Re: [PATCH v6 1/9] ppc64 (le): prepare for -mprofile-kernel
From: AKASHI Takahiro
Date: Thu Feb 04 2016 - 04:32:04 EST
Jiri, Torsten
Thank you for your explanation.
On 02/03/2016 08:24 PM, Torsten Duwe wrote:
On Wed, Feb 03, 2016 at 09:55:11AM +0100, Jiri Kosina wrote:
On Wed, 3 Feb 2016, AKASHI Takahiro wrote:
those efforts, we are proposing[1] a new *generic* gcc option, -fprolog-add=N.
This option will insert N nop instructions at the beginning of each function.
The interesting part of the story with ppc64 is that you indeed want to
create the callsite before the *most* of the prologue, but not really :)
I was silently assuming that GCC would do this right on ppc64le; add the NOPs
right after the TOC load. Or after TOC load and LR save? ...
On arm/arm64, link register must be saved before any function call. So anyhow
we will have to add something, 3 instructions at the minimum, like:
save lr
branch _mcount
restore lr
<prologue>
...
<body>
...
The part of the prologue where TOC pointer is saved needs to happen before
the fentry/profiling call.
Yes, any call, to any profiler/tracer/live patcher is potentially global
and needs the _new_ TOC value.
I don't want to bother you, but for my better understandings, could you show me
an example of asm instructions for a function prologue under -mprofile-kernel, please?
-Takahiro AKASHI
This proposal, if implemented in a too naive fashion, will worsen the problem
we currently discuss: a few NOPs _never_ cause any global reference. GCC might
be even more inclined to not load a new TOC value. That change would need to be
fairly smart on ppc64le.
Torsten