Commit 001eb86

2017-04-04 Filip Pizlo <fpizlo@apple.com>

B3::fixSSA() needs a tune-up

https://bugs.webkit.org/show_bug.cgi?id=170485

Reviewed by Saam Barati.

After the various optimizations to liveness, register allocation, and other phases, the

fixSSA() phase now looks like one of the top offenders. This includes a bunch of

changes to make this phase run faster. This is a ~7% wasm -O1 compile time progression.

Here's what I did:

- We now use IndexSparseSet instead of IndexMap for tracking variable values. This

makes it cheaper to chew through small blocks while there is a non-trivial number of

total variables.

- We now do a "local SSA conversion" pass before anything else. This eliminates

obvious Get's. If we were using temporary Variables, it would eliminate many of

those. That's useful for when we use demoteValues() and duplciateTails(). For wasm

-O1, we mainly care about the fact that it makes a bunch of Set's dead.

- We now do a Set DCE pass after the local SSA but before SSA conversion. This ensures

that any block-local live intervals of Variables disappear and don't need further

consideration.

- We now cache the reaching defs calculation.

- We now perform the reaching defs calculation lazily.

* b3/B3FixSSA.cpp:

(JSC::B3::demoteValues):

(JSC::B3::fixSSA):

* b3/B3SSACalculator.cpp:

(JSC::B3::SSACalculator::reachingDefAtTail):

* b3/B3VariableLiveness.cpp:

(JSC::B3::VariableLiveness::VariableLiveness):

* b3/air/AirLiveness.h:

(JSC::B3::Air::Liveness::Liveness):

* dfg/DFGLivenessAnalysisPhase.cpp:

(JSC::DFG::LivenessAnalysisPhase::LivenessAnalysisPhase): Deleted.

(JSC::DFG::LivenessAnalysisPhase::run): Deleted.

(JSC::DFG::LivenessAnalysisPhase::processBlock): Deleted.