| blob | 4242c9968108ef1fd1b3c2e5f415502203a13b58 |
1 ;;;; function call for the x86 VM
3 ;;;; This software is part of the SBCL system. See the README file for
4 ;;;; more information.
5 ;;;;
6 ;;;; This software is derived from the CMU CL system, which was
7 ;;;; written at Carnegie Mellon University and released into the
8 ;;;; public domain. The software is in the public domain and is
9 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
10 ;;;; files for more information.
13 \f
14 ;;;; interfaces to IR2 conversion
16 ;;; Return a wired TN describing the N'th full call argument passing
17 ;;; location.
25 ;;; Make a passing location TN for a local call return PC.
26 ;;;
27 ;;; Always wire the return PC location to the stack in its standard
28 ;;; location.
32 sap-stack-sc-number return-pc-save-offset))
34 ;;; This is similar to MAKE-RETURN-PC-PASSING-LOCATION, but makes a
35 ;;; location to pass OLD-FP in.
36 ;;;
37 ;;; This is wired in both the standard and the local-call conventions,
38 ;;; because we want to be able to assume it's always there. Besides,
39 ;;; the x86 doesn't have enough registers to really make it profitable
40 ;;; to pass it in a register.
44 ocfp-save-offset))
46 ;;; Make the TNs used to hold OLD-FP and RETURN-PC within the current
47 ;;; function. We treat these specially so that the debugger can find
48 ;;; them at a known location.
49 ;;;
50 ;;; Without using a save-tn - which does not make much sense if it is
51 ;;; wired to the stack?
54 control-stack-sc-number
55 ocfp-save-offset)
56 physenv))
60 sap-stack-sc-number return-pc-save-offset)
61 physenv))
63 ;;; Make a TN for the standard argument count passing location. We only
64 ;;; need to make the standard location, since a count is never passed when we
65 ;;; are using non-standard conventions.
69 ;;; Make a TN to hold the number-stack frame pointer. This is allocated
70 ;;; once per component, and is component-live.
80 ;;; Return a list of TNs that can be used to represent an unknown-values
81 ;;; continuation within a function.
86 ;;; This function is called by the ENTRY-ANALYZE phase, allowing
87 ;;; VM-dependent initialization of the IR2-COMPONENT structure. We
88 ;;; push placeholder entries in the CONSTANTS to leave room for
89 ;;; additional noise in the code object header.
92 ;; The 1+ here is because for the x86 the first constant is a
93 ;; pointer to a list of fixups, or NIL if the code object has none.
94 ;; (If I understand correctly, the fixups are needed at GC copy
95 ;; time because the X86 code isn't relocatable.)
96 ;;
97 ;; KLUDGE: It'd be cleaner to have the fixups entry be a named
98 ;; element of the CODE (aka component) primitive object. However,
99 ;; it's currently a large, tricky, error-prone chore to change
100 ;; the layout of any primitive object, so for the foreseeable future
101 ;; we'll just live with this ugliness. -- WHN 2002-01-02
106 \f
107 ;;;; frame hackery
109 ;;; This is used for setting up the Old-FP in local call.
115 ;;; We don't have a separate NFP, so we don't need to do anything here.
120 nil))
129 ;; Skip space for the function header.
134 ;; The start of the actual code.
135 ;; Save the return-pc.
138 ;; If copy-more-arg follows it will allocate the correct stack
139 ;; size. The stack is not allocated first here as this may expose
140 ;; args on the stack if they take up more space than the frame!
142 ;; The args fit within the frame so just allocate the frame.
150 ;;; This is emitted directly before either a known-call-local, call-local,
151 ;;; or a multiple-call-local. All it does is allocate stack space for the
152 ;;; callee (who has the same size stack as us).
162 ;;; Allocate a partial frame for passing stack arguments in a full
163 ;;; call. NARGS is the number of arguments passed. We allocate at
164 ;;; least 3 slots, because the XEP noise is going to want to use them
165 ;;; before it can extend the stack.
172 \f
173 ;;; Emit code needed at the return-point from an unknown-values call
174 ;;; for a fixed number of values. Values is the head of the TN-REF
175 ;;; list for the locations that the values are to be received into.
176 ;;; Nvals is the number of values that are to be received (should
177 ;;; equal the length of Values).
178 ;;;
179 ;;; MOVE-TEMP is a DESCRIPTOR-REG TN used as a temporary.
180 ;;;
181 ;;; This code exploits the fact that in the unknown-values convention,
182 ;;; a single value return returns at the return PC + 2, whereas a
183 ;;; return of other than one value returns directly at the return PC.
184 ;;;
185 ;;; If 0 or 1 values are expected, then we just emit an instruction to
186 ;;; reset the SP (which will only be executed when other than 1 value
187 ;;; is returned.)
188 ;;;
189 ;;; In the general case we have to do three things:
190 ;;; -- Default unsupplied register values. This need only be done
191 ;;; when a single value is returned, since register values are
192 ;;; defaulted by the called in the non-single case.
193 ;;; -- Default unsupplied stack values. This needs to be done whenever
194 ;;; there are stack values.
195 ;;; -- Reset SP. This must be done whenever other than 1 value is
196 ;;; returned, regardless of the number of values desired.
208 ;; Default the unsupplied registers.
216 for count from 2 below register-arg-count
222 ;; The number of bytes depends on the relative jump instructions.
223 ;; Best case is 31+(n-3)*14, worst case is 35+(n-3)*18. For
224 ;; NVALS=6 that is 73/89 bytes, and for NVALS=7 that is 87/107
225 ;; bytes which is likely better than using the blt below.
230 ;; Branch off to the MV case.
232 ;; Do the single value case.
233 ;; Default the register args
240 ;; Fake other registers so it looks like we returned with all the
241 ;; registers filled in.
287 ;; Branch off to the MV case.
290 ;; Default the register args, and set up the stack as if we
291 ;; entered the MV return point.
297 ;; Compute a pointer to where to put the [defaulted] stack values.
302 ;; Load RAX with NIL so we can quickly store it, and set up
303 ;; stuff for the loop.
307 ;; Jump into the default loop.
310 ;; The regs are defaulted. We need to copy any stack arguments,
311 ;; and then default the remaining stack arguments.
313 ;; Save EDI.
315 ;; Compute the number of stack arguments, and if it's zero or
316 ;; less, don't copy any stack arguments.
320 ;; Throw away any unwanted args.
325 ;; Save the number of stack values.
327 ;; Compute a pointer to where the stack args go.
331 ;; Save ESI, and compute a pointer to where the args come from.
336 ;; Do the copy.
341 ;; Restore RSI.
343 ;; Now we have to default the remaining args. Find out how many.
346 ;; If none, then just blow out of here.
350 ;; Load RAX with NIL for fast storing.
352 ;; Do the store.
356 ;; Restore EDI, and reset the stack.
361 \f
362 ;;;; unknown values receiving
364 ;;; Emit code needed at the return point for an unknown-values call
365 ;;; for an arbitrary number of values.
366 ;;;
367 ;;; We do the single and non-single cases with no shared code: there
368 ;;; doesn't seem to be any potential overlap, and receiving a single
369 ;;; value is more important efficiency-wise.
370 ;;;
371 ;;; When there is a single value, we just push it on the stack,
372 ;;; returning the old SP and 1.
373 ;;;
374 ;;; When there is a variable number of values, we move all of the
375 ;;; argument registers onto the stack, and return ARGS and NARGS.
376 ;;;
377 ;;; ARGS and NARGS are TNs wired to the named locations. We must
378 ;;; explicitly allocate these TNs, since their lifetimes overlap with
379 ;;; the results start and count. (Also, it's nice to be able to target
380 ;;; them.)
396 ;; dtc: this writes the registers onto the stack even if they are
397 ;; not needed, only the number specified in rcx are used and have
398 ;; stack allocated to them. No harm is done.
401 for i downfrom -1
409 ;;; VOP that can be inherited by unknown values receivers. The main thing this
410 ;;; handles is allocation of the result temporaries.
414 values-start)
417 nvals)
420 \f
421 ;;;; local call with unknown values convention return
423 ;;; Non-TR local call for a fixed number of values passed according to
424 ;;; the unknown values convention.
425 ;;;
426 ;;; FP is the frame pointer in install before doing the call.
427 ;;;
428 ;;; NFP would be the number-stack frame pointer if we had a separate
429 ;;; number stack.
430 ;;;
431 ;;; Args are the argument passing locations, which are specified only
432 ;;; to terminate their lifetimes in the caller.
433 ;;;
434 ;;; VALUES are the return value locations (wired to the standard
435 ;;; passing locations). NVALS is the number of values received.
436 ;;;
437 ;;; Save is the save info, which we can ignore since saving has been
438 ;;; done.
439 ;;;
440 ;;; TARGET is a continuation pointing to the start of the called
441 ;;; function.
458 #+nil
463 ;; Is the return-pc on the stack or in a register?
475 RETURN
479 ;;; Non-TR local call for a variable number of return values passed according
480 ;;; to the unknown values convention. The results are the start of the values
481 ;;; glob and the number of values received.
497 #+nil
502 ;; Is the return-pc on the stack or in a register?
507 ;; Stack
511 ;; Register
516 RETURN
520 \f
521 ;;;; local call with known values return
523 ;;; Non-TR local call with known return locations. Known-value return
524 ;;; works just like argument passing in local call.
525 ;;;
526 ;;; Note: we can't use normal load-tn allocation for the fixed args,
527 ;;; since all registers may be tied up by the more operand. Instead,
528 ;;; we use MAYBE-LOAD-STACK-TN.
546 #+nil
551 ;; Is the return-pc on the stack or in a register?
556 ;; Stack
560 ;; Register
565 RETURN
568 \f
569 ;;; Return from known values call. We receive the return locations as
570 ;;; arguments to terminate their lifetimes in the returning function. We
571 ;;; restore FP and CSP and jump to the Return-PC.
572 ;;;
573 ;;; We can assume we know exactly where old-fp and return-pc are because
574 ;;; make-old-fp-save-location and make-return-pc-save-location always
575 ;;; return the same place.
576 #+nil
588 ;; Save the return-pc in a register 'cause the frame-pointer is
589 ;; going away. Note this not in the usual stack location so we
590 ;; can't use RET
592 ;; Restore the stack.
594 ;; Restore the old fp. We know OLD-FP is going to be in its stack
595 ;; save slot, which is a different frame that than this one,
596 ;; so we don't have to worry about having just cleared
597 ;; most of the stack.
601 \f
602 ;;; From Douglas Crosher
603 ;;; Return from known values call. We receive the return locations as
604 ;;; arguments to terminate their lifetimes in the returning function. We
605 ;;; restore FP and CSP and jump to the Return-PC.
606 ;;;
607 ;;; The old-fp may be either in a register or on the stack in its
608 ;;; standard save locations - slot 0.
609 ;;;
610 ;;; The return-pc may be in a register or on the stack in any slot.
621 ;; return-pc may be either in a register or on the stack.
627 ;; Zot all of the stack except for the old-fp.
630 n-word-bytes))))
631 ;; Restore the old fp from its save location on the stack,
632 ;; and zot the stack.
637 "VOP return-local doesn't work if old-fp (in slot ~
638 ~S) is not in slot 0"
642 ;; Zot all the stack.
644 ;; Restore the old-fp.
647 ;; Return; return-pc is in a register.
658 \f
659 ;;;; full call
660 ;;;
661 ;;; There is something of a cross-product effect with full calls.
662 ;;; Different versions are used depending on whether we know the
663 ;;; number of arguments or the name of the called function, and
664 ;;; whether we want fixed values, unknown values, or a tail call.
665 ;;;
666 ;;; In full call, the arguments are passed creating a partial frame on
667 ;;; the stack top and storing stack arguments into that frame. On
668 ;;; entry to the callee, this partial frame is pointed to by FP.
670 ;;; This macro helps in the definition of full call VOPs by avoiding
671 ;;; code replication in defining the cross-product VOPs.
672 ;;;
673 ;;; NAME is the name of the VOP to define.
674 ;;;
675 ;;; NAMED is true if the first argument is an fdefinition object whose
676 ;;; definition is to be called.
677 ;;;
678 ;;; RETURN is either :FIXED, :UNKNOWN or :TAIL:
679 ;;; -- If :FIXED, then the call is for a fixed number of values, returned in
680 ;;; the standard passing locations (passed as result operands).
681 ;;; -- If :UNKNOWN, then the result values are pushed on the stack, and the
682 ;;; result values are specified by the Start and Count as in the
683 ;;; unknown-values continuation representation.
684 ;;; -- If :TAIL, then do a tail-recursive call. No values are returned.
685 ;;; The Old-Fp and Return-PC are passed as the second and third arguments.
686 ;;;
687 ;;; In non-tail calls, the pointer to the stack arguments is passed as
688 ;;; the last fixed argument. If Variable is false, then the passing
689 ;;; locations are passed as a more arg. Variable is true if there are
690 ;;; a variable number of arguments passed on the stack. Variable
691 ;;; cannot be specified with :TAIL return. TR variable argument call
692 ;;; is implemented separately.
693 ;;;
694 ;;; In tail call with fixed arguments, the passing locations are
695 ;;; passed as a more arg, but there is no new-FP, since the arguments
696 ;;; have been set up in the current frame.
702 (:args
724 (:info
728 step-instrumenting)
730 (:ignore
734 ;; We pass either the fdefn object (for named call) or
735 ;; the actual function object (for unnamed call) in
736 ;; RAX. With named call, closure-tramp will replace it
737 ;; with the real function and invoke the real function
738 ;; for closures. Non-closures do not need this value,
739 ;; so don't care what shows up in it.
740 (:temporary
741 (:sc descriptor-reg
742 :offset rax-offset
745 rax)
747 ;; We pass the number of arguments in RCX.
750 ;; With variable call, we have to load the
751 ;; register-args out of the (new) stack frame before
752 ;; doing the call. Therefore, we have to tell the
753 ;; lifetime stuff that we need to use them.
767 old-fp-tmp)))
772 15
776 ;; This has to be done before the frame pointer is
777 ;; changed! RAX stores the 'lexical environment' needed
778 ;; for closures.
783 ;; For variable call, compute the number of
784 ;; arguments and move some of the arguments to
785 ;; registers.
787 ;; Compute the number of arguments.
790 ;; Move the necessary args to registers,
791 ;; this moves them all even if they are
792 ;; not all needed.
795 for index downfrom -1
803 ;; return to so might as well use that clue.
804 ;; This seems really important to the
805 ;; implementation of things like
806 ;; (without-interrupts ...)
807 ;;
808 ;; dtc; Could be doing a tail call from a
809 ;; known-local-call etc in which the old-fp
810 ;; or ret-pc are in regs or in non-standard
811 ;; places. If the passing location were
812 ;; wired to the stack in standard locations
813 ;; then these moves will be un-necessary;
814 ;; this is probably best for the x86.
819 ;; FIXME: FORMAT T for stale
820 ;; diagnostic output (several of
821 ;; them around here), ick
825 rbp-tn
830 rbp-tn
833 ;; For tail call, we have to push the
834 ;; return-pc so that it looks like we CALLed
835 ;; drspite the fact that we are going to JMP.
837 ))
839 ;; For non-tail call, we have to save our
840 ;; frame pointer and install the new frame
841 ;; pointer. We can't load stack tns after this
842 ;; point.
844 ;; here which doesn't leave room for the
845 ;; ofp/ret stuff.
847 ;; The variable args are on the stack and
848 ;; become the frame, but there may be <3
849 ;; args and 3 stack slots are assumed
850 ;; allocate on the call. So need to ensure
851 ;; there are at least 3 slots. This hack
852 ;; just adds 3 more.
856 ;; Save the fp
860 )))
866 DONE
874 n-word-bytes)
875 other-pointer-lowtag)
877 fun-pointer-lowtag))))
879 (:fixed
881 (:unknown
897 ;;; This is defined separately, since it needs special code that BLT's
898 ;;; the arguments down. All the real work is done in the assembly
899 ;;; routine. We just set things up so that it can find what it needs.
908 ; (:ignore ret-addr old-fp)
910 ;; Move these into the passing locations if they are not already there.
914 ;; The following assumes that the return-pc and old-fp are on the
915 ;; stack in their standard save locations - Check this.
927 ;; And jump to the assembly routine.
929 \f
930 ;;;; unknown values return
932 ;;; Return a single-value using the Unknown-Values convention. Specifically,
933 ;;; we jump to clear the stack and jump to return-pc+3.
934 ;;;
935 ;;; We require old-fp to be in a register, because we want to reset RSP before
936 ;;; restoring RBP. If old-fp were still on the stack, it could get clobbered
937 ;;; by a signal.
938 ;;;
939 ;;; pfw--get wired-tn conflicts sometimes if register sc specd for args
940 ;;; having problems targeting args to regs -- using temps instead.
948 ;; Code structure lifted from known-return.
951 ;; return PC in register for some reason (local call?)
952 ;; we jmp to the return pc after fixing the stack and frame.
955 ;; ofp on stack must be in slot 0 (the traditional storage place).
956 ;; Drop the stack above it and pop it off.
960 n-word-bytes))))
963 ;; Should this ever happen, we do the same as above, but
964 ;; using (tn-offset old-fp) instead of ocfp-save-offset
965 ;; (which is 0 anyway, see src/compiler/x86/vm.lisp) and
966 ;; then lea rsp again against itself with a displacement
967 ;; of (* (tn-offset old-fp) n-word-bytes) to clear the
968 ;; rest of the stack.
972 ;; ofp in reg, drop the stack and load the real fp.
976 ;; Set single-value-return flag
978 ;; And return
982 ;; Note that this will only work right if, when old-fp is on
983 ;; the stack, it has a lower tn-offset than return-pc. One of
984 ;; the comments in known-return indicate that this is the case
985 ;; (in that it will be in its save location), but we may wish
986 ;; to assert that (in either the weaker or stronger forms).
987 ;; Should this ever not be the case, we should load old-fp
988 ;; into a temp reg while we fix the stack.
989 ;; Drop stack above return-pc
992 n-word-bytes))))
993 ;; Set single-value return flag
995 ;; Restore the old frame pointer
997 ;; And return, dropping the rest of the stack as we go.
1000 ;;; Do unknown-values return of a fixed (other than 1) number of
1001 ;;; values. The VALUES are required to be set up in the standard
1002 ;;; passing locations. NVALS is the number of values returned.
1003 ;;;
1004 ;;; Basically, we just load RCX with the number of values returned and
1005 ;;; RBX with a pointer to the values, set RSP to point to the end of
1006 ;;; the values, and jump directly to return-pc.
1014 ;; In the case of other than one value, we need these registers to
1015 ;; tell the caller where they are and how many there are.
1019 ;; We need to stretch the lifetime of return-pc past the argument
1020 ;; registers so that we can default the argument registers without
1021 ;; trashing return-pc.
1031 ;; Establish the values pointer and values count.
1036 ;; Restore the frame pointer.
1038 ;; Clear as much of the stack as possible, but not past the return
1039 ;; address.
1042 ;; Pre-default any argument register that need it.
1049 ;; Set the multiple value return flag.
1051 ;; And away we go. Except that return-pc is still on the
1052 ;; stack and we've changed the stack pointer. So we have to
1053 ;; tell it to index off of RBX instead of RBP.
1055 ;; Return popping the return address and the OCFP.
1058 ;; Return popping the return, leaving 1 slot. Can this
1059 ;; happen, or is a single value return handled elsewhere?
1064 n-word-bytes))))))
1068 ;;; Do unknown-values return of an arbitrary number of values (passed
1069 ;;; on the stack.) We check for the common case of a single return
1070 ;;; value, and do that inline using the normal single value return
1071 ;;; convention. Otherwise, we branch off to code that calls an
1072 ;;; assembly-routine.
1073 ;;;
1074 ;;; The assembly routine takes the following args:
1075 ;;; RAX -- the return-pc to finally jump to.
1076 ;;; RBX -- pointer to where to put the values.
1077 ;;; RCX -- number of values to find there.
1078 ;;; RSI -- pointer to where to find the values.
1097 ;; Load the return-pc.
1100 ;; Check for the single case.
1105 ;; Return with one value.
1107 ;; Clear the stack. We load old-fp into a register before clearing
1108 ;; the stack.
1112 ;; clear the multiple-value return flag
1114 ;; Out of here.
1117 ;; Nope, not the single case. Jump to the assembly routine.
1128 \f
1129 ;;;; XEP hackery
1131 ;;; We don't need to do anything special for regular functions.
1136 ;; Don't bother doing anything.
1137 nil))
1139 ;;; Get the lexical environment from its passing location.
1145 ;; Get result.
1148 ;;; Copy a &MORE arg from the argument area to the end of the current
1149 ;;; frame. FIXED is the number of non-&MORE arguments.
1156 ;; Avoid the copy if there are no more args.
1163 ;; Allocate the space on the stack.
1164 ;; stack = rbp - (max 3 frame-size) - (nargs - fixed)
1168 (* n-word-bytes
1173 ;; Now: nargs>=1 && nargs>fixed
1175 ;; Save the original count of args.
1179 ;; We must stop when we run out of stack args, not when we
1180 ;; run out of more args.
1181 ;; Number to copy = nargs-3
1183 ;; Everything of interest in registers.
1186 ;; Number to copy = nargs-fixed
1189 ;; Initialize R8 to be the end of args.
1193 ;; We need to copy from downwards up to avoid overwriting some of
1194 ;; the yet uncopied args. So we need to use R9 as the copy index
1195 ;; and RCX as the loop counter, rather than using RCX for both.
1198 ;; We used to use REP MOVS here, but on modern x86 it performs
1199 ;; much worse than an explicit loop for small blocks.
1200 COPY-LOOP
1207 DO-REGS
1209 ;; Restore RCX
1212 ;; Here: nargs>=1 && nargs>fixed
1214 ;; Now we have to deposit any more args that showed up in
1215 ;; registers.
1218 ;; Store it relative to rbp
1229 ;; Don't deposit any more than there are.
1237 JUST-ALLOC-FRAME
1243 DONE))
1272 ;;; Turn more arg (context, count) into a list.
1274 t)
1295 ;; Check to see whether there are no args, and just return NIL if so.
1303 ;; Set decrement mode (successive args at lower addresses)
1305 ;; Set up the result.
1307 ;; Jump into the middle of the loop, 'cause that's where we want
1308 ;; to start.
1311 ;; Compute a pointer to the next cons.
1313 ;; Store a pointer to this cons in the CDR of the previous cons.
1316 ;; Grab one value and stash it in the car of this cons.
1319 ;; Go back for more.
1322 ;; NIL out the last cons.
1326 ;;; Return the location and size of the &MORE arg glob created by
1327 ;;; COPY-MORE-ARG. SUPPLIED is the total number of arguments supplied
1328 ;;; (originally passed in RCX). FIXED is the number of non-rest
1329 ;;; arguments.
1330 ;;;
1331 ;;; We must duplicate some of the work done by COPY-MORE-ARG, since at
1332 ;;; that time the environment is in a pretty brain-damaged state,
1333 ;;; preventing this info from being returned as values. What we do is
1334 ;;; compute supplied - fixed, and return a pointer that many words
1335 ;;; below the current stack top.
1348 ;; SP at this point points at the last arg pushed.
1349 ;; Point to the first more-arg, not above it.
1356 ;;; Signal wrong argument count error if NARGS isn't equal to COUNT.
1373 ;;; Various other error signallers.
1381 args))
1389 object type)
1391 object layout)
1398 ;;; Single-stepping
1401 ;; We use different ways of representing whether stepping is on on
1402 ;; +SB-THREAD / -SB-THREAD: on +SB-THREAD, we use a slot in the
1403 ;; thread structure. On -SB-THREAD we use the value of a static
1404 ;; symbol. Things are done this way, since reading a thread-local
1405 ;; slot from a symbol would require an extra register on +SB-THREAD,
1406 ;; and reading a slot from a thread structure would require an extra
1407 ;; register on -SB-THREAD. While this isn't critical for x86-64,
1408 ;; it's more serious for x86.
1409 #!+sb-thread
1411 :base thread-base-tn
1413 nil-value)
1414 #!-sb-thread
1420 nil-value))
1429 DONE