* [PATCH] Simplify vector compare-not-select sequence
@ 2015-07-06 13:41 Bill Schmidt
2015-07-09 3:05 ` Jeff Law
2015-08-02 10:18 ` Andreas Schwab
0 siblings, 2 replies; 6+ messages in thread
From: Bill Schmidt @ 2015-07-06 13:41 UTC (permalink / raw)
To: gcc-patches; +Cc: ebotcazou, dje.gcc
Hi,
Due to specifics of the POWER architecture, some forms of a vector
compare followed by a vector select are represented in RTL as a compare,
followed by a logical NOT, followed by the select. This tends to end up
generating an extra instruction. This patch adds a case to
simplify-rtx.c to remove the logical NOT by reversing the outcomes of
the select. I've added a POWER-specific test case that demonstrates
that the issue is fixed.
Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no
regressions. Is this ok for trunk?
Thanks,
Bill
[gcc]
2015-07-06 Bill Schmidt <wschmidt@linux.vnet.ibm.com>
* simplify-rtx.c (simplify_ternary_operation): Add simplification
for (!c) != {0,...,0} ? a : b for vector modes.
[gcc/testsuite]
2015-07-06 Bill Schmidt <wschmidt@linux.vnet.ibm.com>
* gcc.target/powerpc/vec-cmp-sel.c: New test.
Index: gcc/simplify-rtx.c
===================================================================
--- gcc/simplify-rtx.c (revision 225440)
+++ gcc/simplify-rtx.c (working copy)
@@ -5251,6 +5251,32 @@ simplify_ternary_operation (enum rtx_code code, ma
&& rtx_equal_p (XEXP (op0, 1), op1))))
return op2;
+ /* Convert (!c) != {0,...,0} ? a : b into
+ c != {0,...,0} ? b : a for vector modes. */
+ if (VECTOR_MODE_P (GET_MODE (op1))
+ && GET_CODE (op0) == NE
+ && GET_CODE (XEXP (op0, 0)) == NOT
+ && GET_CODE (XEXP (op0, 1)) == CONST_VECTOR)
+ {
+ rtx cv = XEXP (op0, 1);
+ int nunits = CONST_VECTOR_NUNITS (cv);
+ bool ok = true;
+ for (int i = 0; i < nunits; ++i)
+ if (CONST_VECTOR_ELT (cv, i) != const0_rtx)
+ {
+ ok = false;
+ break;
+ }
+ if (ok)
+ {
+ rtx new_op0 = gen_rtx_NE (GET_MODE (op0),
+ XEXP (XEXP (op0, 0), 0),
+ XEXP (op0, 1));
+ rtx retval = gen_rtx_IF_THEN_ELSE (mode, new_op0, op2, op1);
+ return retval;
+ }
+ }
+
if (COMPARISON_P (op0) && ! side_effects_p (op0))
{
machine_mode cmp_mode = (GET_MODE (XEXP (op0, 0)) == VOIDmode
Index: gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c
===================================================================
--- gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (revision 0)
+++ gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (working copy)
@@ -0,0 +1,20 @@
+/* { dg-do compile { target powerpc64*-*-* } } */
+/* { dg-require-effective-target powerpc_p8vector_ok } */
+/* { dg-options "-maltivec -O2" } */
+/* { dg-final { scan-assembler "vcmpgtsd" } } */
+/* { dg-final { scan-assembler-not "xxlnor" } } */
+
+/* Test code in simplify-rtx.c that converts
+ (!c) != {0,...,0} ? a : b
+ into
+ c != {0,...,0} ? b : a */
+
+#include <altivec.h>
+
+vector signed long long foo () {
+ vector signed long long x = { 25399, -12900 };
+ vector signed long long y = { 12178, -9987 };
+ vector bool long long b = vec_cmpge (x, y);
+ vector signed long long z = vec_sel (y, x, b);
+ return z;
+}
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] Simplify vector compare-not-select sequence
2015-07-06 13:41 [PATCH] Simplify vector compare-not-select sequence Bill Schmidt
@ 2015-07-09 3:05 ` Jeff Law
2015-08-02 10:18 ` Andreas Schwab
1 sibling, 0 replies; 6+ messages in thread
From: Jeff Law @ 2015-07-09 3:05 UTC (permalink / raw)
To: Bill Schmidt, gcc-patches; +Cc: ebotcazou, dje.gcc
On 07/06/2015 07:41 AM, Bill Schmidt wrote:
> Hi,
>
> Due to specifics of the POWER architecture, some forms of a vector
> compare followed by a vector select are represented in RTL as a compare,
> followed by a logical NOT, followed by the select. This tends to end up
> generating an extra instruction. This patch adds a case to
> simplify-rtx.c to remove the logical NOT by reversing the outcomes of
> the select. I've added a POWER-specific test case that demonstrates
> that the issue is fixed.
>
> Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no
> regressions. Is this ok for trunk?
>
> Thanks,
> Bill
>
>
> [gcc]
>
> 2015-07-06 Bill Schmidt <wschmidt@linux.vnet.ibm.com>
>
> * simplify-rtx.c (simplify_ternary_operation): Add simplification
> for (!c) != {0,...,0} ? a : b for vector modes.
>
> [gcc/testsuite]
>
> 2015-07-06 Bill Schmidt <wschmidt@linux.vnet.ibm.com>
>
> * gcc.target/powerpc/vec-cmp-sel.c: New test.
OK.
jeff
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] Simplify vector compare-not-select sequence
2015-07-06 13:41 [PATCH] Simplify vector compare-not-select sequence Bill Schmidt
2015-07-09 3:05 ` Jeff Law
@ 2015-08-02 10:18 ` Andreas Schwab
2015-08-03 15:16 ` Bill Schmidt
1 sibling, 1 reply; 6+ messages in thread
From: Andreas Schwab @ 2015-08-02 10:18 UTC (permalink / raw)
To: Bill Schmidt; +Cc: gcc-patches, ebotcazou, dje.gcc
Bill Schmidt <wschmidt@linux.vnet.ibm.com> writes:
> * gcc.target/powerpc/vec-cmp-sel.c: New test.
FAIL: gcc.target/powerpc/vec-cmp-sel.c (test for excess errors)
Excess errors:
/daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:14:1: error: use of 'long long' in AltiVec types is invalid without -mvsx
/daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:15:3: error: use of 'long long' in AltiVec types is invalid without -mvsx
/daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:16:3: error: use of 'long long' in AltiVec types is invalid without -mvsx
/daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:17:3: error: use of 'long long' in AltiVec types is invalid without -mvsx
/daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:17:29: error: incompatible types when initializing type '__vector unsigned long long' using type '__vector __bool long long'
/daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:18:3: error: use of 'long long' in AltiVec types is invalid without -mvsx
Andreas.
--
Andreas Schwab, schwab@linux-m68k.org
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756 01D3 44D5 214B 8276 4ED5
"And now for something completely different."
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] Simplify vector compare-not-select sequence
2015-08-02 10:18 ` Andreas Schwab
@ 2015-08-03 15:16 ` Bill Schmidt
2015-08-03 17:35 ` Andreas Schwab
0 siblings, 1 reply; 6+ messages in thread
From: Bill Schmidt @ 2015-08-03 15:16 UTC (permalink / raw)
To: Andreas Schwab; +Cc: gcc-patches, ebotcazou, dje.gcc
On Sun, 2015-08-02 at 12:18 +0200, Andreas Schwab wrote:
> Bill Schmidt <wschmidt@linux.vnet.ibm.com> writes:
>
> > * gcc.target/powerpc/vec-cmp-sel.c: New test.
>
> FAIL: gcc.target/powerpc/vec-cmp-sel.c (test for excess errors)
> Excess errors:
> /daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:14:1: error: use of 'long long' in AltiVec types is invalid without -mvsx
> /daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:15:3: error: use of 'long long' in AltiVec types is invalid without -mvsx
> /daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:16:3: error: use of 'long long' in AltiVec types is invalid without -mvsx
> /daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:17:3: error: use of 'long long' in AltiVec types is invalid without -mvsx
> /daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:17:29: error: incompatible types when initializing type '__vector unsigned long long' using type '__vector __bool long long'
> /daten/gcc/gcc-20150801/gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c:18:3: error: use of 'long long' in AltiVec types is invalid without -mvsx
>
> Andreas.
>
Hi Andreas,
Can you please verify that this patch works for you?
Thanks,
Bill
Index: gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c
===================================================================
--- gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (revision 226505)
+++ gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (working copy)
@@ -1,6 +1,7 @@
/* { dg-do compile { target powerpc64*-*-* } } */
/* { dg-require-effective-target powerpc_p8vector_ok } */
-/* { dg-options "-maltivec -O2" } */
+/* { dg-require-effective-target powerpc_vsx_ok } */
+/* { dg-options "-maltivec -O2 -mvsx -mpower8-vector" } */
/* { dg-final { scan-assembler "vcmpgtsd" } } */
/* { dg-final { scan-assembler-not "xxlnor" } } */
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] Simplify vector compare-not-select sequence
2015-08-03 15:16 ` Bill Schmidt
@ 2015-08-03 17:35 ` Andreas Schwab
2015-08-04 14:12 ` Bill Schmidt
0 siblings, 1 reply; 6+ messages in thread
From: Andreas Schwab @ 2015-08-03 17:35 UTC (permalink / raw)
To: Bill Schmidt; +Cc: gcc-patches, ebotcazou, dje.gcc
Bill Schmidt <wschmidt@linux.vnet.ibm.com> writes:
> Index: gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c
> ===================================================================
> --- gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (revision 226505)
> +++ gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (working copy)
> @@ -1,6 +1,7 @@
> /* { dg-do compile { target powerpc64*-*-* } } */
If you want -m64 you need dg-require-effective-target lp64, but I see no
need for that.
> /* { dg-require-effective-target powerpc_p8vector_ok } */
> -/* { dg-options "-maltivec -O2" } */
> +/* { dg-require-effective-target powerpc_vsx_ok } */
> +/* { dg-options "-maltivec -O2 -mvsx -mpower8-vector" } */
> /* { dg-final { scan-assembler "vcmpgtsd" } } */
> /* { dg-final { scan-assembler-not "xxlnor" } } */
Looks good.
PASS: gcc.target/powerpc/vec-cmp-sel.c (test for excess errors)
PASS: gcc.target/powerpc/vec-cmp-sel.c scan-assembler vcmpgtsd
PASS: gcc.target/powerpc/vec-cmp-sel.c scan-assembler-not xxlnor
PASS: gcc.target/powerpc/vec-cmp-sel.c (test for excess errors)
PASS: gcc.target/powerpc/vec-cmp-sel.c scan-assembler vcmpgtsd
PASS: gcc.target/powerpc/vec-cmp-sel.c scan-assembler-not xxlnor
Andreas.
--
Andreas Schwab, schwab@linux-m68k.org
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756 01D3 44D5 214B 8276 4ED5
"And now for something completely different."
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] Simplify vector compare-not-select sequence
2015-08-03 17:35 ` Andreas Schwab
@ 2015-08-04 14:12 ` Bill Schmidt
0 siblings, 0 replies; 6+ messages in thread
From: Bill Schmidt @ 2015-08-04 14:12 UTC (permalink / raw)
To: Andreas Schwab; +Cc: gcc-patches, ebotcazou, dje.gcc
Thanks for verifying!
Also verified it still works on powerpc64le-unknown-linux-gnu.
Committed as obvious.
Thanks,
Bill
2015-08-04 Bill Schmidt <wschmidt@vnet.linux.ibm.com>
* gcc.target/powerpc/vec-cmp-sel.c: Avoid test failure on machines
without VSX an Power8 vector support.
Index: gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c
===================================================================
--- gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (revision 226505)
+++ gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (working copy)
@@ -1,6 +1,7 @@
/* { dg-do compile { target powerpc64*-*-* } } */
/* { dg-require-effective-target powerpc_p8vector_ok } */
-/* { dg-options "-maltivec -O2" } */
+/* { dg-require-effective-target powerpc_vsx_ok } */
+/* { dg-options "-maltivec -O2 -mvsx -mpower8-vector" } */
/* { dg-final { scan-assembler "vcmpgtsd" } } */
/* { dg-final { scan-assembler-not "xxlnor" } } */
On Mon, 2015-08-03 at 19:35 +0200, Andreas Schwab wrote:
> Bill Schmidt <wschmidt@linux.vnet.ibm.com> writes:
>
> > Index: gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c
> > ===================================================================
> > --- gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (revision 226505)
> > +++ gcc/testsuite/gcc.target/powerpc/vec-cmp-sel.c (working copy)
> > @@ -1,6 +1,7 @@
> > /* { dg-do compile { target powerpc64*-*-* } } */
>
> If you want -m64 you need dg-require-effective-target lp64, but I see no
> need for that.
>
> > /* { dg-require-effective-target powerpc_p8vector_ok } */
> > -/* { dg-options "-maltivec -O2" } */
> > +/* { dg-require-effective-target powerpc_vsx_ok } */
> > +/* { dg-options "-maltivec -O2 -mvsx -mpower8-vector" } */
> > /* { dg-final { scan-assembler "vcmpgtsd" } } */
> > /* { dg-final { scan-assembler-not "xxlnor" } } */
>
> Looks good.
>
> PASS: gcc.target/powerpc/vec-cmp-sel.c (test for excess errors)
> PASS: gcc.target/powerpc/vec-cmp-sel.c scan-assembler vcmpgtsd
> PASS: gcc.target/powerpc/vec-cmp-sel.c scan-assembler-not xxlnor
> PASS: gcc.target/powerpc/vec-cmp-sel.c (test for excess errors)
> PASS: gcc.target/powerpc/vec-cmp-sel.c scan-assembler vcmpgtsd
> PASS: gcc.target/powerpc/vec-cmp-sel.c scan-assembler-not xxlnor
>
> Andreas.
>
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2015-08-04 14:12 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2015-07-06 13:41 [PATCH] Simplify vector compare-not-select sequence Bill Schmidt
2015-07-09 3:05 ` Jeff Law
2015-08-02 10:18 ` Andreas Schwab
2015-08-03 15:16 ` Bill Schmidt
2015-08-03 17:35 ` Andreas Schwab
2015-08-04 14:12 ` Bill Schmidt
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).