================
@@ -1823,10 +1819,11 @@ mlir::Attribute
ConstantEmitter::emitForMemory(CIRGenModule &cgm,
cgm.errorNYI("emitForMemory: zero-extend HLSL bool vectors");
}
- if (destType->isBitIntType()) {
- cgm.errorNYI("emitForMemory: _BitInt type");
- }
-
+ // CIR represents a _BitInt(N) value in memory at its exact width
----------------
adams381 wrote:
Confirmed at the byte level: `signed _BitInt(6) = -1` emits `0x3f` under
`-fclangir` where classic/GCC write `0xff` (and `signed _BitInt(17) = -1` ->
`ff ff 01 00` vs `ff ff ff ff`). The exact-width value reaches LLVM and the
backend zero-fills the storage padding instead of sign-extending it; unsigned
and non-negative values happen to match, which is why the value-level tests
missed it.
This matches how `cir.bool` already works, so `return c` here is fine -- the
gap is that lowering doesn't expand `_BitInt` yet. I'll add that to LowerToLLVM
(`convertTypeForMemory` + `emitToMemory`/`emitFromMemory`, sign/zero-extending
to the storage integer), erroring out on the long-`_BitInt` byte-array case for
now instead of emitting wrong bytes, and rebuild the PR on that.
https://github.com/llvm/llvm-project/pull/205605
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
