Re: [AARCH64, PATCH] Fix ICE in aarch64_float_const_representable_p

[Tom de Vries]

On 30-05-14 10:20, Ramana Radhakrishnan wrote:

On Fri, May 30, 2014 at 9:14 AM, Tom de Vries <tom_devr...@mentor.com> wrote:

Marcus,

when building for aarch64-linux-gnu with --enable-checking=yes,rtl, I run
into the following error:
...
In file included from src/libgcc/libgcc2.c:56:0:
src/libgcc/libgcc2.c: In function ‘__floattisf’:
src/libgcc/libgcc2.h:200:20: internal compiler error: RTL check: expected
code 'const_double' and not mode 'VOID', have code 'const_double' and mode
'VOID' in aarch64_float_const_representable_p, at
config/aarch64/aarch64.c:8481
  #define __NDW(a,b) __ ## a ## ti ## b
                     ^
src/libgcc/libgcc2.h:293:21: note: in expansion of macro ‘__NDW’
  #define __floatdisf __NDW(float,sf)
                      ^
src/libgcc/libgcc2.c:1569:14: note: in expansion of macro ‘__floatdisf’
  #define FUNC __floatdisf
               ^
src/libgcc/libgcc2.c:1579:1: note: in expansion of macro ‘FUNC’
  FUNC (DWtype u)
  ^
0xc1a278 rtl_check_failed_code_mode(rtx_def const*, rtx_code, machine_mode,
bool, char const*, int, char const*)
         src/gcc/rtl.c:833
0x1062b7b aarch64_float_const_representable_p(rtx_def*)
         src/gcc/config/aarch64/aarch64.c:8481
0x1052ae2 aarch64_rtx_costs
         src/gcc/config/aarch64/aarch64.c:5003
0x1059a9c aarch64_rtx_costs_wrapper
         src/gcc/config/aarch64/aarch64.c:5804
0xc38f1a rtx_cost(rtx_def*, rtx_code, int, bool)
         src/gcc/rtlanal.c:3878
0xb159a9 avoid_expensive_constant
         src/gcc/optabs.c:1389
0xb15ca9 expand_binop_directly
         src/gcc/optabs.c:1441
0xb161e7 expand_binop(machine_mode, optab_tag, rtx_def*, rtx_def*, rtx_def*,
int, optab_methods)
         src/gcc/optabs.c:1568
0x87f27f expand_expr_real_2(separate_ops*, rtx_def*, machine_mode,
expand_modifier)
         src/gcc/expr.c:9200
0x87fe4b expand_expr_real_1(tree_node*, rtx_def*, machine_mode,
expand_modifier, rtx_def**, bool)
         src/gcc/expr.c:9388
0x879534 expand_expr_real(tree_node*, rtx_def*, machine_mode,
expand_modifier, rtx_def**, bool)
         src/gcc/expr.c:7921
0x79a720 expand_normal
         src/gcc/expr.h:471
0x79cdb3 do_jump_by_parts_greater
         src/gcc/dojump.c:741
0x79b682 do_jump_1(tree_code, tree_node*, tree_node*, rtx_def*, rtx_def*,
int)
         src/gcc/dojump.c:280
0x79a8ce jumpifnot_1(tree_code, tree_node*, tree_node*, rtx_def*, int)
         src/gcc/dojump.c:138
0x700de4 expand_gimple_cond
         src/gcc/cfgexpand.c:2145
0x70ed7f expand_gimple_basic_block
         src/gcc/cfgexpand.c:5063
0x71189c execute
         src/gcc/cfgexpand.c:5803
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <http://gcc.gnu.org/bugs.html> for instructions.
...

We assert here during REAL_VALUE_FROM_CONST_DOUBLE because x has VOIDmode,
which means it has integer rather than floating point type:
...
bool
aarch64_float_const_representable_p (rtx x)
{
   /* This represents our current view of how many bits
      make up the mantissa.  */
   int point_pos = 2 * HOST_BITS_PER_WIDE_INT - 1;
   int exponent;
   unsigned HOST_WIDE_INT mantissa, mask;
   REAL_VALUE_TYPE r, m;
   bool fail;

   if (!CONST_DOUBLE_P (x))
     return false;

   REAL_VALUE_FROM_CONST_DOUBLE (r, x);
...

Attached patch fixes the ICE by handling the case that x is VOIDmode, and
allows me to complete the build.

IIUC it would be better to shift the AArch64 port to
TARGET_SUPPORTS_WIDE_INT and not worry about x being VOIDmode at all.
In this case we should not have any const_double's containing
integers.

I have a couple of patches in testing that did this but with all the
other breakages in AArch64 and ARM land in the past couple of weeks I
have been unable to submit this.

I'd say it's still a good idea to fix build-breakers.

Now build and reg-tested on aarch64-linux-gnu.

OK for trunk?

Thanks,
- Tom

regards
Ramana

OK for trunk if build and test succeeds?

I think the call site in aarch64_rtx_costs may need a separate fix.

Thanks,
- Tom