Disable auto-vectorization of P224 ReduceLarge() under GCC.

Something about automatic tree vectorization breaks ReduceLarge(), so
disable the optimization for this function until we determine the real
fix.

BUG=439566

Review URL: https://codereview.chromium.org/814273004

Cr-Commit-Position: refs/heads/master@{#309335}

1 files changed, 14 insertions, 0 deletions

diff --git a/crypto/p224.cc b/crypto/p224.cc
index 5f76fbc..74ad7e0 100644
--- a/crypto/p224.cc
+++ b/crypto/p224.cc
@@ -123,6 +123,15 @@ typedef uint64 LargeFieldElement[15];
 // ReduceLarge converts a LargeFieldElement to a FieldElement.
 //
 // in[i] < 2**62
+
+// GCC 4.9 incorrectly vectorizes the first coefficient elimination loop, so
+// disable that optimization via pragma. Don't use the pragma under Clang, since
+// clang doesn't understand it.
+// TODO(wez): Remove this when crbug.com/439566 is fixed.
+#if defined(__GNUC__) && !defined(__clang__)
+#pragma GCC optimize("no-tree-vectorize")
+#endif
+
 void ReduceLarge(FieldElement* out, LargeFieldElement* inptr) {
   LargeFieldElement& in(*inptr);
 
@@ -164,6 +173,11 @@ void ReduceLarge(FieldElement* out, LargeFieldElement* inptr) {
   // out[5..7] < 2**28
 }
 
+// TODO(wez): Remove this when crbug.com/439566 is fixed.
+#if defined(__GNUC__) && !defined(__clang__)
+#pragma GCC optimize("tree-vectorize")
+#endif
+
 // Mul computes *out = a*b
 //
 // a[i] < 2**29, b[i] < 2**30 (or vice versa)