Re: [PATCH 2/9] binutils: Be more forgiving of targets with large numbers of registers

["Palmer Dabbelt via binutils" <binutils at sourceware dot org>]

On Fri, 22 Nov 2019 04:10:26 PST (-0800), andrew.burgess@embecosm.com wrote:
> Currently if a target has a large ( > 1024 ) number of registers then
> we get a warning when dumping the DWARF whenever a register over the
> 1024 limit is referenced, this occurs in dwarf.c:frame_need_space.
>
> This check was initially introduced to guard against corrupted DWARF
> referencing stupidly large numbers of registers.
>
> The frame_need_space function already has a check in place so that, if
> a target specifies a set of known DWARF register names then we must
> only reference a register within this set, it is only after this check
> that we check for the 1024 limit.
>
> What this means is that if a target DOES NOT define a set of known
> register names and if we reference more than 1024 registers
> frame_need_space will give a warning.
>
> If a target DOES define a set of known registers and there are more
> than 1024 defined registers, and we try to reference a register beyond
> 1024 we will again get an error.
>
> This second case feels wrong to me.  My thinking is that if a target
> defines a set of registers then it is not unreasonable to assume the
> tools can cope with that number of registers.  And so, if the target
> defines 2000 named DWARF registers, frame_need_space should allow
> access to all of these registers.
>
> If a target does not define a set of named registers then the 1024
> limit should remain.  This is pretty arbitrary, but we do need to have
> some limit in place I think, so for now that seems as good as any.
>
> This is an entirely theoretical fix - there are no targets that define
> such large numbers of registers, but while experimenting with adding
> support for RISC-V CSRs I ran into this issue and felt like it was a
> good improvement.
>
> binutils/ChangeLog:
>
> 	* dwarf.c (frame_need_space): Compare dwarf_regnames_count against
> 	0, and only warn about large numbers of registers if the number is
> 	more than the dwarf_regnames_count.
>
> Change-Id: Ifac1a999ff0677676e81ee373c4c044b6a700827
> ---
>  binutils/ChangeLog | 6 ++++++
>  binutils/dwarf.c   | 4 ++--
>  2 files changed, 8 insertions(+), 2 deletions(-)
>
> diff --git a/binutils/dwarf.c b/binutils/dwarf.c
> index 2fe469f0603..62f2817d183 100644
> --- a/binutils/dwarf.c
> +++ b/binutils/dwarf.c
> @@ -7378,7 +7378,7 @@ frame_need_space (Frame_Chunk *fc, unsigned int reg)
>    if (reg < (unsigned int) fc->ncols)
>      return 0;
>
> -  if (dwarf_regnames_count
> +  if (dwarf_regnames_count > 0
>        && reg > dwarf_regnames_count)
>      return -1;

If I understand correctly, this doesn't actually change any behavior because
dwarf_regnames_count is never negative.  I'm fine with either way, though.

> @@ -7389,7 +7389,7 @@ frame_need_space (Frame_Chunk *fc, unsigned int reg)
>      return -1;
>
>    /* PR 17512: file: 2844a11d.  */
> -  if (fc->ncols > 1024)
> +  if (fc->ncols > 1024 && dwarf_regnames_count == 0)
>      {
>        error (_("Unfeasibly large register number: %u\n"), reg);
>        fc->ncols = 0;

Reviewed-by: Palmer Dabbelt <palmerdabbelt@google.com>