diff mbox

[i386,AVX512,75/n] Update vec_init.

Message ID 20141009121324.GL25028@msticlxl57.ims.intel.com
State New
Headers show

Commit Message

Kirill Yukhin Oct. 9, 2014, 12:13 p.m. UTC 
Hello,
This patch extends vec_init-related routines/patterns.

Bootstrapped.
AVX-512* tests on top of patch-set all pass
under simulator.

Is it ok for trunk?

gcc/
	* config/i386/i386.c
	(ix86_expand_vector_init_duplicate): Handle V64QI and V32HI modes,
	update V8HI, V16QI, V32QI modes handling.
	(ix86_expand_vector_init_general): Handle V64QI and V32HI modes.
	* config/i386/sse.md
	(define_mode_iterator VI48F_512): Rename to ...
	(define_mode_iterator VI48F_I12_AVX512BW): ... this. Extend to
	AVX-512BW modes.
	(define_expand "vec_init<mode>"): Use VI48F_I12_AVX512BW.

--
Thanks, K

Comments

Uros Bizjak Oct. 9, 2014, 3:53 p.m. UTC | #1 
On Thu, Oct 9, 2014 at 2:13 PM, Kirill Yukhin <kirill.yukhin@gmail.com> wrote:
> Hello,
> This patch extends vec_init-related routines/patterns.
>
> Bootstrapped.
> AVX-512* tests on top of patch-set all pass
> under simulator.
>
> Is it ok for trunk?
>
> gcc/
>         * config/i386/i386.c
>         (ix86_expand_vector_init_duplicate): Handle V64QI and V32HI modes,
>         update V8HI, V16QI, V32QI modes handling.
>         (ix86_expand_vector_init_general): Handle V64QI and V32HI modes.
>         * config/i386/sse.md
>         (define_mode_iterator VI48F_512): Rename to ...
>         (define_mode_iterator VI48F_I12_AVX512BW): ... this. Extend to
>         AVX-512BW modes.
>         (define_expand "vec_init<mode>"): Use VI48F_I12_AVX512BW.

LGTM, but I'd like to ask Jakub for his opinion on vec_init stuff.

Uros.
Jakub Jelinek Oct. 15, 2014, 4:23 p.m. UTC | #2 
On Thu, Oct 09, 2014 at 04:13:25PM +0400, Kirill Yukhin wrote:
> --- a/gcc/config/i386/i386.c
> +++ b/gcc/config/i386/i386.c
> @@ -39791,6 +39791,8 @@ ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
>      case V8SFmode:
>      case V8SImode:
>      case V2DFmode:
> +    case V64QImode:
> +    case V32HImode:
>      case V2DImode:
>      case V4SFmode:
>      case V4SImode:
> @@ -39821,6 +39823,9 @@ ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
>        goto widen;
>  
>      case V8HImode:
> +      if (TARGET_AVX512VL)
> +        return ix86_vector_duplicate_value (mode, target, val);
> +

Shouldn't that be TARGET_AVX512VL && TARGET_AVX512BW ?

>        if (TARGET_SSE2)
>  	{
>  	  struct expand_vec_perm_d dperm;
> @@ -39851,6 +39856,9 @@ ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
>        goto widen;
>  
>      case V16QImode:
> +      if (TARGET_AVX512VL)
> +        return ix86_vector_duplicate_value (mode, target, val);
> +

Ditto.

>        if (TARGET_SSE2)
>  	goto permute;
>        goto widen;
> @@ -39880,16 +39888,19 @@ ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
>  
>      case V16HImode:
>      case V32QImode:
> -      {
> -	enum machine_mode hvmode = (mode == V16HImode ? V8HImode : V16QImode);
> -	rtx x = gen_reg_rtx (hvmode);
> +      if (TARGET_AVX512VL)
> +        return ix86_vector_duplicate_value (mode, target, val);

Ditto.

> @@ -40503,6 +40515,42 @@ half:
>  			      gen_rtx_VEC_CONCAT (mode, op0, op1)));
>        return;
>  
> +    case V64QImode:
> +      quarter_mode = V16QImode;
> +      half_mode = V32QImode;
> +      goto quarter;
> +
> +    case V32HImode:
> +      quarter_mode = V8HImode;
> +      half_mode = V16HImode;
> +      goto quarter;

I wonder whether for these modes it can ever be beneficial to build them
through interleaves/concatenations etc., if it wouldn't be better to build
them by storing all values into memory and just reading it back.

> -(define_mode_iterator VI48F_512 [V16SI V16SF V8DI V8DF])
> +(define_mode_iterator VI48F_I12_AVX512BW
> +  [V16SI V16SF V8DI V8DF
> +  (V32HI "TARGET_AVX512BW") (V64QI "TARGET_AVX512BW")])

What does the I12 stand for?  Wasn't it meant to be VI48F_512_AVX512BW
or I512?

>  (define_mode_iterator VI48F
>    [V16SI V16SF V8DI V8DF
>     (V8SI "TARGET_AVX512VL") (V8SF "TARGET_AVX512VL")
> @@ -17472,7 +17474,7 @@
>  })
>  
>  (define_expand "vec_init<mode>"
> -  [(match_operand:VI48F_512 0 "register_operand")
> +  [(match_operand:VI48F_I12_AVX512BW 0 "register_operand")
>     (match_operand 1)]
>    "TARGET_AVX512F"
>  {

	Jakub
diff mbox

Patch

diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index e79210c..7c34431 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -39791,6 +39791,8 @@  ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
     case V8SFmode:
     case V8SImode:
     case V2DFmode:
+    case V64QImode:
+    case V32HImode:
     case V2DImode:
     case V4SFmode:
     case V4SImode:
@@ -39821,6 +39823,9 @@  ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
       goto widen;
 
     case V8HImode:
+      if (TARGET_AVX512VL)
+        return ix86_vector_duplicate_value (mode, target, val);
+
       if (TARGET_SSE2)
 	{
 	  struct expand_vec_perm_d dperm;
@@ -39851,6 +39856,9 @@  ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
       goto widen;
 
     case V16QImode:
+      if (TARGET_AVX512VL)
+        return ix86_vector_duplicate_value (mode, target, val);
+
       if (TARGET_SSE2)
 	goto permute;
       goto widen;
@@ -39880,16 +39888,19 @@  ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
 
     case V16HImode:
     case V32QImode:
-      {
-	enum machine_mode hvmode = (mode == V16HImode ? V8HImode : V16QImode);
-	rtx x = gen_reg_rtx (hvmode);
+      if (TARGET_AVX512VL)
+        return ix86_vector_duplicate_value (mode, target, val);
+      else
+	{
+	  enum machine_mode hvmode = (mode == V16HImode ? V8HImode : V16QImode);
+	  rtx x = gen_reg_rtx (hvmode);
 
-	ok = ix86_expand_vector_init_duplicate (false, hvmode, x, val);
-	gcc_assert (ok);
+	  ok = ix86_expand_vector_init_duplicate (false, hvmode, x, val);
+	  gcc_assert (ok);
 
-	x = gen_rtx_VEC_CONCAT (mode, x, x);
-	emit_insn (gen_rtx_SET (VOIDmode, target, x));
-      }
+	  x = gen_rtx_VEC_CONCAT (mode, x, x);
+	  emit_insn (gen_rtx_SET (VOIDmode, target, x));
+	}
       return true;
 
     default:
@@ -40451,8 +40462,9 @@  static void
 ix86_expand_vector_init_general (bool mmx_ok, enum machine_mode mode,
 				 rtx target, rtx vals)
 {
-  rtx ops[64], op0, op1;
+  rtx ops[64], op0, op1, op2, op3, op4, op5;
   enum machine_mode half_mode = VOIDmode;
+  enum machine_mode quarter_mode = VOIDmode;
   int n, i;
 
   switch (mode)
@@ -40503,6 +40515,42 @@  half:
 			      gen_rtx_VEC_CONCAT (mode, op0, op1)));
       return;
 
+    case V64QImode:
+      quarter_mode = V16QImode;
+      half_mode = V32QImode;
+      goto quarter;
+
+    case V32HImode:
+      quarter_mode = V8HImode;
+      half_mode = V16HImode;
+      goto quarter;
+
+quarter:
+      n = GET_MODE_NUNITS (mode);
+      for (i = 0; i < n; i++)
+	ops[i] = XVECEXP (vals, 0, i);
+      op0 = gen_reg_rtx (quarter_mode);
+      op1 = gen_reg_rtx (quarter_mode);
+      op2 = gen_reg_rtx (quarter_mode);
+      op3 = gen_reg_rtx (quarter_mode);
+      op4 = gen_reg_rtx (half_mode);
+      op5 = gen_reg_rtx (half_mode);
+      ix86_expand_vector_init_interleave (quarter_mode, op0, ops,
+					  n >> 3);
+      ix86_expand_vector_init_interleave (quarter_mode, op1,
+					  &ops [n >> 2], n >> 3);
+      ix86_expand_vector_init_interleave (quarter_mode, op2,
+					  &ops [n >> 1], n >> 3);
+      ix86_expand_vector_init_interleave (quarter_mode, op3,
+					  &ops [(n >> 1) | (n >> 2)], n >> 3);
+      emit_insn (gen_rtx_SET (VOIDmode, op4,
+			      gen_rtx_VEC_CONCAT (half_mode, op0, op1)));
+      emit_insn (gen_rtx_SET (VOIDmode, op5,
+			      gen_rtx_VEC_CONCAT (half_mode, op2, op3)));
+      emit_insn (gen_rtx_SET (VOIDmode, target,
+			      gen_rtx_VEC_CONCAT (mode, op4, op5)));
+      return;
+
     case V16QImode:
       if (!TARGET_SSE4_1)
 	break;
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 106fa00..a60b0d1 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -524,7 +524,9 @@ 
   (V16SI "TARGET_AVX512F") (V16SF "TARGET_AVX512F")
   (V8DI  "TARGET_AVX512F") (V8DF  "TARGET_AVX512F")
   (V4DI  "TARGET_AVX512VL") (V4DF  "TARGET_AVX512VL")])
-(define_mode_iterator VI48F_512 [V16SI V16SF V8DI V8DF])
+(define_mode_iterator VI48F_I12_AVX512BW
+  [V16SI V16SF V8DI V8DF
+  (V32HI "TARGET_AVX512BW") (V64QI "TARGET_AVX512BW")])
 (define_mode_iterator VI48F
   [V16SI V16SF V8DI V8DF
    (V8SI "TARGET_AVX512VL") (V8SF "TARGET_AVX512VL")
@@ -17472,7 +17474,7 @@ 
 })
 
 (define_expand "vec_init<mode>"
-  [(match_operand:VI48F_512 0 "register_operand")
+  [(match_operand:VI48F_I12_AVX512BW 0 "register_operand")
    (match_operand 1)]
   "TARGET_AVX512F"
 {