core.math, bignum/f, shift subnormals before rounding. Fixes #1782

diff --git a/core/math/integers/integers-tests.factor b/core/math/integers/integers-tests.factor
index 3e91077c1fabed83b76f6b3b4057b9bda172107c..263a329e261c22f99e20218d9016e91e7d997a06 100644 (file)
--- a/core/math/integers/integers-tests.factor
+++ b/core/math/integers/integers-tests.factor
@@ -280,5 +280,13 @@ IN: math.integers.tests
 { 0x0.6p-1022 } [ 6 1026 2^ /f ] unit-test
 { 0x0.4p-1022 } [ 4 1026 2^ /f ] unit-test
 
+! bignum/f didn't round subnormals
+! biggest subnormal to smallest normal rounding
+{ 0x1.0p-1022 } [ 0xfffffffffffffffffffffffff 1122 2^ /f ] unit-test
+! almost half less than smallest subnormal to smallest subnormal rounding
+{ 0x1.0p-1074 } [ 0x8000000000000000000000001 1122 52 + 2^ /f ] unit-test
+! half less than smallest subnormal to 0
+{ 0.0 } [ 0x8000000000000000000000000 1122 52 + 2^ /f ] unit-test
+
 ! rounding triggering special case in post-scale
 { 1.0 } [ 300 2^ 1 - 300 2^ /f ] unit-test

diff --git a/core/math/integers/integers.factor b/core/math/integers/integers.factor
index bf36076aceda8812502a3137692be9a31cbc2c4b..6467361a608b785c6a4eb7b8a764c2497ee194b6 100644 (file)
--- a/core/math/integers/integers.factor
+++ b/core/math/integers/integers.factor
@@ -119,8 +119,8 @@ M: bignum (log2) bignum-log2 ; inline
 ! As an optimization to minimize the size of the operands of the bignum
 ! divisions we will do, we start by stripping any trailing zeros from
 ! the denominator and move it into the scale factor.
-! We want a result in ]2^54;2^53] to find the mantissa
-! in ]2^53,2^52] with 1 extra "guard" bit for rounding.
+! We want a 54 bit result, starting with leading 1, followed by
+! the 52 bit mantissa and then a guard bit: 1mmmmmmmmmm...mmmmmmmmmmmg
 ! So we shift the numerator to get the result of the integer division
 ! "num/den" in the range ]2^54; 2^53]; Our shift is only a guess
 ! based on the magnitude of the inputs, so it
@@ -148,12 +148,23 @@ M: bignum (log2) bignum-log2 ; inline
 ! "num/den" would be in the range ]2^55; 2^53]. After this step
 ! it will be in the range ]2^54; 2^53]. Compute "num/den" and the
 ! reminder used for rounding
+! For subnormals, after we know the final value of the exponent,
+! we shift the numerator again to get the correct precision.
+! We do it before rounding so that subnormals are correctly rounded.
 : (2/-with-epsilon) ( epsilon? num -- epsilon?' num' )
     [ 1 bitand zero? not or ] [ 2/ ] bi ; inline
 
+: (shift-with-epsilon) ( epsilon? num den scale -- epsilon?' num' den scale )
+    [
+        nip 1021 +
+        [ neg 2^ 1 - bitand zero? not or ] [ shift ] 2bi
+    ] 2keep ; inline
+
 : mantissa-and-guard ( epsilon? num den scale -- epsilon?' mantissa-and-guard rem scale' )
     2over /i log2 53 >
     [ [ (2/-with-epsilon) ] [ ] [ 1 + ] tri* ] when
+    ! At this point, the scale value is the exponent minus 1.
+    dup -1021 < [ (shift-with-epsilon) ] when
     [ /mod ] dip ; inline
 
 ! Third step: rounding
@@ -178,13 +189,18 @@ M: bignum (log2) bignum-log2 ; inline
     ] [ drop nip ] if ; inline
 
 ! Fourth step: post-scaling
-! Because of rounding, our mantissa with guard bit is now in the
-! range [2^54;2^53], so we have to handle 2^54 specially.
+! Because of rounding, our mantissa with guard bit may have overflowed
+! the 54 bit precision to 2^54 so we have to handle it specially.
+! For subnormals, the rounding may also have overflowed the precision,
+! but the overflowed value is actually the correct value by chance
+! (even in the case when the biggest subnormal is rounded up to
+! the smallest normal float) because we interpret it directly
+! as the bits of the resulting double.
 : scale-float ( mantissa scale -- float' )
-    ! At this point, the scale value is the exponent minus 1.
+    ! the scale value is the exponent minus 1.
     {
         { [ dup 1024 > ] [ 2drop 1/0. ] }
-        { [ dup -1021 < ] [ 1021 + shift bits>double ] } ! subnormals and underflow
+        { [ dup -1021 < ] [ drop bits>double ] } ! subnormals and underflow
         [ [ 52 2^ 1 - bitand ] dip 1022 + 52 shift bitor bits>double ]
     } cond ; inline
 

diff --git a/core/math/parser/parser.factor b/core/math/parser/parser.factor
index cb12ad196c39d52c14715a729998db3ccb6f76e7..49ebf8383bf7d9aeadec2266238cfef6b33c9256 100644 (file)
--- a/core/math/parser/parser.factor
+++ b/core/math/parser/parser.factor
@@ -114,7 +114,7 @@ TUPLE: float-parse
 ! Also, take some margin as the current float parsing algorithm
 ! does some rounding; For example,
 ! 0x1.0p-1074 is the smallest IE754 double, but floats down to
-! 0x0.fffffffffffffcp-1074 are parsed as 0x1.0p-1074
+! 0x0.8p-1074 (excluded) are parsed as 0x1.0p-1074
 CONSTANT: max-magnitude-10 309
 CONSTANT: min-magnitude-10 -323
 CONSTANT: max-magnitude-2 1027