[PATCH] D11977: Use modulo semantic to generate non-integer-overflow assumptions

[Mailing List "llvm-commits" via llvm-commits]

llvm-commits added a comment.

Attached 2 json files, without_overflow_assumptions.json
and with_overflow_assumptions.json. The former is a recent upstream
commit with (I think) 4 lnt runs, the latter the current integer
wrapping patch applied to the upstream head with 2 lnt runs.

We have 5 benchmarks that have more than 15% compile time regression:

  ; regress ; prev   ; Current;   σ   

MultiSource/Applications/SIBsim4/SIBsim4         ; 136.24% ; 2.5200 ; 5.9533 ; 0.2801 
 MultiSource/Benchmarks/ASC_Sequoia/IRSmk/IRSmk   ; 21.86%  ; 0.5334 ; 0.6500 ; 0.0016 
 SingleSource/Regression/C++/2003-05-14-expr_stmt ; 16.67%  ; 0.1200 ; 0.1400 ; -      
 MultiSource/Applications/oggenc/oggenc           ; 16.34%  ; 7.0400 ; 8.1900 ; 0.2633 
 SingleSource/Regression/C++/EH/dead_try_block    ; 15.38%  ; 0.1300 ; 0.1500 ; 0.0016

Also the execution time regressed by more than 15% a few times:

                                                                             ;    Δ   ; Previous;  Current;   σ
  SingleSource/Benchmarks/Polybench/stencils/fdtd-apml/fdtd-apml             ; 50.00% ; 0.3400  ; 0.5100  ; 0.0016
  SingleSource/Benchmarks/Polybench/linear-algebra/kernels/trisolv/trisolv   ; 35.27% ; 0.0567  ; 0.0767  ; -     
  SingleSource/Benchmarks/Shootout-C++/strcat                                ; 33.33% ; 0.0300  ; 0.0400  ; -     
  MultiSource/Benchmarks/Olden/treeadd/treeadd                               ; 24.01% ; 0.0833  ; 0.1033  ; 0.0050
  SingleSource/Benchmarks/Polybench/linear-algebra/kernels/bicg/bicg         ; 22.22% ; 0.0900  ; 0.1100  ; 0.0016
  SingleSource/Benchmarks/Polybench/linear-algebra/kernels/symm/symm         ; 19.73% ; 12.7233 ; 15.2333 ; 0.2967
  SingleSource/Benchmarks/Polybench/linear-algebra/kernels/cholesky/cholesky ; 18.05% ; 0.4433  ; 0.5233  ; 0.0050
  SingleSource/Benchmarks/Polybench/linear-algebra/kernels/gesummv/gesummv   ; 16.70% ; 0.1000  ; 0.1167  ; 0.0016

However I am not sure how much, especially of the runtime results, is
real or only some noise.

All in all I think this is as good as it gets for now.

- F746084: msg-3916-965.dat <http://reviews.llvm.org/F746084>
- F746083: msg-3916-966.txt <http://reviews.llvm.org/F746083>


http://reviews.llvm.org/D11977