[PATCH] Fix ICE with vector 16-bit rotate (PR middle-end/83623)

[PATCH] Fix ICE with vector 16-bit rotate (PR middle-end/83623)

[Jakub Jelinek]

Hi!

Ths chunk of code wants to expand 16-bit rotate by 8 as bswaphi, but
if mode is a vector mode (or complex mode), trying to use bswaphi pattern to
swap say V16HImode vector is just wrong.

This patch arranges to use bswapv16hi etc. instead in those cases, which
isn't 100% clear to me from the docs if they only swap bytes of each of the
vector element, but from looking at a few backends it seems like that is the
case and if it wasn't, e.g. simplify-rtx.c would missimplify stuff.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2017-12-30  Jakub Jelinek  <jakub@redhat.com>

	PR middle-end/83623
	* expmed.c (expand_shift_1): For 2-byte rotates by BITS_PER_UNIT,
	check for bswap in mode rather than HImode and use that in expand_unop
	too.

	* gcc.dg/pr83623.c: New test.

--- gcc/expmed.c.jj	2017-12-20 20:40:18.000000000 +0100
+++ gcc/expmed.c	2017-12-30 11:44:47.045433528 +0100
@@ -2490,9 +2490,8 @@ expand_shift_1 (enum tree_code code, mac
       && CONST_INT_P (op1)
       && INTVAL (op1) == BITS_PER_UNIT
       && GET_MODE_SIZE (scalar_mode) == 2
-      && optab_handler (bswap_optab, HImode) != CODE_FOR_nothing)
-    return expand_unop (HImode, bswap_optab, shifted, NULL_RTX,
-				  unsignedp);
+      && optab_handler (bswap_optab, mode) != CODE_FOR_nothing)
+    return expand_unop (mode, bswap_optab, shifted, NULL_RTX, unsignedp);
 
   if (op1 == const0_rtx)
     return shifted;
--- gcc/testsuite/gcc.dg/pr83623.c.jj	2017-12-30 11:49:23.259470538 +0100
+++ gcc/testsuite/gcc.dg/pr83623.c	2017-12-30 11:49:01.309467592 +0100
@@ -0,0 +1,12 @@
+/* PR middle-end/83623 */
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+/* { dg-additional-options "-mmovbe" { target i?86-*-* x86_64-*-* } } */
+
+unsigned short __attribute__ ((__vector_size__ (16))) x;
+
+void
+foo (void)
+{
+  x = x << 8 | x >> 8;
+}

	Jakub

Re: [PATCH] Fix ICE with vector 16-bit rotate (PR middle-end/83623)

[Richard Biener]

On December 30, 2017 8:27:44 PM GMT+01:00, Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!
>
>Ths chunk of code wants to expand 16-bit rotate by 8 as bswaphi, but
>if mode is a vector mode (or complex mode), trying to use bswaphi
>pattern to
>swap say V16HImode vector is just wrong.
>
>This patch arranges to use bswapv16hi etc. instead in those cases,
>which
>isn't 100% clear to me from the docs if they only swap bytes of each of
>the
>vector element, but from looking at a few backends it seems like that
>is the
>case and if it wasn't, e.g. simplify-rtx.c would missimplify stuff.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

OK. 

Richard. 

>2017-12-30  Jakub Jelinek  <jakub@redhat.com>
>
>	PR middle-end/83623
>	* expmed.c (expand_shift_1): For 2-byte rotates by BITS_PER_UNIT,
>	check for bswap in mode rather than HImode and use that in expand_unop
>	too.
>
>	* gcc.dg/pr83623.c: New test.
>
>--- gcc/expmed.c.jj	2017-12-20 20:40:18.000000000 +0100
>+++ gcc/expmed.c	2017-12-30 11:44:47.045433528 +0100
>@@ -2490,9 +2490,8 @@ expand_shift_1 (enum tree_code code, mac
>       && CONST_INT_P (op1)
>       && INTVAL (op1) == BITS_PER_UNIT
>       && GET_MODE_SIZE (scalar_mode) == 2
>-      && optab_handler (bswap_optab, HImode) != CODE_FOR_nothing)
>-    return expand_unop (HImode, bswap_optab, shifted, NULL_RTX,
>-				  unsignedp);
>+      && optab_handler (bswap_optab, mode) != CODE_FOR_nothing)
>+    return expand_unop (mode, bswap_optab, shifted, NULL_RTX,
>unsignedp);
> 
>   if (op1 == const0_rtx)
>     return shifted;
>--- gcc/testsuite/gcc.dg/pr83623.c.jj	2017-12-30 11:49:23.259470538
>+0100
>+++ gcc/testsuite/gcc.dg/pr83623.c	2017-12-30 11:49:01.309467592 +0100
>@@ -0,0 +1,12 @@
>+/* PR middle-end/83623 */
>+/* { dg-do compile } */
>+/* { dg-options "-O2" } */
>+/* { dg-additional-options "-mmovbe" { target i?86-*-* x86_64-*-* } }
>*/
>+
>+unsigned short __attribute__ ((__vector_size__ (16))) x;
>+
>+void
>+foo (void)
>+{
>+  x = x << 8 | x >> 8;
>+}
>
>	Jakub