Register self loops in the loop forest

ir/function.cpp

@@ -1201,9 +1201,16 @@ void LoopAnalysis::run() {
   // Construct the loop forest (0 or more loop trees)
   for (unsigned i = 0; i < bb_count; ++i) {
     auto h = header[i];
+    // Give every loop header a key of its own. A header only ever acquires one
+    // as a side effect of a body block naming it as parent, so a loop whose
+    // body is just the header itself (a self loop) would otherwise be absent
+    // from the forest and consumers walking it top-down would not see a loop
+    // there at all.
+    if (type[i] != nonheader)
+      (void)forest[node[i]];
+
     if (h == 0 && type[i] != nonheader) {
       roots.emplace_back(node[i]);
-      (void)forest[node[i]];
     }
     else if (h != 0 || type[i] != nonheader) {
       parent.emplace(node[i], node[h]);

tests/alive-tv/loops/nested-self-loop-fail.srctgt.ll

@@ -0,0 +1,50 @@
+; TEST-ARGS: -src-unroll=8 -tgt-unroll=8
+; ERROR: Value mismatch
+
+; A loop whose body is a single basic block used not to be registered as a loop
+; header in the loop forest, so the unroller skipped it when it was nested in
+; another loop. Its backedge then survived into the unrolled function, where it
+; is redirected to #sink, so the executions that run the inner loop more than
+; once dropped out of the refinement query and this pair verified.
+; tgt returns 3 * %n, not %n.
+
+define i32 @src(i32 %n) {
+entry:
+  %c = icmp slt i32 %n, 1
+  br i1 %c, label %zero, label %pos
+
+zero:
+  ret i32 0
+
+pos:
+  ret i32 %n
+}
+
+define i32 @tgt(i32 %n) {
+entry:
+  %c = icmp slt i32 %n, 1
+  br i1 %c, label %done, label %outer
+
+outer:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %latch ]
+  %acc = phi i32 [ 0, %entry ], [ %acc.next, %latch ]
+  br label %inner
+
+inner:
+  %j = phi i32 [ 0, %outer ], [ %j.next, %inner ]
+  %a = phi i32 [ %acc, %outer ], [ %a.next, %inner ]
+  %a.next = add i32 %a, 1
+  %j.next = add i32 %j, 1
+  %jc = icmp slt i32 %j.next, 3
+  br i1 %jc, label %inner, label %latch
+
+latch:
+  %acc.next = phi i32 [ %a.next, %inner ]
+  %i.next = add i32 %i, 1
+  %ic = icmp slt i32 %i.next, %n
+  br i1 %ic, label %outer, label %done
+
+done:
+  %r = phi i32 [ 0, %entry ], [ %acc.next, %latch ]
+  ret i32 %r
+}

tests/alive-tv/loops/nested-self-loop.srctgt.ll

@@ -0,0 +1,62 @@
+; TEST-ARGS: -src-unroll=8 -tgt-unroll=8
+
+; The counterpart of nested-self-loop-fail.srctgt.ll: unrolling a nested self
+; loop must not make an equivalent pair unverifiable.
+
+define i32 @src(i32 %n) {
+entry:
+  %c = icmp slt i32 %n, 1
+  br i1 %c, label %done, label %outer
+
+outer:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %latch ]
+  %acc = phi i32 [ 0, %entry ], [ %acc.next, %latch ]
+  br label %inner
+
+inner:
+  %j = phi i32 [ 0, %outer ], [ %j.next, %inner ]
+  %a = phi i32 [ %acc, %outer ], [ %a.next, %inner ]
+  %a.next = add i32 %a, 1
+  %j.next = add i32 %j, 1
+  %jc = icmp slt i32 %j.next, 3
+  br i1 %jc, label %inner, label %latch
+
+latch:
+  %acc.next = phi i32 [ %a.next, %inner ]
+  %i.next = add i32 %i, 1
+  %ic = icmp slt i32 %i.next, %n
+  br i1 %ic, label %outer, label %done
+
+done:
+  %r = phi i32 [ 0, %entry ], [ %acc.next, %latch ]
+  ret i32 %r
+}
+
+define i32 @tgt(i32 %n) {
+entry:
+  %c = icmp slt i32 %n, 1
+  br i1 %c, label %done, label %outer
+
+outer:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %latch ]
+  %acc = phi i32 [ 0, %entry ], [ %acc.next, %latch ]
+  br label %inner
+
+inner:
+  %j = phi i32 [ 0, %outer ], [ %j.next, %inner ]
+  %a = phi i32 [ %acc, %outer ], [ %a.next, %inner ]
+  %a.next = sub i32 %a, -1
+  %j.next = add i32 %j, 1
+  %jc = icmp slt i32 %j.next, 3
+  br i1 %jc, label %inner, label %latch
+
+latch:
+  %acc.next = phi i32 [ %a.next, %inner ]
+  %i.next = add i32 %i, 1
+  %ic = icmp slt i32 %i.next, %n
+  br i1 %ic, label %outer, label %done
+
+done:
+  %r = phi i32 [ 0, %entry ], [ %acc.next, %latch ]
+  ret i32 %r
+}