[LLVM][CodeGen][DirectX] Fix scalarisation when vector ConstantFP is used.

[paulwalker-arm]

When using -use-constant-fp-for-fixed-length-splat splat (float C) becomes ConstantFP(C) rather than ConstantVector(C, C, C...).

[llvmbot]

@llvm/pr-subscribers-backend-directx

Author: Paul Walker (paulwalker-arm)

Changes

When using -use-constant-fp-for-fixed-length-splat splat (float C) becomes ConstantFP(C) rather than ConstantVector(C, C, C...).

---

Full diff: https://github.com/llvm/llvm-project/pull/172684.diff

3 Files Affected:

• (modified) llvm/lib/Target/DirectX/DXILDataScalarization.cpp (+7-11)
• (modified) llvm/lib/Target/DirectX/DXILWriter/DXILBitcodeWriter.cpp (+1-1)
• (modified) llvm/test/CodeGen/DirectX/scalar-bug-117273.ll (+1)

diff --git a/llvm/lib/Target/DirectX/DXILDataScalarization.cpp b/llvm/lib/Target/DirectX/DXILDataScalarization.cpp
index 5f18c37ef1125..d1de0cbe6b75c 100644
--- a/llvm/lib/Target/DirectX/DXILDataScalarization.cpp
+++ b/llvm/lib/Target/DirectX/DXILDataScalarization.cpp
@@ -366,18 +366,14 @@ static Constant *transformInitializer(Constant *Init, Type *OrigType,
   if (isa<VectorType>(OrigType) && isa<ArrayType>(NewType)) {
     // Convert vector initializer to array initializer
     SmallVector<Constant *, MaxVecSize> ArrayElements;
-    if (ConstantVector *ConstVecInit = dyn_cast<ConstantVector>(Init)) {
-      for (unsigned I = 0; I < ConstVecInit->getNumOperands(); ++I)
-        ArrayElements.push_back(ConstVecInit->getOperand(I));
-    } else if (ConstantDataVector *ConstDataVecInit =
-                   llvm::dyn_cast<llvm::ConstantDataVector>(Init)) {
-      for (unsigned I = 0; I < ConstDataVecInit->getNumElements(); ++I)
-        ArrayElements.push_back(ConstDataVecInit->getElementAsConstant(I));
-    } else {
-      assert(false && "Expected a ConstantVector or ConstantDataVector for "
-                      "vector initializer!");
-    }
 
+    unsigned E = cast<FixedVectorType>(OrigType)->getNumElements();
+    for (unsigned I = 0; I != E; ++I)
+      if (Constant *Elt = Init->getAggregateElement(I))
+        ArrayElements.push_back(Elt);
+
+    assert(ArrayElements.size() == E &&
+           "Expected fixed length constant aggregate for vector initializer!");
     return ConstantArray::get(cast<ArrayType>(NewType), ArrayElements);
   }
 
diff --git a/llvm/lib/Target/DirectX/DXILWriter/DXILBitcodeWriter.cpp b/llvm/lib/Target/DirectX/DXILWriter/DXILBitcodeWriter.cpp
index 48a9085820471..57b94ebe19375 100644
--- a/llvm/lib/Target/DirectX/DXILWriter/DXILBitcodeWriter.cpp
+++ b/llvm/lib/Target/DirectX/DXILWriter/DXILBitcodeWriter.cpp
@@ -2016,7 +2016,7 @@ void DXILBitcodeWriter::writeConstants(unsigned FirstVal, unsigned LastVal,
       }
     } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
       Code = bitc::CST_CODE_FLOAT;
-      Type *Ty = CFP->getType();
+      Type *Ty = CFP->getType()->getScalarType();
       if (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy()) {
         Record.push_back(CFP->getValueAPF().bitcastToAPInt().getZExtValue());
       } else if (Ty->isX86_FP80Ty()) {
diff --git a/llvm/test/CodeGen/DirectX/scalar-bug-117273.ll b/llvm/test/CodeGen/DirectX/scalar-bug-117273.ll
index 43bbe9249aac0..c77a3043303e5 100644
--- a/llvm/test/CodeGen/DirectX/scalar-bug-117273.ll
+++ b/llvm/test/CodeGen/DirectX/scalar-bug-117273.ll
@@ -1,5 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
 ; RUN: opt -S -passes='dxil-data-scalarization,dxil-flatten-arrays,function(scalarizer<load-store>),dxil-op-lower' -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
+; RUN: opt -S -passes='dxil-data-scalarization,dxil-flatten-arrays,function(scalarizer<load-store>),dxil-op-lower' -mtriple=dxil-pc-shadermodel6.3-library -use-constant-fp-for-fixed-length-splat %s | FileCheck %s
 
 
 @StaticArr = internal constant [8 x <3 x float>] [<3 x float> zeroinitializer, <3 x float> splat (float 5.000000e-01), <3 x float> <float 1.000000e+00, float 5.000000e-01, float 5.000000e-01>, <3 x float> <float 5.000000e-01, float 1.000000e+00, float 5.000000e-01>, <3 x float> <float 5.000000e-01, float 5.000000e-01, float 1.000000e+00>, <3 x float> <float 5.000000e-01, float 1.000000e+00, float 1.000000e+00>, <3 x float> <float 1.000000e+00, float 5.000000e-01, float 1.000000e+00>, <3 x float> <float 1.000000e+00, float 1.000000e+00, float 5.000000e-01>], align 16