[DirectX] Remove intrinsic definitions with no use #133459
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-backend-directx Author: Farzon Lotfi (farzonl) ChangesI was planning to use GlobalDCE, but that does more than whats needed to fix the backend. GlobalDCE also doesn't honor This change reduces "Unsupported intrinsic for DXIL lowering" errors when compiling DML shaders from 12218 to 415. and improves our compilation success rate from less than 1% to 44%. Full diff: https://github.com/llvm/llvm-project/pull/133459.diff 5 Files Affected:
diff --git a/llvm/lib/Target/DirectX/DXILLegalizePass.cpp b/llvm/lib/Target/DirectX/DXILLegalizePass.cpp
index 395311e430fbb..8eb0a0458d478 100644
--- a/llvm/lib/Target/DirectX/DXILLegalizePass.cpp
+++ b/llvm/lib/Target/DirectX/DXILLegalizePass.cpp
@@ -12,6 +12,7 @@
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/Instruction.h"
+#include "llvm/IR/Module.h"
#include "llvm/Pass.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include <functional>
@@ -20,6 +21,12 @@
using namespace llvm;
+static void removeDeadIntrinsics(Function &F,
+ SmallVectorImpl<Function *> &ToRemove) {
+ if (F.isIntrinsic() && F.use_empty())
+ ToRemove.push_back(&F);
+}
+
static void fixI8TruncUseChain(Instruction &I,
SmallVectorImpl<Instruction *> &ToRemove,
DenseMap<Value *, Value *> &ReplacedValues) {
@@ -146,11 +153,26 @@ class DXILLegalizationPipeline {
public:
DXILLegalizationPipeline() { initializeLegalizationPipeline(); }
- bool runLegalizationPipeline(Function &F) {
+ bool runLegalizationPipeline(Module &M) {
+ bool Changes = false;
+ SmallVector<Function *> ToRemove;
+ for (auto &F : make_early_inc_range(M.functions())) {
+ Changes |= runFunctionLegalizationPipeline(F);
+ for (auto &LegalizationFn : ModuleLegalizationPipeline)
+ LegalizationFn(F, ToRemove);
+ }
+
+ for (Function *F : ToRemove)
+ F->eraseFromParent();
+
+ return Changes && !ToRemove.empty();
+ }
+
+ bool runFunctionLegalizationPipeline(Function &F) {
SmallVector<Instruction *> ToRemove;
DenseMap<Value *, Value *> ReplacedValues;
for (auto &I : instructions(F)) {
- for (auto &LegalizationFn : LegalizationPipeline)
+ for (auto &LegalizationFn : FunctionLegalizationPipeline)
LegalizationFn(I, ToRemove, ReplacedValues);
}
@@ -164,37 +186,40 @@ class DXILLegalizationPipeline {
SmallVector<
std::function<void(Instruction &, SmallVectorImpl<Instruction *> &,
DenseMap<Value *, Value *> &)>>
- LegalizationPipeline;
+ FunctionLegalizationPipeline;
+ SmallVector<std::function<void(Function &, SmallVectorImpl<Function *> &)>>
+ ModuleLegalizationPipeline;
void initializeLegalizationPipeline() {
- LegalizationPipeline.push_back(fixI8TruncUseChain);
- LegalizationPipeline.push_back(downcastI64toI32InsertExtractElements);
+ FunctionLegalizationPipeline.push_back(fixI8TruncUseChain);
+ FunctionLegalizationPipeline.push_back(
+ downcastI64toI32InsertExtractElements);
+ ModuleLegalizationPipeline.push_back(removeDeadIntrinsics);
}
};
-class DXILLegalizeLegacy : public FunctionPass {
+class DXILLegalizeLegacy : public ModulePass {
public:
- bool runOnFunction(Function &F) override;
- DXILLegalizeLegacy() : FunctionPass(ID) {}
+ bool runOnModule(Module &M) override;
+ DXILLegalizeLegacy() : ModulePass(ID) {}
static char ID; // Pass identification.
};
} // namespace
-PreservedAnalyses DXILLegalizePass::run(Function &F,
- FunctionAnalysisManager &FAM) {
+PreservedAnalyses DXILLegalizePass::run(Module &M, ModuleAnalysisManager &MAM) {
DXILLegalizationPipeline DXLegalize;
- bool MadeChanges = DXLegalize.runLegalizationPipeline(F);
+ bool MadeChanges = DXLegalize.runLegalizationPipeline(M);
if (!MadeChanges)
return PreservedAnalyses::all();
PreservedAnalyses PA;
return PA;
}
-bool DXILLegalizeLegacy::runOnFunction(Function &F) {
+bool DXILLegalizeLegacy::runOnModule(Module &M) {
DXILLegalizationPipeline DXLegalize;
- return DXLegalize.runLegalizationPipeline(F);
+ return DXLegalize.runLegalizationPipeline(M);
}
char DXILLegalizeLegacy::ID = 0;
@@ -204,6 +229,6 @@ INITIALIZE_PASS_BEGIN(DXILLegalizeLegacy, DEBUG_TYPE, "DXIL Legalizer", false,
INITIALIZE_PASS_END(DXILLegalizeLegacy, DEBUG_TYPE, "DXIL Legalizer", false,
false)
-FunctionPass *llvm::createDXILLegalizeLegacyPass() {
+ModulePass *llvm::createDXILLegalizeLegacyPass() {
return new DXILLegalizeLegacy();
}
diff --git a/llvm/lib/Target/DirectX/DXILLegalizePass.h b/llvm/lib/Target/DirectX/DXILLegalizePass.h
index 9d6d1cd19081d..dfa8124be7108 100644
--- a/llvm/lib/Target/DirectX/DXILLegalizePass.h
+++ b/llvm/lib/Target/DirectX/DXILLegalizePass.h
@@ -15,7 +15,7 @@ namespace llvm {
class DXILLegalizePass : public PassInfoMixin<DXILLegalizePass> {
public:
- PreservedAnalyses run(Function &F, FunctionAnalysisManager &FAM);
+ PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
};
} // namespace llvm
diff --git a/llvm/lib/Target/DirectX/DirectX.h b/llvm/lib/Target/DirectX/DirectX.h
index 96a8a08c875f8..98985db39fc61 100644
--- a/llvm/lib/Target/DirectX/DirectX.h
+++ b/llvm/lib/Target/DirectX/DirectX.h
@@ -52,7 +52,7 @@ void initializeDXILLegalizeLegacyPass(PassRegistry &);
/// Pass to Legalize DXIL by remove i8 truncations and i64 insert/extract
/// elements
-FunctionPass *createDXILLegalizeLegacyPass();
+ModulePass *createDXILLegalizeLegacyPass();
/// Initializer for DXILOpLowering
void initializeDXILOpLoweringLegacyPass(PassRegistry &);
diff --git a/llvm/lib/Target/DirectX/DirectXPassRegistry.def b/llvm/lib/Target/DirectX/DirectXPassRegistry.def
index 87d91ead1896f..473f142589d1f 100644
--- a/llvm/lib/Target/DirectX/DirectXPassRegistry.def
+++ b/llvm/lib/Target/DirectX/DirectXPassRegistry.def
@@ -26,6 +26,7 @@ MODULE_ANALYSIS("dxil-root-signature-analysis", dxil::RootSignatureAnalysis())
MODULE_PASS("dxil-data-scalarization", DXILDataScalarization())
MODULE_PASS("dxil-flatten-arrays", DXILFlattenArrays())
MODULE_PASS("dxil-intrinsic-expansion", DXILIntrinsicExpansion())
+MODULE_PASS("dxil-legalize", DXILLegalizePass())
MODULE_PASS("dxil-op-lower", DXILOpLowering())
MODULE_PASS("dxil-pretty-printer", DXILPrettyPrinterPass(dbgs()))
MODULE_PASS("dxil-translate-metadata", DXILTranslateMetadata())
@@ -38,5 +39,4 @@ MODULE_PASS("print<dxil-root-signature>", dxil::RootSignatureAnalysisPrinter(dbg
#define FUNCTION_PASS(NAME, CREATE_PASS)
#endif
FUNCTION_PASS("dxil-resource-access", DXILResourceAccess())
-FUNCTION_PASS("dxil-legalize", DXILLegalizePass())
#undef FUNCTION_PASS
diff --git a/llvm/test/CodeGen/DirectX/remove-dead-intriniscs.ll b/llvm/test/CodeGen/DirectX/remove-dead-intriniscs.ll
new file mode 100644
index 0000000000000..e3b1f0b69d0b7
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/remove-dead-intriniscs.ll
@@ -0,0 +1,22 @@
+
+; RUN: llc %s -mtriple=dxil-pc-shadermodel6.3-library --filetype=asm -o - | FileCheck %s
+
+declare void @llvm.lifetime.start.p0(i64, ptr) #1
+declare void @llvm.lifetime.end.p0(i64, ptr) #1
+declare i32 @llvm.dx.udot.v4i32(<4 x i32>, <4 x i32>) #2
+declare void @llvm.memset.p0.i32(ptr, i8, i32, i1) #3
+
+; CHECK-NOT: declare void @llvm.lifetime.start.p0(i64, ptr)
+; CHECK-NOT: declare void @llvm.lifetime.end.p0(i64, ptr)
+; CHECK-NOT: declare i32 @llvm.dx.udot.v4i32(<4 x i32>, <4 x i32>)
+; CHECK-NOT: declare void @llvm.memset.p0.i32(ptr, i8, i32, i1)
+
+; CHECK-LABEL: empty_fn
+define void @empty_fn () local_unnamed_addr #0 {
+ ret void
+ }
+
+attributes #0 = { convergent norecurse nounwind "hlsl.export"}
+attributes #1 = { nounwind memory(argmem: readwrite) }
+attributes #2 = { nounwind memory(none) }
+attributes #3 = { nounwind memory(argmem: write) }
|
I was planning to use GlobalDCE, but that does more than whats needed to fix the backend. GlobalDCE also doesn't honor `hlsl.export` and making it be able to is more work than expanding the DXIL legalizer. This change reduces "Unsupported intrinsic for DXIL lowering" errors when compiling DML shaders from 12218 to 415. and improves our compilation success rate from less than 1% to 44%.
|
A couple of concerns here:
|
FileCheck Doesn't care about the whitespace differences and the FunctionPass Manager was already invoked for |
…nkage.cpp where intrinsics are getting orphaned
farzonl
force-pushed
the
remove-dead-intrinsic-defines
branch
from
5d8498f to
82cac25
Compare
farzonl
commented
Mar 28, 2025
| @@ -18,15 +18,15 @@ | |||
| using namespace llvm; | |||
|
|
|||
| static bool finalizeLinkage(Module &M) { | |||
| SmallPtrSet<Function *, 8> Funcs; | |||
| SmallVector<Function *> Funcs; | |||
There was a problem hiding this comment.
just a cleanup since I'm touching this file.
llvm-beanz
approved these changes
Mar 28, 2025
V-FEXrt
approved these changes
Mar 28, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
I was planning to use GlobalDCE, but that does more than whats needed to fix the backend. GlobalDCE also doesn't honor
hlsl.exportand making it be able to is more work than expanding the DXIL legalizer.This change reduces "Unsupported intrinsic for DXIL lowering" errors when compiling DML shaders from 12218 to 415. and improves our compilation success rate from less than 1% to 44%.