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: [tree-ssa][PATCH]: Alias stats

But in terms of percentages, the queries vs resolved of each alias info provider here are average for the bootstrap of gcc (IE pta resolves about 80% of queries thrown at it, tbaa, about 50% thrown at it).

I'll be working the PTA results into other info soon (like using PTA to determine if a variable accesses global memory, or is actually call clobbered), which hopefully should increase our spec scores some.

I forgot to mention. PTA seems to do even better on C++ than it does C (which it does pretty well).

For example, on gerald's testcase, some of the numbers are like this:

Total alias queries:    64054
Total alias mayalias results:   368
Total alias noalias results:    63686
Total simple queries:   64054
Total simple resolved:  91
Total pointer TBAA queries:     2347
Total pointer TBAA resolved:    25
Total TBAA queries:     63938
Total TBAA resolved:    14855
Total PTA queries:      49083
Total PTA resolved:     48715

That's 99.2% of the PTA queries returning no alias, where TBAA only got 23.2% of it's queries resolved to no alias.

Sometimes TBAA does better, but even then PTA still beats it percentage wise:

Total alias queries:    370012
Total alias mayalias results:   649
Total alias noalias results:    369363
Total simple queries:   370012
Total simple resolved:  193
Total pointer TBAA queries:     87699
Total pointer TBAA resolved:    27
Total TBAA queries:     369792
Total TBAA resolved:    289036
Total PTA queries:      80756
Total PTA resolved:     80107

(TBAA resolved 78%, PTA 99.2)

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