[llvm] [AMDGPU] Fix SIFoldOperands miscompiling values that leave a divergent loop (PR #203256)

[Sameer Sahasrabuddhe via llvm-commits]

================
@@ -49,8 +49,137 @@ body:             |
 
   bb.2:
     SI_END_CF %4, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
-    %9:vgpr_32 = COPY %7
-    %10:sreg_64_xexec = IMPLICIT_DEF
-    %11:vgpr_32 = V_SET_INACTIVE_B32 0, %9, 0, 0, killed %10, implicit $exec
+    %8:vgpr_32 = COPY %7
+    %9:sreg_64_xexec = IMPLICIT_DEF
+    %10:vgpr_32 = V_SET_INACTIVE_B32 0, %8, 0, 0, killed %9, implicit $exec
     S_ENDPGM 0
 ...
+
----------------
ssahasra wrote:

Can you pre-submit this test, so that this PR will display the fix better?

https://github.com/llvm/llvm-project/pull/203256