On Thu, Nov 13, 2014 at 1:35 PM, mliska <mli...@suse.cz> wrote:
> gcc/ChangeLog:
>
> 2014-11-13 Martin Liska <mli...@suse.cz>
>
> * predict.c (propagate_freq): More elegant sreal API is used.
> (estimate_bb_frequencies): New static constants defined by sreal
> replace precomputed ones.
> * sreal.c (sreal::normalize): New function.
> (sreal::to_int): Likewise.
> (sreal::operator+): Likewise.
> (sreal::operator-): Likewise.
> * sreal.h: Definition of new functions added.
Please use gcc_checking_assert()s everywhere. sreal is supposed
to be fast... (I see it has current uses of gcc_assert - you may want
to mass-convert them as a followup).
> ---
> gcc/predict.c | 30 +++++++++++-------------
> gcc/sreal.c | 56 ++++++++++++++++++++++++++++++++++++--------
> gcc/sreal.h | 75
> ++++++++++++++++++++++++++++++++++++++++++++++++++++-------
> 3 files changed, 126 insertions(+), 35 deletions(-)
>
> diff --git a/gcc/predict.c b/gcc/predict.c
> index 0215e91..0f640f5 100644
> --- a/gcc/predict.c
> +++ b/gcc/predict.c
> @@ -82,7 +82,7 @@ along with GCC; see the file COPYING3. If not see
>
> /* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE,
> 1/REG_BR_PROB_BASE, 0.5, BB_FREQ_MAX. */
> -static sreal real_zero, real_one, real_almost_one, real_br_prob_base,
> +static sreal real_almost_one, real_br_prob_base,
> real_inv_br_prob_base, real_one_half, real_bb_freq_max;
>
> static void combine_predictions_for_insn (rtx_insn *, basic_block);
> @@ -2528,13 +2528,13 @@ propagate_freq (basic_block head, bitmap tovisit)
> bb->count = bb->frequency = 0;
> }
>
> - BLOCK_INFO (head)->frequency = real_one;
> + BLOCK_INFO (head)->frequency = sreal::one ();
> last = head;
> for (bb = head; bb; bb = nextbb)
> {
> edge_iterator ei;
> - sreal cyclic_probability = real_zero;
> - sreal frequency = real_zero;
> + sreal cyclic_probability = sreal::zero ();
> + sreal frequency = sreal::zero ();
>
> nextbb = BLOCK_INFO (bb)->next;
> BLOCK_INFO (bb)->next = NULL;
> @@ -2559,13 +2559,13 @@ propagate_freq (basic_block head, bitmap tovisit)
> * BLOCK_INFO (e->src)->frequency /
> REG_BR_PROB_BASE); */
>
> - sreal tmp (e->probability, 0);
> + sreal tmp = e->probability;
> tmp *= BLOCK_INFO (e->src)->frequency;
> tmp *= real_inv_br_prob_base;
> frequency += tmp;
> }
>
> - if (cyclic_probability == real_zero)
> + if (cyclic_probability == sreal::zero ())
> {
> BLOCK_INFO (bb)->frequency = frequency;
> }
> @@ -2577,7 +2577,7 @@ propagate_freq (basic_block head, bitmap tovisit)
> /* BLOCK_INFO (bb)->frequency = frequency
> / (1 - cyclic_probability) */
>
> - cyclic_probability = real_one - cyclic_probability;
> + cyclic_probability = sreal::one () - cyclic_probability;
> BLOCK_INFO (bb)->frequency = frequency / cyclic_probability;
> }
> }
> @@ -2591,7 +2591,7 @@ propagate_freq (basic_block head, bitmap tovisit)
> = ((e->probability * BLOCK_INFO (bb)->frequency)
> / REG_BR_PROB_BASE); */
>
> - sreal tmp (e->probability, 0);
> + sreal tmp = e->probability;
> tmp *= BLOCK_INFO (bb)->frequency;
> EDGE_INFO (e)->back_edge_prob = tmp * real_inv_br_prob_base;
> }
> @@ -2873,13 +2873,11 @@ estimate_bb_frequencies (bool force)
> if (!real_values_initialized)
> {
> real_values_initialized = 1;
> - real_zero = sreal (0, 0);
> - real_one = sreal (1, 0);
> - real_br_prob_base = sreal (REG_BR_PROB_BASE, 0);
> - real_bb_freq_max = sreal (BB_FREQ_MAX, 0);
> + real_br_prob_base = REG_BR_PROB_BASE;
> + real_bb_freq_max = BB_FREQ_MAX;
> real_one_half = sreal (1, -1);
> - real_inv_br_prob_base = real_one / real_br_prob_base;
> - real_almost_one = real_one - real_inv_br_prob_base;
> + real_inv_br_prob_base = sreal::one () / real_br_prob_base;
> + real_almost_one = sreal::one () - real_inv_br_prob_base;
> }
>
> mark_dfs_back_edges ();
> @@ -2897,7 +2895,7 @@ estimate_bb_frequencies (bool force)
>
> FOR_EACH_EDGE (e, ei, bb->succs)
> {
> - EDGE_INFO (e)->back_edge_prob = sreal (e->probability, 0);
> + EDGE_INFO (e)->back_edge_prob = e->probability;
> EDGE_INFO (e)->back_edge_prob *= real_inv_br_prob_base;
> }
> }
> @@ -2906,7 +2904,7 @@ estimate_bb_frequencies (bool force)
> to outermost to examine frequencies for back edges. */
> estimate_loops ();
>
> - freq_max = real_zero;
> + freq_max = sreal::zero ();
> FOR_EACH_BB_FN (bb, cfun)
> if (freq_max < BLOCK_INFO (bb)->frequency)
> freq_max = BLOCK_INFO (bb)->frequency;
> diff --git a/gcc/sreal.c b/gcc/sreal.c
> index 3f5456a..89b9c4d 100644
> --- a/gcc/sreal.c
> +++ b/gcc/sreal.c
> @@ -1,4 +1,4 @@
> -/* Simple data type for positive real numbers for the GNU compiler.
> +/* Simple data type for real numbers for the GNU compiler.
> Copyright (C) 2002-2014 Free Software Foundation, Inc.
>
> This file is part of GCC.
> @@ -17,7 +17,7 @@ You should have received a copy of the GNU General Public
> License
> along with GCC; see the file COPYING3. If not see
> <http://www.gnu.org/licenses/>. */
>
> -/* This library supports positive real numbers and 0;
> +/* This library supports real numbers;
> inf and nan are NOT supported.
> It is written to be simple and fast.
>
> @@ -102,6 +102,7 @@ sreal::normalize ()
> {
> if (m_sig == 0)
> {
> + m_negative = 0;
> m_exp = -SREAL_MAX_EXP;
> }
> else if (m_sig < SREAL_MIN_SIG)
> @@ -153,15 +154,17 @@ sreal::normalize ()
> int64_t
> sreal::to_int () const
> {
> + int64_t sign = m_negative ? -1 : 1;
> +
> if (m_exp <= -SREAL_BITS)
> return 0;
> if (m_exp >= SREAL_PART_BITS)
> - return INTTYPE_MAXIMUM (int64_t);
> + return sign * INTTYPE_MAXIMUM (int64_t);
> if (m_exp > 0)
> - return m_sig << m_exp;
> + return sign * (m_sig << m_exp);
> if (m_exp < 0)
> - return m_sig >> -m_exp;
> - return m_sig;
> + return sign * (m_sig >> -m_exp);
> + return sign * m_sig;
> }
>
> /* Return *this + other. */
> @@ -169,9 +172,19 @@ sreal::to_int () const
> sreal
> sreal::operator+ (const sreal &other) const
> {
> + const sreal *a_p = this, *b_p = &other, *bb;
> +
> + if (m_negative && !other.m_negative)
> + return other + *a_p;
> +
> + if (!m_negative && other.m_negative)
> + return *a_p - -other;
> +
> + gcc_assert (m_negative == other.m_negative);
> +
> int dexp;
> sreal tmp, r;
> -const sreal *a_p = this, *b_p = &other, *bb;
> + r.m_negative = a_p->m_negative;
>
> if (*a_p < *b_p)
> {
> @@ -211,10 +224,30 @@ sreal::operator- (const sreal &other) const
> int dexp;
> sreal tmp, r;
> const sreal *bb;
> + const sreal *a_p = this;
> +
> + /* -a - b => -a + (-b). */
> + if (m_negative && !other.m_negative)
> + return *a_p + -other;
>
> - gcc_assert (*this >= other);
> + /* a - (-b) => a + b. */
> + if (!m_negative && other.m_negative)
> + return *a_p + -other;
> +
> + gcc_assert (m_negative == other.m_negative);
> +
> + /* We want to substract a smaller number from bigger
> + for nonegative numbers. */
> + if (!m_negative && *this < other)
> + return -(other - *this);
> +
> + /* Example: -2 - (-3) => 3 - 2 */
> + if (m_negative && *this > other)
> + return -other - -(*this);
>
> dexp = m_exp - other.m_exp;
> +
> + r.m_negative = m_negative;
> r.m_exp = m_exp;
> if (dexp > SREAL_BITS)
> {
> @@ -240,7 +273,7 @@ sreal::operator- (const sreal &other) const
> sreal
> sreal::operator* (const sreal &other) const
> {
> -sreal r;
> + sreal r;
> if (m_sig < SREAL_MIN_SIG || other.m_sig < SREAL_MIN_SIG)
> {
> r.m_sig = 0;
> @@ -252,6 +285,8 @@ sreal r;
> r.m_exp = m_exp + other.m_exp;
> r.normalize ();
> }
> +
> + r.m_negative = m_negative ^ other.m_negative;
> return r;
> }
>
> @@ -261,9 +296,10 @@ sreal
> sreal::operator/ (const sreal &other) const
> {
> gcc_assert (other.m_sig != 0);
> -sreal r;
> + sreal r;
> r.m_sig = (m_sig << SREAL_PART_BITS) / other.m_sig;
> r.m_exp = m_exp - other.m_exp - SREAL_PART_BITS;
> + r.m_negative = m_negative ^ other.m_negative;
> r.normalize ();
> return r;
> }
> diff --git a/gcc/sreal.h b/gcc/sreal.h
> index 461e28b..bfed3c7 100644
> --- a/gcc/sreal.h
> +++ b/gcc/sreal.h
> @@ -1,4 +1,4 @@
> -/* Definitions for simple data type for positive real numbers.
> +/* Definitions for simple data type for real numbers.
> Copyright (C) 2002-2014 Free Software Foundation, Inc.
>
> This file is part of GCC.
> @@ -25,7 +25,7 @@ along with GCC; see the file COPYING3. If not see
>
> #define SREAL_MIN_SIG ((uint64_t) 1 << (SREAL_PART_BITS - 1))
> #define SREAL_MAX_SIG (((uint64_t) 1 << SREAL_PART_BITS) - 1)
> -#define SREAL_MAX_EXP (INT_MAX / 4)
> +#define SREAL_MAX_EXP (INT_MAX / 8)
>
> #define SREAL_BITS SREAL_PART_BITS
>
> @@ -34,10 +34,21 @@ class sreal
> {
> public:
> /* Construct an uninitialized sreal. */
> - sreal () : m_sig (-1), m_exp (-1) {}
> + sreal () : m_sig (-1), m_exp (-1), m_negative (0) {}
>
> /* Construct a sreal. */
> - sreal (uint64_t sig, int exp) : m_sig (sig), m_exp (exp) { normalize (); }
> + sreal (int64_t sig, int exp = 0) : m_exp (exp)
> + {
> + m_negative = sig < 0;
> +
> + // TODO: speed up
> + if (sig < 0)
> + sig = -sig;
also undefined behavior for sig == int64_min ...
> +
> + m_sig = (uint64_t) sig;
any reason for not making m_sig signed and using the sign of m_sig
as sign of the sreal?
> +
> + normalize ();
> + }
>
> void dump (FILE *) const;
> int64_t to_int () const;
> @@ -49,13 +60,58 @@ public:
>
> bool operator< (const sreal &other) const
> {
> - return m_exp < other.m_exp
> + if (m_negative != other.m_negative)
> + return m_negative > other.m_negative;
> +
> + bool r = m_exp < other.m_exp
> || (m_exp == other.m_exp && m_sig < other.m_sig);
> +
> + return m_negative ? !r : r;
> }
>
> bool operator== (const sreal &other) const
> {
> - return m_exp == other.m_exp && m_sig == other.m_sig;
> + return m_exp == other.m_exp && m_sig == other.m_sig
> + && m_negative == other.m_negative;
> + }
> +
> + sreal operator- () const
> + {
> + if (m_sig == 0)
> + return *this;
> +
> + sreal tmp = *this;
> + tmp.m_negative = !tmp.m_negative;
> +
> + return tmp;
> + }
> +
> + sreal shift (int sig) const
> + {
> + sreal tmp = *this;
> + tmp.m_sig += sig;
> +
> + return tmp;
> + }
> +
> + /* Return zero constant. */
> + inline static sreal zero ()
> + {
> + static const sreal zero = sreal (0);
> + return zero;
> + }
> +
> + /* Return one constant. */
> + inline static sreal one ()
> + {
> + static const sreal one = sreal (1);
> + return one;
> + }
> +
> + /* Global minimum sreal can hold. */
> + inline static sreal min ()
> + {
> + return sreal (LONG_MIN, 0);
> }
>
> private:
> @@ -63,7 +119,8 @@ private:
> void shift_right (int amount);
>
> uint64_t m_sig; /* Significant. */
> - signed int m_exp; /* Exponent. */
> + signed int m_exp: 31; /* Exponent. */
> + unsigned int m_negative: 1; /* Negative sign. */
As this gets padded to 2 * 64bits I wonder if it is necessary to
get the slowdowns for using bitfields here. I'd have just used
uint64_t m_sig; /* Significant. */
signed int m_exp; /* Exponent. */
bool m_negative;
or making m_sig signed...
Thanks,
Richard.
> };
>
> extern void debug (sreal &ref);
> @@ -76,12 +133,12 @@ inline sreal &operator+= (sreal &a, const sreal &b)
>
> inline sreal &operator-= (sreal &a, const sreal &b)
> {
> -return a = a - b;
> + return a = a - b;
> }
>
> inline sreal &operator/= (sreal &a, const sreal &b)
> {
> -return a = a / b;
> + return a = a / b;
> }
>
> inline sreal &operator*= (sreal &a, const sreal &b)
> --
> 2.1.2
>
>
