+#
+\f#################################################
+# Win64 SEH, Structured Exception Handling.
+#
+# Unlike on Unix systems(*) lack of Win64 stack unwinding information
+# has undesired side-effect at run-time: if an exception is raised in
+# assembler subroutine such as those in question (basically we're
+# referring to segmentation violations caused by malformed input
+# parameters), the application is briskly terminated without invoking
+# any exception handlers, most notably without generating memory dump
+# or any user notification whatsoever. This poses a problem. It's
+# possible to address it by registering custom language-specific
+# handler that would restore processor context to the state at
+# subroutine entry point and return "exception is not handled, keep
+# unwinding" code. Writing such handler can be a challenge... But it's
+# doable, though requires certain coding convention. Consider following
+# snippet:
+#
+# .type function,@function
+# function:
+# movq %rsp,%rax # copy rsp to volatile register
+# pushq %r15 # save non-volatile registers
+# pushq %rbx
+# pushq %rbp
+# movq %rsp,%r11
+# subq %rdi,%r11 # prepare [variable] stack frame
+# andq $-64,%r11
+# movq %rax,0(%r11) # check for exceptions
+# movq %r11,%rsp # allocate [variable] stack frame
+# movq %rax,0(%rsp) # save original rsp value
+# magic_point:
+# ...
+# movq 0(%rsp),%rcx # pull original rsp value
+# movq -24(%rcx),%rbp # restore non-volatile registers
+# movq -16(%rcx),%rbx
+# movq -8(%rcx),%r15
+# movq %rcx,%rsp # restore original rsp
+# ret
+# .size function,.-function
+#
+# The key is that up to magic_point copy of original rsp value remains
+# in chosen volatile register and no non-volatile register, except for
+# rsp, is modified. While past magic_point rsp remains constant till
+# the very end of the function. In this case custom language-specific
+# exception handler would look like this:
+#
+# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
+# CONTEXT *context,DISPATCHER_CONTEXT *disp)
+# { ULONG64 *rsp = (ULONG64 *)context->Rax;
+# if (context->Rip >= magic_point)
+# { rsp = ((ULONG64 **)context->Rsp)[0];
+# context->Rbp = rsp[-3];
+# context->Rbx = rsp[-2];
+# context->R15 = rsp[-1];
+# }
+# context->Rsp = (ULONG64)rsp;
+# context->Rdi = rsp[1];
+# context->Rsi = rsp[2];
+#
+# memcpy (disp->ContextRecord,context,sizeof(CONTEXT));
+# RtlVirtualUnwind(UNW_FLAG_NHANDLER,disp->ImageBase,
+# dips->ControlPc,disp->FunctionEntry,disp->ContextRecord,
+# &disp->HandlerData,&disp->EstablisherFrame,NULL);
+# return ExceptionContinueSearch;
+# }
+#
+# It's appropriate to implement this handler in assembler, directly in
+# function's module. In order to do that one has to know members'
+# offsets in CONTEXT and DISPATCHER_CONTEXT structures and some constant
+# values. Here they are:
+#
+# CONTEXT.Rax 120
+# CONTEXT.Rcx 128
+# CONTEXT.Rdx 136
+# CONTEXT.Rbx 144
+# CONTEXT.Rsp 152
+# CONTEXT.Rbp 160
+# CONTEXT.Rsi 168
+# CONTEXT.Rdi 176
+# CONTEXT.R8 184
+# CONTEXT.R9 192
+# CONTEXT.R10 200
+# CONTEXT.R11 208
+# CONTEXT.R12 216
+# CONTEXT.R13 224
+# CONTEXT.R14 232
+# CONTEXT.R15 240
+# CONTEXT.Rip 248
+# CONTEXT.Xmm6 512
+# sizeof(CONTEXT) 1232
+# DISPATCHER_CONTEXT.ControlPc 0
+# DISPATCHER_CONTEXT.ImageBase 8
+# DISPATCHER_CONTEXT.FunctionEntry 16
+# DISPATCHER_CONTEXT.EstablisherFrame 24
+# DISPATCHER_CONTEXT.TargetIp 32
+# DISPATCHER_CONTEXT.ContextRecord 40
+# DISPATCHER_CONTEXT.LanguageHandler 48
+# DISPATCHER_CONTEXT.HandlerData 56
+# UNW_FLAG_NHANDLER 0
+# ExceptionContinueSearch 1
+#
+# In order to tie the handler to the function one has to compose
+# couple of structures: one for .xdata segment and one for .pdata.
+#
+# UNWIND_INFO structure for .xdata segment would be
+#
+# function_unwind_info:
+# .byte 9,0,0,0
+# .rva handler
+#
+# This structure designates exception handler for a function with
+# zero-length prologue, no stack frame or frame register.
+#
+# To facilitate composing of .pdata structures, auto-generated "gear"
+# prologue copies rsp value to rax and denotes next instruction with
+# .LSEH_begin_{function_name} label. This essentially defines the SEH
+# styling rule mentioned in the beginning. Position of this label is
+# chosen in such manner that possible exceptions raised in the "gear"
+# prologue would be accounted to caller and unwound from latter's frame.
+# End of function is marked with respective .LSEH_end_{function_name}
+# label. To summarize, .pdata segment would contain
+#
+# .rva .LSEH_begin_function
+# .rva .LSEH_end_function
+# .rva function_unwind_info
+#
+# Reference to functon_unwind_info from .xdata segment is the anchor.
+# In case you wonder why references are 32-bit .rvas and not 64-bit
+# .quads. References put into these two segments are required to be
+# *relative* to the base address of the current binary module, a.k.a.
+# image base. No Win64 module, be it .exe or .dll, can be larger than
+# 2GB and thus such relative references can be and are accommodated in
+# 32 bits.
+#
+# Having reviewed the example function code, one can argue that "movq
+# %rsp,%rax" above is redundant. It is not! Keep in mind that on Unix
+# rax would contain an undefined value. If this "offends" you, use
+# another register and refrain from modifying rax till magic_point is
+# reached, i.e. as if it was a non-volatile register. If more registers
+# are required prior [variable] frame setup is completed, note that
+# nobody says that you can have only one "magic point." You can
+# "liberate" non-volatile registers by denoting last stack off-load
+# instruction and reflecting it in finer grade unwind logic in handler.
+# After all, isn't it why it's called *language-specific* handler...
+#
+# Attentive reader can notice that exceptions would be mishandled in
+# auto-generated "gear" epilogue. Well, exception effectively can't
+# occur there, because if memory area used by it was subject to
+# segmentation violation, then it would be raised upon call to the
+# function (and as already mentioned be accounted to caller, which is
+# not a problem). If you're still not comfortable, then define tail
+# "magic point" just prior ret instruction and have handler treat it...
+#
+# (*) Note that we're talking about run-time, not debug-time. Lack of
+# unwind information makes debugging hard on both Windows and
+# Unix. "Unlike" referes to the fact that on Unix signal handler
+# will always be invoked, core dumped and appropriate exit code
+# returned to parent (for user notification).