[llvm] [InstCombine] Canonicalize constant GEPs to i8 source element type (PR #68882)

[Nikita Popov via llvm-commits]

================
@@ -382,15 +382,16 @@ define <4 x i1> @compare_geps_same_indices_scalar_vector_base_mismatch(ptr %ptr,
 define ptr @indexed_compare(ptr %A, i64 %offset) {
 ; CHECK-LABEL: @indexed_compare(
 ; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[OFFSET:%.*]]
 ; CHECK-NEXT:    br label [[BB:%.*]]
 ; CHECK:       bb:
-; CHECK-NEXT:    [[RHS_IDX:%.*]] = phi i64 [ [[RHS_ADD:%.*]], [[BB]] ], [ [[OFFSET:%.*]], [[ENTRY:%.*]] ]
-; CHECK-NEXT:    [[RHS_ADD]] = add nsw i64 [[RHS_IDX]], 1
-; CHECK-NEXT:    [[COND:%.*]] = icmp sgt i64 [[RHS_IDX]], 100
+; CHECK-NEXT:    [[RHS:%.*]] = phi ptr [ [[RHS_NEXT:%.*]], [[BB]] ], [ [[TMP]], [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[LHS:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 400
+; CHECK-NEXT:    [[RHS_NEXT]] = getelementptr inbounds i8, ptr [[RHS]], i64 4
+; CHECK-NEXT:    [[COND:%.*]] = icmp ult ptr [[LHS]], [[RHS]]
----------------
nikic wrote:

Index compare fold regression.

https://github.com/llvm/llvm-project/pull/68882