[passmgr] Eliminate use of typeid/dynamic_cast

[lattner]

Bugzilla Link	888
Resolution	FIXED
Resolved on	Jan 22, 2010 00:36
Version	1.0
OS	All

Extended Description

The passmgr is the only significant user of RTTI data. If we could eliminate its use, we could build all of
LLVM with -fno-rtti, which is good for a 5-6% size reduction. This would also provide (slightly) faster
builds for everyone! :)

-Chris

[llvmbot]

The new PassManager continues to use typeid:

Pass.h: return lookupPassInfo(typeid(AnalysisClass));
PassSupport.h: : RegisterPassBase(Name, PassArg, typeid(PassName),
PassSupport.h: : RegisterAGBase(typeid(Interface),
&RPB.getPassInfo()->getTypeInfo(),
PassSupport.h: : RegisterAGBase(typeid(Interface)) {

and dynamic_cast:
PassAnalysisSupport.h: return dynamic_cast<AnalysisType*>
PassAnalysisSupport.h: // AnalysisGroups), we must use dynamic_cast here to
potentially adjust the
PassAnalysisSupport.h: AnalysisType Result =
dynamic_cast<AnalysisType>(ResultPass);
PassAnalysisSupport.h: // AnalysisGroups), we must use dynamic_cast here to
potentially adjust the
PassAnalysisSupport.h: AnalysisType Result =
dynamic_cast<AnalysisType>(ResultPass);

It would be nice if we could get rid of these and start turning on -no-rtti.

[llvmbot]

Devang,

Any chance this PR (888) could get implemented for 2.0?

Reid

[llvmbot]

I have not yet thought much about this. IMO, it is impractical to use isa,classof approach used by Value
and Type classes because it will require each pass to identify itself.

thoughts ? ideas ? suggestions ?

[llvmbot]

Every Pass already identifies itself when it registers. I don't see it as being
a big deal.

[lattner]

One suggested implementation approach: Whenever a pass registration object's static ctor triggers, it
could assign a unique ID to the pass. The passmgr could use this.

Another approach: we could use the address of the pass registration object as a unique ID.

[llvmbot]

I eliminated typeid uses.

[llvmbot]

Cool! Thanks.

When I get a chance to test it, I'll turn typeid support off in the Makefiles
and see if we're "clean" without it. If so, I'll resolve this PR.

Thanks, Devang!

Reid

[llvmbot]

Thanks Reid. I still need to deal with two remaining dynamic_cast in PassAnalysisSupport.h though.

[llvmbot]

Okay, I'll watch for the commits. :)

[llvmbot]

FYI, I turned off typeid support when building llvm-gcc and all hell broke
loose. It seems there are many other subtle uses. I'm just going on memory right
now but will provide details (or a fix) next week.

[lattner]

Passmanager no longer uses RTTI.