[libclc] Optimize isfpclass-like CLC builtins

[frasercrmck]

The builtins we were using to implement __clc_is(finite|inf|nan|normal) -- __builtin_isfinite, etc. -- don't take vector types so we were previously scalarizing. The __builtin_isfpclass builtin does take vector types and thus allows us to keep things in vectors.

There is no change in codegen to the scalar versions of any of these builtins.

[arsenm]

How does using isfpclass avoid scalarization here? I think it's somewhat preferably to use the named operations here, they are subtly different since they canonicalize the input unlike is.fpclass

[frasercrmck]

How does using isfpclass avoid scalarization here? I think it's somewhat preferably to use the named operations here, they are subtly different since they canonicalize the input unlike is.fpclass

The builtins we were using before, like __builtin_isnan, don't take vector types so we were forced to scalarize.

I actually started looked into adding __builtin_elementwise_isnan etc. to clang before realizing that __builtin_isfpclass(x, 0x3) accepts vector types and generates the same code as __builtin_isnan(x) does for scalar types (and essentially the same for vectors). I don't see any input canonicalization going on before this change.

in function _Z5isnanDv2_f:
  in block %entry:
    >   %0 = fcmp uno <2 x float> %a, zeroinitializer
    >   %sext.i = sext <2 x i1> %0 to <2 x i32>
    >   ret <2 x i32> %sext.i
    <   %0 = extractelement <2 x float> %a, i64 0
    <   %1 = fcmp uno float %0, 0.000000e+00
    <   %2 = zext i1 %1 to i32
    <   %vecinit.i = insertelement <2 x i32> poison, i32 %2, i64 0
    <   %3 = extractelement <2 x float> %a, i64 1
    <   %4 = fcmp uno float %3, 0.000000e+00
    <   %5 = zext i1 %4 to i32
    <   %vecinit2.i = insertelement <2 x i32> %vecinit.i, i32 %5, i64 1
    <   %cmp.i = icmp ne <2 x i32> %vecinit2.i, zeroinitializer
    <   %sext.i = sext <2 x i1> %cmp.i to <2 x i32>
    <   ret <2 x i32> %sext.i

[arsenm]

I remember debugging inconsistent behavior on different platforms in treatment of signaling nans and denormal flushing with these queries. I think these days we emit them as is.fpclass anyway, and then instcombine turns them into fcmp when valid. This may be a backwards system, so this may need revisiting in the future