Incorrect callback resolution in pointer analysis

[Frankenween]

While using the wonderful SVF tool for building callgraphs I noticed that some pretty obvious callbacks are not resolved.
Problematic pattern looks like this:

struct t1 {
	void(*ptr)();
};

struct t1 o1 = {
	.ptr = f1
};

void test1() {
	o1.ptr();
}

Call test1 -> f1 is not resolved. However, when there are several callbacks in a struct, the rest seem to be resolved nicely. I thought the problem is with load vs gep:

; Function Attrs: noinline nounwind optnone uwtable
define dso_local void @test6() #1 {
  %1 = load ptr, ptr @t6, align 8
  call void (...) %1()
  %2 = load ptr, ptr getelementptr inbounds (%struct.s1, ptr @t6, i32 0, i32 1), align 8
  call void (...) %2()
  ret void
}

But changing first load to gep manually didn't help.
In my task the problem appeared with different fields, not with the first only. Looks like the problem is not Andersen-specific as almost every analysis failed. On the provided test Steensgaard succeeded, but it also failed in my original problem.

test.zip - here test file, its IR and the result of wpa -ander -dump-callgraph.

[yuleisui]

@jumormt could you take a look at this case?

[yuleisui]

It should be correct if you rename test1 as main since the program has to have a main driver function.

[Frankenween]

Nope, it didn't help.
test.zip

[jumormt]

Nope, it didn't help. test.zip

@Frankenween I believe your observation is correct. The opaque pointer version bitcode miss a gep before loading. This can be solved by turning the ff-eq-base option on: wpa -ander -ff-eq-base -dump-callgraph /path/to/your/bitcode

[jumormt]

This option basically treats the first field equally as the base object.

Nope, it didn't help. test.zip

@Frankenween I believe your observation is correct. The opaque pointer version bitcode miss a gep before loading. This can be solved by turning the ff-eq-base option on: wpa -ander -ff-eq-base -dump-callgraph /path/to/your/bitcode

[Frankenween]

Life-saving, thanks! I'll try it on my main project later