[llvm] [AMDGPU] Fix to prevent sinking of PERMLANE_SWAP instruction (PR #144423)

[Matt Arsenault via llvm-commits]

================
@@ -733,3 +733,64 @@ body:             |
     liveins: $vgpr0, $vgpr1, $vgpr2_vgpr3, $vcc
     S_ENDPGM 0
 ...
+---
+name: test_no_sink_permlane_swap
+tracksRegLiveness: true
+machineFunctionInfo:
+  isEntryFunction: true
+body: |
+  ; GFX9-LABEL: name: test_no_sink_permlane_swap
+  ; GFX9: bb.0:
+  ; GFX9-NEXT:   successors: %bb.2(0x40000000), %bb.1(0x40000000)
+  ; GFX9-NEXT:   liveins: $vgpr0
+  ; GFX9-NEXT: {{  $}}
+  ; GFX9-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX9-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX9-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 0
+  ; GFX9-NEXT:   [[COPY1:%[0-9]+]]:vreg_64 = COPY [[S_MOV_B64_]]
+  ; GFX9-NEXT:   [[GLOBAL_LOAD_DWORD:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD killed [[COPY1]], 0, 0, implicit $exec :: (load (s32), addrspace 1)
+  ; GFX9-NEXT:   [[V_PERMLANE32_SWAP_B32_e64_:%[0-9]+]]:vgpr_32, [[V_PERMLANE32_SWAP_B32_e64_1:%[0-9]+]]:vgpr_32 = V_PERMLANE32_SWAP_B32_e64 [[GLOBAL_LOAD_DWORD]], [[GLOBAL_LOAD_DWORD]], 0, 0, implicit $exec
+  ; GFX9-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr0
+  ; GFX9-NEXT:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 1
+  ; GFX9-NEXT:   [[V_CMP_LT_I32_e64_:%[0-9]+]]:sreg_64 = V_CMP_LT_I32_e64 [[COPY2]](s32), [[S_MOV_B32_]], implicit $exec
+  ; GFX9-NEXT:   [[SI_IF:%[0-9]+]]:sreg_64 = SI_IF [[V_CMP_LT_I32_e64_]], %bb.2, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+  ; GFX9-NEXT:   S_BRANCH %bb.1
+  ; GFX9-NEXT: {{  $}}
+  ; GFX9-NEXT: bb.1:
+  ; GFX9-NEXT:   successors: %bb.2(0x80000000)
+  ; GFX9-NEXT: {{  $}}
+  ; GFX9-NEXT:   [[V_MAX_I32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_I32_e64 [[V_PERMLANE32_SWAP_B32_e64_]], [[V_PERMLANE32_SWAP_B32_e64_1]], implicit $exec
+  ; GFX9-NEXT: {{  $}}
+  ; GFX9-NEXT: bb.2:
+  ; GFX9-NEXT:   successors: %bb.3(0x80000000)
+  ; GFX9-NEXT: {{  $}}
+  ; GFX9-NEXT:   [[PHI:%[0-9]+]]:vgpr_32 = PHI [[V_MOV_B32_e32_]], %bb.0, [[V_MAX_I32_e64_]], %bb.1
+  ; GFX9-NEXT:   SI_END_CF [[SI_IF]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+  ; GFX9-NEXT: {{  $}}
+  ; GFX9-NEXT: bb.3:
+  ; GFX9-NEXT:   S_ENDPGM 0, implicit [[PHI]]
+  bb.0:
+    liveins: $vgpr0
+    %1:vgpr_32 = COPY $vgpr0
----------------
arsenm wrote:

Should run -run-pass=none to compact the register numbers 

https://github.com/llvm/llvm-project/pull/144423