ref: 5e1154d69efc86f5fe3831b047e3531d9cfd3478
parent: 6fb7d34b848c598e6398f960dc0a7e240e09a6f5
author: S. Gilles <sgilles@math.umd.edu>
date: Mon Mar 12 03:09:31 EDT 2018
Make floor & friends slower, but more readable
--- a/lib/math/fpmath-trunc-impl.myr
+++ b/lib/math/fpmath-trunc-impl.myr
@@ -9,152 +9,95 @@
pkglocal const ceil64 : (f : flt64 -> flt64)
;;
-pkglocal const trunc32 = {f : flt32- var u : uint32 = std.flt32bits(f)
- var e : uint32 = (((u >> 23) & 0xff) : uint32) - 127
- var e_lt_zero : uint32 = ((e >> 31) : uint32) & 0x1
- var e_ge_zero : uint32 = 1 - e_lt_zero
- var e_ge_23 : uint32 = 1 - ((e - 23) >> 31)
- /*
- The significand 1 . m1 m2 ... m23 needs to be
- truncated, which corresponds to zeroing all mi
- bits where i is beyond the exponent e (they are
- the actual sub-integer portion).
- */
- var m : uint32 = ~(((1 << 23) - 1) >> e)
- m |= (-1 : uint32) * e_ge_23
- var v_ge_zero : uint32 = (u & m) * e_ge_zero
+const Flt32NegMask : uint32 = (1 << 31)
+const Flt32SigMask : uint32 = (1 << 23) - 1
- /*
- On the other hand, if the exponent is < 0, "23
- - e" is garbage, and we should just return +/-
- zero
- */
- var v_lt_zero : uint32 = (u & (1 << 31)) * e_lt_zero
+const Flt64NegMask : uint64 = (1 << 63)
+const Flt64SigMask : uint64 = (1 << 52) - 1
- /* Try to save a branch */
- var v : uint32 = v_ge_zero + v_lt_zero
- -> std.flt32frombits(v)
-}
-
pkglocal const floor32 = {f : flt32- var u : uint32 = std.flt32bits(f)
- var e : int32 = (((u >> 23) & 0xff) : int32) - 127
- var shift_e : uint32 = (e : uint32)
+ var n, e, s
+ (n, e, s) = std.flt32explode(f)
/* Many special cases */
- if e >= 23 || u == 0x80000000
+ if e >= 23 || f == -0.0
-> f
- elif (e < 0) && (u & (1 << 31) != 0)
- -> -1.0
elif e < 0
- -> 0.0
+ if n
+ -> -1.0
+ else
+ -> 0.0
+ ;;
;;
- if u & (1 << 31) != 0
- var fractional_mask : uint32 = (((1 << 23) - 1) >> shift_e)
- var v : uint32 = u & ~fractional_mask
- if (u & fractional_mask) != 0
- v += ((1 << 23) >> shift_e)
+ if n
+ var fractional_mask = Flt32SigMask >> (e : uint32)
+ if s & fractional_mask == 0
+ -> f
+ else
+ /* Turns out the packing of exp and sig is useful */
+ var u : uint32 = std.flt32bits(f) & ~fractional_mask
+ u += ((1 << 23) >> (e : uint32))
+ -> std.flt32frombits(u)
;;
- -> std.flt32frombits(v)
;;
- var m : uint32 = ~(((1 << 23) - 1) >> shift_e)
- var v : uint32 = u & m
- -> std.flt32frombits(v)
+ var m : uint32 = (Flt32SigMask >> (e : uint32))
+ -> std.flt32assem(n, e, s & ~m)
}
-pkglocal const ceil32 = {f;- var u : uint32 = std.flt32bits(f)
- var e : int32 = (((u >> 23) & 0xff) : int32) - 127
- var shift_e : uint32 = (e : uint32)
- if e >= 23 || u == 0x0
- -> f
- elif (e < 0) && (u & (1 << 31) == 0)
- -> 1.0
- elif e < 0
- -> -0.0
+pkglocal const trunc32 = {f : flt32+ if std.flt32bits(f) & Flt32NegMask != 0
+ -> -floor32(-f)
+ else
+ -> floor32(f)
;;
-
- if u & (1 << 31) == 0
- var fractional_mask : uint32 = (((1 << 23) - 1) >> shift_e)
- var v : uint32 = u & ~fractional_mask
- if (u & fractional_mask) != 0
- v += ((1 << 23) >> shift_e)
- ;;
- -> std.flt32frombits(v)
- ;;
-
- var m : uint32 = ~(((1 << 23) - 1) >> shift_e)
- var v : uint32 = u & m
- -> std.flt32frombits(v)
}
-pkglocal const trunc64 = {f : flt64- var u : uint64 = std.flt64bits(f)
- var e : uint64 = (((u >> 52) & 0x7ff) : uint64) - 1023
- var e_lt_zero : uint64 = ((e >> 63) : uint64) & 0x1
- var e_ge_zero : uint64 = 1 - e_lt_zero
- var e_ge_52 : uint64 = 1 - ((e - 52) >> 63)
- var m : uint64 = ~(((1 << 52) - 1) >> e)
- m |= (-1 : uint64) * e_ge_52
- var v_ge_zero : uint64 = (u & m) * e_ge_zero
- var v_lt_zero : uint64 = (u & (1 << 63)) * e_lt_zero
- var v : uint64 = v_ge_zero + v_lt_zero
- -> std.flt64frombits(v)
+pkglocal const ceil32 = {f : flt32+ -> -floor32(-f)
}
pkglocal const floor64 = {f : flt64- var u : uint64 = std.flt64bits(f)
- var e : int64 = (((u >> 52) & 0x7ff) : int64) - 1023
- var shift_e : uint64 = (e : uint64)
+ var n, e, s
+ (n, e, s) = std.flt64explode(f)
- if e >= 52 || u == 0x8000000000000000ul
+ /* Many special cases */
+ if e >= 52 || f == -0.0
-> f
- elif (e < 0) && (u & (1 << 63) != 0)
- -> -1.0
elif e < 0
- -> 0.0
+ if n
+ -> -1.0
+ else
+ -> 0.0
+ ;;
;;
- if u & (1 << 63) != 0
- var fractional_mask : uint64 = (((1 << 52) - 1) >> shift_e)
- var v : uint64 = u & ~fractional_mask
- if (u & fractional_mask) != 0
- v += ((1 << 52) >> shift_e)
+ if n
+ var fractional_mask = Flt64SigMask >> (e : uint64)
+ if s & fractional_mask == 0
+ -> f
+ else
+ /* Turns out the packing of exp and sig is useful */
+ var u : uint64 = std.flt64bits(f) & ~fractional_mask
+ u += ((1 << 52) >> (e : uint64))
+ -> std.flt64frombits(u)
;;
- -> std.flt64frombits(v)
;;
- var m : uint64 = ~(((1 << 52) - 1) >> shift_e)
- var v : uint64 = u & m
- -> std.flt64frombits(v)
+ var m : uint64 = (Flt64SigMask >> (e : uint64))
+ -> std.flt64assem(n, e, s & ~m)
}
-pkglocal const ceil64 = {f;- var u : uint64 = std.flt64bits(f)
- var e : int64 = (((u >> 52) & 0x7ff) : int64) - 1023
- var shift_e : uint64 = (e : uint64)
-
- if e >= 52 || u == 0x0ul
- -> f
- elif (e < 0) && (u & (1 << 63) == 0)
- -> 1.0
- elif e < 0
- -> -0.0
+pkglocal const trunc64 = {f : flt64+ if std.flt64bits(f) & Flt64NegMask != 0
+ -> -floor64(-f)
+ else
+ -> floor64(f)
;;
+}
- if u & (1 << 63) == 0
- var fractional_mask : uint64 = (((1 << 52) - 1) >> shift_e)
- var v : uint64 = u & ~fractional_mask
- if (u & fractional_mask) != 0
- v += ((1 << 52) >> shift_e)
- ;;
- -> std.flt64frombits(v)
- ;;
-
- var m : uint64 = ~(((1 << 52) - 1) >> shift_e)
- var v : uint64 = u & m
- -> std.flt64frombits(v)
+pkglocal const ceil64 = {f : flt64+ -> -floor64(-f)
}
+
--- a/lib/math/test/fpmath-trunc-impl.myr
+++ b/lib/math/test/fpmath-trunc-impl.myr
@@ -58,6 +58,9 @@
(10664524000000000000.0, 10664524000000000000.0),
(-3.5, -4.0),
(-101.999, -102.0),
+ (-126.999, -127.0),
+ (-127.999, -128.0),
+ (-128.999, -129.0),
(std.flt32nan(), std.flt32nan()),
][:]