[LiveDebugValues] Disable spill tracking on AMDGPU (llvm#561)

The spill tracking in MLocTracker has quite poor performance on AMDGPU because of the large number of subregister indices. Futhermore, range extension doesn't even work on spills since AMDGPU currently doesn't implement is{LoadFrom,StoreTo}StackSlotPostFE. Disable it for now until we can come up with a better way of handling this.

Fixes SWDEV-511880

Author: epilk

llvm/lib/CodeGen/LiveDebugValues/InstrRefBasedImpl.cpp

@@ -156,6 +156,15 @@ static cl::opt<unsigned>
                          cl::desc("livedebugvalues-stack-ws-limit"),
                          cl::init(250));
 
+// Limit for the maximum number of stack slot indexes. On targets where this is
+// exceeded, this effectivly disables tracking debug locations across spills.
+// The spill tracking in MLocTracker performs quite poorly in terms of memory
+// and time on targets with a more complicated register file (FIXME).
+static cl::opt<unsigned>
+    StackSlotIdxesLimit("livedebugvalues-max-stack-slot-idxes", cl::Hidden,
+                        cl::desc("livedebugvalues-max-stack-slot-idxes"),
+                        cl::init(128));
+
 DbgOpID DbgOpID::UndefID = DbgOpID(0xffffffff);
 
 /// Tracker for converting machine value locations and variable values into
@@ -1122,6 +1131,10 @@ void MLocTracker::writeRegMask(const MachineOperand *MO, unsigned CurBB,
 }
 
 std::optional<SpillLocationNo> MLocTracker::getOrTrackSpillLoc(SpillLoc L) {
+  // Disable spill tracking on targets with a large number of slot idxes.
+  if (NumSlotIdxes >= StackSlotIdxesLimit)
+    return std::nullopt;
+
   SpillLocationNo SpillID(SpillLocs.idFor(L));
 
   if (SpillID.id() == 0) {
@@ -3721,6 +3734,15 @@ bool InstrRefBasedLDV::ExtendRanges(MachineFunction &MF,
   VTracker = nullptr;
   TTracker = nullptr;
 
+  if (MTracker->NumSlotIdxes >= StackSlotIdxesLimit) {
+    LLVM_DEBUG(
+        dbgs() << "Disabling InstrRefBasedLDV spill tracking for "
+               << MF.getName()
+               << " since target has too many potential stack slot indexes ("
+               << MTracker->NumSlotIdxes << ", limit is " << StackSlotIdxesLimit
+               << ")\n");
+  }
+
   SmallVector<MLocTransferMap, 32> MLocTransfer;
   SmallVector<VLocTracker, 8> vlocs;
   LiveInsT SavedLiveIns;

llvm/test/DebugInfo/AMDGPU/live-debug-values-spill-tracking.mir

@@ -0,0 +1,71 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -run-pass livedebugvalues %s -o - -debug-only livedebugvalues 2>&1 | FileCheck %s
+
+# Verify that spill tracking is disabled on amdgcn.
+
+# CHECK: Disabling InstrRefBasedLDV spill tracking for kern since target has too many potential stack slot indexes
+
+--- |
+  define void @kern() #0 !dbg !9 {
+    ret void, !dbg !15
+  }
+
+  attributes #0 = { noinline nounwind optnone "target-cpu"="gfx1100" }
+
+  !llvm.dbg.cu = !{!0}
+  !llvm.module.flags = !{!2, !3, !4, !5, !6, !7}
+  !llvm.ident = !{!8}
+
+  !0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !1, producer: "clang version 19.0.0", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, splitDebugInlining: false, nameTableKind: None)
+  !1 = !DIFile(filename: "t.cpp", directory: "/")
+  !2 = !{i32 1, !"amdhsa_code_object_version", i32 500}
+  !3 = !{i32 7, !"Dwarf Version", i32 5}
+  !4 = !{i32 2, !"Debug Info Version", i32 3}
+  !5 = !{i32 1, !"wchar_size", i32 4}
+  !6 = !{i32 8, !"PIC Level", i32 2}
+  !7 = !{i32 7, !"frame-pointer", i32 2}
+  !8 = !{!"clang version 19.0.0"}
+  !9 = distinct !DISubprogram(name: "kern", linkageName: "kern", scope: !1, file: !1, line: 1, type: !10, scopeLine: 1, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition, unit: !0, retainedNodes: !12)
+  !10 = !DISubroutineType(types: !11)
+  !11 = !{}
+  !12 = !{!13}
+  !13 = !DILocalVariable(name: "var", scope: !9, file: !1, line: 1, type: !14)
+  !14 = !DIBasicType(name: "i32", size: 32, encoding: DW_ATE_signed)
+  !15 = !DILocation(line: 1, column: 1, scope: !9)
+
+...
+---
+name: kern
+tracksRegLiveness: true
+debugInstrRef: true
+stack:
+  - { id: 0, name: '', type: spill-slot, offset: 0, size: 4, alignment: 4 }
+machineFunctionInfo:
+  scratchRSrcReg:  $sgpr0_sgpr1_sgpr2_sgpr3
+  stackPtrOffsetReg: '$sgpr32'
+  hasSpilledVGPRs: true
+body:             |
+  bb.0:
+    ; CHECK-LABEL: name: kern
+    ; CHECK: frame-setup CFI_INSTRUCTION escape 0x0f, 0x09, 0x90, 0x40, 0x94, 0x04, 0x36, 0x24, 0x36, 0xe9, 0x02
+    ; CHECK-NEXT: frame-setup CFI_INSTRUCTION llvm_register_pair $pc_reg, $sgpr30, 32, $sgpr31, 32
+    ; CHECK-NEXT: frame-setup CFI_INSTRUCTION undefined $vgpr0
+    ; CHECK-NEXT: DBG_INSTR_REF !13, !DIExpression(DIOpArg(0, i32)), dbg-instr-ref(1, 0), debug-location !15
+    ; CHECK-NEXT: DBG_VALUE_LIST !13, !DIExpression(DIOpArg(0, i32)), $noreg, debug-location !15
+    ; CHECK-NEXT: $vgpr0 = V_MOV_B32_e32 0, implicit $exec, debug-instr-number 1, debug-location !15
+    ; CHECK-NEXT: DBG_VALUE_LIST !13, !DIExpression(DIOpArg(0, i32)), $vgpr0, debug-location !15
+    ; CHECK-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, debug-location !15 :: (store (s32) into %stack.0, addrspace 5)
+    ; CHECK-NEXT: S_NOP 0, debug-location !15
+    ; CHECK-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, debug-location !15 :: (load (s32) from %stack.0, addrspace 5)
+    ; CHECK-NEXT: S_ENDPGM 0, debug-location !15
+    frame-setup CFI_INSTRUCTION escape 0x0f, 0x09, 0x90, 0x40, 0x94, 0x04, 0x36, 0x24, 0x36, 0xe9, 0x02
+    frame-setup CFI_INSTRUCTION llvm_register_pair $pc_reg, $sgpr30, 32, $sgpr31, 32
+    frame-setup CFI_INSTRUCTION undefined $vgpr0
+    DBG_INSTR_REF !13, !DIExpression(DIOpArg(0, i32)), dbg-instr-ref(1, 0), debug-location !15
+    $vgpr0 = V_MOV_B32_e32 0, implicit $exec, debug-instr-number 1, debug-location !15
+    SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, debug-location !15 :: (store (s32) into %stack.0, addrspace 5)
+    S_NOP 0, debug-location !15
+    $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, debug-location !15 :: (load (s32) from %stack.0, addrspace 5)
+    S_ENDPGM 0, debug-location !15
+
+...