[patch] Produce .weak_def_can_be_hidden for some linkonce_odr values

[Nick Lewycky]

On 28 October 2013 19:56, Rafael Espíndola <rafael.espindola at gmail.com>wrote:

> Ping.
>

-      if (Linkage != GlobalValue::LinkOnceODRAutoHideLinkage)
+
+      bool CanBeHidden = Linkage ==
GlobalValue::LinkOnceODRAutoHideLinkage;
+
+      if (!CanBeHidden && Linkage == GlobalValue::LinkOnceODRLinkage) {
+        if (GV->hasUnnamedAddr()) {
+          CanBeHidden = true;
+        } else {
+          GlobalStatus GS;
+          if (!GlobalStatus::analyzeGlobal(GV, GS) && !GS.IsCompared)

Ah, really? Put another way, why should we pay this penalty *here*? This is
something we can compute in the mid-level optimizer and should compute in
the mid-level optimizer, and we shouldn't waste our time second-guessing it
here.

I assume the real problem is that we don't do a second run of globalopt at
the end of the TU to mark more things with unnamed_addr? How much do we
lose if you don't do this recomparison, is that tolerable? I realize that
getting a second run of globalopt at the end is going to be time consuming
to prove worthwhile, but I really think it's the right way to go.

+            CanBeHidden = true;
+        }
+      }
+
+      if (!CanBeHidden)

Everything else about the patch LGTM.

I just tested this with the linker in Xcode 5 and the results are:
>
> $ nm  build-with-master/bin/clang-3.4  | grep  ' T ' | wc
>
>    21053 63159 1790901
>
> $ nm  build-with-patch/bin/clang-3.4  | grep  ' T ' | wc
>
>    19049   57147 1547425
>

Nice!!

Nick
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20131028/43c0f989/attachment.html>