Split-up compile-time CPU feature detect w/compile-time CPU feature use.

Emit a warning when things might be so slow that the main optimization
idea of nsort may fail.

adix/nsort.nim

@@ -60,30 +60,29 @@ const nTinySort{.intdefine.} = 24 # Usually 16-32 is good
 const hMaxBits{.intdefine.} = 15 #2*counter-size*2**hMaxBits =~ L2 here
 const dGiantSort{.intdefine.} = 15 shl 30 #~50% of uncontended DIMM storage
 
-when defined(amd64) and (defined(gcc) or defined(clang)) and#XXX Run-Time CPU
- not defined(portablensort):
- {.passc: "-march=native".} # ..feature test.
+const haveBMI2 =
+ when defined(amd64) and not defined(noSIMD):
+ when defined(gcc): # ||||||||||||| gcc.options here (for X-compiles)
+ staticExec("touch j.c;gcc -march=native -dM -E j.c|grep __BMI2__;rm -f j.c").len > 0
+ elif defined(clang):
+ staticExec("touch j.c;clang -march=native -dM -E j.c|grep __BMI2__;rm -f j.c").len > 0
+ else: false
+
+when haveBMI2:
  proc pext(val,msk:uint32):uint32 {.importc:"_pext_u32", header:"x86intrin.h".}
  proc pext(val,msk:uint64):uint64 {.importc:"_pext_u64", header:"x86intrin.h".}
 else:
- proc pext(val, msk: uint32): uint32 =
- var val = cast[int32](val)
- var msk = cast[int32](msk)
- var bb = 1'u32
- while msk != 0:
- if (val and msk and -msk) != 0:
- result = result or bb
- msk = msk and (msk - 1)
- bb += bb
- proc pext(val, msk: uint64): uint64 =
- var val = cast[int64](val)
- var msk = cast[int64](msk)
- var bb = 1'u64
- while msk != 0:
- if (val and msk and -msk) != 0:
- result = result or bb
- msk = msk and (msk - 1)
- bb += bb
+ {.warning: "nsort compiling without BMI2; Likely slow".}
+ template pextPortable(Ts, Tu) =
+ proc pext(val, msk: Tu): Tu =
+ let val = val.Ts
+ var msk = msk.Ts
+ var bb = 1'u32
+ while msk != 0:
+ if (val and msk and -msk) != 0: result = result or bb
+ msk = msk and (msk - 1)
+ bb += bb
+ pextPortable(int32, uint32); pextPortable(int64, uint64)
 
 when defined(windows): #import alloca
  proc alloca(n:int): pointer{.importc,header:"<malloc.h>".}