llvm-project

1ba2b22d - [X86][CostModel] Price predicate mask-expansion fanout for wide predicated ops (#201666)

[X86][CostModel] Price predicate mask-expansion fanout for wide predicated ops (#201666) Teach the X86 TTI cost model to price the *predicate mask-expansion fanout* of wide predicated operations. When a masked memory op, gather/scatter, or select produces its predicate as a compact `<N x i1>` but the data type legalizes into more register parts than that predicate, codegen must materialize a sub-mask for every extra data part — a `kshiftr` out of the k-register on AVX-512, or a vector unpack/sign-extend on SSE/AVX. The per-part cost tables price these ops linearly in data parts and miss this fanout.