rnk added a comment.
I think this is a reasonable stop-gap fix since the code isn't trying to return
EAX:EDX or XMM0 from the inline asm blob. This affects any function that
contains inline asm and returns a vector, which is potentially a lot of stuff.
================
Comment at: test/CodeGen/pr34021.c:2
+// RUN: %clang_cc1 -fms-extensions %s -triple=i686-unknown-unknown -emit-llvm
-o -
+// RUN: %clang_cc1 -fms-extensions %s -triple=x86_64-unknown-unknown
-emit-llvm -o -
+// REQUIRES: asserts
----------------
Please FileCheck the IR. We should see a pattern like this:
```
define <4 x float> @rep()
%retval = alloca <4 x float>
call i64 inlineasm sideeffect {{.*}}
store {{.*}}, <4 x float>* %retval
%v = load {{.*}}, <4 x float>* %res
store <4 x float> %v, <4 x float>* %retval
%rv = load <4 x float>* %retval
ret <4 x float> %rv
```
That IR pattern makes it clear that the store of the asm output is immediately
killed by the store of `res` to `retval`.
Repository:
rL LLVM
https://reviews.llvm.org/D37448
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
