Invalid C generation

[xTachyon]

I was playing with llvm-cbe, and with this input file transformed with llvm-cbe, I get the following first error:

/home/tachyon/.cargo/registry/src/index.crates.io-6f17d22bba15001f/adler-1.0.2/src/algo.rs: In function ‘_ZN5adler4algo32__EC_LT_EC_impl_EC_u20_EC_adler_OC__OC_Adler32_EC_GT_EC_7compute17h715de753f96090aaE’:
/home/tachyon/.cargo/registry/src/index.crates.io-6f17d22bba15001f/adler-1.0.2/src/algo.rs:89:51: error: ‘struct l_struct_algo_KD__KD_U32X4’ has no member named ‘array’
   89 |         b_vec.0[1] += MOD - a_vec.0[1];
      |

And without #lines:

main.cbe.c: In function ‘_ZN5adler4algo32__EC_LT_EC_impl_EC_u20_EC_adler_OC__OC_Adler32_EC_GT_EC_7compute17h715de753f96090aaE’:
main.cbe.c:1085:51: error: ‘struct l_struct_algo_KD__KD_U32X4’ has no member named ‘array’
 1085 |   llvm_cbe__79 = *(uint32_t*)(((&(&llvm_cbe_a_vec)->array[((int64_t)1)])));
      | 

The line in cause:

llvm_cbe__79 = *(uint32_t*)(((&(&llvm_cbe_a_vec)->array[((int64_t)1)])));

And the struct:

struct l_struct_algo_KD__KD_U32X4 {
  struct l_array_4_uint32_t field0;
};

I tried miniming the input as much as I could, but I don't know how I'd do it more.

[zlfn]

I think llvm-cbe does not support vector type ([4 x i32]) of LLVM-IR so well.

If you are using Rust, disable vectorization is a not perfect solution, but it likely help.

[xTachyon]

I'm not sure this is a vector type. I think these are the instructions:

%"algo::U32X4" = type { [4 x i32] }
...
%a_vec = alloca %"algo::U32X4", align 4
...
%41 = getelementptr inbounds [4 x i32], ptr %a_vec, i64 0, i64 1, !dbg !554
%_79 = load i32, ptr %41, align 4, !dbg !554, !noundef !34

Accoding to the manual, these are arrays. Vector types have <>, not [].

That being said, the function seems to contain vector types. Maybe that's what it confuses it?

And yes, this is Rust, but neither no-vectorize-loops nor no-vectorize-slp seem to change anything, probably because this is unoptimized code, and those opts never ran.

[vtjnash]

There are some issues people have observed with GEP emission (getting the type wrong and/or forgetting casts) which seems like what you are seeing. PRs welcome