F.5 Rationale

Some of the design decisions apparent above are arguable.

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.

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 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.

Why don't you have > or <= operators?

I want to keep the interpreter small, and we don't need them. We can combine the less_ opcodes with log_not, and swap the order of the operands, yielding all four asymmetrical comparison operators. For example, (x <= y) is ! (x > y), which is ! (y < x).

Why do you have log_not?

Why do you have ext?

Why do you have 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.

log_not is equivalent to const8 0 equal; it's used in half the relational operators.

ext n is equivalent to const8 s-n lsh const8 s-n rsh_signed, where s is the size of the stack elements; it follows refm and reg bytecodes when the value should be signed. See the next bulleted item.

zero_ext n is equivalent to constm 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.

Why not have sign-extending variants of the ref operators?

Because that would double the number of ref operators, and we need the ext bytecode anyway for accessing bitfields.

Why not have constant-address variants of the ref operators?

Because that would double the number of ref operators again, and const32 address ref32 is only one byte longer.

Why do the refn 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.

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.

Why aren't the goto ops PC-relative?

The interpreter has the base address around anyway for PC bounds checking, and it seemed simpler.

Why is there only one offset size for the 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.

Where is the function call bytecode?

When we add side-effects, we should add this.

Why does the 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.

Why do we need trace and trace_quick?

Because GDB needs to record all the memory contents and registers an expression touches. If the user wants to evaluate an expression x->y->z, the agent must record the values of x and x->y as well as the value of x->y->z.

Don't the 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 trace bytecodes, they don't get in its way.

Why doesn't 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, trace_quick is a kludge to save space; it only exists so we needn't write dup const8 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.

Why does 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 dup const16 size trace in those cases.

Whatever we decide to do with trace16, we should at least leave opcode 0x30 reserved, to remain compatible with the customer who added it.