-enable-correct-eh-support very slow

[llvmbot]

Bugzilla Link	631
Resolution	FIXED
Resolved on	Mar 06, 2010 13:58
Version	1.0
OS	All
Reporter	LLVM Bugzilla Contributor

Extended Description

Enabling this options makes the exectable a couple of times slower for PyPy.
Currently it works best to use llc -march=c ... and then compile the c file with
gcc. The odd thing is that this c file also contains setjmp/longjmp. Seems gcc
(3.3.6) knows how to optimize this a lot better.

[llvmbot]

assigned to @lattner

[lattner]

Fundamentally, the problem is that C has no notion of exceptions, so invoke/unwind need to be
implemented in terms of setjmp/longjmp.

This PR is due to the fact that we're currently inserting one setjmp for each invoke instruction, which is
really slow in some cases. I have a plan to make it insert one setjmp call into each function that
contains an invoke, which should make functions like this:

void foo() {
while (...) {
invoke bar() ...
}
}

much much faster. I think I can even work a solution to Bug 240 into the fix as well.

Stay tuned.

-Chris

[lattner]

As an aside, this PR is why 253.perlbmk is so slow. It contains a much much larger chunk of code that
boils down to the pseudo code I pasted.

-Chris

[lattner]

Fixed. Patch here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20050926/028386.html

-Chris