FTL B3 should be just as fast as FTL LLVM on Octane/crypto
https://bugs.webkit.org/show_bug.cgi?id=153113
Reviewed by Saam Barati.
Source/JavaScriptCore:
This is the result of a hacking rampage to close the gap between FTL B3 and FTL LLVM on
Octane/crypto. It was a very successful rampage.
The biggest change in this patch is the introduction of a phase called fixObviousSpills()
that fixes patterns like:
Store register to stack slot and then use stack slot:
Move %rcx, (stack42)
Foo use:(stack42) // replace (stack42) with %rcx here.
Load stack slot into register and then use stack slot:
Move (stack42), %rcx
Foo use:(stack42) // replace (stack42) with %rcx here.
Store constant into stack slot and then use stack slot:
Move $42, %rcx
Move %rcx, (stack42)
Bar def:%rcx // %rcx isn't available anymore, but we still know that (stack42) is $42
Foo use:(stack42) // replace (stack42) with $42 here.
This phases does these fixups by doing a global forward flow that propagates sets of
must-aliases.
Also added a phase to report register pressure. It pretty-prints code alongside the set of
in-use registers above each instruction. Using this phase, I found that our register
allocator is actually doing a pretty awesome job. I had previously feared that we'd have to
make substantial changes to register allocation. I don't have such a fear anymore, at least
for Octane/crypto. In the future, we can check how the regalloc is performing just by
enabling logAirRegisterPressure.
Also fixed some FTL codegen pathologies. We were using bitOr where we meant to use a
conditional or. LLVM likes to canonicalize boolean expressions this way. B3, on the other
hand, doesn't do this canonicalization and doesn't have logic to decompose it into sequences
of branches.
Also added strength reductions for checked arithmetic. It turns out that LLVM learned how to
reduce checked multiply to unchecked multiply in some obvious cases that our existing DFG
optimizations lacked. Ideally, our DFG integer range optimization phase would cover this. But
the cases of interest were dead simple - the incoming values to the CheckMul were obviously
too small to cause overflow. I added such reasoning to B3's strength reduction.
Finally, this fixes some bugs with how we were handling subwidth spill slots. The register
allocator was making two mistakes. First, it might cause a Width64 def or use of a 4-byte
spill slot. In that case, it would extend the size of the spill slot to ensure that the use
or def is safe. Second, it emulates ZDef on Tmp behavior by emitting a Move32 to initialize
the high bits of a spill slot. But this is unsound because of the liveness semantics of spill
slots. They cannot have more than one def to initialize their value. I fixed that by making
allocateStack() be the thing that fixes ZDefs. That's a change to ZDef semantics: now, ZDef
on an anonymous stack slot means that the high bits are zero-filled. I wasn't able to
construct a test for this. It might be a hypothetical bug, but still, I like how this
simplifies the register allocator.
This is a ~0.7% speed-up on Octane.
* CMakeLists.txt:
* JavaScriptCore.xcodeproj/project.pbxproj:
* b3/B3CheckSpecial.cpp:
(JSC::B3::CheckSpecial::hiddenBranch):
(JSC::B3::CheckSpecial::forEachArg):
(JSC::B3::CheckSpecial::commitHiddenBranch): Deleted.
* b3/B3CheckSpecial.h:
* b3/B3LowerToAir.cpp:
(JSC::B3::Air::LowerToAir::fillStackmap):
(JSC::B3::Air::LowerToAir::lower):
* b3/B3StackmapValue.h:
* b3/air/AirAllocateStack.cpp:
(JSC::B3::Air::allocateStack):
* b3/air/AirAllocateStack.h:
* b3/air/AirArg.h:
(JSC::B3::Air::Arg::callArg):
(JSC::B3::Air::Arg::stackAddr):
(JSC::B3::Air::Arg::isValidScale):
* b3/air/AirBasicBlock.cpp:
(JSC::B3::Air::BasicBlock::deepDump):
(JSC::B3::Air::BasicBlock::dumpHeader):
(JSC::B3::Air::BasicBlock::dumpFooter):
* b3/air/AirBasicBlock.h:
* b3/air/AirCCallSpecial.cpp:
(JSC::B3::Air::CCallSpecial::CCallSpecial):
(JSC::B3::Air::CCallSpecial::~CCallSpecial):
* b3/air/AirCode.h:
(JSC::B3::Air::Code::lastPhaseName):
(JSC::B3::Air::Code::setEnableRCRS):
(JSC::B3::Air::Code::enableRCRS):
* b3/air/AirCustom.cpp:
(JSC::B3::Air::PatchCustom::isValidForm):
(JSC::B3::Air::CCallCustom::isValidForm):
* b3/air/AirCustom.h:
(JSC::B3::Air::PatchCustom::isValidFormStatic):
(JSC::B3::Air::PatchCustom::admitsStack):
(JSC::B3::Air::PatchCustom::isValidForm): Deleted.
* b3/air/AirEmitShuffle.cpp:
(JSC::B3::Air::ShufflePair::dump):
(JSC::B3::Air::createShuffle):
(JSC::B3::Air::emitShuffle):
* b3/air/AirEmitShuffle.h:
* b3/air/AirFixObviousSpills.cpp: Added.
(JSC::B3::Air::fixObviousSpills):
* b3/air/AirFixObviousSpills.h: Added.
* b3/air/AirFixSpillSlotZDef.h: Removed.
* b3/air/AirGenerate.cpp:
(JSC::B3::Air::prepareForGeneration):
(JSC::B3::Air::generate):
* b3/air/AirHandleCalleeSaves.cpp:
(JSC::B3::Air::handleCalleeSaves):
* b3/air/AirInst.h:
* b3/air/AirInstInlines.h:
(JSC::B3::Air::Inst::reportUsedRegisters):
(JSC::B3::Air::Inst::admitsStack):
(JSC::B3::Air::isShiftValid):
* b3/air/AirIteratedRegisterCoalescing.cpp:
* b3/air/AirLiveness.h:
(JSC::B3::Air::AbstractLiveness::AbstractLiveness):
(JSC::B3::Air::AbstractLiveness::LocalCalc::Iterable::begin):
(JSC::B3::Air::AbstractLiveness::LocalCalc::Iterable::end):
(JSC::B3::Air::AbstractLiveness::LocalCalc::Iterable::contains):
(JSC::B3::Air::AbstractLiveness::LocalCalc::live):
(JSC::B3::Air::AbstractLiveness::LocalCalc::isLive):
(JSC::B3::Air::AbstractLiveness::LocalCalc::execute):
(JSC::B3::Air::AbstractLiveness::rawLiveAtHead):
(JSC::B3::Air::AbstractLiveness::Iterable::begin):
(JSC::B3::Air::AbstractLiveness::Iterable::end):
(JSC::B3::Air::AbstractLiveness::Iterable::contains):
(JSC::B3::Air::AbstractLiveness::liveAtTail):
(JSC::B3::Air::AbstractLiveness::workset):
* b3/air/AirLogRegisterPressure.cpp: Added.
(JSC::B3::Air::logRegisterPressure):
* b3/air/AirLogRegisterPressure.h: Added.
* b3/air/AirOptimizeBlockOrder.cpp:
(JSC::B3::Air::blocksInOptimizedOrder):
(JSC::B3::Air::optimizeBlockOrder):
* b3/air/AirOptimizeBlockOrder.h:
* b3/air/AirReportUsedRegisters.cpp:
(JSC::B3::Air::reportUsedRegisters):
* b3/air/AirReportUsedRegisters.h:
* b3/air/AirSpillEverything.cpp:
(JSC::B3::Air::spillEverything):
* b3/air/AirStackSlot.h:
(JSC::B3::Air::StackSlot::isLocked):
(JSC::B3::Air::StackSlot::index):
(JSC::B3::Air::StackSlot::ensureSize):
(JSC::B3::Air::StackSlot::alignment):
* b3/air/AirValidate.cpp:
* ftl/FTLB3Compile.cpp:
(JSC::FTL::compile):
* ftl/FTLLowerDFGToLLVM.cpp:
(JSC::FTL::DFG::LowerDFGToLLVM::compileArithMul):
(JSC::FTL::DFG::LowerDFGToLLVM::compileArithDiv):
(JSC::FTL::DFG::LowerDFGToLLVM::compileArithMod):
* jit/RegisterSet.h:
(JSC::RegisterSet::get):
(JSC::RegisterSet::setAll):
(JSC::RegisterSet::merge):
(JSC::RegisterSet::filter):
* runtime/Options.h:
Source/WTF:
* wtf/IndexSparseSet.h:
(WTF::IndexSparseSet<OverflowHandler>::IndexSparseSet):
(WTF::IndexSparseSet<OverflowHandler>::add):
(WTF::IndexSparseSet<OverflowHandler>::remove):
* wtf/StringPrintStream.h:
(WTF::StringPrintStream::length):
git-svn-id: http://svn.webkit.org/repository/webkit/trunk@195298 268f45cc-cd09-0410-ab3c-d52691b4dbfc
diff --git a/Source/JavaScriptCore/ftl/FTLB3Compile.cpp b/Source/JavaScriptCore/ftl/FTLB3Compile.cpp
index 2cbbc35..97f34e0 100644
--- a/Source/JavaScriptCore/ftl/FTLB3Compile.cpp
+++ b/Source/JavaScriptCore/ftl/FTLB3Compile.cpp
@@ -28,6 +28,7 @@
#if ENABLE(FTL_JIT) && FTL_USES_B3
+#include "AirCode.h"
#include "B3Generate.h"
#include "B3ProcedureInlines.h"
#include "B3StackSlotValue.h"
@@ -57,7 +58,7 @@
Graph& graph = state.graph;
CodeBlock* codeBlock = graph.m_codeBlock;
VM& vm = graph.m_vm;
-
+
{
GraphSafepoint safepoint(state.graph, safepointResult);
