[llvm] [LiveDebugVariables] Fix a DBG_VALUE reordering issue (PR #111124)

[Jeremy Morse via llvm-commits]

================
@@ -12,11 +12,11 @@ define <4 x float> @Scene_transformT(i32 %subshapeIdx, <4 x float> %v, float %ti
 entry:
   ; CHECK: v_mov_b32_e32 v[[COPIED_ARG_PIECE:[0-9]+]], v9
 
-  ; CHECK: ;DEBUG_VALUE: Scene_transformT:gScene <- [DW_OP_constu 1, DW_OP_swap, DW_OP_xderef, DW_OP_LLVM_fragment 0 32] $vgpr6
+  ; CHECK: ;DEBUG_VALUE: Scene_transformT:gSceneOffsets <- [DW_OP_constu 1, DW_OP_swap, DW_OP_xderef, DW_OP_LLVM_fragment 32 32] $vgpr[[COPIED_ARG_PIECE]]
+  ; CHECK: ;DEBUG_VALUE: Scene_transformT:gSceneOffsets <- [DW_OP_constu 1, DW_OP_swap, DW_OP_xderef, DW_OP_LLVM_fragment 0 32] $vgpr8
   ; CHECK: ;DEBUG_VALUE: Scene_transformT:gScene <- [DW_OP_constu 1, DW_OP_swap, DW_OP_xderef, DW_OP_LLVM_fragment 32 32] $vgpr7
+  ; CHECK: ;DEBUG_VALUE: Scene_transformT:gScene <- [DW_OP_constu 1, DW_OP_swap, DW_OP_xderef, DW_OP_LLVM_fragment 0 32] $vgpr6
   call void @llvm.dbg.value(metadata ptr addrspace(1) %gScene, metadata !120, metadata !DIExpression(DW_OP_constu, 1, DW_OP_swap, DW_OP_xderef)), !dbg !154
-  ; CHECK: ;DEBUG_VALUE: Scene_transformT:gSceneOffsets <- [DW_OP_constu 1, DW_OP_swap, DW_OP_xderef, DW_OP_LLVM_fragment 0 32] $vgpr8
-  ; CHECK: ;DEBUG_VALUE: Scene_transformT:gSceneOffsets <- [DW_OP_constu 1, DW_OP_swap, DW_OP_xderef, DW_OP_LLVM_fragment 32 32] $vgpr[[COPIED_ARG_PIECE]]
   call void @llvm.dbg.value(metadata ptr addrspace(1) %gSceneOffsets, metadata !121, metadata !DIExpression(DW_OP_constu, 1, DW_OP_swap, DW_OP_xderef)), !dbg !155
   %call = tail call ptr addrspace(1) @Scene_getSubShapeData(i32 %subshapeIdx, ptr addrspace(1) %gScene, ptr addrspace(1) %gSceneOffsets)
----------------
jmorse wrote:

Hard for me to analyse just from the github UI as this isn't a MIR test, but !120 = gScene, !121 = gSceneOffsets, which are ordered !120, !121 in the two pre-isel dbg.values. In the old check lines the DEBUG_VALUE lines come out in the right order, but now they're flipped?

Chances are I'm reading it wrong, or there's some other re-ordering happening somewhere in LLVM that LDVars was just accidentally undoing, seeing how the order in other test files changes is fixed.



https://github.com/llvm/llvm-project/pull/111124