X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcompiler%2Fx86%2Farith.lisp;h=ade76894452a3f916a0ae8dd25f7fc6d008b2467;hb=bffa99d35c7d50ac46b9eb7dbe25d1ab1a0e6145;hp=36b869d71e2bcfd0a114f4c51ad413c651c3ca52;hpb=d131dfb49a3e6522d2358d14252f3f52cfcd202a;p=sbcl.git

diff --git a/src/compiler/x86/arith.lisp b/src/compiler/x86/arith.lisp
index 36b869d..ade7689 100644
--- a/src/compiler/x86/arith.lisp
+++ b/src/compiler/x86/arith.lisp
@@ -1,4 +1,4 @@
-;;;; the VM definition arithmetic VOPs for the x86
+;;;; the VM definition of arithmetic VOPs for the x86
 
 ;;;; This software is part of the SBCL system. See the README file for
 ;;;; more information.
@@ -176,9 +176,16 @@
 		  (:translate ,translate)
 		  (:generator ,untagged-penalty
 		  (move r x)
-		  (inst ,op r y))))))
-
-  ;;(define-binop + 4 add)
+		  ,(if (eq translate 'logand)
+		       ;; for the -C/UNSIGNED=>UNSIGNED VOP, this case
+		       ;; is optimized away as an identity somewhere
+		       ;; along the lines.  However, this VOP is used in
+		       ;; -C/SIGNED=>UNSIGNED, below, when the
+		       ;; higher-level lisp code can't optimize away the
+		       ;; non-trivial identity.
+		       `(unless (= y #.(1- (ash 1 n-word-bits)))
+			  (inst ,op r y))
+		       `(inst ,op r y)))))))
   (define-binop - 4 sub)
   (define-binop logand 2 and)
   (define-binop logior 2 or)
@@ -249,7 +256,6 @@
 	   (move r x)
 	   (inst add r y)))))
 
-
 ;;;; Special logand cases: (logand signed unsigned) => unsigned
 
 (define-vop (fast-logand/signed-unsigned=>unsigned
@@ -689,7 +695,9 @@
                   (if (plusp amount)
                       (inst shl result amount)
                       (inst shr result (- amount))))
-		 (t (inst xor result result)))))))
+		 (t (if (sc-is result unsigned-reg)
+                        (inst xor result result)
+                        (inst mov result 0))))))))
 
 (define-vop (fast-ash-left/signed=>signed)
   (:translate ash)
@@ -792,6 +800,82 @@
 
     DONE))
 
+(in-package "SB!C")
+
+(defknown %lea (integer integer (member 1 2 4 8) (signed-byte 32))
+  integer
+  (foldable flushable movable))
+
+(defoptimizer (%lea derive-type) ((base index scale disp))
+  (when (and (constant-lvar-p scale)
+	     (constant-lvar-p disp))
+    (let ((scale (lvar-value scale))
+	  (disp (lvar-value disp))
+	  (base-type (lvar-type base))
+	  (index-type (lvar-type index)))
+      (when (and (numeric-type-p base-type)
+		 (numeric-type-p index-type))
+	(let ((base-lo (numeric-type-low base-type))
+	      (base-hi (numeric-type-high base-type))
+	      (index-lo (numeric-type-low index-type))
+	      (index-hi (numeric-type-high index-type)))
+	  (make-numeric-type :class 'integer
+			     :complexp :real
+			     :low (when (and base-lo index-lo)
+				    (+ base-lo (* index-lo scale) disp))
+			     :high (when (and base-hi index-hi)
+				     (+ base-hi (* index-hi scale) disp))))))))
+
+(defun %lea (base index scale disp)
+  (+ base (* index scale) disp))
+
+(in-package "SB!VM")
+
+(define-vop (%lea/unsigned=>unsigned)
+  (:translate %lea)
+  (:policy :fast-safe)
+  (:args (base :scs (unsigned-reg))
+	 (index :scs (unsigned-reg)))
+  (:info scale disp)
+  (:arg-types unsigned-num unsigned-num
+	      (:constant (member 1 2 4 8))
+	      (:constant (signed-byte 32)))
+  (:results (r :scs (unsigned-reg)))
+  (:result-types unsigned-num)
+  (:generator 5
+    (inst lea r (make-ea :dword :base base :index index
+			 :scale scale :disp disp))))
+
+(define-vop (%lea/signed=>signed)
+  (:translate %lea)
+  (:policy :fast-safe)
+  (:args (base :scs (signed-reg))
+	 (index :scs (signed-reg)))
+  (:info scale disp)
+  (:arg-types signed-num signed-num
+	      (:constant (member 1 2 4 8))
+	      (:constant (signed-byte 32)))
+  (:results (r :scs (signed-reg)))
+  (:result-types signed-num)
+  (:generator 4
+    (inst lea r (make-ea :dword :base base :index index
+			 :scale scale :disp disp))))
+
+(define-vop (%lea/fixnum=>fixnum)
+  (:translate %lea)
+  (:policy :fast-safe)
+  (:args (base :scs (any-reg))
+	 (index :scs (any-reg)))
+  (:info scale disp)
+  (:arg-types tagged-num tagged-num
+	      (:constant (member 1 2 4 8))
+	      (:constant (signed-byte 32)))
+  (:results (r :scs (any-reg)))
+  (:result-types tagged-num)
+  (:generator 3
+    (inst lea r (make-ea :dword :base base :index index
+			 :scale scale :disp disp))))
+
 ;;; FIXME: before making knowledge of this too public, it needs to be
 ;;; fixed so that it's actually _faster_ than the non-CMOV version; at
 ;;; least on my Celeron-XXX laptop, this version is marginally slower
@@ -1080,39 +1164,6 @@
     (move result prev)
     (inst shrd result next :cl)))
 
-(define-source-transform 32bit-logical-not (x)
-  `(logand (lognot (the (unsigned-byte 32) ,x)) #.(1- (ash 1 32))))
-
-(deftransform 32bit-logical-and ((x y))
-  '(logand x y))
-
-(define-source-transform 32bit-logical-nand (x y)
-  `(32bit-logical-not (32bit-logical-and ,x ,y)))
-
-(deftransform 32bit-logical-or ((x y))
-  '(logior x y))
-
-(define-source-transform 32bit-logical-nor (x y)
-  `(32bit-logical-not (32bit-logical-or ,x ,y)))
-
-(deftransform 32bit-logical-xor ((x y))
-  '(logxor x y))
-
-(define-source-transform 32bit-logical-eqv (x y)
-  `(32bit-logical-not (32bit-logical-xor ,x ,y)))
-
-(define-source-transform 32bit-logical-orc1 (x y)
-  `(32bit-logical-or (32bit-logical-not ,x) ,y))
-
-(define-source-transform 32bit-logical-orc2 (x y)
-  `(32bit-logical-or ,x (32bit-logical-not ,y)))
-
-(define-source-transform 32bit-logical-andc1 (x y)
-  `(32bit-logical-and (32bit-logical-not ,x) ,y))
-
-(define-source-transform 32bit-logical-andc2 (x y)
-  `(32bit-logical-and ,x (32bit-logical-not ,y)))
-
 ;;; Only the lower 5 bits of the shift amount are significant.
 (define-vop (shift-towards-someplace)
   (:policy :fast-safe)
@@ -1152,10 +1203,54 @@
 (define-vop (fast---mod32-c/unsigned=>unsigned fast---c/unsigned=>unsigned)
   (:translate --mod32))
 
+(define-modular-fun *-mod32 (x y) * 32)
+(define-vop (fast-*-mod32/unsigned=>unsigned fast-*/unsigned=>unsigned)
+  (:translate *-mod32))
+;;; (no -C variant as x86 MUL instruction doesn't take an immediate)
+
 (define-vop (fast-ash-left-mod32-c/unsigned=>unsigned
              fast-ash-c/unsigned=>unsigned)
   (:translate ash-left-mod32))
 
+(define-vop (fast-ash-left-mod32/unsigned=>unsigned
+             fast-ash-left/unsigned=>unsigned))
+(deftransform ash-left-mod32 ((integer count)
+			      ((unsigned-byte 32) (unsigned-byte 5)))
+  (when (sb!c::constant-lvar-p count)
+    (sb!c::give-up-ir1-transform))
+  '(%primitive fast-ash-left-mod32/unsigned=>unsigned integer count))
+
+(in-package "SB!C")
+
+(defknown sb!vm::%lea-mod32 (integer integer (member 1 2 4 8) (signed-byte 32))
+  (unsigned-byte 32)
+  (foldable flushable movable))
+
+(define-modular-fun-optimizer %lea ((base index scale disp) :width width)
+  (when (and (<= width 32)
+	     (constant-lvar-p scale)
+	     (constant-lvar-p disp))
+    (cut-to-width base width)
+    (cut-to-width index width)
+    'sb!vm::%lea-mod32))
+
+#+sb-xc-host
+(defun sb!vm::%lea-mod32 (base index scale disp)
+  (ldb (byte 32 0) (%lea base index scale disp)))
+#-sb-xc-host
+(defun sb!vm::%lea-mod32 (base index scale disp)
+  (let ((base (logand base #xffffffff))
+	(index (logand index #xffffffff)))
+    ;; can't use modular version of %LEA, as we only have VOPs for
+    ;; constant SCALE and DISP.
+    (ldb (byte 32 0) (+ base (* index scale) disp))))
+
+(in-package "SB!VM")
+
+(define-vop (%lea-mod32/unsigned=>unsigned
+	     %lea/unsigned=>unsigned)
+  (:translate %lea-mod32))
+
 ;;; logical operations
 (define-modular-fun lognot-mod32 (x) lognot 32)
 (define-vop (lognot-mod32/unsigned=>unsigned)
@@ -1203,21 +1298,21 @@
 ;;;; bignum stuff
 
 (define-vop (bignum-length get-header-data)
-  (:translate sb!bignum::%bignum-length)
+  (:translate sb!bignum:%bignum-length)
   (:policy :fast-safe))
 
 (define-vop (bignum-set-length set-header-data)
-  (:translate sb!bignum::%bignum-set-length)
+  (:translate sb!bignum:%bignum-set-length)
   (:policy :fast-safe))
 
 (define-full-reffer bignum-ref * bignum-digits-offset other-pointer-lowtag
-  (unsigned-reg) unsigned-num sb!bignum::%bignum-ref)
+  (unsigned-reg) unsigned-num sb!bignum:%bignum-ref)
 
 (define-full-setter bignum-set * bignum-digits-offset other-pointer-lowtag
-  (unsigned-reg) unsigned-num sb!bignum::%bignum-set)
+  (unsigned-reg) unsigned-num sb!bignum:%bignum-set)
 
 (define-vop (digit-0-or-plus)
-  (:translate sb!bignum::%digit-0-or-plusp)
+  (:translate sb!bignum:%digit-0-or-plusp)
   (:policy :fast-safe)
   (:args (digit :scs (unsigned-reg)))
   (:arg-types unsigned-num)
@@ -1233,7 +1328,7 @@
 ;;; 4. This is easy to deal with and may save a fixnum-word
 ;;; conversion.
 (define-vop (add-w/carry)
-  (:translate sb!bignum::%add-with-carry)
+  (:translate sb!bignum:%add-with-carry)
   (:policy :fast-safe)
   (:args (a :scs (unsigned-reg) :target result)
 	 (b :scs (unsigned-reg unsigned-stack) :to :eval)
@@ -1254,7 +1349,7 @@
 ;;; Note: the borrow is the oppostite of the x86 convention - 1 for no
 ;;; borrow and 0 for a borrow.
 (define-vop (sub-w/borrow)
-  (:translate sb!bignum::%subtract-with-borrow)
+  (:translate sb!bignum:%subtract-with-borrow)
   (:policy :fast-safe)
   (:args (a :scs (unsigned-reg) :to :eval :target result)
 	 (b :scs (unsigned-reg unsigned-stack) :to :result)
@@ -1273,7 +1368,7 @@
 
 
 (define-vop (bignum-mult-and-add-3-arg)
-  (:translate sb!bignum::%multiply-and-add)
+  (:translate sb!bignum:%multiply-and-add)
   (:policy :fast-safe)
   (:args (x :scs (unsigned-reg) :target eax)
 	 (y :scs (unsigned-reg unsigned-stack))
@@ -1295,7 +1390,7 @@
     (move lo eax)))
 
 (define-vop (bignum-mult-and-add-4-arg)
-  (:translate sb!bignum::%multiply-and-add)
+  (:translate sb!bignum:%multiply-and-add)
   (:policy :fast-safe)
   (:args (x :scs (unsigned-reg) :target eax)
 	 (y :scs (unsigned-reg unsigned-stack))
@@ -1321,7 +1416,7 @@
 
 
 (define-vop (bignum-mult)
-  (:translate sb!bignum::%multiply)
+  (:translate sb!bignum:%multiply)
   (:policy :fast-safe)
   (:args (x :scs (unsigned-reg) :target eax)
 	 (y :scs (unsigned-reg unsigned-stack)))
@@ -1340,10 +1435,10 @@
     (move lo eax)))
 
 (define-vop (bignum-lognot lognot-mod32/unsigned=>unsigned)
-  (:translate sb!bignum::%lognot))
+  (:translate sb!bignum:%lognot))
 
 (define-vop (fixnum-to-digit)
-  (:translate sb!bignum::%fixnum-to-digit)
+  (:translate sb!bignum:%fixnum-to-digit)
   (:policy :fast-safe)
   (:args (fixnum :scs (any-reg control-stack) :target digit))
   (:arg-types tagged-num)
@@ -1357,7 +1452,7 @@
     (inst sar digit 2)))
 
 (define-vop (bignum-floor)
-  (:translate sb!bignum::%floor)
+  (:translate sb!bignum:%floor)
   (:policy :fast-safe)
   (:args (div-high :scs (unsigned-reg) :target edx)
 	 (div-low :scs (unsigned-reg) :target eax)
@@ -1378,7 +1473,7 @@
     (move rem edx)))
 
 (define-vop (signify-digit)
-  (:translate sb!bignum::%fixnum-digit-with-correct-sign)
+  (:translate sb!bignum:%fixnum-digit-with-correct-sign)
   (:policy :fast-safe)
   (:args (digit :scs (unsigned-reg unsigned-stack) :target res))
   (:arg-types unsigned-num)
@@ -1393,7 +1488,7 @@
       (inst shl res 2))))
 
 (define-vop (digit-ashr)
-  (:translate sb!bignum::%ashr)
+  (:translate sb!bignum:%ashr)
   (:policy :fast-safe)
   (:args (digit :scs (unsigned-reg unsigned-stack) :target result)
 	 (count :scs (unsigned-reg) :target ecx))
@@ -1409,14 +1504,14 @@
     (inst sar result :cl)))
 
 (define-vop (digit-lshr digit-ashr)
-  (:translate sb!bignum::%digit-logical-shift-right)
+  (:translate sb!bignum:%digit-logical-shift-right)
   (:generator 1
     (move result digit)
     (move ecx count)
     (inst shr result :cl)))
 
 (define-vop (digit-ashl digit-ashr)
-  (:translate sb!bignum::%ashl)
+  (:translate sb!bignum:%ashl)
   (:generator 1
     (move result digit)
     (move ecx count)
@@ -1504,84 +1599,6 @@
 
 (in-package "SB!C")
 
-(defknown %lea ((or (signed-byte 32) (unsigned-byte 32))
-		(or (signed-byte 32) (unsigned-byte 32))
-		(member 1 2 4 8) (signed-byte 32))
-  (or (signed-byte 32) (unsigned-byte 32))
-  (foldable flushable))
-
-(defoptimizer (%lea derive-type) ((base index scale disp))
-  (when (and (constant-lvar-p scale)
-	     (constant-lvar-p disp))
-    (let ((scale (lvar-value scale))
-	  (disp (lvar-value disp))
-	  (base-type (lvar-type base))
-	  (index-type (lvar-type index)))
-      (when (and (numeric-type-p base-type)
-		 (numeric-type-p index-type))
-	(let ((base-lo (numeric-type-low base-type))
-	      (base-hi (numeric-type-high base-type))
-	      (index-lo (numeric-type-low index-type))
-	      (index-hi (numeric-type-high index-type)))
-	  (make-numeric-type :class 'integer
-			     :complexp :real
-			     :low (when (and base-lo index-lo)
-				    (+ base-lo (* index-lo scale) disp))
-			     :high (when (and base-hi index-hi)
-				     (+ base-hi (* index-hi scale) disp))))))))
-
-(defun %lea (base index scale disp)
-  (+ base (* index scale) disp))
-
-(in-package "SB!VM")
-
-(define-vop (%lea/unsigned=>unsigned)
-  (:translate %lea)
-  (:policy :fast-safe)
-  (:args (base :scs (unsigned-reg))
-	 (index :scs (unsigned-reg)))
-  (:info scale disp)
-  (:arg-types unsigned-num unsigned-num
-	      (:constant (member 1 2 4 8))
-	      (:constant (signed-byte 32)))
-  (:results (r :scs (unsigned-reg)))
-  (:result-types unsigned-num)
-  (:generator 5
-    (inst lea r (make-ea :dword :base base :index index
-			 :scale scale :disp disp))))
-
-(define-vop (%lea/signed=>signed)
-  (:translate %lea)
-  (:policy :fast-safe)
-  (:args (base :scs (signed-reg))
-	 (index :scs (signed-reg)))
-  (:info scale disp)
-  (:arg-types signed-num signed-num
-	      (:constant (member 1 2 4 8))
-	      (:constant (signed-byte 32)))
-  (:results (r :scs (signed-reg)))
-  (:result-types signed-num)
-  (:generator 4
-    (inst lea r (make-ea :dword :base base :index index
-			 :scale scale :disp disp))))
-
-(define-vop (%lea/fixnum=>fixnum)
-  (:translate %lea)
-  (:policy :fast-safe)
-  (:args (base :scs (any-reg))
-	 (index :scs (any-reg)))
-  (:info scale disp)
-  (:arg-types tagged-num tagged-num
-	      (:constant (member 1 2 4 8))
-	      (:constant (signed-byte 32)))
-  (:results (r :scs (any-reg)))
-  (:result-types tagged-num)
-  (:generator 3
-    (inst lea r (make-ea :dword :base base :index index
-			 :scale scale :disp disp))))
-
-(in-package "SB!C")
-
 ;;; This is essentially a straight implementation of the algorithm in
 ;;; "Strength Reduction of Multiplications by Integer Constants",
 ;;; Youfeng Wu, ACM SIGPLAN Notices, Vol. 30, No.2, February 1995.
@@ -1654,25 +1671,35 @@
 	      (t (incf count)))))
     (decompose-multiplication arg x n-bits condensed)))
 
+(defun *-transformer (y)
+  (cond
+    ((= y (ash 1 (integer-length y)))
+     ;; there's a generic transform for y = 2^k
+     (give-up-ir1-transform))
+    ((member y '(3 5 9))
+     ;; we can do these multiplications directly using LEA
+     `(%lea x x ,(1- y) 0))
+    ((member :pentium4 *backend-subfeatures*)
+     ;; the pentium4's multiply unit is reportedly very good
+     (give-up-ir1-transform))
+    ;; FIXME: should make this more fine-grained.  If nothing else,
+    ;; there should probably be a cutoff of about 9 instructions on
+    ;; pentium-class machines.
+    (t (optimize-multiply 'x y))))
+
 (deftransform * ((x y)
 		 ((unsigned-byte 32) (constant-arg (unsigned-byte 32)))
 		 (unsigned-byte 32))
   "recode as leas, shifts and adds"
   (let ((y (lvar-value y)))
-    (cond
-      ((= y (ash 1 (integer-length y)))
-       ;; there's a generic transform for y = 2^k
-       (give-up-ir1-transform))
-      ((member y '(3 5 9))
-       ;; we can do these multiplications directly using LEA
-       `(%lea x x ,(1- y) 0))
-      ((member :pentium4 *backend-subfeatures*)
-       ;; the pentium4's multiply unit is reportedly very good
-       (give-up-ir1-transform))
-      ;; FIXME: should make this more fine-grained.  If nothing else,
-      ;; there should probably be a cutoff of about 9 instructions on
-      ;; pentium-class machines.
-      (t (optimize-multiply 'x y)))))
+    (*-transformer y)))
+
+(deftransform sb!vm::*-mod32
+    ((x y) ((unsigned-byte 32) (constant-arg (unsigned-byte 32)))
+     (unsigned-byte 32))
+  "recode as leas, shifts and adds"
+  (let ((y (lvar-value y)))
+    (*-transformer y)))
 
 ;;; FIXME: we should also be able to write an optimizer or two to
 ;;; convert (+ (* x 2) 17), (- (* x 9) 5) to a %LEA.