...in light of recent activity related to llvm.memcpy flags. I want to

prevent an llvm developer from mistakenly thinking that just because the
intrinsic has volatile flags that volatile operations can be converted
to or folded into them.

Platforms may rely on volatile loads and stores of natively supported
data width to be executed as single instruction. When compiling
C, this expectation likely holds for l-values of volatile primitive
types with native hardware support, but not necessarily for aggregate
types. The frontend upholds these expectations, which are not
specified in the IR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@173974 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/docs/LangRef.rst b/docs/LangRef.rst
index 24a73150dac207646697b49bc841931677dd2d6e..ec34a31cd4cb5d88698223d1ada383feae39242a 100644 (file)
--- a/docs/LangRef.rst
+++ b/docs/LangRef.rst
@@ -1080,6 +1080,11 @@ volatile operations. The optimizers *may* change the order of volatile
 operations relative to non-volatile operations. This is not Java's
 "volatile" and has no cross-thread synchronization behavior.
 
+IR-level volatile loads and stores cannot safely be optimized into
+llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
+flagged volatile. Likewise, the backend should never split or merge
+target-legal volatile load/store instructions.
+
 .. _memmodel:
 
 Memory Model for Concurrent Operations