[PATCH] D16836: [CodeGenPrepare] Don't transform select instructions into branches when both of operands are cheap

[Sanjay Patel via llvm-commits]

spatel added a comment.

F1532296: more_select_cgp - perf.csv <http://reviews.llvm.org/F1532296>

Here's test-suite raw data after removing the branch-on-load heuristic completely while running on AMD Jaguar 1.5 GHz. I ran each config 3 times. I'm filtering out anything with more than 1% relative std dev in a google spreadsheet while reviewing the numbers.

Like Junmo's SandyBridge experiment summary (it looks like the actual data is encrypted?), I don't see any real regressions.

There was a similar (+14%) win for MultiSource/Benchmarks/TSVC/Searching-flt/Searching-flt as Junmo reported testing with Cortex-A57, however, there was only a slight (+1.5%) win for MultiSource/Benchmarks/TSVC/Searching-dbl/Searching-dbl.

The average perf of the filtered tests is +0.2% after removing the branch-on-load transform.


http://reviews.llvm.org/D16836