[MemoryBuiltins] Cache the result of ObjectOffsetSizeVisitor::visit. #64796 (#65326)

visit will skip visiting instructions it already has visited
to avoid issues with cycles in the data graph. However,
the result of this skipping behavior is that if we
encounter the same instruction twice, and that instruction
has a well defined result and isn't part of a cycle, we
will introduce unknowns into the analysis even though we
knew the size and offset of the instruction's result.

Instead of skipping such instructions, keep a cache of
the result of visiting them. This result is initialized
to unknown() before visiting, so if we happen to visit
it again recursively (perhaps as the result of a cycle
or a phi), we will get unknown as the cached result and
exit out.

Author: bevin-hansson

llvm/include/llvm/Analysis/MemoryBuiltins.h

@@ -198,7 +198,7 @@ class ObjectSizeOffsetVisitor
   ObjectSizeOpts Options;
   unsigned IntTyBits;
   APInt Zero;
-  SmallPtrSet<Instruction *, 8> SeenInsts;
+  DenseMap<Instruction *, SizeOffsetType> SeenInsts;
 
   APInt align(APInt Size, MaybeAlign Align);
 

llvm/lib/Analysis/MemoryBuiltins.cpp

@@ -733,10 +733,14 @@ SizeOffsetType ObjectSizeOffsetVisitor::computeImpl(Value *V) {
   if (Instruction *I = dyn_cast<Instruction>(V)) {
     // If we have already seen this instruction, bail out. Cycles can happen in
     // unreachable code after constant propagation.
-    if (!SeenInsts.insert(I).second)
-      return unknown();
-
-    return visit(*I);
+    auto P = SeenInsts.try_emplace(I, unknown());
+    if (!P.second)
+      return P.first->second;
+    SizeOffsetType Res = visit(*I);
+    // Cache the result for later visits. If we happened to visit this during
+    // the above recursion, we would consider it unknown until now.
+    SeenInsts[I] = Res;
+    return Res;
   }
   if (Argument *A = dyn_cast<Argument>(V))
     return visitArgument(*A);

llvm/test/Transforms/LowerConstantIntrinsics/builtin-object-size-phi.ll

@@ -61,3 +61,59 @@ if.end:
   %size = call i64 @llvm.objectsize.i64.p0(ptr %p, i1 true, i1 true, i1 false)
   ret i64 %size
 }
+
+define i64 @pick_max_same(i32 %n) {
+; CHECK-LABEL: @pick_max_same(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[BUFFER:%.*]] = alloca i8, i64 20, align 1
+; CHECK-NEXT:    [[COND:%.*]] = icmp eq i32 [[N:%.*]], 0
+; CHECK-NEXT:    br i1 [[COND]], label [[IF_ELSE:%.*]], label [[IF_END:%.*]]
+; CHECK:       if.else:
+; CHECK-NEXT:    [[OFFSETED:%.*]] = getelementptr i8, ptr [[BUFFER]], i64 10
+; CHECK-NEXT:    br label [[IF_END]]
+; CHECK:       if.end:
+; CHECK-NEXT:    [[P:%.*]] = phi ptr [ [[OFFSETED]], [[IF_ELSE]] ], [ [[BUFFER]], [[ENTRY:%.*]] ]
+; CHECK-NEXT:    ret i64 20
+;
+entry:
+  %buffer = alloca i8, i64 20
+  %cond = icmp eq i32 %n, 0
+  br i1 %cond, label %if.else, label %if.end
+
+if.else:
+  %offseted = getelementptr i8, ptr %buffer, i64 10
+  br label %if.end
+
+if.end:
+  %p = phi ptr [ %offseted, %if.else ], [ %buffer, %entry ]
+  %size = call i64 @llvm.objectsize.i64.p0(ptr %p, i1 false, i1 true, i1 false)
+  ret i64 %size
+}
+
+define i64 @pick_min_same(i32 %n) {
+; CHECK-LABEL: @pick_min_same(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[BUFFER:%.*]] = alloca i8, i64 20, align 1
+; CHECK-NEXT:    [[COND:%.*]] = icmp eq i32 [[N:%.*]], 0
+; CHECK-NEXT:    br i1 [[COND]], label [[IF_ELSE:%.*]], label [[IF_END:%.*]]
+; CHECK:       if.else:
+; CHECK-NEXT:    [[OFFSETED:%.*]] = getelementptr i8, ptr [[BUFFER]], i64 10
+; CHECK-NEXT:    br label [[IF_END]]
+; CHECK:       if.end:
+; CHECK-NEXT:    [[P:%.*]] = phi ptr [ [[OFFSETED]], [[IF_ELSE]] ], [ [[BUFFER]], [[ENTRY:%.*]] ]
+; CHECK-NEXT:    ret i64 10
+;
+entry:
+  %buffer = alloca i8, i64 20
+  %cond = icmp eq i32 %n, 0
+  br i1 %cond, label %if.else, label %if.end
+
+if.else:
+  %offseted = getelementptr i8, ptr %buffer, i64 10
+  br label %if.end
+
+if.end:
+  %p = phi ptr [ %offseted, %if.else ], [ %buffer, %entry ]
+  %size = call i64 @llvm.objectsize.i64.p0(ptr %p, i1 true, i1 true, i1 false)
+  ret i64 %size
+}