bjorn3 commented on Issue #1094:
I believe the new style backends have good SIMD support. In any case there are other issues tracking SIMD support for the individual new style backends.
cfallin closed Issue #1094:
$ cat x.clif
target x86_64 haswell function %vector_split_args(i32x2) -> i32x2 fast { ebb0(v0: i32x2): v1 = bxor v0, v0 return v1 }
$ cranelift/target/debug/clif-util compile -d x.clif
function %vector_split_arg(i32 [%rdi], i32 [%rsi]) -> i32 [%rax], i32 [%rdx] fast { ebb0(v2: i32, v3: i32): [-] v0 = vconcat v2, v3 v4, v5 = isplit v0 v6, v7 = isplit v0 v10, v11 = isplit v4 v12, v13 = isplit v6 v14 = bxor v10, v12 v15 = bxor v11, v13 v8 = iconcat v14, v15 v16, v17 = isplit v5 v18, v19 = isplit v7 v20 = bxor v16, v18 v21 = bxor v17, v19 v9 = iconcat v20, v21 [-] v1 = iconcat v8, v9 v22, v23 = vsplit v1 [Op1ret#c3] return v22, v23 ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ; error: inst1: v22 is a ghost value used by a real [Op1ret#c3] instruction } ; 1 verifier error detected (see above). Compilation aborted.
It looks like vsplit/vconcat are inserted as part of ABI legalization? Are they supposed to be subsequently legalized into something else?
If you use "target arm" it also fails but there is no vsplit of v1, it just attempts to "return v1" and that fails with "Return must have an encoding".
Last updated: Nov 22 2024 at 16:03 UTC