compiler: add fixnum-min/max intrinsics; ~10% speedup on benchmark.yuv-to-rgb

basis/compiler/cfg/hats/hats.factor

@@ -35,6 +35,8 @@ IN: compiler.cfg.hats
 : ^^shr-imm ( src1 src2 -- dst ) ^^r2 ##shr-imm ; inline
 : ^^sar ( src1 src2 -- dst ) ^^r2 ##sar ; inline
 : ^^sar-imm ( src1 src2 -- dst ) ^^r2 ##sar-imm ; inline
+: ^^min ( src1 src2 -- dst ) ^^r2 ##min ; inline
+: ^^max ( src1 src2 -- dst ) ^^r2 ##max ; inline
 : ^^not ( src -- dst ) ^^r1 ##not ; inline
 : ^^log2 ( src -- dst ) ^^r1 ##log2 ; inline
 : ^^bignum>integer ( src -- dst ) ^^r1 next-vreg ##bignum>integer ; inline

basis/compiler/cfg/instructions/instructions.factor

@@ -91,6 +91,8 @@ INSN: ##shr < ##binary ;
 INSN: ##shr-imm < ##binary-imm ;
 INSN: ##sar < ##binary ;
 INSN: ##sar-imm < ##binary-imm ;
+INSN: ##min < ##binary ;
+INSN: ##max < ##binary ;
 INSN: ##not < ##unary ;
 INSN: ##log2 < ##unary ;
 

basis/compiler/cfg/intrinsics/intrinsics.factor

@@ -25,6 +25,9 @@ QUALIFIED: math.floats.private
 QUALIFIED: math.libm
 IN: compiler.cfg.intrinsics
 
+: enable-intrinsics ( words -- )
+    [ t "intrinsic" set-word-prop ] each ;
+
 {
     kernel.private:tag
     kernel.private:getenv
@@ -67,7 +70,7 @@ IN: compiler.cfg.intrinsics
     alien.accessors:set-alien-signed-2
     alien.accessors:alien-cell
     alien.accessors:set-alien-cell
-} [ t "intrinsic" set-word-prop ] each
+} enable-intrinsics
 
 : enable-alien-4-intrinsics ( -- )
     {
@@ -75,7 +78,7 @@ IN: compiler.cfg.intrinsics
         alien.accessors:set-alien-unsigned-4
         alien.accessors:alien-signed-4
         alien.accessors:set-alien-signed-4
-    } [ t "intrinsic" set-word-prop ] each ;
+    } enable-intrinsics ;
 
 : enable-float-intrinsics ( -- )
     {
@@ -94,7 +97,7 @@ IN: compiler.cfg.intrinsics
         alien.accessors:set-alien-float
         alien.accessors:alien-double
         alien.accessors:set-alien-double
-    } [ t "intrinsic" set-word-prop ] each ;
+    } enable-intrinsics ;
 
 : enable-fsqrt ( -- )
     \ math.libm:fsqrt t "intrinsic" set-word-prop ;
@@ -103,10 +106,16 @@ IN: compiler.cfg.intrinsics
     {
         math.floats.private:float-min
         math.floats.private:float-max
-    } [ t "intrinsic" set-word-prop ] each ;
+    } enable-intrinsics ;
+
+: enable-min/max ( -- )
+    {
+        math.integers.private:fixnum-min
+        math.integers.private:fixnum-max
+    } enable-intrinsics ;
 
 : enable-fixnum-log2 ( -- )
-    \ math.integers.private:fixnum-log2 t "intrinsic" set-word-prop ;
+    { math.integers.private:fixnum-log2 } enable-intrinsics ;
 
 : emit-intrinsic ( node word -- )
     {
@@ -130,6 +139,8 @@ IN: compiler.cfg.intrinsics
         { \ math.private:fixnum>= [ drop cc>= emit-fixnum-comparison ] }
         { \ math.private:fixnum> [ drop cc> emit-fixnum-comparison ] }
         { \ kernel:eq? [ drop cc= emit-fixnum-comparison ] }
+        { \ math.integers.private:fixnum-min [ drop [ ^^min ] emit-fixnum-op ] }
+        { \ math.integers.private:fixnum-max [ drop [ ^^max ] emit-fixnum-op ] }
         { \ math.private:bignum>fixnum [ drop emit-bignum>fixnum ] }
         { \ math.private:fixnum>bignum [ drop emit-fixnum>bignum ] }
         { \ math.private:float+ [ drop [ ^^add-float ] emit-float-op ] }

basis/compiler/cfg/two-operand/two-operand.factor

@@ -35,6 +35,8 @@ UNION: two-operand-insn
     ##shr-imm
     ##sar
     ##sar-imm
+    ##min
+    ##max
     ##fixnum-overflow
     ##add-float
     ##sub-float

basis/compiler/codegen/codegen.factor

@@ -149,6 +149,8 @@ M: ##shr     generate-insn dst/src1/src2 %shr     ;
 M: ##shr-imm generate-insn dst/src1/src2 %shr-imm ;
 M: ##sar     generate-insn dst/src1/src2 %sar     ;
 M: ##sar-imm generate-insn dst/src1/src2 %sar-imm ;
+M: ##min     generate-insn dst/src1/src2 %min     ;
+M: ##max     generate-insn dst/src1/src2 %max     ;
 M: ##not     generate-insn dst/src       %not     ;
 M: ##log2    generate-insn dst/src       %log2    ;
 

basis/compiler/tests/intrinsics.factor

@@ -1,11 +1,10 @@
-USING: accessors arrays compiler.units kernel kernel.private math
+USING: accessors arrays compiler.units kernel kernel.private
-math.constants math.private sequences strings tools.test words
+math math.constants math.private math.integers.private sequences
-continuations sequences.private hashtables.private byte-arrays
+strings tools.test words continuations sequences.private
-system random layouts vectors
+hashtables.private byte-arrays system random layouts vectors
 sbufs strings.private slots.private alien math.order
 alien.accessors alien.c-types alien.syntax alien.strings
-namespaces libc io.encodings.ascii
+namespaces libc io.encodings.ascii classes compiler ;
-classes compiler ;
 IN: compiler.tests.intrinsics
 
 ! Make sure that intrinsic ops compile to correct code.
@@ -271,6 +270,15 @@ cell 8 = [
     [ 100000 swap array-nth ] compile-call
 ] unit-test
 
+[ 2 ] [ 2 4 [ fixnum-min ] compile-call ] unit-test
+[ 2 ] [ 4 2 [ fixnum-min ] compile-call ] unit-test
+[ 4 ] [ 2 4 [ fixnum-max ] compile-call ] unit-test
+[ 4 ] [ 4 2 [ fixnum-max ] compile-call ] unit-test
+[ -2 ] [ -2 -4 [ fixnum-max ] compile-call ] unit-test
+[ -2 ] [ -4 -2 [ fixnum-max ] compile-call ] unit-test
+[ -4 ] [ -2 -4 [ fixnum-min ] compile-call ] unit-test
+[ -4 ] [ -4 -2 [ fixnum-min ] compile-call ] unit-test
+
 ! 64-bit overflow
 cell 8 = [
     [ t ] [ 1 59 fixnum-shift dup [ fixnum+ ] compile-call 1 60 fixnum-shift = ] unit-test

basis/compiler/tree/propagation/known-words/known-words.factor

@@ -79,11 +79,16 @@ IN: compiler.tree.propagation.known-words
     ] unless ;
 
 : ensure-math-class ( class must-be -- class' )
-    [ class<= ] 2keep ? ;
+    [ class<= ] most ;
 
 : number-valued ( class interval -- class' interval' )
     [ number ensure-math-class ] dip ;
 
+: fixnum-valued ( class interval -- class' interval' )
+    over null-class? [
+        [ drop fixnum ] dip
+    ] unless ;
+
 : integer-valued ( class interval -- class' interval' )
     [ integer ensure-math-class ] dip ;
 
@@ -304,7 +309,15 @@ flog fpow fsqrt facosh fasinh fatanh } [
     { float } "default-output-classes" set-word-prop
 ] each
 
-{ float-min float-max } [
+! Find a less repetitive way of doing this
-    [ { float float } "input-classes" set-word-prop ]
+\ float-min { float float } "input-classes" set-word-prop
-    [ { float } "default-output-classes" set-word-prop ] bi
+\ float-min [ interval-min ] [ float-valued ] binary-op
-] each
+
+\ float-max { float float } "input-classes" set-word-prop
+\ float-max [ interval-max ] [ float-valued ] binary-op
+
+\ fixnum-min { fixnum fixnum } "input-classes" set-word-prop
+\ fixnum-min [ interval-min ] [ fixnum-valued ] binary-op
+
+\ fixnum-max { fixnum fixnum } "input-classes" set-word-prop
+\ fixnum-max [ interval-max ] [ fixnum-valued ] binary-op

basis/compiler/tree/propagation/transforms/transforms.factor

@@ -1,11 +1,12 @@
 ! Copyright (C) 2008, 2009 Slava Pestov, Daniel Ehrenberg.
 ! See http://factorcode.org/license.txt for BSD license.
-USING: kernel sequences words fry generic accessors classes.tuple
+USING: kernel sequences words fry generic accessors
-classes classes.algebra definitions stack-checker.state quotations
+classes.tuple classes classes.algebra definitions
-classes.tuple.private math math.partial-dispatch math.private
+stack-checker.state quotations classes.tuple.private math
-math.intervals math.floats.private layouts math.order vectors hashtables
+math.partial-dispatch math.private math.intervals
-combinators effects generalizations assocs sets
+math.floats.private math.integers.private layouts math.order
-combinators.short-circuit sequences.private locals
+vectors hashtables combinators effects generalizations assocs
+sets combinators.short-circuit sequences.private locals
 stack-checker namespaces compiler.tree.propagation.info ;
 IN: compiler.tree.propagation.transforms
 
@@ -79,15 +80,25 @@ IN: compiler.tree.propagation.transforms
     ] [ f ] if
 ] "custom-inlining" set-word-prop
 
+! Integrate this with generic arithmetic optimization instead?
+: both-inputs? ( #call class -- ? )
+    [ in-d>> first2 ] dip '[ value-info class>> _ class<= ] both? ;
+
+\ min [
     {
-    { min [ float-min ] }
+        { [ dup fixnum both-inputs? ] [ [ fixnum-min ] ] }
-    { max [ float-max ] }
+        { [ dup float both-inputs? ] [ [ float-min ] ] }
-} [
+        [ f ]
-    '[
+    } cond nip
-        in-d>> first2 [ value-info class>> float class<= ] both?
+] "custom-inlining" set-word-prop
-        [ _ ] [ f ] if
+
+\ max [
+    {
+        { [ dup fixnum both-inputs? ] [ [ fixnum-max ] ] }
+        { [ dup float both-inputs? ] [ [ float-max ] ] }
+        [ f ]
+    } cond nip
 ] "custom-inlining" set-word-prop
-] assoc-each
 
 ! Generate more efficient code for common idiom
 \ clone [

basis/cpu/architecture/architecture.factor

@@ -96,6 +96,8 @@ HOOK: %shr     cpu ( dst src1 src2 -- )
 HOOK: %shr-imm cpu ( dst src1 src2 -- )
 HOOK: %sar     cpu ( dst src1 src2 -- )
 HOOK: %sar-imm cpu ( dst src1 src2 -- )
+HOOK: %min     cpu ( dst src1 src2 -- )
+HOOK: %max     cpu ( dst src1 src2 -- )
 HOOK: %not     cpu ( dst src -- )
 HOOK: %log2    cpu ( dst src -- )
 

basis/cpu/x86/x86.factor

@@ -123,6 +123,10 @@ M: x86 %xor-imm nip XOR ;
 M: x86 %shl-imm nip SHL ;
 M: x86 %shr-imm nip SHR ;
 M: x86 %sar-imm nip SAR ;
+
+M: x86 %min     nip [ CMP ] [ CMOVG ] 2bi ;
+M: x86 %max     nip [ CMP ] [ CMOVL ] 2bi ;
+
 M: x86 %not     drop NOT ;
 M: x86 %log2    BSR ;
 
@@ -579,3 +583,5 @@ M: x86 small-enough? ( n -- ? )
     enable-float-intrinsics
     enable-fsqrt
     enable-float-min/max ;
+
+enable-min/max

basis/math/intervals/intervals-tests.factor

@@ -235,6 +235,10 @@ IN: math.intervals.tests
     interval-contains?
 ] unit-test
 
+[ t ] [ full-interval 10 10 [a,b] interval-max 10 1/0. [a,b] = ] unit-test
+
+[ t ] [ full-interval 10 10 [a,b] interval-min -1/0. 10 [a,b] = ] unit-test
+
 [ t ] [ 1 100 [a,b] -1 1 [a,b] interval/i [-inf,inf] = ] unit-test
 
 ! Accuracy of interval-mod

basis/math/intervals/intervals.factor

@@ -7,7 +7,7 @@ IN: math.intervals
 
 SYMBOL: empty-interval
 
-SYMBOL: full-interval
+SINGLETON: full-interval
 
 TUPLE: interval { from read-only } { to read-only } ;
 
@@ -238,12 +238,24 @@ MEMO: array-capacity-interval ( -- interval )
     ] do-empty-interval ;
 
 : interval-max ( i1 i2 -- i3 )
-    #! Inaccurate; could be tighter
+    {
-    [ [ interval-closure ] bi@ [ max ] interval-op nan-not-ok ] do-empty-interval ;
+        { [ over empty-interval eq? ] [ drop ] }
+        { [ dup empty-interval eq? ] [ nip ] }
+        { [ 2dup [ full-interval eq? ] both? ] [ drop ] }
+        { [ over full-interval eq? ] [ nip from>> first [a,inf] ] }
+        { [ dup full-interval eq? ] [ drop from>> first [a,inf] ] }
+        [ [ interval-closure ] bi@ [ max ] interval-op nan-not-ok ]
+    } cond ;
 
 : interval-min ( i1 i2 -- i3 )
-    #! Inaccurate; could be tighter
+    {
-    [ [ interval-closure ] bi@ [ min ] interval-op nan-not-ok ] do-empty-interval ;
+        { [ over empty-interval eq? ] [ drop ] }
+        { [ dup empty-interval eq? ] [ nip ] }
+        { [ 2dup [ full-interval eq? ] both? ] [ drop ] }
+        { [ over full-interval eq? ] [ nip to>> first [-inf,a] ] }
+        { [ dup full-interval eq? ] [ drop to>> first [-inf,a] ] }
+        [ [ interval-closure ] bi@ [ min ] interval-op nan-not-ok ]
+    } cond ;
 
 : interval-interior ( i1 -- i2 )
     dup special-interval? [

core/math/floats/floats.factor

@@ -3,8 +3,8 @@
 USING: kernel math math.private ;
 IN: math.floats.private
 
-: float-min ( x y -- z ) [ float< ] 2keep ? ;
+: float-min ( x y -- z ) [ float< ] most ; foldable
-: float-max ( x y -- z ) [ float> ] 2keep ? ;
+: float-max ( x y -- z ) [ float> ] most ; foldable
 
 M: fixnum >float fixnum>float ; inline
 M: bignum >float bignum>float ; inline

core/math/integers/integers.factor

@@ -1,10 +1,13 @@
-! Copyright (C) 2004, 2008 Slava Pestov.
+! Copyright (C) 2004, 2009 Slava Pestov.
 ! Copyright (C) 2008, Doug Coleman.
 ! See http://factorcode.org/license.txt for BSD license.
 USING: kernel kernel.private sequences
 sequences.private math math.private combinators ;
 IN: math.integers.private
 
+: fixnum-min ( x y -- z ) [ fixnum< ] most ; foldable
+: fixnum-max ( x y -- z ) [ fixnum> ] most ; foldable
+
 M: integer numerator ; inline
 M: integer denominator drop 1 ; inline