| Issue |
176468
|
| Summary |
Refactor `CodeGenFunction::EmitStoreThroughLValue` in `CGExpr.cpp` to consolidate HLSL-specific behavior into helper functions
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
Icohedron
|
As pointed out by @farzonl in [#176216](https://github.com/llvm/llvm-project/pull/176216/changes/BASE..13d849da4710f37710e909699c6150d7a52ec851#r2699745999) the `CodeGenFunction::EmitStoreThroughLValue` function in `CGExpr.cpp` has three significantly large HLSL-specific if-statements for handling stores through LValues of types `Dst.isVectorElt()`, `Dst.isMatrixElt()`, and `Dst.isMatrixRow()`.
The HLSL-specific behaviors essentially boil down to:
- Zero-extending the Src (specifically, an i1 vector) to its corresponding type for the given destination Address
```c++
llvm::Value *Val = Src.getScalarVal();
if (Val->getType()->getPrimitiveSizeInBits() <
ElemTy->getScalarSizeInBits())
Val = Builder.CreateZExt(Val, ElemTy->getScalarType());
```
- Storing individual elements of vectors/matrices into memory via a vector GEP followed by a scalar store, as opposed to storing the entire vector/matrix as one unit
```c++
llvm::Value *Idx = Dst.getVectorIdx();
llvm::Value *Zero = llvm::ConstantInt::get(Int32Ty, 0);
Address DstElemAddr =
Builder.CreateGEP(DstAddr, {Zero, Idx}, DestAddrTy, ElemAlign);
Builder.CreateStore(Val, DstElemAddr, Dst.isVolatileQualified());
```
It would be preferrable to have these behaviors consolidated into their own (HLSL-specific) helper functions, which should not only make the code easier to read by reducing code duplication, but also potentially make it easier for other language modes to adopt the use of one or both of these behaviors.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
