[llvm] [AMDGPU] Allocate i1 argument to SGPRs (PR #72461)
Jun Wang via llvm-commits
llvm-commits at lists.llvm.org
Mon Dec 4 15:23:44 PST 2023
================
@@ -715,6 +715,14 @@ bool SILowerI1Copies::lowerCopiesToI1() {
assert(!MI.getOperand(1).getSubReg());
if (!SrcReg.isVirtual() || (!isLaneMaskReg(SrcReg) && !isVreg1(SrcReg))) {
+ if (!SrcReg.isVirtual() &&
+ TII->getRegisterInfo().getRegSizeInBits(SrcReg, *MRI) == 64) {
+ // When calling convention allocates SGPR for i1, for GPUs with
+ // wavefront size 64, i1 return value is put in 64b SGPR.
+ assert(ST->isWave64());
+ continue;
+ }
+
----------------
jwanggit86 wrote:
Immediate after these new lines is an assert():
`assert(TII->getRegisterInfo().getRegSizeInBits(SrcReg, *MRI) == 32);`
which makes sure SrcReg is 32b before lowering the copy to `V_CMP_NE_U32_e64`.
Because, with the callingconv change, the i1 return value is in an SGPR_64 (for wave64 GPUs), these lines were inserted to avoid triggering the assert.
https://github.com/llvm/llvm-project/pull/72461
More information about the llvm-commits
mailing list