Changeset 214069 in webkit for trunk/Source/JavaScriptCore/ftl/FTLOSRExitCompiler.cpp

Timestamp:

Mar 16, 2017, 2:19:23 PM (8 years ago)

Author:

[email protected]

Message:

FTL should support global and eval code
​https://bugs.webkit.org/show_bug.cgi?id=169656

Reviewed by Geoffrey Garen and Saam Barati.

JSTests:

Added basic performance tests of global and eval code. These tests will run a lot faster in with
the FTL because of the object allocation.

• microbenchmarks/eval-code-ftl-reentry.js: Added.
• microbenchmarks/eval-code-ftl.js: Added.
• microbenchmarks/global-code-ftl.js: Added.
• stress/arith-log-on-various-types.js: This was a flaky fail with concurrent JIT, so I stopped running it with concurrent JIT. The failure was its assertion about how many times something gets compiled.

Source/JavaScriptCore:

Turned off the restriction against global and eval code running in the FTL, and then fixed all of
the things that didn't work.

This is a big speed-up on microbenchmarks that I wrote for this patch. One of the reasons why we
hadn't done this earlier is that we've never seen a benchmark that needed it. Global and eval
code rarely gets FTL-hot. Still, this seems like possibly a small JetStream speed-up.

• dfg/DFGJITCode.cpp:

(JSC::DFG::JITCode::setOSREntryBlock): I outlined this for better debugging.

• dfg/DFGJITCode.h:

(JSC::DFG::JITCode::setOSREntryBlock): Deleted.

• dfg/DFGNode.h:

(JSC::DFG::Node::isSemanticallySkippable): It turns out that global code often has InvalidationPoints before LoopHints. They are also skippable from the standpoint of OSR entrypoint analysis.

• dfg/DFGOperations.cpp: Don't do any normal compiles of global code - just do OSR compiles.
• ftl/FTLCapabilities.cpp: Enable FTL for global and eval code.

(JSC::FTL::canCompile):

• ftl/FTLCompile.cpp: Just debugging clean-ups.

(JSC::FTL::compile):

• ftl/FTLJITFinalizer.cpp: Implement finalize() and ensure that we only do things with the entrypoint buffer if we have one. We won't have one for eval code that we aren't OSR entering into.

(JSC::FTL::JITFinalizer::finalize):
(JSC::FTL::JITFinalizer::finalizeFunction):
(JSC::FTL::JITFinalizer::finalizeCommon):

• ftl/FTLJITFinalizer.h:
• ftl/FTLLink.cpp: When entering a function normally, we need the "entrypoint" to put the arity check code. Global and eval code don't need this.

(JSC::FTL::link):

• ftl/FTLOSREntry.cpp: Fix a dataLog statement.

(JSC::FTL::prepareOSREntry):

• ftl/FTLOSRExitCompiler.cpp: Remove dead code that happened to assert that we're exiting from a function.

(JSC::FTL::compileStub):

File:

• trunk/Source/JavaScriptCore/ftl/FTLOSRExitCompiler.cpp (modified) (1 diff)

trunk/Source/JavaScriptCore/ftl/FTLOSRExitCompiler.cpp

@@ -395 +395 @@
     }
     
-    jit.load32(CCallHelpers::payloadFor(CallFrameSlot::argumentCount), GPRInfo::regT2);
-    
-    // Let's say that the FTL function had failed its arity check. In that case, the stack will
-    // contain some extra stuff.
-    //
-    // We compute the padded stack space:
-    //
-    //     paddedStackSpace = roundUp(codeBlock->numParameters - regT2 + 1)
-    //
-    // The stack will have regT2 + CallFrameHeaderSize stuff.
-    // We want to make the stack look like this, from higher addresses down:
-    //
-    //     - argument padding
-    //     - actual arguments
-    //     - call frame header
-
-    // This code assumes that we're dealing with FunctionCode.
-    RELEASE_ASSERT(codeBlock->codeType() == FunctionCode);
-    
-    jit.add32(
-        MacroAssembler::TrustedImm32(-codeBlock->numParameters()), GPRInfo::regT2,
-        GPRInfo::regT3);
-    MacroAssembler::Jump arityIntact = jit.branch32(
-        MacroAssembler::GreaterThanOrEqual, GPRInfo::regT3, MacroAssembler::TrustedImm32(0));
-    jit.neg32(GPRInfo::regT3);
-    jit.add32(MacroAssembler::TrustedImm32(1 + stackAlignmentRegisters() - 1), GPRInfo::regT3);
-    jit.and32(MacroAssembler::TrustedImm32(-stackAlignmentRegisters()), GPRInfo::regT3);
-    jit.add32(GPRInfo::regT3, GPRInfo::regT2);
-    arityIntact.link(&jit);
-
     CodeBlock* baselineCodeBlock = jit.baselineCodeBlockFor(exit.m_codeOrigin);