Also, the permute operations should maybe be looked at and all clarified as part of this as well.

from safe_arch.

Format Proposal

• swiz_{inputs}_{lane-size}_{source}_{data-type}
  • The inputs are a, b, and/or i
  • The lane sizes are f32, f64, iX, and possibly with a z at the end if you can zero any lane source.
  • The sources are "all" for "pick from all lanes", or "half" for "each half picks to just that half", optionally with h64 and l64 for higher 64 bits or lower 64 bits.
  • the data types are the normal SIMD types

So here's each C intrinsic and the name it would have under this scheme:

from safe_arch.

Oh, also, we're using swiz for "swizzle" because sometimes Intel calls it "shuffle" and sometimes it calls it "permute" and there's seemingly no logic to why one or the other is used for each particular op/intrinsic:

• It's not based on the number of inputs
• It's not based on if the inputs are immediate or not
• It's not based on the minimum CPUID
• It's not based on the destination register also being an input register or not

So we'll simply forsake both names and then pick a third name that doesn't have any existing baggage.

from safe_arch.

Oh dag we also have to consider that some b values are the varying pattern and some b values are the 2nd register to mix in.

from safe_arch.