@c \input texinfo @c %**start of header @c @setfilename agentexpr.info @c @settitle GDB Agent Expressions @c @setchapternewpage off @c %**end of header @c Revision: $Id$ @node Agent Expressions @appendix The GDB Agent Expression Mechanism In some applications, it is not feasable for the debugger to interrupt the program's execution long enough for the developer to learn anything helpful about its behavior. If the program's correctness depends on its real-time behavior, delays introduced by a debugger might cause the program to fail, even when the code itself is correct. It is useful to be able to observe the program's behavior without interrupting it. Using GDB's @code{trace} and @code{collect} commands, the user can specify locations in the program, and arbitrary expressions to evaluate when those locations are reached. Later, using the @code{tfind} command, she can examine the values those expressions had when the program hit the trace points. The expressions may also denote objects in memory --- structures or arrays, for example --- whose values GDB should record; while visiting a particular tracepoint, the user may inspect those objects as if they were in memory at that moment. However, because GDB records these values without interacting with the user, it can do so quickly and unobtrusively, hopefully not disturbing the program's behavior. When GDB is debugging a remote target, the GDB @dfn{agent} code running on the target computes the values of the expressions itself. To avoid having a full symbolic expression evaluator on the agent, GDB translates expressions in the source language into a simpler bytecode language, and then sends the bytecode to the agent; the agent then executes the bytecode, and records the values for GDB to retrieve later. The bytecode language is simple; there are forty-odd opcodes, the bulk of which are the usual vocabulary of C operands (addition, subtraction, shifts, and so on) and various sizes of literals and memory reference operations. The bytecode interpreter operates strictly on machine-level values --- various sizes of integers and floating point numbers --- and requires no information about types or symbols; thus, the interpreter's internal data structures are simple, and each bytecode requires only a few native machine instructions to implement it. The interpreter is small, and strict limits on the memory and time required to evaluate an expression are easy to determine, making it suitable for use by the debugging agent in real-time applications. @menu * General Bytecode Design:: Overview of the interpreter. * Bytecode Descriptions:: What each one does. * Using Agent Expressions:: How agent expressions fit into the big picture. * Varying Target Capabilities:: How to discover what the target can do. * Tracing on Symmetrix:: Special info for implementation on EMC's boxes. * Rationale:: Why we did it this way. @end menu @c @node Rationale @c @section Rationale @node General Bytecode Design @section General Bytecode Design The agent represents bytecode expressions as an array of bytes. Each instruction is one byte long (thus the term @dfn{bytecode}). Some instructions are followed by operand bytes; for example, the @code{goto} instruction is followed by a destination for the jump. The bytecode interpreter is a stack-based machine; most instructions pop their operands off the stack, perform some operation, and push the result back on the stack for the next instruction to consume. Each element of the stack may contain either a integer or a floating point value; these values are as many bits wide as the largest integer that can be directly manipulated in the source language. Stack elements carry no record of their type; bytecode could push a value as an integer, then pop it as a floating point value. However, GDB will not generate code which does this. In C, one might define the type of a stack element as follows: @example union agent_val @{ LONGEST l; DOUBLEST d; @}; @end example @noindent where @code{LONGEST} and @code{DOUBLEST} are @code{typedef} names for the largest integer and floating point types on the machine. By the time the bytecode interpreter reaches the end of the expression, the value of the expression should be the only value left on the stack. For tracing applications, @code{trace} bytecodes in the expression will have recorded the necessary data, and the value on the stack may be discarded. For other applications, like conditional breakpoints, the value may be useful. Separate from the stack, the interpreter has two registers: @table @code @item pc The address of the next bytecode to execute. @item start The address of the start of the bytecode expression, necessary for interpreting the @code{goto} and @code{if_goto} instructions. @end table @noindent Neither of these registers is directly visible to the bytecode language itself, but they are useful for defining the meanings of the bytecode operations. There are no instructions to perform side effects on the running program, or call the program's functions; we assume that these expressions are only used for unobtrusive debugging, not for patching the running code. Most bytecode instructions do not distinguish between the various sizes of values, and operate on full-width values; the upper bits of the values are simply ignored, since they do not usually make a difference to the value computed. The exceptions to this rule are: @table @asis @item memory reference instructions (@code{ref}@var{n}) There are distinct instructions to fetch different word sizes from memory. Once on the stack, however, the values are treated as full-size integers. They may need to be sign-extended; the @code{ext} instruction exists for this purpose. @item the sign-extension instruction (@code{ext} @var{n}) These clearly need to know which portion of their operand is to be extended to occupy the full length of the word. @end table If the interpreter is unable to evaluate an expression completely for some reason (a memory location is inaccessible, or a divisor is zero, for example), we say that interpretation ``terminates with an error''. This means that the problem is reported back to the interpreter's caller in some helpful way. In general, code using agent expressions should assume that they may attempt to divide by zero, fetch arbitrary memory locations, and misbehave in other ways. Even complicated C expressions compile to a few bytecode instructions; for example, the expression @code{x + y * z} would typically produce code like the following, assuming that @code{x} and @code{y} live in registers, and @code{z} is a global variable holding a 32-bit @code{int}: @example reg 1 reg 2 const32 @i{address of z} ref32 ext 32 mul add end @end example In detail, these mean: @table @code @item reg 1 Push the value of register 1 (presumably holding @code{x}) onto the stack. @item reg 2 Push the value of register 2 (holding @code{y}). @item const32 @i{address of z} Push the address of @code{z} onto the stack. @item ref32 Fetch a 32-bit word from the address at the top of the stack; replace the address on the stack with the value. Thus, we replace the address of @code{z} with @code{z}'s value. @item ext 32 Sign-extend the value on the top of the stack from 32 bits to full length. This is necessary because @code{z} is a signed integer. @item mul Pop the top two numbers on the stack, multiply them, and push their product. Now the top of the stack contains the value of the expression @code{y * z}. @item add Pop the top two numbers, add them, and push the sum. Now the top of the stack contains the value of @code{x + y * z}. @item end Stop executing; the value left on the stack top is the value to be recorded. @end table @node Bytecode Descriptions @section Bytecode Descriptions Each bytecode description has the following form: @table @asis @item @code{add} (0x02): @var{a} @var{b} @result{} @var{a+b} Pop the top two stack items, @var{a} and @var{b}, as integers; push their sum, as an integer. @end table In this example, @code{add} is the name of the bytecode, and @code{(0x02)} is the one-byte value used to encode the bytecode, in hexidecimal. The phrase ``@var{a} @var{b} @result{} @var{a+b}'' shows the stack before and after the bytecode executes. Beforehand, the stack must contain at least two values, @var{a} and @var{b}; since the top of the stack is to the right, @var{b} is on the top of the stack, and @var{a} is underneath it. After execution, the bytecode will have popped @var{a} and @var{b} from the stack, and replaced them with a single value, @var{a+b}. There may be other values on the stack below those shown, but the bytecode affects only those shown. Here is another example: @table @asis @item @code{const8} (0x22) @var{n}: @result{} @var{n} Push the 8-bit integer constant @var{n} on the stack, without sign extension. @end table In this example, the bytecode @code{const8} takes an operand @var{n} directly from the bytecode stream; the operand follows the @code{const8} bytecode itself. We write any such operands immediately after the name of the bytecode, before the colon, and describe the exact encoding of the operand in the bytecode stream in the body of the bytecode description. For the @code{const8} bytecode, there are no stack items given before the @result{}; this simply means that the bytecode consumes no values from the stack. If a bytecode consumes no values, or produces no values, the list on either side of the @result{} may be empty. If a value is written as @var{a}, @var{b}, or @var{n}, then the bytecode treats it as an integer. If a value is written is @var{addr}, then the bytecode treats it as an address. We do not fully describe the floating point operations here; although this design can be extended in a clean way to handle floating point values, they are not of immediate interest to the customer, so we avoid describing them, to save time. @table @asis @item @code{float} (0x01): @result{} Prefix for floating-point bytecodes. Not implemented yet. @item @code{add} (0x02): @var{a} @var{b} @result{} @var{a+b} Pop two integers from the stack, and push their sum, as an integer. @item @code{sub} (0x03): @var{a} @var{b} @result{} @var{a-b} Pop two integers from the stack, subtract the top value from the next-to-top value, and push the difference. @item @code{mul} (0x04): @var{a} @var{b} @result{} @var{a*b} Pop two integers from the stack, multiply them, and push the product on the stack. Note that, when one multiplies two @var{n}-bit numbers yielding another @var{n}-bit number, it is irrelevant whether the numbers are signed or not; the results are the same. @item @code{div_signed} (0x05): @var{a} @var{b} @result{} @var{a/b} Pop two signed integers from the stack; divide the next-to-top value by the top value, and push the quotient. If the divisor is zero, terminate with an error. @item @code{div_unsigned} (0x06): @var{a} @var{b} @result{} @var{a/b} Pop two unsigned integers from the stack; divide the next-to-top value by the top value, and push the quotient. If the divisor is zero, terminate with an error. @item @code{rem_signed} (0x07): @var{a} @var{b} @result{} @var{a modulo b} Pop two signed integers from the stack; divide the next-to-top value by the top value, and push the remainder. If the divisor is zero, terminate with an error. @item @code{rem_unsigned} (0x08): @var{a} @var{b} @result{} @var{a modulo b} Pop two unsigned integers from the stack; divide the next-to-top value by the top value, and push the remainder. If the divisor is zero, terminate with an error. @item @code{lsh} (0x09): @var{a} @var{b} @result{} @var{a<>b} Pop two integers from the stack; let @var{a} be the next-to-top value, and @var{b} be the top value. Shift @var{a} right by @var{b} bits, inserting copies of the top bit at the high end, and push the result. @item @code{rsh_unsigned} (0x0b): @var{a} @var{b} @result{} @var{a>>b} Pop two integers from the stack; let @var{a} be the next-to-top value, and @var{b} be the top value. Shift @var{a} right by @var{b} bits, inserting zero bits at the high end, and push the result. @item @code{log_not} (0x0e): @var{a} @result{} @var{!a} Pop an integer from the stack; if it is zero, push the value one; otherwise, push the value zero. @item @code{bit_and} (0x0f): @var{a} @var{b} @result{} @var{a&b} Pop two integers from the stack, and push their bitwise @code{and}. @item @code{bit_or} (0x10): @var{a} @var{b} @result{} @var{a|b} Pop two integers from the stack, and push their bitwise @code{or}. @item @code{bit_xor} (0x11): @var{a} @var{b} @result{} @var{a^b} Pop two integers from the stack, and push their bitwise exclusive-@code{or}. @item @code{bit_not} (0x12): @var{a} @result{} @var{~a} Pop an integer from the stack, and push its bitwise complement. @item @code{equal} (0x13): @var{a} @var{b} @result{} @var{a=b} Pop two integers from the stack; if they are equal, push the value one; otherwise, push the value zero. @item @code{less_signed} (0x14): @var{a} @var{b} @result{} @var{a @var{a} @var{a} Push another copy of the stack's top element. @item @code{swap} (0x2b): @var{a} @var{b} => @var{b} @var{a} Exchange the top two items on the stack. @item @code{pop} (0x29): @var{a} => Discard the top value on the stack. @item @code{if_goto} (0x20) @var{offset}: @var{a} @result{} Pop an integer off the stack; if it is non-zero, branch to the given offset in the bytecode string. Otherwise, continue to the next instruction in the bytecode stream. In other words, if @var{a} is non-zero, set the @code{pc} register to @code{start} + @var{offset}. Thus, an offset of zero denotes the beginning of the expression. The @var{offset} is stored as a sixteen-bit unsigned value, stored immediately following the @code{if_goto} bytecode. It is always stored most significant byte first, regardless of the target's normal endianness. The offset is not guaranteed to fall at any particular alignment within the bytecode stream; thus, on machines where fetching a 16-bit on an unaligned address raises an exception, you should fetch the offset one byte at a time. @item @code{goto} (0x21) @var{offset}: @result{} Branch unconditionally to @var{offset}; in other words, set the @code{pc} register to @code{start} + @var{offset}. The offset is stored in the same way as for the @code{if_goto} bytecode. @item @code{const8} (0x22) @var{n}: @result{} @var{n} @itemx @code{const16} (0x23) @var{n}: @result{} @var{n} @itemx @code{const32} (0x24) @var{n}: @result{} @var{n} @itemx @code{const64} (0x25) @var{n}: @result{} @var{n} Push the integer constant @var{n} on the stack, without sign extension. To produce a small negative value, push a small twos-complement value, and then sign-extend it using the @code{ext} bytecode. The constant @var{n} is stored in the appropriate number of bytes following the @code{const}@var{b} bytecode. The constant @var{n} is always stored most significant byte first, regardless of the target's normal endianness. The constant is not guaranteed to fall at any particular alignment within the bytecode stream; thus, on machines where fetching a 16-bit on an unaligned address raises an exception, you should fetch @var{n} one byte at a time. @item @code{reg} (0x26) @var{n}: @result{} @var{a} Push the value of register number @var{n}, without sign extension. The registers are numbered following GDB's conventions. The register number @var{n} is encoded as a 16-bit unsigned integer immediately following the @code{reg} bytecode. It is always stored most significant byte first, regardless of the target's normal endianness. The register number is not guaranteed to fall at any particular alignment within the bytecode stream; thus, on machines where fetching a 16-bit on an unaligned address raises an exception, you should fetch the register number one byte at a time. @item @code{trace} (0x0c): @var{addr} @var{size} @result{} Record the contents of the @var{size} bytes at @var{addr} in a trace buffer, for later retrieval by GDB. @item @code{trace_quick} (0x0d) @var{size}: @var{addr} @result{} @var{addr} Record the contents of the @var{size} bytes at @var{addr} in a trace buffer, for later retrieval by GDB. @var{size} is a single byte unsigned integer following the @code{trace} opcode. This bytecode is equivalent to the sequence @code{dup const8 @var{size} trace}, but we provide it anyway to save space in bytecode strings. @item @code{trace16} (0x30) @var{size}: @var{addr} @result{} @var{addr} Identical to trace_quick, except that @var{size} is a 16-bit big-endian unsigned integer, not a single byte. This should probably have been named @code{trace_quick16}, for consistency. @item @code{end} (0x27): @result{} Stop executing bytecode; the result should be the top element of the stack. If the purpose of the expression was to compute an lvalue or a range of memory, then the next-to-top of the stack is the lvalue's address, and the top of the stack is the lvalue's size, in bytes. @end table @node Using Agent Expressions @section Using Agent Expressions Here is a sketch of a full non-stop debugging cycle, showing how agent expressions fit into the process. @itemize @bullet @item The user selects trace points in the program's code at which GDB should collect data. @item The user specifies expressions to evaluate at each trace point. These expressions may denote objects in memory, in which case those objects' contents are recorded as the program runs, or computed values, in which case the values themselves are recorded. @item GDB transmits the tracepoints and their associated expressions to the GDB agent, running on the debugging target. @item The agent arranges to be notified when a trace point is hit. Note that, on some systems, the target operating system is completely responsible for collecting the data; see @ref{Tracing on Symmetrix}. @item When execution on the target reaches a trace point, the agent evaluates the expressions associated with that trace point, and records the resulting values and memory ranges. @item Later, when the user selects a given trace event and inspects the objects and expression values recorded, GDB talks to the agent to retrieve recorded data as necessary to meet the user's requests. If the user asks to see an object whose contents have not been recorded, GDB reports an error. @end itemize @node Varying Target Capabilities @section Varying Target Capabilities Some targets don't support floating-point, and some would rather not have to deal with @code{long long} operations. Also, different targets will have different stack sizes, and different bytecode buffer lengths. Thus, GDB needs a way to ask the target about itself. We haven't worked out the details yet, but in general, GDB should be able to send the target a packet asking it to describe itself. The reply should be a packet whose length is explicit, so we can add new information to the packet in future revisions of the agent, without confusing old versions of GDB, and it should contain a version number. It should contain at least the following information: @itemize @bullet @item whether floating point is supported @item whether @code{long long} is supported @item maximum acceptable size of bytecode stack @item maximum acceptable length of bytecode expressions @item which registers are actually available for collection @item whether the target supports disabled tracepoints @end itemize @node Tracing on Symmetrix @section Tracing on Symmetrix This section documents the API used by the GDB agent to collect data on Symmetrix systems. Cygnus originally implemented these tracing features to help EMC Corporation debug their Symmetrix high-availability disk drives. The Symmetrix application code already includes substantial tracing facilities; the GDB agent for the Symmetrix system uses those facilities for its own data collection, via the API described here. @deftypefn Function DTC_RESPONSE adbg_find_memory_in_frame (FRAME_DEF *@var{frame}, char *@var{address}, char **@var{buffer}, unsigned int *@var{size}) Search the trace frame @var{frame} for memory saved from @var{address}. If the memory is available, provide the address of the buffer holding it; otherwise, provide the address of the next saved area. @itemize @bullet @item If the memory at @var{address} was saved in @var{frame}, set @code{*@var{buffer}} to point to the buffer in which that memory was saved, set @code{*@var{size}} to the number of bytes from @var{address} that are saved at @code{*@var{buffer}}, and return @code{OK_TARGET_RESPONSE}. (Clearly, in this case, the function will always set @code{*@var{size}} to a value greater than zero.) @item If @var{frame} does not record any memory at @var{address}, set @code{*@var{size}} to the distance from @var{address} to the start of the saved region with the lowest address higher than @var{address}. If there is no memory saved from any higher address, set @code{*@var{size}} to zero. Return @code{NOT_FOUND_TARGET_RESPONSE}. @end itemize These two possibilities allow the caller to either retrieve the data, or walk the address space to the next saved area. @end deftypefn This function allows the GDB agent to map the regions of memory saved in a particular frame, and retrieve their contents efficiently. This function also provides a clean interface between the GDB agent and the Symmetrix tracing structures, making it easier to adapt the GDB agent to future versions of the Symmetrix system, and vice versa. This function searches all data saved in @var{frame}, whether the data is there at the request of a bytecode expression, or because it falls in one of the format's memory ranges, or because it was saved from the top of the stack. EMC can arbitrarily change and enhance the tracing mechanism, but as long as this function works properly, all collected memory is visible to GDB. The function itself is straightforward to implement. A single pass over the trace frame's stack area, memory ranges, and expression blocks can yield the address of the buffer (if the requested address was saved), and also note the address of the next higher range of memory, to be returned when the search fails. As an example, suppose the trace frame @code{f} has saved sixteen bytes from address @code{0x8000} in a buffer at @code{0x1000}, and thirty-two bytes from address @code{0xc000} in a buffer at @code{0x1010}. Here are some sample calls, and the effect each would have: @table @code @item adbg_find_memory_in_frame (f, (char*) 0x8000, &buffer, &size) This would set @code{buffer} to @code{0x1000}, set @code{size} to sixteen, and return @code{OK_TARGET_RESPONSE}, since @code{f} saves sixteen bytes from @code{0x8000} at @code{0x1000}. @item adbg_find_memory_in_frame (f, (char *) 0x8004, &buffer, &size) This would set @code{buffer} to @code{0x1004}, set @code{size} to twelve, and return @code{OK_TARGET_RESPONSE}, since @file{f} saves the twelve bytes from @code{0x8004} starting four bytes into the buffer at @code{0x1000}. This shows that request addresses may fall in the middle of saved areas; the function should return the address and size of the remainder of the buffer. @item adbg_find_memory_in_frame (f, (char *) 0x8100, &buffer, &size) This would set @code{size} to @code{0x3f00} and return @code{NOT_FOUND_TARGET_RESPONSE}, since there is no memory saved in @code{f} from the address @code{0x8100}, and the next memory available is at @code{0x8100 + 0x3f00}, or @code{0xc000}. This shows that request addresses may fall outside of all saved memory ranges; the function should indicate the next saved area, if any. @item adbg_find_memory_in_frame (f, (char *) 0x7000, &buffer, &size) This would set @code{size} to @code{0x1000} and return @code{NOT_FOUND_TARGET_RESPONSE}, since the next saved memory is at @code{0x7000 + 0x1000}, or @code{0x8000}. @item adbg_find_memory_in_frame (f, (char *) 0xf000, &buffer, &size) This would set @code{size} to zero, and return @code{NOT_FOUND_TARGET_RESPONSE}. This shows how the function tells the caller that no further memory ranges have been saved. @end table As another example, here is a function which will print out the addresses of all memory saved in the trace frame @code{frame} on the Symmetrix INLINES console: @example void print_frame_addresses (FRAME_DEF *frame) @{ char *addr; char *buffer; unsigned long size; addr = 0; for (;;) @{ /* Either find out how much memory we have here, or discover where the next saved region is. */ if (adbg_find_memory_in_frame (frame, addr, &buffer, &size) == OK_TARGET_RESPONSE) printp ("saved %x to %x\n", addr, addr + size); if (size == 0) break; addr += size; @} @} @end example Note that there is not necessarily any connection between the order in which the data is saved in the trace frame, and the order in which @code{adbg_find_memory_in_frame} will return those memory ranges. The code above will always print the saved memory regions in order of increasing address, while the underlying frame structure might store the data in a random order. [[This section should cover the rest of the Symmetrix functions the stub relies upon, too.]] @node Rationale @section Rationale Some of the design decisions apparent above are arguable. @table @b @item What about stack overflow/underflow? GDB should be able to query the target to discover its stack size. Given that information, GDB can determine at translation time whether a given expression will overflow the stack. But this spec isn't about what kinds of error-checking GDB ought to do. @item Why are you doing everything in LONGEST? Speed isn't important, but agent code size is; using LONGEST brings in a bunch of support code to do things like division, etc. So this is a serious concern. First, note that you don't need different bytecodes for different operand sizes. You can generate code without @emph{knowing} how big the stack elements actually are on the target. If the target only supports 32-bit ints, and you don't send any 64-bit bytecodes, everything just works. The observation here is that the MIPS and the Alpha have only fixed-size registers, and you can still get C's semantics even though most instructions only operate on full-sized words. You just need to make sure everything is properly sign-extended at the right times. So there is no need for 32- and 64-bit variants of the bytecodes. Just implement everything using the largest size you support. GDB should certainly check to see what sizes the target supports, so the user can get an error earlier, rather than later. But this information is not necessary for correctness. @item Why don't you have @code{>} or @code{<=} operators? I want to keep the interpreter small, and we don't need them. We can combine the @code{less_} opcodes with @code{log_not}, and swap the order of the operands, yielding all four asymmetrical comparison operators. For example, @code{(x <= y)} is @code{! (x > y)}, which is @code{! (y < x)}. @item Why do you have @code{log_not}? @itemx Why do you have @code{ext}? @itemx Why do you have @code{zero_ext}? These are all easily synthesized from other instructions, but I expect them to be used frequently, and they're simple, so I include them to keep bytecode strings short. @code{log_not} is equivalent to @code{const8 0 equal}; it's used in half the relational operators. @code{ext @var{n}} is equivalent to @code{const8 @var{s-n} lsh const8 @var{s-n} rsh_signed}, where @var{s} is the size of the stack elements; it follows @code{ref@var{m}} and @var{reg} bytecodes when the value should be signed. See the next bulleted item. @code{zero_ext @var{n}} is equivalent to @code{const@var{m} @var{mask} log_and}; it's used whenever we push the value of a register, because we can't assume the upper bits of the register aren't garbage. @item Why not have sign-extending variants of the @code{ref} operators? Because that would double the number of @code{ref} operators, and we need the @code{ext} bytecode anyway for accessing bitfields. @item Why not have constant-address variants of the @code{ref} operators? Because that would double the number of @code{ref} operators again, and @code{const32 @var{address} ref32} is only one byte longer. @item Why do the @code{ref@var{n}} operators have to support unaligned fetches? GDB will generate bytecode that fetches multi-byte values at unaligned addresses whenever the executable's debugging information tells it to. Furthermore, GDB does not know the value the pointer will have when GDB generates the bytecode, so it cannot determine whether a particular fetch will be aligned or not. In particular, structure bitfields may be several bytes long, but follow no alignment rules; members of packed structures are not necessarily aligned either. In general, there are many cases where unaligned references occur in correct C code, either at the programmer's explicit request, or at the compiler's discretion. Thus, it is simpler to make the GDB agent bytecodes work correctly in all circumstances than to make GDB guess in each case whether the compiler did the usual thing. @item Why are there no side-effecting operators? Because our current client doesn't want them? That's a cheap answer. I think the real answer is that I'm afraid of implementing function calls. We should re-visit this issue after the present contract is delivered. @item Why aren't the @code{goto} ops PC-relative? The interpreter has the base address around anyway for PC bounds checking, and it seemed simpler. @item Why is there only one offset size for the @code{goto} ops? Offsets are currently sixteen bits. I'm not happy with this situation either: Suppose we have multiple branch ops with different offset sizes. As I generate code left-to-right, all my jumps are forward jumps (there are no loops in expressions), so I never know the target when I emit the jump opcode. Thus, I have to either always assume the largest offset size, or do jump relaxation on the code after I generate it, which seems like a big waste of time. I can imagine a reasonable expression being longer than 256 bytes. I can't imagine one being longer than 64k. Thus, we need 16-bit offsets. This kind of reasoning is so bogus, but relaxation is pathetic. The other approach would be to generate code right-to-left. Then I'd always know my offset size. That might be fun. @item Where is the function call bytecode? When we add side-effects, we should add this. @item Why does the @code{reg} bytecode take a 16-bit register number? Intel's IA-64 architecture has 128 general-purpose registers, and 128 floating-point registers, and I'm sure it has some random control registers. @item Why do we need @code{trace} and @code{trace_quick}? Because GDB needs to record all the memory contents and registers an expression touches. If the user wants to evaluate an expression @code{x->y->z}, the agent must record the values of @code{x} and @code{x->y} as well as the value of @code{x->y->z}. @item Don't the @code{trace} bytecodes make the interpreter less general? They do mean that the interpreter contains special-purpose code, but that doesn't mean the interpreter can only be used for that purpose. If an expression doesn't use the @code{trace} bytecodes, they don't get in its way. @item Why doesn't @code{trace_quick} consume its arguments the way everything else does? In general, you do want your operators to consume their arguments; it's consistent, and generally reduces the amount of stack rearrangement necessary. However, @code{trace_quick} is a kludge to save space; it only exists so we needn't write @code{dup const8 @var{SIZE} trace} before every memory reference. Therefore, it's okay for it not to consume its arguments; it's meant for a specific context in which we know exactly what it should do with the stack. If we're going to have a kludge, it should be an effective kludge. @item Why does @code{trace16} exist? That opcode was added by the customer that contracted Cygnus for the data tracing work. I personally think it is unnecessary; objects that large will be quite rare, so it is okay to use @code{dup const16 @var{size} trace} in those cases. Whatever we decide to do with @code{trace16}, we should at least leave opcode 0x30 reserved, to remain compatible with the customer who added it. @end table @bye