CodeGen: Don't lazily construct MachineFunctionInfo

This fixes what I consider to be an API flaw I've tripped over
multiple times. The point this is constructed isn't well defined, so
depending on where this is first called, you can conclude different
information based on the MachineFunction. For example, the AMDGPU
implementation inspected the MachineFrameInfo on construction for the
stack objects and if the frame has calls. This kind of worked in
SelectionDAG which visited all allocas up front, but broke in
GlobalISel which hasn't visited any of the IR when arguments are
lowered.

I've run into similar problems before with the MIR parser and trying
to make use of other MachineFunction fields, so I think it's best to
just categorically disallow dependency on the MachineFunction state in
the constructor and to always construct this at the same time as the
MachineFunction itself.

A missing feature I still could use is a way to access an custom
analysis pass on the IR here.

1 files changed, 7 insertions, 0 deletions

diff --git a/llvm/lib/CodeGen/MachineFunction.cpp b/llvm/lib/CodeGen/MachineFunction.cpp
index 78b49c9..c6653ae 100644
--- a/llvm/lib/CodeGen/MachineFunction.cpp
+++ b/llvm/lib/CodeGen/MachineFunction.cpp
@@ -187,6 +187,7 @@ void MachineFunction::init() {
     RegInfo = nullptr;
 
   MFInfo = nullptr;
+
   // We can realign the stack if the target supports it and the user hasn't
   // explicitly asked us not to.
   bool CanRealignSP = STI->getFrameLowering()->isStackRealignable() &&
@@ -232,6 +233,12 @@ void MachineFunction::init() {
   PSVManager = std::make_unique<PseudoSourceValueManager>(getTarget());
 }
 
+void MachineFunction::initTargetMachineFunctionInfo(
+    const TargetSubtargetInfo &STI) {
+  assert(!MFInfo && "MachineFunctionInfo already set");
+  MFInfo = Target.createMachineFunctionInfo(Allocator, F, &STI);
+}
+
 MachineFunction::~MachineFunction() {
   clear();
 }