[llvm] [LowerMemIntrinsics] Lower llvm.memmove to wide memory accesses (PR #100122)

[Matt Arsenault via llvm-commits]

================
@@ -121,30 +142,51 @@ define amdgpu_kernel void @max_size_small_static_memmove_caller0(ptr addrspace(1
 define amdgpu_kernel void @min_size_large_static_memmove_caller0(ptr addrspace(1) %dst, ptr addrspace(1) %src) #0 {
 ; OPT-LABEL: @min_size_large_static_memmove_caller0(
 ; OPT-NEXT:    [[COMPARE_SRC_DST:%.*]] = icmp ult ptr addrspace(1) [[SRC:%.*]], [[DST:%.*]]
-; OPT-NEXT:    [[COMPARE_N_TO_0:%.*]] = icmp eq i64 1025, 0
-; OPT-NEXT:    br i1 [[COMPARE_SRC_DST]], label [[COPY_BACKWARDS:%.*]], label [[COPY_FORWARD:%.*]]
-; OPT:       copy_backwards:
-; OPT-NEXT:    br i1 [[COMPARE_N_TO_0]], label [[MEMMOVE_DONE:%.*]], label [[COPY_BACKWARDS_LOOP:%.*]]
-; OPT:       copy_backwards_loop:
-; OPT-NEXT:    [[TMP1:%.*]] = phi i64 [ [[INDEX_PTR:%.*]], [[COPY_BACKWARDS_LOOP]] ], [ 1025, [[COPY_BACKWARDS]] ]
-; OPT-NEXT:    [[INDEX_PTR]] = sub i64 [[TMP1]], 1
-; OPT-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i8, ptr addrspace(1) [[SRC]], i64 [[INDEX_PTR]]
+; OPT-NEXT:    br i1 [[COMPARE_SRC_DST]], label [[MEMMOVE_COPY_BACKWARDS:%.*]], label [[MEMMOVE_COPY_FORWARD:%.*]]
+; OPT:       memmove_copy_backwards:
+; OPT-NEXT:    br i1 false, label [[MEMMOVE_BWD_MIDDLE:%.*]], label [[MEMMOVE_BWD_RESIDUAL_LOOP:%.*]]
----------------
arsenm wrote:

Depends what's easiest. This is running late so we're not getting the usual control flow cleanups we would get in the middle end 

https://github.com/llvm/llvm-project/pull/100122