On Sat, Apr 11, 2020 at 02:48:22PM -0700, H.J. Lu via Gcc-patches wrote: > Even though ix86_hard_regno_mode_ok doesn't allow xmm16-xmm31 nor > ymm16-ymm31 in 128/256 bit modes when AVX512VL is disabled, reload > can still generate reg to reg moves with xmm16-xmm31 and ymm16-ymm31 > in 128/256 bit modes. Remove mode size check in ix86_get_ssemov.
Looking at the testcase, LRA doesn't have good choices, (define_insn ("*avx512f_gatherdiv16si") [ (set (match_operand:V16SI 0 ("register_operand") ("=&v")) (unspec:V16SI [ (match_operand:V8SI 1 ("register_operand") ("0")) ... so if the output is chosen to be zmm16+, then LRA is required to stick a V8SI pseudo into that register, even when ix86_hard_regno_mode_ok isn't ok with that. Unless we changed all the patterns with such issues (i.e. where a 128-bit/256-bit operand uses matching constraint to 512-bit one with v) so that it uses just x on the matching operand unless avx512vl. Before your changes, we would just emit the *mov<mode>_internal even in those cases, so your change matches what we used to do. > > gcc/ > > PR target/94561 > * config/i386/i386.c (ix86_get_ssemov): Remove mode size check. > > gcc/testsuite/ > > PR target/94561 > * gcc.target/i386/pr94561.c: New test. Ok. Jakub