[llvm] [LoopInterchange] Add test for extra reduction use in inner loop (NFC) (PR #203240)

[Ryotaro Kasuga via llvm-commits]

================
@@ -0,0 +1,283 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt < %s -passes=loop-interchange -loop-interchange-profitabilities=ignore -S | FileCheck %s
+
+; sum = 0;
+; for (i = 0, sum_i = sum; i < 2; i++)
+;   for (j = 0; j < 2; j++) {
+;     sum += A[j][i];
+;     B[j][i] = sum_i;
+;   }
+;
+; Interchanging the loops will become as follows:
+;
+; sum = 0;
+; for (j = 0, sum_j = sum; j < 2; j++)
+;   for (i = 0; i < 2; i++) {
+;     sum += A[j][i];
+;     B[j][i] = sum_j;
+;   }
+;
+; This is invalid transformation. Consider the case when `A` is as follows:
+;
+;  A = {{ 0, 0 }, { 1, 1 }}
+;
+; In this case, `sum_i` evolves 0 -> 1, while `sum_j` evolves 0 -> 0.
+;
+; FIXME: The loops can be interchanged now.
+;
+define i8 @extra_reduction_use_in_inner0(ptr noalias %A, ptr noalias %B) {
+; CHECK-LABEL: define i8 @extra_reduction_use_in_inner0(
+; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]]) {
+; CHECK-NEXT:  [[INNER_PREHEADER:.*:]]
+; CHECK-NEXT:    br label %[[INNER:.*]]
+; CHECK:       [[OUTER_HEADER_PREHEADER1:.*]]:
+; CHECK-NEXT:    br label %[[OUTER_HEADER_PREHEADER:.*]]
+; CHECK:       [[OUTER_HEADER_PREHEADER]]:
+; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[I_NEXT:%.*]], %[[OUTER_LATCH:.*]] ], [ 0, %[[OUTER_HEADER_PREHEADER1]] ]
+; CHECK-NEXT:    [[SUM_INNER:%.*]] = phi i8 [ [[SUM_NEXT:%.*]], %[[OUTER_LATCH]] ], [ [[SUM:%.*]], %[[OUTER_HEADER_PREHEADER1]] ]
+; CHECK-NEXT:    br label %[[INNER_SPLIT1:.*]]
+; CHECK:       [[INNER]]:
+; CHECK-NEXT:    br label %[[INNER1:.*]]
+; CHECK:       [[INNER1]]:
+; CHECK-NEXT:    [[J:%.*]] = phi i64 [ [[TMP2:%.*]], %[[INNER_SPLIT:.*]] ], [ 0, %[[INNER]] ]
+; CHECK-NEXT:    [[SUM]] = phi i8 [ [[SUM_LCSSA:%.*]], %[[INNER_SPLIT]] ], [ 0, %[[INNER]] ]
+; CHECK-NEXT:    br label %[[OUTER_HEADER_PREHEADER1]]
+; CHECK:       [[INNER_SPLIT1]]:
+; CHECK-NEXT:    [[GEP_A:%.*]] = getelementptr [2 x i8], ptr [[A]], i64 [[J]], i64 [[I]]
+; CHECK-NEXT:    [[A:%.*]] = load i8, ptr [[GEP_A]], align 1
+; CHECK-NEXT:    [[SUM_NEXT]] = add i8 [[SUM_INNER]], [[A]]
+; CHECK-NEXT:    [[GEP_B:%.*]] = getelementptr [2 x i8], ptr [[B]], i64 [[J]], i64 [[I]]
+; CHECK-NEXT:    store i8 [[SUM]], ptr [[GEP_B]], align 1
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[J]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq i64 [[TMP0]], 2
+; CHECK-NEXT:    br label %[[OUTER_LATCH]]
+; CHECK:       [[INNER_SPLIT]]:
+; CHECK-NEXT:    [[SUM_LCSSA]] = phi i8 [ [[SUM_NEXT]], %[[OUTER_LATCH]] ]
+; CHECK-NEXT:    [[TMP2]] = add i64 [[J]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[TMP2]], 2
+; CHECK-NEXT:    br i1 [[TMP3]], label %[[EXIT:.*]], label %[[INNER1]]
+; CHECK:       [[OUTER_LATCH]]:
+; CHECK-NEXT:    [[I_NEXT]] = add i64 [[I]], 1
+; CHECK-NEXT:    [[EC_I:%.*]] = icmp eq i64 [[I_NEXT]], 2
+; CHECK-NEXT:    br i1 [[EC_I]], label %[[INNER_SPLIT]], label %[[OUTER_HEADER_PREHEADER]]
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    [[RES:%.*]] = phi i8 [ [[SUM_LCSSA]], %[[INNER_SPLIT]] ]
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+entry:
+  br label %outer.header
+
+outer.header:
+  %i = phi i64 [ 0, %entry ], [ %i.next, %outer.latch ]
+  %sum.outer = phi i8 [ 0, %entry ], [ %sum.lcssa, %outer.latch ]
+  br label %inner
+
+inner:
+  %j = phi i64 [ 0, %outer.header ], [ %j.next, %inner ]
+  %sum.inner = phi i8 [ %sum.outer, %outer.header ], [ %sum.inner.next, %inner ]
+  %gep.A = getelementptr [2 x i8], ptr %A, i64 %j, i64 %i
+  %a = load i8, ptr %gep.A
+  %sum.inner.next = add i8 %sum.inner, %a
+  %gep.B = getelementptr [2 x i8], ptr %B, i64 %j, i64 %i
+  store i8 %sum.outer, ptr %gep.B
+  %j.next = add i64 %j, 1
+  %ec.j = icmp eq i64 %j.next, 2
+  br i1 %ec.j, label %outer.latch, label %inner
+
+outer.latch:
+  %sum.lcssa = phi i8 [ %sum.inner.next, %inner ]
+  %i.next = add i64 %i, 1
+  %ec.i = icmp eq i64 %i.next, 2
+  br i1 %ec.i, label %exit, label %outer.header
+
+exit:
+  %res = phi i8 [ %sum.lcssa, %outer.latch ]
+  ret i8 %res
+}
+
+; sum = 0;
+; for (i = 0; i < 2; i++)
+;   for (j = 0; j < 2; j++) {
+;     sum += A[j][i];
+;     B[j][i] = sum;
+;   }
+;
+; Interchanging the loops will become as follows:
+;
+; sum = 0;
+; for (j = 0; j < 2; j++)
+;   for (i = 0; i < 2; i++) {
+;     sum += A[j][i];
+;     B[j][i] = sum;
+;   }
+;
+; This is invalid transformation. Consider the case when `A` is as follows:
+;
+;  A = {{ 0, 1 }, { 0, 0 }}
+;
+; In this case, in the original loops, `sum` evolves 0 -> 0 -> 1 -> 1, while in
+; the interchanged loops, `sum` evolves 0 -> 1 -> 1 -> 1
+;
+define i8 @extra_reduction_use_in_inner1(ptr noalias %A, ptr noalias %B) {
----------------
kasuga-fj wrote:

I didn't expect this, but apparently this case isn't recognized as a reduction by RecurrenceDescriptor.

https://github.com/llvm/llvm-project/pull/203240