On Tue, Aug 16, 2022 at 09:14:06AM +0100, Richard Sandiford via Gcc-patches wrote:
> IMO the correct low-effort fix is to save and restore recog_data
> in ix86_vector_duplicate_value. It's a relatively big copy,
> but the current code is pretty wasteful anyway (allocating at
> least a new SET and INSN for every query). Compared to the
> overhead of doing that, a copy to and from the stack shouldn't
> be too bad.
The following patch does that.
It isn't the first spot in the compiler that does that, not even the first
spot in the i386 backend.
In i386-expand.cc beyond these 2 recog_memoized calls there is one in
expand_vselect, but I think it is unlikely we'd run into these issues trying
to expand new permutations from splitters.
2022-12-02 Jakub Jelinek <jakub@redhat.com>
PR target/106577
* config/i386/i386-expand.cc (ix86_vector_duplicate_value): Save/restore
recog_data around recog_memoized calls.
* gcc.target/i386/pr106577.c: New test.
bool ok;
rtx_insn *insn;
rtx dup;
+ /* Save/restore recog_data in case this is called from splitters
+ or other routines where recog_data needs to stay valid across
+ force_reg. See PR106577. */
+ recog_data_d recog_data_save = recog_data;
/* First attempt to recognize VAL as-is. */
dup = gen_vec_duplicate (mode, val);
ok = recog_memoized (insn) >= 0;
gcc_assert (ok);
}
+ recog_data = recog_data_save;
return true;
}
--- /dev/null
+/* PR target/106577 */
+/* { dg-do compile { target int128 } } */
+/* { dg-options "-O2 -mavx" } */
+
+int i;
+void
+foo (void)
+{
+ i ^= !(((unsigned __int128)0xf0f0f0f0f0f0f0f0 << 64 | 0xf0f0f0f0f0f0f0f0) & i);
+}