x86-64: Implement an enhancement for byte array and char array System.arraycopy

[a7ehuo]

Implement an enhancement for byte array and char array System.arraycopy on x86-64.

Also add JIT options that disable array copy enhancement:

• disableArrayCopyEnhancementByteArray: Disable array copy enhancement for 8 bit primitive array
• disableArrayCopyEnhancementCharArray: Disable array copy enhancement for 16 bit primitive array

[a7ehuo]

@0xdaryl May I ask you to review this change? Thank you!

@vijaysun-omr @hzongaro fyi

[a7ehuo]

x86-64 macOS failure is a known issue #7181

The following tests FAILED:
      2 - TestJitBuilderAPIGenerator (Failed)

[0xdaryl]

Can you attach a JIT log for a method that shows the code generated for these functions? Just a tracecg of a single method should be fine. Thanks.

[a7ehuo]

Can you attach a JIT log for a method that shows the code generated for these functions?

Search arrayCopy8BitPrimitiveEnhancementImplRoot8 for byte array or arrayCopy16BitPrimitiveEnhancementImplRoot16 for char array to jump to the code gen section on arraycopy

jitlog.arrayCopy8BitPrimitiveEnhancementImplRoot8.txt

jitlog.arrayCopy16BitPrimitiveEnhancementImplRoot16.txt

[0xdaryl]

Generally, I think the instruction sequences are reasonable. Please consider Brad’s suggestions where appropriate for using wider instructions on hardware that supports it.

My main feedback concerns the naming of these optimizations. Calling them “enhancements”, while technically correct, is too generic for someone who wasn’t involved in its development up front. There aren’t even any code comments to describe what the enhancements are, leaving others to read and understand the code to learn about them. I suggest incorporating something like “inlineSmallArraycopiesWithoutSTOS” into the naming would make it better. For example, “inlineSmall16BitPrimitiveArraycopiesWithoutSTOS”

Longer term, I’m not convinced that inlining this much code at each arraycopy call (over 50 instructions) is worth the simplicity and savings from calling out to a helper to do the same thing. I’m willing to let the current inline design go through now, but I would like a helper approach evaluated for the future. This might be fine for a method with one or two arraycopies, but if there are several arraycopies in a method (and we do see methods like that) the code footprint will add up quickly.

[a7ehuo]

I'm currently running sanity test on the changes that address the above comments. I'll push another update once the sanity test looks good

[a7ehuo]

@0xdaryl Comments are addressed. Ready for another review. Thank you!

• Fixed temp vector register type to TR_VRF
• Updated the function name and their corresponding option names to be more specific
  • arrayCopy8BitPrimitiveInlineSmallSizeWithoutREPMOVS
  • arrayCopy16BitPrimitiveInlineSmallSizeWithoutREPMOVS
• Added comments in the beginning of the implementation to give a detailed description of the inlined code sequence

[0xdaryl]

@BradleyWood : please approve if your comments are addressed

[a7ehuo]

Can you confirm you tested this code internally on both 64-bit and 32-bit?

I tested only on 64-bit machines since I thought the enhancement was guarded with checking if target is 64-bit. I'll test on 32-bit as well after addressing the latest comments.

bool enable8BitPrimitiveArrayCopyInlineSmallSizeWithoutREPMOVS = ((elementSize == 1) &&
                                                                     !disableEnhancement &&
                                                                     cg->comp()->target().cpu.supportsAVX() &&
                                                                     cg->comp()->target().is64Bit()) ? true : false;

bool enable16BitPrimitiveArrayCopyInlineSmallSizeWithoutREPMOVS = (!disableEnhancement &&
                                                                 cg->comp()->target().cpu.supportsAVX() &&
                                                                 cg->comp()->target().is64Bit()) ? true : false;

[a7ehuo]

@0xdaryl Addressed all review comments. Ready for another review. Thank you!

• Internal Jenkins test (ID 22283): JDK 8 sanity.functional,sanity.system,sanity.openjdk,extended.system,extended.functional with x86-32_windows, all passed.
• Internal Jenkins test (ID 22275): JDK11 and JDK 17 sanity.functional,sanity.system,sanity.openjdk,extended.system,extended.functional with x86-64_linux : Seen some failures but the failures are not related to this change

[0xdaryl]

@vijaysun-omr : FYI

[vijaysun-omr]

Since we want to keep all codegens aware of key changes in one, I thought I'd mention @zl-wang @r30shah and @knn-k here so that they can consider whatever feels relevant to their platform

[knn-k]

AArch64 codegen inlines arraycopy only when the copy length is a constant at the compile time. Otherwise it generates instructions to call arraycopy helpers in https://github.com/eclipse/omr/blob/master/compiler/aarch64/runtime/ARM64arrayCopy.spp Maybe I can add special helpers for arrays of short/int/long types that skip the checks for lower bits of the copy length.