===================================================================
@@ -328,13 +328,11 @@ (define_predicate "easy_vector_constant"
if (TARGET_PAIRED_FLOAT)
return false;
- if ((VSX_VECTOR_MODE (mode) || mode == TImode) && zero_constant (op, mode))
- return true;
-
- if (ALTIVEC_VECTOR_MODE (mode))
+ if (VECTOR_MEM_ALTIVEC_OR_VSX_P (mode))
{
if (zero_constant (op, mode))
- return true;
+ return true;
+
return easy_altivec_constant (op, mode);
}
===================================================================
@@ -0,0 +1,16 @@
+/* { dg-do compile { target { powerpc*-*-* } } } */
+/* { dg-skip-if "" { powerpc*-*-darwin* } { "*" } { "" } } */
+/* { dg-require-effective-target powerpc_vsx_ok } */
+/* { dg-options "-O3 -mcpu=power7" } */
+/* { dg-final { scan-assembler "xxlxor" } } */
+/* { dg-final { scan-assembler-not "lxvd2x" } } */
+/* { dg-final { scan-assembler-not "lxvw4x" } } */
+/* { dg-final { scan-assembler-not "lvx" } } */
+
+/* PR 47755: Compiler loads vector constant of 0 from TOC instead of using
+ xxlxor. */
+void
+func (vector long long *p)
+{
+ *p = (vector long long) { 0LL, 0LL };
+}
This patch was discovered when we tried to build glibc with -O3 and -mcpu=power7. The initialization function _dl_start sets an auto array to all 0's, but it is done before the TOC pointer is setup. The compiler was assuming that vector long long constants weren't easy, so it pushed them to the constant pool. The vectorizer converts the DImode 0 into V2DI mode { 0, 0 } for the loop. I have done a bootstrap to the compiler and side by side make check's with no regressions. Is it ok to apply the patch? [gcc] 2011-02-15 Michael Meissner <meissner@linux.vnet.ibm.com> PR target/47755 * config/rs6000/predicates.md (easy_vector_constant): Allow V2DI mode for vector constants. [gcc/testsuite] 2011-02-15 Michael Meissner <meissner@linux.vnet.ibm.com> PR target/47755 * gcc.target/powerpc/pr47755.c: New file, test all 0 vector constant does not generate a load from memory.