| Summary: | Negative strides in dwarf_aggregate_size | ||
|---|---|---|---|
| Product: | elfutils | Reporter: | Tim Haines <thaines.astro> |
| Component: | libdw | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | UNCONFIRMED --- | ||
| Severity: | enhancement | CC: | elfutils-devel, mark |
| Priority: | P2 | ||
| Version: | unspecified | ||
| Target Milestone: | --- | ||
| Host: | Target: | ||
| Build: | Last reconfirmed: | ||
I have to admit that I don't understand what it means for a stride to be negative. It looks to me that a stride is like a size, it has to be positive. It would be good to have an example when the stride can be negative. I asked around and John DelSignore came up with the following fortran example:
= array.f95 =
program f_prog
integer*4, allocatable, target, dimension (:,:) :: big_array
integer, dimension (:,:), pointer :: neg_array
allocate (big_array(2000,1000))
neg_array => big_array(2000:1:-1,1000:1:-1)
end program f_prog
So, neg_array is a pointer to an array section where the elements of big_array are reversed. That is, neg_array(x,y) references big_array(2000-x+1,1000-y+1).
In this example, the stride makes us go "backwards" into the array.
But note that gfortran -g array.f95 produces DW_FORM_exprlocs for the DW_AT_byte_stride because almost all attributes of the array are dynamic at runtime:
[ c9] array_type abbrev: 10
ordering (data1) col_major (1)
data_location (exprloc)
[ 0] push_object_address
[ 1] deref
allocated (exprloc)
[ 0] push_object_address
[ 1] deref
[ 2] lit0
[ 3] ne
type (ref4) [ 72]
sibling (ref4) [ 106]
[ db] subrange_type abbrev: 1
lower_bound (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 48
[ 3] deref
upper_bound (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 56
[ 3] deref
byte_stride (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 40
[ 3] deref
[ 4] push_object_address
[ 5] plus_uconst 32
[ 7] deref
[ 8] mul
[ f0] subrange_type abbrev: 1
lower_bound (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 72
[ 3] deref
upper_bound (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 80
[ 3] deref
byte_stride (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 64
[ 3] deref
[ 4] push_object_address
[ 5] plus_uconst 32
[ 7] deref
[ 8] mul
[ 106] array_type abbrev: 11
ordering (data1) col_major (1)
data_location (exprloc)
[ 0] push_object_address
[ 1] deref
associated (exprloc)
[ 0] push_object_address
[ 1] deref
[ 2] lit0
[ 3] ne
type (ref4) [ 72]
[ 114] subrange_type abbrev: 1
lower_bound (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 48
[ 3] deref
upper_bound (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 56
[ 3] deref
byte_stride (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 40
[ 3] deref
[ 4] push_object_address
[ 5] plus_uconst 32
[ 7] deref
[ 8] mul
[ 129] subrange_type abbrev: 1
lower_bound (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 72
[ 3] deref
upper_bound (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 80
[ 3] deref
byte_stride (exprloc)
[ 0] push_object_address
[ 1] plus_uconst 64
[ 3] deref
[ 4] push_object_address
[ 5] plus_uconst 32
[ 7] deref
[ 8] mul
Something dwarf_aggregate_size doesn't handle. And even if it could handle the expression, it doesn't know the object address or how to deref memory...
I was actually trying to make a Fortran example, but couldn't get one to work! Glad John was able to make one. |
In section 5.13 of DWARF5, the standard says Note that the stride can be negative I don't see anywhere in the standard besides this footnote that indicates it should be a signed value. In the array_size helper, the stride is calculated as Dwarf_Word stride = eltsize; if (INTUSE(dwarf_attr_integrate)(die, DW_AT_byte_stride, attr_mem) != NULL) { if (INTUSE(dwarf_formudata)(attr_mem, &stride) != 0) return -1; } else if (INTUSE(dwarf_attr_integrate)(die, DW_AT_bit_stride, attr_mem) != NULL) { if (INTUSE(dwarf_formudata)(attr_mem, &stride) != 0) return -1; if (stride % 8) /* XXX maybe compute in bits? */ return -1; stride /= 8; } Since we could be reading a negative integer, the use of dwarf_formudata would cause the total size to be incorrect. I don't have a reproducer, and I've not seen any binaries that have run across this. I just noticed it while reading through the source.