From: Richard Sandiford <richard.sandiford@linaro.org>
Date: Fri, 25 May 2018 08:18:42 +0000 (+0000)
Subject: Prefer open-coding vector integer division
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=8f76f377861b4195487416806c4a0eacabc433c9;p=gcc.git

Prefer open-coding vector integer division

vect_recog_divmod_pattern currently bails out if the target has
native support for integer division, but I think in practice
it's always going to be better to open-code it anyway, just as
we usually open-code scalar divisions by constants.

I think the only currently affected targets are MIPS MSA and
powerpcspe (which is currently marked obsolete).  For:

  void
  foo (int *x)
  {
    for (int i = 0; i < 100; ++i)
      x[i] /= 2;
  }

the MSA port previously preferred to use division for powers of 2:

        .set    noreorder
        bnz.w   $w1,1f
        div_s.w $w0,$w0,$w1
        break   7
        .set    reorder
1:

(or just the div_s.w for -mno-check-zero-division), but after the patch
it open-codes them using shifts:

        clt_s.w $w1,$w0,$w2
        subv.w  $w0,$w0,$w1
        srai.w  $w0,$w0,1

MSA doesn't define a high-part pattern, so it still uses a division
instruction for the non-power-of-2 case.

Richard B pointed out that this would disable SLP of division by
different amounts, but I think in practice that's a price worth paying,
since the current cost model can't really tell whether using a general
vector division is better than using open-coded scalar divisions.
The fix would be either to support SLP of mixed open-coded divisions
or to improve the cost model and try SLP again without the patterns.
The patch adds an XFAILed test for this.

2018-05-23  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
	* tree-vect-patterns.c: Include predict.h.
	(vect_recog_divmod_pattern): Restrict check for division support
	to when optimizing for size.

gcc/testsuite/
	* gcc.dg/vect/bb-slp-div-1.c: New XFAILed test.

From-SVN: r260711
---

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index fd187b92d39..67938386d67 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,9 @@
+2018-05-25  Richard Sandiford  <richard.sandiford@linaro.org>
+
+	* tree-vect-patterns.c: Include predict.h.
+	(vect_recog_divmod_pattern): Restrict check for division support
+	to when optimizing for size.
+
 2018-05-25  Richard Sandiford  <richard.sandiford@linaro.org>
 
 	* doc/sourcebuild.texi (vect_double_cond_arith: Document.
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index c5b2c631b5d..66296db447e 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,7 @@
+2018-05-25  Richard Sandiford  <richard.sandiford@linaro.org>
+
+	* gcc.dg/vect/bb-slp-div-1.c: New XFAILed test.
+
 2018-05-25  Richard Sandiford  <richard.sandiford@linaro.org>
 
 	* lib/target-supports.exp
diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-div-1.c b/gcc/testsuite/gcc.dg/vect/bb-slp-div-1.c
new file mode 100644
index 00000000000..65d83a437b6
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/bb-slp-div-1.c
@@ -0,0 +1,19 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-msve-vector-bits=256" { target aarch64_sve } } */
+
+int x[8];
+
+void
+f (void)
+{
+  x[0] /= 2;
+  x[1] /= 3;
+  x[2] /= 4;
+  x[3] /= 5;
+  x[4] /= 6;
+  x[5] /= 7;
+  x[6] /= 8;
+  x[7] /= 9;
+}
+
+/* { dg-final { scan-tree-dump "basic block vectorized" "slp2" { xfail *-*-* } } } */
diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c
index 75bf84b7645..6da784cdc3a 100644
--- a/gcc/tree-vect-patterns.c
+++ b/gcc/tree-vect-patterns.c
@@ -45,6 +45,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "attribs.h"
 #include "cgraph.h"
 #include "omp-simd-clone.h"
+#include "predict.h"
 
 /* Pattern recognition functions  */
 static gimple *vect_recog_widen_sum_pattern (vec<gimple *> *, tree *,
@@ -2674,15 +2675,19 @@ vect_recog_divmod_pattern (vec<gimple *> *stmts,
   if (vectype == NULL_TREE)
     return NULL;
 
-  /* If the target can handle vectorized division or modulo natively,
-     don't attempt to optimize this.  */
-  optab = optab_for_tree_code (rhs_code, vectype, optab_default);
-  if (optab != unknown_optab)
+  if (optimize_bb_for_size_p (gimple_bb (last_stmt)))
     {
-      machine_mode vec_mode = TYPE_MODE (vectype);
-      int icode = (int) optab_handler (optab, vec_mode);
-      if (icode != CODE_FOR_nothing)
-	return NULL;
+      /* If the target can handle vectorized division or modulo natively,
+	 don't attempt to optimize this, since native division is likely
+	 to give smaller code.  */
+      optab = optab_for_tree_code (rhs_code, vectype, optab_default);
+      if (optab != unknown_optab)
+	{
+	  machine_mode vec_mode = TYPE_MODE (vectype);
+	  int icode = (int) optab_handler (optab, vec_mode);
+	  if (icode != CODE_FOR_nothing)
+	    return NULL;
+	}
     }
 
   prec = TYPE_PRECISION (itype);