diff mbox

Fix tree-optimization/49963

Message ID 4E4C019D.1040201@oracle.com
State New
Headers show

Commit Message

Paolo Carlini Aug. 17, 2011, 5:59 p.m. UTC 
Hi,

I prepared the below basing on an hint provided by Joseph on the audit 
trail: essentially, I'm replacing all (but two) uses of abs_hwi outside 
hwint.c by absu_hwi, a variant which returns an unsigned HOST_WIDE_INT. 
All the replacements make sense to me: either we are comparing two abs, 
or we are passing the abs to a function actually expecting an unsigned 
HOST_WIDE_INT, or we are comparing to another unsigned HOST_WIDE_INT, or 
we are just comparing to a constant (I don't feel strongly in this case 
but seems safe to use absu_hwi here too). I'm *not* replacing 2 
occurrences in gimple_expand_builtin_pow, because those are safe anyway 
in terms of range of the argument (it's an HOST_WIDE_INT / 2 or 3) and 
the result is passed to a function expecting a plain HOST_WIDE_INT (ie, 
gimple_expand_builtin_powi).

I sanity checked the patch on x86_64-linux and OP reported that on AVR 
the patch fixes the regression.

Is it ok?

Thanks,
Paolo.

PS: compared to the draft version, the attached uses cast to unsigned in 
both arms of the ? : operator, I think Joseph preferred that 
stylistically in a snippet of him posted in an unrelated recent audit trail.

///////////////////////////
2011-08-17  Paolo Carlini  <paolo.carlini@oracle.com>
	    Joseph Myers  <joseph@codesourcery.com>

	PR tree-optimization/49963
	* hwint.c (absu_hwi): Define.
	* hwint.h (absu_hwi): Declare.
	* fold-const.c (fold_plusminus_mult_expr): Use absu_hwi instead
	of abs_hwi.
	* tree-ssa-math-opts.c (gimple_expand_builtin_pow): Likewise.
	* tree-ssa-loop-prefetch.c (prune_ref_by_group_reuse): Likewise.

Comments

Georg-Johann Lay Aug. 17, 2011, 6:55 p.m. UTC | #1 
Paolo Carlini wrote:
> 
> I sanity checked the patch on x86_64-linux and OP reported that on AVR 
> the patch fixes the regression.

Not really "on" AVR; AVR are just tiny 8-bit microcontrollers ;-)
I obseved the problem when compiling for AVR on a x86-linux-gnu host 
where 0x80000000 is negative.

Johann
Richard Biener Aug. 18, 2011, 8:30 a.m. UTC | #2 
On Wed, 17 Aug 2011, Paolo Carlini wrote:

> Hi,
> 
> I prepared the below basing on an hint provided by Joseph on the audit trail:
> essentially, I'm replacing all (but two) uses of abs_hwi outside hwint.c by
> absu_hwi, a variant which returns an unsigned HOST_WIDE_INT. All the
> replacements make sense to me: either we are comparing two abs, or we are
> passing the abs to a function actually expecting an unsigned HOST_WIDE_INT, or
> we are comparing to another unsigned HOST_WIDE_INT, or we are just comparing
> to a constant (I don't feel strongly in this case but seems safe to use
> absu_hwi here too). I'm *not* replacing 2 occurrences in
> gimple_expand_builtin_pow, because those are safe anyway in terms of range of
> the argument (it's an HOST_WIDE_INT / 2 or 3) and the result is passed to a
> function expecting a plain HOST_WIDE_INT (ie, gimple_expand_builtin_powi).
> 
> I sanity checked the patch on x86_64-linux and OP reported that on AVR the
> patch fixes the regression.
> 
> Is it ok?

Ok.  It looks like we might want to elimiate abs_hwi alltogether
in favor of absu_hwi?

Thanks,
Richard.
Paolo Carlini Aug. 18, 2011, 10:26 a.m. UTC | #3 
On 08/18/2011 10:30 AM, Richard Guenther wrote:
> Ok.
Great. Thus I'm going ahead and committing the patch to fix the regression.
> It looks like we might want to elimiate abs_hwi alltogether in favor 
> of absu_hwi? 
Indeed, you are totally right. But I'd rather work on that as a followup 
patch (maybe with Sebastian' help?) because it doesn't seem completely 
trivial: I'm mostly worried by the existing uses of abs_hwi *inside* 
hwint.c itself not but those two remaining instances outside of it 
(which would be possible to replace immediately).

Paolo.
Richard Biener Aug. 18, 2011, 11:04 a.m. UTC | #4 
On Thu, 18 Aug 2011, Paolo Carlini wrote:

> On 08/18/2011 10:30 AM, Richard Guenther wrote:
> > Ok.
> Great. Thus I'm going ahead and committing the patch to fix the regression.
> > It looks like we might want to elimiate abs_hwi alltogether in favor of
> > absu_hwi? 
> Indeed, you are totally right. But I'd rather work on that as a followup patch
> (maybe with Sebastian' help?) because it doesn't seem completely trivial: I'm
> mostly worried by the existing uses of abs_hwi *inside* hwint.c itself not but
> those two remaining instances outside of it (which would be possible to
> replace immediately).

Sure as a followup.

Richard.
diff mbox

Patch

Index: fold-const.c
===================================================================
--- fold-const.c	(revision 177834)
+++ fold-const.c	(working copy)
@@ -7036,7 +7036,7 @@  fold_plusminus_mult_expr (location_t loc, enum tre
       int11 = TREE_INT_CST_LOW (arg11);
 
       /* Move min of absolute values to int11.  */
-      if (abs_hwi (int01) < abs_hwi (int11))
+      if (absu_hwi (int01) < absu_hwi (int11))
         {
 	  tmp = int01, int01 = int11, int11 = tmp;
 	  alt0 = arg00, arg00 = arg10, arg10 = alt0;
@@ -7046,7 +7046,7 @@  fold_plusminus_mult_expr (location_t loc, enum tre
       else
 	maybe_same = arg11;
 
-      if (exact_log2 (abs_hwi (int11)) > 0 && int01 % int11 == 0
+      if (exact_log2 (absu_hwi (int11)) > 0 && int01 % int11 == 0
 	  /* The remainder should not be a constant, otherwise we
 	     end up folding i * 4 + 2 to (i * 2 + 1) * 2 which has
 	     increased the number of multiplications necessary.  */
Index: tree-ssa-math-opts.c
===================================================================
--- tree-ssa-math-opts.c	(revision 177834)
+++ tree-ssa-math-opts.c	(working copy)
@@ -1231,7 +1231,7 @@  gimple_expand_builtin_pow (gimple_stmt_iterator *g
       /* Attempt to fold powi(arg0, abs(n/2)) into multiplies.  If not
          possible or profitable, give up.  Skip the degenerate case when
          n is 1 or -1, where the result is always 1.  */
-      if (abs_hwi (n) != 1)
+      if (absu_hwi (n) != 1)
 	{
 	  powi_x_ndiv2 = gimple_expand_builtin_powi (gsi, loc, arg0,
 						     abs_hwi (n / 2));
@@ -1243,7 +1243,7 @@  gimple_expand_builtin_pow (gimple_stmt_iterator *g
 	 result of the optimal multiply sequence just calculated.  */
       sqrt_arg0 = build_and_insert_call (gsi, loc, &target, sqrtfn, arg0);
 
-      if (abs_hwi (n) == 1)
+      if (absu_hwi (n) == 1)
 	result = sqrt_arg0;
       else
 	result = build_and_insert_binop (gsi, loc, target, MULT_EXPR,
@@ -1285,7 +1285,7 @@  gimple_expand_builtin_pow (gimple_stmt_iterator *g
       /* Attempt to fold powi(arg0, abs(n/3)) into multiplies.  If not
          possible or profitable, give up.  Skip the degenerate case when
          abs(n) < 3, where the result is always 1.  */
-      if (abs_hwi (n) >= 3)
+      if (absu_hwi (n) >= 3)
 	{
 	  powi_x_ndiv3 = gimple_expand_builtin_powi (gsi, loc, arg0,
 						     abs_hwi (n / 3));
@@ -1298,14 +1298,14 @@  gimple_expand_builtin_pow (gimple_stmt_iterator *g
          either cbrt(x) or cbrt(x) * cbrt(x).  */
       cbrt_x = build_and_insert_call (gsi, loc, &target, cbrtfn, arg0);
 
-      if (abs_hwi (n) % 3 == 1)
+      if (absu_hwi (n) % 3 == 1)
 	powi_cbrt_x = cbrt_x;
       else
 	powi_cbrt_x = build_and_insert_binop (gsi, loc, target, MULT_EXPR,
 					      cbrt_x, cbrt_x);
 
       /* Multiply the two subexpressions, unless powi(x,abs(n)/3) = 1.  */
-      if (abs_hwi (n) < 3)
+      if (absu_hwi (n) < 3)
 	result = powi_cbrt_x;
       else
 	result = build_and_insert_binop (gsi, loc, target, MULT_EXPR,
Index: tree-ssa-loop-prefetch.c
===================================================================
--- tree-ssa-loop-prefetch.c	(revision 177834)
+++ tree-ssa-loop-prefetch.c	(working copy)
@@ -795,7 +795,7 @@  prune_ref_by_group_reuse (struct mem_ref *ref, str
       prefetch_before = (hit_from - delta_r + step - 1) / step;
 
       /* Do not reduce prefetch_before if we meet beyond cache size.  */
-      if (prefetch_before > (unsigned) abs_hwi (L2_CACHE_SIZE_BYTES / step))
+      if (prefetch_before > absu_hwi (L2_CACHE_SIZE_BYTES / step))
         prefetch_before = PREFETCH_ALL;
       if (prefetch_before < ref->prefetch_before)
 	ref->prefetch_before = prefetch_before;
Index: hwint.c
===================================================================
--- hwint.c	(revision 177834)
+++ hwint.c	(working copy)
@@ -109,6 +109,14 @@  abs_hwi (HOST_WIDE_INT x)
   return x >= 0 ? x : -x;
 }
 
+/* Compute the absolute value of X as an unsigned type.  */
+
+unsigned HOST_WIDE_INT
+absu_hwi (HOST_WIDE_INT x)
+{
+  return x >= 0 ? (unsigned HOST_WIDE_INT)x : -(unsigned HOST_WIDE_INT)x;
+}
+
 /* Compute the greatest common divisor of two numbers A and B using
    Euclid's algorithm.  */
 
Index: hwint.h
===================================================================
--- hwint.h	(revision 177834)
+++ hwint.h	(working copy)
@@ -233,6 +233,7 @@  exact_log2 (unsigned HOST_WIDE_INT x)
 #define HOST_WIDE_INT_MAX (~(HOST_WIDE_INT_MIN))
 
 extern HOST_WIDE_INT abs_hwi (HOST_WIDE_INT);
+extern unsigned HOST_WIDE_INT absu_hwi (HOST_WIDE_INT);
 extern HOST_WIDE_INT gcd (HOST_WIDE_INT, HOST_WIDE_INT);
 extern HOST_WIDE_INT pos_mul_hwi (HOST_WIDE_INT, HOST_WIDE_INT);
 extern HOST_WIDE_INT mul_hwi (HOST_WIDE_INT, HOST_WIDE_INT);