Missing support for memory(argmem) with escaped ptrs through ptr2int

[regehr]

I was debugging a issue found by arm-tv and it boils down alive-tv not being happy about this input. I can't figure out what's going on. the problem depends on the (unused) argument %0 being there.

@X = external global i32

define i32 @test2(ptr %0) {
  %2 = load i32, ptr @X, align 4
  ret i32 %2
}

regehr@john-home:~/reduce$ ~/alive2-regehr/build/alive-tv reduced.ll 

----------------------------------------
@X = global 4 bytes, align 4

define i32 @test2(ptr %#0) {
#1:
  %#2 = load i32, ptr @X, align 4
  ret i32 %#2
}
=>
@X = global 4 bytes, align 4

define i32 @test2(ptr nocapture noread nowrite %#0) nofree willreturn memory(globals: read, errno: read, other: read) {
#1:
  %#2 = load i32, ptr @X, align 4
  ret i32 %#2
}
Transformation doesn't verify!

ERROR: Source is more defined than target

Example:
ptr %#0 = pointer(non-local, block_id=1, offset=0)

Source:
i32 %#2 = poison

SOURCE MEMORY STATE
===================
NON-LOCAL BLOCKS:
Block 0 >	size: 0	align: 4	alloc type: 0	alive: false	address: 0
Contents:
else: poison

Block 1 >	size: 4	align: 4	alloc type: 0	alive: true	address: 8
Contents:
else: poison

Block 2 >	size: 0	align: 1	alloc type: 0	alive: true	address: 8
Contents:
else: poison


Target:
i32 %#2 = UB triggered!


Summary:
  0 correct transformations
  1 incorrect transformations
  0 failed-to-prove transformations
  0 Alive2 errors
regehr@john-home:~/reduce$ 

cc @tanmaytirpankar

[nunoplopes]

this requires rewriting the support for memory(argmem). The implementation is semantic, but LangRef uses a "based-on" definition.
The full implementation is non-trivial due to ptr2int/int2ptr round-trips (ptr2int must escape the pointer I guess), but we can start with a simple data-flow propagation only.

[regehr]

I see... so this is about the attribute inference pass. makes sense. as a short-term workaround, should we maybe drop some of the inferred attributes that we get back from -O2 on code lifted from ARM assembly? that would be easy and sound.

[nunoplopes]

Ah, right, because you're running -O2 on the lifted code.
Dropping all the attributes should be fine (or just make them equal to the source).
I should be able to fix this within a few days. Just need to sleep today 😅

[nunoplopes]

Ok, partial implementation in. Should be good enough for your use cases. Doesn't take escape pointers into account.

[regehr]

awesome, thanks!