* [PATCH] i386: Fix up V2DI and V1TI inequality comparisons [PR105613]
@ 2022-05-17 7:00 Jakub Jelinek
2022-05-17 10:05 ` Uros Bizjak
0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2022-05-17 7:00 UTC (permalink / raw)
To: Uros Bizjak; +Cc: gcc-patches
Hi!
The recent r13-458 change to introduce vec_cmpeqv1tiv1ti and
add TARGET_SSE2 support to vec_cmpeqv2div2di works nicely for
equality comparisons, but as the testcase shows doesn't work
for inequality comparisons.
For EQ if we perform comparison with twice as many half-sized elemenets,
the result should be ~0 when both halves are ~0 only (both halves need
to be equal for the whole to be equal), otherwise 0, so AND is the
correct operation for it.
But for NE, the result should be ~0 when either of the halves is ~0
(if either half is not equal, the whole is not equal) and so the right
operation for NE is IOR, not AND.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2022-05-17 Jakub Jelinek <jakub@redhat.com>
PR target/105613
* config/i386/sse.md (vec_cmpeqv2div2di, vec_cmpeqv1tiv1ti): Use
andv4si3 only for EQ, for NE use iorv4si3 instead.
* gcc.c-torture/execute/pr105613.c: New test.
--- gcc/config/i386/sse.md.jj 2022-05-16 09:46:01.962065216 +0200
+++ gcc/config/i386/sse.md 2022-05-16 10:48:45.698038881 +0200
@@ -4407,7 +4407,10 @@ (define_expand "vec_cmpeqv2div2di"
emit_insn (gen_sse2_pshufd (tmp1, ops[0], GEN_INT (0xb1)));
rtx tmp2 = gen_reg_rtx (V4SImode);
- emit_insn (gen_andv4si3 (tmp2, tmp1, ops[0]));
+ if (GET_CODE (operands[1]) == EQ)
+ emit_insn (gen_andv4si3 (tmp2, tmp1, ops[0]));
+ else
+ emit_insn (gen_iorv4si3 (tmp2, tmp1, ops[0]));
emit_move_insn (operands[0], gen_lowpart (V2DImode, tmp2));
}
@@ -4435,7 +4438,10 @@ (define_expand "vec_cmpeqv1tiv1ti"
emit_insn (gen_sse2_pshufd (tmp1, tmp2, GEN_INT (0x4e)));
rtx tmp3 = gen_reg_rtx (V4SImode);
- emit_insn (gen_andv4si3 (tmp3, tmp2, tmp1));
+ if (GET_CODE (operands[1]) == EQ)
+ emit_insn (gen_andv4si3 (tmp3, tmp2, tmp1));
+ else
+ emit_insn (gen_iorv4si3 (tmp3, tmp2, tmp1));
emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp3));
DONE;
--- gcc/testsuite/gcc.c-torture/execute/pr105613.c.jj 2022-05-16 10:42:34.286151601 +0200
+++ gcc/testsuite/gcc.c-torture/execute/pr105613.c 2022-05-16 10:48:07.687562119 +0200
@@ -0,0 +1,26 @@
+/* PR target/105613 */
+/* { dg-do run { target int128 } } */
+
+typedef unsigned __int128 __attribute__((__vector_size__ (16))) V;
+
+void
+foo (V v, V *r)
+{
+ *r = v != 0;
+}
+
+int
+main ()
+{
+ V r;
+ foo ((V) {5}, &r);
+ if (r[0] != ~(unsigned __int128) 0)
+ __builtin_abort ();
+ foo ((V) {0x500000005ULL}, &r);
+ if (r[0] != ~(unsigned __int128) 0)
+ __builtin_abort ();
+ foo ((V) {0}, &r);
+ if (r[0] != 0)
+ __builtin_abort ();
+ return 0;
+}
Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] i386: Fix up V2DI and V1TI inequality comparisons [PR105613]
2022-05-17 7:00 [PATCH] i386: Fix up V2DI and V1TI inequality comparisons [PR105613] Jakub Jelinek
@ 2022-05-17 10:05 ` Uros Bizjak
0 siblings, 0 replies; 2+ messages in thread
From: Uros Bizjak @ 2022-05-17 10:05 UTC (permalink / raw)
To: Jakub Jelinek; +Cc: gcc-patches
On Tue, May 17, 2022 at 9:00 AM Jakub Jelinek <jakub@redhat.com> wrote:
>
> Hi!
>
> The recent r13-458 change to introduce vec_cmpeqv1tiv1ti and
> add TARGET_SSE2 support to vec_cmpeqv2div2di works nicely for
> equality comparisons, but as the testcase shows doesn't work
> for inequality comparisons.
> For EQ if we perform comparison with twice as many half-sized elemenets,
> the result should be ~0 when both halves are ~0 only (both halves need
> to be equal for the whole to be equal), otherwise 0, so AND is the
> correct operation for it.
> But for NE, the result should be ~0 when either of the halves is ~0
> (if either half is not equal, the whole is not equal) and so the right
> operation for NE is IOR, not AND.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2022-05-17 Jakub Jelinek <jakub@redhat.com>
>
> PR target/105613
> * config/i386/sse.md (vec_cmpeqv2div2di, vec_cmpeqv1tiv1ti): Use
> andv4si3 only for EQ, for NE use iorv4si3 instead.
>
> * gcc.c-torture/execute/pr105613.c: New test.
OK.
Thanks,
Uros.
>
> --- gcc/config/i386/sse.md.jj 2022-05-16 09:46:01.962065216 +0200
> +++ gcc/config/i386/sse.md 2022-05-16 10:48:45.698038881 +0200
> @@ -4407,7 +4407,10 @@ (define_expand "vec_cmpeqv2div2di"
> emit_insn (gen_sse2_pshufd (tmp1, ops[0], GEN_INT (0xb1)));
>
> rtx tmp2 = gen_reg_rtx (V4SImode);
> - emit_insn (gen_andv4si3 (tmp2, tmp1, ops[0]));
> + if (GET_CODE (operands[1]) == EQ)
> + emit_insn (gen_andv4si3 (tmp2, tmp1, ops[0]));
> + else
> + emit_insn (gen_iorv4si3 (tmp2, tmp1, ops[0]));
>
> emit_move_insn (operands[0], gen_lowpart (V2DImode, tmp2));
> }
> @@ -4435,7 +4438,10 @@ (define_expand "vec_cmpeqv1tiv1ti"
> emit_insn (gen_sse2_pshufd (tmp1, tmp2, GEN_INT (0x4e)));
>
> rtx tmp3 = gen_reg_rtx (V4SImode);
> - emit_insn (gen_andv4si3 (tmp3, tmp2, tmp1));
> + if (GET_CODE (operands[1]) == EQ)
> + emit_insn (gen_andv4si3 (tmp3, tmp2, tmp1));
> + else
> + emit_insn (gen_iorv4si3 (tmp3, tmp2, tmp1));
>
> emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp3));
> DONE;
> --- gcc/testsuite/gcc.c-torture/execute/pr105613.c.jj 2022-05-16 10:42:34.286151601 +0200
> +++ gcc/testsuite/gcc.c-torture/execute/pr105613.c 2022-05-16 10:48:07.687562119 +0200
> @@ -0,0 +1,26 @@
> +/* PR target/105613 */
> +/* { dg-do run { target int128 } } */
> +
> +typedef unsigned __int128 __attribute__((__vector_size__ (16))) V;
> +
> +void
> +foo (V v, V *r)
> +{
> + *r = v != 0;
> +}
> +
> +int
> +main ()
> +{
> + V r;
> + foo ((V) {5}, &r);
> + if (r[0] != ~(unsigned __int128) 0)
> + __builtin_abort ();
> + foo ((V) {0x500000005ULL}, &r);
> + if (r[0] != ~(unsigned __int128) 0)
> + __builtin_abort ();
> + foo ((V) {0}, &r);
> + if (r[0] != 0)
> + __builtin_abort ();
> + return 0;
> +}
>
> Jakub
>
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2022-05-17 10:05 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-05-17 7:00 [PATCH] i386: Fix up V2DI and V1TI inequality comparisons [PR105613] Jakub Jelinek
2022-05-17 10:05 ` Uros Bizjak
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).