[CIR][CodeGen] Store the old CIRGenFunction when popping to a new one

[lanza]

We diverge from CodeGen here by delaying the function emission that
happens for a global variable. However, due to situations where a global
can be emitted while building out a function the old CGF might not be
invalid. So we need to store it here just in case.

[smeenai]

Is there any way to write a test case for this, and should it use llvm::SaveAndRestore instead to work for all exit paths?

[lanza]

Is there any way to write a test case for this

Not yet, I'm trying to hoist out a bunch of code from a 1000 line patch for static local variables that needs heavily refactored... hence all this string of PRs. But this occurs during non-trivial static locals here for something like this:

int foo();
int bar() {
  static int a = foo();
  return a;
}

which we'll emit (before LoweringPrepare) as something like

cir.global "private"  internal dsolocal @_ZZ3barvE1a !s32i {alignment = 4 : i64, static_local = true }
cir.func  @_Z3barv() -> !s32i {
    %1 = cir.static_local "_ZZ3barvE1a" !cir.ptr<1s32i>
    %2 = cir.load %1 : !cir.ptr<!s32i>, !s32i
    cir.return %3 : !s32i
} 

should it use llvm::SaveAndRestore instead to work for all exit paths?

Yes, couldn't remember the name of this mechanism and just did it manually. Thanks!

[smeenai]

Is there any way to write a test case for this

Not yet, I'm trying to hoist out a bunch of code from a 1000 line patch for static local variables that needs heavily refactored... hence all this string of PRs. But this occurs during non-trivial static locals here for something like this:

int foo();
int bar() {
  static int a = foo();
  return a;
}

which we'll emit (before LoweringPrepare) as something like

cir.global "private"  internal dsolocal @_ZZ3barvE1a !s32i {alignment = 4 : i64, static_local = true }
cir.func  @_Z3barv() -> !s32i {
    %1 = cir.static_local "_ZZ3barvE1a" !cir.ptr<1s32i>
    %2 = cir.load %1 : !cir.ptr<!s32i>, !s32i
    cir.return %3 : !s32i
} 

Got it, LGTM then. Not related to this patch, but shouldn't the cir.global have the initializer associated with it?