llvm dev - Jan 2019 - Textual IR value names

The names are dropped to save memory when a release build of the compiler
is being used. This is what you probably want on a release compiler you
intend to ship since it should be faster. The NDEBUG check is an easy way
to tell the difference between release and debug builds. People probably
don't want to have to remember to set an additional cmake option to make a
release compiler faster.

The -fdiscard-value-names/-fno-discard-value-names was added about a year
ago https://reviews.llvm.org/D42887  not sure if there is more discussion
about the asserts behavior discussed there.

~Craig


On Wed, Jan 9, 2019 at 2:25 PM David Greene <dag at cray.com> wrote:
> Craig Topper via llvm-dev <llvm-dev at lists.llvm.org> writes:
>
> > And the clang behavior can be controlled with -
> > fdiscard-value-names/-fno-discard-value-names
>
> That's a little better for users, but again, why should the default
> behavior change based on whether or not asserts are enabled?
>
>                           -David
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190109/803a322c/attachment.html>

Also, the textual IR is really a developer/debugging tool, not something an
end-user would be expected to look at, so the textual names would be largely
pointless in a release build.
If you are writing lit tests that look at textual IR, and they are failing with
release builds, you can add the "REQUIRES: asserts" directive to the
lit test so that the test won't be run for a release build.
--paulr

From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of Craig
Topper via llvm-dev
Sent: Wednesday, January 09, 2019 5:41 PM
To: David Greene
Cc: llvm-dev at lists.llvm.org
Subject: Re: [llvm-dev] Textual IR value names

The names are dropped to save memory when a release build of the compiler is
being used. This is what you probably want on a release compiler you intend to
ship since it should be faster. The NDEBUG check is an easy way to tell the
difference between release and debug builds. People probably don't want to
have to remember to set an additional cmake option to make a release compiler
faster.

The -fdiscard-value-names/-fno-discard-value-names was added about a year ago
https://reviews.llvm.org/D42887  not sure if there is more discussion about the
asserts behavior discussed there.

~Craig


On Wed, Jan 9, 2019 at 2:25 PM David Greene <dag at cray.com<mailto:dag at
cray.com>> wrote:
Craig Topper via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at
lists.llvm.org>> writes:
> And the clang behavior can be controlled with -
> fdiscard-value-names/-fno-discard-value-names
That's a little better for users, but again, why should the default
behavior change based on whether or not asserts are enabled?

                          -David
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190110/6e3410d6/attachment.html>

Craig Topper via llvm-dev <llvm-dev at lists.llvm.org> writes:
> The names are dropped to save memory when a release build of the
> compiler is being used. This is what you probably want on a release
> compiler you intend to ship since it should be faster. The NDEBUG
> check is an easy way to tell the difference between release and debug
> builds. People probably don't want to have to remember to set an
> additional cmake option to make a release compiler faster.
The CMake option could override the current behavior.  Then users
wouldn't be forced to remember to set an option to make release
compilers faster.

In the end it's not a huge deal, I just found keying off asserts to be
quite surprising and it took some time for me to figure out what was
going on.

                            -David

via llvm-dev <llvm-dev at lists.llvm.org> writes:
> Also, the textual IR is really a developer/debugging tool, not
> something an end-user would be expected to look at, so the textual
> names would be largely pointless in a release build.
Our customers using our (non-LLVM-based) Fortran compiler have the
ability to dump internal IR(-ish) and examine it.  Some find it useful
to understand how the compiler transforms their code.  So it's not quite
right to say such dumps are pointless.  It depends on the user.

                          -David

While I understand the value of turning it off, it is somewhat annoying when
you're trying to run a clang -S -emit-llvm to discover that you're not
getting any value names. I would suggest that, irrespective of the kind of build
you use, -S -emit-llvm should cause the value names to be defaulted to on as a
principle of least surprise to the user.

-----Original Message-----
From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of David
Greene via llvm-dev
Sent: Thursday, January 10, 2019 15:05
To: Craig Topper <craig.topper at gmail.com>
Cc: llvm-dev at lists.llvm.org
Subject: Re: [llvm-dev] Textual IR value names

Craig Topper via llvm-dev <llvm-dev at lists.llvm.org> writes:
> The names are dropped to save memory when a release build of the 
> compiler is being used. This is what you probably want on a release 
> compiler you intend to ship since it should be faster. The NDEBUG 
> check is an easy way to tell the difference between release and debug 
> builds. People probably don't want to have to remember to set an 
> additional cmake option to make a release compiler faster.
The CMake option could override the current behavior.  Then users wouldn't
be forced to remember to set an option to make release compilers faster.

In the end it's not a huge deal, I just found keying off asserts to be quite
surprising and it took some time for me to figure out what was going on.

                            -David
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev