Bug 227509 – Add Air opcode add/sub-and-shift for ARM64 and select this instruction in Air

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Yijia Huang]

...

[Yijia Huang]

Created attachment 432963 [details]
Patch

[Radar WebKit Bug Importer]

<rdar://problem/80233961>

[Yijia Huang]

Created attachment 432983 [details]
Patch

[Yijia Huang]

Created attachment 432991 [details]
Patch

[Yijia Huang]

Created attachment 433001 [details]
Patch

[Yijia Huang]

Created attachment 433015 [details]
Patch

[Yijia Huang]

Created attachment 433016 [details]
Patch

[Filip Pizlo]

Comment on attachment 433016 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=433016&action=review

This is really cool!  I don't feel too strongly about it, but I'd recommend not doing canBeInternal/commitInternal for this one.

> Source/JavaScriptCore/b3/B3LowerToAir.cpp:2616
> +                if (!canBeInternal(right) || !imm(right->child(1)) || right->child(1)->asInt() < 0)

I wouldn't do canBeInternal for this one.  I'd assume that an add-with-shift instruction is no more expensive than an add instruction.  So, if we have a program that does a shift and then uses the result of that shift both directly and for an add, then emitting the add as an add-with-shift instruction is fine.  In fact, it might even be more efficient, since it removes a dependency.  If you have a shift and then an add that depends on the shift, then the add can't run until the shift finishes.  If you have a shift and then an add-with-shift that depends on the same input as the shift, then both the shift and the add-with-shift can execute in parallel.

[Filip Pizlo]

(In reply to Filip Pizlo from comment #8)
> Comment on attachment 433016 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=433016&action=review
> 
> This is really cool!  I don't feel too strongly about it, but I'd recommend
> not doing canBeInternal/commitInternal for this one.
> 
> > Source/JavaScriptCore/b3/B3LowerToAir.cpp:2616
> > +                if (!canBeInternal(right) || !imm(right->child(1)) || right->child(1)->asInt() < 0)
> 
> I wouldn't do canBeInternal for this one.  I'd assume that an add-with-shift
> instruction is no more expensive than an add instruction.  So, if we have a
> program that does a shift and then uses the result of that shift both
> directly and for an add, then emitting the add as an add-with-shift
> instruction is fine.  In fact, it might even be more efficient, since it
> removes a dependency.  If you have a shift and then an add that depends on
> the shift, then the add can't run until the shift finishes.  If you have a
> shift and then an add-with-shift that depends on the same input as the
> shift, then both the shift and the add-with-shift can execute in parallel.

Disregard.  I think you're right to use canBeInternal/commitInternal for this one.

[Saam Barati]

Comment on attachment 433016 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=433016&action=review

> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:556
> +        ASSERT(n != ARM64Registers::sp || m != ARM64Registers::sp);
> +        if (m == ARM64Registers::sp)
> +            std::swap(n, m);

Does this actually work when the shift value isn't zero? Aren't we swapping what gets shifted? Maybe we can assert against this use case instead if my analysis is right?

> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:564
> +        ASSERT(n != ARM64Registers::sp || m != ARM64Registers::sp);
> +        if (m == ARM64Registers::sp)
> +            std::swap(n, m);

ditto

> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:572
> +        if (m == ARM64Registers::sp)
> +            std::swap(n, m);

ditto

> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:580
> +        if (m == ARM64Registers::sp)
> +            std::swap(n, m);

ditto

> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:588
> +        if (m == ARM64Registers::sp)
> +            std::swap(n, m);

ditto

> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:596
> +        if (m == ARM64Registers::sp)
> +            std::swap(n, m);

ditto

[Yijia Huang]

Created attachment 433077 [details]
Patch

[Saam Barati]

Comment on attachment 433077 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=433077&action=review

> Source/JavaScriptCore/b3/B3LowerToAir.cpp:2608
> +                    return Air::Nop;

nit: I'd return Oops to be consistent with tryOpcodeForType

> Source/JavaScriptCore/b3/B3LowerToAir.cpp:2614
> +                if (!isValidForm(opcode, Arg::Tmp, Arg::Imm, Arg::Imm, Arg::Tmp)) 

Am I missing something, or is this patch doing anything? Aren't all your instructions you added of the form: "Tmp, Tmp, Imm, Tmp"?

[Yijia Huang]

(In reply to Saam Barati from comment #12)
> Comment on attachment 433077 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=433077&action=review
> 
> > Source/JavaScriptCore/b3/B3LowerToAir.cpp:2608
> > +                    return Air::Nop;
> 
> nit: I'd return Oops to be consistent with tryOpcodeForType

Air::Opcode only has Nop. And JSC::B3::Opcode has Oops.

> > Source/JavaScriptCore/b3/B3LowerToAir.cpp:2614
> > +                if (!isValidForm(opcode, Arg::Tmp, Arg::Imm, Arg::Imm, Arg::Tmp)) 
> 
> Am I missing something, or is this patch doing anything? Aren't all your
> instructions you added of the form: "Tmp, Tmp, Imm, Tmp"?

My bad. Will fix it.

[Yijia Huang]

(In reply to Yijia Huang from comment #13)
> (In reply to Saam Barati from comment #12)
> > Comment on attachment 433077 [details]
> > Patch
> > 
> > View in context:
> > https://bugs.webkit.org/attachment.cgi?id=433077&action=review
> > 
> > > Source/JavaScriptCore/b3/B3LowerToAir.cpp:2608
> > > +                    return Air::Nop;
> > 
> > nit: I'd return Oops to be consistent with tryOpcodeForType
> 
> Air::Opcode only has Nop. And JSC::B3::Opcode has Oops.

Found it. Sorry, ignore this.

> 
> > > Source/JavaScriptCore/b3/B3LowerToAir.cpp:2614
> > > +                if (!isValidForm(opcode, Arg::Tmp, Arg::Imm, Arg::Imm, Arg::Tmp)) 
> > 
> > Am I missing something, or is this patch doing anything? Aren't all your
> > instructions you added of the form: "Tmp, Tmp, Imm, Tmp"?
> 
> My bad. Will fix it.

[Yijia Huang]

Created attachment 433118 [details]
Patch

[Saam Barati]

Comment on attachment 433118 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=433118&action=review

Nice. r=me

> Source/JavaScriptCore/b3/testb3_2.cpp:4148
> +            checkUsesInstruction(*code, toCString("add      w0, w0, w1, lsl #", amount).data());

I'd be a bit wary of relying on the register allocator always allocating in these registers. Can't we just use a regexp here instead? I think Keith added a way to do that at some point.

Ditto to all the checks below.

[Yijia Huang]

Created attachment 433154 [details]
Patch

[Saam Barati]

Comment on attachment 433154 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=433154&action=review

> Tools/Scripts/webkitpy/common/config/contributors.json:5980
> +   "Yijia Huang" : {
> +      "emails" : [
> +         "[e-mail]"
> +      ],
> +      "nicks" : [
> +         "yijia"
> +      ],
> +      "status" : "committer"
> +   },

Let's do this in its own patch

[Yijia Huang]

Created attachment 433161 [details]
Patch

[Yijia Huang]

Created attachment 433184 [details]
Patch

[Yijia Huang]

Created attachment 433188 [details]
Patch

[EWS]

Committed r279773 (239542@main): <https://commits.webkit.org/239542@main>

All reviewed patches have been landed. Closing bug and clearing flags on attachment 433188 [details].