Disabling RTTI

[hs-apotell]

RTTI is enabled by default for both gcc & cl compiler. Also, dynamic_cast is being used across the codebase. Use of RTTI at library forces enabling RTTI at the program level which, ideally, we would like to disable it for number of reasons including performance and size.

There are options that can be used to restrict RTTI information to only the classes that need it. Here's an implementation. I am endorsing this implementation but just presenting one such option.

Any thoughts on implementing custom RTTI rather than forcing it to be enabled at program level?

P.S. We use CopperSpice (a fork of Qt) for our UI and the output binary sizes grow by a factor of 2.8 times by enabling RTTI at program level. This is not an ideal solution but UDHM/Surelog is the only library we use that forces enabling RTTI.

[alaindargelas]

Antlr is using it all over the place. No chance they'll change the code.
My fast version of Antlr did replace it manually in the most common places by bit fields. That works, it is painful but can be done.
On the other hand you are using the plain vanilla Antlr aren't you? So Surelog is not the only library that uses RTTI in your project.

[hzeller]

Could we just link RTTI inside the ANTLR library the the parser code it generates but switch it off for other object files ?
Of course the question is, outside of ANTLR, how feasible is it to not use RTTI in the rest of Surelog ?

Looks like there are only

find src -type f | grep -v /parser/ | xargs grep dynamic_cast | wc -l
223

I223 places, some UHDM, some other, they could probably be eliminated (also: exceptions outside the code that has to deal with Antlr that throws exceptions).

[hs-apotell]

On the other hand you are using the plain vanilla Antlr aren't you? So Surelog is not the only library that uses RTTI in your project.

No we are using the updated Antlr. The issue was that we wanted to use the updated version but it was part of Surelog and that meant pulling the whole repository which wasn't ideal. We just wanted Antlr pulled out of Surelog so dependencies are more explicit and clear.

Antlr is using it all over the place.

We will go thru' and update that too. We are looking at this very seriously because the generated code size differences. On Windows, the binaries file size differences are too big to be ignored.

If this group can agree on the need for this change and which specific RTTI implementation to use, we will take on the charge to make this change across the board.

[alaindargelas]

I'm OK with it, we might get more speed in Antlr if you use the same bitfield technique I used, if you use the kind of library you are pointing to, I'm not sure there will be speedups.
Now Antlr does use dynamic_casts in the generated code. You can't modify that by hand, every time we modify the grammar, you'd have to reedit the files under Surelog/src/parser/*.cc, that is not practical. You'd have to modify the Antlr parser generator or apply a script that modifies the generated parser as a post-processing.

[hzeller]

Before considering something that mimicks RTTI with some other library (which would always be a weird crutch), I'd consider not using RTTI in our code at all and only compile RTTI into the object files that deal with types (e.g. Antlr and its generated parser).

Is the codesize on windows dependent on if there is RTTI used at all, or is it proportional to the number of classes that need it ?
Can the linker find out at link time that an RTTI identifier is ever used on a class-hierarchy and strip it ?

[hs-apotell]

if you use the kind of library you are pointing to, I'm not sure there will be speedups.

The library I shared was just an example implementation. Here's performance comparison of using dynamic_cast vs. type_id solution.

You'd have to modify the Antlr parser generator or apply a script that modifies the generated parser as a post-processing.

We will evaluate the scope of work required to eliminate use of dynamic_cast all together. The work surely needs to be iterative.

I'd consider not using RTTI in our code at all and only compile RTTI into the object files that deal with types

AFAIK if any object file is compiled with RTTI information enabled, the linker expects it to be enabled all types in the final program. I don't know of a way to disable this.

Is the codesize on windows dependent on if there is RTTI used at all, or is it proportional to the number of classes that need it ?

Don't know off-and but can try and report back.

Can the linker find out at link time that an RTTI identifier is ever used on a class-hierarchy and strip it ?

That would be a kick-ass linker if it does all that reliably and consistently :)

[alaindargelas]

@hs-apotell any update/path to recoup the 24% runtime increase? If not we have to revert back to stock antlr (Not my fork) with RTTI in the main Surelog as the slowdown is not acceptable on Linux.
You can point to your RTTI-less Antlr implementation in your fork. For the antlr generated code in Surelog and the surelog code itself, I think we can use your RTTI replacement as the dynamic casts are not in the millions per run like in Antlr.
But first we need to recoup the 24% loss. Please update on your findings.

[hs-apotell]

Sorry, juggling multiple priorities. I haven't had chance to evaluate the performance of the solution just yet. As a stop gap solution, can we just point Surelog to use the master branch of Antlr instead of the modified Surelog branch. That would mean using pristine Antlr without the bitfield modifications. I will try get the RTTI version working in next few days. This is critical change for us as well which has to get done but at the immediate moment I need to clear my backlog so I can dedicate some exclusive time on this.

If you are comfortable with this stop gap solution, either drop a comment here so I can create the PR with necessary change or update the .gitmodules yourself.

[alaindargelas]

Yes please do create a PR with default main antlr

[alaindargelas]

OK, after merging the stock Antlr here are the runtime measurements:

Stock Antlr runs the regression in 1980s
Antlr with RTTI removed using @hs-apotell's technique runs in 1660s
My version with the bitfields runs in 1306s.

I'm adding back the bitfield version under third_party/antlr_fast and make the Surelog cmake point to it meanwhile @hs-apotell experiments with the stock antlr/RTTI-less antlr.
When I ran the experiment a couple of days ago, maybe I was never running the stock Antlr, maybe I was running the antlr with the RTTI removed.

But one correction is that @hs-apotell's version is faster than the stock antlr, just not as fast as the bitfields.

[alaindargelas]

@hs-apotell Let's stop using this fork: https:/alainmarcel/antlr4
Either make a fork directly under chipsalliance or propose a new location since you are the one experimenting with stock antlr.
Re-add the antlr subsystem in Surelog when you have a working version that has a performance not too far from the bitfield version.

[hs-apotell]

I don't think I have permissions to fork under chipsalliance umbrella.

Optionally, why don't you just transfer the repository to the chipsalliance organization? That would keep the history, issues, and everything intact.

[alaindargelas]

@mgielda do you think we can have an antlr fork under ChipsAlliance. That way Surelog/UHDM won't rely on an aobscure fork in my alainmarcel org.

[hs-apotell]

@alaindargelas When you get a moment could you please try PR #1825 (or optionally use alainmarcel/antlr4#4)

Build times from chipsalliance/Surelog builds, averaging 2558s

Build | Time
-------------
2020 | 2257s
2001 | 2628s
1993 | 2712s
1989 | 2627s
1983 | 2570s

Build time from hs-apotell/Surelog build

Build | Time
------------
99 | 2436s

I am going to kick off a few more runs of the change to get a better average.

[alaindargelas]

@hs-apotell can you give me clear instructions on how to pull a draft PR #1825? I'm reading many posts, I can't make head or tail...

[hs-apotell]

1. Go to antlr repository settings 2. At the bottom of that page, you will see an option called “Transfer Ownership”. Click that and provide the new owner’s git handle. 3. You should receive an email to the registered email address for chipsalliance group. 4. Further steps should be self explanatory. 5. To switch Surelog to the new repository address, update the url in .gitmodules file and then run the following commands ‘’’ git submodule sync third_party/antlr git submodule update third_party/antlr git add third_party/antlr git commit ‘’’

…

________________________________ From: alaindargelas ***@***.***> Sent: Saturday, September 11, 2021 11:19:58 AM To: chipsalliance/Surelog ***@***.***> Cc: Hitesh Shah ***@***.***>; Mention ***@***.***> Subject: Re: [chipsalliance/Surelog] Disabling RTTI (#1718) @hs-apotell<https:/hs-apotell> can you give me clear instructions on how to pull a draft PR #1825<#1825>? I'm reading many posts, I can't make head or tail... — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#1718 (comment)>, or unsubscribe<https:/notifications/unsubscribe-auth/AMFNMJ33DVEYM3KNDKBSCALUBOMU5ANCNFSM5C25XBAQ>. Triage notifications on the go with GitHub Mobile for iOS<https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android<https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[alaindargelas]

@hs-apotell, that was not the question. I still didn't get the OK to transfer the alainmarcel/antlr to chipsalliance. That is on standby.

In the meantime, I need to know how to checkout an unmerged pull request. Do you know how?

[alaindargelas]

Got it!
git clone https:/chipsalliance/Surelog
cd Surelog
git submodule update --init --recursive
git fetch origin pull/1825/head:test-rtti
git submodule update --init --recursive
git checkout test-rtti

[alaindargelas]

It's still 1700s vs 1300s for the bitfield version.
As long as this is used only on Windows, I'm OK. Submit a complete PR that uses the third_party/antlr_fast for Linux and Macos and uses the third_party/antlr for Windows only. You can use your RTTI thing in the Surelog code too. That will not make a difference.