On Wed, May 21, 2025 at 9:44 AM Jakub Jelinek <ja...@redhat.com> wrote:
>
> Hi!
>
> As mentioned by Linus, we can't optimize comparison of otherwise unused
> result of plus with CONST_INT second operand, compared against zero.
> This can be done using just cmp instruction with negated constant and say
> js/jns/je/jne etc. conditional jumps (or setcc).
> We already have *cmp<mode>_minus_1 instruction which handles it when
> (as shown in foo in the testcase) the IL has MINUS rather than PLUS,
> but for constants except for the minimum value the canonical form is
> with PLUS.
>
> The following patch adds a new pattern to handle this.
>
> I'm not really sure about the flag_cf_protection stuff in there,
> in other spots we just don't care and can sometimes emit immediates
> with endbr64 or endbr32 in them.
> E.g. *add<mode>_1<nf_name> when x86_maybe_negate_const_int
> returns true, or in *subborrow<mode>_1 or *sub<mode>3_eq_1 at least.
I think this is an oversight that should eventually be fixed. Perhaps
we should teach x86_maybe_negate_const_int to always negate
problematic immediates, and never negate constants that would result
in problematic immediates.
> So, I can drop that part.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2025-05-20 Jakub Jelinek <ja...@redhat.com>
>
> PR target/120360
> * config/i386/i386.md (*cmp<mode>_plus_1): New pattern.
>
> * gcc.target/i386/pr120360.c: New test.
>
> --- gcc/config/i386/i386.md.jj 2025-05-20 08:14:05.036425035 +0200
> +++ gcc/config/i386/i386.md 2025-05-20 20:21:39.294358658 +0200
> @@ -1599,6 +1599,25 @@ (define_insn "*cmp<mode>_minus_1"
> [(set_attr "type" "icmp")
> (set_attr "mode" "<MODE>")])
>
> +(define_insn "*cmp<mode>_plus_1"
> + [(set (reg FLAGS_REG)
> + (compare
> + (plus:SWI (match_operand:SWI 0 "nonimmediate_operand" "<r>m")
> + (match_operand:SWI 1 "<immediate_operand>" "<i>"))
Please introduce "x86_64_neg_const_int_operand" predicate that will
allow only const_int operands, and will reject negative endbr (and
wide DImode) const_ints. You don't need a constraint if the predicate
allows only const_ints (the const_int won't be reloaded).
> + (const_int 0)))]
> + "CONST_INT_P (operands[1])
> + && ix86_match_ccmode (insn, CCGOCmode)
> + && (<MODE>mode != DImode || !mode_signbit_p (SImode, operands[1]))
> + && ((flag_cf_protection & CF_BRANCH) == 0
> + || ((-INTVAL (operands[1]) & 0xffffffffUL)
> + != (TARGET_64BIT ? 0xfa1e0ff3UL : 0xfb1e0ff3UL)))"
New predicate will also avoid the above horrible insn condition.
Uros.
> +{
> + operands[1] = gen_int_mode (-INTVAL (operands[1]), <MODE>mode);
> + return "cmp{<imodesuffix>}\t{%1, %0|%0, %1}";
> +}
> + [(set_attr "type" "icmp")
> + (set_attr "mode" "<MODE>")])
> +
> (define_insn "*cmpqi_ext<mode>_1"
> [(set (reg FLAGS_REG)
> (compare
> --- gcc/testsuite/gcc.target/i386/pr120360.c.jj 2025-05-20 20:40:32.194962716
> +0200
> +++ gcc/testsuite/gcc.target/i386/pr120360.c 2025-05-20 20:43:03.910930990
> +0200
> @@ -0,0 +1,36 @@
> +/* PR target/120360 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fno-stack-protector -masm=att" } */
> +/* { dg-additional-options "-fno-pic" { target { ! *-*-darwin* } } } */
> +/* { dg-final { scan-assembler-times "\tjn*s\t" 3 } } */
> +/* { dg-final { scan-assembler-times "\tcmp\[lq]\t%" 1 } } */
> +/* { dg-final { scan-assembler-times "\tcmp\[lq]\t\\\$-1234," 1 } } */
> +/* { dg-final { scan-assembler-times "\tcmp\[lq]\t\\\$2345," 1 } } */
> +/* { dg-final { scan-assembler-not "\tadd\[lq]\t" { target { ! *-*-darwin* }
> } } } */
> +/* { dg-final { scan-assembler-not "\tsub\[lq]\t" { target { ! *-*-darwin* }
> } } } */
> +
> +void qux (unsigned long);
> +
> +void
> +foo (unsigned long x, unsigned long y)
> +{
> + unsigned long z = x - y;
> + if ((long) z < 0)
> + qux (x);
> +}
> +
> +void
> +bar (unsigned long x)
> +{
> + unsigned long z = x + 1234;
> + if ((long) z < 0)
> + qux (x);
> +}
> +
> +void
> +baz (unsigned long x)
> +{
> + unsigned long z = x - 2345;
> + if ((long) z < 0)
> + qux (x);
> +}
>
> Jakub
>
