https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78580
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org,
| |uros at gcc dot gnu.org
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
/* PR rtl-optimization/78580 */
/* { dg-do compile } */
/* { dg-options "-O0 -ffixed-ebx" } */
extern const signed char a;
int
foo (signed char x)
{
return x;
}
int
main ()
{
foo (a);
return 0;
}
Apparently with just -m32 we have ira_allocno_classes of:
AREG,DREG,CREG,BREG,SIREG,DIREG,AD_REGS,CLOBBERED_REGS,Q_REGS,NON_Q_REGS,TLS_GOTBASE_REGS,GENERAL_REGS,FP_TOP_REG,FP_SECOND_REG,FLOAT_REGS,FLOAT_INT_REGS
while with -m32 -ffixed-ebx
AREG,DREG,CREG,SIREG,DIREG,AD_REGS,CLOBBERED_REGS,NON_Q_REGS,TLS_GOTBASE_REGS,GENERAL_REGS,FP_TOP_REG,FP_SECOND_REG,FLOAT_REGS,FLOAT_INT_REGS
i.e. BREG and Q_REGS is missing in the latter case. update_bad_spill_attribute
only initializes dead_points for the allocno classes, but then we process
allocno with Q_REGS allocno class and in the latter case (sometimes) ICE
because dead_points[Q_REGS] is random garbage.
So, is the bug that i?86 needs Q_REGS to be an allocno class always (shall
ix86_additional_allocno_class_p return true also for Q_REGS? Just for -m32 or
always?), or is the bug that if Q_REGS is not an allocno class then there
shouldn't be allocno with that class?
