print-rtx.c: add 'h', v' and 'p' prefixes to regnos

On Tue, 2016-09-20 at 15:12 -0400, David Malcolm wrote:
> On Tue, 2016-09-20 at 09:20 -0600, Jeff Law wrote:
> > On 09/20/2016 08:34 AM, Bernd Schmidt wrote:
> > > On 09/20/2016 04:32 PM, David Malcolm wrote:
> > > >
> > > > To summarize so far: you want every pseudo to have its regno
> > > > dumped
> > > > with a 'p' prefix, and renumber them on dump (and then on
> > > > load).
> > > > OK.
> > >
> > > Renumbering is not helpful because it interferes with the view
> > > you
> > > have
> > > in the debugger. So, IMO just a prefix, and maybe
> > Yea, I guess it does since we want the numbers in the dump to be
> > the
> > same that we used to access the arrays.  So prefixing in the dump
> > with
> > adjustment of the number in the reader.
>
> To check I understand: am I right in thinking you want:
> (A) the numbers in the dump to be unmodified when dumping, so that we
> can easily look up values in arrays without confusion, and
> (B) regnums in the dump gain a 'p' prefix for values >=
>  FIRST_PSEUDO_REGISTER, so that humans and parsers can easily see
> when
> the regs are pseudos, and that
> (C) the parser will detect if a 'p'-prefixed regno actually has the
> same number as a hard reg (which can happen e.g. when a .md file
> changes, or when sharing .rtl dumps between targets), and remap the
> values on load accordingly
>
> ?
>
> (in which case we do need the regno_remapper class, or something like
> it)
>
> > >
> > > >   (reg/f:DI v1 virtual-stack-vars)
> > >
> > > this.
> > Doesn't the same apply to the number of virtual stack regs?  Those
> > are
> > in the same array as pseudos.  So ISTM we prefix in the dump, but
> > do
> > adjustment of the number in the reader?
>
> Presumably we could use "v" rather than "p" as the prefix for the
> first
> 5 pseudos (up to LAST_VIRTUAL_REGISTER), doing any adjustment at load
> time, rather than at dump time.  So the above example would look
> like:
>
>    (reg/f:DI v82 virtual-stack-vars)
>
> i.e. the 82 for x86_64's virtual-stack-vars would be prefixed with a
> 'v', and the loader would adjust it to be the current target's value
> for VIRTUAL_STACK_VARS_REGNUM.
>
> Do you like the idea of prefixing regnums of hardregs with 'h'? (so
> that all regnos get a one-char prefix) e.g.
>   (reg/i:SI h0 ax)
>   (reg/i:SF h21 xmm0)

(Replying to myself, in the hope of better demonstrating the idea)

The following patch implements this idea for RTL dumps, so that all REGNO
values in dumps get a one character prefix: 'h' for hard registers, 'v'
for virtual registers, and 'p' for non-virtual pseudos (making it easier
for both humans and parsers to grok the meaning of a REGNO).

For example, on compiling this:

  int test (int i)
  {
    return i * i;
  }

for target==x86_64, the dump for expand looks like this:

(note 1 0 4 NOTE_INSN_DELETED)
(note 4 1 2 2 [bb 2] NOTE_INSN_BASIC_BLOCK)
(insn 2 4 3 2 (set (mem/c:SI (plus:DI (reg/f:DI v82 virtual-stack-vars)
                (const_int -4 [0xfffffffffffffffc])) [1 i+0 S4 A32])
        (reg:SI h5 di [ i ])) ../../test.c:2 -1
     (nil))
(note 3 2 6 2 NOTE_INSN_FUNCTION_BEG)
(insn 6 3 7 2 (set (reg:SI p89)
        (mem/c:SI (plus:DI (reg/f:DI v82 virtual-stack-vars)
                (const_int -4 [0xfffffffffffffffc])) [1 i+0 S4 A32])) ../../test.c:3 -1
     (nil))
(insn 7 6 10 2 (parallel [
            (set (reg:SI p87 [ _2 ])
                (mult:SI (reg:SI p89)
                    (mem/c:SI (plus:DI (reg/f:DI v82 virtual-stack-vars)
                            (const_int -4 [0xfffffffffffffffc])) [1 i+0 S4 A32])))
            (clobber (reg:CC h17 flags))
        ]) ../../test.c:3 -1
     (nil))
(insn 10 7 14 2 (set (reg:SI p88 [ <retval> ])
        (reg:SI p87 [ _2 ])) ../../test.c:3 -1
     (nil))
(insn 14 10 15 2 (set (reg/i:SI h0 ax)
        (reg:SI p88 [ <retval> ])) ../../test.c:4 -1
     (nil))
(insn 15 14 0 2 (use (reg/i:SI h0 ax)) ../../test.c:4 -1
     (nil))

Note the 'v' prefix here:
  (reg/f:DI v82 virtual-stack-vars)
the 'h' prefix here:
  (reg/i:SI h0 ax)
and the 'p' prefix here:
  (reg:SI p88 [ <retval> ])

Successfully bootstrapped on x86_64-pc-linux-gnu.
There are various regression test failures involving scan-rtl-dump
due to regexes not matching e.g.
 PASS -> FAIL : gcc.target/i386/pr20020-1.c scan-rtl-dump expand "\\(set \\(reg/i:TI 0 ax\\)"
 PASS -> FAIL : gcc.target/i386/pr20020-1.c scan-rtl-dump expand "\\(set \\(reg:TI [0-9]* \\[ <retval> \\]\\)"
If the approach is OK, I can do an updated patch that also fixes up the
relevant tests (adding the appropriate prefixes); this would touch
multiple targets.

gcc/ChangeLog:
	* print-rtl.c (print_rtx): When dumping regnos, prefix the
	number with 'h', 'v' or 'p', depending on whether it is a
	hard register, a virtual register, or a non-virtual pseudo.
	* doc/rtl.texi (Registers and Memory): Document the above change,
	adding the prefix to examples that specifically refer to hard or
	pseudo regs in the text.
---
 gcc/doc/rtl.texi | 26 +++++++++++++++++++-------
 gcc/print-rtl.c  | 18 +++++++++---------
 2 files changed, 28 insertions(+), 16 deletions(-)

print-rtx.c: add 'h', v' and 'p' prefixes to regnos

Commit Message

Comments

Patch