1 //===- AMDGPULegalizerInfo.cpp -----------------------------------*- C++ -*-==//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 /// This file implements the targeting of the Machinelegalizer class for
11 /// \todo This should be generated by TableGen.
12 //===----------------------------------------------------------------------===//
15 #include "AMDGPULegalizerInfo.h"
16 #include "AMDGPUTargetMachine.h"
17 #include "SIMachineFunctionInfo.h"
18 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
19 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
20 #include "llvm/CodeGen/TargetOpcodes.h"
21 #include "llvm/CodeGen/ValueTypes.h"
22 #include "llvm/IR/DerivedTypes.h"
23 #include "llvm/IR/Type.h"
24 #include "llvm/Support/Debug.h"
26 #define DEBUG_TYPE "amdgpu-legalinfo"
29 using namespace LegalizeActions;
30 using namespace LegalizeMutations;
31 using namespace LegalityPredicates;
34 static LegalityPredicate isMultiple32(unsigned TypeIdx,
35 unsigned MaxSize = 512) {
36 return [=](const LegalityQuery &Query) {
37 const LLT Ty = Query.Types[TypeIdx];
38 const LLT EltTy = Ty.getScalarType();
39 return Ty.getSizeInBits() <= MaxSize && EltTy.getSizeInBits() % 32 == 0;
43 static LegalityPredicate isSmallOddVector(unsigned TypeIdx) {
44 return [=](const LegalityQuery &Query) {
45 const LLT Ty = Query.Types[TypeIdx];
46 return Ty.isVector() &&
47 Ty.getNumElements() % 2 != 0 &&
48 Ty.getElementType().getSizeInBits() < 32;
52 static LegalizeMutation oneMoreElement(unsigned TypeIdx) {
53 return [=](const LegalityQuery &Query) {
54 const LLT Ty = Query.Types[TypeIdx];
55 const LLT EltTy = Ty.getElementType();
56 return std::make_pair(TypeIdx, LLT::vector(Ty.getNumElements() + 1, EltTy));
60 static LegalizeMutation fewerEltsToSize64Vector(unsigned TypeIdx) {
61 return [=](const LegalityQuery &Query) {
62 const LLT Ty = Query.Types[TypeIdx];
63 const LLT EltTy = Ty.getElementType();
64 unsigned Size = Ty.getSizeInBits();
65 unsigned Pieces = (Size + 63) / 64;
66 unsigned NewNumElts = (Ty.getNumElements() + 1) / Pieces;
67 return std::make_pair(TypeIdx, LLT::scalarOrVector(NewNumElts, EltTy));
71 static LegalityPredicate vectorWiderThan(unsigned TypeIdx, unsigned Size) {
72 return [=](const LegalityQuery &Query) {
73 const LLT QueryTy = Query.Types[TypeIdx];
74 return QueryTy.isVector() && QueryTy.getSizeInBits() > Size;
78 static LegalityPredicate numElementsNotEven(unsigned TypeIdx) {
79 return [=](const LegalityQuery &Query) {
80 const LLT QueryTy = Query.Types[TypeIdx];
81 return QueryTy.isVector() && QueryTy.getNumElements() % 2 != 0;
85 // Any combination of 32 or 64-bit elements up to 512 bits, and multiples of
87 static LegalityPredicate isRegisterType(unsigned TypeIdx) {
88 return [=](const LegalityQuery &Query) {
89 const LLT Ty = Query.Types[TypeIdx];
91 const int EltSize = Ty.getElementType().getSizeInBits();
92 return EltSize == 32 || EltSize == 64 ||
93 (EltSize == 16 && Ty.getNumElements() % 2 == 0) ||
94 EltSize == 128 || EltSize == 256;
97 return Ty.getSizeInBits() % 32 == 0 && Ty.getSizeInBits() <= 512;
101 AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
102 const GCNTargetMachine &TM)
104 using namespace TargetOpcode;
106 auto GetAddrSpacePtr = [&TM](unsigned AS) {
107 return LLT::pointer(AS, TM.getPointerSizeInBits(AS));
110 const LLT S1 = LLT::scalar(1);
111 const LLT S8 = LLT::scalar(8);
112 const LLT S16 = LLT::scalar(16);
113 const LLT S32 = LLT::scalar(32);
114 const LLT S64 = LLT::scalar(64);
115 const LLT S128 = LLT::scalar(128);
116 const LLT S256 = LLT::scalar(256);
117 const LLT S512 = LLT::scalar(512);
119 const LLT V2S16 = LLT::vector(2, 16);
120 const LLT V4S16 = LLT::vector(4, 16);
122 const LLT V2S32 = LLT::vector(2, 32);
123 const LLT V3S32 = LLT::vector(3, 32);
124 const LLT V4S32 = LLT::vector(4, 32);
125 const LLT V5S32 = LLT::vector(5, 32);
126 const LLT V6S32 = LLT::vector(6, 32);
127 const LLT V7S32 = LLT::vector(7, 32);
128 const LLT V8S32 = LLT::vector(8, 32);
129 const LLT V9S32 = LLT::vector(9, 32);
130 const LLT V10S32 = LLT::vector(10, 32);
131 const LLT V11S32 = LLT::vector(11, 32);
132 const LLT V12S32 = LLT::vector(12, 32);
133 const LLT V13S32 = LLT::vector(13, 32);
134 const LLT V14S32 = LLT::vector(14, 32);
135 const LLT V15S32 = LLT::vector(15, 32);
136 const LLT V16S32 = LLT::vector(16, 32);
138 const LLT V2S64 = LLT::vector(2, 64);
139 const LLT V3S64 = LLT::vector(3, 64);
140 const LLT V4S64 = LLT::vector(4, 64);
141 const LLT V5S64 = LLT::vector(5, 64);
142 const LLT V6S64 = LLT::vector(6, 64);
143 const LLT V7S64 = LLT::vector(7, 64);
144 const LLT V8S64 = LLT::vector(8, 64);
146 std::initializer_list<LLT> AllS32Vectors =
147 {V2S32, V3S32, V4S32, V5S32, V6S32, V7S32, V8S32,
148 V9S32, V10S32, V11S32, V12S32, V13S32, V14S32, V15S32, V16S32};
149 std::initializer_list<LLT> AllS64Vectors =
150 {V2S64, V3S64, V4S64, V5S64, V6S64, V7S64, V8S64};
152 const LLT GlobalPtr = GetAddrSpacePtr(AMDGPUAS::GLOBAL_ADDRESS);
153 const LLT ConstantPtr = GetAddrSpacePtr(AMDGPUAS::CONSTANT_ADDRESS);
154 const LLT LocalPtr = GetAddrSpacePtr(AMDGPUAS::LOCAL_ADDRESS);
155 const LLT FlatPtr = GetAddrSpacePtr(AMDGPUAS::FLAT_ADDRESS);
156 const LLT PrivatePtr = GetAddrSpacePtr(AMDGPUAS::PRIVATE_ADDRESS);
158 const LLT CodePtr = FlatPtr;
160 const std::initializer_list<LLT> AddrSpaces64 = {
161 GlobalPtr, ConstantPtr, FlatPtr
164 const std::initializer_list<LLT> AddrSpaces32 = {
168 const std::initializer_list<LLT> FPTypesBase = {
172 const std::initializer_list<LLT> FPTypes16 = {
176 const std::initializer_list<LLT> FPTypesPK16 = {
180 setAction({G_BRCOND, S1}, Legal);
182 // TODO: All multiples of 32, vectors of pointers, all v2s16 pairs, more
183 // elements for v3s16
184 getActionDefinitionsBuilder(G_PHI)
185 .legalFor({S32, S64, V2S16, V4S16, S1, S128, S256})
186 .legalFor(AllS32Vectors)
187 .legalFor(AllS64Vectors)
188 .legalFor(AddrSpaces64)
189 .legalFor(AddrSpaces32)
190 .clampScalar(0, S32, S256)
191 .widenScalarToNextPow2(0, 32)
192 .clampMaxNumElements(0, S32, 16)
193 .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
194 .legalIf(isPointer(0));
196 if (ST.has16BitInsts()) {
197 getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL})
198 .legalFor({S32, S16})
199 .clampScalar(0, S16, S32)
202 getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL})
204 .clampScalar(0, S32, S32)
208 getActionDefinitionsBuilder({G_UMULH, G_SMULH})
210 .clampScalar(0, S32, S32)
213 // Report legal for any types we can handle anywhere. For the cases only legal
214 // on the SALU, RegBankSelect will be able to re-legalize.
215 getActionDefinitionsBuilder({G_AND, G_OR, G_XOR})
216 .legalFor({S32, S1, S64, V2S32, S16, V2S16, V4S16})
217 .clampScalar(0, S32, S64)
218 .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
219 .fewerElementsIf(vectorWiderThan(0, 32), fewerEltsToSize64Vector(0))
220 .widenScalarToNextPow2(0)
223 getActionDefinitionsBuilder({G_UADDO, G_SADDO, G_USUBO, G_SSUBO,
224 G_UADDE, G_SADDE, G_USUBE, G_SSUBE})
225 .legalFor({{S32, S1}})
226 .clampScalar(0, S32, S32);
228 getActionDefinitionsBuilder(G_BITCAST)
229 .legalForCartesianProduct({S32, V2S16})
230 .legalForCartesianProduct({S64, V2S32, V4S16})
231 .legalForCartesianProduct({V2S64, V4S32})
232 // Don't worry about the size constraint.
233 .legalIf(all(isPointer(0), isPointer(1)));
235 if (ST.has16BitInsts()) {
236 getActionDefinitionsBuilder(G_FCONSTANT)
237 .legalFor({S32, S64, S16})
238 .clampScalar(0, S16, S64);
240 getActionDefinitionsBuilder(G_FCONSTANT)
241 .legalFor({S32, S64})
242 .clampScalar(0, S32, S64);
245 getActionDefinitionsBuilder(G_IMPLICIT_DEF)
246 .legalFor({S1, S32, S64, V2S32, V4S32, V2S16, V4S16, GlobalPtr,
247 ConstantPtr, LocalPtr, FlatPtr, PrivatePtr})
248 .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
249 .clampScalarOrElt(0, S32, S512)
250 .legalIf(isMultiple32(0))
251 .widenScalarToNextPow2(0, 32)
252 .clampMaxNumElements(0, S32, 16);
255 // FIXME: i1 operands to intrinsics should always be legal, but other i1
256 // values may not be legal. We need to figure out how to distinguish
257 // between these two scenarios.
258 getActionDefinitionsBuilder(G_CONSTANT)
259 .legalFor({S1, S32, S64, GlobalPtr,
260 LocalPtr, ConstantPtr, PrivatePtr, FlatPtr })
261 .clampScalar(0, S32, S64)
262 .widenScalarToNextPow2(0)
263 .legalIf(isPointer(0));
265 setAction({G_FRAME_INDEX, PrivatePtr}, Legal);
267 auto &FPOpActions = getActionDefinitionsBuilder(
268 { G_FADD, G_FMUL, G_FNEG, G_FABS, G_FMA, G_FCANONICALIZE})
269 .legalFor({S32, S64});
271 if (ST.has16BitInsts()) {
272 if (ST.hasVOP3PInsts())
273 FPOpActions.legalFor({S16, V2S16});
275 FPOpActions.legalFor({S16});
278 auto &MinNumMaxNum = getActionDefinitionsBuilder({
279 G_FMINNUM, G_FMAXNUM, G_FMINNUM_IEEE, G_FMAXNUM_IEEE});
281 if (ST.hasVOP3PInsts()) {
282 MinNumMaxNum.customFor(FPTypesPK16)
283 .clampMaxNumElements(0, S16, 2)
284 .clampScalar(0, S16, S64)
286 } else if (ST.has16BitInsts()) {
287 MinNumMaxNum.customFor(FPTypes16)
288 .clampScalar(0, S16, S64)
291 MinNumMaxNum.customFor(FPTypesBase)
292 .clampScalar(0, S32, S64)
297 getActionDefinitionsBuilder({G_FMINIMUM, G_FMAXIMUM}).lower();
299 if (ST.hasVOP3PInsts())
300 FPOpActions.clampMaxNumElements(0, S16, 2);
303 .clampScalar(0, ST.has16BitInsts() ? S16 : S32, S64);
305 if (ST.has16BitInsts()) {
306 getActionDefinitionsBuilder(G_FSQRT)
307 .legalFor({S32, S64, S16})
309 .clampScalar(0, S16, S64);
311 getActionDefinitionsBuilder(G_FSQRT)
312 .legalFor({S32, S64})
314 .clampScalar(0, S32, S64);
317 getActionDefinitionsBuilder(G_FPTRUNC)
318 .legalFor({{S32, S64}, {S16, S32}})
321 getActionDefinitionsBuilder(G_FPEXT)
322 .legalFor({{S64, S32}, {S32, S16}})
323 .lowerFor({{S64, S16}}) // FIXME: Implement
326 // TODO: Verify V_BFI_B32 is generated from expanded bit ops.
327 getActionDefinitionsBuilder(G_FCOPYSIGN).lower();
329 getActionDefinitionsBuilder(G_FSUB)
330 // Use actual fsub instruction
332 // Must use fadd + fneg
333 .lowerFor({S64, S16, V2S16})
335 .clampScalar(0, S32, S64);
337 getActionDefinitionsBuilder({G_SEXT, G_ZEXT, G_ANYEXT})
338 .legalFor({{S64, S32}, {S32, S16}, {S64, S16},
339 {S32, S1}, {S64, S1}, {S16, S1},
341 {S64, LLT::scalar(33)},
342 {S32, S8}, {S128, S32}, {S128, S64}, {S32, LLT::scalar(24)}})
345 getActionDefinitionsBuilder({G_SITOFP, G_UITOFP})
346 .legalFor({{S32, S32}, {S64, S32}})
347 .lowerFor({{S32, S64}})
348 .customFor({{S64, S64}})
351 getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
352 .legalFor({{S32, S32}, {S32, S64}})
355 getActionDefinitionsBuilder(G_INTRINSIC_ROUND)
356 .legalFor({S32, S64})
359 if (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
360 getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_FCEIL, G_FRINT})
361 .legalFor({S32, S64})
362 .clampScalar(0, S32, S64)
365 getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_FCEIL, G_FRINT})
368 .clampScalar(0, S32, S64)
372 getActionDefinitionsBuilder(G_GEP)
373 .legalForCartesianProduct(AddrSpaces64, {S64})
374 .legalForCartesianProduct(AddrSpaces32, {S32})
377 setAction({G_BLOCK_ADDR, CodePtr}, Legal);
380 getActionDefinitionsBuilder(G_ICMP)
381 .legalForCartesianProduct(
382 {S1}, {S32, S64, GlobalPtr, LocalPtr, ConstantPtr, PrivatePtr, FlatPtr})
383 .legalFor({{S1, S32}, {S1, S64}});
384 if (ST.has16BitInsts()) {
385 CmpBuilder.legalFor({{S1, S16}});
389 .widenScalarToNextPow2(1)
390 .clampScalar(1, S32, S64)
392 .legalIf(all(typeIs(0, S1), isPointer(1)));
394 getActionDefinitionsBuilder(G_FCMP)
395 .legalForCartesianProduct({S1}, ST.has16BitInsts() ? FPTypes16 : FPTypesBase)
396 .widenScalarToNextPow2(1)
397 .clampScalar(1, S32, S64)
400 // FIXME: fexp, flog2, flog10 needs to be custom lowered.
401 getActionDefinitionsBuilder({G_FPOW, G_FEXP, G_FEXP2,
402 G_FLOG, G_FLOG2, G_FLOG10})
406 // The 64-bit versions produce 32-bit results, but only on the SALU.
407 getActionDefinitionsBuilder({G_CTLZ, G_CTLZ_ZERO_UNDEF,
408 G_CTTZ, G_CTTZ_ZERO_UNDEF,
410 .legalFor({{S32, S32}, {S32, S64}})
411 .clampScalar(0, S32, S32)
412 .clampScalar(1, S32, S64)
414 .widenScalarToNextPow2(0, 32)
415 .widenScalarToNextPow2(1, 32);
417 // TODO: Expand for > s32
418 getActionDefinitionsBuilder(G_BSWAP)
420 .clampScalar(0, S32, S32)
423 if (ST.has16BitInsts()) {
424 if (ST.hasVOP3PInsts()) {
425 getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
426 .legalFor({S32, S16, V2S16})
427 .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
428 .clampMaxNumElements(0, S16, 2)
429 .clampScalar(0, S16, S32)
430 .widenScalarToNextPow2(0)
433 getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
434 .legalFor({S32, S16})
435 .widenScalarToNextPow2(0)
436 .clampScalar(0, S16, S32)
440 getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
442 .clampScalar(0, S32, S32)
443 .widenScalarToNextPow2(0)
447 auto smallerThan = [](unsigned TypeIdx0, unsigned TypeIdx1) {
448 return [=](const LegalityQuery &Query) {
449 return Query.Types[TypeIdx0].getSizeInBits() <
450 Query.Types[TypeIdx1].getSizeInBits();
454 auto greaterThan = [](unsigned TypeIdx0, unsigned TypeIdx1) {
455 return [=](const LegalityQuery &Query) {
456 return Query.Types[TypeIdx0].getSizeInBits() >
457 Query.Types[TypeIdx1].getSizeInBits();
461 getActionDefinitionsBuilder(G_INTTOPTR)
462 // List the common cases
463 .legalForCartesianProduct(AddrSpaces64, {S64})
464 .legalForCartesianProduct(AddrSpaces32, {S32})
466 // Accept any address space as long as the size matches
467 .legalIf(sameSize(0, 1))
468 .widenScalarIf(smallerThan(1, 0),
469 [](const LegalityQuery &Query) {
470 return std::make_pair(1, LLT::scalar(Query.Types[0].getSizeInBits()));
472 .narrowScalarIf(greaterThan(1, 0),
473 [](const LegalityQuery &Query) {
474 return std::make_pair(1, LLT::scalar(Query.Types[0].getSizeInBits()));
477 getActionDefinitionsBuilder(G_PTRTOINT)
478 // List the common cases
479 .legalForCartesianProduct(AddrSpaces64, {S64})
480 .legalForCartesianProduct(AddrSpaces32, {S32})
482 // Accept any address space as long as the size matches
483 .legalIf(sameSize(0, 1))
484 .widenScalarIf(smallerThan(0, 1),
485 [](const LegalityQuery &Query) {
486 return std::make_pair(0, LLT::scalar(Query.Types[1].getSizeInBits()));
490 [](const LegalityQuery &Query) {
491 return std::make_pair(0, LLT::scalar(Query.Types[1].getSizeInBits()));
494 if (ST.hasFlatAddressSpace()) {
495 getActionDefinitionsBuilder(G_ADDRSPACE_CAST)
500 // TODO: Should load to s16 be legal? Most loads extend to 32-bits, but we
501 // handle some operations by just promoting the register during
502 // selection. There are also d16 loads on GFX9+ which preserve the high bits.
503 getActionDefinitionsBuilder({G_LOAD, G_STORE})
504 .narrowScalarIf([](const LegalityQuery &Query) {
505 unsigned Size = Query.Types[0].getSizeInBits();
506 unsigned MemSize = Query.MMODescrs[0].SizeInBits;
507 return (Size > 32 && MemSize < Size);
509 [](const LegalityQuery &Query) {
510 return std::make_pair(0, LLT::scalar(32));
512 .fewerElementsIf([=](const LegalityQuery &Query) {
513 unsigned MemSize = Query.MMODescrs[0].SizeInBits;
514 return (MemSize == 96) &&
515 Query.Types[0].isVector() &&
516 !ST.hasDwordx3LoadStores();
518 [=](const LegalityQuery &Query) {
519 return std::make_pair(0, V2S32);
521 .legalIf([=](const LegalityQuery &Query) {
522 const LLT &Ty0 = Query.Types[0];
524 unsigned Size = Ty0.getSizeInBits();
525 unsigned MemSize = Query.MMODescrs[0].SizeInBits;
526 if (Size < 32 || (Size > 32 && MemSize < Size))
529 if (Ty0.isVector() && Size != MemSize)
532 // TODO: Decompose private loads into 4-byte components.
533 // TODO: Illegal flat loads on SI
544 return ST.hasDwordx3LoadStores();
548 // TODO: Possibly support loads of i256 and i512 . This will require
549 // adding i256 and i512 types to MVT in order for to be able to use
551 // TODO: Add support for other vector types, this will require
552 // defining more value mappings for the new types.
553 return Ty0.isVector() && (Ty0.getScalarType().getSizeInBits() == 32 ||
554 Ty0.getScalarType().getSizeInBits() == 64);
560 .clampScalar(0, S32, S64);
563 // FIXME: Handle alignment requirements.
564 auto &ExtLoads = getActionDefinitionsBuilder({G_SEXTLOAD, G_ZEXTLOAD})
565 .legalForTypesWithMemDesc({
566 {S32, GlobalPtr, 8, 8},
567 {S32, GlobalPtr, 16, 8},
568 {S32, LocalPtr, 8, 8},
569 {S32, LocalPtr, 16, 8},
570 {S32, PrivatePtr, 8, 8},
571 {S32, PrivatePtr, 16, 8}});
572 if (ST.hasFlatAddressSpace()) {
573 ExtLoads.legalForTypesWithMemDesc({{S32, FlatPtr, 8, 8},
574 {S32, FlatPtr, 16, 8}});
577 ExtLoads.clampScalar(0, S32, S32)
578 .widenScalarToNextPow2(0)
579 .unsupportedIfMemSizeNotPow2()
582 auto &Atomics = getActionDefinitionsBuilder(
583 {G_ATOMICRMW_XCHG, G_ATOMICRMW_ADD, G_ATOMICRMW_SUB,
584 G_ATOMICRMW_AND, G_ATOMICRMW_OR, G_ATOMICRMW_XOR,
585 G_ATOMICRMW_MAX, G_ATOMICRMW_MIN, G_ATOMICRMW_UMAX,
586 G_ATOMICRMW_UMIN, G_ATOMIC_CMPXCHG})
587 .legalFor({{S32, GlobalPtr}, {S32, LocalPtr},
588 {S64, GlobalPtr}, {S64, LocalPtr}});
589 if (ST.hasFlatAddressSpace()) {
590 Atomics.legalFor({{S32, FlatPtr}, {S64, FlatPtr}});
593 // TODO: Pointer types, any 32-bit or 64-bit vector
594 getActionDefinitionsBuilder(G_SELECT)
595 .legalForCartesianProduct({S32, S64, S16, V2S32, V2S16, V4S16,
596 GlobalPtr, LocalPtr, FlatPtr, PrivatePtr,
597 LLT::vector(2, LocalPtr), LLT::vector(2, PrivatePtr)}, {S1})
598 .clampScalar(0, S16, S64)
599 .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
600 .fewerElementsIf(numElementsNotEven(0), scalarize(0))
602 .clampMaxNumElements(0, S32, 2)
603 .clampMaxNumElements(0, LocalPtr, 2)
604 .clampMaxNumElements(0, PrivatePtr, 2)
606 .widenScalarToNextPow2(0)
607 .legalIf(all(isPointer(0), typeIs(1, S1)));
609 // TODO: Only the low 4/5/6 bits of the shift amount are observed, so we can
610 // be more flexible with the shift amount type.
611 auto &Shifts = getActionDefinitionsBuilder({G_SHL, G_LSHR, G_ASHR})
612 .legalFor({{S32, S32}, {S64, S32}});
613 if (ST.has16BitInsts()) {
614 if (ST.hasVOP3PInsts()) {
615 Shifts.legalFor({{S16, S32}, {S16, S16}, {V2S16, V2S16}})
616 .clampMaxNumElements(0, S16, 2);
618 Shifts.legalFor({{S16, S32}, {S16, S16}});
620 Shifts.clampScalar(1, S16, S32);
621 Shifts.clampScalar(0, S16, S64);
622 Shifts.widenScalarToNextPow2(0, 16);
624 // Make sure we legalize the shift amount type first, as the general
625 // expansion for the shifted type will produce much worse code if it hasn't
626 // been truncated already.
627 Shifts.clampScalar(1, S32, S32);
628 Shifts.clampScalar(0, S32, S64);
629 Shifts.widenScalarToNextPow2(0, 32);
633 for (unsigned Op : {G_EXTRACT_VECTOR_ELT, G_INSERT_VECTOR_ELT}) {
634 unsigned VecTypeIdx = Op == G_EXTRACT_VECTOR_ELT ? 1 : 0;
635 unsigned EltTypeIdx = Op == G_EXTRACT_VECTOR_ELT ? 0 : 1;
636 unsigned IdxTypeIdx = 2;
638 getActionDefinitionsBuilder(Op)
639 .customIf([=](const LegalityQuery &Query) {
640 const LLT EltTy = Query.Types[EltTypeIdx];
641 const LLT VecTy = Query.Types[VecTypeIdx];
642 const LLT IdxTy = Query.Types[IdxTypeIdx];
643 return (EltTy.getSizeInBits() == 16 ||
644 EltTy.getSizeInBits() % 32 == 0) &&
645 VecTy.getSizeInBits() % 32 == 0 &&
646 VecTy.getSizeInBits() <= 512 &&
647 IdxTy.getSizeInBits() == 32;
649 .clampScalar(EltTypeIdx, S32, S64)
650 .clampScalar(VecTypeIdx, S32, S64)
651 .clampScalar(IdxTypeIdx, S32, S32);
654 getActionDefinitionsBuilder(G_EXTRACT_VECTOR_ELT)
655 .unsupportedIf([=](const LegalityQuery &Query) {
656 const LLT &EltTy = Query.Types[1].getElementType();
657 return Query.Types[0] != EltTy;
660 for (unsigned Op : {G_EXTRACT, G_INSERT}) {
661 unsigned BigTyIdx = Op == G_EXTRACT ? 1 : 0;
662 unsigned LitTyIdx = Op == G_EXTRACT ? 0 : 1;
664 // FIXME: Doesn't handle extract of illegal sizes.
665 getActionDefinitionsBuilder(Op)
666 .legalIf([=](const LegalityQuery &Query) {
667 const LLT BigTy = Query.Types[BigTyIdx];
668 const LLT LitTy = Query.Types[LitTyIdx];
669 return (BigTy.getSizeInBits() % 32 == 0) &&
670 (LitTy.getSizeInBits() % 16 == 0);
673 [=](const LegalityQuery &Query) {
674 const LLT BigTy = Query.Types[BigTyIdx];
675 return (BigTy.getScalarSizeInBits() < 16);
677 LegalizeMutations::widenScalarOrEltToNextPow2(BigTyIdx, 16))
679 [=](const LegalityQuery &Query) {
680 const LLT LitTy = Query.Types[LitTyIdx];
681 return (LitTy.getScalarSizeInBits() < 16);
683 LegalizeMutations::widenScalarOrEltToNextPow2(LitTyIdx, 16))
684 .moreElementsIf(isSmallOddVector(BigTyIdx), oneMoreElement(BigTyIdx))
685 .widenScalarToNextPow2(BigTyIdx, 32);
689 getActionDefinitionsBuilder(G_BUILD_VECTOR)
690 .legalForCartesianProduct(AllS32Vectors, {S32})
691 .legalForCartesianProduct(AllS64Vectors, {S64})
692 .clampNumElements(0, V16S32, V16S32)
693 .clampNumElements(0, V2S64, V8S64)
694 .minScalarSameAs(1, 0)
695 .legalIf(isRegisterType(0))
696 .minScalarOrElt(0, S32);
698 getActionDefinitionsBuilder(G_CONCAT_VECTORS)
699 .legalIf(isRegisterType(0));
702 for (unsigned Op : {G_MERGE_VALUES, G_UNMERGE_VALUES}) {
703 unsigned BigTyIdx = Op == G_MERGE_VALUES ? 0 : 1;
704 unsigned LitTyIdx = Op == G_MERGE_VALUES ? 1 : 0;
706 auto notValidElt = [=](const LegalityQuery &Query, unsigned TypeIdx) {
707 const LLT &Ty = Query.Types[TypeIdx];
709 const LLT &EltTy = Ty.getElementType();
710 if (EltTy.getSizeInBits() < 8 || EltTy.getSizeInBits() > 64)
712 if (!isPowerOf2_32(EltTy.getSizeInBits()))
718 getActionDefinitionsBuilder(Op)
719 .widenScalarToNextPow2(LitTyIdx, /*Min*/ 16)
720 // Clamp the little scalar to s8-s256 and make it a power of 2. It's not
721 // worth considering the multiples of 64 since 2*192 and 2*384 are not
723 .clampScalar(LitTyIdx, S16, S256)
724 .widenScalarToNextPow2(LitTyIdx, /*Min*/ 32)
726 // Break up vectors with weird elements into scalars
728 [=](const LegalityQuery &Query) { return notValidElt(Query, 0); },
731 [=](const LegalityQuery &Query) { return notValidElt(Query, 1); },
733 .clampScalar(BigTyIdx, S32, S512)
735 [=](const LegalityQuery &Query) {
736 const LLT &Ty = Query.Types[BigTyIdx];
737 return !isPowerOf2_32(Ty.getSizeInBits()) &&
738 Ty.getSizeInBits() % 16 != 0;
740 [=](const LegalityQuery &Query) {
741 // Pick the next power of 2, or a multiple of 64 over 128.
742 // Whichever is smaller.
743 const LLT &Ty = Query.Types[BigTyIdx];
744 unsigned NewSizeInBits = 1 << Log2_32_Ceil(Ty.getSizeInBits() + 1);
745 if (NewSizeInBits >= 256) {
746 unsigned RoundedTo = alignTo<64>(Ty.getSizeInBits() + 1);
747 if (RoundedTo < NewSizeInBits)
748 NewSizeInBits = RoundedTo;
750 return std::make_pair(BigTyIdx, LLT::scalar(NewSizeInBits));
752 .legalIf([=](const LegalityQuery &Query) {
753 const LLT &BigTy = Query.Types[BigTyIdx];
754 const LLT &LitTy = Query.Types[LitTyIdx];
756 if (BigTy.isVector() && BigTy.getSizeInBits() < 32)
758 if (LitTy.isVector() && LitTy.getSizeInBits() < 32)
761 return BigTy.getSizeInBits() % 16 == 0 &&
762 LitTy.getSizeInBits() % 16 == 0 &&
763 BigTy.getSizeInBits() <= 512;
765 // Any vectors left are the wrong size. Scalarize them.
771 verify(*ST.getInstrInfo());
774 bool AMDGPULegalizerInfo::legalizeCustom(MachineInstr &MI,
775 MachineRegisterInfo &MRI,
776 MachineIRBuilder &MIRBuilder,
777 GISelChangeObserver &Observer) const {
778 switch (MI.getOpcode()) {
779 case TargetOpcode::G_ADDRSPACE_CAST:
780 return legalizeAddrSpaceCast(MI, MRI, MIRBuilder);
781 case TargetOpcode::G_FRINT:
782 return legalizeFrint(MI, MRI, MIRBuilder);
783 case TargetOpcode::G_FCEIL:
784 return legalizeFceil(MI, MRI, MIRBuilder);
785 case TargetOpcode::G_INTRINSIC_TRUNC:
786 return legalizeIntrinsicTrunc(MI, MRI, MIRBuilder);
787 case TargetOpcode::G_SITOFP:
788 return legalizeITOFP(MI, MRI, MIRBuilder, true);
789 case TargetOpcode::G_UITOFP:
790 return legalizeITOFP(MI, MRI, MIRBuilder, false);
791 case TargetOpcode::G_FMINNUM:
792 case TargetOpcode::G_FMAXNUM:
793 case TargetOpcode::G_FMINNUM_IEEE:
794 case TargetOpcode::G_FMAXNUM_IEEE:
795 return legalizeMinNumMaxNum(MI, MRI, MIRBuilder);
796 case TargetOpcode::G_EXTRACT_VECTOR_ELT:
797 return legalizeExtractVectorElt(MI, MRI, MIRBuilder);
798 case TargetOpcode::G_INSERT_VECTOR_ELT:
799 return legalizeInsertVectorElt(MI, MRI, MIRBuilder);
804 llvm_unreachable("expected switch to return");
807 Register AMDGPULegalizerInfo::getSegmentAperture(
809 MachineRegisterInfo &MRI,
810 MachineIRBuilder &MIRBuilder) const {
811 MachineFunction &MF = MIRBuilder.getMF();
812 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
813 const LLT S32 = LLT::scalar(32);
815 if (ST.hasApertureRegs()) {
816 // FIXME: Use inline constants (src_{shared, private}_base) instead of
818 unsigned Offset = AS == AMDGPUAS::LOCAL_ADDRESS ?
819 AMDGPU::Hwreg::OFFSET_SRC_SHARED_BASE :
820 AMDGPU::Hwreg::OFFSET_SRC_PRIVATE_BASE;
821 unsigned WidthM1 = AS == AMDGPUAS::LOCAL_ADDRESS ?
822 AMDGPU::Hwreg::WIDTH_M1_SRC_SHARED_BASE :
823 AMDGPU::Hwreg::WIDTH_M1_SRC_PRIVATE_BASE;
825 AMDGPU::Hwreg::ID_MEM_BASES << AMDGPU::Hwreg::ID_SHIFT_ |
826 Offset << AMDGPU::Hwreg::OFFSET_SHIFT_ |
827 WidthM1 << AMDGPU::Hwreg::WIDTH_M1_SHIFT_;
829 Register ApertureReg = MRI.createGenericVirtualRegister(S32);
830 Register GetReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
832 MIRBuilder.buildInstr(AMDGPU::S_GETREG_B32)
835 MRI.setType(GetReg, S32);
837 auto ShiftAmt = MIRBuilder.buildConstant(S32, WidthM1 + 1);
838 MIRBuilder.buildInstr(TargetOpcode::G_SHL)
841 .addUse(ShiftAmt.getReg(0));
846 Register QueuePtr = MRI.createGenericVirtualRegister(
847 LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64));
849 // FIXME: Placeholder until we can track the input registers.
850 MIRBuilder.buildConstant(QueuePtr, 0xdeadbeef);
852 // Offset into amd_queue_t for group_segment_aperture_base_hi /
853 // private_segment_aperture_base_hi.
854 uint32_t StructOffset = (AS == AMDGPUAS::LOCAL_ADDRESS) ? 0x40 : 0x44;
856 // FIXME: Don't use undef
857 Value *V = UndefValue::get(PointerType::get(
858 Type::getInt8Ty(MF.getFunction().getContext()),
859 AMDGPUAS::CONSTANT_ADDRESS));
861 MachinePointerInfo PtrInfo(V, StructOffset);
862 MachineMemOperand *MMO = MF.getMachineMemOperand(
864 MachineMemOperand::MOLoad |
865 MachineMemOperand::MODereferenceable |
866 MachineMemOperand::MOInvariant,
868 MinAlign(64, StructOffset));
870 Register LoadResult = MRI.createGenericVirtualRegister(S32);
873 MIRBuilder.materializeGEP(LoadAddr, QueuePtr, LLT::scalar(64), StructOffset);
874 MIRBuilder.buildLoad(LoadResult, LoadAddr, *MMO);
878 bool AMDGPULegalizerInfo::legalizeAddrSpaceCast(
879 MachineInstr &MI, MachineRegisterInfo &MRI,
880 MachineIRBuilder &MIRBuilder) const {
881 MachineFunction &MF = MIRBuilder.getMF();
883 MIRBuilder.setInstr(MI);
885 Register Dst = MI.getOperand(0).getReg();
886 Register Src = MI.getOperand(1).getReg();
888 LLT DstTy = MRI.getType(Dst);
889 LLT SrcTy = MRI.getType(Src);
890 unsigned DestAS = DstTy.getAddressSpace();
891 unsigned SrcAS = SrcTy.getAddressSpace();
893 // TODO: Avoid reloading from the queue ptr for each cast, or at least each
895 assert(!DstTy.isVector());
897 const AMDGPUTargetMachine &TM
898 = static_cast<const AMDGPUTargetMachine &>(MF.getTarget());
900 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
901 if (ST.getTargetLowering()->isNoopAddrSpaceCast(SrcAS, DestAS)) {
902 MI.setDesc(MIRBuilder.getTII().get(TargetOpcode::G_BITCAST));
906 if (SrcAS == AMDGPUAS::FLAT_ADDRESS) {
907 assert(DestAS == AMDGPUAS::LOCAL_ADDRESS ||
908 DestAS == AMDGPUAS::PRIVATE_ADDRESS);
909 unsigned NullVal = TM.getNullPointerValue(DestAS);
911 auto SegmentNull = MIRBuilder.buildConstant(DstTy, NullVal);
912 auto FlatNull = MIRBuilder.buildConstant(SrcTy, 0);
914 Register PtrLo32 = MRI.createGenericVirtualRegister(DstTy);
916 // Extract low 32-bits of the pointer.
917 MIRBuilder.buildExtract(PtrLo32, Src, 0);
919 Register CmpRes = MRI.createGenericVirtualRegister(LLT::scalar(1));
920 MIRBuilder.buildICmp(CmpInst::ICMP_NE, CmpRes, Src, FlatNull.getReg(0));
921 MIRBuilder.buildSelect(Dst, CmpRes, PtrLo32, SegmentNull.getReg(0));
923 MI.eraseFromParent();
927 assert(SrcAS == AMDGPUAS::LOCAL_ADDRESS ||
928 SrcAS == AMDGPUAS::PRIVATE_ADDRESS);
931 MIRBuilder.buildConstant(SrcTy, TM.getNullPointerValue(SrcAS));
933 MIRBuilder.buildConstant(DstTy, TM.getNullPointerValue(DestAS));
935 Register ApertureReg = getSegmentAperture(DestAS, MRI, MIRBuilder);
937 Register CmpRes = MRI.createGenericVirtualRegister(LLT::scalar(1));
938 MIRBuilder.buildICmp(CmpInst::ICMP_NE, CmpRes, Src, SegmentNull.getReg(0));
940 Register BuildPtr = MRI.createGenericVirtualRegister(DstTy);
942 // Coerce the type of the low half of the result so we can use merge_values.
943 Register SrcAsInt = MRI.createGenericVirtualRegister(LLT::scalar(32));
944 MIRBuilder.buildInstr(TargetOpcode::G_PTRTOINT)
948 // TODO: Should we allow mismatched types but matching sizes in merges to
949 // avoid the ptrtoint?
950 MIRBuilder.buildMerge(BuildPtr, {SrcAsInt, ApertureReg});
951 MIRBuilder.buildSelect(Dst, CmpRes, BuildPtr, FlatNull.getReg(0));
953 MI.eraseFromParent();
957 bool AMDGPULegalizerInfo::legalizeFrint(
958 MachineInstr &MI, MachineRegisterInfo &MRI,
959 MachineIRBuilder &MIRBuilder) const {
960 MIRBuilder.setInstr(MI);
962 Register Src = MI.getOperand(1).getReg();
963 LLT Ty = MRI.getType(Src);
964 assert(Ty.isScalar() && Ty.getSizeInBits() == 64);
966 APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52");
967 APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51");
969 auto C1 = MIRBuilder.buildFConstant(Ty, C1Val);
970 auto CopySign = MIRBuilder.buildFCopysign(Ty, C1, Src);
972 // TODO: Should this propagate fast-math-flags?
973 auto Tmp1 = MIRBuilder.buildFAdd(Ty, Src, CopySign);
974 auto Tmp2 = MIRBuilder.buildFSub(Ty, Tmp1, CopySign);
976 auto C2 = MIRBuilder.buildFConstant(Ty, C2Val);
977 auto Fabs = MIRBuilder.buildFAbs(Ty, Src);
979 auto Cond = MIRBuilder.buildFCmp(CmpInst::FCMP_OGT, LLT::scalar(1), Fabs, C2);
980 MIRBuilder.buildSelect(MI.getOperand(0).getReg(), Cond, Src, Tmp2);
984 bool AMDGPULegalizerInfo::legalizeFceil(
985 MachineInstr &MI, MachineRegisterInfo &MRI,
986 MachineIRBuilder &B) const {
989 const LLT S1 = LLT::scalar(1);
990 const LLT S64 = LLT::scalar(64);
992 Register Src = MI.getOperand(1).getReg();
993 assert(MRI.getType(Src) == S64);
995 // result = trunc(src)
996 // if (src > 0.0 && src != result)
999 auto Trunc = B.buildInstr(TargetOpcode::G_INTRINSIC_TRUNC, {S64}, {Src});
1001 const auto Zero = B.buildFConstant(S64, 0.0);
1002 const auto One = B.buildFConstant(S64, 1.0);
1003 auto Lt0 = B.buildFCmp(CmpInst::FCMP_OGT, S1, Src, Zero);
1004 auto NeTrunc = B.buildFCmp(CmpInst::FCMP_ONE, S1, Src, Trunc);
1005 auto And = B.buildAnd(S1, Lt0, NeTrunc);
1006 auto Add = B.buildSelect(S64, And, One, Zero);
1008 // TODO: Should this propagate fast-math-flags?
1009 B.buildFAdd(MI.getOperand(0).getReg(), Trunc, Add);
1013 static MachineInstrBuilder extractF64Exponent(unsigned Hi,
1014 MachineIRBuilder &B) {
1015 const unsigned FractBits = 52;
1016 const unsigned ExpBits = 11;
1017 LLT S32 = LLT::scalar(32);
1019 auto Const0 = B.buildConstant(S32, FractBits - 32);
1020 auto Const1 = B.buildConstant(S32, ExpBits);
1022 auto ExpPart = B.buildIntrinsic(Intrinsic::amdgcn_ubfe, {S32}, false)
1023 .addUse(Const0.getReg(0))
1024 .addUse(Const1.getReg(0));
1026 return B.buildSub(S32, ExpPart, B.buildConstant(S32, 1023));
1029 bool AMDGPULegalizerInfo::legalizeIntrinsicTrunc(
1030 MachineInstr &MI, MachineRegisterInfo &MRI,
1031 MachineIRBuilder &B) const {
1034 const LLT S1 = LLT::scalar(1);
1035 const LLT S32 = LLT::scalar(32);
1036 const LLT S64 = LLT::scalar(64);
1038 Register Src = MI.getOperand(1).getReg();
1039 assert(MRI.getType(Src) == S64);
1041 // TODO: Should this use extract since the low half is unused?
1042 auto Unmerge = B.buildUnmerge({S32, S32}, Src);
1043 Register Hi = Unmerge.getReg(1);
1045 // Extract the upper half, since this is where we will find the sign and
1047 auto Exp = extractF64Exponent(Hi, B);
1049 const unsigned FractBits = 52;
1051 // Extract the sign bit.
1052 const auto SignBitMask = B.buildConstant(S32, UINT32_C(1) << 31);
1053 auto SignBit = B.buildAnd(S32, Hi, SignBitMask);
1055 const auto FractMask = B.buildConstant(S64, (UINT64_C(1) << FractBits) - 1);
1057 const auto Zero32 = B.buildConstant(S32, 0);
1059 // Extend back to 64-bits.
1060 auto SignBit64 = B.buildMerge(S64, {Zero32.getReg(0), SignBit.getReg(0)});
1062 auto Shr = B.buildAShr(S64, FractMask, Exp);
1063 auto Not = B.buildNot(S64, Shr);
1064 auto Tmp0 = B.buildAnd(S64, Src, Not);
1065 auto FiftyOne = B.buildConstant(S32, FractBits - 1);
1067 auto ExpLt0 = B.buildICmp(CmpInst::ICMP_SLT, S1, Exp, Zero32);
1068 auto ExpGt51 = B.buildICmp(CmpInst::ICMP_SGT, S1, Exp, FiftyOne);
1070 auto Tmp1 = B.buildSelect(S64, ExpLt0, SignBit64, Tmp0);
1071 B.buildSelect(MI.getOperand(0).getReg(), ExpGt51, Src, Tmp1);
1075 bool AMDGPULegalizerInfo::legalizeITOFP(
1076 MachineInstr &MI, MachineRegisterInfo &MRI,
1077 MachineIRBuilder &B, bool Signed) const {
1080 Register Dst = MI.getOperand(0).getReg();
1081 Register Src = MI.getOperand(1).getReg();
1083 const LLT S64 = LLT::scalar(64);
1084 const LLT S32 = LLT::scalar(32);
1086 assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S64);
1088 auto Unmerge = B.buildUnmerge({S32, S32}, Src);
1090 auto CvtHi = Signed ?
1091 B.buildSITOFP(S64, Unmerge.getReg(1)) :
1092 B.buildUITOFP(S64, Unmerge.getReg(1));
1094 auto CvtLo = B.buildUITOFP(S64, Unmerge.getReg(0));
1096 auto ThirtyTwo = B.buildConstant(S32, 32);
1097 auto LdExp = B.buildIntrinsic(Intrinsic::amdgcn_ldexp, {S64}, false)
1098 .addUse(CvtHi.getReg(0))
1099 .addUse(ThirtyTwo.getReg(0));
1101 // TODO: Should this propagate fast-math-flags?
1102 B.buildFAdd(Dst, LdExp, CvtLo);
1103 MI.eraseFromParent();
1107 bool AMDGPULegalizerInfo::legalizeMinNumMaxNum(
1108 MachineInstr &MI, MachineRegisterInfo &MRI,
1109 MachineIRBuilder &B) const {
1110 MachineFunction &MF = B.getMF();
1111 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1113 const bool IsIEEEOp = MI.getOpcode() == AMDGPU::G_FMINNUM_IEEE ||
1114 MI.getOpcode() == AMDGPU::G_FMAXNUM_IEEE;
1116 // With ieee_mode disabled, the instructions have the correct behavior
1117 // already for G_FMINNUM/G_FMAXNUM
1118 if (!MFI->getMode().IEEE)
1124 MachineIRBuilder HelperBuilder(MI);
1125 GISelObserverWrapper DummyObserver;
1126 LegalizerHelper Helper(MF, DummyObserver, HelperBuilder);
1127 HelperBuilder.setMBB(*MI.getParent());
1128 return Helper.lowerFMinNumMaxNum(MI) == LegalizerHelper::Legalized;
1131 bool AMDGPULegalizerInfo::legalizeExtractVectorElt(
1132 MachineInstr &MI, MachineRegisterInfo &MRI,
1133 MachineIRBuilder &B) const {
1134 // TODO: Should move some of this into LegalizerHelper.
1136 // TODO: Promote dynamic indexing of s16 to s32
1137 // TODO: Dynamic s64 indexing is only legal for SGPR.
1138 Optional<int64_t> IdxVal = getConstantVRegVal(MI.getOperand(2).getReg(), MRI);
1139 if (!IdxVal) // Dynamic case will be selected to register indexing.
1142 Register Dst = MI.getOperand(0).getReg();
1143 Register Vec = MI.getOperand(1).getReg();
1145 LLT VecTy = MRI.getType(Vec);
1146 LLT EltTy = VecTy.getElementType();
1147 assert(EltTy == MRI.getType(Dst));
1151 if (IdxVal.getValue() < VecTy.getNumElements())
1152 B.buildExtract(Dst, Vec, IdxVal.getValue() * EltTy.getSizeInBits());
1156 MI.eraseFromParent();
1160 bool AMDGPULegalizerInfo::legalizeInsertVectorElt(
1161 MachineInstr &MI, MachineRegisterInfo &MRI,
1162 MachineIRBuilder &B) const {
1163 // TODO: Should move some of this into LegalizerHelper.
1165 // TODO: Promote dynamic indexing of s16 to s32
1166 // TODO: Dynamic s64 indexing is only legal for SGPR.
1167 Optional<int64_t> IdxVal = getConstantVRegVal(MI.getOperand(3).getReg(), MRI);
1168 if (!IdxVal) // Dynamic case will be selected to register indexing.
1171 Register Dst = MI.getOperand(0).getReg();
1172 Register Vec = MI.getOperand(1).getReg();
1173 Register Ins = MI.getOperand(2).getReg();
1175 LLT VecTy = MRI.getType(Vec);
1176 LLT EltTy = VecTy.getElementType();
1177 assert(EltTy == MRI.getType(Ins));
1181 if (IdxVal.getValue() < VecTy.getNumElements())
1182 B.buildInsert(Dst, Vec, Ins, IdxVal.getValue() * EltTy.getSizeInBits());
1186 MI.eraseFromParent();
1190 // Return the use branch instruction, otherwise null if the usage is invalid.
1191 static MachineInstr *verifyCFIntrinsic(MachineInstr &MI,
1192 MachineRegisterInfo &MRI) {
1193 Register CondDef = MI.getOperand(0).getReg();
1194 if (!MRI.hasOneNonDBGUse(CondDef))
1197 MachineInstr &UseMI = *MRI.use_instr_nodbg_begin(CondDef);
1198 return UseMI.getParent() == MI.getParent() &&
1199 UseMI.getOpcode() == AMDGPU::G_BRCOND ? &UseMI : nullptr;
1202 Register AMDGPULegalizerInfo::getLiveInRegister(MachineRegisterInfo &MRI,
1203 Register Reg, LLT Ty) const {
1204 Register LiveIn = MRI.getLiveInVirtReg(Reg);
1208 Register NewReg = MRI.createGenericVirtualRegister(Ty);
1209 MRI.addLiveIn(Reg, NewReg);
1213 bool AMDGPULegalizerInfo::loadInputValue(Register DstReg, MachineIRBuilder &B,
1214 const ArgDescriptor *Arg) const {
1215 if (!Arg->isRegister())
1216 return false; // TODO: Handle these
1218 assert(Arg->getRegister() != 0);
1219 assert(Arg->getRegister().isPhysical());
1221 MachineRegisterInfo &MRI = *B.getMRI();
1223 LLT Ty = MRI.getType(DstReg);
1224 Register LiveIn = getLiveInRegister(MRI, Arg->getRegister(), Ty);
1226 if (Arg->isMasked()) {
1227 // TODO: Should we try to emit this once in the entry block?
1228 const LLT S32 = LLT::scalar(32);
1229 const unsigned Mask = Arg->getMask();
1230 const unsigned Shift = countTrailingZeros<unsigned>(Mask);
1232 auto ShiftAmt = B.buildConstant(S32, Shift);
1233 auto LShr = B.buildLShr(S32, LiveIn, ShiftAmt);
1234 B.buildAnd(DstReg, LShr, B.buildConstant(S32, Mask >> Shift));
1236 B.buildCopy(DstReg, LiveIn);
1238 // Insert the argument copy if it doens't already exist.
1239 // FIXME: It seems EmitLiveInCopies isn't called anywhere?
1240 if (!MRI.getVRegDef(LiveIn)) {
1241 MachineBasicBlock &EntryMBB = B.getMF().front();
1242 EntryMBB.addLiveIn(Arg->getRegister());
1243 B.setInsertPt(EntryMBB, EntryMBB.begin());
1244 B.buildCopy(LiveIn, Arg->getRegister());
1250 bool AMDGPULegalizerInfo::legalizePreloadedArgIntrin(
1252 MachineRegisterInfo &MRI,
1253 MachineIRBuilder &B,
1254 AMDGPUFunctionArgInfo::PreloadedValue ArgType) const {
1257 const SIMachineFunctionInfo *MFI = B.getMF().getInfo<SIMachineFunctionInfo>();
1259 const ArgDescriptor *Arg;
1260 const TargetRegisterClass *RC;
1261 std::tie(Arg, RC) = MFI->getPreloadedValue(ArgType);
1263 LLVM_DEBUG(dbgs() << "Required arg register missing\n");
1267 if (loadInputValue(MI.getOperand(0).getReg(), B, Arg)) {
1268 MI.eraseFromParent();
1275 bool AMDGPULegalizerInfo::legalizeImplicitArgPtr(MachineInstr &MI,
1276 MachineRegisterInfo &MRI,
1277 MachineIRBuilder &B) const {
1278 const SIMachineFunctionInfo *MFI = B.getMF().getInfo<SIMachineFunctionInfo>();
1279 if (!MFI->isEntryFunction()) {
1280 return legalizePreloadedArgIntrin(MI, MRI, B,
1281 AMDGPUFunctionArgInfo::IMPLICIT_ARG_PTR);
1287 ST.getTargetLowering()->getImplicitParameterOffset(
1288 B.getMF(), AMDGPUTargetLowering::FIRST_IMPLICIT);
1289 Register DstReg = MI.getOperand(0).getReg();
1290 LLT DstTy = MRI.getType(DstReg);
1291 LLT IdxTy = LLT::scalar(DstTy.getSizeInBits());
1293 const ArgDescriptor *Arg;
1294 const TargetRegisterClass *RC;
1296 = MFI->getPreloadedValue(AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR);
1300 Register KernargPtrReg = MRI.createGenericVirtualRegister(DstTy);
1301 if (!loadInputValue(KernargPtrReg, B, Arg))
1304 B.buildGEP(DstReg, KernargPtrReg, B.buildConstant(IdxTy, Offset).getReg(0));
1305 MI.eraseFromParent();
1309 bool AMDGPULegalizerInfo::legalizeIntrinsic(MachineInstr &MI,
1310 MachineRegisterInfo &MRI,
1311 MachineIRBuilder &B) const {
1312 // Replace the use G_BRCOND with the exec manipulate and branch pseudos.
1313 switch (MI.getOperand(MI.getNumExplicitDefs()).getIntrinsicID()) {
1314 case Intrinsic::amdgcn_if: {
1315 if (MachineInstr *BrCond = verifyCFIntrinsic(MI, MRI)) {
1316 const SIRegisterInfo *TRI
1317 = static_cast<const SIRegisterInfo *>(MRI.getTargetRegisterInfo());
1319 B.setInstr(*BrCond);
1320 Register Def = MI.getOperand(1).getReg();
1321 Register Use = MI.getOperand(3).getReg();
1322 B.buildInstr(AMDGPU::SI_IF)
1325 .addMBB(BrCond->getOperand(1).getMBB());
1327 MRI.setRegClass(Def, TRI->getWaveMaskRegClass());
1328 MRI.setRegClass(Use, TRI->getWaveMaskRegClass());
1329 MI.eraseFromParent();
1330 BrCond->eraseFromParent();
1336 case Intrinsic::amdgcn_loop: {
1337 if (MachineInstr *BrCond = verifyCFIntrinsic(MI, MRI)) {
1338 const SIRegisterInfo *TRI
1339 = static_cast<const SIRegisterInfo *>(MRI.getTargetRegisterInfo());
1341 B.setInstr(*BrCond);
1342 Register Reg = MI.getOperand(2).getReg();
1343 B.buildInstr(AMDGPU::SI_LOOP)
1345 .addMBB(BrCond->getOperand(1).getMBB());
1346 MI.eraseFromParent();
1347 BrCond->eraseFromParent();
1348 MRI.setRegClass(Reg, TRI->getWaveMaskRegClass());
1354 case Intrinsic::amdgcn_kernarg_segment_ptr:
1355 return legalizePreloadedArgIntrin(
1356 MI, MRI, B, AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR);
1357 case Intrinsic::amdgcn_implicitarg_ptr:
1358 return legalizeImplicitArgPtr(MI, MRI, B);
1359 case Intrinsic::amdgcn_workitem_id_x:
1360 return legalizePreloadedArgIntrin(MI, MRI, B,
1361 AMDGPUFunctionArgInfo::WORKITEM_ID_X);
1362 case Intrinsic::amdgcn_workitem_id_y:
1363 return legalizePreloadedArgIntrin(MI, MRI, B,
1364 AMDGPUFunctionArgInfo::WORKITEM_ID_Y);
1365 case Intrinsic::amdgcn_workitem_id_z:
1366 return legalizePreloadedArgIntrin(MI, MRI, B,
1367 AMDGPUFunctionArgInfo::WORKITEM_ID_Z);
1368 case Intrinsic::amdgcn_workgroup_id_x:
1369 return legalizePreloadedArgIntrin(MI, MRI, B,
1370 AMDGPUFunctionArgInfo::WORKGROUP_ID_X);
1371 case Intrinsic::amdgcn_workgroup_id_y:
1372 return legalizePreloadedArgIntrin(MI, MRI, B,
1373 AMDGPUFunctionArgInfo::WORKGROUP_ID_Y);
1374 case Intrinsic::amdgcn_workgroup_id_z:
1375 return legalizePreloadedArgIntrin(MI, MRI, B,
1376 AMDGPUFunctionArgInfo::WORKGROUP_ID_Z);
1377 case Intrinsic::amdgcn_dispatch_ptr:
1378 return legalizePreloadedArgIntrin(MI, MRI, B,
1379 AMDGPUFunctionArgInfo::DISPATCH_PTR);
1380 case Intrinsic::amdgcn_queue_ptr:
1381 return legalizePreloadedArgIntrin(MI, MRI, B,
1382 AMDGPUFunctionArgInfo::QUEUE_PTR);
1383 case Intrinsic::amdgcn_implicit_buffer_ptr:
1384 return legalizePreloadedArgIntrin(
1385 MI, MRI, B, AMDGPUFunctionArgInfo::IMPLICIT_BUFFER_PTR);
1386 case Intrinsic::amdgcn_dispatch_id:
1387 return legalizePreloadedArgIntrin(MI, MRI, B,
1388 AMDGPUFunctionArgInfo::DISPATCH_ID);