PR82816: Widening multiplies of bitfields
authorRichard Sandiford <richard.sandiford@linaro.org>
Mon, 6 Nov 2017 14:47:43 +0000 (14:47 +0000)
committerRichard Sandiford <rsandifo@gcc.gnu.org>
Mon, 6 Nov 2017 14:47:43 +0000 (14:47 +0000)
In this PR we tried to create a widening multiply of two 3-bit numbers,
but that isn't a widening multiply at the optab/rtl level, since both
the input and output still have the same mode.

We could trap this either in is_widening_mult_p or (as the patch does)
in the routines that actually ask for an optab.  The latter seemed
more natural since is_widening_mult_p doesn't otherwise care about modes.

2017-11-03  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
PR tree-optimization/82816
* tree-ssa-math-opts.c (convert_mult_to_widen): Return false
if the modes of the two types are the same.
(convert_plusminus_to_widen): Likewise.

gcc/testsuite/
* gcc.c-torture/compile/pr82816.c: New test.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r254454

gcc/ChangeLog
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.c-torture/compile/pr82816.c [new file with mode: 0644]
gcc/tree-ssa-math-opts.c

index 6ba0a80..4ff37f4 100644 (file)
@@ -1,3 +1,12 @@
+2017-11-06  Richard Sandiford  <richard.sandiford@linaro.org>
+           Alan Hayward  <alan.hayward@arm.com>
+           David Sherwood  <david.sherwood@arm.com>
+
+       PR tree-optimization/82816
+       * tree-ssa-math-opts.c (convert_mult_to_widen): Return false
+       if the modes of the two types are the same.
+       (convert_plusminus_to_widen): Likewise.
+
 2017-11-06  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
 
        * config/rs6000/altivec.md (*p9_vadu<mode>3) Rename to
index dde8a35..d098cfc 100644 (file)
@@ -1,3 +1,9 @@
+2017-11-06  Richard Sandiford  <richard.sandiford@linaro.org>
+           Alan Hayward  <alan.hayward@arm.com>
+           David Sherwood  <david.sherwood@arm.com>
+
+       * gcc.c-torture/compile/pr82816.c: New test.
+
 2017-11-06  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
 
        * gcc.target/powerpc/sad-vectorize-1.c: New file.
diff --git a/gcc/testsuite/gcc.c-torture/compile/pr82816.c b/gcc/testsuite/gcc.c-torture/compile/pr82816.c
new file mode 100644 (file)
index 0000000..8e9bd00
--- /dev/null
@@ -0,0 +1,12 @@
+struct A
+{
+  int b:3;
+} d, e;
+
+int c;
+
+void f ()
+{
+  char g = d.b * e.b;
+  c = g;
+}
index 493f4e2..5986ac1 100644 (file)
@@ -3259,6 +3259,9 @@ convert_mult_to_widen (gimple *stmt, gimple_stmt_iterator *gsi)
 
   to_mode = SCALAR_INT_TYPE_MODE (type);
   from_mode = SCALAR_INT_TYPE_MODE (type1);
+  if (to_mode == from_mode)
+    return false;
+
   from_unsigned1 = TYPE_UNSIGNED (type1);
   from_unsigned2 = TYPE_UNSIGNED (type2);
 
@@ -3449,6 +3452,9 @@ convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple *stmt,
 
   to_mode = SCALAR_TYPE_MODE (type);
   from_mode = SCALAR_TYPE_MODE (type1);
+  if (to_mode == from_mode)
+    return false;
+
   from_unsigned1 = TYPE_UNSIGNED (type1);
   from_unsigned2 = TYPE_UNSIGNED (type2);
   optype = type1;