[Bug target/58853] [4.9 regression] ICE in expand_set_or_movmem_prologue_epilogue_by_misaligned_moves

[hubicka at ucw dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58853

--- Comment #7 from Jan Hubicka <hubicka at ucw dot cz> ---
> diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
> index 53e04c4..dd8d943 100644
> --- a/gcc/config/i386/i386.c
> +++ b/gcc/config/i386/i386.c
> @@ -23766,6 +23766,7 @@ ix86_expand_set_or_movmem (rtx dst, rtx src, rtx
> count_exp, rtx 
> val_exp,
>       also avoids redundant job when sizes are known precisely.  */
>    misaligned_prologue_used = (TARGET_MISALIGNED_MOVE_STRING_PROLOGUES
>                    && MAX (desired_align, epilogue_size_needed) <= 32
> +                  && desired_align <= epilogue_size_needed
>                    && ((desired_align > align && !align_bytes)
>                    || (!count && epilogue_size_needed > 1)));
> 
> avoids crash.

Yes, we can add this, too, for correctness with manual changes to tunning flags
(so the patch is approved).
What confuses me is why TARGET_MISALIGNED_MOVE_STRING_PROLOGUES seems to be
true
given that

DEF_TUNE (TARGET_MISALIGNED_MOVE_STRING_PROLOGUES,
          "misaligned_move_string_prologues",
          m_386 | m_486 | m_CORE_ALL | m_AMD_MULTIPLE | m_GENERIC)

and we tune for pentiumpro (that was intentionally left out for this purpose)

Honza