The instruction pattern:
and $-16, %esp
sub $imm, %esp
...
lea imm(%ebp), %esp
appears when the compiler is realigning the stack (for example in
main(), or almost everywhere with -mstackrealign switch). The "and"
instruction is very difficult to model, but that's not necessary, as
these frames are always %ebp-based (the compiler also needs a way to
restore the original %esp). Therefore the plans we were generating for
these function were almost correct already. The only place we were doing
it wrong were the last instructions of the epilogue (usually just
"ret"), where we had to revert to %esp-based unwinding, as the %ebp had
been popped already.
This was wrong because our "distance of esp from cfa" counter had picked
up the "sub" instruction (and incremented the counter) but it had not
seen that the register was reset by the "lea" instruction.
This patch fixes that shortcoming, and adds a test for handling
functions like this.
I have not been able to tickle the compiler into producing a 64-bit
function with this pattern, but I don't see a reason why it couldn't
produce it, if it chose to, so I add a x86_64 test as well.
Shouldn't you change the unwind information for the CFA here? For me saying CFA=rbp seems like an incorrect thing to do, but not sure what would be the correct value (Undefined? IsSame?). The impact is if an other register (or a local variable) have a location specified as CFA+off then after this instruction it will point to bogus location.