[WebAssembly] SIMD operations on `half` make `llc` hang

[folkertdev]

Float comparisons (at least) on half make llc hang when using simd128

https://godbolt.org/z/nd8M99bnM

define <8 x i16> @foo(<8 x half> %x) {
start:
  %1 = fcmp oeq <8 x half> %x, zeroinitializer
  %2 = sext <8 x i1> %1 to <8 x i16>
  ret <8 x i16> %2
}

this one also hangs

define <8 x i1> @compare_oeq_v8f16 (<8 x half> %x, <8 x half> %y) {
  %res = fcmp oeq <8 x half> %x, %y
  ret <8 x i1> %res
}

based on some basic debugging it might be a loop in type legalization. But I'm not really sure what's going on or how to narrow this down really.

found in rust-lang/portable-simd#513.

[llvmbot]

@llvm/issue-subscribers-backend-webassembly

Author: Folkert de Vries (folkertdev)

Details

Float comparisons (at least) on `half` make `llc` hang when using `simd128`

https://godbolt.org/z/nd8M99bnM

define &lt;8 x i16&gt; @<!-- -->foo(&lt;8 x half&gt; %x) {
start:
  %1 = fcmp oeq &lt;8 x half&gt; %x, zeroinitializer
  %2 = sext &lt;8 x i1&gt; %1 to &lt;8 x i16&gt;
  ret &lt;8 x i16&gt; %2
}

this one also hangs

define &lt;8 x i1&gt; @<!-- -->compare_oeq_v8f16 (&lt;8 x half&gt; %x, &lt;8 x half&gt; %y) {
  %res = fcmp oeq &lt;8 x half&gt; %x, %y
  ret &lt;8 x i1&gt; %res
}

based on some basic debugging it might be a loop in type legalization. But I'm not really sure what's going on or how to narrow this down really.

found in rust-lang/portable-simd#513.

[folkertdev]

I've narrowed it down somewhat. The issue seems to be that SETCC won't lower v8f16 properly when simd128 is active, but fp16 is not. I have an explicit fix for that, but there might be a more high-level fix for the "yes simd128 but not fp16" case?

[badumbatish]

this seems related to #171908 but i'm not sure. I looked into the mentioned issue a bit and its a bit nasty

[ParkHanbum]

define <4 x i1> @f(ptr %out, <4 x half> %v) {
  %cmp = fcmp une <4 x half> %v, zeroinitializer
  ret <4 x i1> %cmp
}

this also causes llc hang.
https://godbolt.org/z/K1v8GoM64

but not causes hang older version of llc. (under 21.1.0)

[ParkHanbum]

define <4 x i1>  @f(ptr %out, half %x) {
  %v = insertelement <4 x half> poison, half %x, i32 0
  %cmp = fcmp une <4 x half> %v, zeroinitializer 
  ret <4 x i1> %cmp
}

similar case but causes crash not hang.
https://godbolt.org/z/acz1EY6n8