EarlyCSE: create casts on type-mismatch

getOrCreateResult suffers from the deficiency that it doesn't attempt to
create casts when types mismatch. Fix this deficiency, making EarlyCSE
more powerful.

Author: artagnon

llvm/lib/Transforms/Scalar/EarlyCSE.cpp

@@ -31,6 +31,7 @@
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/Dominators.h"
 #include "llvm/IR/Function.h"
+#include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/InstrTypes.h"
 #include "llvm/IR/Instruction.h"
 #include "llvm/IR/Instructions.h"
@@ -964,32 +965,45 @@ class EarlyCSE {
   bool overridingStores(const ParseMemoryInst &Earlier,
                         const ParseMemoryInst &Later);
 
-  Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const {
-    // TODO: We could insert relevant casts on type mismatch here.
-    if (auto *LI = dyn_cast<LoadInst>(Inst))
-      return LI->getType() == ExpectedType ? LI : nullptr;
-    if (auto *SI = dyn_cast<StoreInst>(Inst)) {
-      Value *V = SI->getValueOperand();
-      return V->getType() == ExpectedType ? V : nullptr;
+  Value *getOrCreateResult(Instruction *Inst, Type *ExpectedType) const {
+    if (!isa<IntrinsicInst, LoadInst, StoreInst>(Inst))
+      llvm_unreachable("Instruction not supported");
+
+    // The load or the store's first operand.
+    Value *V;
+    if (auto *II = dyn_cast<IntrinsicInst>(Inst)) {
+      if (isHandledNonTargetIntrinsic(II->getIntrinsicID()))
+        switch (II->getIntrinsicID()) {
+        case Intrinsic::masked_load:
+          V = II;
+          break;
+        case Intrinsic::masked_store:
+          V = II->getOperand(0);
+          break;
+        default:
+          return nullptr;
+        }
+      else
+        return TTI.getOrCreateResultFromMemIntrinsic(II, ExpectedType);
+    } else {
+      V = isa<LoadInst>(Inst) ? Inst : cast<StoreInst>(Inst)->getValueOperand();
     }
-    assert(isa<IntrinsicInst>(Inst) && "Instruction not supported");
-    auto *II = cast<IntrinsicInst>(Inst);
-    if (isHandledNonTargetIntrinsic(II->getIntrinsicID()))
-      return getOrCreateResultNonTargetMemIntrinsic(II, ExpectedType);
-    return TTI.getOrCreateResultFromMemIntrinsic(II, ExpectedType);
-  }
 
-  Value *getOrCreateResultNonTargetMemIntrinsic(IntrinsicInst *II,
-                                                Type *ExpectedType) const {
-    // TODO: We could insert relevant casts on type mismatch here.
-    switch (II->getIntrinsicID()) {
-    case Intrinsic::masked_load:
-      return II->getType() == ExpectedType ? II : nullptr;
-    case Intrinsic::masked_store: {
-      Value *V = II->getOperand(0);
-      return V->getType() == ExpectedType ? V : nullptr;
-    }
-    }
+    Type *ActualType = V->getType();
+    BasicBlock *TheBB = Inst->getParent();
+
+    // First handle the case when no cast is required.
+    if (ActualType == ExpectedType)
+      return V;
+
+    // Try to create BitCast, SExt, or Trunc.
+    IRBuilder<> Builder(TheBB, std::next(Inst->getIterator()));
+    if (CastInst::castIsValid(Instruction::BitCast, V, ExpectedType))
+      return Builder.CreateBitCast(V, ExpectedType);
+    if (CastInst::castIsValid(Instruction::SExt, V, ExpectedType))
+      return Builder.CreateSExt(V, ExpectedType);
+    if (CastInst::castIsValid(Instruction::Trunc, V, ExpectedType))
+      return Builder.CreateTrunc(V, ExpectedType);
     return nullptr;
   }
 

llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-call.ll

@@ -3683,7 +3683,7 @@ define amdgpu_kernel void @test_call_external_void_func_v32i32_i8_i8_i16() #0 {
   ; CHECK-NEXT:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[DEF]](p4) :: (invariant load (p1) from `ptr addrspace(4) undef`, addrspace 4)
   ; CHECK-NEXT:   [[LOAD1:%[0-9]+]]:_(<32 x s32>) = G_LOAD [[LOAD]](p1) :: ("amdgpu-noclobber" load (<32 x s32>) from %ir.ptr0, addrspace 1)
   ; CHECK-NEXT:   [[LOAD2:%[0-9]+]]:_(s8) = G_LOAD [[DEF1]](p1) :: ("amdgpu-noclobber" load (s8) from `ptr addrspace(1) undef`, addrspace 1)
-  ; CHECK-NEXT:   [[LOAD3:%[0-9]+]]:_(s16) = G_LOAD [[DEF1]](p1) :: ("amdgpu-noclobber" load (s16) from `ptr addrspace(1) undef`, addrspace 1)
+  ; CHECK-NEXT:   [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[LOAD2]](s8)
   ; CHECK-NEXT:   ADJCALLSTACKUP 0, 0, implicit-def $scc
   ; CHECK-NEXT:   [[GV:%[0-9]+]]:_(p0) = G_GLOBAL_VALUE @external_void_func_v32i32_i8_i8_i16
   ; CHECK-NEXT:   [[COPY9:%[0-9]+]]:_(p4) = COPY [[COPY7]]
@@ -3720,7 +3720,7 @@ define amdgpu_kernel void @test_call_external_void_func_v32i32_i8_i8_i16() #0 {
   ; CHECK-NEXT:   G_STORE [[COPY18]](s16), [[PTR_ADD3]](p5) :: (store (s16) into stack + 8, align 8, addrspace 5)
   ; CHECK-NEXT:   [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
   ; CHECK-NEXT:   [[PTR_ADD4:%[0-9]+]]:_(p5) = G_PTR_ADD [[AMDGPU_WAVE_ADDRESS]], [[C6]](s32)
-  ; CHECK-NEXT:   G_STORE [[LOAD3]](s16), [[PTR_ADD4]](p5) :: (store (s16) into stack + 12, align 4, addrspace 5)
+  ; CHECK-NEXT:   G_STORE [[SEXT]](s16), [[PTR_ADD4]](p5) :: (store (s16) into stack + 12, align 4, addrspace 5)
   ; CHECK-NEXT:   $vgpr0 = COPY [[UV]](s32)
   ; CHECK-NEXT:   $vgpr1 = COPY [[UV1]](s32)
   ; CHECK-NEXT:   $vgpr2 = COPY [[UV2]](s32)