C++: Consider columns in Location.isBefore

[MathiasVP]

I was hacking on a solution to #8497 and was surprised to learn that Location.isBefore is really Location.lineIsBefore.

As an added bonus to this PR, this will improve our IR generation for code that uses gotos:

Currently, our only use of Location.isBefore predicate is to determine whether a goto statement creates a back edge in the control-flow graph. So on main, the following program:

int test() {
  int x = 42;
  goto next; next:

  return x;
}

gets translated as:

int test()
  Block 0
    v1_1(void)       = EnterFunction      :
    m1_2(unknown)    = AliasedDefinition  :
    m1_3(unknown)    = InitializeNonLocal :
    m1_4(unknown)    = Chi                : total:m1_2, partial:m1_3
    r2_1(glval<int>) = VariableAddress[x] :
    r2_2(int)        = Constant[42]       :
    m2_3(int)        = Store[x]           : &:r2_1, r2_2
    v3_1(void)       = NoOp               :
  Goto (back edge) -> Block 1

  Block 1
    v3_2(void)       = NoOp                     :
    r6_1(glval<int>) = VariableAddress[#return] :
    r6_2(glval<int>) = VariableAddress[x]       :
    r6_3(int)        = Load[x]                  : &:r6_2, m2_3
    m6_4(int)        = Store[#return]           : &:r6_1, r6_3
    r1_5(glval<int>) = VariableAddress[#return] :
    v1_6(void)       = ReturnValue              : &:r1_5, m6_4
    v1_7(void)       = AliasedUse               : m1_3
    v1_8(void)       = ExitFunction             :

i.e., we conservatively assume that the goto could be a back edge. However, with this PR the same code is translated as:

int test()
  Block 0
    v1_1(void)       = EnterFunction            :
    m1_2(unknown)    = AliasedDefinition        :
    m1_3(unknown)    = InitializeNonLocal       :
    m1_4(unknown)    = Chi                      : total:m1_2, partial:m1_3
    r2_1(glval<int>) = VariableAddress[x]       :
    r2_2(int)        = Constant[42]             :
    m2_3(int)        = Store[x]                 : &:r2_1, r2_2
    v3_1(void)       = NoOp                     :
    v3_2(void)       = NoOp                     :
    r6_1(glval<int>) = VariableAddress[#return] :
    r6_2(glval<int>) = VariableAddress[x]       :
    r6_3(int)        = Load[x]                  : &:r6_2, m2_3
    m6_4(int)        = Store[#return]           : &:r6_1, r6_3
    r1_5(glval<int>) = VariableAddress[#return] :
    v1_6(void)       = ReturnValue              : &:r1_5, m6_4
    v1_7(void)       = AliasedUse               : m1_3
    v1_8(void)       = ExitFunction             :

with no back edge!

I admit this is a bit of a niche case. But it's nice to see that it actually does have the intended effect.

[MathiasVP]

@geoffw0 convinced me that we shouldn't change the behavior of the existing predicate. So I've force-pushed a backward-compatible version.

I've changed IR generation so that it uses the new predicate to get the behavior I described in the PR description.

[jketema]

You might want to add the goto example to the IR tests?

[MathiasVP]

You might want to add the goto example to the IR tests?

Good point! I'll do that.

[MathiasVP]

Done. I did another force-push to insert the test in the right commit.

[jketema]

It shows a warning about an implicit this here. Should that be fixed?

[MathiasVP]

Yeah, it really should. I'll add the explicit this, but I won't force-push the change since that will mess up the DCA run.

[jketema]

Otherwise this LGTM.

[MathiasVP]

Fixed in c35b385.

[jketema]

LGTM.