This patch fixes a bug in which the widening multiply-and-accumulate
optimization failed to take the intermediate types into account.
The effect of this is that the compiler would do what the programmer
expected to happen, rather than what the C standard requires to happen
(in many cases), so obviously this needed fixing. :)
It still needs to optimize the cases where the optimization doesn't
actually change anything (because it's known that overflow cannot
occur), so I don't want to completely disallow extends between multiply
and plus, and I believe this patch achieves this.
OK?
Andrew
2011-10-15 Andrew Stubbs <a...@codesourcery.com>
gcc/
* tree-ssa-math-opts.c (is_widening_mult_p): Remove the 'type'
parameter. Calculate 'type' from stmt.
(convert_mult_to_widen): Update call the is_widening_mult_p.
(convert_plusminus_to_widen): Likewise.
gcc/testsuite/
* gcc.dg/pr50717-1.c: New file.
* gcc.target/arm/wmul-12.c: Correct types.
* gcc.target/arm/wmul-8.c: Correct types.
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr50717-1.c
@@ -0,0 +1,26 @@
+/* PR tree-optimization/50717 */
+/* Ensure that widening multiply-and-accumulate is not used where integer
+ type promotion or users' casts should prevent it. */
+
+/* { dg-options "-O2 -fdump-tree-widening_mul" } */
+
+long long
+f (unsigned int a, char b, long long c)
+{
+ return (a * b) + c;
+}
+
+int
+g (short a, short b, int c)
+{
+ return (short)(a * b) + c;
+}
+
+int
+h (char a, char b, int c)
+{
+ return (char)(a * b) + c;
+}
+
+/* { dg-final { scan-tree-dump-times "WIDEN_MULT_PLUS_EXPR" 0 "widening_mul" } } */
+/* { dg-final { cleanup-tree-dump "widening_mul" } } */
--- a/gcc/testsuite/gcc.target/arm/wmul-12.c
+++ b/gcc/testsuite/gcc.target/arm/wmul-12.c
@@ -5,8 +5,8 @@
long long
foo (int *b, int *c)
{
- int tmp = *b * *c;
- return 10 + (long long)tmp;
+ long long tmp = (long long)*b * *c;
+ return 10 + tmp;
}
/* { dg-final { scan-assembler "smlal" } } */
--- a/gcc/testsuite/gcc.target/arm/wmul-8.c
+++ b/gcc/testsuite/gcc.target/arm/wmul-8.c
@@ -5,7 +5,7 @@
long long
foo (long long a, int *b, int *c)
{
- return a + *b * *c;
+ return a + (long long)*b * *c;
}
/* { dg-final { scan-assembler "smlal" } } */
--- a/gcc/tree-ssa-math-opts.c
+++ b/gcc/tree-ssa-math-opts.c
@@ -2039,10 +2039,12 @@ is_widening_mult_rhs_p (tree type, tree rhs, tree *type_out,
and *TYPE2_OUT would give the operands of the multiplication. */
static bool
-is_widening_mult_p (tree type, gimple stmt,
+is_widening_mult_p (gimple stmt,
tree *type1_out, tree *rhs1_out,
tree *type2_out, tree *rhs2_out)
{
+ tree type = TREE_TYPE (gimple_assign_lhs (stmt));
+
if (TREE_CODE (type) != INTEGER_TYPE
&& TREE_CODE (type) != FIXED_POINT_TYPE)
return false;
@@ -2104,7 +2106,7 @@ convert_mult_to_widen (gimple stmt, gimple_stmt_iterator *gsi)
if (TREE_CODE (type) != INTEGER_TYPE)
return false;
- if (!is_widening_mult_p (type, stmt, &type1, &rhs1, &type2, &rhs2))
+ if (!is_widening_mult_p (stmt, &type1, &rhs1, &type2, &rhs2))
return false;
to_mode = TYPE_MODE (type);
@@ -2281,7 +2283,7 @@ convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple stmt,
if (code == PLUS_EXPR
&& (rhs1_code == MULT_EXPR || rhs1_code == WIDEN_MULT_EXPR))
{
- if (!is_widening_mult_p (type, rhs1_stmt, &type1, &mult_rhs1,
+ if (!is_widening_mult_p (rhs1_stmt, &type1, &mult_rhs1,
&type2, &mult_rhs2))
return false;
add_rhs = rhs2;
@@ -2289,7 +2291,7 @@ convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple stmt,
}
else if (rhs2_code == MULT_EXPR || rhs2_code == WIDEN_MULT_EXPR)
{
- if (!is_widening_mult_p (type, rhs2_stmt, &type1, &mult_rhs1,
+ if (!is_widening_mult_p (rhs2_stmt, &type1, &mult_rhs1,
&type2, &mult_rhs2))
return false;
add_rhs = rhs1;