AMDGPU pass does not respect register class constraint when folding immediate into copies #130020
Description
# RUN: llc -mtriple=amdgcn -mcpu=gfx90a -run-pass=si-fold-operands -verify-machineinstrs -o - %s
# av_class constraint is not respected, and copy is replaced with an illegal v_mov_b32
---
name: s_mov_b32_imm_literal_copy_s_to_av_32
tracksRegLiveness: true
body: |
bb.0:
%0:sreg_32 = S_MOV_B32 999
%1:av_32 = COPY %0
$agpr0 = COPY %1
S_ENDPGM 0
...
---
name: v_mov_b32_imm_literal_copy_v_to_av_32
tracksRegLiveness: true
body: |
bb.0:
%0:vgpr_32 = V_MOV_B32_e32 999, implicit $exec
%1:av_32 = COPY %0
$agpr0 = COPY %1
S_ENDPGM 0
...
This hits a machine verifier failure:
Append normal operand 1
%1:av_32 = V_MOV_B32_e32 %0:vgpr_32
Folded source from %0:vgpr_32 = V_MOV_B32_e32 999, implicit $exec
into OpNo 1 of %1:av_32 = V_MOV_B32_e32 999, implicit $exec
# After SI Fold Operands
# Machine code for function v_mov_b32_imm_literal_copy_v_to_av_32: IsSSA, NoPHIs, TracksLiveness
bb.0:
%1:av_32 = V_MOV_B32_e32 999, implicit $exec
$agpr0 = COPY %1:av_32
S_ENDPGM 0
# End machine code for function v_mov_b32_imm_literal_copy_v_to_av_32.
*** Bad machine code: Illegal virtual register for instruction ***
- function: v_mov_b32_imm_literal_copy_v_to_av_32
- basic block: %bb.0 (0x13b009928)
- instruction: %1:av_32 = V_MOV_B32_e32 999, implicit $exec
- operand 0: %1:av_32
Expected a VGPR_32 register, but got a AV_32 register
It is not possible to directly write a literal value to an AGPR, so it is not legal to fold these