AuroraMiddleware/v8
1
0
Fork 0
Code Issues 2 Pull Requests Releases Wiki Activity

[x64] Teach ArgumentsAdaptorTrampoline about is_safe_to_skip_adapt_arguments.

Browse Source 
This adds support for dynamically checking the newly introduced
is_safe_to_skip_adapt_arguments flag on the SharedFunctionInfo inside
the ArgumentsAdaptorTrampoline builtin, so that if there's an arguments
mismatch, but said bit is set, we can just massage the current stack
frame according to the SharedFunctionInfo and don't need to create an
arguments adaptor frame.

This improves the general (baseline) performance of calls with arguments
mismatches by up to 35%, as can be seen with the test case from the tracking
bug, which goes from around

  callUnderApplication: 608 ms.
  callOverApplication1: 633 ms.
  callOverApplication2: 635 ms.

to

  callUnderApplication: 390 ms.
  callOverApplication1: 439 ms.
  callOverApplication2: 443 ms.

when the TurboFan fast-path (that was previously introduced) is forcibly
disabled. That's a pretty solid baseline improvement and there's probably
still some room for further improvement (i.e. avoid loading the SFI multiple
times, etc.).

Drive-by-fix: Refactor and cleanup the ArgumentsAdaptorTrampoline
builtin a bit to make it more efficient and easier to read (it's still
pretty messy, but I don't wanna change too many things at once).

This is only the initial x64 port, since for ia32 we're really short on
registers and it might not even pay off (initial prototype shows a lot
less performance impact). The arm and arm64 port is in progress and will
land once ready.

Bug: v8:8834, v8:8895
Change-Id: I2e1c8e4767a808a43b932487fdedc0e44cdd35c5
Reviewed-on: https://chromium-review.googlesource.com/c/1488751
Commit-Queue: Benedikt Meurer <bmeurer@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#59881}
This commit is contained in:
Benedikt Meurer 2019-02-26 21:38:36 +01:00 committed by Commit Bot
parent 47210f480f
commit 07ee7eadd4
1 changed files with 119 additions and 75 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

194
src/builtins/x64/builtins-x64.cc
View File

@ -1889,85 +1889,129 @@ void Builtins::Generate_ArgumentsAdaptorTrampoline(MacroAssembler* masm) {
// -- rdi : function (passed through to callee)
// -----------------------------------
Label invoke, dont_adapt_arguments, stack_overflow, enough, too_few;
Label dont_adapt_arguments, stack_overflow, skip_adapt_arguments;
__ cmpq(rbx, Immediate(SharedFunctionInfo::kDontAdaptArgumentsSentinel));
__ j(equal, &dont_adapt_arguments);
__ cmpq(rax, rbx);
__ j(less, &too_few);
{ // Enough parameters: Actual >= expected.
__ bind(&enough);
EnterArgumentsAdaptorFrame(masm);
// The registers rcx and r8 will be modified. The register rbx is only read.
Generate_StackOverflowCheck(masm, rbx, rcx, &stack_overflow);
// Copy receiver and all expected arguments.
const int offset = StandardFrameConstants::kCallerSPOffset;
__ leaq(rax, Operand(rbp, rax, times_system_pointer_size, offset));
__ Set(r8, -1); // account for receiver
Label copy;
__ bind(&copy);
__ incq(r8);
__ Push(Operand(rax, 0));
__ subq(rax, Immediate(kSystemPointerSize));
__ cmpq(r8, rbx);
__ j(less, &copy);
__ jmp(&invoke);
}
{ // Too few parameters: Actual < expected.
__ bind(&too_few);
EnterArgumentsAdaptorFrame(masm);
// The registers rcx and r8 will be modified. The register rbx is only read.
Generate_StackOverflowCheck(masm, rbx, rcx, &stack_overflow);
// Copy receiver and all actual arguments.
const int offset = StandardFrameConstants::kCallerSPOffset;
__ leaq(rdi, Operand(rbp, rax, times_system_pointer_size, offset));
__ Set(r8, -1); // account for receiver
Label copy;
__ bind(&copy);
__ incq(r8);
__ Push(Operand(rdi, 0));
__ subq(rdi, Immediate(kSystemPointerSize));
__ cmpq(r8, rax);
__ j(less, &copy);
// Fill remaining expected arguments with undefined values.
Label fill;
__ LoadRoot(kScratchRegister, RootIndex::kUndefinedValue);
__ bind(&fill);
__ incq(r8);
__ Push(kScratchRegister);
__ cmpq(r8, rbx);
__ j(less, &fill);
// Restore function pointer.
__ movq(rdi, Operand(rbp, ArgumentsAdaptorFrameConstants::kFunctionOffset));
}
// Call the entry point.
__ bind(&invoke);
__ movq(rax, rbx);
// rax : expected number of arguments
// rdx : new target (passed through to callee)
// rdi : function (passed through to callee)
static_assert(kJavaScriptCallCodeStartRegister == rcx, "ABI mismatch");
__ LoadTaggedPointerField(rcx, FieldOperand(rdi, JSFunction::kCodeOffset));
__ CallCodeObject(rcx);
// Store offset of return address for deoptimizer.
masm->isolate()->heap()->SetArgumentsAdaptorDeoptPCOffset(masm->pc_offset());
// Leave frame and return.
LeaveArgumentsAdaptorFrame(masm);
__ ret(0);
__ LoadTaggedPointerField(
rcx, FieldOperand(rdi, JSFunction::kSharedFunctionInfoOffset));
__ testl(
FieldOperand(rcx, SharedFunctionInfo::kFlagsOffset),
Immediate(SharedFunctionInfo::IsSafeToSkipArgumentsAdaptorBit::kMask));
__ j(not_zero, &skip_adapt_arguments);
// -------------------------------------------
// Dont adapt arguments.
// Adapt arguments.
// -------------------------------------------
{
EnterArgumentsAdaptorFrame(masm);
Generate_StackOverflowCheck(masm, rbx, rcx, &stack_overflow);
Label under_application, over_application, invoke;
__ cmpq(rax, rbx);
__ j(less, &under_application, Label::kNear);
// Enough parameters: Actual >= expected.
__ bind(&over_application);
{
// Copy receiver and all expected arguments.
const int offset = StandardFrameConstants::kCallerSPOffset;
__ leaq(r8, Operand(rbp, rax, times_system_pointer_size, offset));
__ Set(rax, -1); // account for receiver
Label copy;
__ bind(&copy);
__ incq(rax);
__ Push(Operand(r8, 0));
__ subq(r8, Immediate(kSystemPointerSize));
__ cmpq(rax, rbx);
__ j(less, &copy);
__ jmp(&invoke, Label::kNear);
}
// Too few parameters: Actual < expected.
__ bind(&under_application);
{
// Copy receiver and all actual arguments.
const int offset = StandardFrameConstants::kCallerSPOffset;
__ leaq(r9, Operand(rbp, rax, times_system_pointer_size, offset));
__ Set(r8, -1); // account for receiver
Label copy;
__ bind(&copy);
__ incq(r8);
__ Push(Operand(r9, 0));
__ subq(r9, Immediate(kSystemPointerSize));
__ cmpq(r8, rax);
__ j(less, &copy);
// Fill remaining expected arguments with undefined values.
Label fill;
__ LoadRoot(kScratchRegister, RootIndex::kUndefinedValue);
__ bind(&fill);
__ incq(rax);
__ Push(kScratchRegister);
__ cmpq(rax, rbx);
__ j(less, &fill);
}
// Call the entry point.
__ bind(&invoke);
// rax : expected number of arguments
// rdx : new target (passed through to callee)
// rdi : function (passed through to callee)
static_assert(kJavaScriptCallCodeStartRegister == rcx, "ABI mismatch");
__ LoadTaggedPointerField(rcx, FieldOperand(rdi, JSFunction::kCodeOffset));
__ CallCodeObject(rcx);
// Store offset of return address for deoptimizer.
masm->isolate()->heap()->SetArgumentsAdaptorDeoptPCOffset(
masm->pc_offset());
// Leave frame and return.
LeaveArgumentsAdaptorFrame(masm);
__ ret(0);
}
// -------------------------------------------
// Skip adapt arguments.
// -------------------------------------------
__ bind(&skip_adapt_arguments);
{
// The callee cannot observe the actual arguments, so it's safe to just
// pass the expected arguments by massaging the stack appropriately. See
// http://bit.ly/v8-faster-calls-with-arguments-mismatch for details.
Label under_application, over_application, invoke;
__ PopReturnAddressTo(rcx);
__ cmpq(rax, rbx);
__ j(less, &under_application, Label::kNear);
__ bind(&over_application);
{
// Remove superfluous parameters from the stack.
__ xchgq(rax, rbx);
__ subq(rbx, rax);
__ leaq(rsp, Operand(rsp, rbx, times_system_pointer_size, 0));
__ jmp(&invoke, Label::kNear);
}
__ bind(&under_application);
{
// Fill remaining expected arguments with undefined values.
Label fill;
__ LoadRoot(kScratchRegister, RootIndex::kUndefinedValue);
__ bind(&fill);
__ incq(rax);
__ Push(kScratchRegister);
__ cmpq(rax, rbx);
__ j(less, &fill);
}
__ bind(&invoke);
__ PushReturnAddressFrom(rcx);
}
// -------------------------------------------
// Don't adapt arguments.
// -------------------------------------------
__ bind(&dont_adapt_arguments);
static_assert(kJavaScriptCallCodeStartRegister == rcx, "ABI mismatch");