Replace hasSimd macro with explicit when dispatch

[treeform]

Replaces the {.hasSimd.} / {.simd.} macros with plain when dispatch written directly in each proc.

The macro kept a {.compileTime.} table that one module wrote and another read. That breaks under --ic:on (see guzba/nimsimd#38). With this change, no compile-time state is shared between modules and the dispatch is visible in the source.

Each former hasSimd proc now looks like this:

proc invert*(image: Image) {.raises: [].} =
  ## Inverts all of the colors and alpha.
  when allowAvx:
    if cpuHasAvx2:
      invertAvx2(image)
      return
  when allowSse2:
    invertSse2(image)
  elif allowNeon:
    invertNeon(image)
  else:
    # plain path

The plain path sits under else:. A bare return inside when followed by more code makes Nim print an UnreachableCode warning. Most of the diff is that re-indent.

Changes

• simd.nim defines allowSse2, allowAvx and allowNeon next to allowSimd.
• The {.simd.} pragma and the nimsimd/hassimd import are removed from the SIMD modules. The SIMD bodies are unchanged.
• Dispatch order and variant choice are the same as before. fillUnsafe checks AVX, the rest check AVX2, and SSE2 is the amd64 baseline.
• New -d:pixieAvx2 flag. It turns cpuHasAvx and cpuHasAvx2 into const true, so builds for a fixed architecture skip the runtime CPU check. Without the flag the runtime check stays as before.

Testing

• nim check has no errors for arm64, -d:pixieNoSimd, amd64, amd64 -d:pixieNoAvx and amd64 -d:pixieAvx2.
• tests/tests.nim runs through to the end on arm64 (NEON) and with -d:pixieNoSimd.
• An x86_64 release build exits 0 under Rosetta. Rosetta reports no AVX2, so that run only covered the SSE2 path. The AVX2 path still needs a run on real x86 hardware or CI.

🤖 Generated with Claude Code