[backport approved but not yet applied patch for an __atomic_load
 problem on ARMv7 <mikpe@it.uu.se>]

Date: Thu, 29 Nov 2012 16:42:41 +0100
Subject: [PATCH] Fix for PR55492 : __atomic_load doesn't match ACQUIRE memory model
From: Yvan Roux <yvan dot roux at linaro dot org>
List-Archive: <http://gcc.gnu.org/ml/gcc-patches/>

Hi,

on ARMv7, the code generated for the __atomic_load builtins in the
__ATOMIC_ACQUIRE memory model, puts a memory barrier before the load, whereas
the semantic of the acquire memory model implies a barrier after.

The issue seems to be in expand_atomic_load which puts a memory fence before
the load in any memory model. The attached patch fixes the problem.

Thanks,
Yvan

gcc/

2012-11-29  Yvan Roux  <yvan.roux@st.com>

	PR target/55942
	* optabs.c (expand_atomic_load): Condition the memory fence to the
	the right memory model.

--- gcc-4.7.2/gcc/optabs.c.~1~	2012-03-04 03:34:55.000000000 +0100
+++ gcc-4.7.2/gcc/optabs.c	2012-12-02 17:31:06.000000000 +0100
@@ -7701,12 +7701,13 @@ expand_atomic_load (rtx target, rtx mem,
     target = gen_reg_rtx (mode);
 
   /* Emit the appropriate barrier before the load.  */
-  expand_mem_thread_fence (model);
+  if (model == MEMMODEL_SEQ_CST)
+    expand_mem_thread_fence (model);
 
   emit_move_insn (target, mem);
 
-  /* For SEQ_CST, also emit a barrier after the load.  */
-  if (model == MEMMODEL_SEQ_CST)
+  /* For SEQ_CST or ACQUIRE, also emit a barrier after the load.  */
+  if (model == MEMMODEL_SEQ_CST || model == MEMMODEL_ACQUIRE)
     expand_mem_thread_fence (model);
 
   return target;