[llvm] r243253 - Roll forward r243250

[Jingyue Wu via llvm-commits]

FYI, searching the code base for getUserCost, I noticed getGEPCost is used
in not only inlining but also loop unrolling and SimplifyCFG to compute
speculation cost.

On Wed, Sep 9, 2015 at 1:51 PM, Hans Wennborg <hans at chromium.org> wrote:

> Ah, I didn't realize this feeds into the inlining cost. That makes
> sense. Thanks.
>
> On Wed, Sep 9, 2015 at 1:45 PM, Jingyue Wu <jingyue at google.com> wrote:
> > I guess this is because more functions are inlined.
> >
> > This patch changes (improves IMO) how the cost of GEPs is computed. If a
> GEP
> > can be complete folded into an addressing mode, TTI reports its cost is
> > zero. Therefore, the cost of a function tends to be smaller than before,
> > causing more functions to be inlined and thus increasing the binary size.
> >
> > On Wed, Sep 9, 2015 at 1:26 PM, Hans Wennborg <hans at chromium.org> wrote:
> >>
> >> (cc'ing the new list; sorry for the duplicate email)
> >>
> >> On Wed, Sep 9, 2015 at 1:24 PM, Hans Wennborg <hans at chromium.org>
> wrote:
> >> > On Sun, Jul 26, 2015 at 12:10 PM, Jingyue Wu <jingyue at google.com>
> wrote:
> >> >> Author: jingyue
> >> >> Date: Sun Jul 26 14:10:03 2015
> >> >> New Revision: 243253
> >> >>
> >> >> URL: http://llvm.org/viewvc/llvm-project?rev=243253&view=rev
> >> >> Log:
> >> >> Roll forward r243250
> >> >>
> >> >> r243250 appeared to break clang/test/Analysis/dead-store.c on one of
> >> >> the build
> >> >> slaves, but I couldn't reproduce this failure locally. Probably a
> false
> >> >> positive as I saw this test was broken by r243246 or r243247 too but
> >> >> passed
> >> >> later without people fixing anything.
> >> >
> >> > This caused a 150 KB binary size increase in Chromium. Is that
> expected?
> >> >
> >> > I'm not familiar with this code, but from my reading of the change
> >> > description, it sounds more like it should decrease size by folding
> >> > more address computations into mov instructions?
> >
> >
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150909/d394e4c2/attachment.html>