Add Pre/Post-Indexed Address Mode to Air for ARM64

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Yijia Huang]

...

[Yijia Huang]

Created attachment 433730 [details]
Patch

[Yijia Huang]

Created attachment 433734 [details]
Patch

[Yijia Huang]

Created attachment 433738 [details]
Patch

[Yijia Huang]

Created attachment 433859 [details]
Patch

[Yijia Huang]

Created attachment 433864 [details]
Patch

[Yijia Huang]

Created attachment 433897 [details]
Patch

[Yijia Huang]

Created attachment 434069 [details]
Patch

[Yijia Huang]

Created attachment 434096 [details]
Patch

[Yijia Huang]

Comment on attachment 434069 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434069&action=review

> Source/JavaScriptCore/b3/B3ReduceStrength.cpp:522
> +                                m_potentialPreIndex.remove(it);

This remove() doesn't destroy the MemoryValue*.

> Source/JavaScriptCore/b3/B3ReduceStrength.cpp:550
> +                m_potentialPreIndex.remove(it);

Why does this remove() destroy the MemoryValue* here?

[Yijia Huang]

Created attachment 434098 [details]
Patch

[Radar WebKit Bug Importer]

<rdar://problem/81053101>

[Yijia Huang]

Created attachment 434166 [details]
Patch

[Yijia Huang]

Created attachment 434167 [details]
Patch

[Yijia Huang]

Created attachment 434174 [details]
Patch

[Yijia Huang]

Created attachment 434178 [details]
Patch

[Yijia Huang]

Created attachment 434179 [details]
Patch

[Yijia Huang]

Created attachment 434181 [details]
Patch

[Yijia Huang]

Created attachment 434182 [details]
Patch

[Yijia Huang]

Created attachment 434197 [details]
Patch

[Filip Pizlo]

Comment on attachment 434197 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434197&action=review

> Source/JavaScriptCore/b3/B3LowerToAir.cpp:916
> +                    auto tryPotentialPostIndexAddress = [&] () {
> +                        if (!right->hasIntPtr())
> +                            return;
> +                        intptr_t offset = right->asIntPtr();
> +                        Value::OffsetType smallOffset = static_cast<Value::OffsetType>(offset);
> +                        if (opcode != Air::Add64 || smallOffset != offset || !Arg::isValidPostIndexForm(smallOffset))
> +                            return;
> +                        m_baseToInst.add(left, &m_insts.last().last());
> +                    };

I think it's weird to have this code here.  Seems like you want it at the top of the Add case in the main lowering switch statement.

> Source/JavaScriptCore/b3/B3LowerToAir.cpp:1164
> -                    if (isValidForm(Store32, Arg::ZeroReg, dest.kind()) && dest.isValidForm(Width32))
> -                        return Inst(Store32, m_value, zeroReg(), dest);
> +                    if (isValidForm(Move32, Arg::ZeroReg, dest.kind()) && dest.isValidForm(Width32))
> +                        return Inst(Move32, m_value, zeroReg(), dest);

Why do you need to name stores as such?

> Source/JavaScriptCore/b3/B3LowerToAir.cpp:2597
> +                if (totalUseCounts(address) == 1 || totalUseCounts(base) > 1)

I don't get this.  What are you trying to achieve here?

Isn't the whole point of pre-increment loads that the address has many uses?

[Yijia Huang]

Created attachment 434272 [details]
Patch

[Yijia Huang]

Created attachment 434273 [details]
Patch

[Yijia Huang]

Created attachment 434294 [details]
Patch

[Yijia Huang]

Created attachment 434376 [details]
Patch

[Yijia Huang]

Created attachment 434416 [details]
Patch

[Yijia Huang]

This is the 5th version.

[Yijia Huang]

Created attachment 434567 [details]
Patch

[Yijia Huang]

Created attachment 434573 [details]
Patch

[Filip Pizlo]

Comment on attachment 434573 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434573&action=review

I think you have a bug in how your new Arg works - it says that it just Uses its input.

Also, you should factor your phase out into its own file.

> Source/JavaScriptCore/b3/B3ReduceStrength.cpp:3165
> +bool canonicalize(Procedure& proc)
> +{
> +    PhaseScope phaseScope(proc, "canonicalize");
> +    ReduceStrength reduceStrength(proc);
> +    return reduceStrength.canonicalizeIncrementAddress();
> +}

I think you need a more descriptive name at the B3 level namespace than "canonicalize".  I would call it something like "canonicalizePrePostIncrements" and maybe I would even conditionalize it so it only runs on ARM64.

Also, I wouldn't make this a method of ReduceStrength.  It doesn't seem like it needs to be.  Seems like this should be a standalone function, in its own file.  Like "B3CanonicalizePrePostIncrements.h" and "B3CanonicalizePrePostIncrements.cpp".

> Source/JavaScriptCore/b3/air/AirArg.h:1395
>              functor(m_base, Use, GP, argRole == UseAddr ? argWidth : pointerWidth());

This is wrong now - it'll claim that pre/post increment "uses" are just Uses.  They are UseDef.

[Yijia Huang]

Created attachment 434582 [details]
Patch

[Filip Pizlo]

Comment on attachment 434582 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434582&action=review

I think this is right?

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.cpp:167
> +    return canonicalizePrePostIncrementsImpl(proc);

You could inline the body of the impl here.

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.h:34
> +bool canonicalizePrePostIncrementsImpl(Procedure&);

Don't need this declared here since you don't call it from anywhere but the non-impl variant.

> Source/JavaScriptCore/b3/air/AirArg.h:1398
> +        case PreIndex:
> +        case PostIndex:
> +            functor(m_base, UseDef, GP, argRole == UseAddr ? argWidth : pointerWidth());
> +            break;

Fingers crossed that this works!

Ideally we'd audit the compiler to make sure nobody is assuming that only a top-level Arg can be a Def.

[Yijia Huang]

Created attachment 434610 [details]
Patch for fixing the layout test failure

[Yijia Huang]

Created attachment 434641 [details]
Patch for fixing the layout test failure

[Yijia Huang]

Created attachment 434656 [details]
Patch for landing

[EWS]

Committed r280493 (240125@main): <https://commits.webkit.org/240125@main>

All reviewed patches have been landed. Closing bug and clearing flags on attachment 434656 [details].

[Saam Barati]

Comment on attachment 434656 [details]
Patch for landing

View in context: https://bugs.webkit.org/attachment.cgi?id=434656&action=review

> Source/JavaScriptCore/b3/B3LowerToAir.cpp:2568
> +            //     memory = load(base, 0)
> +            //     address = add(base, offset)

This comment is not what the code is doing. The order here is backwards of the canonical form.

Might also be worth noting what Air code we emit for clarity?

> Source/JavaScriptCore/b3/B3LowerToAir.cpp:2573
> +                Value* address = m_block->at(m_index - 1);

Why do we do this if address is already locked? I guess it's sound because we are always writing to its temp? That said, if it's locked, some other operation already sucked up this add internal to the other operation. And nobody would use the result of the add. But we're emitting a move, which may not get elided.

[Yijia Huang]

Comment on attachment 434656 [details]
Patch for landing

View in context: https://bugs.webkit.org/attachment.cgi?id=434656&action=review

Thanks for reviewing.

>> Source/JavaScriptCore/b3/B3LowerToAir.cpp:2568
>> +            //     address = add(base, offset)
> 
> This comment is not what the code is doing. The order here is backwards of the canonical form.
> 
> Might also be worth noting what Air code we emit for clarity?

The order is fixed in the patch for store instruction. I will add the corresponding Air code in the store patch.

>> Source/JavaScriptCore/b3/B3LowerToAir.cpp:2573
>> +                Value* address = m_block->at(m_index - 1);
> 
> Why do we do this if address is already locked? I guess it's sound because we are always writing to its temp? That said, if it's locked, some other operation already sucked up this add internal to the other operation. And nobody would use the result of the add. But we're emitting a move, which may not get elided.

I will add a m_locked check on address in the store patch.

[Yijia Huang]

Comment on attachment 434656 [details]
Patch for landing

Thanks for reviewing.

[Saam Barati]

Comment on attachment 434656 [details]
Patch for landing

View in context: https://bugs.webkit.org/attachment.cgi?id=434656&action=review

still reviewing, but a few more comments.

> Source/JavaScriptCore/assembler/AbstractMacroAssembler.h:250
> +    // Describes an address with base address and pre-increment/decrement index.

nit: index => offset everywhere in these classes

> Source/JavaScriptCore/assembler/AbstractMacroAssembler.h:264
> +    // Describes an address with base address and post-increment/decrement index.

ditto

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.cpp:55
> +    BasicBlock* block { nullptr };

this field is totally unused except for one store to it in the candidate search loop

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.cpp:76
> +    auto tryPrePostIndex = [&] (Value* value) {

nit: this should be called something like "tryAddPrePostIndexCandidate"

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.cpp:110
> +                if (smallOffset != offset || !Arg::isValidPreIndexForm(smallOffset))

I think we should just have one function, but if we stick with two, this should be isValidPostIndexForm, not isValidPreIndexForm

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.cpp:144
> +    auto detect = [&] (HashMap<MemoryValue*, Value*> candidates) {

this should be taking like a "const HashMap&" or HashMap&. No need to take the hash map by value.

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.cpp:145
> +        for (auto itr = candidates.begin(); itr != candidates.end(); ++itr) {

you should be using the good style for loops:

for (auto pair : candidates)

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.cpp:146
> +            MemoryValue* memory = itr->key;

then pair.key

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.cpp:147
> +            Value* address = itr->value;

and pair.value

> Source/JavaScriptCore/b3/air/AirArg.h:1296
> +    template<typename Int, typename = Value::IsLegalOffset<Int>>
> +    static bool isValidPreIndexForm(Int offset)
> +    {
> +        if (isARM64())
> +            return isValidSignedImm9(offset);
> +        return false;
> +    }
> +
> +    template<typename Int, typename = Value::IsLegalOffset<Int>>
> +    static bool isValidPostIndexForm(Int offset)
> +    {
> +        if (isARM64())
> +            return isValidSignedImm9(offset);
> +        return false;
> +    }

Can we just make one function "isValidIncrementIndexForm"?

[Saam Barati]

Comment on attachment 434656 [details]
Patch for landing

View in context: https://bugs.webkit.org/attachment.cgi?id=434656&action=review

> Source/JavaScriptCore/b3/B3CanonicalizePrePostIncrements.cpp:151
> +            if (candidates == preIndexCandidates) {

This is doing a hash table comparison by value each iteration. That's a crazy O(n) operation. We should do this by passing in a flag or something.
Also, if we pass in the flag, we don't need to check full control flow equivalence. We know which should forward dominate, which should backwards dominate.