https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102834
--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Rainer Orth from comment #1)
> Created attachment 51632 [details]
> 32-bit avx512f-pr96891-3.s
It relies on the existence of the pattern one_cmpldi2_doubleword, so that the
avx512bw_cmpv64qi3 and one_cmpldi2_doubleword will be combined to
avx512bw_cmpv64qi3 w/ adjustment of imm operand, but on 32-bit Solaris/x86
one_cmpldi2_doubleword is not existed.
(define_expand "one_cmpl<mode>2"
[(set (match_operand:SWIM1248s 0 "nonimmediate_operand")
(not:SWIM1248s (match_operand:SWIM1248s 1 "nonimmediate_operand")))]
(define_mode_iterator SWIM1248s
[(QI "TARGET_QIMODE_MATH")
(HI "TARGET_HIMODE_MATH")
SI (DI "TARGET_64BIT || (TARGET_STV && TARGET_SSE2)")])
mstv
Target Mask(STV) Save
Disable Scalar to Vector optimization pass transforming 64-bit integer
computations into a vector ones.
Guess stv is disabled on on 32-bit Solaris/x86, if that, the fail should be
expected and xfail should be added for 32-bit Solaris/x86