Re: [PATCH] Improve *avx_vperm_broadcast_*

[Jakub Jelinek]

On Tue, May 31, 2016 at 06:54:14AM -0700, H.J. Lu wrote:
> On Mon, May 23, 2016 at 10:15 AM, Jakub Jelinek  wrote:
> > Hi!
> >
> > The vbroadcastss and vpermilps insns are already in AVX512F & AVX512VL,
> > so can be used with v instead of x, the splitter case where we for AVX
> > emit vpermilps plus vpermf128 is more problematic, because the latter
> > insn isn't available in EVEX.  But, we can get the same effect with
> > vshuff32x4 when both source operands are the same.
> > Alternatively, we could replace the vpermilps and vshuff32x4 insns
> > with the AVX512VL arbitrary permutations I think, the question is
> > what is faster, because we'd need to load the mask from memory.
> >
> > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
> >
> > 2016-05-23  Jakub Jelinek  
> >
> > * config/i386/sse.md
> > (avx512vl_shuf_32x4_1): Rename
> > to ...
> > (avx512vl_shuf_32x4_1): ... this.
> > (*avx_vperm_broadcast_v4sf): Use v constraint instead of x.  Use
> > maybe_evex prefix instead of vex.
> > (*avx_vperm_broadcast_): Use v constraint instead of x.  
> > Handle
> > EXT_REX_SSE_REG_P (op0) case in the splitter.
> >
> > * gcc.target/i386/avx512vl-vbroadcast-3.c: New test.
> >
> 
> The new test fails on x32 due to 32-bit register in address.  This
> patch fixes it.  Tested on x86-64.  OK for trunk?

Ok, thanks.
> 2016-05-31  H.J. Lu  
> 
> * gcc.target/i386/avx512vl-vbroadcast-3.c: Scan %\[re\]di
> instead of %rdi.
> * gcc.target/i386/avx512vl-vcvtps2ph-3.c: Likewise.

Jakub

Re: [PATCH] Improve *avx_vperm_broadcast_*

[H.J. Lu]

On Mon, May 23, 2016 at 10:15 AM, Jakub Jelinek  wrote:
> Hi!
>
> The vbroadcastss and vpermilps insns are already in AVX512F & AVX512VL,
> so can be used with v instead of x, the splitter case where we for AVX
> emit vpermilps plus vpermf128 is more problematic, because the latter
> insn isn't available in EVEX.  But, we can get the same effect with
> vshuff32x4 when both source operands are the same.
> Alternatively, we could replace the vpermilps and vshuff32x4 insns
> with the AVX512VL arbitrary permutations I think, the question is
> what is faster, because we'd need to load the mask from memory.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2016-05-23  Jakub Jelinek  
>
> * config/i386/sse.md
> (avx512vl_shuf_32x4_1): Rename
> to ...
> (avx512vl_shuf_32x4_1): ... this.
> (*avx_vperm_broadcast_v4sf): Use v constraint instead of x.  Use
> maybe_evex prefix instead of vex.
> (*avx_vperm_broadcast_): Use v constraint instead of x.  Handle
> EXT_REX_SSE_REG_P (op0) case in the splitter.
>
> * gcc.target/i386/avx512vl-vbroadcast-3.c: New test.
>

The new test fails on x32 due to 32-bit register in address.  This
patch fixes it.  Tested on x86-64.  OK for trunk?

Thanks.

H.J.

2016-05-31  H.J. Lu  

* gcc.target/i386/avx512vl-vbroadcast-3.c: Scan %\[re\]di
instead of %rdi.
* gcc.target/i386/avx512vl-vcvtps2ph-3.c: Likewise.

diff --git a/gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-3.c
b/gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-3.c
index d981fe4..7233398 100644
--- a/gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-3.c
+++ b/gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-3.c
@@ -150,9 +150,9 @@ f16 (V2 *x)
   asm volatile ("" : "+v" (a));
 }

-/* { dg-final { scan-assembler-times
"vbroadcastss\[^\n\r]*%rdi\[^\n\r]*%xmm16" 4 } } */
+/* { dg-final { scan-assembler-times
"vbroadcastss\[^\n\r]*%\[re\]di\[^\n\r]*%xmm16" 4 } } */
 /* { dg-final { scan-assembler-times
"vbroadcastss\[^\n\r]*%xmm16\[^\n\r]*%ymm16" 3 } } */
-/* { dg-final { scan-assembler-times
"vbroadcastss\[^\n\r]*%rdi\[^\n\r]*%ymm16" 3 } } */
+/* { dg-final { scan-assembler-times
"vbroadcastss\[^\n\r]*%\[re\]di\[^\n\r]*%ymm16" 3 } } */
 /* { dg-final { scan-assembler-times
"vpermilps\[^\n\r]*\\\$0\[^\n\r]*%xmm16\[^\n\r]*%xmm16" 1 } } */
 /* { dg-final { scan-assembler-times
"vpermilps\[^\n\r]*\\\$85\[^\n\r]*%xmm16\[^\n\r]*%xmm16" 1 } } */
 /* { dg-final { scan-assembler-times
"vpermilps\[^\n\r]*\\\$170\[^\n\r]*%xmm16\[^\n\r]*%xmm16" 1 } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx512vl-vcvtps2ph-3.c
b/gcc/testsuite/gcc.target/i386/avx512vl-vcvtps2ph-3.c
index 2fd2215..c2e3f01 100644
--- a/gcc/testsuite/gcc.target/i386/avx512vl-vcvtps2ph-3.c
+++ b/gcc/testsuite/gcc.target/i386/avx512vl-vcvtps2ph-3.c
@@ -38,4 +38,4 @@ f3 (__m256 x, __v8hi *y)
   *y = (__v8hi) _mm256_cvtps_ph (a, 1);
 }

-/* { dg-final { scan-assembler
"vcvtps2ph\[^\n\r]*\\\$1\[^\n\r]*%ymm16\[^\n\r]*%rdi" } } */
+/* { dg-final { scan-assembler
"vcvtps2ph\[^\n\r]*\\\$1\[^\n\r]*%ymm16\[^\n\r]*%\[re\]di" } } */

Re: [PATCH] Improve *avx_vperm_broadcast_*

[Kirill Yukhin]

Hi Jakub,
On 23 May 19:15, Jakub Jelinek wrote:
> Hi!
> 
> The vbroadcastss and vpermilps insns are already in AVX512F & AVX512VL,
> so can be used with v instead of x, the splitter case where we for AVX
> emit vpermilps plus vpermf128 is more problematic, because the latter
> insn isn't available in EVEX.  But, we can get the same effect with
> vshuff32x4 when both source operands are the same.
> Alternatively, we could replace the vpermilps and vshuff32x4 insns
> with the AVX512VL arbitrary permutations I think, the question is
> what is faster, because we'd need to load the mask from memory.
> 
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
patch is OK.

--
Thanks, K

[PATCH] Improve *avx_vperm_broadcast_*

[Jakub Jelinek]

Hi!

The vbroadcastss and vpermilps insns are already in AVX512F & AVX512VL,
so can be used with v instead of x, the splitter case where we for AVX
emit vpermilps plus vpermf128 is more problematic, because the latter
insn isn't available in EVEX.  But, we can get the same effect with
vshuff32x4 when both source operands are the same.
Alternatively, we could replace the vpermilps and vshuff32x4 insns
with the AVX512VL arbitrary permutations I think, the question is
what is faster, because we'd need to load the mask from memory.

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

2016-05-23  Jakub Jelinek  

* config/i386/sse.md
(avx512vl_shuf_32x4_1): Rename
to ...
(avx512vl_shuf_32x4_1): ... this.
(*avx_vperm_broadcast_v4sf): Use v constraint instead of x.  Use
maybe_evex prefix instead of vex.
(*avx_vperm_broadcast_): Use v constraint instead of x.  Handle
EXT_REX_SSE_REG_P (op0) case in the splitter.

* gcc.target/i386/avx512vl-vbroadcast-3.c: New test.

--- gcc/config/i386/sse.md.jj   2016-05-22 12:27:34.0 +0200
+++ gcc/config/i386/sse.md  2016-05-23 13:54:22.211998751 +0200
@@ -12380,7 +12380,7 @@
   DONE;
 })
 
-(define_insn "avx512vl_shuf_32x4_1"
+(define_insn "avx512vl_shuf_32x4_1"
   [(set (match_operand:VI4F_256 0 "register_operand" "=v")
(vec_select:VI4F_256
  (vec_concat:
@@ -17247,9 +17247,9 @@
 ;; If it so happens that the input is in memory, use vbroadcast.
 ;; Otherwise use vpermilp (and in the case of 256-bit modes, vperm2f128).
 (define_insn "*avx_vperm_broadcast_v4sf"
-  [(set (match_operand:V4SF 0 "register_operand" "=x,x,x")
+  [(set (match_operand:V4SF 0 "register_operand" "=v,v,v")
(vec_select:V4SF
- (match_operand:V4SF 1 "nonimmediate_operand" "m,o,x")
+ (match_operand:V4SF 1 "nonimmediate_operand" "m,o,v")
  (match_parallel 2 "avx_vbroadcast_operand"
[(match_operand 3 "const_int_operand" "C,n,n")])))]
   "TARGET_AVX"
@@ -17271,13 +17271,13 @@
   [(set_attr "type" "ssemov,ssemov,sselog1")
(set_attr "prefix_extra" "1")
(set_attr "length_immediate" "0,0,1")
-   (set_attr "prefix" "vex")
+   (set_attr "prefix" "maybe_evex")
(set_attr "mode" "SF,SF,V4SF")])
 
 (define_insn_and_split "*avx_vperm_broadcast_"
-  [(set (match_operand:VF_256 0 "register_operand" "=x,x,x")
+  [(set (match_operand:VF_256 0 "register_operand" "=v,v,v")
(vec_select:VF_256
- (match_operand:VF_256 1 "nonimmediate_operand" "m,o,?x")
+ (match_operand:VF_256 1 "nonimmediate_operand" "m,o,?v")
  (match_parallel 2 "avx_vbroadcast_operand"
[(match_operand 3 "const_int_operand" "C,n,n")])))]
   "TARGET_AVX"
@@ -17309,6 +17309,23 @@
 
   /* Shuffle the lane we care about into both lanes of the dest.  */
   mask = (elt / ( / 2)) * 0x11;
+  if (EXT_REX_SSE_REG_P (op0))
+   {
+ /* There is no EVEX VPERM2F128, but we can use either VBROADCASTSS
+or VSHUFF128.  */
+ gcc_assert (mode == V8SFmode);
+ if ((mask & 1) == 0)
+   emit_insn (gen_avx2_vec_dupv8sf (op0,
+gen_lowpart (V4SFmode, op0)));
+ else
+   emit_insn (gen_avx512vl_shuf_f32x4_1 (op0, op0, op0,
+ GEN_INT (4), GEN_INT (5),
+ GEN_INT (6), GEN_INT (7),
+ GEN_INT (12), GEN_INT (13),
+ GEN_INT (14), GEN_INT (15)));
+ DONE;
+   }
+
   emit_insn (gen_avx_vperm2f1283 (op0, op0, op0, GEN_INT (mask)));
   DONE;
 }
--- gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-3.c.jj2016-05-23 
14:07:36.266695992 +0200
+++ gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-3.c   2016-05-23 
14:14:49.495012459 +0200
@@ -0,0 +1,162 @@
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2 -mavx512vl -masm=att" } */
+
+typedef float V1 __attribute__((vector_size (16)));
+typedef float V2 __attribute__((vector_size (32)));
+typedef int V4 __attribute__((vector_size (16)));
+typedef int V5 __attribute__((vector_size (32)));
+
+void
+f1 (V1 x)
+{
+  register V1 a __asm ("xmm16");
+  a = x;
+  asm volatile ("" : "+v" (a));
+  a = __builtin_shuffle (a, (V4) { 0, 0, 0, 0 });
+  asm volatile ("" : "+v" (a));
+}
+
+void
+f2 (V1 x)
+{
+  register V1 a __asm ("xmm16");
+  a = x;
+  asm volatile ("" : "+v" (a));
+  a = __builtin_shuffle (a, (V4) { 1, 1, 1, 1 });
+  asm volatile ("" : "+v" (a));
+}
+
+void
+f3 (V1 x)
+{
+  register V1 a __asm ("xmm16");
+  a = x;
+  asm volatile ("" : "+v" (a));
+  a = __builtin_shuffle (a, (V4) { 2, 2, 2, 2 });
+  asm volatile ("" : "+v" (a));
+}
+
+void
+f4 (V1 x)
+{
+  register V1 a __asm ("xmm16");
+  a = x;
+  asm volatile ("" : "+v" (a));
+  a = __builtin_shuffle (a, (V4) { 3,