[X86] Poor codegen when bitcasting boolean vectors larger than 64 bits

[jhuber6]

These types are usually represented by integers, a bitcast from a vector of 128 bools to a __uint128_t type generates over 800 lines of assembly https://godbolt.org/z/74vbacq48

Fixing this isn't a huge priority, as 64-bits (a mask for 64 bytes in AVX512) is likely the highest that will be practically used, but I figured I'd make a note of this.

[bala-bhargav]

@jhuber6
can i work on it ??

[RKSimon]

define i128 @b(<128 x i1> %m) {
  %t = bitcast <128 x i1> %m to i128
  ret i128 %t
}

This is due to the <128 x i1> being represented as 128 i8 elements in the ABI - the first few in registers, then the rest on the stack.

What we're needing is to realise that this is something similar to this pattern:

define i128 @c(i8 %a0, i8 %a1, i8 %a2, ptr %p) {
  %v = load <128 x i8>, ptr %p
  %v0 = insertelement <128 x i8> %v , i8 %a0, i32 0
  %v1 = insertelement <128 x i8> %v0, i8 %a1, i32 1
  %v2 = insertelement <128 x i8> %v1, i8 %a2, i32 2
  %t0 = and <128 x i8> %v2, splat (i8 1)
  %t1 = icmp ne <128 x i8> %t0, zeroinitializer
  %t2 = bitcast <128 x i1> %t1 to i128
  ret i128 %t2
}

But load combining an odd subvector of i1 anyextended as i8 isn't going to be easy.....