From: Martin Sebor <msebor@gmail.com>
To: Jeff Law <law@redhat.com>, gcc-patches <gcc-patches@gcc.gnu.org>
Subject: Re: [RFA] [tree-optimization/80576] Handle non-constant sizes in DSE
Date: Fri, 16 Aug 2019 22:49:00 -0000 [thread overview]
Message-ID: <845c672d-6d48-69a9-3210-eafb5e3b374b@gmail.com> (raw)
In-Reply-To: <72ac1813-0ca3-6798-a77b-7f1ab914d91c@redhat.com>
On 8/16/19 12:15 PM, Jeff Law wrote:
> On 8/16/19 12:09 PM, Marc Glisse wrote:
>> On Fri, 16 Aug 2019, Jeff Law wrote:
>>
>>> This patch improves our ability to detect dead stores by handling cases
>>> where the size memcpy, memset, strncpy, etc call is not constant. This
>>> addresses some, but not all, of the issues in 80576.
>>>
>>> The key here is when the size is not constant we can make conservative
>>> decisions that still give us a chance to analyze the code for dead
>>> stores.
>>>
>>> Remember that for dead store elimination, we're trying to prove that
>>> given two stores, the second store overwrites (partially or fully) the
>>> same memory locations as the first store. That makes the first store
>>> either partially or fully dead.
>>>
>>> When we encounter the first store, we set up a bitmap of bytes written
>>> by that store (live_bytes). We then look at subsequent stores and clear
>>> the appropriate entries in the bitmap.
>>>
>>> If the first store has a nonconstant length argument we can use the
>>> range of the length argument (max) and the size of the destination
>>> object to make a conservative estimation of how many bytes are written.
>>>
>>> For the second store the conservative thing to do for a non-constant
>>> length is to use the minimum of the range of the length argument.
>>
>> So I guess it won't handle things like
>>
>> void f(char*p,int n){
>> Â __builtin_memset(p,3,n);
>> Â __builtin_memset(p,7,n);
>> }
>>
>> where we know nothing about the length, except that it is the same? Or
>> do you look at symbolic ranges?
> Nope. I think ao_ref can represent that, so it'd just be a matter of
> recording "n" as the length, then verifying that the second call's
> length is "n" as well. That makes the first call dead. We'd have to
> bypass the byte tracking in that case, but I think that's trivial
> because we already have a means to do that when the sizes are too large.
>
>>
>>> This doesn't come up a lot in practice. But it also happens to put some
>>> of the infrastructure in place to handle strcpy and strcpy_chk which are
>>> needed to fully resolve 80576.
>>>
>>> Bootstrapped and regression tested on x86, x86_64, ppc64le, ppc64,
>>> ppc32, aarch64, sparc, s390x and probably others. Also verified that
>>> the tests work on the various *-elf targets in my tester.
>>>
>>> OK for the trunk?
>>
>> ENOPATCH
> Opps. Attached.
It's an improvement and I realize you said it doesn't handle
everything and that you don't think it comes up a lot, but...
I would actually expect the following example (from the bug)
not to be that uncommon:
void g (char *s)
{
char a[8];
__builtin_strcpy (a, s);
__builtin_memset (a, 0, sizeof a);
f (a);
}
or at least to be more common than the equivalent alternative
the improvement does optimize:
void g (char *s)
{
char a[8];
__builtin_memcpy (a, s, __builtin_strlen (s));
__builtin_memset (a, 0, 8);
f (a);
}
It seems that making the first one work should be just a matter
of handling strcpy analogously (the string length doesn't matter).
As an aside, the new tests make me realize that -Wstringop-overflow
should be enhanced to detect this problem (i.e., a consider
the largest size in a PHI).
Martin
next prev parent reply other threads:[~2019-08-16 22:21 UTC|newest]
Thread overview: 21+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-08-16 18:10 Jeff Law
2019-08-16 19:55 ` Marc Glisse
2019-08-16 20:41 ` Jeff Law
2019-08-16 22:49 ` Martin Sebor [this message]
2019-08-22 0:30 ` Jeff Law
2019-08-22 18:50 ` Martin Sebor
2019-08-23 16:50 ` Jeff Law
2019-08-19 14:23 ` Richard Biener
2019-08-22 2:12 ` Jeff Law
2019-08-22 11:14 ` Richard Biener
2019-08-23 20:27 ` Jeff Law
2019-08-26 10:07 ` Richard Biener
2019-09-03 21:24 ` Jeff Law
2019-09-09 20:10 ` Jeff Law
2019-09-16 9:12 ` Richard Biener
2019-09-16 9:18 ` Richard Biener
2019-08-22 15:53 ` Martin Sebor
2019-08-23 16:50 ` Jeff Law
2019-08-16 21:50 ` Jeff Law
2019-08-16 22:19 ` Marc Glisse
2019-08-16 22:43 ` Jeff Law
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=845c672d-6d48-69a9-3210-eafb5e3b374b@gmail.com \
--to=msebor@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=law@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).