[llvm] [PreISelIntrinsicLowering] Look through ptrtoint for a constant (e.g. global ptr) when trying to produce memset_pattern16 (PR #129220)

[Matt Arsenault via llvm-commits]

================
@@ -144,6 +147,56 @@ define void @memset_pattern_i64_128_tbaa(ptr %a) nounwind {
 !7 = !{!"omnipotent char", !8, i64 0}
 !8 = !{!"Simple C++ TBAA"}
 
+ at G = global i32 5
+
+define void @memset_pattern_i64_16_fromptr(ptr %a) nounwind {
+; CHECK-LABEL: define void @memset_pattern_i64_16_fromptr(
+; CHECK-SAME: ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint ptr @G to i64
+; CHECK-NEXT:    call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.2, i64 128)
+; CHECK-NEXT:    ret void
+;
+  %1 = ptrtoint ptr @G to i64
+  tail call void @llvm.experimental.memset.pattern(ptr %a, i64 %1, i64 16, i1 false)
+  ret void
+}
+
+define void @memset_pattern_i64_x_fromptr(ptr %a, i64 %x) nounwind {
+; CHECK-LABEL: define void @memset_pattern_i64_x_fromptr(
+; CHECK-SAME: ptr [[A:%.*]], i64 [[X:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint ptr @G to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = mul i64 8, [[X]]
+; CHECK-NEXT:    call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.3, i64 [[TMP2]])
+; CHECK-NEXT:    ret void
+;
+  %1 = ptrtoint ptr @G to i64
+  tail call void @llvm.experimental.memset.pattern(ptr %a, i64 %1, i64 %x, i1 false)
+  ret void
+}
+
+; memset_pattern16 shouldn't be used for this example (at least not by just
+; creating a constantarray global at compile time), as the ptr isn't constant.
+define void @memset_pattern_i64_x_fromnonconstptr(ptr %a, i64 %x, ptr %p) nounwind {
+; CHECK-LABEL: define void @memset_pattern_i64_x_fromnonconstptr(
+; CHECK-SAME: ptr [[A:%.*]], i64 [[X:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint ptr [[P]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = icmp eq i64 0, [[X]]
+; CHECK-NEXT:    br i1 [[TMP2]], label %[[SPLIT:.*]], label %[[LOADSTORELOOP:.*]]
+; CHECK:       [[LOADSTORELOOP]]:
+; CHECK-NEXT:    [[TMP3:%.*]] = phi i64 [ 0, [[TMP0:%.*]] ], [ [[TMP5:%.*]], %[[LOADSTORELOOP]] ]
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i64, ptr [[A]], i64 [[TMP3]]
+; CHECK-NEXT:    store i64 [[TMP1]], ptr [[TMP4]], align 1
+; CHECK-NEXT:    [[TMP5]] = add i64 [[TMP3]], 1
+; CHECK-NEXT:    [[TMP6:%.*]] = icmp ult i64 [[TMP5]], [[X]]
+; CHECK-NEXT:    br i1 [[TMP6]], label %[[LOADSTORELOOP]], label %[[SPLIT]]
+; CHECK:       [[SPLIT]]:
+; CHECK-NEXT:    ret void
+;
+  %1 = ptrtoint ptr %p to i64
+  tail call void @llvm.experimental.memset.pattern(ptr %a, i64 %1, i64 %x, i1 false)
+  ret void
----------------
arsenm wrote:

(finding the new no-required mangling suffix behavior confusing)

Could instcombine fold this to pull the cast in, and adjust the call signature? 

https://github.com/llvm/llvm-project/pull/129220