X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcompiler%2Fx86-64%2Farith.lisp;h=922962632595cac1b7a0356b6575adcab7a2f7c8;hb=91a5cbf7375439309fede4776d8debc7a132dc20;hp=37cc2a2b9b905416ecd1f548eea5eda5fa801024;hpb=f7aa9f04c9d2124f094f4b9e857e8ce3b92035e4;p=sbcl.git diff --git a/src/compiler/x86-64/arith.lisp b/src/compiler/x86-64/arith.lisp index 37cc2a2..9229626 100644 --- a/src/compiler/x86-64/arith.lisp +++ b/src/compiler/x86-64/arith.lisp @@ -1,4 +1,4 @@ -;;;; the VM definition of arithmetic VOPs for the x86 +;;;; the VM definition of arithmetic VOPs for the x86-64 ;;;; This software is part of the SBCL system. See the README file for ;;;; more information. @@ -11,6 +11,26 @@ (in-package "SB!VM") + +;; If chopping X to 32 bits and sign-extending is equal to the original X, +;; return the chopped X, which the CPU will always treat as signed. +;; Notably this allows MOST-POSITIVE-WORD to be an immediate constant. +(defun immediate32-p (x) + (typecase x + ((signed-byte 32) x) + ((unsigned-byte 64) + (let ((chopped (sb!c::mask-signed-field 32 x))) + (and (= x (ldb (byte 64 0) chopped)) + chopped))) + (t nil))) + +;; If 'immediate32-p' is true, use it; otherwise use a RIP-relative constant. +;; I couldn't think of a more accurate name for this other than maybe +;; 'signed-immediate32-or-rip-relativize' which is just too awful. +(defun constantize (x) + (or (immediate32-p x) + (register-inline-constant :qword x))) + ;;;; unary operations (define-vop (fast-safe-arith-op) @@ -46,13 +66,13 @@ (define-vop (fast-lognot/fixnum fixnum-unop) (:translate lognot) - (:generator 2 + (:generator 1 (move res x) (inst xor res (fixnumize -1)))) (define-vop (fast-lognot/signed signed-unop) (:translate lognot) - (:generator 1 + (:generator 2 (move res x) (inst not res))) @@ -109,85 +129,110 @@ (:note "inline (signed-byte 64) arithmetic")) (define-vop (fast-fixnum-binop-c fast-safe-arith-op) - (:args (x :target r :scs (any-reg control-stack))) + (:args (x :target r :scs (any-reg) :load-if t)) (:info y) - (:arg-types tagged-num (:constant (signed-byte 29))) - (:results (r :scs (any-reg) - :load-if (not (location= x r)))) + (:arg-types tagged-num (:constant fixnum)) + (:results (r :scs (any-reg) :load-if t)) (:result-types tagged-num) (:note "inline fixnum arithmetic")) -;; 31 not 64 because it's hard work loading 64 bit constants, and since -;; sign-extension of immediates causes problems with 32. (define-vop (fast-unsigned-binop-c fast-safe-arith-op) - (:args (x :target r :scs (unsigned-reg unsigned-stack))) + (:args (x :target r :scs (unsigned-reg) :load-if t)) (:info y) - (:arg-types unsigned-num (:constant (unsigned-byte 31))) - (:results (r :scs (unsigned-reg) - :load-if (not (location= x r)))) + (:arg-types unsigned-num (:constant (unsigned-byte 64))) + (:results (r :scs (unsigned-reg) :load-if t)) (:result-types unsigned-num) (:note "inline (unsigned-byte 64) arithmetic")) (define-vop (fast-signed-binop-c fast-safe-arith-op) - (:args (x :target r :scs (signed-reg signed-stack))) + (:args (x :target r :scs (signed-reg) :load-if t)) (:info y) - (:arg-types signed-num (:constant (signed-byte 32))) - (:results (r :scs (signed-reg) - :load-if (not (location= x r)))) + (:arg-types signed-num (:constant (signed-byte 64))) + (:results (r :scs (signed-reg) :load-if t)) (:result-types signed-num) (:note "inline (signed-byte 64) arithmetic")) -(macrolet ((define-binop (translate untagged-penalty op) +(macrolet ((define-binop (translate untagged-penalty op + &key fixnum=>fixnum c/fixnum=>fixnum + signed=>signed c/signed=>signed + unsigned=>unsigned c/unsigned=>unsigned) + `(progn (define-vop (,(symbolicate "FAST-" translate "/FIXNUM=>FIXNUM") fast-fixnum-binop) (:translate ,translate) (:generator 2 - (move r x) - (inst ,op r y))) + ,@(or fixnum=>fixnum `((move r x) (inst ,op r y))))) (define-vop (,(symbolicate 'fast- translate '-c/fixnum=>fixnum) fast-fixnum-binop-c) (:translate ,translate) (:generator 1 - (move r x) - (inst ,op r (fixnumize y)))) + ,@(or c/fixnum=>fixnum + `((move r x) + (inst ,op r (constantize (fixnumize y))))))) (define-vop (,(symbolicate "FAST-" translate "/SIGNED=>SIGNED") fast-signed-binop) (:translate ,translate) (:generator ,(1+ untagged-penalty) - (move r x) - (inst ,op r y))) + ,@(or signed=>signed `((move r x) (inst ,op r y))))) (define-vop (,(symbolicate 'fast- translate '-c/signed=>signed) fast-signed-binop-c) (:translate ,translate) (:generator ,untagged-penalty - (move r x) - (inst ,op r y))) + ,@(or c/signed=>signed + `((move r x) (inst ,op r (constantize y)))))) (define-vop (,(symbolicate "FAST-" translate "/UNSIGNED=>UNSIGNED") - fast-unsigned-binop) + fast-unsigned-binop) (:translate ,translate) (:generator ,(1+ untagged-penalty) - (move r x) - (inst ,op r y))) + ,@(or unsigned=>unsigned `((move r x) (inst ,op r y))))) (define-vop (,(symbolicate 'fast- translate '-c/unsigned=>unsigned) fast-unsigned-binop-c) (:translate ,translate) (:generator ,untagged-penalty - (move r x) - (inst ,op r y)))))) + ,@(or c/unsigned=>unsigned + `((move r x) (inst ,op r (constantize y))))))))) ;;(define-binop + 4 add) (define-binop - 4 sub) - (define-binop logand 2 and) - (define-binop logior 2 or) - (define-binop logxor 2 xor)) + + ;; The following have microoptimizations for some special cases + ;; not caught by the front end. + + (define-binop logand 2 and + :c/unsigned=>unsigned + ((move r x) + (let ((y (constantize y))) + ;; ANDing with #xFFFF_FFFF_FFFF_FFFF is a no-op, other than + ;; the eflags state which we don't care about. + (unless (eql y -1) ; do nothing if this is true + (inst and r y))))) + + (define-binop logior 2 or + :c/unsigned=>unsigned + ((let ((y (constantize y))) + (cond ((and (register-p r) (eql y -1)) ; special-case "OR reg, all-ones" + ;; I have yet to elicit this case. Can it happen? + (inst mov r -1)) + (t + (move r x) + (inst or r y)))))) + + (define-binop logxor 2 xor + :c/unsigned=>unsigned + ((move r x) + (let ((y (constantize y))) + (if (eql y -1) ; special-case "XOR reg, [all-ones]" + (inst not r) + (inst xor r y)))))) ;;; Special handling of add on the x86; can use lea to avoid a ;;; register load, otherwise it uses add. +;;; FIXME: either inherit from fast-foo-binop or explain why not. (define-vop (fast-+/fixnum=>fixnum fast-safe-arith-op) (:translate +) (:args (x :scs (any-reg) :target r @@ -214,19 +259,20 @@ (define-vop (fast-+-c/fixnum=>fixnum fast-safe-arith-op) (:translate +) - (:args (x :target r :scs (any-reg control-stack))) + (:args (x :target r :scs (any-reg) :load-if t)) (:info y) - (:arg-types tagged-num (:constant (signed-byte 29))) - (:results (r :scs (any-reg) - :load-if (not (location= x r)))) + (:arg-types tagged-num (:constant fixnum)) + (:results (r :scs (any-reg) :load-if t)) (:result-types tagged-num) (:note "inline fixnum arithmetic") (:generator 1 - (cond ((and (sc-is x any-reg) (sc-is r any-reg) (not (location= x r))) - (inst lea r (make-ea :qword :base x :disp (fixnumize y)))) - (t - (move r x) - (inst add r (fixnumize y)))))) + (let ((y (fixnumize y))) + (cond ((and (not (location= x r)) + (typep y '(signed-byte 32))) + (inst lea r (make-ea :qword :base x :disp y))) + (t + (move r x) + (inst add r (constantize y))))))) (define-vop (fast-+/signed=>signed fast-safe-arith-op) (:translate +) @@ -251,7 +297,6 @@ (move r x) (inst add r y))))) - ;;;; Special logand cases: (logand signed unsigned) => unsigned (define-vop (fast-logand/signed-unsigned=>unsigned @@ -264,10 +309,13 @@ (y :scs (unsigned-reg unsigned-stack))) (:arg-types signed-num unsigned-num)) +;; This special case benefits from the special case for c/unsigned=>unsigned. +;; In particular, converting a (signed-byte 64) to (unsigned-byte 64) by +;; way of (LDB (byte 64 0)) doesn't need an AND instruction. (define-vop (fast-logand-c/signed-unsigned=>unsigned fast-logand-c/unsigned=>unsigned) - (:args (x :target r :scs (signed-reg signed-stack))) - (:arg-types signed-num (:constant (unsigned-byte 31)))) + (:args (x :target r :scs (signed-reg))) + (:arg-types signed-num (:constant (unsigned-byte 64)))) (define-vop (fast-logand/unsigned-signed=>unsigned fast-logand/unsigned=>unsigned) @@ -282,22 +330,27 @@ (define-vop (fast-+-c/signed=>signed fast-safe-arith-op) (:translate +) - (:args (x :target r :scs (signed-reg signed-stack))) + (:args (x :target r :scs (signed-reg) + :load-if (or (not (typep y '(signed-byte 32))) + (not (sc-is r signed-reg signed-stack))))) (:info y) - (:arg-types signed-num (:constant (signed-byte 32))) + (:arg-types signed-num (:constant (signed-byte 64))) (:results (r :scs (signed-reg) - :load-if (not (location= x r)))) + :load-if (or (not (location= x r)) + (not (typep y '(signed-byte 32)))))) (:result-types signed-num) (:note "inline (signed-byte 64) arithmetic") (:generator 4 (cond ((and (sc-is x signed-reg) (sc-is r signed-reg) - (not (location= x r))) + (not (location= x r)) + (typep y '(signed-byte 32))) (inst lea r (make-ea :qword :base x :disp y))) (t (move r x) - (if (= y 1) - (inst inc r) - (inst add r y)))))) + (cond ((= y 1) + (inst inc r)) + (t + (inst add r (constantize y)))))))) (define-vop (fast-+/unsigned=>unsigned fast-safe-arith-op) (:translate +) @@ -325,22 +378,27 @@ (define-vop (fast-+-c/unsigned=>unsigned fast-safe-arith-op) (:translate +) - (:args (x :target r :scs (unsigned-reg unsigned-stack))) + (:args (x :target r :scs (unsigned-reg) + :load-if (or (not (typep y '(unsigned-byte 31))) + (not (sc-is x unsigned-reg unsigned-stack))))) (:info y) - (:arg-types unsigned-num (:constant (unsigned-byte 31))) + (:arg-types unsigned-num (:constant (unsigned-byte 64))) (:results (r :scs (unsigned-reg) - :load-if (not (location= x r)))) + :load-if (or (not (location= x r)) + (not (typep y '(unsigned-byte 31)))))) (:result-types unsigned-num) (:note "inline (unsigned-byte 64) arithmetic") (:generator 4 (cond ((and (sc-is x unsigned-reg) (sc-is r unsigned-reg) - (not (location= x r))) + (not (location= x r)) + (typep y '(unsigned-byte 31))) (inst lea r (make-ea :qword :base x :disp y))) (t (move r x) - (if (= y 1) - (inst inc r) - (inst add r y)))))) + (cond ((= y 1) + (inst inc r)) + (t + (inst add r (constantize y)))))))) ;;;; multiplication and division @@ -355,20 +413,26 @@ (:note "inline fixnum arithmetic") (:generator 4 (move r x) - (inst sar r 3) + (inst sar r n-fixnum-tag-bits) (inst imul r y))) (define-vop (fast-*-c/fixnum=>fixnum fast-safe-arith-op) (:translate *) ;; We need different loading characteristics. - (:args (x :scs (any-reg control-stack))) + (:args (x :scs (any-reg) + :load-if (or (not (typep y '(signed-byte 32))) + (not (sc-is x any-reg control-stack))))) (:info y) - (:arg-types tagged-num (:constant (signed-byte 29))) + (:arg-types tagged-num (:constant fixnum)) (:results (r :scs (any-reg))) (:result-types tagged-num) (:note "inline fixnum arithmetic") (:generator 3 - (inst imul r x y))) + (cond ((typep y '(signed-byte 32)) + (inst imul r x y)) + (t + (move r x) + (inst imul r (register-inline-constant :qword y)))))) (define-vop (fast-*/signed=>signed fast-safe-arith-op) (:translate *) @@ -386,14 +450,20 @@ (define-vop (fast-*-c/signed=>signed fast-safe-arith-op) (:translate *) ;; We need different loading characteristics. - (:args (x :scs (signed-reg signed-stack))) + (:args (x :scs (signed-reg) + :load-if (or (not (typep y '(signed-byte 32))) + (not (sc-is x signed-reg signed-stack))))) (:info y) - (:arg-types signed-num (:constant (signed-byte 32))) + (:arg-types signed-num (:constant (signed-byte 64))) (:results (r :scs (signed-reg))) (:result-types signed-num) (:note "inline (signed-byte 64) arithmetic") (:generator 4 - (inst imul r x y))) + (cond ((typep y '(signed-byte 32)) + (inst imul r x y)) + (t + (move r x) + (inst imul r (register-inline-constant :qword y)))))) (define-vop (fast-*/unsigned=>unsigned fast-safe-arith-op) (:translate *) @@ -415,6 +485,26 @@ (inst mul eax y) (move r eax))) +(define-vop (fast-*-c/unsigned=>unsigned fast-safe-arith-op) + (:translate *) + (:args (x :scs (unsigned-reg) :target eax)) + (:info y) + (:arg-types unsigned-num (:constant (unsigned-byte 64))) + (:temporary (:sc unsigned-reg :offset eax-offset :target r + :from (:argument 0) :to :result) eax) + (:temporary (:sc unsigned-reg :offset edx-offset + :from :eval :to :result) edx) + (:ignore edx) + (:results (r :scs (unsigned-reg))) + (:result-types unsigned-num) + (:note "inline (unsigned-byte 64) arithmetic") + (:vop-var vop) + (:save-p :compute-only) + (:generator 6 + (move eax x) + (inst mul eax (register-inline-constant :qword y)) + (move r eax))) + (define-vop (fast-truncate/fixnum=>fixnum fast-safe-arith-op) (:translate truncate) @@ -432,7 +522,7 @@ (:vop-var vop) (:save-p :compute-only) (:generator 31 - (let ((zero (generate-error-code vop division-by-zero-error x y))) + (let ((zero (generate-error-code vop 'division-by-zero-error x y))) (if (sc-is y any-reg) (inst test y y) ; smaller instruction (inst cmp y 0)) @@ -441,15 +531,18 @@ (inst cqo) (inst idiv eax y) (if (location= quo eax) - (inst shl eax 3) - (inst lea quo (make-ea :qword :index eax :scale 8))) + (inst shl eax n-fixnum-tag-bits) + (if (= n-fixnum-tag-bits 1) + (inst lea quo (make-ea :qword :base eax :index eax)) + (inst lea quo (make-ea :qword :index eax + :scale (ash 1 n-fixnum-tag-bits))))) (move rem edx))) (define-vop (fast-truncate-c/fixnum=>fixnum fast-safe-arith-op) (:translate truncate) (:args (x :scs (any-reg) :target eax)) (:info y) - (:arg-types tagged-num (:constant (signed-byte 29))) + (:arg-types tagged-num (:constant fixnum)) (:temporary (:sc signed-reg :offset eax-offset :target quo :from :argument :to (:result 0)) eax) (:temporary (:sc any-reg :offset edx-offset :target rem @@ -467,8 +560,11 @@ (inst mov y-arg (fixnumize y)) (inst idiv eax y-arg) (if (location= quo eax) - (inst shl eax 3) - (inst lea quo (make-ea :qword :index eax :scale 8))) + (inst shl eax n-fixnum-tag-bits) + (if (= n-fixnum-tag-bits 1) + (inst lea quo (make-ea :qword :base eax :index eax)) + (inst lea quo (make-ea :qword :index eax + :scale (ash 1 n-fixnum-tag-bits))))) (move rem edx))) (define-vop (fast-truncate/unsigned=>unsigned fast-safe-arith-op) @@ -487,7 +583,7 @@ (:vop-var vop) (:save-p :compute-only) (:generator 33 - (let ((zero (generate-error-code vop division-by-zero-error x y))) + (let ((zero (generate-error-code vop 'division-by-zero-error x y))) (if (sc-is y unsigned-reg) (inst test y y) ; smaller instruction (inst cmp y 0)) @@ -502,7 +598,7 @@ (:translate truncate) (:args (x :scs (unsigned-reg) :target eax)) (:info y) - (:arg-types unsigned-num (:constant (unsigned-byte 31))) + (:arg-types unsigned-num (:constant (unsigned-byte 64))) (:temporary (:sc unsigned-reg :offset eax-offset :target quo :from :argument :to (:result 0)) eax) (:temporary (:sc unsigned-reg :offset edx-offset :target rem @@ -538,7 +634,7 @@ (:vop-var vop) (:save-p :compute-only) (:generator 33 - (let ((zero (generate-error-code vop division-by-zero-error x y))) + (let ((zero (generate-error-code vop 'division-by-zero-error x y))) (if (sc-is y signed-reg) (inst test y y) ; smaller instruction (inst cmp y 0)) @@ -553,7 +649,7 @@ (:translate truncate) (:args (x :scs (signed-reg) :target eax)) (:info y) - (:arg-types signed-num (:constant (signed-byte 32))) + (:arg-types signed-num (:constant (signed-byte 64))) (:temporary (:sc signed-reg :offset eax-offset :target quo :from :argument :to (:result 0)) eax) (:temporary (:sc signed-reg :offset edx-offset :target rem @@ -591,6 +687,8 @@ (location= number result))))) (:result-types tagged-num) (:note "inline ASH") + (:variant nil) + (:variant-vars modularp) (:generator 2 (cond ((and (= amount 1) (not (location= number result))) (inst lea result (make-ea :qword :base number :index number))) @@ -600,19 +698,25 @@ (inst lea result (make-ea :qword :index number :scale 8))) (t (move result number) - (cond ((plusp amount) - ;; We don't have to worry about overflow because of the - ;; result type restriction. - (inst shl result amount)) - (t - ;; Since the shift instructions take the shift amount - ;; modulo 64 we must special case amounts of 64 and more. - ;; Because fixnums have only 61 bits, the result is 0 or - ;; -1 for all amounts of 60 or more, so use this as the - ;; limit instead. - (inst sar result (min (- n-word-bits n-fixnum-tag-bits 1) - (- amount))) - (inst and result (lognot fixnum-tag-mask)))))))) + (cond ((< -64 amount 64) + ;; this code is used both in ASH and ASH-MODFX, so + ;; be careful + (if (plusp amount) + (inst shl result amount) + (progn + (inst sar result (- amount)) + (inst and result (lognot fixnum-tag-mask))))) + ;; shifting left (zero fill) + ((plusp amount) + (unless modularp + (aver (not "Impossible: fixnum ASH should not be called with +constant shift greater than word length"))) + (if (sc-is result any-reg) + (zeroize result) + (inst mov result 0))) + ;; shifting right (sign fill) + (t (inst sar result 63) + (inst and result (lognot fixnum-tag-mask)))))))) (define-vop (fast-ash-left/fixnum=>fixnum) (:translate ash) @@ -752,7 +856,7 @@ (:generator 5 (move result number) (move ecx amount) - (inst or ecx ecx) + (inst test ecx ecx) (inst jmp :ns POSITIVE) (inst neg ecx) (inst cmp ecx 63) @@ -781,7 +885,7 @@ (:generator 5 (move result number) (move ecx amount) - (inst or ecx ecx) + (inst test ecx ecx) (inst jmp :ns POSITIVE) (inst neg ecx) (inst cmp ecx 63) @@ -798,6 +902,52 @@ DONE)) +#!+ash-right-vops +(define-vop (fast-%ash/right/unsigned) + (:translate %ash/right) + (:policy :fast-safe) + (:args (number :scs (unsigned-reg) :target result) + (amount :scs (unsigned-reg) :target rcx)) + (:arg-types unsigned-num unsigned-num) + (:results (result :scs (unsigned-reg) :from (:argument 0))) + (:result-types unsigned-num) + (:temporary (:sc signed-reg :offset rcx-offset :from (:argument 1)) rcx) + (:generator 4 + (move result number) + (move rcx amount) + (inst shr result :cl))) + +#!+ash-right-vops +(define-vop (fast-%ash/right/signed) + (:translate %ash/right) + (:policy :fast-safe) + (:args (number :scs (signed-reg) :target result) + (amount :scs (unsigned-reg) :target rcx)) + (:arg-types signed-num unsigned-num) + (:results (result :scs (signed-reg) :from (:argument 0))) + (:result-types signed-num) + (:temporary (:sc signed-reg :offset rcx-offset :from (:argument 1)) rcx) + (:generator 4 + (move result number) + (move rcx amount) + (inst sar result :cl))) + +#!+ash-right-vops +(define-vop (fast-%ash/right/fixnum) + (:translate %ash/right) + (:policy :fast-safe) + (:args (number :scs (any-reg) :target result) + (amount :scs (unsigned-reg) :target rcx)) + (:arg-types tagged-num unsigned-num) + (:results (result :scs (any-reg) :from (:argument 0))) + (:result-types tagged-num) + (:temporary (:sc signed-reg :offset rcx-offset :from (:argument 1)) rcx) + (:generator 3 + (move result number) + (move rcx amount) + (inst sar result :cl) + (inst and result (lognot fixnum-tag-mask)))) + (in-package "SB!C") (defknown %lea (integer integer (member 1 2 4 8 16) (signed-byte 64)) @@ -893,7 +1043,7 @@ (:generator 4 (move result number) (move ecx amount) - (inst or ecx ecx) + (inst test ecx ecx) (inst jmp :ns POSITIVE) (inst neg ecx) (zeroize zero) @@ -918,7 +1068,7 @@ (:result-types unsigned-num) (:generator 28 (move res arg) - (inst cmp res 0) + (inst test res res) (inst jmp :ge POS) (inst not res) POS @@ -947,6 +1097,52 @@ (zeroize res) DONE)) +;; INTEGER-LENGTH is implemented by using the BSR instruction, which +;; returns the position of the first 1-bit from the right. And that needs +;; to be incremented to get the width of the integer, and BSR doesn't +;; work on 0, so it needs a branch to handle 0. + +;; But fixnums are tagged by being shifted left n-fixnum-tag-bits times, +;; untagging by shifting right n-fixnum-tag-bits-1 times (and if +;; n-fixnum-tag-bits = 1, no shifting is required), will make the +;; resulting integer one bit wider, making the increment unnecessary. +;; Then, to avoid calling BSR on 0, OR the result with 1. That sets the +;; first bit to 1, and if all other bits are 0, BSR will return 0, +;; which is the correct value for INTEGER-LENGTH. +(define-vop (positive-fixnum-len) + (:translate integer-length) + (:note "inline positive fixnum integer-length") + (:policy :fast-safe) + (:args (arg :scs (any-reg))) + (:arg-types positive-fixnum) + (:results (res :scs (unsigned-reg))) + (:result-types unsigned-num) + (:generator 24 + (move res arg) + (when (> n-fixnum-tag-bits 1) + (inst shr res (1- n-fixnum-tag-bits))) + (inst or res 1) + (inst bsr res res))) + +(define-vop (fixnum-len) + (:translate integer-length) + (:note "inline fixnum integer-length") + (:policy :fast-safe) + (:args (arg :scs (any-reg) :target res)) + (:arg-types tagged-num) + (:results (res :scs (unsigned-reg))) + (:result-types unsigned-num) + (:generator 25 + (move res arg) + (when (> n-fixnum-tag-bits 1) + (inst sar res (1- n-fixnum-tag-bits))) + (inst test res res) + (inst jmp :ge POS) + (inst not res) + POS + (inst or res 1) + (inst bsr res res))) + (define-vop (unsigned-byte-64-count) (:translate logcount) (:note "inline (unsigned-byte 64) logcount") @@ -1005,15 +1201,12 @@ ;;;; binary conditional VOPs (define-vop (fast-conditional) - (:conditional) - (:info target not-p) + (:conditional :e) + (:info) (:effects) (:affected) (:policy :fast-safe)) -;;; constant variants are declared for 32 bits not 64 bits, because -;;; loading a 64 bit constant is silly - (define-vop (fast-conditional/fixnum fast-conditional) (:args (x :scs (any-reg) :load-if (not (and (sc-is x control-stack) @@ -1023,9 +1216,9 @@ (:note "inline fixnum comparison")) (define-vop (fast-conditional-c/fixnum fast-conditional/fixnum) - (:args (x :scs (any-reg control-stack))) - (:arg-types tagged-num (:constant (signed-byte 29))) - (:info target not-p y)) + (:args (x :scs (any-reg) :load-if t)) + (:arg-types tagged-num (:constant fixnum)) + (:info y)) (define-vop (fast-conditional/signed fast-conditional) (:args (x :scs (signed-reg) @@ -1036,9 +1229,9 @@ (:note "inline (signed-byte 64) comparison")) (define-vop (fast-conditional-c/signed fast-conditional/signed) - (:args (x :scs (signed-reg signed-stack))) - (:arg-types signed-num (:constant (signed-byte 31))) - (:info target not-p y)) + (:args (x :scs (signed-reg) :load-if t)) + (:arg-types signed-num (:constant (signed-byte 64))) + (:info y)) (define-vop (fast-conditional/unsigned fast-conditional) (:args (x :scs (unsigned-reg) @@ -1049,9 +1242,77 @@ (:note "inline (unsigned-byte 64) comparison")) (define-vop (fast-conditional-c/unsigned fast-conditional/unsigned) - (:args (x :scs (unsigned-reg unsigned-stack))) - (:arg-types unsigned-num (:constant (unsigned-byte 31))) - (:info target not-p y)) + (:args (x :scs (unsigned-reg) :load-if t)) + (:arg-types unsigned-num (:constant (unsigned-byte 64))) + (:info y)) + +;; Stolen liberally from the x86 32-bit implementation. +(macrolet ((define-logtest-vops () + `(progn + ,@(loop for suffix in '(/fixnum -c/fixnum + /signed -c/signed + /unsigned -c/unsigned) + for cost in '(4 3 6 5 6 5) + collect + `(define-vop (,(symbolicate "FAST-LOGTEST" suffix) + ,(symbolicate "FAST-CONDITIONAL" suffix)) + (:translate logtest) + (:conditional :ne) + (:generator ,cost + (emit-optimized-test-inst x + ,(case suffix + (-c/fixnum + `(constantize (fixnumize y))) + ((-c/signed -c/unsigned) + `(constantize y)) + (t + 'y))))))))) + (define-logtest-vops)) + +(defknown %logbitp (integer unsigned-byte) boolean + (movable foldable flushable always-translatable)) + +;;; only for constant folding within the compiler +(defun %logbitp (integer index) + (logbitp index integer)) + +;;; too much work to do the non-constant case (maybe?) +(define-vop (fast-logbitp-c/fixnum fast-conditional-c/fixnum) + (:translate %logbitp) + (:conditional :c) + (:arg-types tagged-num (:constant (integer 0 #.(- 63 n-fixnum-tag-bits)))) + (:generator 4 + (inst bt x (+ y n-fixnum-tag-bits)))) + +(define-vop (fast-logbitp/signed fast-conditional/signed) + (:args (x :scs (signed-reg signed-stack)) + (y :scs (signed-reg))) + (:translate %logbitp) + (:conditional :c) + (:generator 6 + (inst bt x y))) + +(define-vop (fast-logbitp-c/signed fast-conditional-c/signed) + (:translate %logbitp) + (:conditional :c) + (:arg-types signed-num (:constant (integer 0 63))) + (:generator 5 + (inst bt x y))) + +(define-vop (fast-logbitp/unsigned fast-conditional/unsigned) + (:args (x :scs (unsigned-reg unsigned-stack)) + (y :scs (unsigned-reg))) + (:translate %logbitp) + (:conditional :c) + (:generator 6 + (inst bt x y))) + +(define-vop (fast-logbitp-c/unsigned fast-conditional-c/unsigned) + (:translate %logbitp) + (:conditional :c) + (:arg-types unsigned-num (:constant (integer 0 63))) + (:generator 5 + (inst bt x y))) (macrolet ((define-conditional-vop (tran cond unsigned not-cond not-unsigned) `(progn @@ -1065,19 +1326,15 @@ (format nil "~:@(FAST-CONDITIONAL~A~)" suffix))) (:translate ,tran) + (:conditional ,(if signed cond unsigned)) (:generator ,cost (inst cmp x - ,(if (eq suffix '-c/fixnum) - '(fixnumize y) - 'y)) - (inst jmp (if not-p - ,(if signed - not-cond - not-unsigned) - ,(if signed - cond - unsigned)) - target)))) + ,(case suffix + (-c/fixnum + `(constantize (fixnumize y))) + ((-c/signed -c/unsigned) + `(constantize y)) + (t 'y)))))) '(/fixnum -c/fixnum /signed -c/signed /unsigned -c/unsigned) ; '(/fixnum /signed /unsigned) '(4 3 6 5 6 5) @@ -1089,8 +1346,7 @@ (define-vop (fast-if-eql/signed fast-conditional/signed) (:translate eql) (:generator 6 - (inst cmp x y) - (inst jmp (if not-p :ne :e) target))) + (inst cmp x y))) (define-vop (fast-if-eql-c/signed fast-conditional-c/signed) (:translate eql) @@ -1098,14 +1354,12 @@ (cond ((and (sc-is x signed-reg) (zerop y)) (inst test x x)) ; smaller instruction (t - (inst cmp x y))) - (inst jmp (if not-p :ne :e) target))) + (inst cmp x (constantize y)))))) (define-vop (fast-if-eql/unsigned fast-conditional/unsigned) (:translate eql) (:generator 6 - (inst cmp x y) - (inst jmp (if not-p :ne :e) target))) + (inst cmp x y))) (define-vop (fast-if-eql-c/unsigned fast-conditional-c/unsigned) (:translate eql) @@ -1113,8 +1367,7 @@ (cond ((and (sc-is x unsigned-reg) (zerop y)) (inst test x x)) ; smaller instruction (t - (inst cmp x y))) - (inst jmp (if not-p :ne :e) target))) + (inst cmp x (constantize y)))))) ;;; EQL/FIXNUM is funny because the first arg can be of any type, not just a ;;; known fixnum. @@ -1134,8 +1387,8 @@ (:note "inline fixnum comparison") (:translate eql) (:generator 4 - (inst cmp x y) - (inst jmp (if not-p :ne :e) target))) + (inst cmp x y))) + (define-vop (generic-eql/fixnum fast-eql/fixnum) (:args (x :scs (any-reg descriptor-reg) :load-if (not (and (sc-is x control-stack) @@ -1144,41 +1397,26 @@ (:arg-types * tagged-num) (:variant-cost 7)) - -(define-vop (fast-eql-c/fixnum fast-conditional/fixnum) - (:args (x :scs (any-reg control-stack))) - (:arg-types tagged-num (:constant (signed-byte 29))) - (:info target not-p y) +(define-vop (fast-eql-c/fixnum fast-conditional-c/fixnum) + (:args (x :scs (any-reg) :load-if t)) + (:arg-types tagged-num (:constant fixnum)) + (:info y) + (:conditional :e) + (:policy :fast-safe) (:translate eql) (:generator 2 - (cond ((and (sc-is x any-reg) (zerop y)) + (cond ((and (sc-is x any-reg descriptor-reg) (zerop y)) (inst test x x)) ; smaller instruction (t - (inst cmp x (fixnumize y)))) - (inst jmp (if not-p :ne :e) target))) + (inst cmp x (constantize (fixnumize y))))))) (define-vop (generic-eql-c/fixnum fast-eql-c/fixnum) - (:args (x :scs (any-reg descriptor-reg control-stack))) - (:arg-types * (:constant (signed-byte 29))) + (:args (x :scs (any-reg descriptor-reg) :load-if t)) + (:arg-types * (:constant fixnum)) (:variant-cost 6)) ;;;; 32-bit logical operations -(define-vop (merge-bits) - (:translate merge-bits) - (:args (shift :scs (signed-reg unsigned-reg) :target ecx) - (prev :scs (unsigned-reg) :target result) - (next :scs (unsigned-reg))) - (:arg-types tagged-num unsigned-num unsigned-num) - (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 0)) ecx) - (:results (result :scs (unsigned-reg) :from (:argument 1))) - (:result-types unsigned-num) - (:policy :fast-safe) - (:generator 4 - (move ecx shift) - (move result prev) - (inst shrd result next :cl))) - ;;; Only the lower 6 bits of the shift amount are significant. (define-vop (shift-towards-someplace) (:policy :fast-safe) @@ -1232,19 +1470,11 @@ (defmacro define-mod-binop-c ((name prototype) function) `(define-vop (,name ,prototype) (:args (x :target r :scs (unsigned-reg signed-reg) - :load-if (not (and (or (sc-is x unsigned-stack) - (sc-is x signed-stack)) - (or (sc-is r unsigned-stack) - (sc-is r signed-stack)) - (location= x r))))) + :load-if t)) (:info y) (:arg-types untagged-num (:constant (or (unsigned-byte 64) (signed-byte 64)))) (:results (r :scs (unsigned-reg signed-reg) :from (:argument 0) - :load-if (not (and (or (sc-is x unsigned-stack) - (sc-is x signed-stack)) - (or (sc-is r unsigned-stack) - (sc-is r unsigned-stack)) - (location= x r))))) + :load-if t)) (:result-types unsigned-num) (:translate ,function))) @@ -1258,22 +1488,22 @@ (vop64f (intern (format nil "FAST-~S-MOD64/FIXNUM=>FIXNUM" name))) (vop64cu (intern (format nil "FAST-~S-MOD64-C/WORD=>UNSIGNED" name))) (vop64cf (intern (format nil "FAST-~S-MOD64-C/FIXNUM=>FIXNUM" name))) - (sfun61 (intern (format nil "~S-SMOD61" name))) - (svop61f (intern (format nil "FAST-~S-SMOD61/FIXNUM=>FIXNUM" name))) - (svop61cf (intern (format nil "FAST-~S-SMOD61-C/FIXNUM=>FIXNUM" name)))) + (funfx (intern (format nil "~S-MODFX" name))) + (vopfxf (intern (format nil "FAST-~S-MODFX/FIXNUM=>FIXNUM" name))) + (vopfxcf (intern (format nil "FAST-~S-MODFX-C/FIXNUM=>FIXNUM" name)))) `(progn - (define-modular-fun ,fun64 (x y) ,name :unsigned 64) - (define-modular-fun ,sfun61 (x y) ,name :signed 61) + (define-modular-fun ,fun64 (x y) ,name :untagged nil 64) + (define-modular-fun ,funfx (x y) ,name :tagged t + #.(- n-word-bits n-fixnum-tag-bits)) (define-mod-binop (,vop64u ,vopu) ,fun64) (define-vop (,vop64f ,vopf) (:translate ,fun64)) - (define-vop (,svop61f ,vopf) (:translate ,sfun61)) + (define-vop (,vopfxf ,vopf) (:translate ,funfx)) ,@(when -c-p `((define-mod-binop-c (,vop64cu ,vopcu) ,fun64) - (define-vop (,svop61cf ,vopcf) (:translate ,sfun61)))))))) + (define-vop (,vopfxcf ,vopcf) (:translate ,funfx)))))))) (def + t) (def - t) - ;; (no -C variant as x86 MUL instruction doesn't take an immediate) - (def * nil)) + (def * t)) (define-vop (fast-ash-left-mod64-c/unsigned=>unsigned fast-ash-c/unsigned=>unsigned) @@ -1286,47 +1516,49 @@ (sb!c::give-up-ir1-transform)) '(%primitive fast-ash-left-mod64/unsigned=>unsigned integer count)) -(define-vop (fast-ash-left-smod61-c/fixnum=>fixnum +(define-vop (fast-ash-left-modfx-c/fixnum=>fixnum fast-ash-c/fixnum=>fixnum) - (:translate ash-left-smod61)) -(define-vop (fast-ash-left-smod61/fixnum=>fixnum + (:variant :modular) + (:translate ash-left-modfx)) +(define-vop (fast-ash-left-modfx/fixnum=>fixnum fast-ash-left/fixnum=>fixnum)) -(deftransform ash-left-smod61 ((integer count) - ((signed-byte 61) (unsigned-byte 6))) +(deftransform ash-left-modfx ((integer count) + (fixnum (unsigned-byte 6))) (when (sb!c::constant-lvar-p count) (sb!c::give-up-ir1-transform)) - '(%primitive fast-ash-left-smod61/fixnum=>fixnum integer count)) + '(%primitive fast-ash-left-modfx/fixnum=>fixnum integer count)) (in-package "SB!C") (defknown sb!vm::%lea-mod64 (integer integer (member 1 2 4 8) (signed-byte 64)) (unsigned-byte 64) (foldable flushable movable)) -(defknown sb!vm::%lea-smod61 (integer integer (member 1 2 4 8) (signed-byte 64)) - (signed-byte 61) +(defknown sb!vm::%lea-modfx (integer integer (member 1 2 4 8) (signed-byte 64)) + fixnum (foldable flushable movable)) -(define-modular-fun-optimizer %lea ((base index scale disp) :unsigned :width width) +(define-modular-fun-optimizer %lea ((base index scale disp) :untagged nil :width width) (when (and (<= width 64) (constant-lvar-p scale) (constant-lvar-p disp)) - (cut-to-width base :unsigned width) - (cut-to-width index :unsigned width) + (cut-to-width base :untagged width nil) + (cut-to-width index :untagged width nil) 'sb!vm::%lea-mod64)) -(define-modular-fun-optimizer %lea ((base index scale disp) :signed :width width) - (when (and (<= width 61) +(define-modular-fun-optimizer %lea ((base index scale disp) :tagged t :width width) + (when (and (<= width (- sb!vm:n-word-bits sb!vm:n-fixnum-tag-bits)) (constant-lvar-p scale) (constant-lvar-p disp)) - (cut-to-width base :signed width) - (cut-to-width index :signed width) - 'sb!vm::%lea-smod61)) + (cut-to-width base :tagged width t) + (cut-to-width index :tagged width t) + 'sb!vm::%lea-modfx)) #+sb-xc-host (progn (defun sb!vm::%lea-mod64 (base index scale disp) (ldb (byte 64 0) (%lea base index scale disp))) - (defun sb!vm::%lea-smod61 (base index scale disp) - (mask-signed-field 61 (%lea base index scale disp)))) + (defun sb!vm::%lea-modfx (base index scale disp) + (mask-signed-field (- sb!vm:n-word-bits sb!vm:n-fixnum-tag-bits) + (%lea base index scale disp)))) #-sb-xc-host (progn (defun sb!vm::%lea-mod64 (base index scale disp) @@ -1335,24 +1567,25 @@ ;; can't use modular version of %LEA, as we only have VOPs for ;; constant SCALE and DISP. (ldb (byte 64 0) (+ base (* index scale) disp)))) - (defun sb!vm::%lea-smod61 (base index scale disp) - (let ((base (mask-signed-field 61 base)) - (index (mask-signed-field 61 index))) + (defun sb!vm::%lea-modfx (base index scale disp) + (let* ((fixnum-width (- sb!vm:n-word-bits sb!vm:n-fixnum-tag-bits)) + (base (mask-signed-field fixnum-width base)) + (index (mask-signed-field fixnum-width index))) ;; can't use modular version of %LEA, as we only have VOPs for ;; constant SCALE and DISP. - (mask-signed-field 61 (+ base (* index scale) disp))))) + (mask-signed-field fixnum-width (+ base (* index scale) disp))))) (in-package "SB!VM") (define-vop (%lea-mod64/unsigned=>unsigned %lea/unsigned=>unsigned) (:translate %lea-mod64)) -(define-vop (%lea-smod61/fixnum=>fixnum +(define-vop (%lea-modfx/fixnum=>fixnum %lea/fixnum=>fixnum) - (:translate %lea-smod61)) + (:translate %lea-modfx)) ;;; logical operations -(define-modular-fun lognot-mod64 (x) lognot :unsigned 64) +(define-modular-fun lognot-mod64 (x) lognot :untagged nil 64) (define-vop (lognot-mod64/unsigned=>unsigned) (:translate lognot-mod64) (:args (x :scs (unsigned-reg unsigned-stack) :target r @@ -1370,20 +1603,6 @@ (move r x) (inst not r))) -(define-modular-fun logxor-mod64 (x y) logxor :unsigned 64) -(define-mod-binop (fast-logxor-mod64/word=>unsigned - fast-logxor/unsigned=>unsigned) - logxor-mod64) -(define-mod-binop-c (fast-logxor-mod64-c/word=>unsigned - fast-logxor-c/unsigned=>unsigned) - logxor-mod64) -(define-vop (fast-logxor-mod64/fixnum=>fixnum - fast-logxor/fixnum=>fixnum) - (:translate logxor-mod64)) -(define-vop (fast-logxor-mod64-c/fixnum=>fixnum - fast-logxor-c/fixnum=>fixnum) - (:translate logxor-mod64)) - (define-source-transform logeqv (&rest args) (if (oddp (length args)) `(logxor ,@args) @@ -1413,7 +1632,9 @@ (define-full-reffer bignum-ref * bignum-digits-offset other-pointer-lowtag (unsigned-reg) unsigned-num sb!bignum:%bignum-ref) - +(define-full-reffer+offset bignum--ref-with-offset * bignum-digits-offset + other-pointer-lowtag (unsigned-reg) unsigned-num + sb!bignum:%bignum-ref-with-offset) (define-full-setter bignum-set * bignum-digits-offset other-pointer-lowtag (unsigned-reg) unsigned-num sb!bignum:%bignum-set) @@ -1422,11 +1643,9 @@ (:policy :fast-safe) (:args (digit :scs (unsigned-reg))) (:arg-types unsigned-num) - (:conditional) - (:info target not-p) + (:conditional :ns) (:generator 3 - (inst or digit digit) - (inst jmp (if not-p :s :ns) target))) + (inst test digit digit))) ;;; For add and sub with carry the sc of carry argument is any-reg so @@ -1539,6 +1758,42 @@ (move hi edx) (move lo eax))) +#!+multiply-high-vops +(define-vop (mulhi) + (:translate sb!kernel:%multiply-high) + (:policy :fast-safe) + (:args (x :scs (unsigned-reg) :target eax) + (y :scs (unsigned-reg unsigned-stack))) + (:arg-types unsigned-num unsigned-num) + (:temporary (:sc unsigned-reg :offset eax-offset :from (:argument 0)) + eax) + (:temporary (:sc unsigned-reg :offset edx-offset :from (:argument 1) + :to (:result 0) :target hi) edx) + (:results (hi :scs (unsigned-reg))) + (:result-types unsigned-num) + (:generator 20 + (move eax x) + (inst mul eax y) + (move hi edx))) + +#!+multiply-high-vops +(define-vop (mulhi/fx) + (:translate sb!kernel:%multiply-high) + (:policy :fast-safe) + (:args (x :scs (any-reg) :target eax) + (y :scs (unsigned-reg unsigned-stack))) + (:arg-types positive-fixnum unsigned-num) + (:temporary (:sc any-reg :offset eax-offset :from (:argument 0)) eax) + (:temporary (:sc any-reg :offset edx-offset :from (:argument 1) + :to (:result 0) :target hi) edx) + (:results (hi :scs (any-reg))) + (:result-types positive-fixnum) + (:generator 15 + (move eax x) + (inst mul eax y) + (move hi edx) + (inst and hi (lognot fixnum-tag-mask)))) + (define-vop (bignum-lognot lognot-mod64/unsigned=>unsigned) (:translate sb!bignum:%lognot)) @@ -1554,10 +1809,10 @@ (:result-types unsigned-num) (:generator 1 (move digit fixnum) - (inst sar digit 3))) + (inst sar digit n-fixnum-tag-bits))) (define-vop (bignum-floor) - (:translate sb!bignum:%floor) + (:translate sb!bignum:%bigfloor) (:policy :fast-safe) (:args (div-high :scs (unsigned-reg) :target edx) (div-low :scs (unsigned-reg) :target eax) @@ -1590,7 +1845,7 @@ (:generator 1 (move res digit) (when (sc-is res any-reg control-stack) - (inst shl res 3)))) + (inst shl res n-fixnum-tag-bits)))) (define-vop (digit-ashr) (:translate sb!bignum:%ashr) @@ -1635,6 +1890,77 @@ (move result digit) (move ecx count) (inst shl result :cl))) + +(define-vop (logand-bignum/c) + (:translate logand) + (:policy :fast-safe) + (:args (x :scs (descriptor-reg))) + (:arg-types bignum (:constant word)) + (:results (r :scs (unsigned-reg))) + (:info mask) + (:result-types unsigned-num) + (:generator 4 + (let ((mask (constantize mask))) + (cond ((or (integerp mask) + (location= x r)) + (loadw r x bignum-digits-offset other-pointer-lowtag) + (unless (eql mask -1) + (inst and r mask))) + (t + (inst mov r mask) + (inst and r (make-ea-for-object-slot x + bignum-digits-offset + other-pointer-lowtag))))))) + +;; Specialised mask-signed-field VOPs. +(define-vop (mask-signed-field-word/c) + (:translate sb!c::mask-signed-field) + (:policy :fast-safe) + (:args (x :scs (signed-reg unsigned-reg) :target r)) + (:arg-types (:constant (integer 0 64)) untagged-num) + (:results (r :scs (signed-reg))) + (:result-types signed-num) + (:info width) + (:generator 3 + (cond ((zerop width) + (zeroize r)) + ((= width 64) + (move r x)) + ((member width '(32 16 8)) + (inst movsx r (reg-in-size x (ecase width + (32 :dword) + (16 :word) + (8 :byte))))) + (t + (move r x) + (let ((delta (- n-word-bits width))) + (inst shl r delta) + (inst sar r delta)))))) + +(define-vop (mask-signed-field-bignum/c) + (:translate sb!c::mask-signed-field) + (:policy :fast-safe) + (:args (x :scs (descriptor-reg) :target r)) + (:arg-types (:constant (integer 0 64)) bignum) + (:results (r :scs (signed-reg))) + (:result-types signed-num) + (:info width) + (:generator 4 + (cond ((zerop width) + (zeroize r)) + ((member width '(8 16 32 64)) + (ecase width + (64 (loadw r x bignum-digits-offset other-pointer-lowtag)) + ((32 16 8) + (inst movsx r (make-ea (ecase width (32 :dword) (16 :word) (8 :byte)) + :base x + :disp (- (* bignum-digits-offset n-word-bytes) + other-pointer-lowtag)))))) + (t + (loadw r x bignum-digits-offset other-pointer-lowtag) + (let ((delta (- n-word-bits width))) + (inst shl r delta) + (inst sar r delta)))))) ;;;; static functions @@ -1679,14 +2005,14 @@ (*-transformer y))) (deftransform * ((x y) - ((signed-byte 61) (constant-arg (unsigned-byte 64))) - (signed-byte 61)) + (fixnum (constant-arg (unsigned-byte 64))) + fixnum) "recode as leas, shifts and adds" (let ((y (lvar-value y))) (*-transformer y))) -(deftransform sb!vm::*-smod61 - ((x y) ((signed-byte 61) (constant-arg (unsigned-byte 64))) - (signed-byte 61)) +(deftransform sb!vm::*-modfx + ((x y) (fixnum (constant-arg (unsigned-byte 64))) + fixnum) "recode as leas, shifts and adds" (let ((y (lvar-value y))) (*-transformer y)))