| Issue |
171724
|
| Summary |
AMDGPU: Select of double with fneg inputs does not use source modifiers with v_cndmask_b32
|
| Labels |
backend:AMDGPU,
missed-optimization
|
| Assignees |
|
| Reporter |
arsenm
|
https://godbolt.org/z/n9GYarEof
Select of f64 with fneg/fabs inputs does not produce the ideal output. It does not make use of source modifiers, and we get bit instructions:
```
; Missed source modifiers on the high half
; v_xor_b32_e32 v2, 0x80000000, v2
; v_xor_b32_e32 v4, 0x80000000, v4
; v_cmp_eq_u32_e32 vcc, 0, v0
; v_cndmask_b32_e32 v0, v1, v3, vcc
; v_cndmask_b32_e32 v1, v2, v4, vcc
define double @select_f64_fneg(i32 %cond, double %arg, double %arg1) {
bb:
%neg.0 = fneg double %arg
%neg.1 = fneg double %arg1
%cmp = icmp eq i32 %cond, 0
%select = select i1 %cmp, double %neg.1, double %neg.0
ret double %select
}
```
Oddly if the fneg is pulled out of the select, we get the ideal result, with the fneg pushed back into the select:
```
; Does produce the ideal result
; v_cmp_eq_u32_e32 vcc, 0, v0
; v_cndmask_b32_e32 v0, v3, v1, vcc
; v_cndmask_b32_e64 v1, -v4, -v2, vcc
define double @fneg_select_f64(i32 %cond, double %arg, double %arg1) {
bb:
%cmp = icmp eq i32 %cond, 0
%select = select i1 %cmp, double %arg, double %arg1
%neg.select = fneg double %select
ret double %neg.select
}
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
