1905609 - Clean-up various floating point conversion methods

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, task)

Description

[André Bargull [:anba]]

Perform some clean-up in preparation for bug 1835034.

Comment 1

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 1: Add CacheIRCompiler::liveVolatileRegs() method. r=jandem!

Add liveVolatileRegs to avoid some code duplication.

Comment 2

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 1: Add CacheIRCompiler::liveVolatileRegs() method. r=jandem!

Add liveVolatileRegs to avoid some code duplication.

Comment 3

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 2: Remove an outdated restriction. r=jandem!

The issue mentioned in the comment no longer applies to CacheIR. See also bug 836000.

Depends on D215327

Comment 4

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 3: Use scratch register when available. r=jandem!

Depends on D215328

Comment 5

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 4: Remove unused MToFPInstruction::ConversionKind. r=jandem!

And avoid code duplication in MToFPInstruction sub-classes.

Depends on D215329

Comment 6

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 5: Remove unused IntConversionInputKind::NumbersOrBoolsOnly. r=jandem!

Depends on D215330

Comment 7

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 6: Merge code to convert JS::Value to Float32/Double. r=jandem!

Also remove DISPATCH_FLOATING_POINT_OP macro.

Depends on D215331

Comment 8

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 7: Remove {bool,int32}ValueToDouble calls from GenerateJitEntry. r=jandem!

And change GenerateJitEntry to avoid using ScratchTagScopeRelease.

Depends on D215332

Comment 9

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 8: Replace int32ValueTo{Float32,Double} with direct calls to convertInt32To{Float32,Double}. r=jandem!

Depends on D215333

Comment 10

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 9: Replace boolValueTo{Float32,Double} with direct calls to convertInt32To{Float32,Double}. r=jandem!

Code-gen changes:

• ARM32 loaded 1.0 into the output register and conditionally subtracted the
  output with itself to compute 0.0. The new code-gen uses int32-to-float
  conversions instead.

• MIPS had an extra call to convertBoolToInt32 before convertInt32ToFloat32.
  It's unclear why this was performed, because JS::TrueValue and JS::FalseValue
  don't have any extra bits set which require convertBoolToInt32.

• LOONG64 and RISCV code was copied from MIPS, so it also had the extra call to
  convertBoolToInt32.

Depends on D215334

Comment 11

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 10: Merge Int32 and Bool codepaths in convertValueToFloatingPoint. r=jandem!

Depends on D215335

Comment 12

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 11: Avoid an extra move in MacroAssembler::convertInt32ValueToDouble. r=jandem!

Directly read from payloadOrValueReg to match convertValueToFloatingPoint,
which also saves a move in X64 and ARM64.

Depends on D215336

Comment 13

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 12: Clean-up fuzzilli codegen. r=jandem!

Depends on D215337

Comment 14

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 13: Share MacroAssembler::ensureDouble. r=jandem!

64, RISCV, and MIPS64 performed an additional unboxInt32 to first
move the value into a scratch register. This doesn't look necessary,
for example ARM64 directly converted the value into a double.

Depends on D215338

Comment 15

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 14: Rename convertValueToFloat to convertValueToFloat32. r=jandem!

Depends on D215339

Comment 16

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 13: Share MacroAssembler::ensureDouble. r=jandem!

64, RISCV, and MIPS64 performed an additional unboxInt32 to first
move the value into a scratch register. This doesn't look necessary,
for example ARM64 directly converted the value into a double.

Depends on D215338

Comment 17

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 14: Rename convertValueToFloat to convertValueToFloat32. r=jandem!

Depends on D215339

Comment 18

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 13: Share MacroAssembler::ensureDouble. r=jandem!

64, RISCV, and MIPS64 performed an additional unboxInt32 to first
move the value into a scratch register. This doesn't look necessary,
for example ARM64 directly converted the value into a double.

Depends on D215338

Comment 19

[André Bargull [:anba]]

Attachment: Bug 1905609 - Part 14: Rename convertValueToFloat to convertValueToFloat32. r=jandem!

Depends on D215339

Comment 20

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5c9d6193adbf
Part 1: Add CacheIRCompiler::liveVolatileRegs() method. r=jandem
https://hg.mozilla.org/integration/autoland/rev/938c62c4b746
Part 2: Remove an outdated restriction. r=jandem
https://hg.mozilla.org/integration/autoland/rev/5958fa2e4e82
Part 3: Use scratch register when available. r=jandem
https://hg.mozilla.org/integration/autoland/rev/e436cecf3988
Part 4: Remove unused MToFPInstruction::ConversionKind. r=jandem
https://hg.mozilla.org/integration/autoland/rev/3b9b53aa8e11
Part 5: Remove unused IntConversionInputKind::NumbersOrBoolsOnly. r=jandem
https://hg.mozilla.org/integration/autoland/rev/6ae199d650a2
Part 6: Merge code to convert JS::Value to Float32/Double. r=jandem
https://hg.mozilla.org/integration/autoland/rev/1fb1c737d9df
Part 7: Remove {bool,int32}ValueToDouble calls from GenerateJitEntry. r=jandem
https://hg.mozilla.org/integration/autoland/rev/1f59cc9362d7
Part 8: Replace int32ValueTo{Float32,Double} with direct calls to convertInt32To{Float32,Double}. r=jandem
https://hg.mozilla.org/integration/autoland/rev/2333bfe3a36b
Part 9: Replace boolValueTo{Float32,Double} with direct calls to convertInt32To{Float32,Double}. r=jandem
https://hg.mozilla.org/integration/autoland/rev/62e3c8983d41
Part 10: Merge Int32 and Bool codepaths in convertValueToFloatingPoint. r=jandem
https://hg.mozilla.org/integration/autoland/rev/29a868060e95
Part 11: Avoid an extra move in MacroAssembler::convertInt32ValueToDouble. r=jandem
https://hg.mozilla.org/integration/autoland/rev/39f29404d94d
Part 12: Clean-up fuzzilli codegen. r=jandem
https://hg.mozilla.org/integration/autoland/rev/00d1a13b7fb4
Part 13: Share MacroAssembler::ensureDouble. r=jandem
https://hg.mozilla.org/integration/autoland/rev/c788f13a3a12
Part 14: Rename convertValueToFloat to convertValueToFloat32. r=jandem
https://hg.mozilla.org/integration/autoland/rev/9efa1f43808a
apply code formatting via Lando

Comment 21

[Iulian Moraru]

https://hg.mozilla.org/mozilla-central/rev/5c9d6193adbf
https://hg.mozilla.org/mozilla-central/rev/938c62c4b746
https://hg.mozilla.org/mozilla-central/rev/5958fa2e4e82
https://hg.mozilla.org/mozilla-central/rev/e436cecf3988
https://hg.mozilla.org/mozilla-central/rev/3b9b53aa8e11
https://hg.mozilla.org/mozilla-central/rev/6ae199d650a2
https://hg.mozilla.org/mozilla-central/rev/1fb1c737d9df
https://hg.mozilla.org/mozilla-central/rev/1f59cc9362d7
https://hg.mozilla.org/mozilla-central/rev/2333bfe3a36b
https://hg.mozilla.org/mozilla-central/rev/62e3c8983d41
https://hg.mozilla.org/mozilla-central/rev/29a868060e95
https://hg.mozilla.org/mozilla-central/rev/39f29404d94d
https://hg.mozilla.org/mozilla-central/rev/00d1a13b7fb4
https://hg.mozilla.org/mozilla-central/rev/c788f13a3a12
https://hg.mozilla.org/mozilla-central/rev/9efa1f43808a