[llvm] [LV] Extend FindLastIV to unsigned case (PR #141752)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -843,66 +843,184 @@ exit:                                             ; preds = %for.body, %entry
 ; This test cannot be vectorized, even with a runtime check.
 ; The construct that are introduced by IndVarSimplify is:
 ;   %1 = trunc i64 %iv to i32
-; However, the loop exit condition is a constant that overflows i32:
+; The loop exit condition is a constant that overflows signed i32,
+; but not unsigned i32:
 ;   %exitcond.not = icmp eq i64 %inc, 4294967294
-; Hence, the i32 will most certainly wrap and hit the sentinel value, and we
-; cannot vectorize this case.
-define i32 @not_vectorized_select_icmp_truncated_iv_out_of_bound(ptr %a) {
-; CHECK-VF4IC1-LABEL: define i32 @not_vectorized_select_icmp_truncated_iv_out_of_bound(
+; Hence, we can vectorize with the unsigned variant of FindLastIV.
+define i32 @select_icmp_truncated_unsigned_iv_range(ptr %a) {
+; CHECK-VF4IC1-LABEL: define i32 @select_icmp_truncated_unsigned_iv_range(
 ; CHECK-VF4IC1-SAME: ptr [[A:%.*]]) {
 ; CHECK-VF4IC1-NEXT:  [[ENTRY:.*]]:
+; CHECK-VF4IC1-NEXT:    br i1 false, label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK-VF4IC1:       [[VECTOR_PH]]:
+; CHECK-VF4IC1-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK-VF4IC1:       [[VECTOR_BODY]]:
+; CHECK-VF4IC1-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-VF4IC1-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP3:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-VF4IC1-NEXT:    [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 2147483646, i32 2147483647, i32 -2147483648, i32 -2147483647>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-VF4IC1-NEXT:    [[IV:%.*]] = add i64 2147483646, [[INDEX]]
+; CHECK-VF4IC1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV]]
+; CHECK-VF4IC1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[ARRAYIDX]], i32 0
+; CHECK-VF4IC1-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP1]], align 4
+; CHECK-VF4IC1-NEXT:    [[TMP2:%.*]] = icmp sgt <4 x i32> [[WIDE_LOAD]], splat (i32 3)
+; CHECK-VF4IC1-NEXT:    [[TMP3]] = select <4 x i1> [[TMP2]], <4 x i32> [[VEC_IND]], <4 x i32> [[VEC_PHI]]
+; CHECK-VF4IC1-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-VF4IC1-NEXT:    [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
+; CHECK-VF4IC1-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 2147483648
+; CHECK-VF4IC1-NEXT:    br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
+; CHECK-VF4IC1:       [[MIDDLE_BLOCK]]:
+; CHECK-VF4IC1-NEXT:    [[TMP5:%.*]] = call i32 @llvm.vector.reduce.smax.v4i32(<4 x i32> [[TMP3]])
----------------
artagnon wrote:

Oops, thanks for catching this serious bug! Fixed now.

https://github.com/llvm/llvm-project/pull/141752