[PATCH 2/2] x86: Add cmpmemsi for -minline-all-stringops

[H.J. Lu]

On Mon, Jun 1, 2020 at 3:19 AM Alexander Monakov <amonakov@ispras.ru> wrote:
>
> On Sun, 31 May 2020, H.J. Lu via Gcc-patches wrote:
>
> > --- a/gcc/config/i386/i386-expand.c
> > +++ b/gcc/config/i386/i386-expand.c
> > @@ -7656,6 +7656,90 @@ ix86_expand_set_or_cpymem (rtx dst, rtx src, rtx count_exp, rtx val_exp,
> >    return true;
> >  }
> >
> > +/* Expand cmpstrn or memcmp.  */
> > +
> > +bool
> > +ix86_expand_cmpstrn_or_cmpmem (rtx result, rtx src1, rtx src2,
> > +                            rtx length, rtx align, bool is_cmpstrn)
> > +{
> > +  if (optimize_insn_for_size_p () && !TARGET_INLINE_ALL_STRINGOPS)
> > +    return false;
> > +
> > +  /* Can't use this if the user has appropriated ecx, esi or edi.  */
> > +  if (fixed_regs[CX_REG] || fixed_regs[SI_REG] || fixed_regs[DI_REG])
> > +    return false;
> > +
> > +  if (is_cmpstrn)
> > +    {
> > +      /* For strncmp, length is the maximum length, which can be larger
> > +      than actual string lengths.  We can expand the cmpstrn pattern
> > +      to "repz cmpsb" only if one of the strings is a constant so
> > +      that expand_builtin_strncmp() can write the length argument to
> > +      be the minimum of the const string length and the actual length
> > +      argument.  Otherwise, "repz cmpsb" may pass the 0 byte.  */
> > +      tree t1 = MEM_EXPR (src1);
> > +      tree t2 = MEM_EXPR (src2);
> > +      if (!((t1 && TREE_CODE (t1) == MEM_REF
> > +          && TREE_CODE (TREE_OPERAND (t1, 0)) == ADDR_EXPR
> > +          && (TREE_CODE (TREE_OPERAND (TREE_OPERAND (t1, 0), 0))
> > +              == STRING_CST))
> > +         || (t2 && TREE_CODE (t2) == MEM_REF
> > +             && TREE_CODE (TREE_OPERAND (t2, 0)) == ADDR_EXPR
> > +             && (TREE_CODE (TREE_OPERAND (TREE_OPERAND (t2, 0), 0))
> > +                 == STRING_CST))))
> > +     return false;
> > +    }
> > +  else
> > +    {
> > +      /* Expand memcmp to "repz cmpsb" only for -minline-all-stringops
> > +      since "repz cmpsb" can be much slower than memcmp function
> > +      implemented with vector instructions, see
> > +
> > +      https://gcc.gnu.org/bugzilla/show_bug.cgi?id=43052
> > +       */
> > +      if (!TARGET_INLINE_ALL_STRINGOPS)
> > +     return false;
> > +    }
>
> This check seems to be misplaced, "rep cmps" is slower than either memcmp or
> strcmp. The test for TARGET_INLINE_ALL_STRINGOPS should happen regardless of
> is_cmpstrn, so it should go earlier in the function.
>

My patch doesn't change strncmp at all.   I opened:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95458


-- 
H.J.