-##--------------------------------------------------------------------##
-##--- The core dispatch loop, for jumping to a code address. ---##
-##--- dispatch-x86.S ---##
-##--------------------------------------------------------------------##
+/*--------------------------------------------------------------------*/
+/*--- The core dispatch loop, for jumping to a code address. ---*/
+/*--- dispatch-x86.S ---*/
+/*--------------------------------------------------------------------*/
/*
This file is part of Valgrind, a dynamic binary instrumentation
/*--- The dispatch loop. ---*/
/*------------------------------------------------------------*/
-/* signature: UWord VG_(run_innerloop) ( void* guest_state ) */
+/*----------------------------------------------------*/
+/*--- Preamble (set everything up) ---*/
+/*----------------------------------------------------*/
+/* signature:
+UWord VG_(run_innerloop) ( void* guest_state, UWord do_profiling );
+*/
+.text
.globl VG_(run_innerloop)
VG_(run_innerloop):
/* 4(%esp) holds guest_state */
+ /* 8(%esp) holds do_profiling */
/* ----- entry point to VG_(run_innerloop) ----- */
pushl %ebx
pushl %ebp
/* 28(%esp) holds guest_state */
+ /* 32(%esp) holds do_profiling */
/* Set up the guest state pointer */
movl 28(%esp), %ebp
/* set dir flag to known value */
cld
- /* fall into main loop */
+ /* fall into main loop (the right one) */
+ cmpl $0, 32(%esp) /* do_profiling */
+ je VG_(run_innerloop__dispatch_unprofiled)
+ jmp VG_(run_innerloop__dispatch_profiled)
+ /*NOTREACHED*/
+
+/*----------------------------------------------------*/
+/*--- NO-PROFILING (standard) dispatcher ---*/
+/*----------------------------------------------------*/
+
+.align 16
+.global VG_(run_innerloop__dispatch_unprofiled)
+VG_(run_innerloop__dispatch_unprofiled):
+ /* AT ENTRY: %eax is next guest addr, %ebp is possibly
+ modified guest state ptr */
+
+ /* Has the guest state pointer been messed with? If yes, exit. */
+ cmpl 28(%esp), %ebp
+ jnz gsp_changed
+
+ /* save the jump address in the guest state */
+ movl %eax, OFFSET_x86_EIP(%ebp)
- /* Here, %eax is the only live (real) register. The entire
- simulated state is saved in the ThreadState. */
+ /* Are we out of timeslice? If yes, defer to scheduler. */
+ subl $1, VG_(dispatch_ctr)
+ jz counter_is_zero
+
+ /* try a fast lookup in the translation cache */
+ movl %eax, %ebx
+ andl $VG_TT_FAST_MASK, %ebx
+ movl VG_(tt_fast)(,%ebx,4), %ecx
+ cmpl %eax, (%ecx)
+ jnz fast_lookup_failed
+
+ /* Found a match. Jump to tce[1], which is 8 bytes along,
+ since each tce element is a 64-bit int. */
+ addl $8, %ecx
+ jmp *%ecx
+ ud2 /* persuade insn decoders not to speculate past here */
+ /* generated code should run, then jump back to
+ VG_(run_innerloop__dispatch_unprofiled). */
+ /*NOTREACHED*/
+
+/*----------------------------------------------------*/
+/*--- PROFILING dispatcher (can be much slower) ---*/
+/*----------------------------------------------------*/
+
+.align 16
+.global VG_(run_innerloop__dispatch_profiled)
+VG_(run_innerloop__dispatch_profiled):
+ /* AT ENTRY: %eax is next guest addr, %ebp is possibly
+ modified guest state ptr */
+
+ /* Has the guest state pointer been messed with? If yes, exit. */
+ cmpl 28(%esp), %ebp
+ jnz gsp_changed
-dispatch_boring:
/* save the jump address in the guest state */
movl %eax, OFFSET_x86_EIP(%ebp)
/* Are we out of timeslice? If yes, defer to scheduler. */
- subl $1, VG_(dispatch_ctr)
+ subl $1, VG_(dispatch_ctr)
jz counter_is_zero
/* try a fast lookup in the translation cache */
- movl %eax, %ebx
- andl $VG_TT_FAST_MASK, %ebx
- movl VG_(tt_fast)(,%ebx,4), %ecx
- cmpl %eax, (%ecx)
- jnz fast_lookup_failed
+ movl %eax, %ebx
+ andl $VG_TT_FAST_MASK, %ebx
+ movl VG_(tt_fast)(,%ebx,4), %ecx
+ cmpl %eax, (%ecx)
+ jnz fast_lookup_failed
/* increment bb profile counter */
- movl VG_(tt_fastN)(,%ebx,4), %edx
+ /* note: innocuous as this sounds, it causes a huge amount more
+ stress on D1 and significantly slows everything down. */
+ movl VG_(tt_fastN)(,%ebx,4), %edx
/* Use "addl $1", not "incl", to avoid partial-flags stall on P4 */
- addl $1, (%edx)
+ addl $1, (%edx)
- /* Found a match. Call tce[1], which is 8 bytes along, since
- each tce element is a 64-bit int. */
+ /* Found a match. Jump to tce[1], which is 8 bytes along,
+ since each tce element is a 64-bit int. */
addl $8, %ecx
- call *%ecx
-
- /*
- %eax holds destination (original) address.
- %ebp indicates further details of the control transfer
- requested to the address in %eax.
-
- If ebp is unchanged (== * 28(%esp)), just jump next to %eax.
+ jmp *%ecx
+ ud2 /* persuade insn decoders not to speculate past here */
+ /* generated code should run, then jump back to
+ VG_(run_innerloop__dispatch_profiled). */
+ /*NOTREACHED*/
+
+/*----------------------------------------------------*/
+/*--- exit points ---*/
+/*----------------------------------------------------*/
+
+gsp_changed:
+ /* Someone messed with the gsp. Have to
+ defer to scheduler to resolve this. dispatch ctr
+ is not yet decremented, so no need to increment. */
+ /* %EIP is NOT up to date here. First, need to write
+ %eax back to %EIP, but without trashing %ebp since
+ that holds the value we want to return to the scheduler.
+ Hence use %esi transiently for the guest state pointer. */
+ movl 28(%esp), %esi
+ movl %eax, OFFSET_x86_EIP(%esi)
+ movl %ebp, %eax
+ jmp run_innerloop_exit
+ /*NOTREACHED*/
- Otherwise fall out, back to the scheduler, and let it
- figure out what to do next.
- */
+counter_is_zero:
+ /* %EIP is up to date here */
+ /* back out decrement of the dispatch counter */
+ addl $1, VG_(dispatch_ctr)
+ movl $VG_TRC_INNER_COUNTERZERO, %eax
+ jmp run_innerloop_exit
+ /*NOTREACHED*/
- cmpl 28(%esp), %ebp
- jz dispatch_boring
+fast_lookup_failed:
+ /* %EIP is up to date here */
+ /* back out decrement of the dispatch counter */
+ addl $1, VG_(dispatch_ctr)
+ movl $VG_TRC_INNER_FASTMISS, %eax
+ jmp run_innerloop_exit
+ /*NOTREACHED*/
- jmp dispatch_exceptional
-
/* All exits from the dispatcher go through here. %eax holds
the return value.
popl %ebx
ret
-
-
-/* Other ways of getting out of the inner loop. Placed out-of-line to
- make it look cleaner.
-*/
-dispatch_exceptional:
- /* this is jumped to only, not fallen-through from above */
-
- /* save %eax in %EIP and defer to sched */
- movl 28(%esp), %edi
- movl %eax, OFFSET_x86_EIP(%edi)
- movl %ebp, %eax
- jmp run_innerloop_exit
-
-fast_lookup_failed:
- /* %EIP is up to date here since dispatch_boring dominates */
- addl $1, VG_(dispatch_ctr)
- movl $VG_TRC_INNER_FASTMISS, %eax
- jmp run_innerloop_exit
-
-counter_is_zero:
- /* %EIP is up to date here since dispatch_boring dominates */
- addl $1, VG_(dispatch_ctr)
- movl $VG_TRC_INNER_COUNTERZERO, %eax
- jmp run_innerloop_exit
-
/* Let the linker know we don't need an executable stack */
.section .note.GNU-stack,"",@progbits
-##--------------------------------------------------------------------##
-##--- end ---##
-##--------------------------------------------------------------------##
+/*--------------------------------------------------------------------*/
+/*--- end ---*/
+/*--------------------------------------------------------------------*/
#include "pub_core_basics.h"
#include "pub_core_aspacemgr.h"
-#include "pub_core_machine.h" // For VG_(machine_get_VexArchInfo)
- // and VG_(get_SP)
+#include "pub_core_machine.h" // For VG_(machine_get_VexArchInfo)
+ // and VG_(get_SP)
#include "pub_core_libcbase.h"
#include "pub_core_libcassert.h"
#include "pub_core_libcprint.h"
#include "pub_core_options.h"
#include "pub_core_profile.h"
-#include "pub_core_debuginfo.h" // Needed for pub_core_redir :(
-#include "pub_core_redir.h" // For VG_(code_redirect)()
+#include "pub_core_debuginfo.h" // Needed for pub_core_redir :(
+#include "pub_core_redir.h" // For VG_(code_redirect)()
-#include "pub_core_signals.h" // For VG_(synth_fault_{perms,mapping})()
-#include "pub_core_stacks.h" // For VG_(unknown_SP_update)()
-#include "pub_core_tooliface.h" // For VG_(tdict)
+#include "pub_core_signals.h" // For VG_(synth_fault_{perms,mapping})()
+#include "pub_core_stacks.h" // For VG_(unknown_SP_update)()
+#include "pub_core_tooliface.h" // For VG_(tdict)
#include "pub_core_translate.h"
#include "pub_core_transtab.h"
+#include "pub_core_dispatch.h" // VG_(run_innerloop__dispatch_{un}profiled)
+
/*------------------------------------------------------------*/
/*--- Stats ---*/
VexArch vex_arch;
VexArchInfo vex_archinfo;
VexGuestExtents vge;
+ VexTranslateArgs vta;
VexTranslateResult tres;
/* Make sure Vex is initialised right. */
/* Set up closure arg for "chase_into_ok" */
chase_into_ok__CLOSURE_tid = tid;
- tres = LibVEX_Translate (
- vex_arch, &vex_archinfo,
- vex_arch, &vex_archinfo,
- (UChar*)ULong_to_Ptr(orig_addr),
- (Addr64)orig_addr,
- (Addr64)orig_addr_noredir,
- chase_into_ok,
- &vge,
- tmpbuf, N_TMPBUF, &tmpbuf_used,
- VG_(tdict).tool_instrument,
- need_to_handle_SP_assignment()
- ? vg_SP_update_pass
- : NULL,
- True, /* cleanup after instrumentation */
- do_self_check,
- NULL,
- verbosity
- );
+ vta.arch_guest = vex_arch;
+ vta.archinfo_guest = vex_archinfo;
+ vta.arch_host = vex_arch;
+ vta.archinfo_host = vex_archinfo;
+ vta.guest_bytes = (UChar*)ULong_to_Ptr(orig_addr);
+ vta.guest_bytes_addr = (Addr64)orig_addr;
+ vta.guest_bytes_addr_noredir = (Addr64)orig_addr_noredir;
+ vta.chase_into_ok = chase_into_ok;
+ vta.guest_extents = &vge;
+ vta.host_bytes = tmpbuf;
+ vta.host_bytes_size = N_TMPBUF;
+ vta.host_bytes_used = &tmpbuf_used;
+ vta.instrument1 = VG_(tdict).tool_instrument;
+ vta.instrument2 = need_to_handle_SP_assignment()
+ ? vg_SP_update_pass
+ : NULL;
+ vta.do_self_check = do_self_check;
+ vta.traceflags = verbosity;
+
+ /* Set up the dispatch-return info. For archs without a link
+ register, vex generates a jump back to the specified dispatch
+ address. Else, it just generates a branch-to-LR. */
+# if defined(VGA_x86) || defined(VGA_amd64)
+ vta.dispatch = VG_(clo_profile_flags) > 0
+ ? (void*) &VG_(run_innerloop__dispatch_profiled)
+ : (void*) &VG_(run_innerloop__dispatch_unprofiled);
+# elif defined(VGA_ppc32) || defined(VGA_ppc64)
+ vta.dispatch = NULL;
+# else
+# error "Unknown arch"
+# endif
+
+ /* Sheesh. Finally, actually _do_ the translation! */
+ tres = LibVEX_Translate ( &vta );
vg_assert(tres == VexTransOK);
vg_assert(tmpbuf_used <= N_TMPBUF);
projects / thirdparty / valgrind.git / commitdiff
