1 # x86/x86_64 support for -fsplit-stack.
2 # Copyright (C) 2009-2015 Free Software Foundation, Inc.
3 # Contributed by Ian Lance Taylor <iant@google.com>.
5 # This file is part of GCC.
7 # GCC is free software; you can redistribute it and/or modify it under
8 # the terms of the GNU General Public License as published by the Free
9 # Software Foundation; either version 3, or (at your option) any later
12 # GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 # WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 # FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 # Under Section 7 of GPL version 3, you are granted additional
18 # permissions described in the GCC Runtime Library Exception, version
19 # 3.1, as published by the Free Software Foundation.
21 # You should have received a copy of the GNU General Public License and
22 # a copy of the GCC Runtime Library Exception along with this program;
23 # see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
24 # <http://www.gnu.org/licenses/>.
27 # Support for allocating more stack space when using -fsplit-stack.
28 # When a function discovers that it needs more stack space, it will
29 # call __morestack with the size of the stack frame and the size of
30 # the parameters to copy from the old stack frame to the new one.
31 # The __morestack function preserves the parameter registers and
32 # calls __generic_morestack to actually allocate the stack space.
34 # When this is called stack space is very low, but we ensure that
35 # there is enough space to push the parameter registers and to call
36 # __generic_morestack.
38 # When calling __generic_morestack, FRAME_SIZE points to the size of
39 # the desired frame when the function is called, and the function
40 # sets it to the size of the allocated stack. OLD_STACK points to
41 # the parameters on the old stack and PARAM_SIZE is the number of
42 # bytes of parameters to copy to the new stack. These are the
43 # parameters of the function that called __morestack. The
44 # __generic_morestack function returns the new stack pointer,
45 # pointing to the address of the first copied parameter. The return
46 # value minus the returned *FRAME_SIZE will be the first address on
47 # the stack which we should not use.
49 # void *__generic_morestack (size_t *frame_size, void *old_stack,
52 # The __morestack routine has to arrange for the caller to return to a
53 # stub on the new stack. The stub is responsible for restoring the
54 # old stack pointer and returning to the caller's caller. This calls
55 # __generic_releasestack to retrieve the old stack pointer and release
56 # the newly allocated stack.
58 # void *__generic_releasestack (size_t *available);
60 # We do a little dance so that the processor's call/return return
61 # address prediction works out. The compiler arranges for the caller
63 # call __generic_morestack
66 # // carry on with function
67 # After we allocate more stack, we call L, which is in our caller.
68 # When that returns (to the predicted instruction), we release the
69 # stack segment and reset the stack pointer. We then return to the
70 # predicted instruction, namely the ret instruction immediately after
71 # the call to __generic_morestack. That then returns to the caller of
72 # the original caller.
75 # The amount of extra space we ask for. In general this has to be
76 # enough for the dynamic loader to find a symbol and for a signal
80 #define BACKOFF (1024)
82 #define BACKOFF (1536)
86 # The amount of space we ask for when calling non-split-stack code.
87 #define NON_SPLIT_STACK 0x100000
89 # This entry point is for split-stack code which calls non-split-stack
90 # code. When the linker sees this case, it converts the call to
91 # __morestack to call __morestack_non_split instead. We just bump the
92 # requested stack space by 16K.
94 .global __morestack_non_split
95 .hidden __morestack_non_split
98 .type __morestack_non_split,@function
101 __morestack_non_split:
106 # See below for an extended explanation of this.
109 pushl %eax # Save %eax in case it is a parameter.
111 .cfi_adjust_cfa_offset 4 # Account for pushed register.
113 movl %esp,%eax # Current stack,
114 subl 8(%esp),%eax # less required stack frame size,
115 subl $NON_SPLIT_STACK,%eax # less space for non-split code.
116 cmpl %gs:0x30,%eax # See if we have enough space.
117 jb 2f # Get more space if we need it.
120 # %esp + 20: stack pointer after two returns
121 # %esp + 16: return address of morestack caller's caller
122 # %esp + 12: size of parameters
123 # %esp + 8: new stack frame size
124 # %esp + 4: return address of this function
127 # Since we aren't doing a full split stack, we don't need to
128 # do anything when our caller returns. So we return to our
129 # caller rather than calling it, and let it return as usual.
130 # To make that work we adjust the return address.
132 # This breaks call/return address prediction for the call to
133 # this function. I can't figure out a way to make it work
134 # short of copying the parameters down the stack, which will
135 # probably take more clock cycles than we will lose breaking
136 # call/return address prediction. We will only break
137 # prediction for this call, not for our caller.
139 movl 4(%esp),%eax # Increment the return address
140 cmpb $0xc3,(%eax) # to skip the ret instruction;
145 # If the instruction that we return to is
146 # leal 20(%ebp),{%eax,%ecx,%edx}
147 # then we have been called by a varargs function that expects
148 # %ebp to hold a real value. That can only work if we do the
149 # full stack split routine. FIXME: This is fragile.
162 movl %eax,4(%esp) # Update return address.
164 popl %eax # Restore %eax and stack.
166 .cfi_adjust_cfa_offset -4 # Account for popped register.
168 ret $8 # Return to caller, popping args.
171 .cfi_adjust_cfa_offset 4 # Back to where we were.
173 popl %eax # Restore %eax and stack.
175 .cfi_adjust_cfa_offset -4 # Account for popped register.
177 # Increment space we request.
178 addl $NON_SPLIT_STACK+0x1000+BACKOFF,4(%esp)
180 # Fall through into morestack.
184 # See below for an extended explanation of this.
187 pushq %rax # Save %rax in case caller is using
188 # it to preserve original %r10.
189 .cfi_adjust_cfa_offset 8 # Adjust for pushed register.
191 movq %rsp,%rax # Current stack,
192 subq %r10,%rax # less required stack frame size,
193 subq $NON_SPLIT_STACK,%rax # less space for non-split code.
196 cmpq %fs:0x70,%rax # See if we have enough space.
201 jb 2f # Get more space if we need it.
203 # If the instruction that we return to is
204 # leaq 24(%rbp), %r11n
205 # then we have been called by a varargs function that expects
206 # %ebp to hold a real value. That can only work if we do the
207 # full stack split routine. FIXME: This is fragile.
209 incq %rax # Skip ret instruction in caller.
210 cmpl $0x185d8d4c,(%rax)
213 # This breaks call/return prediction, as described above.
214 incq 8(%rsp) # Increment the return address.
216 popq %rax # Restore register.
218 .cfi_adjust_cfa_offset -8 # Adjust for popped register.
220 ret # Return to caller.
223 popq %rax # Restore register.
225 .cfi_adjust_cfa_offset -8 # Adjust for popped register.
227 # Increment space we request.
228 addq $NON_SPLIT_STACK+0x1000+BACKOFF,%r10
230 # Fall through into morestack.
236 .size __morestack_non_split, . - __morestack_non_split
239 # __morestack_non_split falls through into __morestack.
242 # The __morestack function.
248 .type __morestack,@function
259 # The 32-bit __morestack function.
261 # We use a cleanup to restore the stack guard if an exception
262 # is thrown through this code.
264 .cfi_personality 0,__gcc_personality_v0
267 .cfi_personality 0x9b,DW.ref.__gcc_personality_v0
268 .cfi_lsda 0x1b,.LLSDA1
271 # We return below with a ret $8. We will return to a single
272 # return instruction, which will return to the caller of our
273 # caller. We let the unwinder skip that single return
274 # instruction, and just return to the real caller.
276 # Here CFA points just past the return address on the stack,
277 # e.g., on function entry it is %esp + 4. The stack looks
279 # CFA + 12: stack pointer after two returns
280 # CFA + 8: return address of morestack caller's caller
281 # CFA + 4: size of parameters
282 # CFA: new stack frame size
283 # CFA - 4: return address of this function
284 # CFA - 8: previous value of %ebp; %ebp points here
285 # Setting the new CFA to be the current CFA + 12 (i.e., %esp +
286 # 16) will make the unwinder pick up the right return address.
291 .cfi_adjust_cfa_offset 4
292 .cfi_offset %ebp, -20
294 .cfi_def_cfa_register %ebp
296 # In 32-bit mode the parameters are pushed on the stack. The
297 # argument size is pushed then the new stack frame size is
300 # In the body of a non-leaf function, the stack pointer will
301 # be aligned to a 16-byte boundary. That is CFA + 12 in the
302 # stack picture above: (CFA + 12) % 16 == 0. At this point we
303 # have %esp == CFA - 8, so %esp % 16 == 12. We need some
304 # space for saving registers and passing parameters, and we
305 # need to wind up with %esp % 16 == 0.
308 # Because our cleanup code may need to clobber %ebx, we need
309 # to save it here so the unwinder can restore the value used
310 # by the caller. Note that we don't have to restore the
311 # register, since we don't change it, we just have to save it
314 .cfi_offset %ebx, -24
316 # In 32-bit mode the registers %eax, %edx, and %ecx may be
317 # used for parameters, depending on the regparm and fastcall
324 call __morestack_block_signals
326 movl 12(%ebp),%eax # The size of the parameters.
328 leal 20(%ebp),%eax # Address of caller's parameters.
330 addl $BACKOFF,8(%ebp) # Ask for backoff bytes.
331 leal 8(%ebp),%eax # The address of the new frame size.
334 call __generic_morestack
336 movl %eax,%esp # Switch to the new stack.
337 subl 8(%ebp),%eax # The end of the stack space.
338 addl $BACKOFF,%eax # Back off 512 bytes.
341 # FIXME: The offset must match
342 # TARGET_THREAD_SPLIT_STACK_OFFSET in
343 # gcc/config/i386/linux.h.
344 movl %eax,%gs:0x30 # Save the new stack boundary.
346 call __morestack_unblock_signals
348 movl -12(%ebp),%edx # Restore registers.
351 movl 4(%ebp),%eax # Increment the return address
352 cmpb $0xc3,(%eax) # to skip the ret instruction;
357 movl %eax,-12(%ebp) # Store return address in an
360 movl -8(%ebp),%eax # Restore the last register.
362 call *-12(%ebp) # Call our caller!
364 # The caller will return here, as predicted.
366 # Save the registers which may hold a return value. We
367 # assume that __generic_releasestack does not touch any
368 # floating point or vector registers.
372 # Push the arguments to __generic_releasestack now so that the
373 # stack is at a 16-byte boundary for
374 # __morestack_block_signals.
375 pushl $0 # Where the available space is returned.
376 leal 0(%esp),%eax # Push its address.
379 call __morestack_block_signals
381 call __generic_releasestack
383 subl 4(%esp),%eax # Subtract available space.
384 addl $BACKOFF,%eax # Back off 512 bytes.
386 movl %eax,%gs:0x30 # Save the new stack boundary.
388 addl $8,%esp # Remove values from stack.
390 # We need to restore the old stack pointer, which is in %rbp,
391 # before we unblock signals. We also need to restore %eax and
392 # %edx after we unblock signals but before we return. Do this
393 # by moving %eax and %edx from the current stack to the old
396 popl %edx # Pop return value from current stack.
399 movl %ebp,%esp # Restore stack pointer.
401 # As before, we now have %esp % 16 == 12.
403 pushl %eax # Push return value on old stack.
405 subl $4,%esp # Align stack to 16-byte boundary.
407 call __morestack_unblock_signals
410 popl %edx # Restore return value.
415 # We never changed %ebx, so we don't have to actually restore it.
420 .cfi_def_cfa %esp, 16
421 ret $8 # Return to caller, which will
422 # immediately return. Pop
423 # arguments as we go.
425 # This is the cleanup code called by the stack unwinder when unwinding
426 # through the code between .LEHB0 and .LEHE0 above.
430 subl $16,%esp # Maintain 16 byte alignment.
431 movl %eax,4(%esp) # Save exception header.
432 movl %ebp,(%esp) # Stack pointer after resume.
433 call __generic_findstack
434 movl %ebp,%ecx # Get the stack pointer.
435 subl %eax,%ecx # Subtract available space.
436 addl $BACKOFF,%ecx # Back off 512 bytes.
437 movl %ecx,%gs:0x30 # Save new stack boundary.
438 movl 4(%esp),%eax # Function argument.
441 call __x86.get_pc_thunk.bx # %ebx may not be set up for us.
442 addl $_GLOBAL_OFFSET_TABLE_, %ebx
443 call _Unwind_Resume@PLT # Resume unwinding.
448 #else /* defined(__x86_64__) */
451 # The 64-bit __morestack function.
453 # We use a cleanup to restore the stack guard if an exception
454 # is thrown through this code.
456 .cfi_personality 0x3,__gcc_personality_v0
457 .cfi_lsda 0x3,.LLSDA1
459 .cfi_personality 0x9b,DW.ref.__gcc_personality_v0
460 .cfi_lsda 0x1b,.LLSDA1
463 # We will return a single return instruction, which will
464 # return to the caller of our caller. Let the unwinder skip
465 # that single return instruction, and just return to the real
469 # Set up a normal backtrace.
471 .cfi_adjust_cfa_offset 8
472 .cfi_offset %rbp, -24
474 .cfi_def_cfa_register %rbp
476 # In 64-bit mode the new stack frame size is passed in r10
477 # and the argument size is passed in r11.
479 addq $BACKOFF,%r10 # Ask for backoff bytes.
480 pushq %r10 # Save new frame size.
482 # In 64-bit mode the registers %rdi, %rsi, %rdx, %rcx, %r8,
483 # and %r9 may be used for parameters. We also preserve %rax
484 # which the caller may use to hold %r10.
496 # We entered morestack with the stack pointer aligned to a
497 # 16-byte boundary (the call to morestack's caller used 8
498 # bytes, and the call to morestack used 8 bytes). We have now
499 # pushed 10 registers, so we are still aligned to a 16-byte
502 call __morestack_block_signals
504 leaq -8(%rbp),%rdi # Address of new frame size.
505 leaq 24(%rbp),%rsi # The caller's parameters.
506 popq %rdx # The size of the parameters.
508 subq $8,%rsp # Align stack.
510 call __generic_morestack
512 movq -8(%rbp),%r10 # Reload modified frame size
513 movq %rax,%rsp # Switch to the new stack.
514 subq %r10,%rax # The end of the stack space.
515 addq $BACKOFF,%rax # Back off 1024 bytes.
518 # FIXME: The offset must match
519 # TARGET_THREAD_SPLIT_STACK_OFFSET in
520 # gcc/config/i386/linux64.h.
521 # Macro to save the new stack boundary.
523 #define X86_64_SAVE_NEW_STACK_BOUNDARY(reg) movq %r##reg,%fs:0x70
525 #define X86_64_SAVE_NEW_STACK_BOUNDARY(reg) movl %e##reg,%fs:0x40
527 X86_64_SAVE_NEW_STACK_BOUNDARY (ax)
529 call __morestack_unblock_signals
531 movq -24(%rbp),%rdi # Restore registers.
538 movq 8(%rbp),%r10 # Increment the return address
539 incq %r10 # to skip the ret instruction;
542 movq -16(%rbp),%rax # Restore caller's %rax.
544 call *%r10 # Call our caller!
546 # The caller will return here, as predicted.
548 # Save the registers which may hold a return value. We
549 # assume that __generic_releasestack does not touch any
550 # floating point or vector registers.
554 call __morestack_block_signals
556 pushq $0 # For alignment.
557 pushq $0 # Where the available space is returned.
558 leaq 0(%rsp),%rdi # Pass its address.
560 call __generic_releasestack
562 subq 0(%rsp),%rax # Subtract available space.
563 addq $BACKOFF,%rax # Back off 1024 bytes.
565 X86_64_SAVE_NEW_STACK_BOUNDARY (ax)
567 addq $16,%rsp # Remove values from stack.
569 # We need to restore the old stack pointer, which is in %rbp,
570 # before we unblock signals. We also need to restore %rax and
571 # %rdx after we unblock signals but before we return. Do this
572 # by moving %rax and %rdx from the current stack to the old
575 popq %rdx # Pop return value from current stack.
578 movq %rbp,%rsp # Restore stack pointer.
580 # Now (%rsp & 16) == 8.
582 subq $8,%rsp # For alignment.
583 pushq %rax # Push return value on old stack.
586 call __morestack_unblock_signals
588 popq %rdx # Restore return value.
595 .cfi_def_cfa %rsp, 16
596 ret # Return to caller, which will
597 # immediately return.
599 # This is the cleanup code called by the stack unwinder when unwinding
600 # through the code between .LEHB0 and .LEHE0 above.
604 subq $16,%rsp # Maintain 16 byte alignment.
605 movq %rax,(%rsp) # Save exception header.
606 movq %rbp,%rdi # Stack pointer after resume.
607 call __generic_findstack
608 movq %rbp,%rcx # Get the stack pointer.
609 subq %rax,%rcx # Subtract available space.
610 addq $BACKOFF,%rcx # Back off 1024 bytes.
611 X86_64_SAVE_NEW_STACK_BOUNDARY (cx)
612 movq (%rsp),%rdi # Restore exception data for call.
614 call _Unwind_Resume@PLT # Resume unwinding.
616 call _Unwind_Resume # Resume unwinding.
619 #endif /* defined(__x86_64__) */
623 .size __morestack, . - __morestack
626 #if !defined(__x86_64__) && defined(__PIC__)
627 # Output the thunk to get PC into bx, since we use it above.
628 .section .text.__x86.get_pc_thunk.bx,"axG",@progbits,__x86.get_pc_thunk.bx,comdat
629 .globl __x86.get_pc_thunk.bx
630 .hidden __x86.get_pc_thunk.bx
632 .type __x86.get_pc_thunk.bx, @function
634 __x86.get_pc_thunk.bx:
640 .size __x86.get_pc_thunk.bx, . - __x86.get_pc_thunk.bx
644 # The exception table. This tells the personality routine to execute
645 # the exception handler.
647 .section .gcc_except_table,"a",@progbits
650 .byte 0xff # @LPStart format (omit)
651 .byte 0xff # @TType format (omit)
652 .byte 0x1 # call-site format (uleb128)
653 .uleb128 .LLSDACSE1-.LLSDACSB1 # Call-site table length
655 .uleb128 .LEHB0-.LFB1 # region 0 start
656 .uleb128 .LEHE0-.LEHB0 # length
657 .uleb128 .L1-.LFB1 # landing pad
662 .global __gcc_personality_v0
664 # Build a position independent reference to the basic
665 # personality function.
666 .hidden DW.ref.__gcc_personality_v0
667 .weak DW.ref.__gcc_personality_v0
668 .section .data.DW.ref.__gcc_personality_v0,"awG",@progbits,DW.ref.__gcc_personality_v0,comdat
669 .type DW.ref.__gcc_personality_v0, @object
670 DW.ref.__gcc_personality_v0:
673 .size DW.ref.__gcc_personality_v0, 4
674 .long __gcc_personality_v0
677 .size DW.ref.__gcc_personality_v0, 8
678 .quad __gcc_personality_v0
682 #if defined __x86_64__ && defined __LP64__
684 # This entry point is used for the large model. With this entry point
685 # the upper 32 bits of %r10 hold the argument size and the lower 32
686 # bits hold the new stack frame size. There doesn't seem to be a way
687 # to know in the assembler code that we are assembling for the large
688 # model, and there doesn't seem to be a large model multilib anyhow.
689 # If one is developed, then the non-PIC code is probably OK since we
690 # will probably be close to the morestack code, but the PIC code
691 # almost certainly needs to be changed. FIXME.
694 .global __morestack_large_model
695 .hidden __morestack_large_model
698 .type __morestack_large_model,@function
701 __morestack_large_model:
706 andl $0xffffffff, %r10d
712 .size __morestack_large_model, . - __morestack_large_model
715 #endif /* __x86_64__ && __LP64__ */
717 # Initialize the stack test value when the program starts or when a
718 # new thread starts. We don't know how large the main stack is, so we
719 # guess conservatively. We might be able to use getrlimit here.
722 .global __stack_split_initialize
723 .hidden __stack_split_initialize
726 .type __stack_split_initialize, @function
729 __stack_split_initialize:
733 leal -16000(%esp),%eax # We should have at least 16K.
738 call __generic_morestack_set_initial_sp@PLT
740 call __generic_morestack_set_initial_sp
745 #else /* defined(__x86_64__) */
747 leaq -16000(%rsp),%rax # We should have at least 16K.
748 X86_64_SAVE_NEW_STACK_BOUNDARY (ax)
752 call __generic_morestack_set_initial_sp@PLT
754 call __generic_morestack_set_initial_sp
758 #endif /* defined(__x86_64__) */
761 .size __stack_split_initialize, . - __stack_split_initialize
764 # Routines to get and set the guard, for __splitstack_getcontext,
765 # __splitstack_setcontext, and __splitstack_makecontext.
767 # void *__morestack_get_guard (void) returns the current stack guard.
769 .global __morestack_get_guard
770 .hidden __morestack_get_guard
773 .type __morestack_get_guard,@function
776 __morestack_get_guard:
790 .size __morestack_get_guard, . - __morestack_get_guard
793 # void __morestack_set_guard (void *) sets the stack guard.
794 .global __morestack_set_guard
795 .hidden __morestack_set_guard
798 .type __morestack_set_guard,@function
801 __morestack_set_guard:
807 X86_64_SAVE_NEW_STACK_BOUNDARY (di)
812 .size __morestack_set_guard, . - __morestack_set_guard
815 # void *__morestack_make_guard (void *, size_t) returns the stack
816 # guard value for a stack.
817 .global __morestack_make_guard
818 .hidden __morestack_make_guard
821 .type __morestack_make_guard,@function
824 __morestack_make_guard:
838 .size __morestack_make_guard, . - __morestack_make_guard
841 # Make __stack_split_initialize a high priority constructor. FIXME:
842 # This is ELF specific.
844 .section .ctors.65535,"aw",@progbits
848 .long __stack_split_initialize
849 .long __morestack_load_mmap
852 .quad __stack_split_initialize
853 .quad __morestack_load_mmap
857 .section .note.GNU-stack,"",@progbits
858 .section .note.GNU-split-stack,"",@progbits
859 .section .note.GNU-no-split-stack,"",@progbits