[AMDGPU] Model distance to instruction in bundle

[rampitec]

Summary

This change allows to model the height of the instruction
within a bundle for latency adjustment purposes.

[foad]

LGTM

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp

void GCNSubtarget::adjustSchedDependency(SUnit *Src, SUnit *Dst,
   MachineInstr *SrcI = Src->getInstr();
   MachineInstr *DstI = Dst->getInstr();
 
   if (SrcI->isBundle()) {
     const SIRegisterInfo *TRI = getRegisterInfo();
     auto Reg = Dep.getReg();
     MachineBasicBlock::const_instr_iterator I(SrcI->getIterator());
     MachineBasicBlock::const_instr_iterator E(SrcI->getParent()->instr_end());
+    unsigned Lat = 0;
     for (++I; I != E && I->isBundledWithPred(); ++I) {
-      if (!I->modifiesRegister(Reg, TRI))
-        continue;
-      Dep.setLatency(InstrInfo.getInstrLatency(getInstrItineraryData(), *I));
-      break;
+      if (I->modifiesRegister(Reg, TRI))
+        Lat = InstrInfo.getInstrLatency(getInstrItineraryData(), *I);
+      else if (Lat)
+        --Lat;
     }
+    Dep.setLatency(Lat);
   } else if (DstI->isBundle()) {
-    Dep.setLatency(InstrInfo.getInstrLatency(getInstrItineraryData(), *SrcI));
+    const SIRegisterInfo *TRI = getRegisterInfo();
+    auto Reg = Dep.getReg();
+    MachineBasicBlock::const_instr_iterator I(DstI->getIterator());
+    MachineBasicBlock::const_instr_iterator E(DstI->getParent()->instr_end());
+    unsigned Lat = InstrInfo.getInstrLatency(getInstrItineraryData(), *SrcI);
+    for (++I; I != E && I->isBundledWithPred() && Lat; ++I) {
+      if (I->readsRegister(Reg, TRI))
+        break;
+      --Lat;
+    }
+    Dep.setLatency(Lat);
   }
 }
 
 namespace {
 struct MemOpClusterMutation : ScheduleDAGMutation {
   const SIInstrInfo *TII;
 
   MemOpClusterMutation(const SIInstrInfo *tii) : TII(tii) {}

llvm/test/CodeGen/AMDGPU/bundle-latency.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs -run-pass=post-RA-sched %s -o - | FileCheck -check-prefix=GCN %s
+
+# Check that we move consumer further from producer, even if one of them is in a bundle.
+
+---
+name:            src_bundle_latency
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    ; GCN-LABEL: name: src_bundle_latency
+    ; GCN: $vgpr0, $vgpr1 = BUNDLE undef $vgpr3_vgpr4, implicit $exec {
+    ; GCN:   $vgpr0 = GLOBAL_LOAD_DWORD undef $vgpr3_vgpr4, 0, 0, 0, 0, implicit $exec
+    ; GCN:   $vgpr1 = GLOBAL_LOAD_DWORD undef $vgpr3_vgpr4, 4, 0, 0, 0, implicit $exec
+    ; GCN: }
+    ; GCN: $vgpr6 = V_ADD_F32_e32 killed $vgpr0, $vgpr0, implicit $exec
+    ; GCN: $vgpr5 = V_ADD_F32_e32 killed $vgpr1, $vgpr1, implicit $exec
+    $vgpr0, $vgpr1 = BUNDLE undef $vgpr3_vgpr4, implicit $exec {
+      $vgpr0 = GLOBAL_LOAD_DWORD undef $vgpr3_vgpr4, 0, 0, 0, 0, implicit $exec
+      $vgpr1 = GLOBAL_LOAD_DWORD undef $vgpr3_vgpr4, 4, 0, 0, 0, implicit $exec
+    }
+    $vgpr5 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $exec
+    $vgpr6 = V_ADD_F32_e32 $vgpr0, $vgpr0, implicit $exec
+...
+
+---
+name:            dst_bundle_latency
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    ; GCN-LABEL: name: dst_bundle_latency
+    ; GCN: $vgpr1 = V_ADD_F32_e32 undef $vgpr6, undef $vgpr6, implicit $exec
+    ; GCN: $vgpr0 = V_ADD_F32_e32 undef $vgpr5, undef $vgpr5, implicit $exec
+    ; GCN: BUNDLE killed $vgpr0, killed $vgpr1, undef $vgpr3_vgpr4, implicit $exec {
+    ; GCN:   GLOBAL_STORE_DWORD undef $vgpr3_vgpr4, killed $vgpr1, 0, 0, 0, 0, implicit $exec
+    ; GCN:   GLOBAL_STORE_DWORD undef $vgpr3_vgpr4, killed $vgpr0, 4, 0, 0, 0, implicit $exec
+    ; GCN: }
+    $vgpr0 = V_ADD_F32_e32 undef $vgpr5, undef $vgpr5, implicit $exec
+    $vgpr1 = V_ADD_F32_e32 undef $vgpr6, undef $vgpr6, implicit $exec
+    BUNDLE $vgpr0, $vgpr1, undef $vgpr3_vgpr4, implicit $exec {
+      GLOBAL_STORE_DWORD undef $vgpr3_vgpr4, $vgpr1, 0, 0, 0, 0, implicit $exec
+      GLOBAL_STORE_DWORD undef $vgpr3_vgpr4, $vgpr0, 4, 0, 0, 0, implicit $exec
+    }
+...