[DirectX] Remove intrinsic definitions with no use

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%.

Author: farzonl

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();
 }

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
 

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 &);

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

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) }