ScheduleDAGInstrs: Skip debug instructions at end of scheduling region

[arsenm]

Summary

If the end instruction of the scheduling region was a DBG_VALUE, the
uses of the debug instruction were tracked as if they were real
uses. This would then hit the deadDefHasNoUse assertion in
addVRegDefDeps if the only use was the debug instruction.

[rampitec]

LGTM

[arsenm]

188df1742042610a4c9af1fff9943d3d2a2740c6

llvm/lib/CodeGen/ScheduleDAGInstrs.cpp

void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
   NumRegionInstrs = regioninstrs;
 }
 
 void ScheduleDAGInstrs::exitRegion() {
   // Nothing to do.
 }
 
 void ScheduleDAGInstrs::addSchedBarrierDeps() {
-  MachineInstr *ExitMI = RegionEnd != BB->end() ? &*RegionEnd : nullptr;
+  MachineInstr *ExitMI =
+      RegionEnd != BB->end()
+          ? &*skipDebugInstructionsBackward(RegionEnd, RegionBegin)
+          : nullptr;
   ExitSU.setInstr(ExitMI);
   // Add dependencies on the defs and uses of the instruction.
   if (ExitMI) {
     for (const MachineOperand &MO : ExitMI->operands()) {
       if (!MO.isReg() || MO.isDef()) continue;
       Register Reg = MO.getReg();
       if (Register::isPhysicalRegister(Reg)) {
         Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));

 /// Adds a register data dependency if the instruction that defines the
 /// virtual register used at OperIdx is mapped to an SUnit. Add a register
 /// antidependency from this SUnit to instructions that occur later in the same
 /// scheduling region if they write the virtual register.
 ///
 /// TODO: Handle ExitSU "uses" properly.
 void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
   const MachineInstr *MI = SU->getInstr();
+  assert(!MI->isDebugInstr());
+
   const MachineOperand &MO = MI->getOperand(OperIdx);
   Register Reg = MO.getReg();
 
   // Remember the use. Data dependencies will be added when we find the def.
   LaneBitmask LaneMask = TrackLaneMasks ? getLaneMaskForMO(MO)
                                         : LaneBitmask::getAll();
   CurrentVRegUses.insert(VReg2SUnitOperIdx(Reg, LaneMask, OperIdx, SU));
 

llvm/test/CodeGen/AMDGPU/dbg-value-ends-sched-region.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 -verify-machineinstrs -run-pass=machine-scheduler -o - %s | FileCheck %s
+# The DBG_VALUE in bb.5 ends a scheduling region, and its uses should
+# not be tracked like a normal instruction.
+
+--- |
+  define void @kernel_direct_lighting() {
+    unreachable
+  }
+
+  !llvm.dbg.cu = !{!0}
+  !llvm.module.flags = !{!3, !4}
+
+  !0 = distinct !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang version 12.0.0 (/src/external/llvm-project/clang ff12a3768d17240b01cbe9957e3c9e8f8ffe9854)", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, retainedTypes: !2, globals: !2, splitDebugInlining: false, nameTableKind: None)
+  !1 = !DIFile(filename: "/tmp/comgr-054202/input/CompileSource", directory: "/home/matt/bugs/SWDEV-253507")
+  !2 = !{}
+  !3 = !{i32 2, !"Debug Info Version", i32 3}
+  !4 = !{i32 7, !"PIC Level", i32 1}
+  !5 = !{!6, !9, i64 68}
+  !6 = !{!"ShaderData", !7, i64 0, !7, i64 16, !7, i64 32, !7, i64 48, !9, i64 64, !9, i64 68, !9, i64 72, !9, i64 76, !9, i64 80, !10, i64 84, !10, i64 88, !9, i64 92, !9, i64 96, !10, i64 100, !10, i64 104, !11, i64 112, !11, i64 144, !12, i64 176, !12, i64 184, !7, i64 192, !7, i64 208, !7, i64 224, !11, i64 240, !9, i64 272, !9, i64 276, !9, i64 280, !10, i64 284, !7, i64 288, !7, i64 304, !7, i64 320, !7, i64 336}
+  !7 = !{!"omnipotent char", !8, i64 0}
+  !8 = !{!"Simple C/C++ TBAA"}
+  !9 = !{!"int", !7, i64 0}
+  !10 = !{!"float", !7, i64 0}
+  !11 = !{!"differential3", !7, i64 0, !7, i64 16}
+  !12 = !{!"differential", !10, i64 0, !10, i64 4}
+  !13 = !{!6, !9, i64 280}
+  !16 = !DILocalVariable(name: "data", scope: !17, file: !18, line: 24)
+  !17 = distinct !DISubprogram(name: "svm_node_geometry", scope: !18, file: !18, line: 21, type: !19, scopeLine: 23, flags: DIFlagPrototyped | DIFlagAllCallsDescribed, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
+  !18 = !DIFile(filename: "source/kernel/svm/svm_geometry.h", directory: "/home/matt/bugs/SWDEV-253507")
+  !19 = !DISubroutineType(types: !20)
+  !20 = !{null}
+  !21 = !DILocation(line: 0, scope: !17)
+
+...
+---
+name:            kernel_direct_lighting
+tracksRegLiveness: true
+machineFunctionInfo:
+  scratchRSrcReg:  '$sgpr0_sgpr1_sgpr2_sgpr3'
+  frameOffsetReg:  '$sgpr33'
+  stackPtrOffsetReg: '$sgpr32'
+body:             |
+  ; CHECK-LABEL: name: kernel_direct_lighting
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.1(0x80000000)
+  ; CHECK:   liveins: $vgpr0_vgpr1, $vgpr2_vgpr3
+  ; CHECK:   [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+  ; CHECK:   [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr2_vgpr3
+  ; CHECK:   [[DEF:%[0-9]+]]:vreg_64 = IMPLICIT_DEF
+  ; CHECK:   [[DEF1:%[0-9]+]]:vreg_64 = IMPLICIT_DEF
+  ; CHECK:   [[DEF2:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+  ; CHECK:   [[DEF3:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+  ; CHECK:   [[DEF4:%[0-9]+]]:vreg_64 = IMPLICIT_DEF
+  ; CHECK:   [[DEF5:%[0-9]+]]:vreg_64 = IMPLICIT_DEF
+  ; CHECK:   [[DEF6:%[0-9]+]]:vreg_64 = IMPLICIT_DEF
+  ; CHECK:   [[DEF7:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+  ; CHECK:   [[DEF8:%[0-9]+]]:vreg_64 = IMPLICIT_DEF
+  ; CHECK:   undef %11.sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec
+  ; CHECK: bb.1:
+  ; CHECK:   successors: %bb.2(0x80000000)
+  ; CHECK:   dead %12:vreg_64 = COPY [[DEF]]
+  ; CHECK:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[DEF3]]
+  ; CHECK:   dead %14:vgpr_32 = COPY [[DEF2]]
+  ; CHECK:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[DEF5]].sub1
+  ; CHECK:   dead %16:vgpr_32 = COPY %11.sub0
+  ; CHECK:   undef %17.sub0:vreg_64, %18:sreg_64_xexec = V_ADD_CO_U32_e64 [[DEF4]].sub0, [[DEF6]].sub0, 0, implicit $exec
+  ; CHECK:   dead undef %17.sub1:vreg_64, dead %19:sreg_64_xexec = V_ADDC_U32_e64 [[DEF4]].sub1, [[DEF6]].sub1, %18, 0, implicit $exec
+  ; CHECK:   [[GLOBAL_LOAD_DWORDX2_:%[0-9]+]]:vreg_64 = GLOBAL_LOAD_DWORDX2 [[DEF1]], 0, 0, 0, 0, implicit $exec :: (load 8, addrspace 1)
+  ; CHECK:   dead %20:sreg_64 = V_CMP_GT_I32_e64 4, [[DEF7]], implicit $exec
+  ; CHECK:   GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF8]], 288, 0, 0, 0, implicit $exec :: (store 8, addrspace 1)
+  ; CHECK: bb.2:
+  ; CHECK:   successors: %bb.3(0x80000000)
+  ; CHECK:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY [[COPY2]]
+  ; CHECK: bb.3:
+  ; CHECK:   successors: %bb.2(0x40000000), %bb.4(0x40000000)
+  ; CHECK:   undef [[DEF5]].sub1:vreg_64 = COPY [[COPY3]]
+  ; CHECK:   S_CBRANCH_EXECZ %bb.2, implicit $exec
+  ; CHECK: bb.4:
+  ; CHECK:   successors: %bb.5(0x80000000)
+  ; CHECK:   dead %21:sreg_64 = COPY $exec
+  ; CHECK:   dead %22:vreg_128 = GLOBAL_LOAD_DWORDX4 [[COPY1]], 0, 0, 0, 0, implicit $exec :: (load 16, addrspace 1)
+  ; CHECK:   DBG_VALUE %22, $noreg, <0x8c16360>, !DIExpression(DW_OP_constu, 1, DW_OP_swap, DW_OP_xderef), debug-location !DILocation(line: 0, scope: <0x8c166d0>)
+  ; CHECK: bb.5:
+  ; CHECK:   successors: %bb.3(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   S_CBRANCH_EXECZ %bb.3, implicit $exec
+  ; CHECK:   S_BRANCH %bb.1
+  bb.0:
+    liveins: $vgpr0_vgpr1, $vgpr2_vgpr3
+
+    %0:vreg_64 = COPY $vgpr0_vgpr1
+    %1:vreg_64 = COPY $vgpr2_vgpr3
+    %2:vreg_64 = IMPLICIT_DEF
+    %3:vreg_64 = IMPLICIT_DEF
+    %4:vgpr_32 = IMPLICIT_DEF
+    %5:vgpr_32 = IMPLICIT_DEF
+    %6:vreg_64 = IMPLICIT_DEF
+    %7:vreg_64 = IMPLICIT_DEF
+    %8:vreg_64 = IMPLICIT_DEF
+    %9:vgpr_32 = IMPLICIT_DEF
+    %10:vreg_64 = IMPLICIT_DEF
+    undef %11.sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec
+
+  bb.1:
+    %12:vreg_64 = COPY %2
+    %13:vgpr_32 = COPY %5
+    dead %14:vgpr_32 = COPY %4
+    %15:vgpr_32 = COPY %7.sub1
+    dead %16:vgpr_32 = COPY %11.sub0
+    undef %17.sub0:vreg_64, %18:sreg_64_xexec = V_ADD_CO_U32_e64 %6.sub0, %8.sub0, 0, implicit $exec
+    dead %17.sub1:vreg_64, dead %19:sreg_64_xexec = V_ADDC_U32_e64 %6.sub1, %8.sub1, %18, 0, implicit $exec
+    %6:vreg_64 = GLOBAL_LOAD_DWORDX2 %3, 0, 0, 0, 0, implicit $exec :: (load 8, addrspace 1)
+    dead %20:sreg_64 = V_CMP_GT_I32_e64 4, %9, implicit $exec
+    GLOBAL_STORE_DWORDX2 %0, %10, 288, 0, 0, 0, implicit $exec :: (store 8, addrspace 1)
+
+  bb.2:
+    %5:vgpr_32 = COPY %13
+
+  bb.3:
+    undef %7.sub1:vreg_64 = COPY %15
+    S_CBRANCH_EXECZ %bb.2, implicit $exec
+
+  bb.4:
+    dead %21:sreg_64 = COPY $exec
+    %22:vreg_128 = GLOBAL_LOAD_DWORDX4 %1, 0, 0, 0, 0, implicit $exec :: (load 16, addrspace 1)
+    DBG_VALUE %22, $noreg, !16, !DIExpression(DW_OP_constu, 1, DW_OP_swap, DW_OP_xderef), debug-location !21
+
+  bb.5:
+    S_CBRANCH_EXECZ %bb.3, implicit $exec
+    S_BRANCH %bb.1
+
+...