Text file src/runtime/asm_amd64.s

     1  // Copyright 2009 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  #include "go_asm.h"
     6  #include "go_tls.h"
     7  #include "funcdata.h"
     8  #include "textflag.h"
     9  #include "cgo/abi_amd64.h"
    10  
    11  // _rt0_amd64 is common startup code for most amd64 systems when using
    12  // internal linking. This is the entry point for the program from the
    13  // kernel for an ordinary -buildmode=exe program. The stack holds the
    14  // number of arguments and the C-style argv.
    15  TEXT _rt0_amd64(SB),NOSPLIT,$-8
    16  	MOVQ	0(SP), DI	// argc
    17  	LEAQ	8(SP), SI	// argv
    18  	JMP	runtime·rt0_go(SB)
    19  
    20  // main is common startup code for most amd64 systems when using
    21  // external linking. The C startup code will call the symbol "main"
    22  // passing argc and argv in the usual C ABI registers DI and SI.
    23  TEXT main(SB),NOSPLIT,$-8
    24  	JMP	runtime·rt0_go(SB)
    25  
    26  // _rt0_amd64_lib is common startup code for most amd64 systems when
    27  // using -buildmode=c-archive or -buildmode=c-shared. The linker will
    28  // arrange to invoke this function as a global constructor (for
    29  // c-archive) or when the shared library is loaded (for c-shared).
    30  // We expect argc and argv to be passed in the usual C ABI registers
    31  // DI and SI.
    32  TEXT _rt0_amd64_lib(SB),NOSPLIT|NOFRAME,$0
    33  	// Transition from C ABI to Go ABI.
    34  	PUSH_REGS_HOST_TO_ABI0()
    35  
    36  	MOVQ	DI, _rt0_amd64_lib_argc<>(SB)
    37  	MOVQ	SI, _rt0_amd64_lib_argv<>(SB)
    38  
    39  #ifdef GOOS_windows
    40  	// Set up a dummy TLS value on Windows so that the autogenerated
    41  	// ABI wrappers don't crash when trying to load G from TLS before
    42  	// wintls has set up the real TLS slot in rt0_go.
    43  	MOVQ	$zeroTLS<>(SB), DI
    44  	CALL	runtime·settls(SB)
    45  #endif
    46  
    47  	CALL	runtime·libInit(SB)
    48  
    49  	POP_REGS_HOST_TO_ABI0()
    50  	RET
    51  
    52  // rt0_lib_go initializes the Go runtime.
    53  // This is started in a separate thread by _rt0_amd64_lib.
    54  TEXT runtime·rt0_lib_go<ABIInternal>(SB),NOSPLIT,$0
    55  	MOVQ	_rt0_amd64_lib_argc<>(SB), DI
    56  	MOVQ	_rt0_amd64_lib_argv<>(SB), SI
    57  	JMP	runtime·rt0_go(SB)
    58  
    59  DATA _rt0_amd64_lib_argc<>(SB)/8, $0
    60  GLOBL _rt0_amd64_lib_argc<>(SB),NOPTR, $8
    61  DATA _rt0_amd64_lib_argv<>(SB)/8, $0
    62  GLOBL _rt0_amd64_lib_argv<>(SB),NOPTR, $8
    63  
    64  #ifdef GOAMD64_v2
    65  DATA bad_cpu_msg<>+0x00(SB)/84, $"This program can only be run on AMD64 processors with v2 microarchitecture support.\n"
    66  #endif
    67  
    68  #ifdef GOAMD64_v3
    69  DATA bad_cpu_msg<>+0x00(SB)/84, $"This program can only be run on AMD64 processors with v3 microarchitecture support.\n"
    70  #endif
    71  
    72  #ifdef GOAMD64_v4
    73  DATA bad_cpu_msg<>+0x00(SB)/84, $"This program can only be run on AMD64 processors with v4 microarchitecture support.\n"
    74  #endif
    75  
    76  GLOBL bad_cpu_msg<>(SB), RODATA, $84
    77  
    78  // Define a list of AMD64 microarchitecture level features
    79  // https://en.wikipedia.org/wiki/X86-64#Microarchitecture_levels
    80  
    81                       // SSE3     SSSE3    CMPXCHNG16 SSE4.1    SSE4.2    POPCNT
    82  #define V2_FEATURES_CX (1 << 0 | 1 << 9 | 1 << 13  | 1 << 19 | 1 << 20 | 1 << 23)
    83                           // LAHF/SAHF
    84  #define V2_EXT_FEATURES_CX (1 << 0)
    85                                        // FMA       MOVBE     OSXSAVE   AVX       F16C
    86  #define V3_FEATURES_CX (V2_FEATURES_CX | 1 << 12 | 1 << 22 | 1 << 27 | 1 << 28 | 1 << 29)
    87                                                // ABM (FOR LZNCT)
    88  #define V3_EXT_FEATURES_CX (V2_EXT_FEATURES_CX | 1 << 5)
    89                           // BMI1     AVX2     BMI2
    90  #define V3_EXT_FEATURES_BX (1 << 3 | 1 << 5 | 1 << 8)
    91                         // XMM      YMM
    92  #define V3_OS_SUPPORT_AX (1 << 1 | 1 << 2)
    93  
    94  #define V4_FEATURES_CX V3_FEATURES_CX
    95  
    96  #define V4_EXT_FEATURES_CX V3_EXT_FEATURES_CX
    97                                                // AVX512F   AVX512DQ  AVX512CD  AVX512BW  AVX512VL
    98  #define V4_EXT_FEATURES_BX (V3_EXT_FEATURES_BX | 1 << 16 | 1 << 17 | 1 << 28 | 1 << 30 | 1 << 31)
    99                                            // OPMASK   ZMM
   100  #define V4_OS_SUPPORT_AX (V3_OS_SUPPORT_AX | 1 << 5 | (1 << 6 | 1 << 7))
   101  
   102  #ifdef GOAMD64_v2
   103  #define NEED_MAX_CPUID 0x80000001
   104  #define NEED_FEATURES_CX V2_FEATURES_CX
   105  #define NEED_EXT_FEATURES_CX V2_EXT_FEATURES_CX
   106  #endif
   107  
   108  #ifdef GOAMD64_v3
   109  #define NEED_MAX_CPUID 0x80000001
   110  #define NEED_FEATURES_CX V3_FEATURES_CX
   111  #define NEED_EXT_FEATURES_CX V3_EXT_FEATURES_CX
   112  #define NEED_EXT_FEATURES_BX V3_EXT_FEATURES_BX
   113  #define NEED_OS_SUPPORT_AX V3_OS_SUPPORT_AX
   114  #endif
   115  
   116  #ifdef GOAMD64_v4
   117  #define NEED_MAX_CPUID 0x80000001
   118  #define NEED_FEATURES_CX V4_FEATURES_CX
   119  #define NEED_EXT_FEATURES_CX V4_EXT_FEATURES_CX
   120  #define NEED_EXT_FEATURES_BX V4_EXT_FEATURES_BX
   121  
   122  // Darwin requires a different approach to check AVX512 support, see CL 285572.
   123  #ifdef GOOS_darwin
   124  #define NEED_OS_SUPPORT_AX V3_OS_SUPPORT_AX
   125  // These values are from:
   126  // https://github.com/apple/darwin-xnu/blob/xnu-4570.1.46/osfmk/i386/cpu_capabilities.h
   127  #define commpage64_base_address         0x00007fffffe00000
   128  #define commpage64_cpu_capabilities64   (commpage64_base_address+0x010)
   129  #define commpage64_version              (commpage64_base_address+0x01E)
   130  #define AVX512F                         0x0000004000000000
   131  #define AVX512CD                        0x0000008000000000
   132  #define AVX512DQ                        0x0000010000000000
   133  #define AVX512BW                        0x0000020000000000
   134  #define AVX512VL                        0x0000100000000000
   135  #define NEED_DARWIN_SUPPORT             (AVX512F | AVX512DQ | AVX512CD | AVX512BW | AVX512VL)
   136  #else
   137  #define NEED_OS_SUPPORT_AX V4_OS_SUPPORT_AX
   138  #endif
   139  
   140  #endif
   141  
   142  TEXT runtime·rt0_go(SB),NOSPLIT|NOFRAME|TOPFRAME,$0
   143  	// copy arguments forward on an even stack
   144  	MOVQ	DI, AX		// argc
   145  	MOVQ	SI, BX		// argv
   146  	SUBQ	$(5*8), SP		// 3args 2auto
   147  	ANDQ	$~15, SP
   148  	MOVQ	AX, 24(SP)
   149  	MOVQ	BX, 32(SP)
   150  
   151  	// This is typically the entry point for Go programs.
   152  	// Call stack unwinding must not proceed past this frame.
   153  	// Set the frame pointer register to 0 so that frame pointer-based unwinders
   154  	// (which don't use debug info for performance reasons)
   155  	// won't attempt to unwind past this function.
   156  	// See go.dev/issue/63630
   157  	MOVQ	$0, BP
   158  
   159  	// create istack out of the given (operating system) stack.
   160  	// _cgo_init may update stackguard.
   161  	MOVQ	$runtime·g0(SB), DI
   162  	LEAQ	(-64*1024)(SP), BX
   163  	MOVQ	BX, g_stackguard0(DI)
   164  	MOVQ	BX, g_stackguard1(DI)
   165  	MOVQ	BX, (g_stack+stack_lo)(DI)
   166  	MOVQ	SP, (g_stack+stack_hi)(DI)
   167  
   168  	// find out information about the processor we're on
   169  	MOVL	$0, AX
   170  	CPUID
   171  	CMPL	AX, $0
   172  	JE	nocpuinfo
   173  
   174  	CMPL	BX, $0x756E6547  // "Genu"
   175  	JNE	notintel
   176  	CMPL	DX, $0x49656E69  // "ineI"
   177  	JNE	notintel
   178  	CMPL	CX, $0x6C65746E  // "ntel"
   179  	JNE	notintel
   180  	MOVB	$1, runtime·isIntel(SB)
   181  
   182  notintel:
   183  	// Load EAX=1 cpuid flags
   184  	MOVL	$1, AX
   185  	CPUID
   186  	MOVL	AX, runtime·processorVersionInfo(SB)
   187  
   188  nocpuinfo:
   189  	// if there is an _cgo_init, call it.
   190  	MOVQ	_cgo_init(SB), AX
   191  	TESTQ	AX, AX
   192  	JZ	needtls
   193  	// arg 1: g0, already in DI
   194  	MOVQ	$setg_gcc<>(SB), SI // arg 2: setg_gcc
   195  	MOVQ	$0, DX	// arg 3, 4: not used when using platform's TLS
   196  	MOVQ	$0, CX
   197  #ifdef GOOS_android
   198  	MOVQ	$runtime·tls_g(SB), DX 	// arg 3: &tls_g
   199  	// arg 4: TLS base, stored in slot 0 (Android's TLS_SLOT_SELF).
   200  	// Compensate for tls_g (+16).
   201  	MOVQ	-16(TLS), CX
   202  #endif
   203  #ifdef GOOS_windows
   204  	MOVQ	$runtime·tls_g(SB), DX 	// arg 3: &tls_g
   205  	// Adjust for the Win64 calling convention.
   206  	MOVQ	CX, R9 // arg 4
   207  	MOVQ	DX, R8 // arg 3
   208  	MOVQ	SI, DX // arg 2
   209  	MOVQ	DI, CX // arg 1
   210  #endif
   211  	CALL	AX
   212  
   213  	// update stackguard after _cgo_init
   214  	MOVQ	$runtime·g0(SB), CX
   215  	MOVQ	(g_stack+stack_lo)(CX), AX
   216  	ADDQ	$const_stackGuard, AX
   217  	MOVQ	AX, g_stackguard0(CX)
   218  	MOVQ	AX, g_stackguard1(CX)
   219  
   220  #ifndef GOOS_windows
   221  	JMP ok
   222  #endif
   223  needtls:
   224  #ifdef GOOS_plan9
   225  	// skip TLS setup on Plan 9
   226  	JMP ok
   227  #endif
   228  #ifdef GOOS_solaris
   229  	// skip TLS setup on Solaris
   230  	JMP ok
   231  #endif
   232  #ifdef GOOS_illumos
   233  	// skip TLS setup on illumos
   234  	JMP ok
   235  #endif
   236  #ifdef GOOS_darwin
   237  	// skip TLS setup on Darwin
   238  	JMP ok
   239  #endif
   240  #ifdef GOOS_openbsd
   241  	// skip TLS setup on OpenBSD
   242  	JMP ok
   243  #endif
   244  
   245  #ifdef GOOS_windows
   246  	CALL	runtime·wintls(SB)
   247  #endif
   248  
   249  	LEAQ	runtime·m0+m_tls(SB), DI
   250  	CALL	runtime·settls(SB)
   251  
   252  	// store through it, to make sure it works
   253  	get_tls(BX)
   254  	MOVQ	$0x123, g(BX)
   255  	MOVQ	runtime·m0+m_tls(SB), AX
   256  	CMPQ	AX, $0x123
   257  	JEQ 2(PC)
   258  	CALL	runtime·abort(SB)
   259  ok:
   260  	// set the per-goroutine and per-mach "registers"
   261  	get_tls(BX)
   262  	LEAQ	runtime·g0(SB), CX
   263  	MOVQ	CX, g(BX)
   264  	LEAQ	runtime·m0(SB), AX
   265  
   266  	// save m->g0 = g0
   267  	MOVQ	CX, m_g0(AX)
   268  	// save m0 to g0->m
   269  	MOVQ	AX, g_m(CX)
   270  
   271  	CLD				// convention is D is always left cleared
   272  
   273  	// Check GOAMD64 requirements
   274  	// We need to do this after setting up TLS, so that
   275  	// we can report an error if there is a failure. See issue 49586.
   276  #ifdef NEED_FEATURES_CX
   277  	MOVL	$0, AX
   278  	CPUID
   279  	CMPL	AX, $0
   280  	JE	bad_cpu
   281  	MOVL	$1, AX
   282  	CPUID
   283  	ANDL	$NEED_FEATURES_CX, CX
   284  	CMPL	CX, $NEED_FEATURES_CX
   285  	JNE	bad_cpu
   286  #endif
   287  
   288  #ifdef NEED_MAX_CPUID
   289  	MOVL	$0x80000000, AX
   290  	CPUID
   291  	CMPL	AX, $NEED_MAX_CPUID
   292  	JL	bad_cpu
   293  #endif
   294  
   295  #ifdef NEED_EXT_FEATURES_BX
   296  	MOVL	$7, AX
   297  	MOVL	$0, CX
   298  	CPUID
   299  	ANDL	$NEED_EXT_FEATURES_BX, BX
   300  	CMPL	BX, $NEED_EXT_FEATURES_BX
   301  	JNE	bad_cpu
   302  #endif
   303  
   304  #ifdef NEED_EXT_FEATURES_CX
   305  	MOVL	$0x80000001, AX
   306  	CPUID
   307  	ANDL	$NEED_EXT_FEATURES_CX, CX
   308  	CMPL	CX, $NEED_EXT_FEATURES_CX
   309  	JNE	bad_cpu
   310  #endif
   311  
   312  #ifdef NEED_OS_SUPPORT_AX
   313  	XORL    CX, CX
   314  	XGETBV
   315  	ANDL	$NEED_OS_SUPPORT_AX, AX
   316  	CMPL	AX, $NEED_OS_SUPPORT_AX
   317  	JNE	bad_cpu
   318  #endif
   319  
   320  #ifdef NEED_DARWIN_SUPPORT
   321  	MOVQ	$commpage64_version, BX
   322  	CMPW	(BX), $13  // cpu_capabilities64 undefined in versions < 13
   323  	JL	bad_cpu
   324  	MOVQ	$commpage64_cpu_capabilities64, BX
   325  	MOVQ	(BX), BX
   326  	MOVQ	$NEED_DARWIN_SUPPORT, CX
   327  	ANDQ	CX, BX
   328  	CMPQ	BX, CX
   329  	JNE	bad_cpu
   330  #endif
   331  
   332  	CALL	runtime·check(SB)
   333  
   334  	MOVL	24(SP), AX		// copy argc
   335  	MOVL	AX, 0(SP)
   336  	MOVQ	32(SP), AX		// copy argv
   337  	MOVQ	AX, 8(SP)
   338  	CALL	runtime·args(SB)
   339  	CALL	runtime·osinit(SB)
   340  	CALL	runtime·schedinit(SB)
   341  
   342  	// create a new goroutine to start program
   343  	MOVQ	$runtime·mainPC(SB), AX		// entry
   344  	PUSHQ	AX
   345  	CALL	runtime·newproc(SB)
   346  	POPQ	AX
   347  
   348  	// start this M
   349  	CALL	runtime·mstart(SB)
   350  
   351  	CALL	runtime·abort(SB)	// mstart should never return
   352  	RET
   353  
   354  bad_cpu: // show that the program requires a certain microarchitecture level.
   355  	MOVQ	$2, 0(SP)
   356  	MOVQ	$bad_cpu_msg<>(SB), AX
   357  	MOVQ	AX, 8(SP)
   358  	MOVQ	$84, 16(SP)
   359  	CALL	runtime·write(SB)
   360  	MOVQ	$1, 0(SP)
   361  	CALL	runtime·exit(SB)
   362  	CALL	runtime·abort(SB)
   363  	RET
   364  
   365  	// Prevent dead-code elimination of debugCallV2 and debugPinnerV1, which are
   366  	// intended to be called by debuggers.
   367  	MOVQ	$runtime·debugPinnerV1<ABIInternal>(SB), AX
   368  	MOVQ	$runtime·debugCallV2<ABIInternal>(SB), AX
   369  	RET
   370  
   371  // mainPC is a function value for runtime.main, to be passed to newproc.
   372  // The reference to runtime.main is made via ABIInternal, since the
   373  // actual function (not the ABI0 wrapper) is needed by newproc.
   374  DATA	runtime·mainPC+0(SB)/8,$runtime·main<ABIInternal>(SB)
   375  GLOBL	runtime·mainPC(SB),RODATA,$8
   376  
   377  TEXT runtime·breakpoint(SB),NOSPLIT,$0-0
   378  	BYTE	$0xcc
   379  	RET
   380  
   381  TEXT runtime·asminit(SB),NOSPLIT,$0-0
   382  	// No per-thread init.
   383  	RET
   384  
   385  TEXT runtime·mstart(SB),NOSPLIT|TOPFRAME|NOFRAME,$0
   386  	// This is the root frame of new Go-created OS threads.
   387  	// Call stack unwinding must not proceed past this frame.
   388  	// Set the frame pointer register to 0 so that frame pointer-based unwinders
   389  	// (which don't use debug info for performance reasons)
   390  	// won't attempt to unwind past this function.
   391  	// See go.dev/issue/63630
   392  	MOVD	$0, BP
   393  	CALL	runtime·mstart0(SB)
   394  	RET // not reached
   395  
   396  /*
   397   *  go-routine
   398   */
   399  
   400  // func gogo(buf *gobuf)
   401  // restore state from Gobuf; longjmp
   402  TEXT runtime·gogo(SB), NOSPLIT, $0-8
   403  	MOVQ	buf+0(FP), BX		// gobuf
   404  	MOVQ	gobuf_g(BX), DX
   405  	MOVQ	0(DX), CX		// make sure g != nil
   406  	JMP	gogo<>(SB)
   407  
   408  TEXT gogo<>(SB), NOSPLIT, $0
   409  	get_tls(CX)
   410  	MOVQ	DX, g(CX)
   411  	MOVQ	DX, R14		// set the g register
   412  	MOVQ	gobuf_sp(BX), SP	// restore SP
   413  	MOVQ	gobuf_ctxt(BX), DX
   414  	MOVQ	gobuf_bp(BX), BP
   415  	MOVQ	$0, gobuf_sp(BX)	// clear to help garbage collector
   416  	MOVQ	$0, gobuf_ctxt(BX)
   417  	MOVQ	$0, gobuf_bp(BX)
   418  	MOVQ	gobuf_pc(BX), BX
   419  	JMP	BX
   420  
   421  // func mcall(fn func(*g))
   422  // Switch to m->g0's stack, call fn(g).
   423  // Fn must never return. It should gogo(&g->sched)
   424  // to keep running g.
   425  TEXT runtime·mcall<ABIInternal>(SB), NOSPLIT, $0-8
   426  #ifdef GOEXPERIMENT_runtimesecret
   427  	CMPL	g_secret(R14), $0
   428  	JEQ	nosecret
   429  	CALL	·secretEraseRegistersMcall(SB)
   430  nosecret:
   431  #endif
   432  
   433  	MOVQ	AX, DX	// DX = fn
   434  
   435  	// Save state in g->sched. The caller's SP and PC are restored by gogo to
   436  	// resume execution in the caller's frame (implicit return). The caller's BP
   437  	// is also restored to support frame pointer unwinding.
   438  	MOVQ	SP, BX	// hide (SP) reads from vet
   439  	MOVQ	8(BX), BX	// caller's PC
   440  	MOVQ	BX, (g_sched+gobuf_pc)(R14)
   441  	LEAQ	fn+0(FP), BX	// caller's SP
   442  	MOVQ	BX, (g_sched+gobuf_sp)(R14)
   443  	// Get the caller's frame pointer by dereferencing BP. Storing BP as it is
   444  	// can cause a frame pointer cycle, see CL 476235.
   445  	MOVQ	(BP), BX // caller's BP
   446  	MOVQ	BX, (g_sched+gobuf_bp)(R14)
   447  
   448  	// switch to m->g0 & its stack, call fn
   449  	MOVQ	g_m(R14), BX
   450  	MOVQ	m_g0(BX), SI	// SI = g.m.g0
   451  	CMPQ	SI, R14	// if g == m->g0 call badmcall
   452  	JNE	goodm
   453  	JMP	runtime·badmcall(SB)
   454  goodm:
   455  	MOVQ	R14, AX		// AX (and arg 0) = g
   456  	MOVQ	SI, R14		// g = g.m.g0
   457  	get_tls(CX)		// Set G in TLS
   458  	MOVQ	R14, g(CX)
   459  	MOVQ	(g_sched+gobuf_sp)(R14), SP	// sp = g0.sched.sp
   460  	MOVQ	$0, BP	// clear frame pointer, as caller may execute on another M
   461  	PUSHQ	AX	// open up space for fn's arg spill slot
   462  	MOVQ	0(DX), R12
   463  	CALL	R12		// fn(g)
   464  	// The Windows native stack unwinder incorrectly classifies the next instruction
   465  	// as part of the function epilogue, producing a wrong call stack.
   466  	// Add a NOP to work around this issue. See go.dev/issue/67007.
   467  	BYTE	$0x90
   468  	POPQ	AX
   469  	JMP	runtime·badmcall2(SB)
   470  	RET
   471  
   472  // systemstack_switch is a dummy routine that systemstack leaves at the bottom
   473  // of the G stack. We need to distinguish the routine that
   474  // lives at the bottom of the G stack from the one that lives
   475  // at the top of the system stack because the one at the top of
   476  // the system stack terminates the stack walk (see topofstack()).
   477  // The frame layout needs to match systemstack
   478  // so that it can pretend to be systemstack_switch.
   479  TEXT runtime·systemstack_switch(SB), NOSPLIT, $0-0
   480  	// Align for consistency with offset used in gosave_systemstack_switch
   481  	PCALIGN	$8
   482  	UNDEF
   483  	// Make sure this function is not leaf,
   484  	// so the frame is saved.
   485  	CALL	runtime·abort(SB)
   486  	RET
   487  
   488  // func systemstack(fn func())
   489  TEXT runtime·systemstack(SB), NOSPLIT, $0-8
   490  #ifdef GOEXPERIMENT_runtimesecret
   491  	// If in secret mode, erase registers on transition
   492  	// from G stack to M stack,
   493  	get_tls(CX)
   494  	MOVQ	g(CX), AX
   495  	CMPL	g_secret(AX), $0
   496  	JEQ	nosecret
   497  	CALL	·secretEraseRegisters(SB)
   498  nosecret:
   499  #endif
   500  
   501  	MOVQ	fn+0(FP), DI	// DI = fn
   502  	get_tls(CX)
   503  	MOVQ	g(CX), AX	// AX = g
   504  	MOVQ	g_m(AX), BX	// BX = m
   505  
   506  	CMPQ	AX, m_gsignal(BX)
   507  	JEQ	noswitch
   508  
   509  	MOVQ	m_g0(BX), DX	// DX = g0
   510  	CMPQ	AX, DX
   511  	JEQ	noswitch
   512  
   513  	CMPQ	AX, m_curg(BX)
   514  	JNE	bad
   515  
   516  	// Switch stacks.
   517  	// The original frame pointer is stored in BP,
   518  	// which is useful for stack unwinding.
   519  	// Save our state in g->sched. Pretend to
   520  	// be systemstack_switch if the G stack is scanned.
   521  	CALL	gosave_systemstack_switch<>(SB)
   522  
   523  	// switch to g0
   524  	MOVQ	DX, g(CX)
   525  	MOVQ	DX, R14 // set the g register
   526  	MOVQ	(g_sched+gobuf_sp)(DX), SP
   527  
   528  	// call target function
   529  	MOVQ	DI, DX
   530  	MOVQ	0(DI), DI
   531  	CALL	DI
   532  
   533  	// switch back to g
   534  	get_tls(CX)
   535  	MOVQ	g(CX), AX
   536  	MOVQ	g_m(AX), BX
   537  	MOVQ	m_curg(BX), AX
   538  	MOVQ	AX, g(CX)
   539  	MOVQ	(g_sched+gobuf_sp)(AX), SP
   540  	MOVQ	(g_sched+gobuf_bp)(AX), BP
   541  	MOVQ	$0, (g_sched+gobuf_sp)(AX)
   542  	MOVQ	$0, (g_sched+gobuf_bp)(AX)
   543  	RET
   544  
   545  noswitch:
   546  	// already on m stack; tail call the function
   547  	// Using a tail call here cleans up tracebacks since we won't stop
   548  	// at an intermediate systemstack.
   549  	MOVQ	DI, DX
   550  	MOVQ	0(DI), DI
   551  	// The function epilogue is not called on a tail call.
   552  	// Pop BP from the stack to simulate it.
   553  	POPQ	BP
   554  	JMP	DI
   555  
   556  bad:
   557  	// Bad: g is not gsignal, not g0, not curg. What is it?
   558  	MOVQ	$runtime·badsystemstack(SB), AX
   559  	CALL	AX
   560  	INT	$3
   561  
   562  // func switchToCrashStack0(fn func())
   563  TEXT runtime·switchToCrashStack0<ABIInternal>(SB), NOSPLIT, $0-8
   564  	MOVQ	g_m(R14), BX // curm
   565  
   566  	// set g to gcrash
   567  	LEAQ	runtime·gcrash(SB), R14 // g = &gcrash
   568  	MOVQ	BX, g_m(R14)            // g.m = curm
   569  	MOVQ	R14, m_g0(BX)           // curm.g0 = g
   570  	get_tls(CX)
   571  	MOVQ	R14, g(CX)
   572  
   573  	// switch to crashstack
   574  	MOVQ	(g_stack+stack_hi)(R14), BX
   575  	SUBQ	$(4*8), BX
   576  	MOVQ	BX, SP
   577  
   578  	// call target function
   579  	MOVQ	AX, DX
   580  	MOVQ	0(AX), AX
   581  	CALL	AX
   582  
   583  	// should never return
   584  	CALL	runtime·abort(SB)
   585  	UNDEF
   586  
   587  /*
   588   * support for morestack
   589   */
   590  
   591  // Called during function prolog when more stack is needed.
   592  //
   593  // The traceback routines see morestack on a g0 as being
   594  // the top of a stack (for example, morestack calling newstack
   595  // calling the scheduler calling newm calling gc), so we must
   596  // record an argument size. For that purpose, it has no arguments.
   597  TEXT runtime·morestack(SB),NOSPLIT|NOFRAME,$0-0
   598  	// Cannot grow scheduler stack (m->g0).
   599  	get_tls(CX)
   600  	MOVQ	g(CX), DI     // DI = g
   601  	MOVQ	g_m(DI), BX   // BX = m
   602  
   603  	// Set g->sched to context in f.
   604  	MOVQ	0(SP), AX // f's PC
   605  	MOVQ	AX, (g_sched+gobuf_pc)(DI)
   606  	LEAQ	8(SP), AX // f's SP
   607  	MOVQ	AX, (g_sched+gobuf_sp)(DI)
   608  	MOVQ	BP, (g_sched+gobuf_bp)(DI)
   609  	MOVQ	DX, (g_sched+gobuf_ctxt)(DI)
   610  
   611  	MOVQ	m_g0(BX), SI  // SI = m.g0
   612  	CMPQ	DI, SI
   613  	JNE	3(PC)
   614  	CALL	runtime·badmorestackg0(SB)
   615  	CALL	runtime·abort(SB)
   616  
   617  	// Cannot grow signal stack (m->gsignal).
   618  	MOVQ	m_gsignal(BX), SI
   619  	CMPQ	DI, SI
   620  	JNE	3(PC)
   621  	CALL	runtime·badmorestackgsignal(SB)
   622  	CALL	runtime·abort(SB)
   623  
   624  	// Called from f.
   625  	// Set m->morebuf to f's caller.
   626  	NOP	SP	// tell vet SP changed - stop checking offsets
   627  	MOVQ	8(SP), AX	// f's caller's PC
   628  	MOVQ	AX, (m_morebuf+gobuf_pc)(BX)
   629  	LEAQ	16(SP), AX	// f's caller's SP
   630  	MOVQ	AX, (m_morebuf+gobuf_sp)(BX)
   631  	MOVQ	DI, (m_morebuf+gobuf_g)(BX)
   632  
   633  	// If in secret mode, erase registers on transition
   634  	// from G stack to M stack,
   635  #ifdef GOEXPERIMENT_runtimesecret
   636  	CMPL	g_secret(DI), $0
   637  	JEQ	nosecret
   638  	CALL	·secretEraseRegisters(SB)
   639  	get_tls(CX)
   640  	MOVQ	g(CX), DI     // DI = g
   641  	MOVQ	g_m(DI), BX   // BX = m
   642  nosecret:
   643  #endif
   644  
   645  	// Call newstack on m->g0's stack.
   646  	MOVQ	m_g0(BX), BX
   647  	MOVQ	BX, g(CX)
   648  	MOVQ	(g_sched+gobuf_sp)(BX), SP
   649  	MOVQ	$0, BP			// clear frame pointer, as caller may execute on another M
   650  	CALL	runtime·newstack(SB)
   651  	CALL	runtime·abort(SB)	// crash if newstack returns
   652  	RET
   653  
   654  // morestack but not preserving ctxt.
   655  TEXT runtime·morestack_noctxt(SB),NOSPLIT,$0
   656  	MOVL	$0, DX
   657  	JMP	runtime·morestack(SB)
   658  
   659  // spillArgs stores return values from registers to a *internal/abi.RegArgs in R12.
   660  TEXT ·spillArgs(SB),NOSPLIT,$0-0
   661  	MOVQ AX, 0(R12)
   662  	MOVQ BX, 8(R12)
   663  	MOVQ CX, 16(R12)
   664  	MOVQ DI, 24(R12)
   665  	MOVQ SI, 32(R12)
   666  	MOVQ R8, 40(R12)
   667  	MOVQ R9, 48(R12)
   668  	MOVQ R10, 56(R12)
   669  	MOVQ R11, 64(R12)
   670  	MOVQ X0, 72(R12)
   671  	MOVQ X1, 80(R12)
   672  	MOVQ X2, 88(R12)
   673  	MOVQ X3, 96(R12)
   674  	MOVQ X4, 104(R12)
   675  	MOVQ X5, 112(R12)
   676  	MOVQ X6, 120(R12)
   677  	MOVQ X7, 128(R12)
   678  	MOVQ X8, 136(R12)
   679  	MOVQ X9, 144(R12)
   680  	MOVQ X10, 152(R12)
   681  	MOVQ X11, 160(R12)
   682  	MOVQ X12, 168(R12)
   683  	MOVQ X13, 176(R12)
   684  	MOVQ X14, 184(R12)
   685  	RET
   686  
   687  // unspillArgs loads args into registers from a *internal/abi.RegArgs in R12.
   688  TEXT ·unspillArgs(SB),NOSPLIT,$0-0
   689  	MOVQ 0(R12), AX
   690  	MOVQ 8(R12), BX
   691  	MOVQ 16(R12), CX
   692  	MOVQ 24(R12), DI
   693  	MOVQ 32(R12), SI
   694  	MOVQ 40(R12), R8
   695  	MOVQ 48(R12), R9
   696  	MOVQ 56(R12), R10
   697  	MOVQ 64(R12), R11
   698  	MOVQ 72(R12), X0
   699  	MOVQ 80(R12), X1
   700  	MOVQ 88(R12), X2
   701  	MOVQ 96(R12), X3
   702  	MOVQ 104(R12), X4
   703  	MOVQ 112(R12), X5
   704  	MOVQ 120(R12), X6
   705  	MOVQ 128(R12), X7
   706  	MOVQ 136(R12), X8
   707  	MOVQ 144(R12), X9
   708  	MOVQ 152(R12), X10
   709  	MOVQ 160(R12), X11
   710  	MOVQ 168(R12), X12
   711  	MOVQ 176(R12), X13
   712  	MOVQ 184(R12), X14
   713  	RET
   714  
   715  // reflectcall: call a function with the given argument list
   716  // func call(stackArgsType *_type, f *FuncVal, stackArgs *byte, stackArgsSize, stackRetOffset, frameSize uint32, regArgs *abi.RegArgs).
   717  // we don't have variable-sized frames, so we use a small number
   718  // of constant-sized-frame functions to encode a few bits of size in the pc.
   719  // Caution: ugly multiline assembly macros in your future!
   720  
   721  #define DISPATCH(NAME,MAXSIZE)		\
   722  	CMPQ	CX, $MAXSIZE;		\
   723  	JA	3(PC);			\
   724  	MOVQ	$NAME(SB), AX;		\
   725  	JMP	AX
   726  // Note: can't just "JMP NAME(SB)" - bad inlining results.
   727  
   728  TEXT ·reflectcall(SB), NOSPLIT, $0-48
   729  	MOVLQZX frameSize+32(FP), CX
   730  	DISPATCH(runtime·call16, 16)
   731  	DISPATCH(runtime·call32, 32)
   732  	DISPATCH(runtime·call64, 64)
   733  	DISPATCH(runtime·call128, 128)
   734  	DISPATCH(runtime·call256, 256)
   735  	DISPATCH(runtime·call512, 512)
   736  	DISPATCH(runtime·call1024, 1024)
   737  	DISPATCH(runtime·call2048, 2048)
   738  	DISPATCH(runtime·call4096, 4096)
   739  	DISPATCH(runtime·call8192, 8192)
   740  	DISPATCH(runtime·call16384, 16384)
   741  	DISPATCH(runtime·call32768, 32768)
   742  	DISPATCH(runtime·call65536, 65536)
   743  	DISPATCH(runtime·call131072, 131072)
   744  	DISPATCH(runtime·call262144, 262144)
   745  	DISPATCH(runtime·call524288, 524288)
   746  	DISPATCH(runtime·call1048576, 1048576)
   747  	DISPATCH(runtime·call2097152, 2097152)
   748  	DISPATCH(runtime·call4194304, 4194304)
   749  	DISPATCH(runtime·call8388608, 8388608)
   750  	DISPATCH(runtime·call16777216, 16777216)
   751  	DISPATCH(runtime·call33554432, 33554432)
   752  	DISPATCH(runtime·call67108864, 67108864)
   753  	DISPATCH(runtime·call134217728, 134217728)
   754  	DISPATCH(runtime·call268435456, 268435456)
   755  	DISPATCH(runtime·call536870912, 536870912)
   756  	DISPATCH(runtime·call1073741824, 1073741824)
   757  	MOVQ	$runtime·badreflectcall(SB), AX
   758  	JMP	AX
   759  
   760  #define CALLFN(NAME,MAXSIZE)			\
   761  TEXT NAME(SB), WRAPPER, $MAXSIZE-48;		\
   762  	NO_LOCAL_POINTERS;			\
   763  	/* copy arguments to stack */		\
   764  	MOVQ	stackArgs+16(FP), SI;		\
   765  	MOVLQZX stackArgsSize+24(FP), CX;		\
   766  	MOVQ	SP, DI;				\
   767  	REP;MOVSB;				\
   768  	/* set up argument registers */		\
   769  	MOVQ    regArgs+40(FP), R12;		\
   770  	CALL    ·unspillArgs(SB);		\
   771  	/* call function */			\
   772  	MOVQ	f+8(FP), DX;			\
   773  	PCDATA  $PCDATA_StackMapIndex, $0;	\
   774  	MOVQ	(DX), R12;			\
   775  	CALL	R12;				\
   776  	/* copy register return values back */		\
   777  	MOVQ    regArgs+40(FP), R12;		\
   778  	CALL    ·spillArgs(SB);		\
   779  	MOVLQZX	stackArgsSize+24(FP), CX;		\
   780  	MOVLQZX	stackRetOffset+28(FP), BX;		\
   781  	MOVQ	stackArgs+16(FP), DI;		\
   782  	MOVQ	stackArgsType+0(FP), DX;		\
   783  	MOVQ	SP, SI;				\
   784  	ADDQ	BX, DI;				\
   785  	ADDQ	BX, SI;				\
   786  	SUBQ	BX, CX;				\
   787  	CALL	callRet<>(SB);			\
   788  	RET
   789  
   790  // callRet copies return values back at the end of call*. This is a
   791  // separate function so it can allocate stack space for the arguments
   792  // to reflectcallmove. It does not follow the Go ABI; it expects its
   793  // arguments in registers.
   794  TEXT callRet<>(SB), NOSPLIT, $40-0
   795  	NO_LOCAL_POINTERS
   796  	MOVQ	DX, 0(SP)
   797  	MOVQ	DI, 8(SP)
   798  	MOVQ	SI, 16(SP)
   799  	MOVQ	CX, 24(SP)
   800  	MOVQ	R12, 32(SP)
   801  	CALL	runtime·reflectcallmove(SB)
   802  	RET
   803  
   804  CALLFN(·call16, 16)
   805  CALLFN(·call32, 32)
   806  CALLFN(·call64, 64)
   807  CALLFN(·call128, 128)
   808  CALLFN(·call256, 256)
   809  CALLFN(·call512, 512)
   810  CALLFN(·call1024, 1024)
   811  CALLFN(·call2048, 2048)
   812  CALLFN(·call4096, 4096)
   813  CALLFN(·call8192, 8192)
   814  CALLFN(·call16384, 16384)
   815  CALLFN(·call32768, 32768)
   816  CALLFN(·call65536, 65536)
   817  CALLFN(·call131072, 131072)
   818  CALLFN(·call262144, 262144)
   819  CALLFN(·call524288, 524288)
   820  CALLFN(·call1048576, 1048576)
   821  CALLFN(·call2097152, 2097152)
   822  CALLFN(·call4194304, 4194304)
   823  CALLFN(·call8388608, 8388608)
   824  CALLFN(·call16777216, 16777216)
   825  CALLFN(·call33554432, 33554432)
   826  CALLFN(·call67108864, 67108864)
   827  CALLFN(·call134217728, 134217728)
   828  CALLFN(·call268435456, 268435456)
   829  CALLFN(·call536870912, 536870912)
   830  CALLFN(·call1073741824, 1073741824)
   831  
   832  TEXT runtime·procyieldAsm(SB),NOSPLIT,$0-0
   833  	MOVL	cycles+0(FP), AX
   834  	TESTL	AX, AX
   835  	JZ	done
   836  again:
   837  	PAUSE
   838  	SUBL	$1, AX
   839  	JNZ	again
   840  done:
   841  	RET
   842  
   843  
   844  TEXT ·publicationBarrier<ABIInternal>(SB),NOSPLIT,$0-0
   845  	// Stores are already ordered on x86, so this is just a
   846  	// compile barrier.
   847  	RET
   848  
   849  // Save state of caller into g->sched,
   850  // but using fake PC from systemstack_switch.
   851  // Must only be called from functions with frame pointer
   852  // and without locals ($0) or else unwinding from
   853  // systemstack_switch is incorrect.
   854  // Smashes R9.
   855  TEXT gosave_systemstack_switch<>(SB),NOSPLIT|NOFRAME,$0
   856  	// Take systemstack_switch PC and add 8 bytes to skip
   857  	// the prologue. Keep 8 bytes offset consistent with
   858  	// PCALIGN $8 in systemstack_swtich, pointing start of
   859  	// UNDEF instruction beyond prologue.
   860  	MOVQ	$runtime·systemstack_switch+8(SB), R9
   861  	MOVQ	R9, (g_sched+gobuf_pc)(R14)
   862  	LEAQ	8(SP), R9
   863  	MOVQ	R9, (g_sched+gobuf_sp)(R14)
   864  	MOVQ	BP, (g_sched+gobuf_bp)(R14)
   865  	// Assert ctxt is zero. See func save.
   866  	MOVQ	(g_sched+gobuf_ctxt)(R14), R9
   867  	TESTQ	R9, R9
   868  	JZ	2(PC)
   869  	CALL	runtime·abort(SB)
   870  	RET
   871  
   872  // func asmcgocall_no_g(fn, arg unsafe.Pointer)
   873  // Call fn(arg) aligned appropriately for the gcc ABI.
   874  // Called on a system stack, and there may be no g yet (during needm).
   875  TEXT ·asmcgocall_no_g(SB),NOSPLIT,$32-16
   876  	MOVQ	fn+0(FP), AX
   877  	MOVQ	arg+8(FP), BX
   878  	MOVQ	SP, DX
   879  	ANDQ	$~15, SP	// alignment
   880  	MOVQ	DX, 8(SP)
   881  	MOVQ	BX, DI		// DI = first argument in AMD64 ABI
   882  	MOVQ	BX, CX		// CX = first argument in Win64
   883  	CALL	AX
   884  	MOVQ	8(SP), DX
   885  	MOVQ	DX, SP
   886  	RET
   887  
   888  // asmcgocall_landingpad calls AX with BX as argument.
   889  // Must be called on the system stack.
   890  TEXT ·asmcgocall_landingpad(SB),NOSPLIT,$0-0
   891  #ifdef GOOS_windows
   892  	// Make sure we have enough room for 4 stack-backed fast-call
   893  	// registers as per Windows amd64 calling convention.
   894  	ADJSP	$32
   895  	// On Windows, asmcgocall_landingpad acts as landing pad for exceptions
   896  	// thrown in the cgo call. Exceptions that reach this function will be
   897  	// handled by runtime.sehtramp thanks to the SEH metadata added
   898  	// by the compiler.
   899  	// Note that runtime.sehtramp can't be attached directly to asmcgocall
   900  	// because its initial stack pointer can be outside the system stack bounds,
   901  	// and Windows stops the stack unwinding without calling the exception handler
   902  	// when it reaches that point.
   903  	MOVQ	BX, CX		// CX = first argument in Win64
   904  	CALL	AX
   905  	// The exception handler is not called if the next instruction is part of
   906  	// the epilogue, which includes the RET instruction, so we need to add a NOP here.
   907  	BYTE	$0x90
   908  	ADJSP	$-32
   909  	RET
   910  #endif
   911  	// Tail call AX on non-Windows, as the extra stack frame is not needed.
   912  	MOVQ	BX, DI		// DI = first argument in AMD64 ABI
   913  	JMP	AX
   914  
   915  // func asmcgocall(fn, arg unsafe.Pointer) int32
   916  // Call fn(arg) on the scheduler stack,
   917  // aligned appropriately for the gcc ABI.
   918  // See cgocall.go for more details.
   919  TEXT ·asmcgocall(SB),NOSPLIT,$0-20
   920  	// Figure out if we need to switch to m->g0 stack.
   921  	// We get called to create new OS threads too, and those
   922  	// come in on the m->g0 stack already. Or we might already
   923  	// be on the m->gsignal stack.
   924  	get_tls(CX)
   925  	MOVQ	g(CX), DI
   926  	CMPQ	DI, $0
   927  	JEQ	nosave
   928  	MOVQ	g_m(DI), R8
   929  	MOVQ	m_gsignal(R8), SI
   930  	CMPQ	DI, SI
   931  	JEQ	nosave
   932  	MOVQ	m_g0(R8), SI
   933  	CMPQ	DI, SI
   934  	JEQ	nosave
   935  
   936  	// Running on a user G
   937  	// Figure out if we're running secret code and clear the registers
   938  	// so that the C code we're about to call doesn't spill confidential
   939  	// information into memory
   940  #ifdef GOEXPERIMENT_runtimesecret
   941  	CMPL	g_secret(DI), $0
   942  	JEQ	nosecret
   943  	CALL	·secretEraseRegisters(SB)
   944  
   945  nosecret:
   946  #endif
   947  	MOVQ	fn+0(FP), AX
   948  	MOVQ	arg+8(FP), BX
   949  	MOVQ	SP, DX
   950  
   951  	// Switch to system stack.
   952  	// The original frame pointer is stored in BP,
   953  	// which is useful for stack unwinding.
   954  	CALL	gosave_systemstack_switch<>(SB)
   955  	MOVQ	SI, g(CX)
   956  	MOVQ	(g_sched+gobuf_sp)(SI), SP
   957  
   958  	// Now on a scheduling stack (a pthread-created stack).
   959  	SUBQ	$16, SP
   960  	ANDQ	$~15, SP	// alignment for gcc ABI
   961  	MOVQ	DI, 8(SP)	// save g
   962  	MOVQ	(g_stack+stack_hi)(DI), DI
   963  	SUBQ	DX, DI
   964  	MOVQ	DI, 0(SP)	// save depth in stack (can't just save SP, as stack might be copied during a callback)
   965  	CALL	runtime·asmcgocall_landingpad(SB)
   966  
   967  	// Restore registers, g, stack pointer.
   968  	get_tls(CX)
   969  	MOVQ	8(SP), DI
   970  	MOVQ	(g_stack+stack_hi)(DI), SI
   971  	SUBQ	0(SP), SI
   972  	MOVQ	DI, g(CX)
   973  	MOVQ	SI, SP
   974  
   975  	MOVL	AX, ret+16(FP)
   976  	RET
   977  
   978  nosave:
   979  	// Running on a system stack, perhaps even without a g.
   980  	// Having no g can happen during thread creation or thread teardown
   981  	// (see needm/dropm on Solaris, for example).
   982  	// This code is like the above sequence but without saving/restoring g
   983  	// and without worrying about the stack moving out from under us
   984  	// (because we're on a system stack, not a goroutine stack).
   985  	MOVQ	fn+0(FP), AX
   986  	MOVQ	arg+8(FP), BX
   987  	MOVQ	SP, DX
   988  
   989  	SUBQ	$16, SP
   990  	ANDQ	$~15, SP
   991  	MOVQ	$0, 8(SP)		// where above code stores g, in case someone looks during debugging
   992  	MOVQ	DX, 0(SP)	// save original stack pointer
   993  	CALL	runtime·asmcgocall_landingpad(SB)
   994  	MOVQ	0(SP), SI	// restore original stack pointer
   995  	MOVQ	SI, SP
   996  	MOVL	AX, ret+16(FP)
   997  	RET
   998  
   999  #ifdef GOOS_windows
  1000  // Dummy TLS that's used on Windows so that we don't crash trying
  1001  // to restore the G register in needm. needm and its callees are
  1002  // very careful never to actually use the G, the TLS just can't be
  1003  // unset since we're in Go code.
  1004  GLOBL zeroTLS<>(SB),RODATA,$const_tlsSize
  1005  #endif
  1006  
  1007  // func cgocallback(fn, frame unsafe.Pointer, ctxt uintptr)
  1008  // See cgocall.go for more details.
  1009  TEXT ·cgocallback(SB),NOSPLIT,$24-24
  1010  	NO_LOCAL_POINTERS
  1011  
  1012  	// Skip cgocallbackg, just dropm when fn is nil, and frame is the saved g.
  1013  	// It is used to dropm while thread is exiting.
  1014  	MOVQ	fn+0(FP), AX
  1015  	CMPQ	AX, $0
  1016  	JNE	loadg
  1017  	// Restore the g from frame.
  1018  	get_tls(CX)
  1019  	MOVQ	frame+8(FP), BX
  1020  	MOVQ	BX, g(CX)
  1021  	JMP	dropm
  1022  
  1023  loadg:
  1024  	// If g is nil, Go did not create the current thread,
  1025  	// or if this thread never called into Go on pthread platforms.
  1026  	// Call needm to obtain one m for temporary use.
  1027  	// In this case, we're running on the thread stack, so there's
  1028  	// lots of space, but the linker doesn't know. Hide the call from
  1029  	// the linker analysis by using an indirect call through AX.
  1030  	get_tls(CX)
  1031  #ifdef GOOS_windows
  1032  	MOVL	$0, BX
  1033  	CMPQ	CX, $0
  1034  	JEQ	2(PC)
  1035  #endif
  1036  	MOVQ	g(CX), BX
  1037  	CMPQ	BX, $0
  1038  	JEQ	needm
  1039  	MOVQ	g_m(BX), BX
  1040  	MOVQ	BX, savedm-8(SP)	// saved copy of oldm
  1041  	JMP	havem
  1042  needm:
  1043  #ifdef GOOS_windows
  1044  	// Set up a dummy TLS value. needm is careful not to use it,
  1045  	// but it needs to be there to prevent autogenerated code from
  1046  	// crashing when it loads from it.
  1047  	// We don't need to clear it or anything later because needm
  1048  	// will set up TLS properly.
  1049  	MOVQ	$zeroTLS<>(SB), DI
  1050  	CALL	runtime·settls(SB)
  1051  #endif
  1052  	// On some platforms (Windows) we cannot call needm through
  1053  	// an ABI wrapper because there's no TLS set up, and the ABI
  1054  	// wrapper will try to restore the G register (R14) from TLS.
  1055  	// Clear X15 because Go expects it and we're not calling
  1056  	// through a wrapper, but otherwise avoid setting the G
  1057  	// register in the wrapper and call needm directly. It
  1058  	// takes no arguments and doesn't return any values so
  1059  	// there's no need to handle that. Clear R14 so that there's
  1060  	// a bad value in there, in case needm tries to use it.
  1061  	XORPS	X15, X15
  1062  	XORQ    R14, R14
  1063  	MOVQ	$runtime·needAndBindM<ABIInternal>(SB), AX
  1064  	CALL	AX
  1065  	MOVQ	$0, savedm-8(SP)
  1066  	get_tls(CX)
  1067  	MOVQ	g(CX), BX
  1068  	MOVQ	g_m(BX), BX
  1069  
  1070  	// Set m->sched.sp = SP, so that if a panic happens
  1071  	// during the function we are about to execute, it will
  1072  	// have a valid SP to run on the g0 stack.
  1073  	// The next few lines (after the havem label)
  1074  	// will save this SP onto the stack and then write
  1075  	// the same SP back to m->sched.sp. That seems redundant,
  1076  	// but if an unrecovered panic happens, unwindm will
  1077  	// restore the g->sched.sp from the stack location
  1078  	// and then systemstack will try to use it. If we don't set it here,
  1079  	// that restored SP will be uninitialized (typically 0) and
  1080  	// will not be usable.
  1081  	MOVQ	m_g0(BX), SI
  1082  	MOVQ	SP, (g_sched+gobuf_sp)(SI)
  1083  
  1084  havem:
  1085  	// Now there's a valid m, and we're running on its m->g0.
  1086  	// Save current m->g0->sched.sp on stack and then set it to SP.
  1087  	// Save current sp in m->g0->sched.sp in preparation for
  1088  	// switch back to m->curg stack.
  1089  	// NOTE: unwindm knows that the saved g->sched.sp is at 0(SP).
  1090  	MOVQ	m_g0(BX), SI
  1091  	MOVQ	(g_sched+gobuf_sp)(SI), AX
  1092  	MOVQ	AX, 0(SP)
  1093  	MOVQ	SP, (g_sched+gobuf_sp)(SI)
  1094  
  1095  	// Switch to m->curg stack and call runtime.cgocallbackg.
  1096  	// Because we are taking over the execution of m->curg
  1097  	// but *not* resuming what had been running, we need to
  1098  	// save that information (m->curg->sched) so we can restore it.
  1099  	// We can restore m->curg->sched.sp easily, because calling
  1100  	// runtime.cgocallbackg leaves SP unchanged upon return.
  1101  	// To save m->curg->sched.pc, we push it onto the curg stack and
  1102  	// open a frame the same size as cgocallback's g0 frame.
  1103  	// Once we switch to the curg stack, the pushed PC will appear
  1104  	// to be the return PC of cgocallback, so that the traceback
  1105  	// will seamlessly trace back into the earlier calls.
  1106  	MOVQ	m_curg(BX), SI
  1107  	MOVQ	SI, g(CX)
  1108  	MOVQ	(g_sched+gobuf_sp)(SI), DI  // prepare stack as DI
  1109  	MOVQ	(g_sched+gobuf_pc)(SI), BX
  1110  	MOVQ	BX, -8(DI)  // "push" return PC on the g stack
  1111  	// Gather our arguments into registers.
  1112  	MOVQ	fn+0(FP), BX
  1113  	MOVQ	frame+8(FP), CX
  1114  	MOVQ	ctxt+16(FP), DX
  1115  	// Compute the size of the frame, including return PC and, if
  1116  	// GOEXPERIMENT=framepointer, the saved base pointer
  1117  	LEAQ	fn+0(FP), AX
  1118  	SUBQ	SP, AX   // AX is our actual frame size
  1119  	SUBQ	AX, DI   // Allocate the same frame size on the g stack
  1120  	MOVQ	DI, SP
  1121  
  1122  	MOVQ	BX, 0(SP)
  1123  	MOVQ	CX, 8(SP)
  1124  	MOVQ	DX, 16(SP)
  1125  	MOVQ	$runtime·cgocallbackg(SB), AX
  1126  	CALL	AX	// indirect call to bypass nosplit check. We're on a different stack now.
  1127  
  1128  	// Compute the size of the frame again. FP and SP have
  1129  	// completely different values here than they did above,
  1130  	// but only their difference matters.
  1131  	LEAQ	fn+0(FP), AX
  1132  	SUBQ	SP, AX
  1133  
  1134  	// Restore g->sched (== m->curg->sched) from saved values.
  1135  	get_tls(CX)
  1136  	MOVQ	g(CX), SI
  1137  	MOVQ	SP, DI
  1138  	ADDQ	AX, DI
  1139  	MOVQ	-8(DI), BX
  1140  	MOVQ	BX, (g_sched+gobuf_pc)(SI)
  1141  	MOVQ	DI, (g_sched+gobuf_sp)(SI)
  1142  
  1143  	// Switch back to m->g0's stack and restore m->g0->sched.sp.
  1144  	// (Unlike m->curg, the g0 goroutine never uses sched.pc,
  1145  	// so we do not have to restore it.)
  1146  	MOVQ	g(CX), BX
  1147  	MOVQ	g_m(BX), BX
  1148  	MOVQ	m_g0(BX), SI
  1149  	MOVQ	SI, g(CX)
  1150  	MOVQ	(g_sched+gobuf_sp)(SI), SP
  1151  	MOVQ	0(SP), AX
  1152  	MOVQ	AX, (g_sched+gobuf_sp)(SI)
  1153  
  1154  	// If the m on entry was nil, we called needm above to borrow an m,
  1155  	// 1. for the duration of the call on non-pthread platforms,
  1156  	// 2. or the duration of the C thread alive on pthread platforms.
  1157  	// If the m on entry wasn't nil,
  1158  	// 1. the thread might be a Go thread,
  1159  	// 2. or it wasn't the first call from a C thread on pthread platforms,
  1160  	//    since then we skip dropm to reuse the m in the first call.
  1161  	MOVQ	savedm-8(SP), BX
  1162  	CMPQ	BX, $0
  1163  	JNE	done
  1164  
  1165  	// Skip dropm to reuse it in the next call, when a pthread key has been created.
  1166  	MOVQ	_cgo_pthread_key_created(SB), AX
  1167  	// It means cgo is disabled when _cgo_pthread_key_created is a nil pointer, need dropm.
  1168  	CMPQ	AX, $0
  1169  	JEQ	dropm
  1170  	CMPQ	(AX), $0
  1171  	JNE	done
  1172  
  1173  dropm:
  1174  	MOVQ	$runtime·dropm(SB), AX
  1175  	CALL	AX
  1176  #ifdef GOOS_windows
  1177  	// We need to clear the TLS pointer in case the next
  1178  	// thread that comes into Go tries to reuse that space
  1179  	// but uses the same M.
  1180  	XORQ	DI, DI
  1181  	CALL	runtime·settls(SB)
  1182  #endif
  1183  done:
  1184  
  1185  	// Done!
  1186  	RET
  1187  
  1188  // func setg(gg *g)
  1189  // set g. for use by needm.
  1190  TEXT runtime·setg(SB), NOSPLIT, $0-8
  1191  	MOVQ	gg+0(FP), BX
  1192  	get_tls(CX)
  1193  	MOVQ	BX, g(CX)
  1194  	RET
  1195  
  1196  // void setg_gcc(G*); set g called from gcc.
  1197  TEXT setg_gcc<>(SB),NOSPLIT,$0
  1198  	get_tls(AX)
  1199  	MOVQ	DI, g(AX)
  1200  	MOVQ	DI, R14 // set the g register
  1201  	RET
  1202  
  1203  TEXT runtime·abort(SB),NOSPLIT,$0-0
  1204  	INT	$3
  1205  loop:
  1206  	JMP	loop
  1207  
  1208  // check that SP is in range [g->stack.lo, g->stack.hi)
  1209  TEXT runtime·stackcheck(SB), NOSPLIT|NOFRAME, $0-0
  1210  	get_tls(CX)
  1211  	MOVQ	g(CX), AX
  1212  	CMPQ	(g_stack+stack_hi)(AX), SP
  1213  	JHI	2(PC)
  1214  	CALL	runtime·abort(SB)
  1215  	CMPQ	SP, (g_stack+stack_lo)(AX)
  1216  	JHI	2(PC)
  1217  	CALL	runtime·abort(SB)
  1218  	RET
  1219  
  1220  // func cputicks() int64
  1221  TEXT runtime·cputicks(SB),NOSPLIT,$0-0
  1222  	CMPB	internal∕cpu·X86+const_offsetX86HasRDTSCP(SB), $1
  1223  	JNE	fences
  1224  	// Instruction stream serializing RDTSCP is supported.
  1225  	// RDTSCP is supported by Intel Nehalem (2008) and
  1226  	// AMD K8 Rev. F (2006) and newer.
  1227  	RDTSCP
  1228  done:
  1229  	SHLQ	$32, DX
  1230  	ADDQ	DX, AX
  1231  	MOVQ	AX, ret+0(FP)
  1232  	RET
  1233  fences:
  1234  	// MFENCE is instruction stream serializing and flushes the
  1235  	// store buffers on AMD. The serialization semantics of LFENCE on AMD
  1236  	// are dependent on MSR C001_1029 and CPU generation.
  1237  	// LFENCE on Intel does wait for all previous instructions to have executed.
  1238  	// Intel recommends MFENCE;LFENCE in its manuals before RDTSC to have all
  1239  	// previous instructions executed and all previous loads and stores to globally visible.
  1240  	// Using MFENCE;LFENCE here aligns the serializing properties without
  1241  	// runtime detection of CPU manufacturer.
  1242  	MFENCE
  1243  	LFENCE
  1244  	RDTSC
  1245  	JMP done
  1246  
  1247  // func memhash(p unsafe.Pointer, h, s uintptr) uintptr
  1248  // hash function using AES hardware instructions
  1249  TEXT runtime·memhash<ABIInternal>(SB),NOSPLIT,$0-32
  1250  	// AX = ptr to data
  1251  	// BX = seed
  1252  	// CX = size
  1253  	CMPB	runtime·useAeshash(SB), $0
  1254  	JEQ	noaes
  1255  	JMP	runtime·aeshashbody<>(SB)
  1256  noaes:
  1257  	JMP	runtime·memhashFallback<ABIInternal>(SB)
  1258  
  1259  // func strhash(p unsafe.Pointer, h uintptr) uintptr
  1260  TEXT runtime·strhash<ABIInternal>(SB),NOSPLIT,$0-24
  1261  	// AX = ptr to string struct
  1262  	// BX = seed
  1263  	CMPB	runtime·useAeshash(SB), $0
  1264  	JEQ	noaes
  1265  	MOVQ	8(AX), CX	// length of string
  1266  	MOVQ	(AX), AX	// string data
  1267  	JMP	runtime·aeshashbody<>(SB)
  1268  noaes:
  1269  	JMP	runtime·strhashFallback<ABIInternal>(SB)
  1270  
  1271  // AX: data
  1272  // BX: hash seed
  1273  // CX: length
  1274  // At return: AX = return value
  1275  TEXT runtime·aeshashbody<>(SB),NOSPLIT,$0-0
  1276  	// Fill an SSE register with our seeds.
  1277  	MOVQ	BX, X0				// 64 bits of per-table hash seed
  1278  	PINSRW	$4, CX, X0			// 16 bits of length
  1279  	PSHUFHW $0, X0, X0			// repeat length 4 times total
  1280  	MOVO	X0, X1				// save unscrambled seed
  1281  	PXOR	runtime·aeskeysched(SB), X0	// xor in per-process seed
  1282  	AESENC	X0, X0				// scramble seed
  1283  
  1284  	CMPQ	CX, $16
  1285  	JB	aes0to15
  1286  	JE	aes16
  1287  	CMPQ	CX, $32
  1288  	JBE	aes17to32
  1289  	CMPQ	CX, $64
  1290  	JBE	aes33to64
  1291  	CMPQ	CX, $128
  1292  	JBE	aes65to128
  1293  	JMP	aes129plus
  1294  
  1295  aes0to15:
  1296  	TESTQ	CX, CX
  1297  	JE	aes0
  1298  
  1299  	ADDQ	$16, AX
  1300  	TESTW	$0xff0, AX
  1301  	JE	endofpage
  1302  
  1303  	// 16 bytes loaded at this address won't cross
  1304  	// a page boundary, so we can load it directly.
  1305  	MOVOU	-16(AX), X1
  1306  	ADDQ	CX, CX
  1307  	MOVQ	$masks<>(SB), AX
  1308  	PAND	(AX)(CX*8), X1
  1309  final1:
  1310  	PXOR	X0, X1	// xor data with seed
  1311  	AESENC	X1, X1	// scramble combo 3 times
  1312  	AESENC	X1, X1
  1313  	AESENC	X1, X1
  1314  	MOVQ	X1, AX	// return X1
  1315  	RET
  1316  
  1317  endofpage:
  1318  	// address ends in 1111xxxx. Might be up against
  1319  	// a page boundary, so load ending at last byte.
  1320  	// Then shift bytes down using pshufb.
  1321  	MOVOU	-32(AX)(CX*1), X1
  1322  	ADDQ	CX, CX
  1323  	MOVQ	$shifts<>(SB), AX
  1324  	PSHUFB	(AX)(CX*8), X1
  1325  	JMP	final1
  1326  
  1327  aes0:
  1328  	// Return scrambled input seed
  1329  	AESENC	X0, X0
  1330  	MOVQ	X0, AX	// return X0
  1331  	RET
  1332  
  1333  aes16:
  1334  	MOVOU	(AX), X1
  1335  	JMP	final1
  1336  
  1337  aes17to32:
  1338  	// make second starting seed
  1339  	PXOR	runtime·aeskeysched+16(SB), X1
  1340  	AESENC	X1, X1
  1341  
  1342  	// load data to be hashed
  1343  	MOVOU	(AX), X2
  1344  	MOVOU	-16(AX)(CX*1), X3
  1345  
  1346  	// xor with seed
  1347  	PXOR	X0, X2
  1348  	PXOR	X1, X3
  1349  
  1350  	// scramble 3 times
  1351  	AESENC	X2, X2
  1352  	AESENC	X3, X3
  1353  	AESENC	X2, X2
  1354  	AESENC	X3, X3
  1355  	AESENC	X2, X2
  1356  	AESENC	X3, X3
  1357  
  1358  	// combine results
  1359  	PXOR	X3, X2
  1360  	MOVQ	X2, AX	// return X2
  1361  	RET
  1362  
  1363  aes33to64:
  1364  	// make 3 more starting seeds
  1365  	MOVO	X1, X2
  1366  	MOVO	X1, X3
  1367  	PXOR	runtime·aeskeysched+16(SB), X1
  1368  	PXOR	runtime·aeskeysched+32(SB), X2
  1369  	PXOR	runtime·aeskeysched+48(SB), X3
  1370  	AESENC	X1, X1
  1371  	AESENC	X2, X2
  1372  	AESENC	X3, X3
  1373  
  1374  	MOVOU	(AX), X4
  1375  	MOVOU	16(AX), X5
  1376  	MOVOU	-32(AX)(CX*1), X6
  1377  	MOVOU	-16(AX)(CX*1), X7
  1378  
  1379  	PXOR	X0, X4
  1380  	PXOR	X1, X5
  1381  	PXOR	X2, X6
  1382  	PXOR	X3, X7
  1383  
  1384  	AESENC	X4, X4
  1385  	AESENC	X5, X5
  1386  	AESENC	X6, X6
  1387  	AESENC	X7, X7
  1388  
  1389  	AESENC	X4, X4
  1390  	AESENC	X5, X5
  1391  	AESENC	X6, X6
  1392  	AESENC	X7, X7
  1393  
  1394  	AESENC	X4, X4
  1395  	AESENC	X5, X5
  1396  	AESENC	X6, X6
  1397  	AESENC	X7, X7
  1398  
  1399  	PXOR	X6, X4
  1400  	PXOR	X7, X5
  1401  	PXOR	X5, X4
  1402  	MOVQ	X4, AX	// return X4
  1403  	RET
  1404  
  1405  aes65to128:
  1406  	// make 7 more starting seeds
  1407  	MOVO	X1, X2
  1408  	MOVO	X1, X3
  1409  	MOVO	X1, X4
  1410  	MOVO	X1, X5
  1411  	MOVO	X1, X6
  1412  	MOVO	X1, X7
  1413  	PXOR	runtime·aeskeysched+16(SB), X1
  1414  	PXOR	runtime·aeskeysched+32(SB), X2
  1415  	PXOR	runtime·aeskeysched+48(SB), X3
  1416  	PXOR	runtime·aeskeysched+64(SB), X4
  1417  	PXOR	runtime·aeskeysched+80(SB), X5
  1418  	PXOR	runtime·aeskeysched+96(SB), X6
  1419  	PXOR	runtime·aeskeysched+112(SB), X7
  1420  	AESENC	X1, X1
  1421  	AESENC	X2, X2
  1422  	AESENC	X3, X3
  1423  	AESENC	X4, X4
  1424  	AESENC	X5, X5
  1425  	AESENC	X6, X6
  1426  	AESENC	X7, X7
  1427  
  1428  	// load data
  1429  	MOVOU	(AX), X8
  1430  	MOVOU	16(AX), X9
  1431  	MOVOU	32(AX), X10
  1432  	MOVOU	48(AX), X11
  1433  	MOVOU	-64(AX)(CX*1), X12
  1434  	MOVOU	-48(AX)(CX*1), X13
  1435  	MOVOU	-32(AX)(CX*1), X14
  1436  	MOVOU	-16(AX)(CX*1), X15
  1437  
  1438  	// xor with seed
  1439  	PXOR	X0, X8
  1440  	PXOR	X1, X9
  1441  	PXOR	X2, X10
  1442  	PXOR	X3, X11
  1443  	PXOR	X4, X12
  1444  	PXOR	X5, X13
  1445  	PXOR	X6, X14
  1446  	PXOR	X7, X15
  1447  
  1448  	// scramble 3 times
  1449  	AESENC	X8, X8
  1450  	AESENC	X9, X9
  1451  	AESENC	X10, X10
  1452  	AESENC	X11, X11
  1453  	AESENC	X12, X12
  1454  	AESENC	X13, X13
  1455  	AESENC	X14, X14
  1456  	AESENC	X15, X15
  1457  
  1458  	AESENC	X8, X8
  1459  	AESENC	X9, X9
  1460  	AESENC	X10, X10
  1461  	AESENC	X11, X11
  1462  	AESENC	X12, X12
  1463  	AESENC	X13, X13
  1464  	AESENC	X14, X14
  1465  	AESENC	X15, X15
  1466  
  1467  	AESENC	X8, X8
  1468  	AESENC	X9, X9
  1469  	AESENC	X10, X10
  1470  	AESENC	X11, X11
  1471  	AESENC	X12, X12
  1472  	AESENC	X13, X13
  1473  	AESENC	X14, X14
  1474  	AESENC	X15, X15
  1475  
  1476  	// combine results
  1477  	PXOR	X12, X8
  1478  	PXOR	X13, X9
  1479  	PXOR	X14, X10
  1480  	PXOR	X15, X11
  1481  	PXOR	X10, X8
  1482  	PXOR	X11, X9
  1483  	PXOR	X9, X8
  1484  	// X15 must be zero on return
  1485  	PXOR	X15, X15
  1486  	MOVQ	X8, AX	// return X8
  1487  	RET
  1488  
  1489  aes129plus:
  1490  	// make 7 more starting seeds
  1491  	MOVO	X1, X2
  1492  	MOVO	X1, X3
  1493  	MOVO	X1, X4
  1494  	MOVO	X1, X5
  1495  	MOVO	X1, X6
  1496  	MOVO	X1, X7
  1497  	PXOR	runtime·aeskeysched+16(SB), X1
  1498  	PXOR	runtime·aeskeysched+32(SB), X2
  1499  	PXOR	runtime·aeskeysched+48(SB), X3
  1500  	PXOR	runtime·aeskeysched+64(SB), X4
  1501  	PXOR	runtime·aeskeysched+80(SB), X5
  1502  	PXOR	runtime·aeskeysched+96(SB), X6
  1503  	PXOR	runtime·aeskeysched+112(SB), X7
  1504  	AESENC	X1, X1
  1505  	AESENC	X2, X2
  1506  	AESENC	X3, X3
  1507  	AESENC	X4, X4
  1508  	AESENC	X5, X5
  1509  	AESENC	X6, X6
  1510  	AESENC	X7, X7
  1511  
  1512  	// start with last (possibly overlapping) block
  1513  	MOVOU	-128(AX)(CX*1), X8
  1514  	MOVOU	-112(AX)(CX*1), X9
  1515  	MOVOU	-96(AX)(CX*1), X10
  1516  	MOVOU	-80(AX)(CX*1), X11
  1517  	MOVOU	-64(AX)(CX*1), X12
  1518  	MOVOU	-48(AX)(CX*1), X13
  1519  	MOVOU	-32(AX)(CX*1), X14
  1520  	MOVOU	-16(AX)(CX*1), X15
  1521  
  1522  	// xor in seed
  1523  	PXOR	X0, X8
  1524  	PXOR	X1, X9
  1525  	PXOR	X2, X10
  1526  	PXOR	X3, X11
  1527  	PXOR	X4, X12
  1528  	PXOR	X5, X13
  1529  	PXOR	X6, X14
  1530  	PXOR	X7, X15
  1531  
  1532  	// compute number of remaining 128-byte blocks
  1533  	DECQ	CX
  1534  	SHRQ	$7, CX
  1535  
  1536  	PCALIGN $16
  1537  aesloop:
  1538  	// scramble state
  1539  	AESENC	X8, X8
  1540  	AESENC	X9, X9
  1541  	AESENC	X10, X10
  1542  	AESENC	X11, X11
  1543  	AESENC	X12, X12
  1544  	AESENC	X13, X13
  1545  	AESENC	X14, X14
  1546  	AESENC	X15, X15
  1547  
  1548  	// scramble state, xor in a block
  1549  	MOVOU	(AX), X0
  1550  	MOVOU	16(AX), X1
  1551  	MOVOU	32(AX), X2
  1552  	MOVOU	48(AX), X3
  1553  	AESENC	X0, X8
  1554  	AESENC	X1, X9
  1555  	AESENC	X2, X10
  1556  	AESENC	X3, X11
  1557  	MOVOU	64(AX), X4
  1558  	MOVOU	80(AX), X5
  1559  	MOVOU	96(AX), X6
  1560  	MOVOU	112(AX), X7
  1561  	AESENC	X4, X12
  1562  	AESENC	X5, X13
  1563  	AESENC	X6, X14
  1564  	AESENC	X7, X15
  1565  
  1566  	ADDQ	$128, AX
  1567  	DECQ	CX
  1568  	JNE	aesloop
  1569  
  1570  	// 3 more scrambles to finish
  1571  	AESENC	X8, X8
  1572  	AESENC	X9, X9
  1573  	AESENC	X10, X10
  1574  	AESENC	X11, X11
  1575  	AESENC	X12, X12
  1576  	AESENC	X13, X13
  1577  	AESENC	X14, X14
  1578  	AESENC	X15, X15
  1579  	AESENC	X8, X8
  1580  	AESENC	X9, X9
  1581  	AESENC	X10, X10
  1582  	AESENC	X11, X11
  1583  	AESENC	X12, X12
  1584  	AESENC	X13, X13
  1585  	AESENC	X14, X14
  1586  	AESENC	X15, X15
  1587  	AESENC	X8, X8
  1588  	AESENC	X9, X9
  1589  	AESENC	X10, X10
  1590  	AESENC	X11, X11
  1591  	AESENC	X12, X12
  1592  	AESENC	X13, X13
  1593  	AESENC	X14, X14
  1594  	AESENC	X15, X15
  1595  
  1596  	PXOR	X12, X8
  1597  	PXOR	X13, X9
  1598  	PXOR	X14, X10
  1599  	PXOR	X15, X11
  1600  	PXOR	X10, X8
  1601  	PXOR	X11, X9
  1602  	PXOR	X9, X8
  1603  	// X15 must be zero on return
  1604  	PXOR	X15, X15
  1605  	MOVQ	X8, AX	// return X8
  1606  	RET
  1607  
  1608  // func memhash32(p unsafe.Pointer, h uintptr) uintptr
  1609  // ABIInternal for performance.
  1610  TEXT runtime·memhash32<ABIInternal>(SB),NOSPLIT,$0-24
  1611  	// AX = ptr to data
  1612  	// BX = seed
  1613  	CMPB	runtime·useAeshash(SB), $0
  1614  	JEQ	noaes
  1615  	MOVQ	BX, X0	// X0 = seed
  1616  	PINSRD	$2, (AX), X0	// data
  1617  	AESENC	runtime·aeskeysched+0(SB), X0
  1618  	AESENC	runtime·aeskeysched+16(SB), X0
  1619  	AESENC	runtime·aeskeysched+32(SB), X0
  1620  	MOVQ	X0, AX	// return X0
  1621  	RET
  1622  noaes:
  1623  	JMP	runtime·memhash32Fallback<ABIInternal>(SB)
  1624  
  1625  // func memhash64(p unsafe.Pointer, h uintptr) uintptr
  1626  // ABIInternal for performance.
  1627  TEXT runtime·memhash64<ABIInternal>(SB),NOSPLIT,$0-24
  1628  	// AX = ptr to data
  1629  	// BX = seed
  1630  	CMPB	runtime·useAeshash(SB), $0
  1631  	JEQ	noaes
  1632  	MOVQ	BX, X0	// X0 = seed
  1633  	PINSRQ	$1, (AX), X0	// data
  1634  	AESENC	runtime·aeskeysched+0(SB), X0
  1635  	AESENC	runtime·aeskeysched+16(SB), X0
  1636  	AESENC	runtime·aeskeysched+32(SB), X0
  1637  	MOVQ	X0, AX	// return X0
  1638  	RET
  1639  noaes:
  1640  	JMP	runtime·memhash64Fallback<ABIInternal>(SB)
  1641  
  1642  // simple mask to get rid of data in the high part of the register.
  1643  DATA masks<>+0x00(SB)/8, $0x0000000000000000
  1644  DATA masks<>+0x08(SB)/8, $0x0000000000000000
  1645  DATA masks<>+0x10(SB)/8, $0x00000000000000ff
  1646  DATA masks<>+0x18(SB)/8, $0x0000000000000000
  1647  DATA masks<>+0x20(SB)/8, $0x000000000000ffff
  1648  DATA masks<>+0x28(SB)/8, $0x0000000000000000
  1649  DATA masks<>+0x30(SB)/8, $0x0000000000ffffff
  1650  DATA masks<>+0x38(SB)/8, $0x0000000000000000
  1651  DATA masks<>+0x40(SB)/8, $0x00000000ffffffff
  1652  DATA masks<>+0x48(SB)/8, $0x0000000000000000
  1653  DATA masks<>+0x50(SB)/8, $0x000000ffffffffff
  1654  DATA masks<>+0x58(SB)/8, $0x0000000000000000
  1655  DATA masks<>+0x60(SB)/8, $0x0000ffffffffffff
  1656  DATA masks<>+0x68(SB)/8, $0x0000000000000000
  1657  DATA masks<>+0x70(SB)/8, $0x00ffffffffffffff
  1658  DATA masks<>+0x78(SB)/8, $0x0000000000000000
  1659  DATA masks<>+0x80(SB)/8, $0xffffffffffffffff
  1660  DATA masks<>+0x88(SB)/8, $0x0000000000000000
  1661  DATA masks<>+0x90(SB)/8, $0xffffffffffffffff
  1662  DATA masks<>+0x98(SB)/8, $0x00000000000000ff
  1663  DATA masks<>+0xa0(SB)/8, $0xffffffffffffffff
  1664  DATA masks<>+0xa8(SB)/8, $0x000000000000ffff
  1665  DATA masks<>+0xb0(SB)/8, $0xffffffffffffffff
  1666  DATA masks<>+0xb8(SB)/8, $0x0000000000ffffff
  1667  DATA masks<>+0xc0(SB)/8, $0xffffffffffffffff
  1668  DATA masks<>+0xc8(SB)/8, $0x00000000ffffffff
  1669  DATA masks<>+0xd0(SB)/8, $0xffffffffffffffff
  1670  DATA masks<>+0xd8(SB)/8, $0x000000ffffffffff
  1671  DATA masks<>+0xe0(SB)/8, $0xffffffffffffffff
  1672  DATA masks<>+0xe8(SB)/8, $0x0000ffffffffffff
  1673  DATA masks<>+0xf0(SB)/8, $0xffffffffffffffff
  1674  DATA masks<>+0xf8(SB)/8, $0x00ffffffffffffff
  1675  GLOBL masks<>(SB),RODATA,$256
  1676  
  1677  // func checkASM() bool
  1678  TEXT ·checkASM(SB),NOSPLIT,$0-1
  1679  	// check that masks<>(SB) and shifts<>(SB) are aligned to 16-byte
  1680  	MOVQ	$masks<>(SB), AX
  1681  	MOVQ	$shifts<>(SB), BX
  1682  	ORQ	BX, AX
  1683  	TESTQ	$15, AX
  1684  	SETEQ	ret+0(FP)
  1685  	RET
  1686  
  1687  // these are arguments to pshufb. They move data down from
  1688  // the high bytes of the register to the low bytes of the register.
  1689  // index is how many bytes to move.
  1690  DATA shifts<>+0x00(SB)/8, $0x0000000000000000
  1691  DATA shifts<>+0x08(SB)/8, $0x0000000000000000
  1692  DATA shifts<>+0x10(SB)/8, $0xffffffffffffff0f
  1693  DATA shifts<>+0x18(SB)/8, $0xffffffffffffffff
  1694  DATA shifts<>+0x20(SB)/8, $0xffffffffffff0f0e
  1695  DATA shifts<>+0x28(SB)/8, $0xffffffffffffffff
  1696  DATA shifts<>+0x30(SB)/8, $0xffffffffff0f0e0d
  1697  DATA shifts<>+0x38(SB)/8, $0xffffffffffffffff
  1698  DATA shifts<>+0x40(SB)/8, $0xffffffff0f0e0d0c
  1699  DATA shifts<>+0x48(SB)/8, $0xffffffffffffffff
  1700  DATA shifts<>+0x50(SB)/8, $0xffffff0f0e0d0c0b
  1701  DATA shifts<>+0x58(SB)/8, $0xffffffffffffffff
  1702  DATA shifts<>+0x60(SB)/8, $0xffff0f0e0d0c0b0a
  1703  DATA shifts<>+0x68(SB)/8, $0xffffffffffffffff
  1704  DATA shifts<>+0x70(SB)/8, $0xff0f0e0d0c0b0a09
  1705  DATA shifts<>+0x78(SB)/8, $0xffffffffffffffff
  1706  DATA shifts<>+0x80(SB)/8, $0x0f0e0d0c0b0a0908
  1707  DATA shifts<>+0x88(SB)/8, $0xffffffffffffffff
  1708  DATA shifts<>+0x90(SB)/8, $0x0e0d0c0b0a090807
  1709  DATA shifts<>+0x98(SB)/8, $0xffffffffffffff0f
  1710  DATA shifts<>+0xa0(SB)/8, $0x0d0c0b0a09080706
  1711  DATA shifts<>+0xa8(SB)/8, $0xffffffffffff0f0e
  1712  DATA shifts<>+0xb0(SB)/8, $0x0c0b0a0908070605
  1713  DATA shifts<>+0xb8(SB)/8, $0xffffffffff0f0e0d
  1714  DATA shifts<>+0xc0(SB)/8, $0x0b0a090807060504
  1715  DATA shifts<>+0xc8(SB)/8, $0xffffffff0f0e0d0c
  1716  DATA shifts<>+0xd0(SB)/8, $0x0a09080706050403
  1717  DATA shifts<>+0xd8(SB)/8, $0xffffff0f0e0d0c0b
  1718  DATA shifts<>+0xe0(SB)/8, $0x0908070605040302
  1719  DATA shifts<>+0xe8(SB)/8, $0xffff0f0e0d0c0b0a
  1720  DATA shifts<>+0xf0(SB)/8, $0x0807060504030201
  1721  DATA shifts<>+0xf8(SB)/8, $0xff0f0e0d0c0b0a09
  1722  GLOBL shifts<>(SB),RODATA,$256
  1723  
  1724  // Called from cgo wrappers, this function returns g->m->curg.stack.hi.
  1725  // Must obey the gcc calling convention.
  1726  TEXT _cgo_topofstack(SB),NOSPLIT,$0
  1727  	get_tls(CX)
  1728  	MOVQ	g(CX), AX
  1729  	MOVQ	g_m(AX), AX
  1730  	MOVQ	m_curg(AX), AX
  1731  	MOVQ	(g_stack+stack_hi)(AX), AX
  1732  	RET
  1733  
  1734  // The top-most function running on a goroutine
  1735  // returns to goexit+PCQuantum.
  1736  TEXT runtime·goexit(SB),NOSPLIT|TOPFRAME|NOFRAME,$0-0
  1737  	BYTE	$0x90	// NOP
  1738  	CALL	runtime·goexit1(SB)	// does not return
  1739  	// traceback from goexit1 must hit code range of goexit
  1740  	BYTE	$0x90	// NOP
  1741  
  1742  // This is called from .init_array and follows the platform, not Go, ABI.
  1743  TEXT runtime·addmoduledata(SB),NOSPLIT,$0-0
  1744  	PUSHQ	R15 // The access to global variables below implicitly uses R15, which is callee-save
  1745  	MOVQ	runtime·lastmoduledatap(SB), AX
  1746  	MOVQ	DI, moduledata_next(AX)
  1747  	MOVQ	DI, runtime·lastmoduledatap(SB)
  1748  	POPQ	R15
  1749  	RET
  1750  
  1751  // Initialize special registers then jump to sigpanic.
  1752  // This function is injected from the signal handler for panicking
  1753  // signals. It is quite painful to set X15 in the signal context,
  1754  // so we do it here.
  1755  TEXT ·sigpanic0(SB),NOSPLIT,$0-0
  1756  	get_tls(R14)
  1757  	MOVQ	g(R14), R14
  1758  	XORPS	X15, X15
  1759  	JMP	·sigpanic<ABIInternal>(SB)
  1760  
  1761  // gcWriteBarrier informs the GC about heap pointer writes.
  1762  //
  1763  // gcWriteBarrier returns space in a write barrier buffer which
  1764  // should be filled in by the caller.
  1765  // gcWriteBarrier does NOT follow the Go ABI. It accepts the
  1766  // number of bytes of buffer needed in R11, and returns a pointer
  1767  // to the buffer space in R11.
  1768  // It clobbers FLAGS. It does not clobber any general-purpose registers,
  1769  // but may clobber others (e.g., SSE registers).
  1770  // Typical use would be, when doing *(CX+88) = AX
  1771  //     CMPL    $0, runtime.writeBarrier(SB)
  1772  //     JEQ     dowrite
  1773  //     CALL    runtime.gcBatchBarrier2(SB)
  1774  //     MOVQ    AX, (R11)
  1775  //     MOVQ    88(CX), DX
  1776  //     MOVQ    DX, 8(R11)
  1777  // dowrite:
  1778  //     MOVQ    AX, 88(CX)
  1779  TEXT gcWriteBarrier<>(SB),NOSPLIT,$112
  1780  	// Save the registers clobbered by the fast path. This is slightly
  1781  	// faster than having the caller spill these.
  1782  	MOVQ	R12, 96(SP)
  1783  	MOVQ	R13, 104(SP)
  1784  retry:
  1785  	// TODO: Consider passing g.m.p in as an argument so they can be shared
  1786  	// across a sequence of write barriers.
  1787  	MOVQ	g_m(R14), R13
  1788  	MOVQ	m_p(R13), R13
  1789  	// Get current buffer write position.
  1790  	MOVQ	(p_wbBuf+wbBuf_next)(R13), R12	// original next position
  1791  	ADDQ	R11, R12			// new next position
  1792  	// Is the buffer full?
  1793  	CMPQ	R12, (p_wbBuf+wbBuf_end)(R13)
  1794  	JA	flush
  1795  	// Commit to the larger buffer.
  1796  	MOVQ	R12, (p_wbBuf+wbBuf_next)(R13)
  1797  	// Make return value (the original next position)
  1798  	SUBQ	R11, R12
  1799  	MOVQ	R12, R11
  1800  	// Restore registers.
  1801  	MOVQ	96(SP), R12
  1802  	MOVQ	104(SP), R13
  1803  	RET
  1804  
  1805  flush:
  1806  	// Save all general purpose registers since these could be
  1807  	// clobbered by wbBufFlush and were not saved by the caller.
  1808  	// It is possible for wbBufFlush to clobber other registers
  1809  	// (e.g., SSE registers), but the compiler takes care of saving
  1810  	// those in the caller if necessary. This strikes a balance
  1811  	// with registers that are likely to be used.
  1812  	//
  1813  	// We don't have type information for these, but all code under
  1814  	// here is NOSPLIT, so nothing will observe these.
  1815  	//
  1816  	// TODO: We could strike a different balance; e.g., saving X0
  1817  	// and not saving GP registers that are less likely to be used.
  1818  	MOVQ	DI, 0(SP)
  1819  	MOVQ	AX, 8(SP)
  1820  	MOVQ	BX, 16(SP)
  1821  	MOVQ	CX, 24(SP)
  1822  	MOVQ	DX, 32(SP)
  1823  	// DI already saved
  1824  	MOVQ	SI, 40(SP)
  1825  	MOVQ	BP, 48(SP)
  1826  	MOVQ	R8, 56(SP)
  1827  	MOVQ	R9, 64(SP)
  1828  	MOVQ	R10, 72(SP)
  1829  	MOVQ	R11, 80(SP)
  1830  	// R12 already saved
  1831  	// R13 already saved
  1832  	// R14 is g
  1833  	MOVQ	R15, 88(SP)
  1834  
  1835  	CALL	runtime·wbBufFlush(SB)
  1836  
  1837  	MOVQ	0(SP), DI
  1838  	MOVQ	8(SP), AX
  1839  	MOVQ	16(SP), BX
  1840  	MOVQ	24(SP), CX
  1841  	MOVQ	32(SP), DX
  1842  	MOVQ	40(SP), SI
  1843  	MOVQ	48(SP), BP
  1844  	MOVQ	56(SP), R8
  1845  	MOVQ	64(SP), R9
  1846  	MOVQ	72(SP), R10
  1847  	MOVQ	80(SP), R11
  1848  	MOVQ	88(SP), R15
  1849  	JMP	retry
  1850  
  1851  TEXT runtime·gcWriteBarrier1<ABIInternal>(SB),NOSPLIT|NOFRAME,$0
  1852  	MOVL   $8, R11
  1853  	JMP     gcWriteBarrier<>(SB)
  1854  TEXT runtime·gcWriteBarrier2<ABIInternal>(SB),NOSPLIT|NOFRAME,$0
  1855  	MOVL   $16, R11
  1856  	JMP     gcWriteBarrier<>(SB)
  1857  TEXT runtime·gcWriteBarrier3<ABIInternal>(SB),NOSPLIT|NOFRAME,$0
  1858  	MOVL   $24, R11
  1859  	JMP     gcWriteBarrier<>(SB)
  1860  TEXT runtime·gcWriteBarrier4<ABIInternal>(SB),NOSPLIT|NOFRAME,$0
  1861  	MOVL   $32, R11
  1862  	JMP     gcWriteBarrier<>(SB)
  1863  TEXT runtime·gcWriteBarrier5<ABIInternal>(SB),NOSPLIT|NOFRAME,$0
  1864  	MOVL   $40, R11
  1865  	JMP     gcWriteBarrier<>(SB)
  1866  TEXT runtime·gcWriteBarrier6<ABIInternal>(SB),NOSPLIT|NOFRAME,$0
  1867  	MOVL   $48, R11
  1868  	JMP     gcWriteBarrier<>(SB)
  1869  TEXT runtime·gcWriteBarrier7<ABIInternal>(SB),NOSPLIT|NOFRAME,$0
  1870  	MOVL   $56, R11
  1871  	JMP     gcWriteBarrier<>(SB)
  1872  TEXT runtime·gcWriteBarrier8<ABIInternal>(SB),NOSPLIT|NOFRAME,$0
  1873  	MOVL   $64, R11
  1874  	JMP     gcWriteBarrier<>(SB)
  1875  
  1876  DATA	debugCallFrameTooLarge<>+0x00(SB)/20, $"call frame too large"
  1877  GLOBL	debugCallFrameTooLarge<>(SB), RODATA, $20	// Size duplicated below
  1878  
  1879  // debugCallV2 is the entry point for debugger-injected function
  1880  // calls on running goroutines. It informs the runtime that a
  1881  // debug call has been injected and creates a call frame for the
  1882  // debugger to fill in.
  1883  //
  1884  // To inject a function call, a debugger should:
  1885  // 1. Check that the goroutine is in state _Grunning and that
  1886  //    there are at least 256 bytes free on the stack.
  1887  // 2. Push the current PC on the stack (updating SP).
  1888  // 3. Write the desired argument frame size at SP-16 (using the SP
  1889  //    after step 2).
  1890  // 4. Save all machine registers (including flags and XMM registers)
  1891  //    so they can be restored later by the debugger.
  1892  // 5. Set the PC to debugCallV2 and resume execution.
  1893  //
  1894  // If the goroutine is in state _Grunnable, then it's not generally
  1895  // safe to inject a call because it may return out via other runtime
  1896  // operations. Instead, the debugger should unwind the stack to find
  1897  // the return to non-runtime code, add a temporary breakpoint there,
  1898  // and inject the call once that breakpoint is hit.
  1899  //
  1900  // If the goroutine is in any other state, it's not safe to inject a call.
  1901  //
  1902  // This function communicates back to the debugger by setting R12 and
  1903  // invoking INT3 to raise a breakpoint signal. See the comments in the
  1904  // implementation for the protocol the debugger is expected to
  1905  // follow. InjectDebugCall in the runtime tests demonstrates this protocol.
  1906  //
  1907  // The debugger must ensure that any pointers passed to the function
  1908  // obey escape analysis requirements. Specifically, it must not pass
  1909  // a stack pointer to an escaping argument. debugCallV2 cannot check
  1910  // this invariant.
  1911  //
  1912  // This is ABIInternal because Go code injects its PC directly into new
  1913  // goroutine stacks.
  1914  TEXT runtime·debugCallV2<ABIInternal>(SB),NOSPLIT,$152-0
  1915  	// Save all registers that may contain pointers so they can be
  1916  	// conservatively scanned.
  1917  	//
  1918  	// We can't do anything that might clobber any of these
  1919  	// registers before this.
  1920  	MOVQ	R15, r15-(14*8+8)(SP)
  1921  	MOVQ	R14, r14-(13*8+8)(SP)
  1922  	MOVQ	R13, r13-(12*8+8)(SP)
  1923  	MOVQ	R12, r12-(11*8+8)(SP)
  1924  	MOVQ	R11, r11-(10*8+8)(SP)
  1925  	MOVQ	R10, r10-(9*8+8)(SP)
  1926  	MOVQ	R9, r9-(8*8+8)(SP)
  1927  	MOVQ	R8, r8-(7*8+8)(SP)
  1928  	MOVQ	DI, di-(6*8+8)(SP)
  1929  	MOVQ	SI, si-(5*8+8)(SP)
  1930  	MOVQ	BP, bp-(4*8+8)(SP)
  1931  	MOVQ	BX, bx-(3*8+8)(SP)
  1932  	MOVQ	DX, dx-(2*8+8)(SP)
  1933  	// Save the frame size before we clobber it. Either of the last
  1934  	// saves could clobber this depending on whether there's a saved BP.
  1935  	MOVQ	frameSize-24(FP), DX	// aka -16(RSP) before prologue
  1936  	MOVQ	CX, cx-(1*8+8)(SP)
  1937  	MOVQ	AX, ax-(0*8+8)(SP)
  1938  
  1939  	// Save the argument frame size.
  1940  	MOVQ	DX, frameSize-128(SP)
  1941  
  1942  	// Perform a safe-point check.
  1943  	MOVQ	retpc-8(FP), AX	// Caller's PC
  1944  	MOVQ	AX, 0(SP)
  1945  	CALL	runtime·debugCallCheck(SB)
  1946  	MOVQ	8(SP), AX
  1947  	TESTQ	AX, AX
  1948  	JZ	good
  1949  	// The safety check failed. Put the reason string at the top
  1950  	// of the stack.
  1951  	MOVQ	AX, 0(SP)
  1952  	MOVQ	16(SP), AX
  1953  	MOVQ	AX, 8(SP)
  1954  	// Set R12 to 8 and invoke INT3. The debugger should get the
  1955  	// reason a call can't be injected from the top of the stack
  1956  	// and resume execution.
  1957  	MOVQ	$8, R12
  1958  	BYTE	$0xcc
  1959  	JMP	restore
  1960  
  1961  good:
  1962  	// Registers are saved and it's safe to make a call.
  1963  	// Open up a call frame, moving the stack if necessary.
  1964  	//
  1965  	// Once the frame is allocated, this will set R12 to 0 and
  1966  	// invoke INT3. The debugger should write the argument
  1967  	// frame for the call at SP, set up argument registers, push
  1968  	// the trapping PC on the stack, set the PC to the function to
  1969  	// call, set RDX to point to the closure (if a closure call),
  1970  	// and resume execution.
  1971  	//
  1972  	// If the function returns, this will set R12 to 1 and invoke
  1973  	// INT3. The debugger can then inspect any return value saved
  1974  	// on the stack at SP and in registers and resume execution again.
  1975  	//
  1976  	// If the function panics, this will set R12 to 2 and invoke INT3.
  1977  	// The interface{} value of the panic will be at SP. The debugger
  1978  	// can inspect the panic value and resume execution again.
  1979  #define DEBUG_CALL_DISPATCH(NAME,MAXSIZE)	\
  1980  	CMPQ	AX, $MAXSIZE;			\
  1981  	JA	5(PC);				\
  1982  	MOVQ	$NAME(SB), AX;			\
  1983  	MOVQ	AX, 0(SP);			\
  1984  	CALL	runtime·debugCallWrap(SB);	\
  1985  	JMP	restore
  1986  
  1987  	MOVQ	frameSize-128(SP), AX
  1988  	DEBUG_CALL_DISPATCH(debugCall32<>, 32)
  1989  	DEBUG_CALL_DISPATCH(debugCall64<>, 64)
  1990  	DEBUG_CALL_DISPATCH(debugCall128<>, 128)
  1991  	DEBUG_CALL_DISPATCH(debugCall256<>, 256)
  1992  	DEBUG_CALL_DISPATCH(debugCall512<>, 512)
  1993  	DEBUG_CALL_DISPATCH(debugCall1024<>, 1024)
  1994  	DEBUG_CALL_DISPATCH(debugCall2048<>, 2048)
  1995  	DEBUG_CALL_DISPATCH(debugCall4096<>, 4096)
  1996  	DEBUG_CALL_DISPATCH(debugCall8192<>, 8192)
  1997  	DEBUG_CALL_DISPATCH(debugCall16384<>, 16384)
  1998  	DEBUG_CALL_DISPATCH(debugCall32768<>, 32768)
  1999  	DEBUG_CALL_DISPATCH(debugCall65536<>, 65536)
  2000  	// The frame size is too large. Report the error.
  2001  	MOVQ	$debugCallFrameTooLarge<>(SB), AX
  2002  	MOVQ	AX, 0(SP)
  2003  	MOVQ	$20, 8(SP) // length of debugCallFrameTooLarge string
  2004  	MOVQ	$8, R12
  2005  	BYTE	$0xcc
  2006  	JMP	restore
  2007  
  2008  restore:
  2009  	// Calls and failures resume here.
  2010  	//
  2011  	// Set R12 to 16 and invoke INT3. The debugger should restore
  2012  	// all registers except RIP and RSP and resume execution.
  2013  	MOVQ	$16, R12
  2014  	BYTE	$0xcc
  2015  	// We must not modify flags after this point.
  2016  
  2017  	// Restore pointer-containing registers, which may have been
  2018  	// modified from the debugger's copy by stack copying.
  2019  	MOVQ	ax-(0*8+8)(SP), AX
  2020  	MOVQ	cx-(1*8+8)(SP), CX
  2021  	MOVQ	dx-(2*8+8)(SP), DX
  2022  	MOVQ	bx-(3*8+8)(SP), BX
  2023  	MOVQ	bp-(4*8+8)(SP), BP
  2024  	MOVQ	si-(5*8+8)(SP), SI
  2025  	MOVQ	di-(6*8+8)(SP), DI
  2026  	MOVQ	r8-(7*8+8)(SP), R8
  2027  	MOVQ	r9-(8*8+8)(SP), R9
  2028  	MOVQ	r10-(9*8+8)(SP), R10
  2029  	MOVQ	r11-(10*8+8)(SP), R11
  2030  	MOVQ	r12-(11*8+8)(SP), R12
  2031  	MOVQ	r13-(12*8+8)(SP), R13
  2032  	MOVQ	r14-(13*8+8)(SP), R14
  2033  	MOVQ	r15-(14*8+8)(SP), R15
  2034  
  2035  	RET
  2036  
  2037  // runtime.debugCallCheck assumes that functions defined with the
  2038  // DEBUG_CALL_FN macro are safe points to inject calls.
  2039  #define DEBUG_CALL_FN(NAME,MAXSIZE)		\
  2040  TEXT NAME(SB),WRAPPER,$MAXSIZE-0;		\
  2041  	NO_LOCAL_POINTERS;			\
  2042  	MOVQ	$0, R12;				\
  2043  	BYTE	$0xcc;				\
  2044  	MOVQ	$1, R12;				\
  2045  	BYTE	$0xcc;				\
  2046  	RET
  2047  DEBUG_CALL_FN(debugCall32<>, 32)
  2048  DEBUG_CALL_FN(debugCall64<>, 64)
  2049  DEBUG_CALL_FN(debugCall128<>, 128)
  2050  DEBUG_CALL_FN(debugCall256<>, 256)
  2051  DEBUG_CALL_FN(debugCall512<>, 512)
  2052  DEBUG_CALL_FN(debugCall1024<>, 1024)
  2053  DEBUG_CALL_FN(debugCall2048<>, 2048)
  2054  DEBUG_CALL_FN(debugCall4096<>, 4096)
  2055  DEBUG_CALL_FN(debugCall8192<>, 8192)
  2056  DEBUG_CALL_FN(debugCall16384<>, 16384)
  2057  DEBUG_CALL_FN(debugCall32768<>, 32768)
  2058  DEBUG_CALL_FN(debugCall65536<>, 65536)
  2059  
  2060  // func debugCallPanicked(val interface{})
  2061  TEXT runtime·debugCallPanicked(SB),NOSPLIT,$16-16
  2062  	// Copy the panic value to the top of stack.
  2063  	MOVQ	val_type+0(FP), AX
  2064  	MOVQ	AX, 0(SP)
  2065  	MOVQ	val_data+8(FP), AX
  2066  	MOVQ	AX, 8(SP)
  2067  	MOVQ	$2, R12
  2068  	BYTE	$0xcc
  2069  	RET
  2070  
  2071  TEXT runtime·panicBounds<ABIInternal>(SB),NOSPLIT,$144-0
  2072  	NO_LOCAL_POINTERS
  2073  	// Save all 14 int registers that could have an index in them.
  2074  	// They may be pointers, but if they are they are dead.
  2075  	MOVQ	AX, 16(SP)
  2076  	MOVQ	CX, 24(SP)
  2077  	MOVQ	DX, 32(SP)
  2078  	MOVQ	BX, 40(SP)
  2079  	// skip SP @ 48(SP)
  2080  	MOVQ	BP, 56(SP)
  2081  	MOVQ	SI, 64(SP)
  2082  	MOVQ	DI, 72(SP)
  2083  	MOVQ	R8, 80(SP)
  2084  	MOVQ	R9, 88(SP)
  2085  	MOVQ	R10, 96(SP)
  2086  	MOVQ	R11, 104(SP)
  2087  	MOVQ	R12, 112(SP)
  2088  	MOVQ	R13, 120(SP)
  2089  	// skip R14 @ 128(SP) (aka G)
  2090  	MOVQ	R15, 136(SP)
  2091  
  2092  	MOVQ	SP, AX		// hide SP read from vet
  2093  	MOVQ	152(AX), AX	// PC immediately after call to panicBounds
  2094  	LEAQ	16(SP), BX
  2095  	CALL	runtime·panicBounds64<ABIInternal>(SB)
  2096  	RET
  2097  
  2098  #ifdef GOOS_android
  2099  // Use the free TLS_SLOT_APP slot #2 on Android Q.
  2100  // Earlier androids are set up in gcc_android.c.
  2101  DATA runtime·tls_g+0(SB)/8, $16
  2102  GLOBL runtime·tls_g+0(SB), NOPTR, $8
  2103  #endif
  2104  #ifdef GOOS_windows
  2105  GLOBL runtime·tls_g+0(SB), NOPTR, $8
  2106  #endif
  2107  
  2108  // The compiler and assembler's -spectre=ret mode rewrites
  2109  // all indirect CALL AX / JMP AX instructions to be
  2110  // CALL retpolineAX / JMP retpolineAX.
  2111  // See https://support.google.com/faqs/answer/7625886.
  2112  #define RETPOLINE(reg) \
  2113  	/*   CALL setup */     BYTE $0xE8; BYTE $(2+2); BYTE $0; BYTE $0; BYTE $0;	\
  2114  	/* nospec: */									\
  2115  	/*   PAUSE */           BYTE $0xF3; BYTE $0x90;					\
  2116  	/*   JMP nospec */      BYTE $0xEB; BYTE $-(2+2);				\
  2117  	/* setup: */									\
  2118  	/*   MOVQ AX, 0(SP) */  BYTE $0x48|((reg&8)>>1); BYTE $0x89;			\
  2119  	                        BYTE $0x04|((reg&7)<<3); BYTE $0x24;			\
  2120  	/*   RET */             BYTE $0xC3
  2121  
  2122  TEXT runtime·retpolineAX(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(0)
  2123  TEXT runtime·retpolineCX(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(1)
  2124  TEXT runtime·retpolineDX(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(2)
  2125  TEXT runtime·retpolineBX(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(3)
  2126  /* SP is 4, can't happen / magic encodings */
  2127  TEXT runtime·retpolineBP(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(5)
  2128  TEXT runtime·retpolineSI(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(6)
  2129  TEXT runtime·retpolineDI(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(7)
  2130  TEXT runtime·retpolineR8(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(8)
  2131  TEXT runtime·retpolineR9(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(9)
  2132  TEXT runtime·retpolineR10(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(10)
  2133  TEXT runtime·retpolineR11(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(11)
  2134  TEXT runtime·retpolineR12(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(12)
  2135  TEXT runtime·retpolineR13(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(13)
  2136  TEXT runtime·retpolineR14(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(14)
  2137  TEXT runtime·retpolineR15(SB),NOSPLIT|NOFRAME,$0; RETPOLINE(15)
  2138  
  2139  TEXT ·getfp<ABIInternal>(SB),NOSPLIT|NOFRAME,$0
  2140  	MOVQ BP, AX
  2141  	RET
  2142  

View as plain text