On Wed, Aug 29, 2018 at 05:39:26PM +0100, Vlad Lazar wrote:
> r263591 introduced the following regressions on multiple platforms:
> +FAIL: c-c++-common/torture/builtin-arith-overflow-17.c -O0 execution test
> +FAIL: c-c++-common/torture/builtin-arith-overflow-17.c -O2 execution test
> +FAIL: c-c++-common/torture/builtin-arith-overflow-17.c -O2 -flto
> execution test
> +FAIL: c-c++-common/torture/builtin-arith-overflow-17.c -O2 -flto
> -flto-partition=none execution test
> +FAIL: c-c++-common/torture/builtin-arith-overflow-p-17.c -O0 execution
> test
> +FAIL: c-c++-common/torture/builtin-arith-overflow-p-17.c -O2 execution
> test
> +FAIL: c-c++-common/torture/builtin-arith-overflow-p-17.c -O2 -flto
> execution test
> +FAIL: c-c++-common/torture/builtin-arith-overflow-p-17.c -O2 -flto
> -flto-partition=none execution test
>
> The cause for this was that in canonicalize_comparison wi::add was used to
> add a negative number. This meant
> that in case of underflow the flag would not have been set. The solution is
> to use wi::sub if the immediate
> needs to be decremented.
>
> The patch fixes the mentioned regressions.
> Bootstrapped and regtested on aarch64-none-linux-gnu and x86_64-pc-linux-gnu.
LGTM, but ChangeLog entry is missing, can you please provide one before I
can ack it?
> diff --git a/gcc/expmed.c b/gcc/expmed.c
> index be9f0ec9011..84f58f540ab 100644
> --- a/gcc/expmed.c
> +++ b/gcc/expmed.c
> @@ -6235,7 +6235,13 @@ canonicalize_comparison (machine_mode mode, enum
> rtx_code *code, rtx *imm)
> wrapping around in the case of unsigned values. If any occur
> cancel the optimization. */
> wi::overflow_type overflow = wi::OVF_NONE;
> - wide_int imm_modif = wi::add (imm_val, to_add, sgn, &overflow);
> + wide_int imm_modif;
> +
> + if (to_add == 1)
> + imm_modif = wi::add (imm_val, 1, sgn, &overflow);
> + else
> + imm_modif = wi::sub (imm_val, 1, sgn, &overflow);
> +
> if (overflow)
> return;
Jakub
