This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Remove no-longer-needed fp-bit target macros
- From: "Joseph S. Myers" <joseph at codesourcery dot com>
- To: <gcc-patches at gcc dot gnu dot org>
- Cc: <hp at axis dot com>, <richard dot earnshaw at arm dot com>
- Date: Fri, 5 Sep 2014 17:21:04 +0000
- Subject: Remove no-longer-needed fp-bit target macros
- Authentication-results: sourceware.org; auth=none
This patch removes some fp-bit target macros that are no longer
needed:
* __make_dp was not really designed as a target macro, but CRIS
defined it in cris.h anyway for optimization purposes (so making it
show up on lists of target macros). Since CRIS no longer uses
fp-bit, that definition is no longer needed.
* There were no non-default definitions of ROUND_TOWARDS_ZERO.
* There were no valid non-default definitions of
LARGEST_EXPONENT_IS_NORMAL. (One definition present for ARM to
handle half-precision treated it as if it were a host-side macro,
referring to definitions only present in host-side code, but this
macro is only used in code built for the target, not the host. In
any case, fp-bit does not support half-precision, and as far as I
can tell no ARM targets use fp-bit.)
The latter two macros are poisoned by this patch; it doesn't seem
worth poisoning the first given it wasn't intended as a target macro
anyway.
Note there is still plenty more dead code in fp-bit that's outside the
scope of the present patch series cleaning up uses of host-side target
macros in target-side code. Specifically, all the TFmode support is
dead (both IEEE binary128 and IBM long double); fp-bit is only used
for SFmode and DFmode now. (IBM long double support in fp-bit became
dead when the IRIX port was removed, binary128 support when MIPS
GNU/Linux moved to using soft-fp.)
Tested with no regressions with cross to powerpc-none-eabi. OK to
commit?
gcc:
2014-09-05 Joseph Myers <joseph@codesourcery.com>
* defaults.h (LARGEST_EXPONENT_IS_NORMAL, ROUND_TOWARDS_ZERO):
Remove.
* doc/tm.texi.in (ROUND_TOWARDS_ZERO, LARGEST_EXPONENT_IS_NORMAL):
Remove.
* doc/tm.texi: Regenerate.
* system.h (LARGEST_EXPONENT_IS_NORMAL, ROUND_TOWARDS_ZERO):
Poison.
* config/arm/arm.h (LARGEST_EXPONENT_IS_NORMAL): Remove.
* config/cris/cris.h (__make_dp): Remove.
libgcc:
2014-09-05 Joseph Myers <joseph@codesourcery.com>
* fp-bit.c (pack_d, unpack_d): Remove LARGEST_EXPONENT_IS_NORMAL
and ROUND_TOWARDS_ZERO conditionals.
Index: gcc/config/arm/arm.h
===================================================================
--- gcc/config/arm/arm.h (revision 214954)
+++ gcc/config/arm/arm.h (working copy)
@@ -440,9 +440,6 @@ extern int arm_fpu_attr;
#define TARGET_DEFAULT_FLOAT_ABI ARM_FLOAT_ABI_SOFT
#endif
-#define LARGEST_EXPONENT_IS_NORMAL(bits) \
- ((bits) == 16 && arm_fp16_format == ARM_FP16_FORMAT_ALTERNATIVE)
-
#ifndef ARM_DEFAULT_ABI
#define ARM_DEFAULT_ABI ARM_ABI_APCS
#endif
Index: gcc/config/cris/cris.h
===================================================================
--- gcc/config/cris/cris.h (revision 214954)
+++ gcc/config/cris/cris.h (working copy)
@@ -80,15 +80,7 @@ along with GCC; see the file COPYING3. If not see
/* Which CPU version this is. The parsed and adjusted cris_cpu_str. */
extern int cris_cpu_version;
-/* Changing the order used to be necessary to put the fourth __make_dp
- argument (a DImode parameter) in registers, to fit with the libfunc
- parameter passing scheme used for intrinsic functions. FIXME: Check
- performance. */
-#ifdef IN_LIBGCC2
-#define __make_dp(a,b,c,d) __cris_make_dp(d,a,b,c)
-#endif
-
/* Node: Driver */
/* Also provide canonical vN definitions when user specifies an alias. */
Index: gcc/defaults.h
===================================================================
--- gcc/defaults.h (revision 214954)
+++ gcc/defaults.h (working copy)
@@ -926,14 +926,6 @@ see the files COPYING3 and COPYING.RUNTIME respect
#define PREFERRED_DEBUGGING_TYPE NO_DEBUG
#endif
-#ifndef LARGEST_EXPONENT_IS_NORMAL
-#define LARGEST_EXPONENT_IS_NORMAL(SIZE) 0
-#endif
-
-#ifndef ROUND_TOWARDS_ZERO
-#define ROUND_TOWARDS_ZERO 0
-#endif
-
#ifndef FLOAT_LIB_COMPARE_RETURNS_BOOL
#define FLOAT_LIB_COMPARE_RETURNS_BOOL(MODE, COMPARISON) false
#endif
Index: gcc/doc/tm.texi
===================================================================
--- gcc/doc/tm.texi (revision 214954)
+++ gcc/doc/tm.texi (working copy)
@@ -1341,27 +1341,6 @@ Return machine mode to be used for @code{_Unwind_W
The default is to use @code{word_mode}.
@end deftypefn
-@defmac ROUND_TOWARDS_ZERO
-If defined, this macro should be true if the prevailing rounding
-mode is towards zero.
-
-Defining this macro only affects the way @file{libgcc.a} emulates
-floating-point arithmetic.
-
-Not defining this macro is equivalent to returning zero.
-@end defmac
-
-@defmac LARGEST_EXPONENT_IS_NORMAL (@var{size})
-This macro should return true if floats with @var{size}
-bits do not have a NaN or infinity representation, but use the largest
-exponent for normal numbers instead.
-
-Defining this macro only affects the way @file{libgcc.a} emulates
-floating-point arithmetic.
-
-The default definition of this macro returns false for all sizes.
-@end defmac
-
@deftypefn {Target Hook} bool TARGET_MS_BITFIELD_LAYOUT_P (const_tree @var{record_type})
This target hook returns @code{true} if bit-fields in the given
@var{record_type} are to be laid out following the rules of Microsoft
Index: gcc/doc/tm.texi.in
===================================================================
--- gcc/doc/tm.texi.in (revision 214954)
+++ gcc/doc/tm.texi.in (working copy)
@@ -1255,27 +1255,6 @@ pattern needs to support both a 32- and a 64-bit m
@hook TARGET_UNWIND_WORD_MODE
-@defmac ROUND_TOWARDS_ZERO
-If defined, this macro should be true if the prevailing rounding
-mode is towards zero.
-
-Defining this macro only affects the way @file{libgcc.a} emulates
-floating-point arithmetic.
-
-Not defining this macro is equivalent to returning zero.
-@end defmac
-
-@defmac LARGEST_EXPONENT_IS_NORMAL (@var{size})
-This macro should return true if floats with @var{size}
-bits do not have a NaN or infinity representation, but use the largest
-exponent for normal numbers instead.
-
-Defining this macro only affects the way @file{libgcc.a} emulates
-floating-point arithmetic.
-
-The default definition of this macro returns false for all sizes.
-@end defmac
-
@hook TARGET_MS_BITFIELD_LAYOUT_P
@hook TARGET_DECIMAL_FLOAT_SUPPORTED_P
Index: gcc/system.h
===================================================================
--- gcc/system.h (revision 214954)
+++ gcc/system.h (working copy)
@@ -934,7 +934,8 @@ extern void fancy_abort (const char *, int, const
REG_CLASS_FROM_CONSTRAINT REG_CLASS_FOR_CONSTRAINT \
EXTRA_CONSTRAINT_STR EXTRA_MEMORY_CONSTRAINT \
EXTRA_ADDRESS_CONSTRAINT CONST_DOUBLE_OK_FOR_CONSTRAINT_P \
- CALLER_SAVE_PROFITABLE
+ CALLER_SAVE_PROFITABLE LARGEST_EXPONENT_IS_NORMAL \
+ ROUND_TOWARDS_ZERO
/* Hooks that are no longer used. */
#pragma GCC poison LANG_HOOKS_FUNCTION_MARK LANG_HOOKS_FUNCTION_FREE \
Index: libgcc/fp-bit.c
===================================================================
--- libgcc/fp-bit.c (revision 214954)
+++ libgcc/fp-bit.c (working copy)
@@ -202,17 +202,9 @@ pack_d (const fp_number_type *src)
int sign = src->sign;
int exp = 0;
- if (LARGEST_EXPONENT_IS_NORMAL (FRAC_NBITS) && (isnan (src) || isinf (src)))
+ if (isnan (src))
{
- /* We can't represent these values accurately. By using the
- largest possible magnitude, we guarantee that the conversion
- of infinity is at least as big as any finite number. */
exp = EXPMAX;
- fraction = ((fractype) 1 << FRACBITS) - 1;
- }
- else if (isnan (src))
- {
- exp = EXPMAX;
/* Restore the NaN's payload. */
fraction >>= NGARDS;
fraction &= QUIET_NAN - 1;
@@ -291,8 +283,7 @@ pack_d (const fp_number_type *src)
fraction >>= NGARDS;
#endif /* NO_DENORMALS */
}
- else if (!LARGEST_EXPONENT_IS_NORMAL (FRAC_NBITS)
- && __builtin_expect (src->normal_exp > EXPBIAS, 0))
+ else if (__builtin_expect (src->normal_exp > EXPBIAS, 0))
{
exp = EXPMAX;
fraction = 0;
@@ -300,35 +291,25 @@ pack_d (const fp_number_type *src)
else
{
exp = src->normal_exp + EXPBIAS;
- if (!ROUND_TOWARDS_ZERO)
+ /* IF the gard bits are the all zero, but the first, then we're
+ half way between two numbers, choose the one which makes the
+ lsb of the answer 0. */
+ if ((fraction & GARDMASK) == GARDMSB)
{
- /* IF the gard bits are the all zero, but the first, then we're
- half way between two numbers, choose the one which makes the
- lsb of the answer 0. */
- if ((fraction & GARDMASK) == GARDMSB)
- {
- if (fraction & (1 << NGARDS))
- fraction += GARDROUND + 1;
- }
- else
- {
- /* Add a one to the guards to round up */
- fraction += GARDROUND;
- }
- if (fraction >= IMPLICIT_2)
- {
- fraction >>= 1;
- exp += 1;
- }
+ if (fraction & (1 << NGARDS))
+ fraction += GARDROUND + 1;
}
- fraction >>= NGARDS;
-
- if (LARGEST_EXPONENT_IS_NORMAL (FRAC_NBITS) && exp > EXPMAX)
+ else
{
- /* Saturate on overflow. */
- exp = EXPMAX;
- fraction = ((fractype) 1 << FRACBITS) - 1;
+ /* Add a one to the guards to round up */
+ fraction += GARDROUND;
}
+ if (fraction >= IMPLICIT_2)
+ {
+ fraction >>= 1;
+ exp += 1;
+ }
+ fraction >>= NGARDS;
}
}
@@ -556,8 +537,7 @@ unpack_d (FLO_union_type * src, fp_number_type * d
dst->fraction.ll = fraction;
}
}
- else if (!LARGEST_EXPONENT_IS_NORMAL (FRAC_NBITS)
- && __builtin_expect (exp == EXPMAX, 0))
+ else if (__builtin_expect (exp == EXPMAX, 0))
{
/* Huge exponent*/
if (fraction == 0)
@@ -915,7 +895,7 @@ _fpmul_parts ( fp_number_type * a,
low <<= 1;
}
- if (!ROUND_TOWARDS_ZERO && (high & GARDMASK) == GARDMSB)
+ if ((high & GARDMASK) == GARDMSB)
{
if (high & (1 << NGARDS))
{
@@ -1035,7 +1015,7 @@ _fpdiv_parts (fp_number_type * a,
numerator *= 2;
}
- if (!ROUND_TOWARDS_ZERO && (quotient & GARDMASK) == GARDMSB)
+ if ((quotient & GARDMASK) == GARDMSB)
{
if (quotient & (1 << NGARDS))
{
--
Joseph S. Myers
joseph@codesourcery.com