[PATCH] D43093: [FastISel] Sink local value materializations to first use

[Reid Kleckner via Phabricator via llvm-commits]

rnk added a comment.

I did some timings on the sqlite3 amalgamation, and I found this adds about 0.13% to compile time. I was hoping it wouldn't be measurable, but there is a repeatable 0.13% difference in the total number of instructions retired reported by the linux perf tool.

I think we should probably remove the cl::opt in this patch and then land it more or less as is. It improves the -O0 debugging experience. If speeding up X86 ISel becomes a priority at some point, I think we'll want to spend that effort working on X86 GlobalISel so that our optimizations help -O0 and -O2 codegen.


https://reviews.llvm.org/D43093