Hi, is there anything else I can improve?

Split irrelevant comments change to 8fccdfa.

Add a comment about parameter DeduplicateRestoreBB of function insertIndirectBranch.

Ping for review :)

Update code and tests for LoongArch.

Add --disable-block-placement for LoongArch tests so that all far branches will not be optimized out.

Add tests for LoongArch.

Currently SIInstrinfo::insertIndirectBranch do something like this:

• Fix typo

Rebase and add CHECK-RV64 checks for relax_jal_spill_32_deduplicate_restore_block, as https://reviews.llvm.org/D130560#3746417 mentioned.

Yes, exactly. Thank you for your time on this patch!

Because if the generated restore blocks have the incorrect BasicBlock information, like the # %dest_3 for .LBB8_6 and .LBB6_10 before this change, they would be confusing. Although I believe it is also acceptable if they contain no corresponding BasicBlock information, it will be easier to understand if they do, such as the '# %dest 1' with '.LBB8_6' and '# %dest 2' with '.LBB6_10' after this change.

Move new tests requiring larger -amdgpu-s-branch-bits value to new file branch-relax-spill-deduplication.ll.

Removes unnecessary #include.

Ping for review :)

Ping for review :)

Ping for review :)

Rebase and ping for review :)

• Rebase
• Update the test with update_llc_test_checks.py

In D130560#3746417, @luismarques wrote:

In D130560#3746379, @luismarques wrote:

The tests don't seem to have been properly updated with update_llc_test_checks.py. llvm/test/CodeGen/RISCV/branch-relaxation.ll contains RV64 RUN lines but the corresponding CHECK lines are missing in some functions.

Looking more closely at this, I guess you tried to only include the CHECK-RV64 and CHECK-RV32 checks when relevant. That's a good instinct but I guess it goes a bit against how we normally use update_llc_test_checks.py. My understanding of the trade-off of using that tool is that the test updates are much easier, even if sometimes the CHECKs aren't as tight as something more tailormade.

Ping for kindly review :)

• Change the type of BasicBlockInfo::RestoreBlocks from SmallVector to TinyPtrVector.

• Rebase onto D130560.

Revise according to @dblaikie's comments.

Ping :) Do we have any updates on this?

Update comments.

Thank you! Would you mind taking a look at D129999? This patch depends on it for estimating functions' sizes.

• Revise according to @StephenFan's comments.

Ping :) Is there anything else I can do to improve this?

In D129999#3694257, @craig.topper wrote:

Can you update the title to say make "Mark MachineInstr with FrameIndex as not compressible". The phrase "incomplete MachineInstr" is not very clear. I would interpret it as a MachineInstr with less operands than it's supposed to have. A MachineInstr with a FrameIndex is a completely valid MachineInstr so I wouldn't call it incomplete.

Sure! I'll explain the details about why I want to make this change.

Ping for kindly review :)

In D130560#3684918, @CoelacanthusHex wrote:

There seems to be a problem here: if we perform only some operations on registers without sd after a jump (which can easily happen if only a few short instructions are executed between jumps), we then jump to another place. If we still restore immediately in this case, we are actually wasting resources.

Reuse scavenging spill slot for scratch register's spill slot in branch relaxation.

Remove unnecessary #include.

Fix code style via clang-format.

Add test for a large stack that requires adjusting the offset of BranchRelaxationScratchFrameIndex.

Fix code style via clang-format.

• More conservative estimate for branches in EstimateFunctionSizeInBytes.

ARM hard-coded EK_Inline as jump table encoding in rG8d3ba7349c2137f1a7f68c0ce5c1345496fc0b63. As a result, std::max(JTI->getEntrySize(DL), JTI->getEntryAlignment(DL)) always returns 1, which causes test/CodeGen/Thumb/large-fn-switch.ll to fail.

Update test.

Fix compile error.

(included more context)