Remove "S.build" constructors

[stoklund]

The initial proposal in #1 includes S.build() constructors that create SIMD vectors from dynamic per-lane arguments:

• b8x16.build(x: i32[16]) -> b8x16
• b16x8.build(x: i32[8]) -> b16x8
• b32x4.build(x: i32[4]) -> b32x4
• b64x2.build(x: i32[2]) -> b64x2
• i8x16.build(x: i32[16]) -> v128
• i16x8.build(x: i32[8]) -> v128
• i32x4.build(x: i32[4]) -> v128
• i64x2.build(x: i64[2]) -> v128
• f32x4.build(x: f32[4]) -> v128
• f64x2.build(x: f64[2]) -> v128

@sunfishcode suggested that they are redundant since you can use replace_lane instead to insert one lane at a time, starting from some v128.const.

Prior art

LLVM IR does not have vector constructors like these. The convention is to begin from an undef of the right type and insert one lane at a time with the insertelement instruction.

LLVM's SelectionDAG code generator will pattern-match the long sequences of insertelement instructions and transform them into an ISD::BUILD_VECTOR node which looks more like the build instructions in the proposal.

[billbudge]

Seconded!