[llvm] [AMDGPU] Eliminate redundant SGPR restores (PR #222680)

[Janek van Oirschot via llvm-commits]

================
@@ -40,13 +39,6 @@ body:             |
     ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: ("amdgpu-thread-private" store (s32) into %stack.0, addrspace 5)
     ; CHECK-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: ("amdgpu-thread-private" load (s32) from %stack.2, addrspace 5)
     ; CHECK-NEXT: $exec = S_MOV_B64 killed $sgpr4_sgpr5, implicit killed $vgpr1
-    ; CHECK-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
-    ; CHECK-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr1
-    ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: ("amdgpu-thread-private" store (s32) into %stack.2, addrspace 5)
-    ; CHECK-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: ("amdgpu-thread-private" load (s32) from %stack.0, addrspace 5)
-    ; CHECK-NEXT: $sgpr10 = SI_RESTORE_S32_FROM_VGPR killed $vgpr1, 0
-    ; CHECK-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: ("amdgpu-thread-private" load (s32) from %stack.2, addrspace 5)
-    ; CHECK-NEXT: $exec = S_MOV_B64 killed $sgpr4_sgpr5, implicit killed $vgpr1
----------------
JanekvO wrote:

Does this optimize away what's being tested? I.e., does the change make this test not fit for purpose anymore?

https://github.com/llvm/llvm-project/pull/222680