This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: C++ PATCH: Fix PR 8338

I might have a different fix for this. The fix involves allocating 2 bits on type nodes, one meaning uses_template_params, and the other meaning doesn't use template params, I store these bits on types (the recursive case). It is way better, and way faster than your fix.

You can preview the work in:

Attachment: template-3.diffs
Description: Binary data

These give a 10x speedup on the Abraham's template torture test case at:

for N==200. The changes speed up any template beefy code by large amounts. Once you allocate the bits for this use, there isn't a need for the hashtable you have. I believe this information is static in the tree. Let me know if any of it is not. A hashtable doesn't give the speed advantages using 2 bits gives. Your change I suspect slows down the compiler, instead of speeding it up, well, except of course in the infinite case.

If you want, rip the 2 bit solution out of the patches, and try that out. I hate to use the bits, but if you examine just how often we ask that question, and how expensive that tree walk is (very, very, very expensive), you'd realize that the cost of using those two bits is worth it.

If you want to test and put any of the other work in feel free, if not, I'll finish testing it up and submit it soon.

On Thursday, November 7, 2002, at 05:35 PM, Mark Mitchell wrote:
This patch fixes PR 8338, a regression that caused the compiler to go
into an infinite loop.  The fix may actually speed up the compiler a
touch on some programs by avoiding walking the same nodes more than
once in for_each_template_parm.

A bigger speedup will be to use that function in dramatically fewer
situations; coming once the new parser gets checked in.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]