Drew Kersnar
3255d4d762
Decode small byte constants the same way we encode them. The bitcode writer stores ConstantByte values as signed integers, so the reader must rebuild them using the signed ConstantByte::get path. This has high-bit values like b8 255 round-trip as their canonical signed form, b8 -1, instead of tripping the APInt width assertion. This matches current i8 behavior. Before the fix, the new test crashes in llvm-dis with: "APInt.h: Assertion `llvm::isUIntN(BitWidth, val) && "Value is not an N-bit unsigned value"' failed." Bug found while investigating this PR (https://github.com/llvm/llvm-project/pull/177908), which transitions the LSV to emitting the byte type. Fix assisted by AI.
8 lines
245 B
LLVM
8 lines
245 B
LLVM
; RUN: llvm-as < %s | llvm-dis | FileCheck %s
|
|
|
|
; CHECK: @byte.highbit = constant b8 -1
|
|
@byte.highbit = constant b8 255
|
|
|
|
; CHECK: @byte.vector.highbit = constant <2 x b8> <b8 -128, b8 -1>
|
|
@byte.vector.highbit = constant <2 x b8> <b8 128, b8 255>
|