[PATCH] Fix PR64822: incorrect folding of bitfield in union on big endian targets

["Thomas Preud'homme" <thomas dot preudhomme at arm dot com>]

Hi,

Incorrect folding happen when a bitfield in a union is accessed and the union was initialized via another field. This is a reminiscence of GCC32RM-325 due to sccvn not passing the right size to fold_ctor_reference. sccvn uses TYPE_SIZE which gives the size of the container integer for a bitfield instead of TYPE_PRECISION that gives the actual size occupied by the bitfield. This patch fixes that and modify bitfld-6 so that it catch this issue in addition to the old one.

ChangeLog entries are as follows:

*** gcc/ChangeLog ***

2015-01-28  Thomas Preud'homme  <thomas.preudhomme@arm.com>

        * tree-ssa-sccvn.c (fully_constant_vn_reference_p): Use TYPE_PRECISION
        to compute size of referenced value in the constant case.

*** gcc/testsuite/ChangeLog ***

2015-01-28  Thomas Preud'homme  <thomas.preudhomme@arm.com>

        * gcc.c-torture/execute/bitfld-6.c: Use 24 bits for bitfield b.  Adapt
        expected values accordingly.

diff --git a/gcc/testsuite/gcc.c-torture/execute/bitfld-6.c b/gcc/testsuite/gcc.c-torture/execute/bitfld-6.c
index 50927dc..e9a61df 100644
--- a/gcc/testsuite/gcc.c-torture/execute/bitfld-6.c
+++ b/gcc/testsuite/gcc.c-torture/execute/bitfld-6.c
@@ -1,7 +1,7 @@
 union U
 {
   const int a;
-  unsigned b : 20;
+  unsigned b : 24;
 };
 
 static union U u = { 0x12345678 };
@@ -14,9 +14,9 @@ main (void)
 {
 #ifdef __BYTE_ORDER__
 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
-  return u.b - 0x45678;
+  return u.b - 0x345678;
 #else
-  return u.b - 0x12345;
+  return u.b - 0x123456;
 #endif
 #endif
   return 0;
diff --git a/gcc/tree-ssa-sccvn.c b/gcc/tree-ssa-sccvn.c
index 25c67d0..0f1299a 100644
--- a/gcc/tree-ssa-sccvn.c
+++ b/gcc/tree-ssa-sccvn.c
@@ -1352,7 +1352,7 @@ fully_constant_vn_reference_p (vn_reference_t ref)
 	       || TYPE_PRECISION (ref->type) % BITS_PER_UNIT == 0))
     {
       HOST_WIDE_INT off = 0;
-      HOST_WIDE_INT size = tree_to_shwi (TYPE_SIZE (ref->type));
+      HOST_WIDE_INT size = TYPE_PRECISION (ref->type);
       if (size % BITS_PER_UNIT != 0
 	  || size > MAX_BITSIZE_MODE_ANY_MODE)
 	return NULL_TREE;

Tested by doing a native x86_64 bootstrap + testsuite run and building a arm-none-eabi cross-compiler + testsuite run (on QEMU, compiling for Cortex-M3) without any regression.

Is this ok for trunk?

Best regards,

Thomas