linuxrocks123 wrote:
@jayfoad I believe I can confirm worse code generation using this command as
well: `opt -O2 -mtriple=amdgcn -mcpu=gfx900 -S < s_bcnt0.ll | llc
-mtriple=amdgcn -mcpu=gfx900`
As is:
```
bcnt064_ctpop_multiple_uses: ; @bcnt064_ctpop_multiple_uses
; %bb.0:
s_bcnt1_i32_b64 s2, s[0:1]
s_bcnt0_i32_b64 s4, s[0:1]
s_cmp_lg_u64 s[0:1], -1
s_mov_b32 s3, 0
s_cselect_b32 s0, 1, 0
s_mov_b32 s5, s3
;;#ASMSTART
; use s[2:3]
;;#ASMEND
;;#ASMSTART
; use s[4:5]
;;#ASMEND
; return to shader part epilog
```
With patch:
```
bcnt064_ctpop_multiple_uses: ; @bcnt064_ctpop_multiple_uses
; %bb.0:
s_bcnt1_i32_b64 s2, s[0:1]
s_sub_u32 s4, 64, s2
s_subb_u32 s5, 0, 0
s_cmp_lg_u64 s[0:1], -1
s_cselect_b32 s0, 1, 0
s_mov_b32 s3, 0
;;#ASMSTART
; use s[2:3]
;;#ASMEND
;;#ASMSTART
; use s[4:5]
;;#ASMEND
; return to shader part epilog
```
I know it's the same number of instructions in this particular test, but it's
clear the pattern isn't firing when it should.
We can't get rid of the pattern I have, because it's firing in cases where
yours doesn't fire. If you think we need both patterns, though, I'll add both.
Would you like me to?
https://github.com/llvm/llvm-project/pull/164847
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
