Move fold_trunc_transparent_mathfn to match.pd
authorRichard Sandiford <richard.sandiford@arm.com>
Fri, 23 Oct 2015 10:01:47 +0000 (10:01 +0000)
committerRichard Sandiford <rsandifo@gcc.gnu.org>
Fri, 23 Oct 2015 10:01:47 +0000 (10:01 +0000)
commit67dbe5829e2f9cac9deb756762cdaf5a8829cb31
tree2104e5a46faba7156714cceac0b8be8dce7c9b63
parent735a559c68e208b9dd5b5a4bb7e109289743fbeb
Move fold_trunc_transparent_mathfn to match.pd

This moves the fold rules for trunc, floor, ceil, round, nearbyint and
rint in one go, since they're tested as a group.  Most of the code is
supporting the f(x)->x fold when x is known to be integer-valued.
Like with the non-negative test, this is probably more elegantly handled
by tracking range information for reals, but until that happens, I think
we should handle it analogously to tree_expr_nonnegative_p.

I've incorporated the fix for PR68031 in the new version of
integer_valued_real_p.  However, it seemed confusing to test for an
SSA name at the head of the function rather than the case statement,
and not fall through to tree_simple_nonnegative_warnv_p (which
conceptually shouldn't care whether an update is in progress).
But tree_simple_nonnegative_warnv_p is a no-op for SSA names,
so I simply changed it to:

          return (!name_registered_for_update_p (t)
      && depth < PARAM_VALUE (PARAM_MAX_SSA_NAME_QUERY_DEPTH)
      && gimple_stmt_nonnegative_warnv_p (SSA_NAME_DEF_STMT (t),
  strict_overflow_p, depth));

and used that in the new code too.

Doing these folds later meant that IPA would start to use information
about the aborting sinf and floor in 20030125-1.c before the folds
kicked in.  I changed them from noinline to weak to stop that.

Tested on x86_64-linux-gnu, aarch64-linux-gnu and arm-linux-gnueabi.

gcc/
* builtins.c (integer_valued_real_p): Move to fold-const.c.
(fold_trunc_transparent_mathfn, fold_builtin_trunc, fold_builtin_floor)
(fold_builtin_ceil, fold_builtin_round): Delete.
(fold_builtin_1): Handle constant trunc, floor, ceil and round
arguments here.
* convert.c (convert_to_real): Remove narrowing of rounding
functions.
* fold-const.h (integer_valued_real_unary_p)
(integer_valued_real_binary_p, integer_valued_real_call_p)
(integer_valued_real_single_p, integer_valued_real_p): Declare.
* fold-const.c (tree_single_nonnegative_warnv_p): Move
name_registered_for_update_p check to SSA_NAME case statement.
Don't call tree_simple_nonnegative_warnv_p for SSA names.
(integer_valued_real_unary_p, integer_valued_real_binary_p)
(integer_valued_real_call_p, integer_valued_real_single_p)
(integer_valued_real_invalid_p): New functions.
(integer_valued_real_p): Move from fold-const.c and rework
to call the functions above.  Handle SSA names.
* gimple-fold.h (gimple_stmt_integer_valued_real_p): Declare.
* gimple-fold.c (gimple_assign_integer_valued_real_p)
(gimple_call_integer_valued_real_p, gimple_phi_integer_valued_real_p)
(gimple_stmt_integer_valued_real_p): New functions.
* match.pd: Fold f(f(x))->f(x) for fp->fp rounding functions f.
Fold f(x)->x for the same f if x is known to be integer-valued.
Fold f(extend(x))->extend(f'(x)) if doing so doesn't affect
the result.  Canonicalize floor(x) as trunc(x) if x is
nonnegative.

gcc/testsuite/
* gcc.c-torture/execute/20030125-1.c (floor, floorf, sin, sinf):
Make weak rather than noinline.
* gcc.dg/builtins-57.c: Compile with -O.
* gcc.dg/torture/builtin-integral-1.c: Skip for -O0.

From-SVN: r229221
12 files changed:
gcc/ChangeLog
gcc/builtins.c
gcc/convert.c
gcc/fold-const.c
gcc/fold-const.h
gcc/gimple-fold.c
gcc/gimple-fold.h
gcc/match.pd
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.c-torture/execute/20030125-1.c
gcc/testsuite/gcc.dg/builtins-57.c
gcc/testsuite/gcc.dg/torture/builtin-integral-1.c