On Fri, Feb 7, 2020 at 8:58 AM Jakub Jelinek <ja...@redhat.com> wrote:
>
> Hi!
>
> The following testcase ICEs. The generated split_insns starts
> with recog_data.insn = NULL and then tries to put various operands into
> recog_data.operand array and checks various splitter conditions.
> The problem is that some atom related tuning splitters indirectly call
> extract_insn_cached on the insn they are used in. This can change
> recog_data.operand, but most likely it will just keep it as is, but
> sets recog_data.insn to the current instruction. If that splitter doesn't
> match, we continue trying some other split conditions and modify
> recog_data.operand array again. If even that doesn't find any usable
> splitter, we punt, but at that point recog_data.insn says that recog_data
> is valid for that particular instruction, even when recog_data.operand array
> can be anything.
> The safest thing would be to copy whole recog_data to a temporary object
> before doing the calls that can call extract_insn_cached and restore it
> afterwards, but it would be also very costly, recog_data has 1280 bytes.
> So, this patch just makes sure to clear recog_data.insn if it has changed
> during the extract_insn_cached call, which means if we extract_insn_cached
> later, we'll extract it properly, while if we call it say from some other
> context than splitter conditions, the insn is already cached, we don't reset
> the cache.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2020-02-07 Jakub Jelinek <ja...@redhat.com>
>
> PR target/93611
> * config/i386/i386.c (ix86_lea_outperforms): Make sure to clear
> recog_data.insn if distance_non_agu_define changed it.
>
> * gcc.target/i386/pr93611.c: New test.
LGTM.
Thanks,
Uros.
> --- gcc/config/i386/i386.c.jj 2020-01-28 08:45:56.781090684 +0100
> +++ gcc/config/i386/i386.c 2020-02-06 16:29:35.548663197 +0100
> @@ -14459,9 +14459,18 @@ ix86_lea_outperforms (rtx_insn *insn, un
> return true;
> }
>
> + rtx_insn *rinsn = recog_data.insn;
> +
> dist_define = distance_non_agu_define (regno1, regno2, insn);
> dist_use = distance_agu_use (regno0, insn);
>
> + /* distance_non_agu_define can call extract_insn_cached. If this function
> + is called from define_split conditions, that can break insn splitting,
> + because split_insns works by clearing recog_data.insn and then modifying
> + recog_data.operand array and match the various split conditions. */
> + if (recog_data.insn != rinsn)
> + recog_data.insn = NULL;
> +
> if (dist_define < 0 || dist_define >= LEA_MAX_STALL)
> {
> /* If there is no non AGU operand definition, no AGU
> --- gcc/testsuite/gcc.target/i386/pr93611.c.jj 2020-02-06 12:24:28.005976435
> +0100
> +++ gcc/testsuite/gcc.target/i386/pr93611.c 2020-02-06 12:24:17.685131826
> +0100
> @@ -0,0 +1,5 @@
> +/* PR target/93611 */
> +/* { dg-do compile } */
> +/* { dg-options "-fira-algorithm=priority -O3 -mtune=bonnell" } */
> +
> +#include "../../gcc.dg/vect/pr58508.c"
>
> Jakub
>
