[llvm] [AMDGPU] Improve StructurizeCFG pass performance by using SSAUpdaterBulk. (PR #130611)

[Valery Pykhtin via llvm-commits]

================
@@ -69,13 +69,13 @@ define amdgpu_kernel void @loop_subregion_misordered(ptr addrspace(1) %arg0) #0
 ; CHECK-NEXT:    [[TMP5:%.*]] = phi i1 [ [[CMP_END_ELSE_BLOCK:%.*]], [[END_ELSE_BLOCK]] ], [ true, [[FLOW2]] ]
 ; CHECK-NEXT:    br i1 [[TMP5]], label [[FLOW4:%.*]], label [[LOOP_HEADER]]
 ; CHECK:       Flow4:
-; CHECK-NEXT:    br i1 [[TMP7:%.*]], label [[BB64:%.*]], label [[RETURN:%.*]]
+; CHECK-NEXT:    br i1 false, label [[BB64:%.*]], label [[RETURN:%.*]]
----------------
vpykhtin wrote:

PSDB failed for original and modified patches. It looks like the problem is that the SSAUpdaterBulk doesn't have processing similar to SSAUpdater::GetValueInMiddleOfBlock: when we add i1 variable value for a block SSAUpdaterBulk  just uses it instead of making a phi in a special case like GetValueInMiddleOfBlock.

https://github.com/llvm/llvm-project/pull/130611