[AIEX] Shift G_CONCAT_VECTORS closer to the user #234
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
andcarminati
requested review from
abhinay-anubola,
abnikant,
gbossu,
khallouh,
konstantinschwarz,
martien-de-jong,
SagarMaheshwari99 and
stephenneuendorffer
as code owners
|
QoR results: The results quite modest at first glance, but the real goal of this PR is remove one PADD from GEMM in its innermost loop. GEMM aie-public: GEMM this PR: |
F-Stuckmann
reviewed
Nov 13, 2024
| void llvm::applyUpdToConcat(MachineInstr &MI, MachineRegisterInfo &MRI, | ||
| MachineIRBuilder &B, | ||
| std::map<unsigned, Register> &IndexRegMap) { | ||
| B.setInstrAndDebugLoc(MI); | ||
| B.setDebugLoc(MI.getDebugLoc()); |
There was a problem hiding this comment.
why aren't the debug location and insertion point for MachineInstr the same?
There was a problem hiding this comment.
Hi @F-Stuckmann , in this case, the goal is to retain the same debug information (related to the instruction that we are replacing), but building the instruction in a different (shifted) position.
gbossu
reviewed
Nov 26, 2024
| MachineInstr &findClosestToUseInsertPoint(MachineInstr &MI, | ||
| MachineRegisterInfo &MRI) { | ||
|
|
||
| for (auto &User : MRI.use_instructions(MI.getOperand(0).getReg())) { |
There was a problem hiding this comment.
Nit: Maybe add a top-level comment like Find a use of \p MI in the same block where it can be moved
gbossu
reviewed
Nov 26, 2024
| for (auto &User : MRI.use_instructions(MI.getOperand(0).getReg())) { | ||
| if (User.isPHI()) | ||
| continue; | ||
| if (User.getParent() == MI.getParent() && canDelayMemOp(MI, User, MRI)) |
There was a problem hiding this comment.
Super-nit: I think we should really rename canDelayMemOp into just canDelayOp
There was a problem hiding this comment.
I agree, because it sounds strange to use it for non-memory instructions. On the other hand, this function check specific side effects of crossing memory operations, differently from canAdvanceOp where the instruction of interest is assumed to not be a load/store. What do you think?
There was a problem hiding this comment.
I think we can leave it like this, it might just be a bit too conservative when encountering a store. We will need to revamp it as to point anyway if we want the combiners to move past certain intrinsics with side effects.
When commiting applyUpdToConcat. This enables more postinc combiner cases.
andcarminati
force-pushed
the
andreu.concat.insert.point
branch
from
1c65cc2 to
cb3924b
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When commiting applyUpdToConcat. This enables more postinc combiner cases.