diff mbox

Improve vec_widen_?mult_odd_* (take 2)

Message ID 20130428214306.GZ28963@tucnak.redhat.com
State New
Headers show

Commit Message

Jakub Jelinek April 28, 2013, 9:43 p.m. UTC 
On Sat, Apr 27, 2013 at 11:07:50AM +0200, Uros Bizjak wrote:
> Yes, please add a new predicate, the pattern is much more descriptive
> this way. (Without the predicate, it looks like an expander that
> generates a RTX fragment, used instead of gen_RTX... sequences).

Ok, updated patch below.  Bootstrapped/regtested again on x86_64-linux and
i686-linux.

> OTOH, does vector mode "general_operand" still accept scalar
> immediates? The predicate, proposed above is effectively

general_operand doesn't accept most of CONST_VECTOR constants (because they
aren't targetm.legitimate_constant_p).  It won't accept CONST_INT
or CONST_DOUBLE due to:
  /* Don't accept CONST_INT or anything similar
     if the caller wants something floating.  */
  if (GET_MODE (op) == VOIDmode && mode != VOIDmode
      && GET_MODE_CLASS (mode) != MODE_INT
      && GET_MODE_CLASS (mode) != MODE_PARTIAL_INT)
    return 0;

2013-04-28  Jakub Jelinek  <jakub@redhat.com>

	* config/i386/predicates.md (general_vector_operand): New predicate.
	* config/i386/i386.c (const_vector_equal_evenodd_p): New function.
	(ix86_expand_mul_widen_evenodd): Force op1 resp. op2 into register
	if they aren't nonimmediate operands.  If their original values
	satisfy const_vector_equal_evenodd_p, don't shift them.
	* config/i386/sse.md (mul<mode>3): Use general_vector_operand
	predicates.  For the SSE4.1 case force operands[{1,2}] into registers if
	not nonimmediate_operand.
	(vec_widen_smult_even_v4si): Use nonimmediate_operand predicates
	instead of register_operand.
	(vec_widen_<s>mult_odd_<mode>): Use general_vector_operand predicates.



	Jakub

Comments

Uros Bizjak April 29, 2013, 6:33 a.m. UTC | #1 
On Sun, Apr 28, 2013 at 11:43 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Sat, Apr 27, 2013 at 11:07:50AM +0200, Uros Bizjak wrote:
>> Yes, please add a new predicate, the pattern is much more descriptive
>> this way. (Without the predicate, it looks like an expander that
>> generates a RTX fragment, used instead of gen_RTX... sequences).
>
> Ok, updated patch below.  Bootstrapped/regtested again on x86_64-linux and
> i686-linux.
>
>> OTOH, does vector mode "general_operand" still accept scalar
>> immediates? The predicate, proposed above is effectively
>
> general_operand doesn't accept most of CONST_VECTOR constants (because they
> aren't targetm.legitimate_constant_p).  It won't accept CONST_INT
> or CONST_DOUBLE due to:
>   /* Don't accept CONST_INT or anything similar
>      if the caller wants something floating.  */
>   if (GET_MODE (op) == VOIDmode && mode != VOIDmode
>       && GET_MODE_CLASS (mode) != MODE_INT
>       && GET_MODE_CLASS (mode) != MODE_PARTIAL_INT)
>     return 0;
>
> 2013-04-28  Jakub Jelinek  <jakub@redhat.com>
>
>         * config/i386/predicates.md (general_vector_operand): New predicate.
>         * config/i386/i386.c (const_vector_equal_evenodd_p): New function.
>         (ix86_expand_mul_widen_evenodd): Force op1 resp. op2 into register
>         if they aren't nonimmediate operands.  If their original values
>         satisfy const_vector_equal_evenodd_p, don't shift them.
>         * config/i386/sse.md (mul<mode>3): Use general_vector_operand
>         predicates.  For the SSE4.1 case force operands[{1,2}] into registers if
>         not nonimmediate_operand.
>         (vec_widen_smult_even_v4si): Use nonimmediate_operand predicates
>         instead of register_operand.
>         (vec_widen_<s>mult_odd_<mode>): Use general_vector_operand predicates.

OK for mainline.

Thanks,
Uros.
diff mbox

Patch

--- gcc/config/i386/predicates.md.jj	2013-04-28 21:29:40.000000000 +0200
+++ gcc/config/i386/predicates.md	2013-04-28 21:32:30.308907622 +0200
@@ -1303,3 +1303,8 @@  (define_predicate "avx2_pblendw_operand"
   HOST_WIDE_INT low = val & 0xff;
   return val == ((low << 8) | low);
 })
+
+;; Return true if OP is nonimmediate_operand or CONST_VECTOR.
+(define_predicate "general_vector_operand"
+  (ior (match_operand 0 "nonimmediate_operand")
+       (match_code "const_vector")))
--- gcc/config/i386/i386.c.jj	2013-04-28 21:29:44.248917523 +0200
+++ gcc/config/i386/i386.c	2013-04-28 21:32:07.274908996 +0200
@@ -40827,6 +40827,24 @@  ix86_expand_vecop_qihi (enum rtx_code co
 		       gen_rtx_fmt_ee (code, qimode, op1, op2));
 }
 
+/* Helper function of ix86_expand_mul_widen_evenodd.  Return true
+   if op is CONST_VECTOR with all odd elements equal to their
+   preceeding element.  */
+
+static bool
+const_vector_equal_evenodd_p (rtx op)
+{
+  enum machine_mode mode = GET_MODE (op);
+  int i, nunits = GET_MODE_NUNITS (mode);
+  if (GET_CODE (op) != CONST_VECTOR
+      || nunits != CONST_VECTOR_NUNITS (op))
+    return false;
+  for (i = 0; i < nunits; i += 2)
+    if (CONST_VECTOR_ELT (op, i) != CONST_VECTOR_ELT (op, i + 1))
+      return false;
+  return true;
+}
+
 void
 ix86_expand_mul_widen_evenodd (rtx dest, rtx op1, rtx op2,
 			       bool uns_p, bool odd_p)
@@ -40834,6 +40852,12 @@  ix86_expand_mul_widen_evenodd (rtx dest,
   enum machine_mode mode = GET_MODE (op1);
   enum machine_mode wmode = GET_MODE (dest);
   rtx x;
+  rtx orig_op1 = op1, orig_op2 = op2;
+
+  if (!nonimmediate_operand (op1, mode))
+    op1 = force_reg (mode, op1);
+  if (!nonimmediate_operand (op2, mode))
+    op2 = force_reg (mode, op2);
 
   /* We only play even/odd games with vectors of SImode.  */
   gcc_assert (mode == V4SImode || mode == V8SImode);
@@ -40852,10 +40876,12 @@  ix86_expand_mul_widen_evenodd (rtx dest,
 	}
 
       x = GEN_INT (GET_MODE_UNIT_BITSIZE (mode));
-      op1 = expand_binop (wmode, lshr_optab, gen_lowpart (wmode, op1),
-			  x, NULL, 1, OPTAB_DIRECT);
-      op2 = expand_binop (wmode, lshr_optab, gen_lowpart (wmode, op2),
-			  x, NULL, 1, OPTAB_DIRECT);
+      if (!const_vector_equal_evenodd_p (orig_op1))
+	op1 = expand_binop (wmode, lshr_optab, gen_lowpart (wmode, op1),
+			    x, NULL, 1, OPTAB_DIRECT);
+      if (!const_vector_equal_evenodd_p (orig_op2))
+	op2 = expand_binop (wmode, lshr_optab, gen_lowpart (wmode, op2),
+			    x, NULL, 1, OPTAB_DIRECT);
       op1 = gen_lowpart (mode, op1);
       op2 = gen_lowpart (mode, op2);
     }
--- gcc/config/i386/sse.md.jj	2013-04-28 21:29:44.244917523 +0200
+++ gcc/config/i386/sse.md	2013-04-28 21:32:07.280908995 +0200
@@ -5631,14 +5631,16 @@  (define_insn "*sse2_pmaddwd"
 (define_expand "mul<mode>3"
   [(set (match_operand:VI4_AVX2 0 "register_operand")
 	(mult:VI4_AVX2
-	  (match_operand:VI4_AVX2 1 "nonimmediate_operand")
-	  (match_operand:VI4_AVX2 2 "nonimmediate_operand")))]
+	  (match_operand:VI4_AVX2 1 "general_vector_operand")
+	  (match_operand:VI4_AVX2 2 "general_vector_operand")))]
   "TARGET_SSE2"
 {
   if (TARGET_SSE4_1)
     {
-      if (CONSTANT_P (operands[2]))
-	operands[2] = force_const_mem (<MODE>mode, operands[2]);
+      if (!nonimmediate_operand (operands[1], <MODE>mode))
+	operands[1] = force_reg (<MODE>mode, operands[1]);
+      if (!nonimmediate_operand (operands[2], <MODE>mode))
+	operands[2] = force_reg (<MODE>mode, operands[2]);
       ix86_fixup_binary_operands_no_copy (MULT, <MODE>mode, operands);
     }
   else
@@ -5702,8 +5704,8 @@  (define_expand "vec_widen_<s>mult_lo_<mo
 ;; named patterns, but signed V4SI needs special help for plain SSE2.
 (define_expand "vec_widen_smult_even_v4si"
   [(match_operand:V2DI 0 "register_operand")
-   (match_operand:V4SI 1 "register_operand")
-   (match_operand:V4SI 2 "register_operand")]
+   (match_operand:V4SI 1 "nonimmediate_operand")
+   (match_operand:V4SI 2 "nonimmediate_operand")]
   "TARGET_SSE2"
 {
   ix86_expand_mul_widen_evenodd (operands[0], operands[1], operands[2],
@@ -5714,8 +5716,8 @@  (define_expand "vec_widen_smult_even_v4s
 (define_expand "vec_widen_<s>mult_odd_<mode>"
   [(match_operand:<sseunpackmode> 0 "register_operand")
    (any_extend:<sseunpackmode>
-     (match_operand:VI4_AVX2 1 "register_operand"))
-   (match_operand:VI4_AVX2 2 "register_operand")]
+     (match_operand:VI4_AVX2 1 "general_vector_operand"))
+   (match_operand:VI4_AVX2 2 "general_vector_operand")]
   "TARGET_SSE2"
 {
   ix86_expand_mul_widen_evenodd (operands[0], operands[1], operands[2],