On Tue, Nov 07, 2017 at 09:40:31AM +0100, Uros Bizjak wrote:
> > - a1 (r93,l0) best MASK_EVEX_REGS, allocno MASK_EVEX_REGS
> > + a1 (r93,l0) best GENERAL_REGS, allocno GENERAL_REGS
> >
> > - a1(r93,l0) costs: AREG:2000,2000 DREG:2000,2000 CREG:2000,2000
> > BREG:2000,2000 SIREG:2000,2000 DIREG:2000,2000 AD_REGS:2000,2000
> > CLOBBERED_REGS:2000,2000 Q_REGS:2000,2000 NON_Q_REGS:2000,2000
> > TLS_GOTBASE_REGS:2000,2000 GENERAL_REGS:2000,2000 MASK_EVEX_REGS:0,0
> > MASK_REGS:2000,2000 ALL_REGS:420000,420000 MEM:10000,10000
> > + a1(r93,l0) costs: AREG:2000,2000 DREG:2000,2000 CREG:2000,2000
> > BREG:2000,2000 SIREG:2000,2000 DIREG:2000,2000 AD_REGS:2000,2000
> > CLOBBERED_REGS:2000,2000 Q_REGS:2000,2000 NON_Q_REGS:2000,2000
> > TLS_GOTBASE_REGS:2000,2000 GENERAL_REGS:2000,2000 MASK_EVEX_REGS:2000,2000
> > MASK_REGS:4000,4000 ALL_REGS:420000,420000 MEM:10000,10000
> >
> > - r93: preferred MASK_EVEX_REGS, alternative GENERAL_REGS, allocno
> > ALL_REGS
> > + r93: preferred GENERAL_REGS, alternative NO_REGS, allocno GENERAL_REGS
> >
> > - a1(r93,l0) costs: AREG:4000,4000 DREG:4000,4000 CREG:4000,4000
> > BREG:4000,4000 SIREG:4000,4000 DIREG:4000,4000 AD_REGS:4000,4000
> > CLOBBERED_REGS:4000,4000 Q_REGS:4000,4000 NON_Q_REGS:4000,4000
> > TLS_GOTBASE_REGS:4000,4000 GENERAL_REGS:4000,4000 MASK_EVEX_REGS:0,0
> > MASK_REGS:2000,2000 ALL_REGS:420000,420000 MEM:10000,10000
> > - a2(r97,l0) costs: SSE_FIRST_REG:6000,6000 NO_REX_SSE_REGS:6000,6000
> > SSE_REGS:6000,6000 EVEX_SSE_REGS:6000,6000 ALL_SSE_REGS:6000,6000
> > MEM:25000,25000
> > - a3(r98,l0) costs: SSE_FIRST_REG:0,0 NO_REX_SSE_REGS:0,0 SSE_REGS:0,0
> > EVEX_SSE_REGS:0,0 ALL_SSE_REGS:0,0 MEM:19000,19000
> > + a1(r93,l0) costs: GENERAL_REGS:4000,4000 MASK_EVEX_REGS:4000,4000
> > MASK_REGS:6000,6000 ALL_REGS:422000,422000 MEM:12000,12000
> > + a2(r97,l0) costs: SSE_FIRST_REG:8000,8000 NO_REX_SSE_REGS:8000,8000
> > SSE_REGS:8000,8000 EVEX_SSE_REGS:8000,8000 ALL_SSE_REGS:8000,8000
> > MEM:27000,27000
> > + a3(r98,l0) costs: SSE_FIRST_REG:2000,2000 NO_REX_SSE_REGS:2000,2000
> > SSE_REGS:2000,2000 EVEX_SSE_REGS:2000,2000 ALL_SSE_REGS:2000,2000
> > MEM:21000,21000
> >
> > Isn't it sufficient that moves disparage slightly the k alternatives?
> > Or are you worried about the case where the pseudo would need to be spilled
> > and LRA would choose to reload it into a %kN register?
>
> The later... Perhaps it is possible to prevent unwanted reloads with
> ?k (or even ??k, we already have ?m<r>, IDK). I have used *k in
> zero-extend patterns to prevent unwanted reloads, which is perhaps too
> strong, but we have ree pass that eliminates unwanted moves in this
> case (*k in move patterns is handled by other regmove elimination
> passes).
?k gives the same IRA decisions as *k, in both cases it considers
GENERAL_REGS as expensive as MASK_EVEX_REGS and picks up GENERAL_REGS
rather than MASK_EVEX_REGS for the pseudo, which results in
kmovb %k1, %eax; test.
Jakub
