[Bug target/113133] [14 Regression] ICE: SIGSEGV in mark_label_nuses(rtx_def*) (emit-rtl.cc:3896) with -O -fno-tree-ter -mavx512f -march=barcelona

[ubizjak at gmail dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113133

--- Comment #8 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Haochen Jiang from comment #6)
> Aha, I see what happened. x/ymm16+ are usable for AVX512F w/o AVX512VL and
> that is why I added that to allow them.
> 
> Let me find a way to see if we can fix this.

It looks to me that ix86_hard_regno_mode_ok should be fixed to allow x/ymm16+
also with EVEX512. Currently we have:

      /* TODO check for QI/HI scalars.  */
      /* AVX512VL allows sse regs16+ for 128/256 bit modes.  */
      if (TARGET_AVX512VL
          && (VALID_AVX256_REG_OR_OI_MODE (mode)
              || VALID_AVX512VL_128_REG_MODE (mode)))
        return true;

so the compiler is unable to change some of the modes of the xmm16 to 128-bit
mode using lowpart_subreg, e.g. DFmode to V4SFmode.

Please also note that your original patch missed to add TARGET_EVEX512 to the
splitter that handles float_truncate with TARGET_USE_VECTOR_FP_CONVERTS.

I propose to proceed with the minimal fix from Comment #3 as a hotfix to
unbreak the testcase in this PR. The real, but more involved fix is to fix
ix86_hard_regno_mode_ok, which I'll leave to you.