[Bug target/69671] [6 Regression] FAIL: gcc.target/i386/avx512vl-vpmovqb-1.c scan-assembler-times vpmovqb[ \\t]+[^{\n]*%ymm[0-9]+[^\n]*%xmm[0-9]+{%k[1-7]}{z}(?

[vmakarov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69671

--- Comment #13 from Vladimir Makarov <vmakarov at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #9)
> But something like that might remove the flexibility from the register
> allocator.
> 
> Wonder why the RA in this case doesn't see that the value loaded into that
> pseudo register is CONST0_RTX which satisfies the C constraint and doesn't
> undo CSE (rematerialize) in that case if it doesn't have that value already
> loaded in the matching register to the output one.

LRA remat is a simple global scope subpass to remove reload loads from memory
by reusing values held in registers.  It does not check constraints at all. 
Adding some checks will complicate the sub-pass much.

I think Kirill proposed a good solution:

"Split all patterns into match_dup and 0_operand by hand"

I believe it will work well for RA.