[CIR] [Lowering] care trailing zero for lowering constant array #976
Conversation
| @@ -2229,7 +2229,10 @@ class CIRGlobalOpLowering | |||
| if (auto constArr = | |||
| mlir::dyn_cast<mlir::cir::ConstArrayAttr>(init.value())) { | |||
| if (auto attr = mlir::dyn_cast<mlir::StringAttr>(constArr.getElts())) { | |||
| init = rewriter.getStringAttr(attr.getValue()); | |||
| llvm::SmallString<256> literal(attr.getValue()); | |||
| if (constArr.getTrailingZerosNum()) | |||
There was a problem hiding this comment.
Hmm. For regular strings we just emit the string directly (https://godbolt.org/z/a6xWGx3hf), but for trailing NUL bytes we use trailing zeros (https://godbolt.org/z/r54vvKPb7). Maybe we shouldn't be generating the trailing zeroes instead (assuming there's no technical limitation around that)?
There was a problem hiding this comment.
LLVM will generate the trailing zero: https://godbolt.org/z/4G75e4P1o . I feel better if we can get align with LLVM in the traditional pipeline. (this is LowerToLLVM.cpp and it shouldn't affect how we generate cir)
There was a problem hiding this comment.
Perhaps OG should be fixed too? Anyways, in the meantime I'm sold on keeping it consistent with OG
| @@ -2229,7 +2229,10 @@ class CIRGlobalOpLowering | |||
| if (auto constArr = | |||
| mlir::dyn_cast<mlir::cir::ConstArrayAttr>(init.value())) { | |||
| if (auto attr = mlir::dyn_cast<mlir::StringAttr>(constArr.getElts())) { | |||
| init = rewriter.getStringAttr(attr.getValue()); | |||
| llvm::SmallString<256> literal(attr.getValue()); | |||
| if (constArr.getTrailingZerosNum()) | |||
There was a problem hiding this comment.
Perhaps OG should be fixed too? Anyways, in the meantime I'm sold on keeping it consistent with OG
Close #975 See the attached test case for example
Close #975
See the attached test case for example