[llvm-branch-commits] [llvm] [AMDGPU] Treat a bitcast between legal types or same-size scalars as free (PR #238412)
Dmitry Sidorov via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Thu Nov 5 05:51:14 PST 2026
================
@@ -1066,6 +1066,17 @@ InstructionCost GCNTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
Elts && Src->isIntegerTy(*Elts))
return InstructionCost(MaskUnpackCostPerElt) * *Elts *
getFullRateInstrCost();
+ // Integer and FP values share registers. The sizes differ for the parts of
+ // a split vector, and a legalized vector may need its lanes repacked.
+ if (Src->getPrimitiveSizeInBits() == Dst->getPrimitiveSizeInBits()) {
----------------
MrSidims wrote:
Okay, I see. We may be scalarizing vector to vector cast and the check is preventing from estimating artificial pairs that might or might be inserted at the scalarization stage. Having this in mind:
1) (nitpick) I'd suggest the following layout):
```
// The generic code scalarizes a vector cast into element casts, which for a bitcast may pair elements of different sizes.
if (Src->getPrimitiveSizeInBits() == Dst->getPrimitiveSizeInBits()) {
// Integer and FP values share registers.
/* the checks */
```
2) if I read the tests correctly - none of them are exercising the check, right? if so, would be nice to have one.
3) the fix now feels like a workaround, the model itself, which currently spells like "bitcast with different number of elements = number of element-wise cast", which doesn't seem right.
https://github.com/llvm/llvm-project/pull/238412
More information about the llvm-branch-commits
mailing list