Signed bitcode values emitted as unsigned C

[artemdinaburg]

Rellic will sometimes forget signedess and emit unsigned casts for signed integers. Simple example below:

signed.c:

#include <stdbool.h>

bool signed_less_than_1(signed short v) {
  return v < 1;
}

Compile:

clang-11 -O1 -std=c99 -emit-llvm -c -o signed.bc signed.c

Rellic:

rellic-decomp-11.0 --input signed.bc --output signed_decomp.c

Output:

unsigned char signed_less_than_1(unsigned short arg0);
unsigned char signed_less_than_1(unsigned short arg0) {
    return (arg0 < (unsigned short)1U);
}

The bitcode in question performs a signed compare, and rellic should emit a signed compare as well:

define dso_local zeroext i1 @signed_less_than_1(i16 signext %0) local_unnamed_addr #0 {
  %2 = icmp slt i16 %0, 1
  ret i1 %2
}

[surovic]

Resolved by #121

Unfortunately I haven't found a good way of regression testing this one using our executable roundtrip tests, because we don't have type inference.

[greenbagels]

I might just be out of the loop, but it seems that the issue still exists; after building 356ecb1 and reproducing the test case, the LLVM IR is identical to before:

define dso_local zeroext i1 @signed_less_than_1(i16 signext %0) local_unnamed_addr #0 {
  %2 = icmp slt i16 %0, 1
  ret i1 %2
}

while the produced roundtrip code is slightly different, but still with the wrong type signature:

unsigned char signed_less_than_1(unsigned short arg0);
unsigned char signed_less_than_1(unsigned short arg0) {
    return (short)arg0 < 1;
}