X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcompiler%2Fx86%2Fsystem.lisp;h=d65cda32e167f90ff39a0e15783a09468e32510f;hb=9b1fade83db8453b75b8c7380eb12ce41b5b889c;hp=b9e9c5fbf0f443f6d7ee6b2d8847508a997226d9;hpb=4d5a8689d1d303f65c2fa933bb8757063641a8f9;p=sbcl.git diff --git a/src/compiler/x86/system.lisp b/src/compiler/x86/system.lisp index b9e9c5f..d65cda3 100644 --- a/src/compiler/x86/system.lisp +++ b/src/compiler/x86/system.lisp @@ -85,9 +85,7 @@ (:result-types positive-fixnum) (:generator 6 (move eax type) - (inst mov - (make-ea :byte :base function :disp (- fun-pointer-lowtag)) - al-tn) + (storeb al-tn function 0 fun-pointer-lowtag) (move result eax))) (define-vop (get-header-data) @@ -126,14 +124,16 @@ (storew eax x 0 other-pointer-lowtag) (move res x))) -(define-vop (make-fixnum) +(define-vop (pointer-hash) + (:translate pointer-hash) (:args (ptr :scs (any-reg descriptor-reg) :target res)) (:results (res :scs (any-reg descriptor-reg))) + (:policy :fast-safe) (:generator 1 - ;; Some code (the hash table code) depends on this returning a - ;; positive number so make sure it does. (move res ptr) - (inst shl res 3) + ;; Mask the lowtag, and shift the whole address into a positive + ;; fixnum. + (inst and res (lognot lowtag-mask)) (inst shr res 1))) (define-vop (make-other-immediate-type) @@ -232,9 +232,6 @@ (define-source-transform %closure-fun (closure) `(%simple-fun-self ,closure)) -(define-source-transform %funcallable-instance-fun (fin) - `(%simple-fun-self ,fin)) - (define-vop (%set-fun-self) (:policy :fast-safe) (:translate (setf %simple-fun-self)) @@ -243,26 +240,10 @@ (:temporary (:sc any-reg :from (:argument 0) :to :result) temp) (:results (result :scs (descriptor-reg))) (:generator 3 - (inst lea temp - (make-ea :byte :base new-self - :disp (- (ash simple-fun-code-offset word-shift) - fun-pointer-lowtag))) + (inst lea temp (make-ea-for-object-slot new-self simple-fun-code-offset + fun-pointer-lowtag)) (storew temp function simple-fun-self-slot fun-pointer-lowtag) (move result new-self))) - -;;; KLUDGE: This seems to be some kind of weird override of the way -;;; that the objdef.lisp code would ordinarily set up the slot -;;; accessor. It's inherited from CMU CL, and it works, and naively -;;; deleting it seemed to cause problems, but it's not obvious why -;;; it's done this way. Any ideas? -- WHN 2001-08-02 -(defknown ((setf %funcallable-instance-fun)) (function function) function - (unsafe)) -;;; CMU CL comment: -;;; We would have really liked to use a source-transform for this, but -;;; they don't work with SETF functions. -;;; FIXME: Can't we just use DEFSETF or something? -(deftransform (setf %funcallable-instance-fun) ((value fin)) - '(setf (%simple-fun-self fin) value)) ;;;; other miscellaneous VOPs @@ -303,13 +284,72 @@ (note-next-instruction vop :internal-error) (inst wait))) -;;;; dynamic vop count collection support +;;;; Miscellany + +;;; the RDTSC instruction (present on Pentium processors and +;;; successors) allows you to access the time-stamp counter, a 64-bit +;;; model-specific register that counts executed cycles. The +;;; instruction returns the low cycle count in EAX and high cycle +;;; count in EDX. +;;; +;;; In order to obtain more significant results on out-of-order +;;; processors (such as the Pentium II and later), we issue a +;;; serializing CPUID instruction before reading the cycle counter. +;;; This instruction is used for its side effect of emptying the +;;; processor pipeline, to ensure that the RDTSC instruction is +;;; executed once all pending instructions have been completed. +;;; CPUID writes to EBX and ECX in addition to EAX and EDX, so +;;; they need to be added as temporaries. +;;; +;;; Note that cache effects mean that the cycle count can vary for +;;; different executions of the same code (it counts cycles, not +;;; retired instructions). Furthermore, the results are per-processor +;;; and not per-process, so are unreliable on multiprocessor machines +;;; where processes can migrate between processors. +;;; +;;; This method of obtaining a cycle count has the advantage of being +;;; very fast (around 20 cycles), and of not requiring a system call. +;;; However, you need to know your processor's clock speed to translate +;;; this into real execution time. +;;; +;;; FIXME: This about the WITH-CYCLE-COUNTER interface a bit, and then +;;; perhaps export it from SB-SYS. + +(defknown %read-cycle-counter () (values (unsigned-byte 32) (unsigned-byte 32)) ()) + +(define-vop (%read-cycle-counter) + (:policy :fast-safe) + (:translate %read-cycle-counter) + (:temporary (:sc unsigned-reg :offset eax-offset :target lo) eax) + (:temporary (:sc unsigned-reg :offset edx-offset :target hi) edx) + (:temporary (:sc unsigned-reg :offset ebx-offset) ebx) + (:temporary (:sc unsigned-reg :offset ecx-offset) ecx) + (:ignore ebx ecx) + (:results (hi :scs (unsigned-reg)) + (lo :scs (unsigned-reg))) + (:result-types unsigned-num unsigned-num) + (:generator 5 + (inst xor eax eax) + ;; Intel docs seem quite consistent on only using CPUID before RDTSC, + ;; not both before and after. Go figure. + (inst cpuid) + (inst rdtsc) + (move lo eax) + (move hi edx))) + +(defmacro with-cycle-counter (&body body) + "Returns the primary value of BODY as the primary value, and the +number of CPU cycles elapsed as secondary value. EXPERIMENTAL." + (with-unique-names (hi0 hi1 lo0 lo1) + `(multiple-value-bind (,hi0 ,lo0) (%read-cycle-counter) + (values (locally ,@body) + (multiple-value-bind (,hi1 ,lo1) (%read-cycle-counter) + (+ (ash (- ,hi1 ,hi0) 32) + (- ,lo1 ,lo0))))))) #!+sb-dyncount (define-vop (count-me) (:args (count-vector :scs (descriptor-reg))) (:info index) (:generator 0 - (inst inc (make-ea :dword :base count-vector - :disp (- (* (+ vector-data-offset index) n-word-bytes) - other-pointer-lowtag))))) + (inst inc (make-ea-for-vector-data count-vector :offset index))))