Patchwork Fix SLSR wrong-code (PR tree-optimization/56962)

login
register
mail settings
Submitter Jakub Jelinek
Date April 15, 2013, 2:28 p.m.
Message ID <20130415142832.GI12880@tucnak.redhat.com>
Download mbox | patch
Permalink /patch/236620/
State New
Headers show

Comments

Jakub Jelinek - April 15, 2013, 2:28 p.m.
Hi!

record_increment failed to verify that initializer is usable, which it is
only if cand_stmt is an addition (CAND_ADD can be e.g. even on a cast of
addition to some type of the same precision etc.) and one of the operands is
c->base_expr (because then the other operand necessarily has to be the rest,
but the code was only checking one of the operands, but cand_stmt e.g. can
be a sum of two SSA_NAMEs where each of those adds some multiply of one of
base_expr operands and some multiply of the c->stride.  If we set
initializer to randomly chosen operand of such stmt, while we'll have the
right multiply of c->stride, the base_expr might be wrong.

Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk/4.8?

2013-04-15  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/56962
	* gimple-ssa-strength-reduction.c (record_increment): Only set
	initializer if gimple_assign_rhs_code is {,POINTER_}PLUS_EXPR and
	either rhs1 or rhs2 is equal to c->base_expr.

	* gcc.c-torture/execute/pr56962.c: New test.


	Jakub
William J. Schmidt - April 15, 2013, 2:44 p.m.
FWIW, I agree with Jakub's fix.  Thanks!

Bill

On Mon, 2013-04-15 at 16:28 +0200, Jakub Jelinek wrote:
> Hi!
> 
> record_increment failed to verify that initializer is usable, which it is
> only if cand_stmt is an addition (CAND_ADD can be e.g. even on a cast of
> addition to some type of the same precision etc.) and one of the operands is
> c->base_expr (because then the other operand necessarily has to be the rest,
> but the code was only checking one of the operands, but cand_stmt e.g. can
> be a sum of two SSA_NAMEs where each of those adds some multiply of one of
> base_expr operands and some multiply of the c->stride.  If we set
> initializer to randomly chosen operand of such stmt, while we'll have the
> right multiply of c->stride, the base_expr might be wrong.
> 
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk/4.8?
> 
> 2013-04-15  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR tree-optimization/56962
> 	* gimple-ssa-strength-reduction.c (record_increment): Only set
> 	initializer if gimple_assign_rhs_code is {,POINTER_}PLUS_EXPR and
> 	either rhs1 or rhs2 is equal to c->base_expr.
> 
> 	* gcc.c-torture/execute/pr56962.c: New test.
> 
> --- gcc/gimple-ssa-strength-reduction.c.jj	2013-01-11 09:02:50.000000000 +0100
> +++ gcc/gimple-ssa-strength-reduction.c	2013-04-15 11:59:46.668463873 +0200
> @@ -1829,16 +1829,20 @@ record_increment (slsr_cand_t c, double_
>        if (c->kind == CAND_ADD
>  	  && c->index == increment
>  	  && (increment.sgt (double_int_one)
> -	      || increment.slt (double_int_minus_one)))
> +	      || increment.slt (double_int_minus_one))
> +	  && (gimple_assign_rhs_code (c->cand_stmt) == PLUS_EXPR
> +	      || gimple_assign_rhs_code (c->cand_stmt) == POINTER_PLUS_EXPR))
>  	{
> -	  tree t0;
> +	  tree t0 = NULL_TREE;
>  	  tree rhs1 = gimple_assign_rhs1 (c->cand_stmt);
>  	  tree rhs2 = gimple_assign_rhs2 (c->cand_stmt);
>  	  if (operand_equal_p (rhs1, c->base_expr, 0))
>  	    t0 = rhs2;
> -	  else
> +	  else if (operand_equal_p (rhs2, c->base_expr, 0))
>  	    t0 = rhs1;
> -	  if (SSA_NAME_DEF_STMT (t0) && gimple_bb (SSA_NAME_DEF_STMT (t0)))
> +	  if (t0
> +	      && SSA_NAME_DEF_STMT (t0)
> +	      && gimple_bb (SSA_NAME_DEF_STMT (t0)))
>  	    {
>  	      incr_vec[incr_vec_len].initializer = t0;
>  	      incr_vec[incr_vec_len++].init_bb
> --- gcc/testsuite/gcc.c-torture/execute/pr56962.c.jj	2013-04-15 12:09:24.781355085 +0200
> +++ gcc/testsuite/gcc.c-torture/execute/pr56962.c	2013-04-15 12:09:19.985381802 +0200
> @@ -0,0 +1,30 @@
> +/* PR tree-optimization/56962 */
> +
> +extern void abort (void);
> +long long v[144];
> +
> +__attribute__((noinline, noclone)) void
> +bar (long long *x)
> +{
> +  if (x != &v[29])
> +    abort ();
> +}
> +
> +__attribute__((noinline, noclone)) void
> +foo (long long *x, long y, long z)
> +{
> +  long long a, b, c;
> +  a = x[z * 4 + y * 3];
> +  b = x[z * 5 + y * 3];
> +  c = x[z * 5 + y * 4];
> +  x[y * 4] = a;
> +  bar (&x[z * 5 + y]);
> +  x[z * 5 + y * 5] = b + c;
> +}
> +
> +int
> +main ()
> +{
> +  foo (v, 24, 1);
> +  return 0;
> +}
> 
> 	Jakub
>
Richard Guenther - April 15, 2013, 2:49 p.m.
On Mon, 15 Apr 2013, Jakub Jelinek wrote:

> Hi!
> 
> record_increment failed to verify that initializer is usable, which it is
> only if cand_stmt is an addition (CAND_ADD can be e.g. even on a cast of
> addition to some type of the same precision etc.) and one of the operands is
> c->base_expr (because then the other operand necessarily has to be the rest,
> but the code was only checking one of the operands, but cand_stmt e.g. can
> be a sum of two SSA_NAMEs where each of those adds some multiply of one of
> base_expr operands and some multiply of the c->stride.  If we set
> initializer to randomly chosen operand of such stmt, while we'll have the
> right multiply of c->stride, the base_expr might be wrong.
> 
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk/4.8?

Ok.

Thanks,
Richard.

> 2013-04-15  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR tree-optimization/56962
> 	* gimple-ssa-strength-reduction.c (record_increment): Only set
> 	initializer if gimple_assign_rhs_code is {,POINTER_}PLUS_EXPR and
> 	either rhs1 or rhs2 is equal to c->base_expr.
> 
> 	* gcc.c-torture/execute/pr56962.c: New test.
> 
> --- gcc/gimple-ssa-strength-reduction.c.jj	2013-01-11 09:02:50.000000000 +0100
> +++ gcc/gimple-ssa-strength-reduction.c	2013-04-15 11:59:46.668463873 +0200
> @@ -1829,16 +1829,20 @@ record_increment (slsr_cand_t c, double_
>        if (c->kind == CAND_ADD
>  	  && c->index == increment
>  	  && (increment.sgt (double_int_one)
> -	      || increment.slt (double_int_minus_one)))
> +	      || increment.slt (double_int_minus_one))
> +	  && (gimple_assign_rhs_code (c->cand_stmt) == PLUS_EXPR
> +	      || gimple_assign_rhs_code (c->cand_stmt) == POINTER_PLUS_EXPR))
>  	{
> -	  tree t0;
> +	  tree t0 = NULL_TREE;
>  	  tree rhs1 = gimple_assign_rhs1 (c->cand_stmt);
>  	  tree rhs2 = gimple_assign_rhs2 (c->cand_stmt);
>  	  if (operand_equal_p (rhs1, c->base_expr, 0))
>  	    t0 = rhs2;
> -	  else
> +	  else if (operand_equal_p (rhs2, c->base_expr, 0))
>  	    t0 = rhs1;
> -	  if (SSA_NAME_DEF_STMT (t0) && gimple_bb (SSA_NAME_DEF_STMT (t0)))
> +	  if (t0
> +	      && SSA_NAME_DEF_STMT (t0)
> +	      && gimple_bb (SSA_NAME_DEF_STMT (t0)))
>  	    {
>  	      incr_vec[incr_vec_len].initializer = t0;
>  	      incr_vec[incr_vec_len++].init_bb
> --- gcc/testsuite/gcc.c-torture/execute/pr56962.c.jj	2013-04-15 12:09:24.781355085 +0200
> +++ gcc/testsuite/gcc.c-torture/execute/pr56962.c	2013-04-15 12:09:19.985381802 +0200
> @@ -0,0 +1,30 @@
> +/* PR tree-optimization/56962 */
> +
> +extern void abort (void);
> +long long v[144];
> +
> +__attribute__((noinline, noclone)) void
> +bar (long long *x)
> +{
> +  if (x != &v[29])
> +    abort ();
> +}
> +
> +__attribute__((noinline, noclone)) void
> +foo (long long *x, long y, long z)
> +{
> +  long long a, b, c;
> +  a = x[z * 4 + y * 3];
> +  b = x[z * 5 + y * 3];
> +  c = x[z * 5 + y * 4];
> +  x[y * 4] = a;
> +  bar (&x[z * 5 + y]);
> +  x[z * 5 + y * 5] = b + c;
> +}
> +
> +int
> +main ()
> +{
> +  foo (v, 24, 1);
> +  return 0;
> +}
> 
> 	Jakub
> 
>

Patch

--- gcc/gimple-ssa-strength-reduction.c.jj	2013-01-11 09:02:50.000000000 +0100
+++ gcc/gimple-ssa-strength-reduction.c	2013-04-15 11:59:46.668463873 +0200
@@ -1829,16 +1829,20 @@  record_increment (slsr_cand_t c, double_
       if (c->kind == CAND_ADD
 	  && c->index == increment
 	  && (increment.sgt (double_int_one)
-	      || increment.slt (double_int_minus_one)))
+	      || increment.slt (double_int_minus_one))
+	  && (gimple_assign_rhs_code (c->cand_stmt) == PLUS_EXPR
+	      || gimple_assign_rhs_code (c->cand_stmt) == POINTER_PLUS_EXPR))
 	{
-	  tree t0;
+	  tree t0 = NULL_TREE;
 	  tree rhs1 = gimple_assign_rhs1 (c->cand_stmt);
 	  tree rhs2 = gimple_assign_rhs2 (c->cand_stmt);
 	  if (operand_equal_p (rhs1, c->base_expr, 0))
 	    t0 = rhs2;
-	  else
+	  else if (operand_equal_p (rhs2, c->base_expr, 0))
 	    t0 = rhs1;
-	  if (SSA_NAME_DEF_STMT (t0) && gimple_bb (SSA_NAME_DEF_STMT (t0)))
+	  if (t0
+	      && SSA_NAME_DEF_STMT (t0)
+	      && gimple_bb (SSA_NAME_DEF_STMT (t0)))
 	    {
 	      incr_vec[incr_vec_len].initializer = t0;
 	      incr_vec[incr_vec_len++].init_bb
--- gcc/testsuite/gcc.c-torture/execute/pr56962.c.jj	2013-04-15 12:09:24.781355085 +0200
+++ gcc/testsuite/gcc.c-torture/execute/pr56962.c	2013-04-15 12:09:19.985381802 +0200
@@ -0,0 +1,30 @@ 
+/* PR tree-optimization/56962 */
+
+extern void abort (void);
+long long v[144];
+
+__attribute__((noinline, noclone)) void
+bar (long long *x)
+{
+  if (x != &v[29])
+    abort ();
+}
+
+__attribute__((noinline, noclone)) void
+foo (long long *x, long y, long z)
+{
+  long long a, b, c;
+  a = x[z * 4 + y * 3];
+  b = x[z * 5 + y * 3];
+  c = x[z * 5 + y * 4];
+  x[y * 4] = a;
+  bar (&x[z * 5 + y]);
+  x[z * 5 + y * 5] = b + c;
+}
+
+int
+main ()
+{
+  foo (v, 24, 1);
+  return 0;
+}