On Thu, Nov 16, 2023 at 8:16 AM Jakub Jelinek <ja...@redhat.com> wrote:
>
> Hi!
>
> The following testcase is miscompiled on x86_64 since PR110551 r14-4968
> commit. That commit added 2 peephole2s, one for
> mov imm,%rXX; mov %rYY,%rax; mulq %rXX -> mov imm,%rax; mulq %rYY
> which I believe is ok, and another one for
> mov imm,%rXX; mov %rYY,%rdx; mulx %rXX, %rZZ, %rWW -> mov imm,%rdx; mulx
> %rYY, %rZZ, %rWW
> which is wrong. Both peephole2s verify that %rXX above is dead at
> the end of the pattern, by checking if %rXX is either one of the
> registers overwritten in the multiplication (%rdx:%rax in the first
> case, the 2 destination registers of mulx in the latter case), because
> we no longer set %rXX to that immediate (we set %rax resp. %rdx to it
> instead) when the peephole2 replaces it. But, we also need to ensure
> that the other register previously set to the value of %rYY and newly
> to imm isn't used after the multiplication, and neither of the peephole2s
> does that. Now, for the first one (at least assuming in the % pattern
> the matching operand (i.e. hardcoded %rax resp. %rdx) after RA will always go
> first) I think it is always the case, because operands[2] if it must be %rax
> register will be overwritten by mulq writing to %rdx:%rax. But in the
> second case, there is no reason why %rdx couldn't be used after the pattern,
> and if it is (like in the testcase), we can't make those changes.
> So, the patch checks similarly to operands[0] that operands[2] (which ought
> to be %rdx if RA puts the % match_dup operand first and nothing swaps it
> afterwards) is either the same register as one of the destination registers
> of mulx or dies at the end of the multiplication.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2023-11-16 Jakub Jelinek <ja...@redhat.com>
>
> PR target/112526
> * config/i386/i386.md
> (mov imm,%rax; mov %rdi,%rdx; mulx %rax -> mov imm,%rdx; mulx %rdi):
> Verify in define_peephole2 that operands[2] dies or is overwritten
> at the end of multiplication.
>
> * gcc.target/i386/bmi2-pr112526.c: New test.
OK.
Thanks,
Uros.
>
> --- gcc/config/i386/i386.md.jj 2023-11-14 21:38:38.667046713 +0100
> +++ gcc/config/i386/i386.md 2023-11-15 17:03:28.308048728 +0100
> @@ -9918,7 +9918,10 @@ (define_peephole2
> && REGNO (operands[0]) != REGNO (operands[3])
> && (REGNO (operands[0]) == REGNO (operands[4])
> || REGNO (operands[0]) == REGNO (operands[5])
> - || peep2_reg_dead_p (3, operands[0]))"
> + || peep2_reg_dead_p (3, operands[0]))
> + && (REGNO (operands[2]) == REGNO (operands[4])
> + || REGNO (operands[2]) == REGNO (operands[5])
> + || peep2_reg_dead_p (3, operands[2]))"
> [(set (match_dup 2) (match_dup 1))
> (parallel [(set (match_dup 4)
> (mult:DWIH (match_dup 2) (match_dup 3)))
> --- gcc/testsuite/gcc.target/i386/bmi2-pr112526.c.jj 2023-11-15
> 16:58:02.230380183 +0100
> +++ gcc/testsuite/gcc.target/i386/bmi2-pr112526.c 2023-11-15
> 17:02:22.478942259 +0100
> @@ -0,0 +1,27 @@
> +/* PR target/112526 */
> +/* { dg-do run { target { bmi2 && int128 } } } */
> +/* { dg-options "-O2 -mbmi2" } */
> +
> +#include "bmi2-check.h"
> +
> +__attribute__((noipa)) void
> +foo (unsigned long x, unsigned __int128 *y, unsigned long z, unsigned long
> *w)
> +{
> + register unsigned long a __asm ("%r10") = x + z;
> + register unsigned __int128 b __asm ("%r8") = ((unsigned __int128) a) *
> 257342423UL;
> + asm volatile ("" : "+r" (b));
> + asm volatile ("" : "+d" (a));
> + *y = b;
> + *w = a;
> +}
> +
> +static void
> +bmi2_test ()
> +{
> + unsigned __int128 y;
> + unsigned long w;
> + foo (10268318293806702989UL, &y, 4702524958196331333UL, &w);
> + if (y != ((((unsigned __int128) 0xc72d2c9UL) << 64) | 0x9586adfdc95b225eUL)
> + || w != 14970843252003034322UL)
> + abort ();
> +}
>
> Jakub
>
