2 * very-low-level utilities for runtime support
6 * This software is part of the SBCL system. See the README file for
9 * This software is derived from the CMU CL system, which was
10 * written at Carnegie Mellon University and released into the
11 * public domain. The software is in the public domain and is
12 * provided with absolutely no warranty. See the COPYING and CREDITS
13 * files for more information.
16 #define LANGUAGE_ASSEMBLY
17 #include "genesis/config.h"
20 #include "genesis/closure.h"
21 #include "genesis/funcallable-instance.h"
22 #include "genesis/fdefn.h"
23 #include "genesis/static-symbols.h"
24 #include "genesis/symbol.h"
25 #include "genesis/thread.h"
27 /* Minimize conditionalization for different OS naming schemes. */
28 #if defined __linux__ || defined __FreeBSD__ /* (but *not* OpenBSD) */
29 #define GNAME(var) var
31 #define GNAME(var) _##var
34 /* Get the right type of alignment. Linux and FreeBSD (but not OpenBSD)
35 * want alignment in bytes. */
36 #if defined(__linux__) || defined(__FreeBSD__)
39 #define align_16byte 16
40 #define align_32byte 32
44 #define align_16byte 4
48 * The assembler used for win32 doesn't like .type or .size directives,
49 * so we want to conditionally kill them out. So let's wrap them in macros
50 * that are defined to be no-ops on win32. Hopefully this still works on
53 #if !defined(LISP_FEATURE_WIN32) && !defined(LISP_FEATURE_DARWIN)
54 #define TYPE(name) .type name,@function
55 #define SIZE(name) .size name,.-name
56 #define DOLLAR(name) $(name)
63 * x86/darwin (as of MacOS X 10.4.5) doesn't reliably fire signal
64 * handlers (SIGTRAP or Mach exception handlers) for 0xCC, wo we have
65 * to use ud2 instead. ud2 is an undefined opcode, #x0b0f, or
66 * 0F 0B in low-endian notation, that causes SIGILL to fire. We check
67 * for this instruction in the SIGILL handler and if we see it, we
68 * advance the EIP by two bytes to skip over ud2 instruction and
69 * call sigtrap_handler. */
70 #if defined(LISP_FEATURE_DARWIN)
77 * More Apple assembler hacks
80 #if defined(LISP_FEATURE_DARWIN)
81 /* global symbol x86-64 sym(%rip) hack:*/
82 #define GSYM(name) name(%rip)
85 #define GSYM(name) $name
91 .globl GNAME(all_threads)
95 /* From lower to higher-numbered addresses, the stack contains
96 * return address, arg 0, arg 1, arg 2 ...
97 * rax contains the address of the function to call
98 * Lisp expects return value in rax, whic is already consistent with C
99 * XXXX correct floating point handling is unimplemented so far
100 * Based on comments cleaned from x86-assem.S, we believe that
101 * Lisp is expecting us to preserve rsi, rdi, rsp (no idea about r8-15)
104 .align align_16byte,0x90
105 .globl GNAME(call_into_c)
106 TYPE(GNAME(call_into_c))
108 /* ABI requires that the direction flag be clear on function
111 push %rbp # Save old frame pointer.
112 mov %rsp,%rbp # Establish new frame.
114 push %rsi # args are going in here
127 SIZE(GNAME(call_into_c))
131 .globl GNAME(call_into_lisp_first_time)
132 TYPE(GNAME(call_into_lisp_first_time))
134 /* The *ALIEN-STACK* pointer is set up on the first call_into_lisp when
135 * the stack changes. We don't worry too much about saving registers
136 * here, because we never expect to return from the initial call to lisp
139 .align align_16byte,0x90
140 GNAME(call_into_lisp_first_time):
141 push %rbp # Save old frame pointer.
142 mov %rsp,%rbp # Establish new frame.
143 mov %rsp,ALIEN_STACK + SYMBOL_VALUE_OFFSET
144 movq GSYM(GNAME(all_threads)),%rax
145 mov THREAD_CONTROL_STACK_START_OFFSET(%rax) ,%rsp
146 /* don't think too hard about what happens if we get interrupted
148 add $(THREAD_CONTROL_STACK_SIZE)-16,%rsp
152 .globl GNAME(call_into_lisp)
153 TYPE(GNAME(call_into_lisp))
156 * amd64 calling convention: C expects that
157 * arguments go in rdi rsi rdx rcx r8 r9
158 * return values in rax rdx
159 * callee saves rbp rbx r12-15 if it uses them
162 .align align_16byte,0x90
163 GNAME(call_into_lisp):
164 push %rbp # Save old frame pointer.
165 mov %rsp,%rbp # Establish new frame.
167 /* FIXME x86 saves FPU state here */
168 push %rbx # these regs are callee-saved according to C
169 push %r12 # so must be preserved and restored when
170 push %r13 # the lisp function returns
174 mov %rsp,%rbx # remember current stack
175 push %rbx # Save entry stack on (maybe) new stack.
177 push %rdi # args from C
180 #ifdef LISP_FEATURE_SB_THREAD
182 call pthread_getspecific
186 pop %rbx # arg vector
187 pop %rax # function ptr/lexenv
189 xor %rdx,%rdx # clear any descriptor registers
190 xor %rdi,%rdi # that we can't be sure we'll
191 xor %rsi,%rsi # initialise properly. XX do r8-r15 too?
192 shl $3,%rcx # (fixnumize num-args)
195 mov 0(%rbx),%rdx # arg0
198 mov 8(%rbx),%rdi # arg1
201 mov 16(%rbx),%rsi # arg2
203 /* Registers rax, rcx, rdx, rdi, and rsi are now live. */
204 xor %rbx,%rbx # available
206 /* Alloc new frame. */
207 mov %rsp,%rbx # The current sp marks start of new frame.
208 push %rbp # fp in save location S0
209 sub $16,%rsp # Ensure 3 slots are allocated, one above.
210 mov %rbx,%rbp # Switch to new frame.
213 call *CLOSURE_FUN_OFFSET(%rax)
215 /* If the function returned multiple values, it will return to
216 this point. Lose them */
221 /* Restore the stack, in case there was a stack change. */
231 /* ABI requires that the direction flag be clear on function
235 /* FIXME Restore the NPX state. */
237 /* return value is already in rax where lisp expects it */
240 SIZE(GNAME(call_into_lisp))
242 /* support for saving and restoring the NPX state from C */
244 .globl GNAME(fpu_save)
245 TYPE(GNAME(fpu_save))
249 fnsave (%rax) # Save the NPX state. (resets NPX)
251 SIZE(GNAME(fpu_save))
253 .globl GNAME(fpu_restore)
254 TYPE(GNAME(fpu_restore))
258 frstor (%rax) # Restore the NPX state.
260 SIZE(GNAME(fpu_restore))
263 * the undefined-function trampoline
266 .align align_8byte,0x90
267 .globl GNAME(undefined_tramp)
268 TYPE(GNAME(undefined_tramp))
269 GNAME(undefined_tramp):
273 .byte UNDEFINED_FUN_ERROR
274 .byte sc_DescriptorReg # eax in the Descriptor-reg SC
276 SIZE(GNAME(undefined_tramp))
280 .align align_8byte,0x90
281 .globl GNAME(alloc_tramp)
282 TYPE(GNAME(alloc_tramp))
284 push %rbp # Save old frame pointer.
285 mov %rsp,%rbp # Establish new frame.
309 SIZE(GNAME(alloc_tramp))
313 * the closure trampoline
316 .align align_8byte,0x90
317 .globl GNAME(closure_tramp)
318 TYPE(GNAME(closure_tramp))
319 GNAME(closure_tramp):
320 mov FDEFN_FUN_OFFSET(%rax),%rax
321 /* FIXME: The '*' after "jmp" in the next line is from PVE's
322 * patch posted to the CMU CL mailing list Oct 6, 1999. It looks
323 * reasonable, and it certainly seems as though if CMU CL needs it,
324 * SBCL needs it too, but I haven't actually verified that it's
325 * right. It would be good to find a way to force the flow of
326 * control through here to test it. */
327 jmp *CLOSURE_FUN_OFFSET(%rax)
328 SIZE(GNAME(closure_tramp))
331 .align align_8byte,0x90
332 .globl GNAME(funcallable_instance_tramp)
333 #if !defined(LISP_FEATURE_DARWIN)
334 .type GNAME(funcallable_instance_tramp),@function
336 GNAME(funcallable_instance_tramp):
337 mov FUNCALLABLE_INSTANCE_FUNCTION_OFFSET(%rax),%rax
338 /* KLUDGE: on this platform, whatever kind of function is in %rax
339 * now, the first word of it contains the address to jump to. */
340 jmp *CLOSURE_FUN_OFFSET(%rax)
341 #if !defined(LISP_FEATURE_DARWIN)
342 .size GNAME(funcallable_instance_tramp), .-GNAME(funcallable_instance_tramp)
345 * fun-end breakpoint magic
348 .globl GNAME(fun_end_breakpoint_guts)
350 GNAME(fun_end_breakpoint_guts):
351 /* Multiple Value return */
352 jc multiple_value_return
353 /* Single value return: The eventual return will now use the
354 multiple values return convention but with a return values
356 mov %rsp,%rbx # Setup ebx - the ofp.
357 sub $8,%rsp # Allocate one stack slot for the return value
358 mov $8,%rcx # Setup ecx for one return value.
359 #if defined(LISP_FEATURE_DARWIN)
360 mov GSYM(NIL),%rdi # default second value
361 mov GSYM(NIL),%rsi # default third value
363 mov $NIL,%rdi # default second value
364 mov $NIL,%rsi # default third value
366 multiple_value_return:
368 .globl GNAME(fun_end_breakpoint_trap)
369 GNAME(fun_end_breakpoint_trap):
371 .byte trap_FunEndBreakpoint
372 hlt # We should never return here.
374 .globl GNAME(fun_end_breakpoint_end)
375 GNAME(fun_end_breakpoint_end):
378 .globl GNAME(do_pending_interrupt)
379 TYPE(GNAME(do_pending_interrupt))
380 .align align_8byte,0x90
381 GNAME(do_pending_interrupt):
383 .byte trap_PendingInterrupt
385 SIZE(GNAME(do_pending_interrupt))
387 .globl GNAME(post_signal_tramp)
388 TYPE(GNAME(post_signal_tramp))
389 .align align_8byte,0x90
390 GNAME(post_signal_tramp):
391 /* this is notionally the second half of a function whose first half
392 * doesn't exist. This is where call_into_lisp returns when called
393 * using return_to_lisp_function */
404 /* skip RBP and RSP */
412 SIZE(GNAME(post_signal_tramp))
415 .align align_8byte,0x90
416 .globl GNAME(fast_bzero)
417 TYPE(GNAME(fast_bzero))
420 /* A fast routine for zero-filling blocks of memory that are
421 * guaranteed to start and end at a 4096-byte aligned address.
423 shr $6, %rsi /* Amount of 64-byte blocks to copy */
424 jz Lend /* If none, stop */
425 mov %rsi, %rcx /* Save start address */
426 movups %xmm7, -16(%rsp) /* Save XMM register */
427 xorps %xmm7, %xmm7 /* Zero the XMM register */
432 /* Copy the 16 zeroes from xmm7 to memory, 4 times. MOVNTDQ is the
433 * non-caching double-quadword moving variant, i.e. the memory areas
434 * we're touching are not fetched into the L1 cache, since we're just
435 * going to overwrite the memory soon anyway.
437 movntdq %xmm7, 0(%rdi)
438 movntdq %xmm7, 16(%rdi)
439 movntdq %xmm7, 32(%rdi)
440 movntdq %xmm7, 48(%rdi)
442 add $64, %rdi /* Advance pointer */
443 dec %rsi /* Decrement 64-byte block count */
445 mfence /* Ensure that the writes are globally visible, since
446 * MOVNTDQ is weakly ordered */
447 movups -16(%rsp), %xmm7 /* Restore the XMM register */
448 prefetcht0 0(%rcx) /* Prefetch the start of the block into cache,
449 * since it's likely to be used immediately. */
452 SIZE(GNAME(fast_bzero))