Stuckmann.multi.slot.pseudos #304

F-Stuckmann · 2025-01-23T10:13:59Z

Statically assign multi-slot pseudo instructions.

This PR needs GPR Renaming (Rename GPR Register in innermost loop body #245 ) to allow the post pipeliner the freedom to actually reach an II=8
Postpipeliner tripps up over conv2d_1 (added unittest) (misses Stage=3, II=8)

todos:

do not refactor SlotCounts into an own utilitiy function

llvm/lib/Target/AIE/AIEBaseInstrInfo.h

llvm/lib/Target/AIE/AIEBaseAddrSpaceInfo.h

llvm/lib/Target/AIE/AIEBaseInstrInfo.cpp

llvm/lib/Target/AIE/AIESlotCounts.cpp

llvm/lib/Target/AIE/AIEHazardRecognizer.cpp

llvm/lib/Target/AIE/AIEInterBlockScheduling.cpp

llvm/lib/Target/AIE/AIESlotCounts.cpp

llvm/lib/Target/AIE/AIESlotCounts.h

llvm/lib/Target/AIE/AIEPostPipeliner.cpp

llvm/lib/Target/AIE/AIEMachineScheduler.cpp

llvm/lib/Target/AIE/AIEMultiSlotPseudoInstrMaterialization.cpp

llvm/lib/Target/AIE/Utils/AIESlotCountsUtils.h

llvm/lib/Target/AIE/AIEMachineScheduler.cpp

llvm/lib/Target/AIE/AIEMultiSlotPseudoInstrMaterialization.cpp

llvm/lib/Target/AIE/AIEPostPipeliner.cpp

llvm/lib/Target/AIE/AIESlotCounts.h

F-Stuckmann · 2025-02-10T09:07:19Z

I moved the multi slot Assignment into the leaveBlock stage of leaveMBB, so Machine Scheduling happens on the materialized VLD_pseudo Instructions (i.e. VLDA/VLDB)
I left the getMemoryBanks in the AIEHazardRecognizer and passed a Reference of AIEHazardRecognizer into my function call.
I now collect all the addrspace and Slots, decide an assign, and then set the Descriptor of the MachineInstr

llvm/lib/Target/AIE/MCTargetDesc/AIEMCFormats.h

martien-de-jong · 2025-02-18T16:21:13Z

llvm/lib/Target/AIE/MCTargetDesc/AIEMCFormats.cpp

@@ -60,4 +60,8 @@ const MCFormatDesc *AIEMCFormats::getMCFormats() const { return AIE::Formats; }

 const PacketFormats &AIEMCFormats::getPacketFormats() const { return Formats; }

+SmallVector<MCSlotKind, 2> AIEMCFormats::getLoadSlotKinds() const {
+  return {AIESlotKind::AIE_SLOT_LDB, AIESlotKind::AIE_SLOT_LDA};
+}


I'd like to forget AIE1. We won't hit the unreachable, will we?

we would hit unreachable, if i don't implement it for AIE1 and if the option is turned on by default, we would break stuff

Would we? I don't think AIE1 uses the MachineScheduler

llvm/lib/Target/AIE/AIEMultiSlotInstrMaterializer.cpp

llvm/lib/Target/AIE/AIEInterBlockScheduling.cpp

llvm/lib/Target/AIE/AIEMultiSlotInstrMaterializer.cpp

martien-de-jong

LGTM. Apply nits to taste

llvm/lib/Target/AIE/AIEMultiSlotInstrMaterializer.cpp

gbossu · 2025-02-19T10:46:25Z

llvm/lib/Target/AIE/AIEMultiSlotInstrMaterializer.cpp

+  /// \return an unused Slot from the mapping.
+  std::optional<MCSlotKind> getUnusedSlot(const AIEBaseInstrInfo *TII) const {
+    const SmallVector<MCSlotKind, 2> UnusedLoadSlots =
+        TII->getFormatInterface()->getLoadSlotKinds();


Nit: just LoadSlots?

Nit2: Maybe also rename the function to getUnusedLoadSlot(), because this is only to be used for multi-slot instructions that load.

It might actually just be better to iterate over the slots of the multi-slot instruction directly? I'm not sure why we need getLoadSlotKinds()

gbossu · 2025-02-19T10:49:11Z

llvm/lib/Target/AIE/AIEMultiSlotInstrMaterializer.cpp

+  std::optional<MCSlotKind>
+  getLeastRecentlyUsedSlot(const AIEBaseInstrInfo *TII) {
+    SmallVector<MCSlotKind, 2> AvailableSlots =
+        TII->getFormatInterface()->getLoadSlotKinds();


Same here: I guess this is fine for now, but I believe we'd rather just collect all the "feasible slots" by iterating over the multi-slot instructions, and create a least-recently-used list out of them, like we do in register-reallocation. This would avoid the need for a new getLoadSlotKinds() hook.

gbossu · 2025-02-19T11:04:51Z

Heads-up: I have now rebased #359 on top of aie-public. I think you can rebase as well.

Also, it would be nice to update the tests in llvm/test/CodeGen/AIE/aie2p/end-to-end/ so that they use --aie-multi-slot-pseudo-instr

.../CodeGen/AIE/aie2p/schedule/postpipeliner/multiSlotAssignment/future_conflict_assignment.mir

llvm/lib/Target/AIE/AIEBaseInstrInfo.h

F-Stuckmann · 2025-02-20T09:48:20Z

fixed githubs rebase issues

F-Stuckmann force-pushed the stuckmann.multi.slot.pseudos branch 2 times, most recently from 1b09f1e to 6f5a268 Compare February 5, 2025 10:15

martien-de-jong reviewed Feb 5, 2025

View reviewed changes

llvm/lib/Target/AIE/AIEBaseInstrInfo.h Outdated Show resolved Hide resolved

martien-de-jong reviewed Feb 5, 2025

View reviewed changes

llvm/lib/Target/AIE/AIEBaseAddrSpaceInfo.h Outdated Show resolved Hide resolved

martien-de-jong reviewed Feb 5, 2025

View reviewed changes

llvm/lib/Target/AIE/AIEBaseInstrInfo.cpp Outdated Show resolved Hide resolved

martien-de-jong reviewed Feb 5, 2025

View reviewed changes

llvm/lib/Target/AIE/AIESlotCounts.cpp Outdated Show resolved Hide resolved

gbossu reviewed Feb 5, 2025

View reviewed changes

martien-de-jong reviewed Feb 5, 2025

View reviewed changes

llvm/lib/Target/AIE/Utils/AIESlotCountsUtils.h Outdated Show resolved Hide resolved

martien-de-jong reviewed Feb 5, 2025

View reviewed changes

llvm/lib/Target/AIE/Utils/AIESlotCountsUtils.h Outdated Show resolved Hide resolved

martien-de-jong reviewed Feb 5, 2025

View reviewed changes

llvm/lib/Target/AIE/AIEMachineScheduler.cpp Outdated Show resolved Hide resolved

martien-de-jong reviewed Feb 5, 2025

View reviewed changes

llvm/lib/Target/AIE/AIEMultiSlotPseudoInstrMaterialization.cpp Outdated Show resolved Hide resolved