https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111798
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2023-10-16
CC|richard.guenther at gmail dot com |rguenth at gcc dot
gnu.org
Ever confirmed|0 |1
Status|UNCONFIRMED |WAITING
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
If I configured correctly the only change (early) SRA does is
unsigned int fn1K (unsigned int x)
{
+ <unnamed-unsigned:6> y$k;
...
@@ -57,13 +53,14 @@
<bb 2> :
y = sK;
- _1 = y.k;
+ y$k_13 = sK.k;
+ _1 = y$k_13;
_2 = (unsigned char) _1;
_3 = (unsigned char) x_9(D);
_4 = _2 + _3;
_5 = (<unnamed-unsigned:6>) _4;
- y.k = _5;
- _6 = y.k;
+ y$k_14 = _5;
+ _6 = y$k_14;
_11 = (unsigned int) _6;
y ={v} {CLOBBER(eol)};
return _11;
and later DSE eliminates 'y' completely. Note I can't see any big assembly
difference -O2 vs -O2 -fno-tree-sra so I wonder if I really reproduced the
significantly worse code generation issue. In fact generated code looks
better, the stack slot is elided:
> diff -u t.s.good t.s.bad
--- t.s.good 2023-10-16 10:18:27.457651977 +0200
+++ t.s.bad 2023-10-16 10:07:49.997656268 +0200
@@ -19,13 +19,11 @@
.export fn1K
.type fn1K, @function
fn1K:
- subi sp,8
lrw r7, sK
ld.b r7,(r7)
addu r2,r7
movi r7,63
and r2,r7
- addi sp,8
jmp r15
.size fn1K, .-fn1K
.align 1
