Re: Early inlining

[Josh Conner <jconner at apple dot com>]

On Jun 23, 2005, at 7:37 PM, Jan Hubicka wrote:

> Hi,
> this patch adds early inlining.  As discussed earlier it helps
> noticeably to cut down expenses of tree profiling on testcases with
> extreme function call overhead (such as tramp3d).  It also save memory
> on Gerald's testcase and makes GCC module compilation test apparently
> tinny bit faster at -O3 (but close to noise).  In the future I would
> like to have more optimizations in between the early inlining and real
> inliing that will make this infrastructure bit more usefull, but at
> the moment it seems to solve some side cases and have no measurable
> overhead otherwise (even tought we rebuild cgraph edges after the  
> early
> passes) so I am enabling it by default now.
...

Jan -

After this patch, I see that static functions called once that are  
inlined are no longer removed from the CFG during actual (not early)  
inlining.  Assuming this is the intended behavior, there is a slight  
issue that is raised in the function cgraph_decide_inlining().  This  
can be seen by compiling the attached test.c with "-O3 -fdump-ipa- 
all", producing a sefault.  The problem is that when inlining  
functions called once (starting around ipa-inline.c:809), we call  
cgraph_mark_inline() which creates a duplicate callee node, resetting  
node->callers to NULL.  With dumps enabled, this causes a null- 
pointer reference immediately after.  The attached patch addresses  
the issue by holding on to the edge across the call to  
cgraph_mark_inline.  I haven't run any regression tests on this yet,  
but I wanted to be sure that the original behavior of no longer  
removing CFG nodes was the intended behavior first.

Shall I kick off the tests?

- Josh
~~~~~~
Josh Conner

Attachment: test.c
Description: Binary data

Attachment: inl-dump.patch
Description: Binary data