Has anybody come across a list of ARM & THUMB instructions that cause deviation from the linear instruction stream?
I've been trying to figure out gdb-stub single stepping using software interrupts, and in single stepping you need to find
the next instruction(s) where the next breakpoint instruction needs to be set.
There are three cases:
1) current instruction doesn't change the execution path. Next instruction is the next word.
2) current instruction is a jump. The operand defines the next instruction address
3) current instruction is conditional branch. One possible next instruction is the next word, the other possible
instruction address is defined by the operand. (That includes conditional add with PC as the target, and the like).
To implement single stepping, I need to tell those cases apart and figure out how to find out the possible branching address.
I could go through manuals of numerous processors instruction by instruction and maybe I'd be done within the next couple of years,
or I could find a list of instructions to check, or a paper that explains how to "decode" the instructions in a useful way.
Also, there doesn't seem to be lots of sources of ARM gdb servers or stubs around that use software breakpoints.
I'm happy to hear that the table method will work.
Normally, instruction groups are very logical.
Yes, you could say that the LDR/STR instruction is just a 'transfer instruction between memory and registers', where there's a flag indicating the direction.
Well, the table method doesn't work, I saw that when I got the spreadsheet about the instructions done.
When you look at the sheet, you can see that with all the instructions only bits 27, 26 and 25 are significant (= not belonging to an operand or modifier of some instruction). When you take a group of instructions that have bits 27, 26 and 25 all zeros, the only significant bit is bit 4. With instruction groups that have bits 27 - 25 and 4 zero, the new significant bits are 24, 23 and 21. If bit 4 is one, the only new significant bit is bit 7 and so on.
Makes one h*** of a decoding logic.
The spreadsheet is pretty nice tool for something like this: you can sort the lines (instructions) by different combinations of columns to find out which decoding steps work.
I got the arithmetic & logic instructions nicely organized by selecting the lines and organizing them as bit 25 as primary, bit 4 as secondary and bit 21 as tertiary sorting "key". The instructions become sorted by addressing modes.
First all register operation, then register shifted register and last the immediates. You can't apply all the keys at the same time, unfortunately, but one by one.
This is what it looks like:
Sorting with bits 24 - 21 in that order gives:
I still think it's possible to use the table, especially, if you're only decoding 32-bit instructions.
Let's make a few rules:
1: If a bit is known to be either 1 or 0, set the corresponding bit in mask to 1, otherwise set it to 0.
2: If a bit is known, set keep the bit in data, otherwise set the corresponding bit in data to 0.
That is ...
Instr: cccc0000010Snnnnddddssss0TT1mmmm
mask : 00001111111000000000000010010000
data : 00000000010000000000000000010000
Using this rule, we can convert the above table using the attached perl-script.
We now get the following (slightly decorated):
static const InstrTab sInstructionTable[] = {
0x0fe00010, 0x00000000, "and", &and_shift, /* AND{S}<c> <Rd>,<Rn>,<Rm>{,<shift>} A1 A8.8.14 */
0x0fe00090, 0x00000010, "and", &and_reg, /* AND{S}<c> <Rd>,<Rn>,<Rm>,<type>,<Rs> A1 A8.8.15 */
0x0fe00000, 0x02000000, "and", &and_imm, /* AND{S}<c> <Rd>,<Rn>,#<const> A1 A8.8.13 */
0x0fe00010, 0x00200000, "eor", &eor_shift, /* EOR{S}<c> <Rd>,<Rn>,<Rm>{,<shift>} A1 A8.8.47 */
0x0fe00090, 0x00200010, "eor", &eor_reg, /* EOR{S}<c> <Rd>,<Rn>,<Rm>,<type>,<Rs> A1 A8.8.48 */
0x0fe00000, 0x02200000, "eor", &eor_imm, /* EOR{S}<c> <Rd>,<Rn>,#<const> A1 A8.8.46 */
0x0fe00010, 0x00400000, "sub", &sub_shift, /* SUB{S}<c> <Rd>,<Rn>,<Rm>{,<shift>} A1 A8.8.223 */
0x0fe00090, 0x00400010, "sub", &sub_reg, /* SUB{S}<c> <Rd>,<Rn>,<Rm>,<type>,<Rs> A1 A8.8.224 */
0x0fe00000, 0x02400000, "sub", &sub_imm, /* SUB{S}<c> <Rd>,<Rn>,#<const> A1 A8.8.222 */
0x0fe00010, 0x00600000, "rsb", &rsb_shift, /* RSB{S}<c> <Rd>,<Rn>,<Rm>{,<shift>} A1 A8.8.153 */
0x00000000, 0x00000000, "", &done, /* end of list */
};
Now, the problem with the table above, is that the order is incorrect. This can be fixed easily:
(First two AND were swapped, first two EOR were swapped and first two SUB were swapped).
-Of course, the Perl-script could be enhanced to do this automatically, generating extra masks and data in those cases where that is needed.
So the following should work:
InstrTab *tab; uint32_t instr; instr = *pc; tab = &sInstructionTable; while((instr & tab->mask) != tab->data) { tab++; } tab->handle(instr,tab);
If writing the code in assembly language, I believe it would be a good idea to re-arrange the contents of the table, so that the instruction name comes first, then the mask, the data and finally the handler.
find_instr:
ldmia r4!,{r1-r3,r12}
and r2,r2,r0
cmp r3,r2
bne find_instr
/* here r0 contains the instruction opcode, r1 contains the name and r12 contains the address of the handler */
pop { ... } /* restore any saved registers */
bx r12 /* jump directly to the handler */
Hmm, I guess I've been working too hard with the project - I didn't come to think adding the handler in the table.
(I don't like to use function pointers if not necessary - the execution path gets fuzzy and some debuggers can't handle them, but using an enum...) I had a different idea about what the table method meas, but I don't remember what I thought any more. What I remember is that I didn't think of comparing an instruction to all instructions of the instruction set.
If applied to the whole instruction set, it may be a bit heavy for single stepping - on the average it means going through half of the table for each instruction each time.
Then again writing the logic in C code takes a lot of time, and the code in not something you'd show to small children...
(I made the pseudocode whith which I'm still not completely happy, but the C-coding of it is not started yet.)
Maybe I should use the table approach now after all, and go for the C code logic later (if I still feel like it).
I really have to rethink this.
Especially now that I have the instruction set in a spreadsheet from which the instruction bit patterns are easily edited (bits into bytes, byte order, ...) and they can be printed out as a text file for simple editing (if any needed) into a table initializer data form. Also the mask and data can be easily generated with spreadsheet formuli.
The table is also quite compact form - the instruction set itself probably fits in 1 kB, and the actual handlers need to be written anyway.
Too bad there is only 'like'-button. There should also be 'Halleluyah'-button.
When I wrote my 68xxx debugger, the table was fine (though these were only 16-bit words).
-But ARM's instruction set is not too complicated either. I have not had a look at Cortex-A yet, but the time will come.
In case the table gets too large, you have an extra approach: To split the words into two 16-bit words, so you find the "main part", which leads to a "sub-tree".
Remember: The execution unit in the processor does the job very, very quickly, so I am convinced that ARM designed the instruction set, so it should be easy to dispatch (even by using code).
Yes, the hard part is to find out how.
The good thing about using pointers, is that you can make your lookup-routine in assembly language, and it can jump directly to your C routine.
You can then call it as a C-function, because it uses "goto-style", thus it'll be completely transparent and your C-code will behave like a normal subroutine-call; just very quickly.
The above look-up example can be unrolled easily; this will save a few clock cycles on each iteration:
.rept 7
ldmiane r4!,{r1-r3,r12}
andne r2,r2,r0
cmpne r3,r2
.endr
-Change the '.rept' count as you like... make it 15 or 31, adjust it to suit your needs. Perhaps a large number may start to cause longer execution time, but it's a question of balance.
If you're lucky, you can place instructions that are used often in the beginning of the table (I did that with my debugger, and it started to become quite quick at disassembling).
This kind of code is something I really like. The table-lookup, masks and AND stuff - it brings out good memories too.
-But of course, sometimes it might be easier or shorter or quicker to write a switch-statement and use enumerations for handling each instruction type.
Some instruction types could be handled by the same handler; eg. AND/ORR/EOR and ADD/SUB.
In many cases, it's useful to think of instructions as being in "instruction groups". Eg. LDR/STR is a good example, AND/ORR/EOR, ASL/ASR/LSL/LSR/ROR, etc.
If you're lucky, you can place instructions that are used often in the beginning of the table
You read my mind.
But ARM's instruction set is not too complicated either.
Assembly is not, but the encoding is.
and:
Oh, and for small assembly routines I've been using inline asm, like
void rpi2_trap_handler()
{
// IRQs need to be enabled for serial I/O
asm volatile (
"push {r0}\n\t"
"mrs r0, cpsr\n\t"
"bic r0, #128 @ enable irqs\n\t"
"msr cpsr, r0\n\t"
"pop {r0}\n\t"
);
gdb_trap_handler();
}
: