[DirectX] replace byte splitting via vector bitcast with scalar

instructions
- instead of bitcasting and extract element lets use trunc or trunc  and logical shift right to split.
- fixes #139020

Author: farzonl

llvm/lib/Target/DirectX/DXILLegalizePass.cpp

@@ -8,6 +8,8 @@
 
 #include "DXILLegalizePass.h"
 #include "DirectX.h"
+#include "llvm/ADT/APInt.h"
+#include "llvm/IR/Constants.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/InstIterator.h"
@@ -510,6 +512,49 @@ static void updateFnegToFsub(Instruction &I,
   ToRemove.push_back(&I);
 }
 
+static void
+legalizeGetHighLowi64Bytes(Instruction &I,
+                           SmallVectorImpl<Instruction *> &ToRemove,
+                           DenseMap<Value *, Value *> &ReplacedValues) {
+  if (auto *BitCast = dyn_cast<BitCastInst>(&I)) {
+    if (BitCast->getDestTy() ==
+            FixedVectorType::get(Type::getInt32Ty(I.getContext()), 2) &&
+        BitCast->getSrcTy()->isIntegerTy(64)) {
+      ToRemove.push_back(BitCast);
+      ReplacedValues[BitCast] = BitCast->getOperand(0);
+    }
+  }
+
+  if (auto *Extract = dyn_cast<ExtractElementInst>(&I)) {
+    auto *VecTy = dyn_cast<FixedVectorType>(Extract->getVectorOperandType());
+    if (VecTy && VecTy->getElementType()->isIntegerTy(32) &&
+        VecTy->getNumElements() == 2) {
+      if (auto *Index = dyn_cast<ConstantInt>(Extract->getIndexOperand())) {
+        unsigned Idx = Index->getZExtValue();
+        IRBuilder<> Builder(&I);
+        assert(dyn_cast<BitCastInst>(Extract->getVectorOperand()));
+        auto *Replacement = ReplacedValues[Extract->getVectorOperand()];
+        if (Idx == 0) {
+          Value *LowBytes = Builder.CreateTrunc(
+              Replacement, Type::getInt32Ty(I.getContext()));
+          ReplacedValues[Extract] = LowBytes;
+        } else {
+          assert(Idx == 1);
+          Value *LogicalShiftRight = Builder.CreateLShr(
+              Replacement,
+              ConstantInt::get(
+                  Replacement->getType(),
+                  APInt(Replacement->getType()->getIntegerBitWidth(), 32)));
+          Value *HighBytes = Builder.CreateTrunc(
+              LogicalShiftRight, Type::getInt32Ty(I.getContext()));
+          ReplacedValues[Extract] = HighBytes;
+        }
+        ToRemove.push_back(Extract);
+      }
+    }
+  }
+}
+
 namespace {
 class DXILLegalizationPipeline {
 
@@ -544,6 +589,7 @@ class DXILLegalizationPipeline {
     LegalizationPipeline.push_back(legalizeMemCpy);
     LegalizationPipeline.push_back(removeMemSet);
     LegalizationPipeline.push_back(updateFnegToFsub);
+    LegalizationPipeline.push_back(legalizeGetHighLowi64Bytes);
   }
 };
 

llvm/test/CodeGen/DirectX/legalize-i64-high-low-vec-spilt.ll

@@ -0,0 +1,18 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes='dxil-legalize' -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
+
+define void @split_via_extract(i64 noundef %a) {
+; CHECK-LABEL: define void @split_via_extract(
+; CHECK-SAME: i64 noundef [[A:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP0:%.*]] = trunc i64 [[A]] to i32
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr i64 [[A]], 32
+; CHECK-NEXT:    [[TMP2:%.*]] = trunc i64 [[TMP1]] to i32
+; CHECK-NEXT:    ret void
+;
+entry:
+  %vecA = bitcast i64 %a to <2 x i32>
+  %low = extractelement <2 x i32> %vecA, i32 0 ; low 32 bits
+  %high = extractelement <2 x i32> %vecA, i32 1 ; high 32 bits
+  ret void
+}