Bug 228543 – [JSC] load/store with BaseIndex is inefficient in ARM64

• Component: New Bugs
• Status: RESOLVED FIXED

[Yusuke Suzuki]

[JSC] load/store with BaseIndex is inefficient in ARM64

[Yusuke Suzuki]

Created attachment 434411 [details]
Patch

[Mark Lam]

Comment on attachment 434411 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434411&action=review

> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:1433
> +        if (!address.scale || address.scale == 3) {

Can you use the Scale enum values instead here?
    if (address.scale == TimesOne || address.scale == TimesEight) {

> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:1435
> +            if (auto baseGPR = tryFoldBaseAndOffsetPart(address)) {
> +                m_assembler.ldr<64>(dest, baseGPR.value(), address.index, indexExtendType(address), address.scale);

I think you can do better: if (scale == TimesEight && isUInt12(address.offset / 8)), then you can use the `ALWAYS_INLINE void ldr(RegisterID rt, RegisterID rn, unsigned pimm)` form and just issue 1 instruction to do this load.  See also section C6.2.101 of the spec.

See also LDUR in section C6.2.136 if scale == TimesOne.

[Mark Lam]

Comment on attachment 434411 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434411&action=review

>> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:1435
>> +                m_assembler.ldr<64>(dest, baseGPR.value(), address.index, indexExtendType(address), address.scale);
> 
> I think you can do better: if (scale == TimesEight && isUInt12(address.offset / 8)), then you can use the `ALWAYS_INLINE void ldr(RegisterID rt, RegisterID rn, unsigned pimm)` form and just issue 1 instruction to do this load.  See also section C6.2.101 of the spec.
> 
> See also LDUR in section C6.2.136 if scale == TimesOne.

I should clarify: my point is that you may be able to do this with just 1 instruction instead of 2.  Ditto for other cases below.  So, I think it's worth checking the spec manual for available instructions.

[Yusuke Suzuki]

Comment on attachment 434411 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434411&action=review

>> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:1433
>> +        if (!address.scale || address.scale == 3) {
> 
> Can you use the Scale enum values instead here?
>     if (address.scale == TimesOne || address.scale == TimesEight) {

Changed.

>>> Source/JavaScriptCore/assembler/MacroAssemblerARM64.h:1435
>>> +                m_assembler.ldr<64>(dest, baseGPR.value(), address.index, indexExtendType(address), address.scale);
>> 
>> I think you can do better: if (scale == TimesEight && isUInt12(address.offset / 8)), then you can use the `ALWAYS_INLINE void ldr(RegisterID rt, RegisterID rn, unsigned pimm)` form and just issue 1 instruction to do this load.  See also section C6.2.101 of the spec.
>> 
>> See also LDUR in section C6.2.136 if scale == TimesOne.
> 
> I should clarify: my point is that you may be able to do this with just 1 instruction instead of 2.  Ditto for other cases below.  So, I think it's worth checking the spec manual for available instructions.

I don't think it is available since here there is base and index registers (since this is BaseIndex). But `ALWAYS_INLINE void ldr(RegisterID rt, RegisterID rn, unsigned pimm)` only takes base register.
So if we would like to use `ALWAYS_INLINE void ldr(RegisterID rt, RegisterID rn, unsigned pimm)`, then we first need to fold index into base. This means that it requires 2 instructions anyway.

[Yusuke Suzuki]

Created attachment 434434 [details]
Patch

[Mark Lam]

Comment on attachment 434434 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434434&action=review

r=me with suggestions.

> Source/JavaScriptCore/assembler/testmasm.cpp:4224
> +        uint64_t array[] = { 1, 2, 3, 4, 5, };
> +        CHECK_EQ(invoke<uint64_t>(test, array, static_cast<UCPURegister>(3)), 3);

nit: can you add  INT64_MAX to each of the values in array so that this test will also catch bugs where the load instruction doesn't load the full 8 bytes?

> Source/JavaScriptCore/assembler/testmasm.cpp:4234
> +        uint64_t array[] = { 1, 2, 3, 4, 5, };
> +        CHECK_EQ(invoke<uint64_t>(test, array, static_cast<UCPURegister>(3)), 5);

Ditto.

> Source/JavaScriptCore/assembler/testmasm.cpp:4247
> +        uint32_t array[] = { 1, 2, 3, 4, 5, };
> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), 3);

Ditto: add INT32_MAX.

> Source/JavaScriptCore/assembler/testmasm.cpp:4257
> +        uint32_t array[] = { 1, 2, 3, 4, 5, };
> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), 5);

Ditto.

> Source/JavaScriptCore/assembler/testmasm.cpp:4269
> +        uint16_t array[] = { 1, 2, 3, 4, 5, };
> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), 3);

Ditto: add INT16_MAX.

> Source/JavaScriptCore/assembler/testmasm.cpp:4279
> +        uint16_t array[] = { 1, 2, 3, 4, static_cast<uint16_t>(-1), };
> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), 0xffff);

Ditto.

> Source/JavaScriptCore/assembler/testmasm.cpp:4291
> +        uint16_t array[] = { 1, 2, 3, 4, 5, };
> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), 3);

nit: add (INT16_MAX - 10)?

> Source/JavaScriptCore/assembler/testmasm.cpp:4301
> +        uint16_t array[] = { 1, 2, 3, 4, static_cast<uint16_t>(-1), };
> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), static_cast<uint32_t>(-1));

nit: add (INT16_MAX - 10) but make leave the last one as -1?

> Source/JavaScriptCore/assembler/testmasm.cpp:4406
> +        invoke<void>(test, array, 3, 42);
> +        CHECK_EQ(array[2], 42);

nit: Add INT64_MAX to 42.

> Source/JavaScriptCore/assembler/testmasm.cpp:4417
> +        invoke<void>(test, array, 3, 42);
> +        CHECK_EQ(array[4], 42);

Ditto.

> Source/JavaScriptCore/assembler/testmasm.cpp:4431
> +        invoke<void>(test, array, 3, 42);
> +        CHECK_EQ(array[2], 42);

nit: Add INT32_MAX to 42.

> Source/JavaScriptCore/assembler/testmasm.cpp:4442
> +        invoke<void>(test, array, 3, 42);
> +        CHECK_EQ(array[4], 42);

Ditto.

> Source/JavaScriptCore/assembler/testmasm.cpp:4455
> +        invoke<void>(test, array, 3, 42);
> +        CHECK_EQ(array[2], 42);

nit: Add INT16_MAX to 42.

> Source/JavaScriptCore/assembler/testmasm.cpp:4466
> +        invoke<void>(test, array, 3, 42);
> +        CHECK_EQ(array[4], 42);

Ditto.

[Yusuke Suzuki]

Comment on attachment 434434 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434434&action=review

Thanks!

>> Source/JavaScriptCore/assembler/testmasm.cpp:4224
>> +        CHECK_EQ(invoke<uint64_t>(test, array, static_cast<UCPURegister>(3)), 3);
> 
> nit: can you add  INT64_MAX to each of the values in array so that this test will also catch bugs where the load instruction doesn't load the full 8 bytes?

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4234
>> +        CHECK_EQ(invoke<uint64_t>(test, array, static_cast<UCPURegister>(3)), 5);
> 
> Ditto.

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4247
>> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), 3);
> 
> Ditto: add INT32_MAX.

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4257
>> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), 5);
> 
> Ditto.

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4279
>> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), 0xffff);
> 
> Ditto.

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4291
>> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), 3);
> 
> nit: add (INT16_MAX - 10)?

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4301
>> +        CHECK_EQ(invoke<uint32_t>(test, array, static_cast<UCPURegister>(3)), static_cast<uint32_t>(-1));
> 
> nit: add (INT16_MAX - 10) but make leave the last one as -1?

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4406
>> +        CHECK_EQ(array[2], 42);
> 
> nit: Add INT64_MAX to 42.

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4417
>> +        CHECK_EQ(array[4], 42);
> 
> Ditto.

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4431
>> +        CHECK_EQ(array[2], 42);
> 
> nit: Add INT32_MAX to 42.

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4442
>> +        CHECK_EQ(array[4], 42);
> 
> Ditto.

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4455
>> +        CHECK_EQ(array[2], 42);
> 
> nit: Add INT16_MAX to 42.

Changed.

>> Source/JavaScriptCore/assembler/testmasm.cpp:4466
>> +        CHECK_EQ(array[4], 42);
> 
> Ditto.

Changed.

[Yusuke Suzuki]

Committed r280391 (240032@main): <https://commits.webkit.org/240032@main>

[Radar WebKit Bug Importer]

<rdar://problem/81227865>

[Yusuke Suzuki]

There is 0% chance that https://ews-build.webkit.org/#/builders/56/builds/12242 failures are related to this patch since

1. This is x64 EWS
2. This patch does nothing on x64. It only affects on ARM64.

[Yusuke Suzuki]

(In reply to Yusuke Suzuki from comment #10)
> There is 0% chance that
> https://ews-build.webkit.org/#/builders/56/builds/12242 failures are related
> to this patch since
> 
> 1. This is x64 EWS
> 2. This patch does nothing on x64. It only affects on ARM64.

Looks like checking-out is failing. So, it is EWS issue.