From f4d455ba6f6017e0b97a648423c858a44783ce2d Mon Sep 17 00:00:00 2001
From: Matthieu Schaller <matthieu.schaller@durham.ac.uk>
Date: Thu, 17 May 2018 12:05:09 +0200
Subject: [PATCH] Only compile the AVX512 version of meswap if using the Intel
compiler.
---
src/memswap.h | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/src/memswap.h b/src/memswap.h
index 92c902eeb1..2f7b9215ed 100644
--- a/src/memswap.h
+++ b/src/memswap.h
@@ -49,11 +49,13 @@
*
* Keep in mind that this function only works when the underlying data
* is aligned to the vector length, e.g. with the @c
- * __attribute__((aligned(32)))
- * syntax!
+ * __attribute__((aligned(32))) syntax!
* Furthermore, register re-labeling only seems to work when the code is
* compiled with @c -funroll-loops.
*
+ * Note that GCC (at least until 7.3) produces incorrect AVX512 code here
+ * by automatically assuming alignment.
+ *
* @param void_a Pointer to the first element.
* @param void_b Pointer to the second element.
* @param bytes Size, in bytes, of the data pointed to by @c a and @c b.
@@ -61,7 +63,7 @@
__attribute__((always_inline)) inline void memswap(void *void_a, void *void_b,
size_t bytes) {
char *a = (char *)void_a, *b = (char *)void_b;
-#ifdef __AVX512F__
+#if defined(__AVX512F__) && defined(__INTEL_COMPILER)
swap_loop(__m512i, a, b, bytes);
#endif
#ifdef __AVX__
--
GitLab