMask(PermMask.Val))
3133 // If VT is integer, try PSHUF* first, then SHUFP*.
3134 if (MVT::isInteger(VT)) {
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);
3144 if (X86::isSHUFPMask(PermMask.Val))
3147 // Handle v8i16 shuffle high / low shuffle node pair.
3148 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3149 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3150 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3151 SmallVector<SDOperand, 8> MaskVec;
3152 for (unsigned i = 0; i != 4; ++i)
3153 MaskVec.push_back(PermMask.getOperand(i));
3154 for (unsigned i = 4; i != 8; ++i)
3155 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3156 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3157 &MaskVec[0], MaskVec.size());
3158 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3160 for (unsigned i = 0; i != 4; ++i)
3161 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3162 for (unsigned i = 4; i != 8; ++i)
3163 MaskVec.push_back(PermMask.getOperand(i));
3164 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3165 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3168 // Floating point cases in the other order.
3169 if (X86::isSHUFPMask(PermMask.Val))
3171 if (X86::isPSHUFDMask(PermMask.Val) ||
3172 X86::isPSHUFHWMask(PermMask.Val) ||
3173 X86::isPSHUFLWMask(PermMask.Val)) {
3174 if (V2.getOpcode() != ISD::UNDEF)
3175 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3176 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3181 if (NumElems == 4) {
3182 MVT::ValueType MaskVT = PermMask.getValueType();
3183 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
3184 SmallVector<std::pair<int, int>, 8> Locs;
3185 Locs.reserve(NumElems);
3186 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3187 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3190 // If no more than two elements come from either vector. This can be
3191 // implemented with two shuffles. First shuffle gather the elements.
3192 // The second shuffle, which takes the first shuffle as both of its
3193 // vector operands, put the elements into the right order.
3194 for (unsigned i = 0; i != NumElems; ++i) {
3195 SDOperand Elt = PermMask.getOperand(i);
3196 if (Elt.getOpcode() == ISD::UNDEF) {
3197 Locs[i] = std::make_pair(-1, -1);
3199 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3200 if (Val < NumElems) {
3201 Locs[i] = std::make_pair(0, NumLo);
3205 Locs[i] = std::make_pair(1, NumHi);
3206 if (2+NumHi < NumElems)
3207 Mask1[2+NumHi] = Elt;
3212 if (NumLo <= 2 && NumHi <= 2) {
3213 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3214 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3215 &Mask1[0], Mask1.size()));
3216 for (unsigned i = 0; i != NumElems; ++i) {
3217 if (Locs[i].first == -1)
3220 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3221 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3222 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3226 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3227 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3228 &Mask2[0], Mask2.size()));
3231 // Break it into (shuffle shuffle_hi, shuffle_lo).
3233 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3234 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3235 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3236 unsigned MaskIdx = 0;
3238 unsigned HiIdx = NumElems/2;
3239 for (unsigned i = 0; i != NumElems; ++i) {
3240 if (i == NumElems/2) {
3246 SDOperand Elt = PermMask.getOperand(i);
3247 if (Elt.getOpcode() == ISD::UNDEF) {
3248 Locs[i] = std::make_pair(-1, -1);
3249 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3250 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3251 (*MaskPtr)[LoIdx] = Elt;
3254 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3255 (*MaskPtr)[HiIdx] = Elt;
3260 SDOperand LoShuffle =
3261 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3262 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3263 &LoMask[0], LoMask.size()));
3264 SDOperand HiShuffle =
3265 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3266 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3267 &HiMask[0], HiMask.size()));
3268 SmallVector<SDOperand, 8> MaskOps;
3269 for (unsigned i = 0; i != NumElems; ++i) {
3270 if (Locs[i].first == -1) {
3271 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3273 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3274 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3277 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3278 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3279 &MaskOps[0], MaskOps.size()));
3286 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3287 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3290 MVT::ValueType VT = Op.getValueType();
3291 // TODO: handle v16i8.
3292 if (MVT::getSizeInBits(VT) == 16) {
3293 // Transform it so it match pextrw which produces a 32-bit result.
3294 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3295 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3296 Op.getOperand(0), Op.getOperand(1));
3297 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3298 DAG.getValueType(VT));
3299 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3300 } else if (MVT::getSizeInBits(VT) == 32) {
3301 SDOperand Vec = Op.getOperand(0);
3302 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3305 // SHUFPS the element to the lowest double word, then movss.
3306 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3307 SmallVector<SDOperand, 8> IdxVec;
3308 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
3309 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3310 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3311 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3312 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3313 &IdxVec[0], IdxVec.size());
3314 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3315 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3316 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3317 DAG.getConstant(0, getPointerTy()));
3318 } else if (MVT::getSizeInBits(VT) == 64) {
3319 SDOperand Vec = Op.getOperand(0);
3320 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3324 // UNPCKHPD the element to the lowest double word, then movsd.
3325 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3326 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3327 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3328 SmallVector<SDOperand, 8> IdxVec;
3329 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
3330 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3331 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3332 &IdxVec[0], IdxVec.size());
3333 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3334 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3335 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3336 DAG.getConstant(0, getPointerTy()));
3343 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3344 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3345 // as its second argument.
3346 MVT::ValueType VT = Op.getValueType();
3347 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
3348 SDOperand N0 = Op.getOperand(0);
3349 SDOperand N1 = Op.getOperand(1);
3350 SDOperand N2 = Op.getOperand(2);
3351 if (MVT::getSizeInBits(BaseVT) == 16) {
3352 if (N1.getValueType() != MVT::i32)
3353 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3354 if (N2.getValueType() != MVT::i32)
3355 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
3356 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3357 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3358 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3361 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3362 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3363 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3364 SmallVector<SDOperand, 8> MaskVec;
3365 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3366 for (unsigned i = 1; i <= 3; ++i)
3367 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3368 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3369 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3370 &MaskVec[0], MaskVec.size()));
3372 // Use two pinsrw instructions to insert a 32 bit value.
3374 if (MVT::isFloatingPoint(N1.getValueType())) {
3375 if (ISD::isNON_EXTLoad(N1.Val)) {
3376 // Just load directly from f32mem to GR32.
3377 LoadSDNode *LD = cast<LoadSDNode>(N1);
3378 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
3379 LD->getSrcValue(), LD->getSrcValueOffset());
3381 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3382 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3383 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3384 DAG.getConstant(0, getPointerTy()));
3387 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3388 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3389 DAG.getConstant(Idx, getPointerTy()));
3390 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3391 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3392 DAG.getConstant(Idx+1, getPointerTy()));
3393 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3401 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3402 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3403 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3406 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3407 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3408 // one of the above mentioned nodes. It has to be wrapped because otherwise
3409 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3410 // be used to form addressing mode. These wrapped nodes will be selected
3413 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3414 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3415 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3417 CP->getAlignment());
3418 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3419 // With PIC, the address is actually $g + Offset.
3420 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3421 !Subtarget->isPICStyleRIPRel()) {
3422 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3423 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3431 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3432 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3433 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3434 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3435 // With PIC, the address is actually $g + Offset.
3436 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3437 !Subtarget->isPICStyleRIPRel()) {
3438 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3439 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3443 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3444 // load the value at address GV, not the value of GV itself. This means that
3445 // the GlobalAddress must be in the base or index register of the address, not
3446 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3447 // The same applies for external symbols during PIC codegen
3448 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3449 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3455 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3456 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3457 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3458 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3459 // With PIC, the address is actually $g + Offset.
3460 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3461 !Subtarget->isPICStyleRIPRel()) {
3462 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3463 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3470 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3471 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3472 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3473 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3474 // With PIC, the address is actually $g + Offset.
3475 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3476 !Subtarget->isPICStyleRIPRel()) {
3477 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3478 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3485 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3486 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3487 "Not an i64 shift!");
3488 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3489 SDOperand ShOpLo = Op.getOperand(0);
3490 SDOperand ShOpHi = Op.getOperand(1);
3491 SDOperand ShAmt = Op.getOperand(2);
3492 SDOperand Tmp1 = isSRA ?
3493 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3494 DAG.getConstant(0, MVT::i32);
3496 SDOperand Tmp2, Tmp3;
3497 if (Op.getOpcode() == ISD::SHL_PARTS) {
3498 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3499 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3501 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3502 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3505 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3506 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3507 DAG.getConstant(32, MVT::i8));
3508 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3509 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3512 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3514 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3515 SmallVector<SDOperand, 4> Ops;
3516 if (Op.getOpcode() == ISD::SHL_PARTS) {
3517 Ops.push_back(Tmp2);
3518 Ops.push_back(Tmp3);
3520 Ops.push_back(InFlag);
3521 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3522 InFlag = Hi.getValue(1);
3525 Ops.push_back(Tmp3);
3526 Ops.push_back(Tmp1);
3528 Ops.push_back(InFlag);
3529 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3531 Ops.push_back(Tmp2);
3532 Ops.push_back(Tmp3);
3534 Ops.push_back(InFlag);
3535 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3536 InFlag = Lo.getValue(1);
3539 Ops.push_back(Tmp3);
3540 Ops.push_back(Tmp1);
3542 Ops.push_back(InFlag);
3543 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3546 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3550 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3553 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3554 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3555 Op.getOperand(0).getValueType() >= MVT::i16 &&
3556 "Unknown SINT_TO_FP to lower!");
3559 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3560 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3561 MachineFunction &MF = DAG.getMachineFunction();
3562 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3563 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3564 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3565 StackSlot, NULL, 0);
3570 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3572 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3573 SmallVector<SDOperand, 8> Ops;
3574 Ops.push_back(Chain);
3575 Ops.push_back(StackSlot);
3576 Ops.push_back(DAG.getValueType(SrcVT));
3577 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3578 Tys, &Ops[0], Ops.size());
3581 Chain = Result.getValue(1);
3582 SDOperand InFlag = Result.getValue(2);
3584 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3585 // shouldn't be necessary except that RFP cannot be live across
3586 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3587 MachineFunction &MF = DAG.getMachineFunction();
3588 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3589 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3590 Tys = DAG.getVTList(MVT::Other);
3591 SmallVector<SDOperand, 8> Ops;
3592 Ops.push_back(Chain);
3593 Ops.push_back(Result);
3594 Ops.push_back(StackSlot);
3595 Ops.push_back(DAG.getValueType(Op.getValueType()));
3596 Ops.push_back(InFlag);
3597 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3598 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3604 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3605 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3606 "Unknown FP_TO_SINT to lower!");
3607 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3609 MachineFunction &MF = DAG.getMachineFunction();
3610 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3611 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3612 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3615 switch (Op.getValueType()) {
3616 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3617 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3618 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3619 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3622 SDOperand Chain = DAG.getEntryNode();
3623 SDOperand Value = Op.getOperand(0);
3625 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3626 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3627 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3629 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3631 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3632 Chain = Value.getValue(1);
3633 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3634 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3637 // Build the FP_TO_INT*_IN_MEM
3638 SDOperand Ops[] = { Chain, Value, StackSlot };
3639 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3642 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3645 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3646 MVT::ValueType VT = Op.getValueType();
3647 const Type *OpNTy = MVT::getTypeForValueType(VT);
3648 std::vector<Constant*> CV;
3649 if (VT == MVT::f64) {
3650 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3651 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3653 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3654 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3655 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3656 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3658 Constant *CS = ConstantStruct::get(CV);
3659 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3660 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3661 SmallVector<SDOperand, 3> Ops;
3662 Ops.push_back(DAG.getEntryNode());
3663 Ops.push_back(CPIdx);
3664 Ops.push_back(DAG.getSrcValue(NULL));
3665 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3666 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3669 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3670 MVT::ValueType VT = Op.getValueType();
3671 const Type *OpNTy = MVT::getTypeForValueType(VT);
3672 std::vector<Constant*> CV;
3673 if (VT == MVT::f64) {
3674 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3675 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3677 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3678 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3679 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3680 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3682 Constant *CS = ConstantStruct::get(CV);
3683 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3684 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3685 SmallVector<SDOperand, 3> Ops;
3686 Ops.push_back(DAG.getEntryNode());
3687 Ops.push_back(CPIdx);
3688 Ops.push_back(DAG.getSrcValue(NULL));
3689 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3690 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3693 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3694 SDOperand Op0 = Op.getOperand(0);
3695 SDOperand Op1 = Op.getOperand(1);
3696 MVT::ValueType VT = Op.getValueType();
3697 MVT::ValueType SrcVT = Op1.getValueType();
3698 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3700 // If second operand is smaller, extend it first.
3701 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3702 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3706 // First get the sign bit of second operand.
3707 std::vector<Constant*> CV;
3708 if (SrcVT == MVT::f64) {
3709 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3710 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3712 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3713 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3714 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3715 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3717 Constant *CS = ConstantStruct::get(CV);
3718 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3719 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3720 SmallVector<SDOperand, 3> Ops;
3721 Ops.push_back(DAG.getEntryNode());
3722 Ops.push_back(CPIdx);
3723 Ops.push_back(DAG.getSrcValue(NULL));
3724 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3725 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3727 // Shift sign bit right or left if the two operands have different types.
3728 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3729 // Op0 is MVT::f32, Op1 is MVT::f64.
3730 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3731 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3732 DAG.getConstant(32, MVT::i32));
3733 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3734 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3735 DAG.getConstant(0, getPointerTy()));
3738 // Clear first operand sign bit.
3740 if (VT == MVT::f64) {
3741 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3742 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3744 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3745 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3746 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3747 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3749 CS = ConstantStruct::get(CV);
3750 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3751 Tys = DAG.getVTList(VT, MVT::Other);
3753 Ops.push_back(DAG.getEntryNode());
3754 Ops.push_back(CPIdx);
3755 Ops.push_back(DAG.getSrcValue(NULL));
3756 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3757 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3759 // Or the value with the sign bit.
3760 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3763 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3765 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3767 SDOperand Op0 = Op.getOperand(0);
3768 SDOperand Op1 = Op.getOperand(1);
3769 SDOperand CC = Op.getOperand(2);
3770 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3771 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3772 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3773 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3776 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3778 SDOperand Ops1[] = { Chain, Op0, Op1 };
3779 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3780 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3781 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3784 assert(isFP && "Illegal integer SetCC!");
3786 SDOperand COps[] = { Chain, Op0, Op1 };
3787 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3789 switch (SetCCOpcode) {
3790 default: assert(false && "Illegal floating point SetCC!");
3791 case ISD::SETOEQ: { // !PF & ZF
3792 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3793 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3794 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3796 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3797 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3799 case ISD::SETUNE: { // PF | !ZF
3800 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3801 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3802 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3804 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3805 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3810 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3811 bool addTest = true;
3812 SDOperand Chain = DAG.getEntryNode();
3813 SDOperand Cond = Op.getOperand(0);
3815 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3817 if (Cond.getOpcode() == ISD::SETCC)
3818 Cond = LowerSETCC(Cond, DAG, Chain);
3820 if (Cond.getOpcode() == X86ISD::SETCC) {
3821 CC = Cond.getOperand(0);
3823 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3824 // (since flag operand cannot be shared). Use it as the condition setting
3825 // operand in place of the X86ISD::SETCC.
3826 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3827 // to use a test instead of duplicating the X86ISD::CMP (for register
3828 // pressure reason)?
3829 SDOperand Cmp = Cond.getOperand(1);
3830 unsigned Opc = Cmp.getOpcode();
3831 bool IllegalFPCMov = !X86ScalarSSE &&
3832 MVT::isFloatingPoint(Op.getValueType()) &&
3833 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3834 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3836 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3837 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3843 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3844 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3845 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3848 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3849 SmallVector<SDOperand, 4> Ops;
3850 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3851 // condition is true.
3852 Ops.push_back(Op.getOperand(2));
3853 Ops.push_back(Op.getOperand(1));
3855 Ops.push_back(Cond.getValue(1));
3856 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3859 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3860 bool addTest = true;
3861 SDOperand Chain = Op.getOperand(0);
3862 SDOperand Cond = Op.getOperand(1);
3863 SDOperand Dest = Op.getOperand(2);
3865 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3867 if (Cond.getOpcode() == ISD::SETCC)
3868 Cond = LowerSETCC(Cond, DAG, Chain);
3870 if (Cond.getOpcode() == X86ISD::SETCC) {
3871 CC = Cond.getOperand(0);
3873 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3874 // (since flag operand cannot be shared). Use it as the condition setting
3875 // operand in place of the X86ISD::SETCC.
3876 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3877 // to use a test instead of duplicating the X86ISD::CMP (for register
3878 // pressure reason)?
3879 SDOperand Cmp = Cond.getOperand(1);
3880 unsigned Opc = Cmp.getOpcode();
3881 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3882 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3883 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3889 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3890 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3891 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3893 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3894 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3897 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3898 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3900 if (Subtarget->is64Bit())
3901 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3903 switch (CallingConv) {
3905 assert(0 && "Unsupported calling convention");
3906 case CallingConv::Fast:
3908 return LowerFastCCCallTo(Op, DAG, CallingConv);
3910 case CallingConv::C:
3911 case CallingConv::X86_StdCall:
3912 return LowerCCCCallTo(Op, DAG, CallingConv);
3913 case CallingConv::X86_FastCall:
3914 return LowerFastCCCallTo(Op, DAG, CallingConv);
3919 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3920 MachineFunction &MF = DAG.getMachineFunction();
3921 const Function* Fn = MF.getFunction();
3922 if (Fn->hasExternalLinkage() &&
3923 Subtarget->isTargetCygMing() &&
3924 Fn->getName() == "main")
3925 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
3927 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3928 if (Subtarget->is64Bit())
3929 return LowerX86_64CCCArguments(Op, DAG);
3933 assert(0 && "Unsupported calling convention");
3934 case CallingConv::Fast:
3936 return LowerFastCCArguments(Op, DAG);
3939 case CallingConv::C:
3940 return LowerCCCArguments(Op, DAG);
3941 case CallingConv::X86_StdCall:
3942 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
3943 return LowerCCCArguments(Op, DAG, true);
3944 case CallingConv::X86_FastCall:
3945 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
3946 return LowerFastCCArguments(Op, DAG, true);
3950 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3951 SDOperand InFlag(0, 0);
3952 SDOperand Chain = Op.getOperand(0);
3954 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3955 if (Align == 0) Align = 1;
3957 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3958 // If not DWORD aligned, call memset if size is less than the threshold.
3959 // It knows how to align to the right boundary first.
3960 if ((Align & 3) != 0 ||
3961 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3962 MVT::ValueType IntPtr = getPointerTy();
3963 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3964 TargetLowering::ArgListTy Args;
3965 TargetLowering::ArgListEntry Entry;
3966 Entry.Node = Op.getOperand(1);
3967 Entry.Ty = IntPtrTy;
3968 Entry.isSigned = false;
3969 Entry.isInReg = false;
3970 Entry.isSRet = false;
3971 Args.push_back(Entry);
3972 // Extend the unsigned i8 argument to be an int value for the call.
3973 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3974 Entry.Ty = IntPtrTy;
3975 Entry.isSigned = false;
3976 Entry.isInReg = false;
3977 Entry.isSRet = false;
3978 Args.push_back(Entry);
3979 Entry.Node = Op.getOperand(3);
3980 Args.push_back(Entry);
3981 std::pair<SDOperand,SDOperand> CallResult =
3982 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3983 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3984 return CallResult.second;
3989 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3990 unsigned BytesLeft = 0;
3991 bool TwoRepStos = false;
3994 uint64_t Val = ValC->getValue() & 255;
3996 // If the value is a constant, then we can potentially use larger sets.
3997 switch (Align & 3) {
3998 case 2: // WORD aligned
4001 Val = (Val << 8) | Val;
4003 case 0: // DWORD aligned
4006 Val = (Val << 8) | Val;
4007 Val = (Val << 16) | Val;
4008 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4011 Val = (Val << 32) | Val;
4014 default: // Byte aligned
4017 Count = Op.getOperand(3);
4021 if (AVT > MVT::i8) {
4023 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4024 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4025 BytesLeft = I->getValue() % UBytes;
4027 assert(AVT >= MVT::i32 &&
4028 "Do not use rep;stos if not at least DWORD aligned");
4029 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4030 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4035 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4037 InFlag = Chain.getValue(1);
4040 Count = Op.getOperand(3);
4041 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4042 InFlag = Chain.getValue(1);
4045 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4047 InFlag = Chain.getValue(1);
4048 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4049 Op.getOperand(1), InFlag);
4050 InFlag = Chain.getValue(1);
4052 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4053 SmallVector<SDOperand, 8> Ops;
4054 Ops.push_back(Chain);
4055 Ops.push_back(DAG.getValueType(AVT));
4056 Ops.push_back(InFlag);
4057 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4060 InFlag = Chain.getValue(1);
4061 Count = Op.getOperand(3);
4062 MVT::ValueType CVT = Count.getValueType();
4063 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4064 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4065 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4067 InFlag = Chain.getValue(1);
4068 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4070 Ops.push_back(Chain);
4071 Ops.push_back(DAG.getValueType(MVT::i8));
4072 Ops.push_back(InFlag);
4073 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4074 } else if (BytesLeft) {
4075 // Issue stores for the last 1 - 7 bytes.
4077 unsigned Val = ValC->getValue() & 255;
4078 unsigned Offset = I->getValue() - BytesLeft;
4079 SDOperand DstAddr = Op.getOperand(1);
4080 MVT::ValueType AddrVT = DstAddr.getValueType();
4081 if (BytesLeft >= 4) {
4082 Val = (Val << 8) | Val;
4083 Val = (Val << 16) | Val;
4084 Value = DAG.getConstant(Val, MVT::i32);
4085 Chain = DAG.getStore(Chain, Value,
4086 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4087 DAG.getConstant(Offset, AddrVT)),
4092 if (BytesLeft >= 2) {
4093 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4094 Chain = DAG.getStore(Chain, Value,
4095 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4096 DAG.getConstant(Offset, AddrVT)),
4101 if (BytesLeft == 1) {
4102 Value = DAG.getConstant(Val, MVT::i8);
4103 Chain = DAG.getStore(Chain, Value,
4104 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4105 DAG.getConstant(Offset, AddrVT)),
4113 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
4114 SDOperand Chain = Op.getOperand(0);
4116 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4117 if (Align == 0) Align = 1;
4119 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4120 // If not DWORD aligned, call memcpy if size is less than the threshold.
4121 // It knows how to align to the right boundary first.
4122 if ((Align & 3) != 0 ||
4123 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4124 MVT::ValueType IntPtr = getPointerTy();
4125 TargetLowering::ArgListTy Args;
4126 TargetLowering::ArgListEntry Entry;
4127 Entry.Ty = getTargetData()->getIntPtrType();
4128 Entry.isSigned = false;
4129 Entry.isInReg = false;
4130 Entry.isSRet = false;
4131 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
4132 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
4133 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
4134 std::pair<SDOperand,SDOperand> CallResult =
4135 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4136 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
4137 return CallResult.second;
4142 unsigned BytesLeft = 0;
4143 bool TwoRepMovs = false;
4144 switch (Align & 3) {
4145 case 2: // WORD aligned
4148 case 0: // DWORD aligned
4150 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4153 default: // Byte aligned
4155 Count = Op.getOperand(3);
4159 if (AVT > MVT::i8) {
4161 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4162 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4163 BytesLeft = I->getValue() % UBytes;
4165 assert(AVT >= MVT::i32 &&
4166 "Do not use rep;movs if not at least DWORD aligned");
4167 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4168 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4173 SDOperand InFlag(0, 0);
4174 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4176 InFlag = Chain.getValue(1);
4177 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4178 Op.getOperand(1), InFlag);
4179 InFlag = Chain.getValue(1);
4180 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4181 Op.getOperand(2), InFlag);
4182 InFlag = Chain.getValue(1);
4184 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4185 SmallVector<SDOperand, 8> Ops;
4186 Ops.push_back(Chain);
4187 Ops.push_back(DAG.getValueType(AVT));
4188 Ops.push_back(InFlag);
4189 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4192 InFlag = Chain.getValue(1);
4193 Count = Op.getOperand(3);
4194 MVT::ValueType CVT = Count.getValueType();
4195 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4196 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4197 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4199 InFlag = Chain.getValue(1);
4200 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4202 Ops.push_back(Chain);
4203 Ops.push_back(DAG.getValueType(MVT::i8));
4204 Ops.push_back(InFlag);
4205 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4206 } else if (BytesLeft) {
4207 // Issue loads and stores for the last 1 - 7 bytes.
4208 unsigned Offset = I->getValue() - BytesLeft;
4209 SDOperand DstAddr = Op.getOperand(1);
4210 MVT::ValueType DstVT = DstAddr.getValueType();
4211 SDOperand SrcAddr = Op.getOperand(2);
4212 MVT::ValueType SrcVT = SrcAddr.getValueType();
4214 if (BytesLeft >= 4) {
4215 Value = DAG.getLoad(MVT::i32, Chain,
4216 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4217 DAG.getConstant(Offset, SrcVT)),
4219 Chain = Value.getValue(1);
4220 Chain = DAG.getStore(Chain, Value,
4221 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4222 DAG.getConstant(Offset, DstVT)),
4227 if (BytesLeft >= 2) {
4228 Value = DAG.getLoad(MVT::i16, Chain,
4229 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4230 DAG.getConstant(Offset, SrcVT)),
4232 Chain = Value.getValue(1);
4233 Chain = DAG.getStore(Chain, Value,
4234 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4235 DAG.getConstant(Offset, DstVT)),
4241 if (BytesLeft == 1) {
4242 Value = DAG.getLoad(MVT::i8, Chain,
4243 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4244 DAG.getConstant(Offset, SrcVT)),
4246 Chain = Value.getValue(1);
4247 Chain = DAG.getStore(Chain, Value,
4248 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4249 DAG.getConstant(Offset, DstVT)),
4258 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4259 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4260 SDOperand TheOp = Op.getOperand(0);
4261 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4262 if (Subtarget->is64Bit()) {
4263 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4264 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4265 MVT::i64, Copy1.getValue(2));
4266 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4267 DAG.getConstant(32, MVT::i8));
4269 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4272 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4273 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4276 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4277 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4278 MVT::i32, Copy1.getValue(2));
4279 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4280 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4281 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4284 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4285 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4287 if (!Subtarget->is64Bit()) {
4288 // vastart just stores the address of the VarArgsFrameIndex slot into the
4289 // memory location argument.
4290 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4291 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4296 // gp_offset (0 - 6 * 8)
4297 // fp_offset (48 - 48 + 8 * 16)
4298 // overflow_arg_area (point to parameters coming in memory).
4300 SmallVector<SDOperand, 8> MemOps;
4301 SDOperand FIN = Op.getOperand(1);
4303 SDOperand Store = DAG.getStore(Op.getOperand(0),
4304 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4305 FIN, SV->getValue(), SV->getOffset());
4306 MemOps.push_back(Store);
4309 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4310 DAG.getConstant(4, getPointerTy()));
4311 Store = DAG.getStore(Op.getOperand(0),
4312 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4313 FIN, SV->getValue(), SV->getOffset());
4314 MemOps.push_back(Store);
4316 // Store ptr to overflow_arg_area
4317 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4318 DAG.getConstant(4, getPointerTy()));
4319 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4320 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4322 MemOps.push_back(Store);
4324 // Store ptr to reg_save_area.
4325 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4326 DAG.getConstant(8, getPointerTy()));
4327 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4328 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4330 MemOps.push_back(Store);
4331 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4335 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4336 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4338 default: return SDOperand(); // Don't custom lower most intrinsics.
4339 // Comparison intrinsics.
4340 case Intrinsic::x86_sse_comieq_ss:
4341 case Intrinsic::x86_sse_comilt_ss:
4342 case Intrinsic::x86_sse_comile_ss:
4343 case Intrinsic::x86_sse_comigt_ss:
4344 case Intrinsic::x86_sse_comige_ss:
4345 case Intrinsic::x86_sse_comineq_ss:
4346 case Intrinsic::x86_sse_ucomieq_ss:
4347 case Intrinsic::x86_sse_ucomilt_ss:
4348 case Intrinsic::x86_sse_ucomile_ss:
4349 case Intrinsic::x86_sse_ucomigt_ss:
4350 case Intrinsic::x86_sse_ucomige_ss:
4351 case Intrinsic::x86_sse_ucomineq_ss:
4352 case Intrinsic::x86_sse2_comieq_sd:
4353 case Intrinsic::x86_sse2_comilt_sd:
4354 case Intrinsic::x86_sse2_comile_sd:
4355 case Intrinsic::x86_sse2_comigt_sd:
4356 case Intrinsic::x86_sse2_comige_sd:
4357 case Intrinsic::x86_sse2_comineq_sd:
4358 case Intrinsic::x86_sse2_ucomieq_sd:
4359 case Intrinsic::x86_sse2_ucomilt_sd:
4360 case Intrinsic::x86_sse2_ucomile_sd:
4361 case Intrinsic::x86_sse2_ucomigt_sd:
4362 case Intrinsic::x86_sse2_ucomige_sd:
4363 case Intrinsic::x86_sse2_ucomineq_sd: {
4365 ISD::CondCode CC = ISD::SETCC_INVALID;
4368 case Intrinsic::x86_sse_comieq_ss:
4369 case Intrinsic::x86_sse2_comieq_sd:
4373 case Intrinsic::x86_sse_comilt_ss:
4374 case Intrinsic::x86_sse2_comilt_sd:
4378 case Intrinsic::x86_sse_comile_ss:
4379 case Intrinsic::x86_sse2_comile_sd:
4383 case Intrinsic::x86_sse_comigt_ss:
4384 case Intrinsic::x86_sse2_comigt_sd:
4388 case Intrinsic::x86_sse_comige_ss:
4389 case Intrinsic::x86_sse2_comige_sd:
4393 case Intrinsic::x86_sse_comineq_ss:
4394 case Intrinsic::x86_sse2_comineq_sd:
4398 case Intrinsic::x86_sse_ucomieq_ss:
4399 case Intrinsic::x86_sse2_ucomieq_sd:
4400 Opc = X86ISD::UCOMI;
4403 case Intrinsic::x86_sse_ucomilt_ss:
4404 case Intrinsic::x86_sse2_ucomilt_sd:
4405 Opc = X86ISD::UCOMI;
4408 case Intrinsic::x86_sse_ucomile_ss:
4409 case Intrinsic::x86_sse2_ucomile_sd:
4410 Opc = X86ISD::UCOMI;
4413 case Intrinsic::x86_sse_ucomigt_ss:
4414 case Intrinsic::x86_sse2_ucomigt_sd:
4415 Opc = X86ISD::UCOMI;
4418 case Intrinsic::x86_sse_ucomige_ss:
4419 case Intrinsic::x86_sse2_ucomige_sd:
4420 Opc = X86ISD::UCOMI;
4423 case Intrinsic::x86_sse_ucomineq_ss:
4424 case Intrinsic::x86_sse2_ucomineq_sd:
4425 Opc = X86ISD::UCOMI;
4431 SDOperand LHS = Op.getOperand(1);
4432 SDOperand RHS = Op.getOperand(2);
4433 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4435 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4436 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4437 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4438 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4439 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4440 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4441 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4446 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4447 // Depths > 0 not supported yet!
4448 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4451 // Just load the return address
4452 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4453 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4456 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4457 // Depths > 0 not supported yet!
4458 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4461 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4462 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4463 DAG.getConstant(4, getPointerTy()));
4466 /// LowerOperation - Provide custom lowering hooks for some operations.
4468 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4469 switch (Op.getOpcode()) {
4470 default: assert(0 && "Should not custom lower this!");
4471 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4472 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4473 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4474 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4475 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4476 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4477 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4478 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4479 case ISD::SHL_PARTS:
4480 case ISD::SRA_PARTS:
4481 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4482 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4483 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4484 case ISD::FABS: return LowerFABS(Op, DAG);
4485 case ISD::FNEG: return LowerFNEG(Op, DAG);
4486 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4487 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4488 case ISD::SELECT: return LowerSELECT(Op, DAG);
4489 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4490 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4491 case ISD::CALL: return LowerCALL(Op, DAG);
4492 case ISD::RET: return LowerRET(Op, DAG);
4493 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4494 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4495 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4496 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4497 case ISD::VASTART: return LowerVASTART(Op, DAG);
4498 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4499 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4500 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4505 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4507 default: return NULL;
4508 case X86ISD::SHLD: return "X86ISD::SHLD";
4509 case X86ISD::SHRD: return "X86ISD::SHRD";
4510 case X86ISD::FAND: return "X86ISD::FAND";
4511 case X86ISD::FOR: return "X86ISD::FOR";
4512 case X86ISD::FXOR: return "X86ISD::FXOR";
4513 case X86ISD::FSRL: return "X86ISD::FSRL";
4514 case X86ISD::FILD: return "X86ISD::FILD";
4515 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4516 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4517 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4518 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4519 case X86ISD::FLD: return "X86ISD::FLD";
4520 case X86ISD::FST: return "X86ISD::FST";
4521 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4522 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4523 case X86ISD::CALL: return "X86ISD::CALL";
4524 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4525 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4526 case X86ISD::CMP: return "X86ISD::CMP";
4527 case X86ISD::COMI: return "X86ISD::COMI";
4528 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4529 case X86ISD::SETCC: return "X86ISD::SETCC";
4530 case X86ISD::CMOV: return "X86ISD::CMOV";
4531 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4532 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4533 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4534 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4535 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4536 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4537 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4538 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4539 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4540 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4541 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4542 case X86ISD::FMAX: return "X86ISD::FMAX";
4543 case X86ISD::FMIN: return "X86ISD::FMIN";
4547 /// isLegalAddressImmediate - Return true if the integer value or
4548 /// GlobalValue can be used as the offset of the target addressing mode.
4549 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
4550 // X86 allows a sign-extended 32-bit immediate field.
4551 return (V > -(1LL << 32) && V < (1LL << 32)-1);
4554 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
4555 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
4556 // field unless we are in small code model.
4557 if (Subtarget->is64Bit() &&
4558 getTargetMachine().getCodeModel() != CodeModel::Small)
4561 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4564 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4565 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4566 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4567 /// are assumed to be legal.
4569 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4570 // Only do shuffles on 128-bit vector types for now.
4571 if (MVT::getSizeInBits(VT) == 64) return false;
4572 return (Mask.Val->getNumOperands() <= 4 ||
4573 isSplatMask(Mask.Val) ||
4574 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4575 X86::isUNPCKLMask(Mask.Val) ||
4576 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4577 X86::isUNPCKHMask(Mask.Val));
4580 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4582 SelectionDAG &DAG) const {
4583 unsigned NumElts = BVOps.size();
4584 // Only do shuffles on 128-bit vector types for now.
4585 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4586 if (NumElts == 2) return true;
4588 return (isMOVLMask(&BVOps[0], 4) ||
4589 isCommutedMOVL(&BVOps[0], 4, true) ||
4590 isSHUFPMask(&BVOps[0], 4) ||
4591 isCommutedSHUFP(&BVOps[0], 4));
4596 //===----------------------------------------------------------------------===//
4597 // X86 Scheduler Hooks
4598 //===----------------------------------------------------------------------===//
4601 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4602 MachineBasicBlock *BB) {
4603 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4604 switch (MI->getOpcode()) {
4605 default: assert(false && "Unexpected instr type to insert");
4606 case X86::CMOV_FR32:
4607 case X86::CMOV_FR64:
4608 case X86::CMOV_V4F32:
4609 case X86::CMOV_V2F64:
4610 case X86::CMOV_V2I64: {
4611 // To "insert" a SELECT_CC instruction, we actually have to insert the
4612 // diamond control-flow pattern. The incoming instruction knows the
4613 // destination vreg to set, the condition code register to branch on, the
4614 // true/false values to select between, and a branch opcode to use.
4615 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4616 ilist<MachineBasicBlock>::iterator It = BB;
4622 // cmpTY ccX, r1, r2
4624 // fallthrough --> copy0MBB
4625 MachineBasicBlock *thisMBB = BB;
4626 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4627 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4629 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4630 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4631 MachineFunction *F = BB->getParent();
4632 F->getBasicBlockList().insert(It, copy0MBB);
4633 F->getBasicBlockList().insert(It, sinkMBB);
4634 // Update machine-CFG edges by first adding all successors of the current
4635 // block to the new block which will contain the Phi node for the select.
4636 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4637 e = BB->succ_end(); i != e; ++i)
4638 sinkMBB->addSuccessor(*i);
4639 // Next, remove all successors of the current block, and add the true
4640 // and fallthrough blocks as its successors.
4641 while(!BB->succ_empty())
4642 BB->removeSuccessor(BB->succ_begin());
4643 BB->addSuccessor(copy0MBB);
4644 BB->addSuccessor(sinkMBB);
4647 // %FalseValue = ...
4648 // # fallthrough to sinkMBB
4651 // Update machine-CFG edges
4652 BB->addSuccessor(sinkMBB);
4655 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4658 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4659 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4660 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4662 delete MI; // The pseudo instruction is gone now.
4666 case X86::FP_TO_INT16_IN_MEM:
4667 case X86::FP_TO_INT32_IN_MEM:
4668 case X86::FP_TO_INT64_IN_MEM: {
4669 // Change the floating point control register to use "round towards zero"
4670 // mode when truncating to an integer value.
4671 MachineFunction *F = BB->getParent();
4672 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4673 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4675 // Load the old value of the high byte of the control word...
4677 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4678 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4680 // Set the high part to be round to zero...
4681 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4684 // Reload the modified control word now...
4685 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4687 // Restore the memory image of control word to original value
4688 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4691 // Get the X86 opcode to use.
4693 switch (MI->getOpcode()) {
4694 default: assert(0 && "illegal opcode!");
4695 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4696 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4697 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4701 MachineOperand &Op = MI->getOperand(0);
4702 if (Op.isRegister()) {
4703 AM.BaseType = X86AddressMode::RegBase;
4704 AM.Base.Reg = Op.getReg();
4706 AM.BaseType = X86AddressMode::FrameIndexBase;
4707 AM.Base.FrameIndex = Op.getFrameIndex();
4709 Op = MI->getOperand(1);
4710 if (Op.isImmediate())
4711 AM.Scale = Op.getImm();
4712 Op = MI->getOperand(2);
4713 if (Op.isImmediate())
4714 AM.IndexReg = Op.getImm();
4715 Op = MI->getOperand(3);
4716 if (Op.isGlobalAddress()) {
4717 AM.GV = Op.getGlobal();
4719 AM.Disp = Op.getImm();
4721 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4722 .addReg(MI->getOperand(4).getReg());
4724 // Reload the original control word now.
4725 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4727 delete MI; // The pseudo instruction is gone now.
4733 //===----------------------------------------------------------------------===//
4734 // X86 Optimization Hooks
4735 //===----------------------------------------------------------------------===//
4737 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4739 uint64_t &KnownZero,
4741 unsigned Depth) const {
4742 unsigned Opc = Op.getOpcode();
4743 assert((Opc >= ISD::BUILTIN_OP_END ||
4744 Opc == ISD::INTRINSIC_WO_CHAIN ||
4745 Opc == ISD::INTRINSIC_W_CHAIN ||
4746 Opc == ISD::INTRINSIC_VOID) &&
4747 "Should use MaskedValueIsZero if you don't know whether Op"
4748 " is a target node!");
4750 KnownZero = KnownOne = 0; // Don't know anything.
4754 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4759 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4760 /// element of the result of the vector shuffle.
4761 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4762 MVT::ValueType VT = N->getValueType(0);
4763 SDOperand PermMask = N->getOperand(2);
4764 unsigned NumElems = PermMask.getNumOperands();
4765 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4767 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4769 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4770 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4771 SDOperand Idx = PermMask.getOperand(i);
4772 if (Idx.getOpcode() == ISD::UNDEF)
4773 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4774 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4779 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4780 /// node is a GlobalAddress + an offset.
4781 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4782 unsigned Opc = N->getOpcode();
4783 if (Opc == X86ISD::Wrapper) {
4784 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4785 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4788 } else if (Opc == ISD::ADD) {
4789 SDOperand N1 = N->getOperand(0);
4790 SDOperand N2 = N->getOperand(1);
4791 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4792 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4794 Offset += V->getSignExtended();
4797 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4798 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4800 Offset += V->getSignExtended();
4808 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4810 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4811 MachineFrameInfo *MFI) {
4812 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4815 SDOperand Loc = N->getOperand(1);
4816 SDOperand BaseLoc = Base->getOperand(1);
4817 if (Loc.getOpcode() == ISD::FrameIndex) {
4818 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4820 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4821 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4822 int FS = MFI->getObjectSize(FI);
4823 int BFS = MFI->getObjectSize(BFI);
4824 if (FS != BFS || FS != Size) return false;
4825 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4827 GlobalValue *GV1 = NULL;
4828 GlobalValue *GV2 = NULL;
4829 int64_t Offset1 = 0;
4830 int64_t Offset2 = 0;
4831 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4832 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4833 if (isGA1 && isGA2 && GV1 == GV2)
4834 return Offset1 == (Offset2 + Dist*Size);
4840 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4841 const X86Subtarget *Subtarget) {
4844 if (isGAPlusOffset(Base, GV, Offset))
4845 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4847 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4848 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4850 // Fixed objects do not specify alignment, however the offsets are known.
4851 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4852 (MFI->getObjectOffset(BFI) % 16) == 0);
4854 return MFI->getObjectAlignment(BFI) >= 16;
4860 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4861 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4862 /// if the load addresses are consecutive, non-overlapping, and in the right
4864 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4865 const X86Subtarget *Subtarget) {
4866 MachineFunction &MF = DAG.getMachineFunction();
4867 MachineFrameInfo *MFI = MF.getFrameInfo();
4868 MVT::ValueType VT = N->getValueType(0);
4869 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4870 SDOperand PermMask = N->getOperand(2);
4871 int NumElems = (int)PermMask.getNumOperands();
4872 SDNode *Base = NULL;
4873 for (int i = 0; i < NumElems; ++i) {
4874 SDOperand Idx = PermMask.getOperand(i);
4875 if (Idx.getOpcode() == ISD::UNDEF) {
4876 if (!Base) return SDOperand();
4879 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4880 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4884 else if (!isConsecutiveLoad(Arg.Val, Base,
4885 i, MVT::getSizeInBits(EVT)/8,MFI))
4890 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4892 LoadSDNode *LD = cast<LoadSDNode>(Base);
4893 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4894 LD->getSrcValueOffset());
4896 // Just use movups, it's shorter.
4897 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4898 SmallVector<SDOperand, 3> Ops;
4899 Ops.push_back(Base->getOperand(0));
4900 Ops.push_back(Base->getOperand(1));
4901 Ops.push_back(Base->getOperand(2));
4902 return DAG.getNode(ISD::BIT_CONVERT, VT,
4903 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4907 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4908 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4909 const X86Subtarget *Subtarget) {
4910 SDOperand Cond = N->getOperand(0);
4912 // If we have SSE[12] support, try to form min/max nodes.
4913 if (Subtarget->hasSSE2() &&
4914 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4915 if (Cond.getOpcode() == ISD::SETCC) {
4916 // Get the LHS/RHS of the select.
4917 SDOperand LHS = N->getOperand(1);
4918 SDOperand RHS = N->getOperand(2);
4919 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4921 unsigned Opcode = 0;
4922 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4925 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4928 if (!UnsafeFPMath) break;
4930 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4932 Opcode = X86ISD::FMIN;
4935 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4938 if (!UnsafeFPMath) break;
4940 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4942 Opcode = X86ISD::FMAX;
4945 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4948 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4951 if (!UnsafeFPMath) break;
4953 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4955 Opcode = X86ISD::FMIN;
4958 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4961 if (!UnsafeFPMath) break;
4963 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4965 Opcode = X86ISD::FMAX;
4971 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4980 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4981 DAGCombinerInfo &DCI) const {
4982 SelectionDAG &DAG = DCI.DAG;
4983 switch (N->getOpcode()) {
4985 case ISD::VECTOR_SHUFFLE:
4986 return PerformShuffleCombine(N, DAG, Subtarget);
4988 return PerformSELECTCombine(N, DAG, Subtarget);
4994 //===----------------------------------------------------------------------===//
4995 // X86 Inline Assembly Support
4996 //===----------------------------------------------------------------------===//
4998 /// getConstraintType - Given a constraint letter, return the type of
4999 /// constraint it is for this target.
5000 X86TargetLowering::ConstraintType
5001 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
5002 switch (ConstraintLetter) {
5011 return C_RegisterClass;
5012 default: return TargetLowering::getConstraintType(ConstraintLetter);
5016 /// isOperandValidForConstraint - Return the specified operand (possibly
5017 /// modified) if the specified SDOperand is valid for the specified target
5018 /// constraint letter, otherwise return null.
5019 SDOperand X86TargetLowering::
5020 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
5021 switch (Constraint) {
5024 // Literal immediates are always ok.
5025 if (isa<ConstantSDNode>(Op)) return Op;
5027 // If we are in non-pic codegen mode, we allow the address of a global to
5028 // be used with 'i'.
5029 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
5030 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
5031 return SDOperand(0, 0);
5033 if (GA->getOpcode() != ISD::TargetGlobalAddress)
5034 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5039 // Otherwise, not valid for this mode.
5040 return SDOperand(0, 0);
5042 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
5046 std::vector<unsigned> X86TargetLowering::
5047 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5048 MVT::ValueType VT) const {
5049 if (Constraint.size() == 1) {
5050 // FIXME: not handling fp-stack yet!
5051 // FIXME: not handling MMX registers yet ('y' constraint).
5052 switch (Constraint[0]) { // GCC X86 Constraint Letters
5053 default: break; // Unknown constraint letter
5054 case 'A': // EAX/EDX
5055 if (VT == MVT::i32 || VT == MVT::i64)
5056 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5058 case 'r': // GENERAL_REGS
5059 case 'R': // LEGACY_REGS
5060 if (VT == MVT::i64 && Subtarget->is64Bit())
5061 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
5062 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
5063 X86::R8, X86::R9, X86::R10, X86::R11,
5064 X86::R12, X86::R13, X86::R14, X86::R15, 0);
5066 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5067 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
5068 else if (VT == MVT::i16)
5069 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5070 X86::SI, X86::DI, X86::BP, X86::SP, 0);
5071 else if (VT == MVT::i8)
5072 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5074 case 'l': // INDEX_REGS
5076 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5077 X86::ESI, X86::EDI, X86::EBP, 0);
5078 else if (VT == MVT::i16)
5079 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5080 X86::SI, X86::DI, X86::BP, 0);
5081 else if (VT == MVT::i8)
5082 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5084 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5087 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5088 else if (VT == MVT::i16)
5089 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5090 else if (VT == MVT::i8)
5091 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5093 case 'x': // SSE_REGS if SSE1 allowed
5094 if (Subtarget->hasSSE1())
5095 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5096 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5098 return std::vector<unsigned>();
5099 case 'Y': // SSE_REGS if SSE2 allowed
5100 if (Subtarget->hasSSE2())
5101 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5102 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5104 return std::vector<unsigned>();
5108 return std::vector<unsigned>();
5111 std::pair<unsigned, const TargetRegisterClass*>
5112 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5113 MVT::ValueType VT) const {
5114 // Use the default implementation in TargetLowering to convert the register
5115 // constraint into a member of a register class.
5116 std::pair<unsigned, const TargetRegisterClass*> Res;
5117 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5119 // Not found as a standard register?
5120 if (Res.second == 0) {
5121 // GCC calls "st(0)" just plain "st".
5122 if (StringsEqualNoCase("{st}", Constraint)) {
5123 Res.first = X86::ST0;
5124 Res.second = X86::RSTRegisterClass;
5130 // Otherwise, check to see if this is a register class of the wrong value
5131 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5132 // turn into {ax},{dx}.
5133 if (Res.second->hasType(VT))
5134 return Res; // Correct type already, nothing to do.
5136 // All of the single-register GCC register classes map their values onto
5137 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5138 // really want an 8-bit or 32-bit register, map to the appropriate register
5139 // class and return the appropriate register.
5140 if (Res.second != X86::GR16RegisterClass)
5143 if (VT == MVT::i8) {
5144 unsigned DestReg = 0;
5145 switch (Res.first) {
5147 case X86::AX: DestReg = X86::AL; break;
5148 case X86::DX: DestReg = X86::DL; break;
5149 case X86::CX: DestReg = X86::CL; break;
5150 case X86::BX: DestReg = X86::BL; break;
5153 Res.first = DestReg;
5154 Res.second = Res.second = X86::GR8RegisterClass;
5156 } else if (VT == MVT::i32) {
5157 unsigned DestReg = 0;
5158 switch (Res.first) {
5160 case X86::AX: DestReg = X86::EAX; break;
5161 case X86::DX: DestReg = X86::EDX; break;
5162 case X86::CX: DestReg = X86::ECX; break;
5163 case X86::BX: DestReg = X86::EBX; break;
5164 case X86::SI: DestReg = X86::ESI; break;
5165 case X86::DI: DestReg = X86::EDI; break;
5166 case X86::BP: DestReg = X86::EBP; break;
5167 case X86::SP: DestReg = X86::ESP; break;
5170 Res.first = DestReg;
5171 Res.second = Res.second = X86::GR32RegisterClass;
5173 } else if (VT == MVT::i64) {
5174 unsigned DestReg = 0;
5175 switch (Res.first) {
5177 case X86::AX: DestReg = X86::RAX; break;
5178 case X86::DX: DestReg = X86::RDX; break;
5179 case X86::CX: DestReg = X86::RCX; break;
5180 case X86::BX: DestReg = X86::RBX; break;
5181 case X86::SI: DestReg = X86::RSI; break;
5182 case X86::DI: DestReg = X86::RDI; break;
5183 case X86::BP: DestReg = X86::RBP; break;
5184 case X86::SP: DestReg = X86::RSP; break;
5187 Res.first = DestReg;
5188 Res.second = Res.second = X86::GR64RegisterClass;