[PATCH] D13363: [DeadStoreElimination] Add support for non-local DSE

[Ivan Baev via llvm-commits]

ivanbaev added a comment.

Hi Bruno,

Thanks for running and sharing x86 results! Deltas look reasonable, in my opinion. Removing dead stores and related instructions across basic blocks should be a good thing by itself. There is a possibility for second order effects - e.g. longer live ranges, increased number of spills.

Results for SPEC2006/sjeng are interesting. On Arm/AArch64/A57 devices at -O2 we have the following run times (in seconds, so lower is better)

- "-mllvm -enable-nonlocal-dse=false": 1795, 1799, 1801, 1804, 1805, 1806
- "-mllvm -enable-nonlocal-dse=true":  1773, 1774,1775, 1778, 1778, 1786

At -O3, we do not see any performance delta for sjeng. What optimization level did you use for compiling SPEC on x86?

Using -stats, for both -O2 and -O3. on sjeng we are able to delete 13 non-local stores in total (found in 4 source files). No changes in the number of regalloc spills and other relevant stats other than decreases in machine instructions emitted and printed and isl dag tries (see below).

Glad that the DSE compile-time is under control, and actually improved for one Trimaran benchmark.
On sjeng, there is a decrease in isel - Number of times dag isel has to try another path - from 2380 to 2348 with this patch. How is this related to compile-time?

It would be good to hear what reviewers think about run- and compile-time results on x86 and AArch64 with this patch, and if we can commit it.

Best regards,
Ivan


http://reviews.llvm.org/D13363