Changeset 192950 in webkit

Timestamp:

Dec 2, 2015, 11:20:12 AM (9 years ago)

Author:

mark.lam@apple.com

Message:

Use the JITAddGenerator snippet in the FTL.
​https://bugs.webkit.org/show_bug.cgi?id=151519

Reviewed by Geoffrey Garen.

One detail about how we choosing to handle operands to the binary snippets that
may be constant: the slow path call to a C++ function still needs the constant
operand loaded in a register. To simplify things, we're choosing to always tell
LLVM to load the operands into registers even if they may be constant. However,
even though a constant operand is preloaded in a register, the snippet generator
will not be made aware of it. It will continue to load the constant as an
immediate.

• ftl/FTLCompile.cpp:
• ftl/FTLCompileBinaryOp.cpp:

(JSC::FTL::generateArithSubFastPath):
(JSC::FTL::generateValueAddFastPath):

• generateValueAddFastPath() currently is an exact copy of generateArithSubFastPath() except that it uses JITAddGenerator instead of JITSubGenerator. When we add support for JITMulGenerator later, the code will start to vary. We'll refactor these functions then when we have more insight into what needs to vary between the implementations.

• ftl/FTLCompileBinaryOp.h:
• ftl/FTLInlineCacheDescriptor.h:
• ftl/FTLInlineCacheDescriptorInlines.h:

(JSC::FTL::ValueAddDescriptor::ValueAddDescriptor):
(JSC::FTL::ValueAddDescriptor::icSize):

• ftl/FTLInlineCacheSize.cpp:

(JSC::FTL::sizeOfValueAdd):

• ftl/FTLInlineCacheSize.h:
• ftl/FTLLowerDFGToLLVM.cpp:

(JSC::FTL::DFG::LowerDFGToLLVM::lower):
(JSC::FTL::DFG::LowerDFGToLLVM::compileValueAdd):

Location:

trunk/Source/JavaScriptCore

Files:

• ChangeLog (modified) (1 diff)
• ftl/FTLCompile.cpp (modified) (1 diff)
• ftl/FTLCompileBinaryOp.cpp (modified) (2 diffs)
• ftl/FTLCompileBinaryOp.h (modified) (1 diff)
• ftl/FTLInlineCacheDescriptor.h (modified) (1 diff)
• ftl/FTLInlineCacheDescriptorInlines.h (modified) (1 diff)
• ftl/FTLInlineCacheSize.cpp (modified) (1 diff)
• ftl/FTLInlineCacheSize.h (modified) (1 diff)
• ftl/FTLLowerDFGToLLVM.cpp (modified) (3 diffs)

trunk/Source/JavaScriptCore/ChangeLog

@@ +1 @@
+2015-12-02  Mark Lam  <mark.lam@apple.com>
+
+        Use the JITAddGenerator snippet in the FTL.
+        https://bugs.webkit.org/show_bug.cgi?id=151519
+
+        Reviewed by Geoffrey Garen.
+
+        One detail about how we choosing to handle operands to the binary snippets that
+        may be constant: the slow path call to a C++ function still needs the constant
+        operand loaded in a register.  To simplify things, we're choosing to always tell
+        LLVM to load the operands into registers even if they may be constant.  However,
+        even though a constant operand is preloaded in a register, the snippet generator
+        will not be made aware of it.  It will continue to load the constant as an
+        immediate.
+
+        * ftl/FTLCompile.cpp:
+        * ftl/FTLCompileBinaryOp.cpp:
+        (JSC::FTL::generateArithSubFastPath):
+        (JSC::FTL::generateValueAddFastPath):
+        - generateValueAddFastPath() currently is an exact copy of generateArithSubFastPath()
+          except that it uses JITAddGenerator instead of JITSubGenerator.  When we add
+          support for JITMulGenerator later, the code will start to vary.  We'll refactor
+          these functions then when we have more insight into what needs to vary between
+          the implementations.
+
+        * ftl/FTLCompileBinaryOp.h:
+        * ftl/FTLInlineCacheDescriptor.h:
+        * ftl/FTLInlineCacheDescriptorInlines.h:
+        (JSC::FTL::ValueAddDescriptor::ValueAddDescriptor):
+        (JSC::FTL::ValueAddDescriptor::icSize):
+        * ftl/FTLInlineCacheSize.cpp:
+        (JSC::FTL::sizeOfValueAdd):
+        * ftl/FTLInlineCacheSize.h:
+        * ftl/FTLLowerDFGToLLVM.cpp:
+        (JSC::FTL::DFG::LowerDFGToLLVM::lower):
+        (JSC::FTL::DFG::LowerDFGToLLVM::compileValueAdd):
+
 2015-12-02  Mark Lam  <mark.lam@apple.com>
 

trunk/Source/JavaScriptCore/ftl/FTLCompile.cpp

@@ -341 +341 @@
             generateArithSubFastPath(ic, fastPathJIT, result, left, right, usedRegisters, done, slowPathStart);
             break;
+        case ValueAdd:
+            generateValueAddFastPath(ic, fastPathJIT, result, left, right, usedRegisters, done, slowPathStart);
+            break;
         default:
             RELEASE_ASSERT_NOT_REACHED();

trunk/Source/JavaScriptCore/ftl/FTLCompileBinaryOp.cpp

@@ -32 +32 @@
 #include "FTLInlineCacheDescriptor.h"
 #include "GPRInfo.h"
+#include "JITAddGenerator.h"
 #include "JITSubGenerator.h"
 #include "ScratchRegisterAllocator.h"
@@ -189 +190 @@
 }
 
+void generateValueAddFastPath(BinaryOpDescriptor& ic, CCallHelpers& jit,
+    GPRReg result, GPRReg left, GPRReg right, RegisterSet usedRegisters,
+    CCallHelpers::Jump& done, CCallHelpers::Jump& slowPathStart)
+{
+    ASSERT(ic.nodeType() == ValueAdd);
+    ScratchRegisterAllocator allocator(usedRegisters);
+
+    BinarySnippetRegisterContext context(allocator, result, left, right);
+
+    GPRReg scratchGPR = allocator.allocateScratchGPR();
+    FPRReg leftFPR = allocator.allocateScratchFPR();
+    FPRReg rightFPR = allocator.allocateScratchFPR();
+    FPRReg scratchFPR = InvalidFPRReg;
+
+    JITAddGenerator gen(ic.leftOperand(), ic.rightOperand(), JSValueRegs(result),
+        JSValueRegs(left), JSValueRegs(right), leftFPR, rightFPR, scratchGPR, scratchFPR);
+
+    auto numberOfBytesUsedToPreserveReusedRegisters =
+    allocator.preserveReusedRegistersByPushing(jit, ScratchRegisterAllocator::ExtraStackSpace::NoExtraSpace);
+
+    context.initializeRegisters(jit);
+    gen.generateFastPath(jit);
+
+    ASSERT(gen.didEmitFastPath());
+    gen.endJumpList().link(&jit);
+    context.restoreRegisters(jit);
+    allocator.restoreReusedRegistersByPopping(jit, numberOfBytesUsedToPreserveReusedRegisters,
+        ScratchRegisterAllocator::ExtraStackSpace::SpaceForCCall);
+    done = jit.jump();
+
+    gen.slowPathJumpList().link(&jit);
+    context.restoreRegisters(jit);
+    allocator.restoreReusedRegistersByPopping(jit, numberOfBytesUsedToPreserveReusedRegisters,
+        ScratchRegisterAllocator::ExtraStackSpace::SpaceForCCall);
+    slowPathStart = jit.jump();
+}
+
 } } // namespace JSC::FTL
 

trunk/Source/JavaScriptCore/ftl/FTLCompileBinaryOp.h

@@ -40 +40 @@
     CCallHelpers::Jump& done, CCallHelpers::Jump& slowPathStart);
 
+void generateValueAddFastPath(BinaryOpDescriptor&, CCallHelpers&,
+    GPRReg result, GPRReg left, GPRReg right, RegisterSet usedRegisters,
+    CCallHelpers::Jump& done, CCallHelpers::Jump& slowPathStart);
+
 } // namespace FTL
 } // namespace JSC

trunk/Source/JavaScriptCore/ftl/FTLInlineCacheDescriptor.h

@@ -172 +172 @@
 };
 
+class ValueAddDescriptor : public BinaryOpDescriptor {
+public:
+    ValueAddDescriptor(unsigned stackmapID, CodeOrigin, const SnippetOperand& leftOperand, const SnippetOperand& rightOperand);
+    static size_t icSize();
+};
+
 // You can create a lazy slow path call in lowerDFGToLLVM by doing:
 // m_ftlState.lazySlowPaths.append(

trunk/Source/JavaScriptCore/ftl/FTLInlineCacheDescriptorInlines.h

@@ -48 +48 @@
 }
 
+ValueAddDescriptor::ValueAddDescriptor(unsigned stackmapID, CodeOrigin codeOrigin,
+    const SnippetOperand& leftOperand, const SnippetOperand& rightOperand)
+    : BinaryOpDescriptor(DFG::ValueAdd, stackmapID, codeOrigin, icSize(),
+        "ValueAdd", "ValueAdd IC fast path", DFG::operationValueAdd, leftOperand, rightOperand)
+{
+}
+
+size_t ValueAddDescriptor::icSize()
+{
+    return sizeOfValueAdd();
+}
+
 } } // namespace JSC::FTL
 

trunk/Source/JavaScriptCore/ftl/FTLInlineCacheSize.cpp

@@ -142 +142 @@
 #else
     return 298; // X86_64 debug.
+#endif
+#endif
+}
+
+size_t sizeOfValueAdd()
+{
+#if CPU(ARM64)
+#ifdef NDEBUG
+    return 180; // ARM64 release.
+#else
+    return 276; // ARM64 debug.
+#endif
+#else // CPU(X86_64)
+#ifdef NDEBUG
+    return 199; // X86_64 release.
+#else
+    return 286; // X86_64 debug.
 #endif
 #endif

trunk/Source/JavaScriptCore/ftl/FTLInlineCacheSize.h

@@ -48 +48 @@
 size_t sizeOfIn();
 size_t sizeOfArithSub();
+size_t sizeOfValueAdd();
 #if ENABLE(MASM_PROBE)
 size_t sizeOfProbe();

trunk/Source/JavaScriptCore/ftl/FTLLowerDFGToLLVM.cpp

@@ -235 +235 @@
                     case ArithSub:
                     case GetById:
-                    case GetByIdFlush: {
+                    case GetByIdFlush:
+                    case ValueAdd: {
                         // We may have to flush one thing for GetByIds/ArithSubs when the base and result or the left/right and the result
                         // are assigned the same register. For a more comprehensive overview, look at the comment in FTLCompile.cpp
@@ -244 +245 @@
                         bool willCatchException = m_graph.willCatchExceptionInMachineFrame(node->origin.forExit, opCatchOrigin, exceptionHandler);
                         if (willCatchException) {
-                            static const size_t numberOfGetByIdOrSubSpills = 1;
-                            maxNumberOfCatchSpills = std::max(maxNumberOfCatchSpills, numberOfGetByIdOrSubSpills);
+                            static const size_t numberOfGetByIdOrBinaryOpSpills = 1;
+                            maxNumberOfCatchSpills = std::max(maxNumberOfCatchSpills, numberOfGetByIdOrBinaryOpSpills);
                         }
                         break;
@@ -1485 +1486 @@
     void compileValueAdd()
     {
-        J_JITOperation_EJJ operation;
-        if (!(provenType(m_node->child1()) & SpecFullNumber)
-            && !(provenType(m_node->child2()) & SpecFullNumber))
-            operation = operationValueAddNotNumber;
-        else
-            operation = operationValueAdd;
-        setJSValue(vmCall(
-            m_out.int64, m_out.operation(operation), m_callFrame,
-            lowJSValue(m_node->child1()), lowJSValue(m_node->child2())));
+        auto leftChild = m_node->child1();
+        auto rightChild = m_node->child2();
+
+        if (!(provenType(leftChild) & SpecFullNumber) || !(provenType(rightChild) & SpecFullNumber)) {
+            setJSValue(vmCall(m_out.int64, m_out.operation(operationValueAddNotNumber), m_callFrame,
+                lowJSValue(leftChild), lowJSValue(rightChild)));
+            return;
+        }
+
+        unsigned stackmapID = m_stackmapIDs++;
+
+        if (Options::verboseCompilation())
+            dataLog("    Emitting ValueAdd patchpoint with stackmap #", stackmapID, "\n");
+
+#if FTL_USES_B3
+        CRASH();
+#else
+        LValue left = lowJSValue(leftChild);
+        LValue right = lowJSValue(rightChild);
+
+        SnippetOperand leftOperand(abstractValue(leftChild).resultType());
+        SnippetOperand rightOperand(abstractValue(rightChild).resultType());
+
+        // The DFG does not always fold the sum of 2 constant int operands together.
+        // Because the snippet does not support both operands being constant, if the left
+        // operand is already a constant, we'll just pretend the right operand is not.
+        if (leftChild->isInt32Constant())
+            leftOperand.setConstInt32(leftChild->asInt32());
+        if (!leftOperand.isConst() && rightChild->isInt32Constant())
+            rightOperand.setConstInt32(rightChild->asInt32());
+
+        // Arguments: id, bytes, target, numArgs, args...
+        StackmapArgumentList arguments;
+        arguments.append(m_out.constInt64(stackmapID));
+        arguments.append(m_out.constInt32(ValueAddDescriptor::icSize()));
+        arguments.append(constNull(m_out.ref8));
+        arguments.append(m_out.constInt32(2));
+        arguments.append(left);
+        arguments.append(right);
+
+        appendOSRExitArgumentsForPatchpointIfWillCatchException(arguments,
+            ExceptionType::BinaryOpGenerator, 3); // left, right, and result show up in the stackmap locations.
+
+        LValue call = m_out.call(m_out.int64, m_out.patchpointInt64Intrinsic(), arguments);
+        setInstructionCallingConvention(call, LLVMAnyRegCallConv);
+
+        m_ftlState.binaryOps.append(ValueAddDescriptor(stackmapID, m_node->origin.semantic, leftOperand, rightOperand));
+
+        setJSValue(call);
+#endif
     }