Skip to content

Commit ba0a52a

Browse files
AlexMacleanAlexMaclean
authored
[InferAS] Support getAssumedAddrSpace for Arguments for NVPTX (#133991)
1 parent d33ae41 commit ba0a52a

File tree

10 files changed

+785
-524
lines changed

10 files changed

+785
-524
lines changed

‎llvm/lib/Target/NVPTX/NVPTXLowerArgs.cpp

Lines changed: 4 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -678,11 +678,8 @@ static bool runOnKernelFunction(const NVPTXTargetMachine &TM, Function &F) {
678678

679679
LLVM_DEBUG(dbgs() << "Lowering kernel args of " << F.getName() << "\n");
680680
for (Argument &Arg : F.args()) {
681-
if (Arg.getType()->isPointerTy()) {
682-
if (Arg.hasByValAttr())
683-
handleByValParam(TM, &Arg);
684-
else if (TM.getDrvInterface() == NVPTX::CUDA)
685-
markPointerAsGlobal(&Arg);
681+
if (Arg.getType()->isPointerTy() && Arg.hasByValAttr()) {
682+
handleByValParam(TM, &Arg);
686683
} else if (Arg.getType()->isIntegerTy() &&
687684
TM.getDrvInterface() == NVPTX::CUDA) {
688685
HandleIntToPtr(Arg);
@@ -699,10 +696,9 @@ static bool runOnDeviceFunction(const NVPTXTargetMachine &TM, Function &F) {
699696
cast<NVPTXTargetLowering>(TM.getSubtargetImpl()->getTargetLowering());
700697

701698
for (Argument &Arg : F.args())
702-
if (Arg.getType()->isPointerTy() && Arg.hasByValAttr()) {
703-
markPointerAsAS(&Arg, ADDRESS_SPACE_LOCAL);
699+
if (Arg.getType()->isPointerTy() && Arg.hasByValAttr())
704700
adjustByValArgAlignment(&Arg, &Arg, TLI);
705-
}
701+
706702
return true;
707703
}
708704

‎llvm/lib/Target/NVPTX/NVPTXTargetTransformInfo.cpp

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -599,6 +599,21 @@ unsigned NVPTXTTIImpl::getAssumedAddrSpace(const Value *V) const {
599599
if (isa<AllocaInst>(V))
600600
return ADDRESS_SPACE_LOCAL;
601601

602+
if (const Argument *Arg = dyn_cast<Argument>(V)) {
603+
if (isKernelFunction(*Arg->getParent())) {
604+
const NVPTXTargetMachine &TM =
605+
static_cast<const NVPTXTargetMachine &>(getTLI()->getTargetMachine());
606+
if (TM.getDrvInterface() == NVPTX::CUDA && !Arg->hasByValAttr())
607+
return ADDRESS_SPACE_GLOBAL;
608+
} else {
609+
// We assume that all device parameters that are passed byval will be
610+
// placed in the local AS. Very simple cases will be updated after ISel to
611+
// use the device param space where possible.
612+
if (Arg->hasByValAttr())
613+
return ADDRESS_SPACE_LOCAL;
614+
}
615+
}
616+
602617
return -1;
603618
}
604619

‎llvm/lib/Transforms/Scalar/InferAddressSpaces.cpp

Lines changed: 36 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -305,10 +305,16 @@ static bool isNoopPtrIntCastPair(const Operator *I2P, const DataLayout &DL,
305305
}
306306

307307
// Returns true if V is an address expression.
308-
// TODO: Currently, we consider only phi, bitcast, addrspacecast, and
309-
// getelementptr operators.
308+
// TODO: Currently, we only consider:
309+
// - arguments
310+
// - phi, bitcast, addrspacecast, and getelementptr operators
310311
static bool isAddressExpression(const Value &V, const DataLayout &DL,
311312
const TargetTransformInfo *TTI) {
313+
314+
if (const Argument *Arg = dyn_cast<Argument>(&V))
315+
return Arg->getType()->isPointerTy() &&
316+
TTI->getAssumedAddrSpace(&V) != UninitializedAddressSpace;
317+
312318
const Operator *Op = dyn_cast<Operator>(&V);
313319
if (!Op)
314320
return false;
@@ -341,6 +347,9 @@ static bool isAddressExpression(const Value &V, const DataLayout &DL,
341347
static SmallVector<Value *, 2>
342348
getPointerOperands(const Value &V, const DataLayout &DL,
343349
const TargetTransformInfo *TTI) {
350+
if (isa<Argument>(&V))
351+
return {};
352+
344353
const Operator &Op = cast<Operator>(V);
345354
switch (Op.getOpcode()) {
346355
case Instruction::PHI: {
@@ -505,13 +514,11 @@ void InferAddressSpacesImpl::appendsFlatAddressExpressionToPostorderStack(
505514
if (Visited.insert(V).second) {
506515
PostorderStack.emplace_back(V, false);
507516

508-
Operator *Op = cast<Operator>(V);
509-
for (unsigned I = 0, E = Op->getNumOperands(); I != E; ++I) {
510-
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op->getOperand(I))) {
511-
if (isAddressExpression(*CE, *DL, TTI) && Visited.insert(CE).second)
512-
PostorderStack.emplace_back(CE, false);
513-
}
514-
}
517+
if (auto *Op = dyn_cast<Operator>(V))
518+
for (auto &O : Op->operands())
519+
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(O))
520+
if (isAddressExpression(*CE, *DL, TTI) && Visited.insert(CE).second)
521+
PostorderStack.emplace_back(CE, false);
515522
}
516523
}
517524
}
@@ -828,6 +835,18 @@ Value *InferAddressSpacesImpl::cloneValueWithNewAddressSpace(
828835
assert(V->getType()->getPointerAddressSpace() == FlatAddrSpace &&
829836
isAddressExpression(*V, *DL, TTI));
830837

838+
if (auto *Arg = dyn_cast<Argument>(V)) {
839+
// Arguments are address space casted in the function body, as we do not
840+
// want to change the function signature.
841+
Function *F = Arg->getParent();
842+
BasicBlock::iterator Insert = F->getEntryBlock().getFirstNonPHIIt();
843+
844+
Type *NewPtrTy = PointerType::get(Arg->getContext(), NewAddrSpace);
845+
auto *NewI = new AddrSpaceCastInst(Arg, NewPtrTy);
846+
NewI->insertBefore(Insert);
847+
return NewI;
848+
}
849+
831850
if (Instruction *I = dyn_cast<Instruction>(V)) {
832851
Value *NewV = cloneInstructionWithNewAddressSpace(
833852
I, NewAddrSpace, ValueWithNewAddrSpace, PredicatedAS, PoisonUsesToFix);
@@ -966,8 +985,12 @@ bool InferAddressSpacesImpl::updateAddressSpace(
966985
// of all its pointer operands.
967986
unsigned NewAS = UninitializedAddressSpace;
968987

969-
const Operator &Op = cast<Operator>(V);
970-
if (Op.getOpcode() == Instruction::Select) {
988+
// isAddressExpression should guarantee that V is an operator or an argument.
989+
assert(isa<Operator>(V) || isa<Argument>(V));
990+
991+
if (isa<Operator>(V) &&
992+
cast<Operator>(V).getOpcode() == Instruction::Select) {
993+
const Operator &Op = cast<Operator>(V);
971994
Value *Src0 = Op.getOperand(1);
972995
Value *Src1 = Op.getOperand(2);
973996

@@ -1258,7 +1281,7 @@ void InferAddressSpacesImpl::performPointerReplacement(
12581281
}
12591282

12601283
// Otherwise, replaces the use with flat(NewV).
1261-
if (Instruction *VInst = dyn_cast<Instruction>(V)) {
1284+
if (isa<Instruction>(V) || isa<Instruction>(NewV)) {
12621285
// Don't create a copy of the original addrspacecast.
12631286
if (U == V && isa<AddrSpaceCastInst>(V))
12641287
return;
@@ -1268,7 +1291,7 @@ void InferAddressSpacesImpl::performPointerReplacement(
12681291
if (Instruction *NewVInst = dyn_cast<Instruction>(NewV))
12691292
InsertPos = std::next(NewVInst->getIterator());
12701293
else
1271-
InsertPos = std::next(VInst->getIterator());
1294+
InsertPos = std::next(cast<Instruction>(V)->getIterator());
12721295

12731296
while (isa<PHINode>(InsertPos))
12741297
++InsertPos;

‎llvm/test/CodeGen/NVPTX/i1-ext-load.ll

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -12,14 +12,14 @@ define ptx_kernel void @foo(ptr noalias readonly %ptr, ptr noalias %retval) {
1212
; CHECK: .reg .b64 %rd<5>;
1313
; CHECK-EMPTY:
1414
; CHECK: ld.param.u64 %rd1, [foo_param_0];
15-
; CHECK: ld.param.u64 %rd2, [foo_param_1];
16-
; CHECK: cvta.to.global.u64 %rd3, %rd2;
17-
; CHECK: cvta.to.global.u64 %rd4, %rd1;
18-
; CHECK: ld.global.nc.u8 %rs1, [%rd4];
15+
; CHECK: cvta.to.global.u64 %rd2, %rd1;
16+
; CHECK: ld.param.u64 %rd3, [foo_param_1];
17+
; CHECK: cvta.to.global.u64 %rd4, %rd3;
18+
; CHECK: ld.global.nc.u8 %rs1, [%rd2];
1919
; CHECK: cvt.u32.u8 %r1, %rs1;
2020
; CHECK: add.s32 %r2, %r1, 1;
2121
; CHECK: and.b32 %r3, %r2, 1;
22-
; CHECK: st.global.u32 [%rd3], %r3;
22+
; CHECK: st.global.u32 [%rd4], %r3;
2323
; CHECK: ret;
2424
%ld = load i1, ptr %ptr, align 1
2525
%zext = zext i1 %ld to i32

‎llvm/test/CodeGen/NVPTX/lower-args-gridconstant.ll

Lines changed: 13 additions & 33 deletions
Original file line numberDiff line numberDiff line change
@@ -12,9 +12,7 @@ define dso_local noundef i32 @non_kernel_function(ptr nocapture noundef readonly
1212
; OPT-LABEL: define dso_local noundef i32 @non_kernel_function(
1313
; OPT-SAME: ptr noundef readonly byval([[STRUCT_UINT4:%.*]]) align 16 captures(none) [[A:%.*]], i1 noundef zeroext [[B:%.*]], i32 noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
1414
; OPT-NEXT: [[ENTRY:.*:]]
15-
; OPT-NEXT: [[A2:%.*]] = addrspacecast ptr [[A]] to ptr addrspace(5)
16-
; OPT-NEXT: [[A1:%.*]] = addrspacecast ptr addrspace(5) [[A2]] to ptr
17-
; OPT-NEXT: [[A_:%.*]] = select i1 [[B]], ptr [[A1]], ptr addrspacecast (ptr addrspace(1) @gi to ptr)
15+
; OPT-NEXT: [[A_:%.*]] = select i1 [[B]], ptr [[A]], ptr addrspacecast (ptr addrspace(1) @gi to ptr)
1816
; OPT-NEXT: [[IDX_EXT:%.*]] = sext i32 [[C]] to i64
1917
; OPT-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds i8, ptr [[A_]], i64 [[IDX_EXT]]
2018
; OPT-NEXT: [[TMP0:%.*]] = load i32, ptr [[ADD_PTR]], align 1
@@ -74,12 +72,10 @@ define ptx_kernel void @grid_const_int(ptr byval(i32) align 4 %input1, i32 %inpu
7472
; PTX-NEXT: ret;
7573
; OPT-LABEL: define ptx_kernel void @grid_const_int(
7674
; OPT-SAME: ptr byval(i32) align 4 [[INPUT1:%.*]], i32 [[INPUT2:%.*]], ptr [[OUT:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
77-
; OPT-NEXT: [[OUT2:%.*]] = addrspacecast ptr [[OUT]] to ptr addrspace(1)
78-
; OPT-NEXT: [[OUT3:%.*]] = addrspacecast ptr addrspace(1) [[OUT2]] to ptr
7975
; OPT-NEXT: [[INPUT11:%.*]] = addrspacecast ptr [[INPUT1]] to ptr addrspace(101)
8076
; OPT-NEXT: [[TMP:%.*]] = load i32, ptr addrspace(101) [[INPUT11]], align 4
8177
; OPT-NEXT: [[ADD:%.*]] = add i32 [[TMP]], [[INPUT2]]
82-
; OPT-NEXT: store i32 [[ADD]], ptr [[OUT3]], align 4
78+
; OPT-NEXT: store i32 [[ADD]], ptr [[OUT]], align 4
8379
; OPT-NEXT: ret void
8480
%tmp = load i32, ptr %input1, align 4
8581
%add = add i32 %tmp, %input2
@@ -105,15 +101,13 @@ define ptx_kernel void @grid_const_struct(ptr byval(%struct.s) align 4 %input, p
105101
; PTX-NEXT: ret;
106102
; OPT-LABEL: define ptx_kernel void @grid_const_struct(
107103
; OPT-SAME: ptr byval([[STRUCT_S:%.*]]) align 4 [[INPUT:%.*]], ptr [[OUT:%.*]]) #[[ATTR0]] {
108-
; OPT-NEXT: [[OUT4:%.*]] = addrspacecast ptr [[OUT]] to ptr addrspace(1)
109-
; OPT-NEXT: [[OUT5:%.*]] = addrspacecast ptr addrspace(1) [[OUT4]] to ptr
110104
; OPT-NEXT: [[INPUT1:%.*]] = addrspacecast ptr [[INPUT]] to ptr addrspace(101)
111105
; OPT-NEXT: [[GEP13:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr addrspace(101) [[INPUT1]], i32 0, i32 0
112106
; OPT-NEXT: [[GEP22:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr addrspace(101) [[INPUT1]], i32 0, i32 1
113107
; OPT-NEXT: [[TMP1:%.*]] = load i32, ptr addrspace(101) [[GEP13]], align 4
114108
; OPT-NEXT: [[TMP2:%.*]] = load i32, ptr addrspace(101) [[GEP22]], align 4
115109
; OPT-NEXT: [[ADD:%.*]] = add i32 [[TMP1]], [[TMP2]]
116-
; OPT-NEXT: store i32 [[ADD]], ptr [[OUT5]], align 4
110+
; OPT-NEXT: store i32 [[ADD]], ptr [[OUT]], align 4
117111
; OPT-NEXT: ret void
118112
%gep1 = getelementptr inbounds %struct.s, ptr %input, i32 0, i32 0
119113
%gep2 = getelementptr inbounds %struct.s, ptr %input, i32 0, i32 1
@@ -233,11 +227,9 @@ define ptx_kernel void @grid_const_memory_escape(ptr byval(%struct.s) align 4 %i
233227
; PTX-NEXT: ret;
234228
; OPT-LABEL: define ptx_kernel void @grid_const_memory_escape(
235229
; OPT-SAME: ptr byval([[STRUCT_S:%.*]]) align 4 [[INPUT:%.*]], ptr [[ADDR:%.*]]) #[[ATTR0]] {
236-
; OPT-NEXT: [[ADDR4:%.*]] = addrspacecast ptr [[ADDR]] to ptr addrspace(1)
237-
; OPT-NEXT: [[ADDR5:%.*]] = addrspacecast ptr addrspace(1) [[ADDR4]] to ptr
238230
; OPT-NEXT: [[INPUT_PARAM:%.*]] = addrspacecast ptr [[INPUT]] to ptr addrspace(101)
239231
; OPT-NEXT: [[INPUT1:%.*]] = call ptr @llvm.nvvm.ptr.param.to.gen.p0.p101(ptr addrspace(101) [[INPUT_PARAM]])
240-
; OPT-NEXT: store ptr [[INPUT1]], ptr [[ADDR5]], align 8
232+
; OPT-NEXT: store ptr [[INPUT1]], ptr [[ADDR]], align 8
241233
; OPT-NEXT: ret void
242234
store ptr %input, ptr %addr, align 8
243235
ret void
@@ -263,14 +255,12 @@ define ptx_kernel void @grid_const_inlineasm_escape(ptr byval(%struct.s) align 4
263255
; PTX-NOT .local
264256
; OPT-LABEL: define ptx_kernel void @grid_const_inlineasm_escape(
265257
; OPT-SAME: ptr byval([[STRUCT_S:%.*]]) align 4 [[INPUT:%.*]], ptr [[RESULT:%.*]]) #[[ATTR0]] {
266-
; OPT-NEXT: [[RESULT4:%.*]] = addrspacecast ptr [[RESULT]] to ptr addrspace(1)
267-
; OPT-NEXT: [[RESULT5:%.*]] = addrspacecast ptr addrspace(1) [[RESULT4]] to ptr
268258
; OPT-NEXT: [[INPUT_PARAM:%.*]] = addrspacecast ptr [[INPUT]] to ptr addrspace(101)
269259
; OPT-NEXT: [[INPUT1:%.*]] = call ptr @llvm.nvvm.ptr.param.to.gen.p0.p101(ptr addrspace(101) [[INPUT_PARAM]])
270260
; OPT-NEXT: [[TMPPTR1:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[INPUT1]], i32 0, i32 0
271261
; OPT-NEXT: [[TMPPTR2:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[INPUT1]], i32 0, i32 1
272262
; OPT-NEXT: [[TMP2:%.*]] = call i64 asm "add.s64 $0, $1, $2
273-
; OPT-NEXT: store i64 [[TMP2]], ptr [[RESULT5]], align 8
263+
; OPT-NEXT: store i64 [[TMP2]], ptr [[RESULT]], align 8
274264
; OPT-NEXT: ret void
275265
%tmpptr1 = getelementptr inbounds %struct.s, ptr %input, i32 0, i32 0
276266
%tmpptr2 = getelementptr inbounds %struct.s, ptr %input, i32 0, i32 1
@@ -311,13 +301,11 @@ define ptx_kernel void @grid_const_partial_escape(ptr byval(i32) %input, ptr %ou
311301
; PTX-NEXT: ret;
312302
; OPT-LABEL: define ptx_kernel void @grid_const_partial_escape(
313303
; OPT-SAME: ptr byval(i32) [[INPUT:%.*]], ptr [[OUTPUT:%.*]]) #[[ATTR0]] {
314-
; OPT-NEXT: [[OUTPUT4:%.*]] = addrspacecast ptr [[OUTPUT]] to ptr addrspace(1)
315-
; OPT-NEXT: [[OUTPUT5:%.*]] = addrspacecast ptr addrspace(1) [[OUTPUT4]] to ptr
316304
; OPT-NEXT: [[INPUT1:%.*]] = addrspacecast ptr [[INPUT]] to ptr addrspace(101)
317305
; OPT-NEXT: [[INPUT1_GEN:%.*]] = call ptr @llvm.nvvm.ptr.param.to.gen.p0.p101(ptr addrspace(101) [[INPUT1]])
318306
; OPT-NEXT: [[VAL1:%.*]] = load i32, ptr [[INPUT1_GEN]], align 4
319307
; OPT-NEXT: [[TWICE:%.*]] = add i32 [[VAL1]], [[VAL1]]
320-
; OPT-NEXT: store i32 [[TWICE]], ptr [[OUTPUT5]], align 4
308+
; OPT-NEXT: store i32 [[TWICE]], ptr [[OUTPUT]], align 4
321309
; OPT-NEXT: [[CALL:%.*]] = call i32 @escape(ptr [[INPUT1_GEN]])
322310
; OPT-NEXT: ret void
323311
%val = load i32, ptr %input
@@ -361,15 +349,13 @@ define ptx_kernel i32 @grid_const_partial_escapemem(ptr byval(%struct.s) %input,
361349
; PTX-NEXT: ret;
362350
; OPT-LABEL: define ptx_kernel i32 @grid_const_partial_escapemem(
363351
; OPT-SAME: ptr byval([[STRUCT_S:%.*]]) [[INPUT:%.*]], ptr [[OUTPUT:%.*]]) #[[ATTR0]] {
364-
; OPT-NEXT: [[OUTPUT4:%.*]] = addrspacecast ptr [[OUTPUT]] to ptr addrspace(1)
365-
; OPT-NEXT: [[OUTPUT5:%.*]] = addrspacecast ptr addrspace(1) [[OUTPUT4]] to ptr
366352
; OPT-NEXT: [[INPUT2:%.*]] = addrspacecast ptr [[INPUT]] to ptr addrspace(101)
367353
; OPT-NEXT: [[INPUT1:%.*]] = call ptr @llvm.nvvm.ptr.param.to.gen.p0.p101(ptr addrspace(101) [[INPUT2]])
368354
; OPT-NEXT: [[PTR1:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[INPUT1]], i32 0, i32 0
369355
; OPT-NEXT: [[VAL1:%.*]] = load i32, ptr [[PTR1]], align 4
370356
; OPT-NEXT: [[PTR2:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[INPUT1]], i32 0, i32 1
371357
; OPT-NEXT: [[VAL2:%.*]] = load i32, ptr [[PTR2]], align 4
372-
; OPT-NEXT: store ptr [[INPUT1]], ptr [[OUTPUT5]], align 8
358+
; OPT-NEXT: store ptr [[INPUT1]], ptr [[OUTPUT]], align 8
373359
; OPT-NEXT: [[ADD:%.*]] = add i32 [[VAL1]], [[VAL2]]
374360
; OPT-NEXT: [[CALL2:%.*]] = call i32 @escape(ptr [[PTR1]])
375361
; OPT-NEXT: ret i32 [[ADD]]
@@ -407,11 +393,9 @@ define ptx_kernel void @grid_const_phi(ptr byval(%struct.s) align 4 %input1, ptr
407393
; PTX-NEXT: ret;
408394
; OPT-LABEL: define ptx_kernel void @grid_const_phi(
409395
; OPT-SAME: ptr byval([[STRUCT_S:%.*]]) align 4 [[INPUT1:%.*]], ptr [[INOUT:%.*]]) #[[ATTR0]] {
410-
; OPT-NEXT: [[INOUT1:%.*]] = addrspacecast ptr [[INOUT]] to ptr addrspace(1)
411-
; OPT-NEXT: [[INOUT2:%.*]] = addrspacecast ptr addrspace(1) [[INOUT1]] to ptr
412396
; OPT-NEXT: [[INPUT1_PARAM:%.*]] = addrspacecast ptr [[INPUT1]] to ptr addrspace(101)
413397
; OPT-NEXT: [[INPUT1_PARAM_GEN:%.*]] = call ptr @llvm.nvvm.ptr.param.to.gen.p0.p101(ptr addrspace(101) [[INPUT1_PARAM]])
414-
; OPT-NEXT: [[VAL:%.*]] = load i32, ptr [[INOUT2]], align 4
398+
; OPT-NEXT: [[VAL:%.*]] = load i32, ptr [[INOUT]], align 4
415399
; OPT-NEXT: [[LESS:%.*]] = icmp slt i32 [[VAL]], 0
416400
; OPT-NEXT: br i1 [[LESS]], label %[[FIRST:.*]], label %[[SECOND:.*]]
417401
; OPT: [[FIRST]]:
@@ -423,7 +407,7 @@ define ptx_kernel void @grid_const_phi(ptr byval(%struct.s) align 4 %input1, ptr
423407
; OPT: [[MERGE]]:
424408
; OPT-NEXT: [[PTRNEW:%.*]] = phi ptr [ [[PTR1]], %[[FIRST]] ], [ [[PTR2]], %[[SECOND]] ]
425409
; OPT-NEXT: [[VALLOADED:%.*]] = load i32, ptr [[PTRNEW]], align 4
426-
; OPT-NEXT: store i32 [[VALLOADED]], ptr [[INOUT2]], align 4
410+
; OPT-NEXT: store i32 [[VALLOADED]], ptr [[INOUT]], align 4
427411
; OPT-NEXT: ret void
428412

429413
%val = load i32, ptr %inout
@@ -470,13 +454,11 @@ define ptx_kernel void @grid_const_phi_ngc(ptr byval(%struct.s) align 4 %input1,
470454
; PTX-NEXT: ret;
471455
; OPT-LABEL: define ptx_kernel void @grid_const_phi_ngc(
472456
; OPT-SAME: ptr byval([[STRUCT_S:%.*]]) align 4 [[INPUT1:%.*]], ptr byval([[STRUCT_S]]) [[INPUT2:%.*]], ptr [[INOUT:%.*]]) #[[ATTR0]] {
473-
; OPT-NEXT: [[INOUT1:%.*]] = addrspacecast ptr [[INOUT]] to ptr addrspace(1)
474-
; OPT-NEXT: [[INOUT2:%.*]] = addrspacecast ptr addrspace(1) [[INOUT1]] to ptr
475457
; OPT-NEXT: [[INPUT2_PARAM:%.*]] = addrspacecast ptr [[INPUT2]] to ptr addrspace(101)
476458
; OPT-NEXT: [[INPUT2_PARAM_GEN:%.*]] = call ptr @llvm.nvvm.ptr.param.to.gen.p0.p101(ptr addrspace(101) [[INPUT2_PARAM]])
477459
; OPT-NEXT: [[INPUT1_PARAM:%.*]] = addrspacecast ptr [[INPUT1]] to ptr addrspace(101)
478460
; OPT-NEXT: [[INPUT1_PARAM_GEN:%.*]] = call ptr @llvm.nvvm.ptr.param.to.gen.p0.p101(ptr addrspace(101) [[INPUT1_PARAM]])
479-
; OPT-NEXT: [[VAL:%.*]] = load i32, ptr [[INOUT2]], align 4
461+
; OPT-NEXT: [[VAL:%.*]] = load i32, ptr [[INOUT]], align 4
480462
; OPT-NEXT: [[LESS:%.*]] = icmp slt i32 [[VAL]], 0
481463
; OPT-NEXT: br i1 [[LESS]], label %[[FIRST:.*]], label %[[SECOND:.*]]
482464
; OPT: [[FIRST]]:
@@ -488,7 +470,7 @@ define ptx_kernel void @grid_const_phi_ngc(ptr byval(%struct.s) align 4 %input1,
488470
; OPT: [[MERGE]]:
489471
; OPT-NEXT: [[PTRNEW:%.*]] = phi ptr [ [[PTR1]], %[[FIRST]] ], [ [[PTR2]], %[[SECOND]] ]
490472
; OPT-NEXT: [[VALLOADED:%.*]] = load i32, ptr [[PTRNEW]], align 4
491-
; OPT-NEXT: store i32 [[VALLOADED]], ptr [[INOUT2]], align 4
473+
; OPT-NEXT: store i32 [[VALLOADED]], ptr [[INOUT]], align 4
492474
; OPT-NEXT: ret void
493475
%val = load i32, ptr %inout
494476
%less = icmp slt i32 %val, 0
@@ -531,17 +513,15 @@ define ptx_kernel void @grid_const_select(ptr byval(i32) align 4 %input1, ptr by
531513
; PTX-NEXT: ret;
532514
; OPT-LABEL: define ptx_kernel void @grid_const_select(
533515
; OPT-SAME: ptr byval(i32) align 4 [[INPUT1:%.*]], ptr byval(i32) [[INPUT2:%.*]], ptr [[INOUT:%.*]]) #[[ATTR0]] {
534-
; OPT-NEXT: [[INOUT1:%.*]] = addrspacecast ptr [[INOUT]] to ptr addrspace(1)
535-
; OPT-NEXT: [[INOUT2:%.*]] = addrspacecast ptr addrspace(1) [[INOUT1]] to ptr
536516
; OPT-NEXT: [[INPUT2_PARAM:%.*]] = addrspacecast ptr [[INPUT2]] to ptr addrspace(101)
537517
; OPT-NEXT: [[INPUT2_PARAM_GEN:%.*]] = call ptr @llvm.nvvm.ptr.param.to.gen.p0.p101(ptr addrspace(101) [[INPUT2_PARAM]])
538518
; OPT-NEXT: [[INPUT1_PARAM:%.*]] = addrspacecast ptr [[INPUT1]] to ptr addrspace(101)
539519
; OPT-NEXT: [[INPUT1_PARAM_GEN:%.*]] = call ptr @llvm.nvvm.ptr.param.to.gen.p0.p101(ptr addrspace(101) [[INPUT1_PARAM]])
540-
; OPT-NEXT: [[VAL:%.*]] = load i32, ptr [[INOUT2]], align 4
520+
; OPT-NEXT: [[VAL:%.*]] = load i32, ptr [[INOUT]], align 4
541521
; OPT-NEXT: [[LESS:%.*]] = icmp slt i32 [[VAL]], 0
542522
; OPT-NEXT: [[PTRNEW:%.*]] = select i1 [[LESS]], ptr [[INPUT1_PARAM_GEN]], ptr [[INPUT2_PARAM_GEN]]
543523
; OPT-NEXT: [[VALLOADED:%.*]] = load i32, ptr [[PTRNEW]], align 4
544-
; OPT-NEXT: store i32 [[VALLOADED]], ptr [[INOUT2]], align 4
524+
; OPT-NEXT: store i32 [[VALLOADED]], ptr [[INOUT]], align 4
545525
; OPT-NEXT: ret void
546526
%val = load i32, ptr %inout
547527
%less = icmp slt i32 %val, 0

0 commit comments

Comments
 (0)