r344957 - Give Multiversion-inline functions linkonce linkage

Mon Oct 29 11:46:09 PDT 2018

From: David Blaikie [mailto:dblaikie at gmail.com]
Sent: Monday, October 29, 2018 11:41 AM
To: Keane, Erich <erich.keane at intel.com>
Cc: Eric Christopher <echristo at gmail.com>; cfe-commits at lists.llvm.org
Subject: Re: r344957 - Give Multiversion-inline functions linkonce linkage


On Mon, Oct 29, 2018 at 11:30 AM Keane, Erich <erich.keane at intel.com<mailto:erich.keane at intel.com>> wrote:
GCC actually doesn’t support function multiversioning in C mode, so this fix is part of our extension to support C with multiversioning.

Ah, what's the motivation for that?
[Keane, Erich] Basically, I identified no good reason to not support it.  GCC’s motivation is lost to time as far as I can tell.  This multiversioning is super useful in my employer’s math/intrinsic libraries (MKL uses this feature extensively), so enabling it in C seemed like a positive improvement.  We’ve made some other behavioral changes to make ‘target’ work more sanely to suppress some of the worst bugs (becoming MV after usage) as well as make it work better with Clang’s emit-functionality.


  I perhaps wonder if this is part of the reason GCC only supports it in C++ mode...

Yeah... could be.

I mean I suppose using linkonce isn't making multiversioning any worse in C than it already is in C++ inline functions (which is where it's pretty easy to misuse, as I understand it - two files, one compiled with a CPU feature, one without, linkonce functions get linked together & the linker picks one - then code built without the feature calls code that uses the feature - but, I mean, that can happen even witohut multiversioning, if you compile some source files with a feature and some without) - I guess the main risk would be if it confuses users by making multiversioned C inline functions end up behaving too differently from non-multiversioned ones...
[Keane, Erich] I suspect that multiversioning is such a sharp edged feature, that a change like this makes it a touch more sane.  Since the individual versions are independently mangled, you don’t have to worry about merging different versions.  The only real issue comes from merging the resolver functions if the visible list of declarations is different, but we treat that as an ODR violation.


- Dave


From: David Blaikie [mailto:dblaikie at gmail.com<mailto:dblaikie at gmail.com>]
Sent: Monday, October 29, 2018 11:25 AM
To: Keane, Erich <erich.keane at intel.com<mailto:erich.keane at intel.com>>; Eric Christopher <echristo at gmail.com<mailto:echristo at gmail.com>>
Cc: cfe-commits at lists.llvm.org<mailto:cfe-commits at lists.llvm.org>
Subject: Re: r344957 - Give Multiversion-inline functions linkonce linkage

Does this match GCC's approach here?

(I ask this sort of as throwaway/conversation starter - because the linkage/behavior around multiversion functions and their inlining is full of sharp corners/risks of code moving out of the areas appropriately restricted based on the cpu features)
On Mon, Oct 22, 2018 at 2:22 PM Erich Keane via cfe-commits <cfe-commits at lists.llvm.org<mailto:cfe-commits at lists.llvm.org>> wrote:
Author: erichkeane
Date: Mon Oct 22 14:20:45 2018
New Revision: 344957

URL: http://llvm.org/viewvc/llvm-project?rev=344957&view=rev
Log:
Give Multiversion-inline functions linkonce linkage

Since multiversion variant functions can be inline, in C they become
available-externally linkage.  This ends up causing the variants to not
be emitted, and not available to the linker.

The solution is to make sure that multiversion functions are always
emitted by marking them linkonce.

Change-Id: I897aa37c7cbba0c1eb2c57ee881d5000a2113b75

Modified:
    cfe/trunk/lib/CodeGen/CodeGenModule.cpp
    cfe/trunk/test/CodeGen/attr-target-mv.c

Modified: cfe/trunk/lib/CodeGen/CodeGenModule.cpp
URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/CodeGen/CodeGenModule.cpp?rev=344957&r1=344956&r2=344957&view=diff
==============================================================================

--- cfe/trunk/lib/CodeGen/CodeGenModule.cpp (original)
+++ cfe/trunk/lib/CodeGen/CodeGenModule.cpp Mon Oct 22 14:20:45 2018
@@ -3669,6 +3669,10 @@ llvm::GlobalValue::LinkageTypes CodeGenM
       return llvm::GlobalVariable::WeakAnyLinkage;
   }

+  if (const auto *FD = D->getAsFunction())
+    if (FD->isMultiVersion() && Linkage == GVA_AvailableExternally)
+      return llvm::GlobalVariable::LinkOnceAnyLinkage;
+
   // We are guaranteed to have a strong definition somewhere else,
   // so we can use available_externally linkage.
   if (Linkage == GVA_AvailableExternally)

Modified: cfe/trunk/test/CodeGen/attr-target-mv.c
URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/test/CodeGen/attr-target-mv.c?rev=344957&r1=344956&r2=344957&view=diff
==============================================================================
--- cfe/trunk/test/CodeGen/attr-target-mv.c (original)
+++ cfe/trunk/test/CodeGen/attr-target-mv.c Mon Oct 22 14:20:45 2018
@@ -88,19 +88,19 @@ void bar4() {

 // CHECK: declare i32 @foo.arch_sandybridge()

-// CHECK: define available_externally i32 @foo_inline.sse4.2()
+// CHECK: define linkonce i32 @foo_inline.sse4.2()
 // CHECK: ret i32 0

 // CHECK: declare i32 @foo_inline.arch_sandybridge()
 //
-// CHECK: define available_externally i32 @foo_inline.arch_ivybridge()
+// CHECK: define linkonce i32 @foo_inline.arch_ivybridge()
 // CHECK: ret i32 1
-// CHECK: define available_externally i32 @foo_inline()
+// CHECK: define linkonce i32 @foo_inline()
 // CHECK: ret i32 2

-// CHECK: define available_externally void @foo_decls()
-// CHECK: define available_externally void @foo_decls.sse4.2()
+// CHECK: define linkonce void @foo_decls()
+// CHECK: define linkonce void @foo_decls.sse4.2()

-// CHECK: define available_externally void @foo_multi.avx_sse4.2()
-// CHECK: define available_externally void @foo_multi.fma4_sse4.2()
-// CHECK: define available_externally void @foo_multi.arch_ivybridge_fma4_sse4.2()
+// CHECK: define linkonce void @foo_multi.avx_sse4.2()
+// CHECK: define linkonce void @foo_multi.fma4_sse4.2()
+// CHECK: define linkonce void @foo_multi.arch_ivybridge_fma4_sse4.2()


_______________________________________________
cfe-commits mailing list
cfe-commits at lists.llvm.org<mailto:cfe-commits at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20181029/7fe1918d/attachment.html>
