On Wed, Apr 30, 2025 at 11:43 PM H.J. Lu <hjl.to...@gmail.com> wrote: > > On Wed, Apr 30, 2025 at 8:12 PM Uros Bizjak <ubiz...@gmail.com> wrote: > > > > On Tue, Apr 29, 2025 at 11:40 PM H.J. Lu <hjl.to...@gmail.com> wrote: > > > > > > SSE_FIRST_REG was added to CLASS_LIKELY_SPILLED_P, which became > > > TARGET_CLASS_LIKELY_SPILLED_P, for > > > > > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=40470 > > > > > > Since RA has been improved and xmm0 is a commonly used register, remove > > > SSE_FIRST_REG from ix86_class_likely_spilled_p to improve xmm0 codegen: > > > > While the AVX version of pblendvb doesn't use XMM0 as an architectural > > register, there are still plenty of other cases, e.g. PCMPESTM and > > PCMPISTRM, SHA and KEYLOCKER insns, so you are risking RA failures > > with these insn if the life of XMM0 is extended. > > This is no different from > > void > foo1 (double x) > { > register double xmm1 __asm ("xmm1") = x; > asm volatile ("# %0" : "+v" (xmm1)); > } > > We don't add XMM1 to ix86_class_likely_spilled_p.
I don't see the relation between the above example and xor %xmm0,%xmm0 ... movaps val,%xmm0 <- needed for pcmpistrm, fails reload due to already live %xmm0 pcmpistrm ... movaps %xmm0, %xmm2 > > OTOH, the (unrelated?) ternlog change changes the expander, where the > > expander does subreg tricks on memory operand, which doesn' look > > correct to me. adjust_address should be used instead. > > > > > > Should I drop the ternlog change and submit a different patch? No, please submit a ternlog change that will avoid subreg with memory. Uros.
