RTL frontend, insn recognition, and pointer equality

David Malcolm dmalcolm@redhat.com
Wed Oct 19 23:49:00 GMT 2016


I've updated the RTL frontend to the new "compact" dump format and have
it "mostly working", for some definition of that term [1].  I'm
focusing on the "__RTL" extension to cc1, rather than having a
true "rtl1" frontend.

I've run into an issue with insn recognition relating to pointer
equality of rtx.

I have a test case for "final", which contains this x86_64 code
(shown here in compact form):

      (cinsn (set (mem/v:BLK (scratch:DI) [0  A8])
                    (unspec:BLK [
                            (mem/v:BLK (scratch:DI) [0  A8])
                        ] UNSPEC_MEMORY_BLOCKAGE)) "test.c":2
                 (nil))

This fails inside "final" with:
gcc.dg/rtl/x86_64/final.c:130:1: error: unrecognizable insn:
(insn 5 4 6 2 (set (mem/v:BLK (scratch:DI) [0  A8])
        (unspec:BLK [
                (mem/v:BLK (scratch:DI) [0  A8])
            ] UNSPEC_MEMORY_BLOCKAGE)) -1
     (nil))

(the error message is showing the non-compact form of the
loaded insn).

The pertinent part of i386.md is this:

;; Do not schedule instructions accessing memory across this point.

(define_expand "memory_blockage"
  [(set (match_dup 0)
	(unspec:BLK [(match_dup 0)] UNSPEC_MEMORY_BLOCKAGE))]
  ""
{
  operands[0] = gen_rtx_MEM (BLKmode, gen_rtx_SCRATCH (Pmode));
  MEM_VOLATILE_P (operands[0]) = 1;
})

(define_insn "*memory_blockage"
  [(set (match_operand:BLK 0)
	(unspec:BLK [(match_dup 0)] UNSPEC_MEMORY_BLOCKAGE))]
  ""
  ""
  [(set_attr "length" "0")])


If I run that directly in memory, the insn *is* recognized:

(gdb) call gen_memory_blockage ()
$19 = (insn 33 0 0 (set (mem/v:BLK (scratch:DI) [0  A8])
        (unspec:BLK [
                (mem/v:BLK (scratch:DI) [0  A8])
            ] UNSPEC_MEMORY_BLOCKAGE)) -1
     (nil))

(gdb) call recog_memoized ((rtx_insn *)$19)
$20 = 695
(gdb) call debug ($19)
(insn 33 0 0 (set (mem/v:BLK (scratch:DI) [0  A8])
        (unspec:BLK [
                (mem/v:BLK (scratch:DI) [0  A8])
            ] UNSPEC_MEMORY_BLOCKAGE)) 695 {*memory_blockage}
     (nil))

By contrast, if I load the rtx from a dump file, the insn is
*not* recognized.

Stepping through insn-recog.c revealed the problem is that the
  (match_dup 0)
fails when run on a reconstructed-from-dump copy:

44388	    case 17:
44389	      if (GET_MODE (x2) != BLKmode)
44390	        return -1;
44391	      x3 = XVECEXP (x2, 0, 0);
44392	      if (!rtx_equal_p (x3, operands[0]))
44393	        return -1;
44394	      return 695; /* *memory_blockage */

(gdb) call debug (x3)
(mem/v:BLK (scratch:DI) [0  A8])

(gdb) call debug (operands[0])
(mem/v:BLK (scratch:DI) [0  A8])

(gdb) p x3
$40 = (rtx) 0x7ffff19ee150
(gdb) p operands[0]
$41 = (rtx) 0x7ffff19ee120

(gdb) call rtx_equal_p ($40, $41)
$42 = 0

The issue is that on loading from a file, we have two distinct
  (scratch:DI)
rtx.

rtl.c:rtx_equal_p has:

    case DEBUG_EXPR:
    case VALUE:
    case SCRATCH:
    CASE_CONST_UNIQUE:
      return 0;

i.e. SCRATCH values always compare as non-equal, unless they
have pointer equality.

It works when built directly from insn-emit.c within gen_memory_blockage:

74308	  emit_insn (gen_rtx_SET (operand0,
74309		gen_rtx_UNSPEC (BLKmode,
74310		gen_rtvec (1,
74311			copy_rtx (operand0)),
74312		17)));

since rtl.c:copy_rtx has:

    case SCRATCH:
      /* SCRATCH must be shared because they represent distinct values.  */
      return orig;

So we have a case where a gen_* builds a pattern that contains a
graph, i.e. where a SCRATCH appears in two places:

(insn 5 4 6 2 (set (mem/v:BLK (scratch:DI) [0  A8])
                               ^^^^^^^^^^
        (unspec:BLK [
                (mem/v:BLK (scratch:DI) [0  A8])
                            ^^^^^^^^^^
            ] UNSPEC_MEMORY_BLOCKAGE)) -1
     (nil))

and for which the recognizer will only recognize the insn
if that graph-like property is preserved: that we have pointer-equality
for the two (scratch:DI).

I'm wondering what the best course of action for the RTL frontend is.

Some ideas:

(a) dump the insn name even in compact mode, and do a string-match to
look it up when reading dumps.  This would isolate us somewhat from .md
file changes, but if we lose the memoized value, it can't be
re-recognized.  This feels like papering over the problem.

(b) for all codes for which rtx_equal_p requires pointer equality, add
some kind of extra ID to the dump, allowing the loader to reconstruct
the graph.  This could be the pointer itself:

  (scratch:DI [0x7ffff19ee150])
  (scratch:DI [ptr:0x7ffff19ee150])

or somesuch, but it would perhaps be better to use some kind of more
human-friendly ID e.g.

  (scratch:DI [ptr-idx: 42])

or similar, rather than subject ourselves to raw hex values.  Probably
need an extra flag to print-rtl.c, to avoid making the dumps more
verbose for everyone who doesn't want this (if so, does "compact" mode
need renaming...)

Or just do this for SCRATCH, maybe?

(c) some kind of special-casing?  (ugh)

(d) something I haven't thought of


I'm leaning towards (b).

Thoughts?
Dave

[1] FWIW, a work-in-progress patch kit is at:
  https://dmalcolm.fedorapeople.org/gcc/2016-10-19/rtl-frontend-v143-relative-to-r241136/
though it's not yet ready for review.



More information about the Gcc-patches mailing list