i1* function argument on x86-64

Project Infrastructure LLVM Dev List Archives
1

I am running into a problem with 'i1*' as a function's argument which seems to have appeared since I switched to LLVM 3.6 (but can have other source, of course). If I look at the assembler that the MCJIT generates for an x86-64 target I see that the array 'i1*' is taken as a sequence of 1 bit wide elements. (I guess that's correct). However, I used to call the function from C passing in a 'bool*' which has 1 byte wideelements, I guess. (not sure if that's a compiler's choice) Now, since I haven't changed my code on these parts but only made the transition from LLVM 3.4/5 -> 3.6 I wonder if the element width has changed when i1* is used as a function's argument..!?

Thanks,
Frank

2

I attach the orginal IR and the generated assembly where one can see that the array elements are treated 1 bit wide. Is this the intended behavior? I doubt it, because I am passing in C pointers, and those have byte granularity. (In C a bit cannot be addressed with a pointer).

Frank

copymask.ll (3.31 KB)

copymask.S (1.74 KB)

3

I am running into a problem with 'i1*' as a function's argument which
seems to have appeared since I switched to LLVM 3.6 (but can have other
source, of course). If I look at the assembler that the MCJIT generates
for an x86-64 target I see that the array 'i1*' is taken as a sequence
of 1 bit wide elements. (I guess that's correct).

Yes, that is correct.

However, I used to
call the function from C passing in a 'bool*' which has 1 byte
wideelements, I guess.

This is the bug, as the caller and callee need to agree on how data for arguments is laid out. 'bool' is i8 in clang when stored to memory (for most targets... IIRC, one of the ppc darwin targets has 32bit bools?).

(not sure if that's a compiler's choice) Now,

Width of 'bool' is a frontend choice, yes. In C, the minimum addressable thing has to have the same size as char, and the assumption that char is 8 bits is very baked into llvm.

since I haven't changed my code on these parts but only made the
transition from LLVM 3.4/5 -> 3.6 I wonder if the element width has
changed when i1* is used as a function's argument..!?

I'm not sure of the history here. Perhaps it could have been a bugfix?

Cheers,

Jon

4

Okay, it might got fixed along the way.
Then, the only portable way of doing this is to change the function argument from 'i1*' to 'i<X>*' (plus the code changing it to an i1) and add a new parameter <X> to my code which specifies the bit width of a bool that the C compiler uses.

Frank