https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121218
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|tree-optimization |rtl-optimization
Target| |x86_64-*-*
Priority|P3 |P2
CC| |jakub at gcc dot gnu.org,
| |rsandifo at gcc dot gnu.org
--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
Besides REG note changes CSE2 (when disabling CSE1) only does two things:
- 189: [r117:DI]=r193:DI
+ 189: [r117:DI]=r154:TI#0
...
- 195: [r120:DI]=r195:DI
+ 195: [r120:DI]=r155:TI#0
But we have
181: r193:DI=r154:TI#0
182: r194:DI=r154:TI#8
REG_DEAD r154:TI
183: r195:DI=r155:TI#0
184: r196:DI=r155:TI#8
REG_DEAD r155:TI
and CSE fails to remove those REG_DEAD notes.
The x86 STV pass is exposing this issue, -mno-stv fixes it. It's also
interesting CSE1 triggers the issue as well, but that runs before STV.
STV only does
- 68: [r99:DI+0x48]=0
+ 205: r207:V1TI=const_vector
+ 68: [r99:DI+0x48]=r207:V1TI
REG_DEAD r99:DI
with or without CSE1, so possibly the const_vector trigges the error.
With CSE1 disabled and -mno-stv CSE2 still performs the bogus transform,
but it doesn't trigger a miscompile.
Jakub and Richard S. fixed the last CSE bugs, I hope they know their way
around cse.cc, esp. where it's supposed to remove REG_DEAD and why that
doesn't happen for SUBREG_REGs. The issue is likely latent for a while.
