On Wed, May 11, 2022 at 4:45 PM Richard Biener via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Mon, May 9, 2022 at 7:19 AM liuhongt <hongtao....@intel.com> wrote:
> >
> > This patch will enable below optimization:
> >
> >  {
> > -  int bit;
> > -  long long unsigned int _1;
> > -  long long unsigned int _2;
> > -
> >    <bb 2> [local count: 46707768]:
> > -
> > -  <bb 3> [local count: 1027034057]:
> > -  # tmp_11 = PHI <tmp_8(3), tmp_6(D)(2)>
> > -  # bit_13 = PHI <bit_9(3), 63(2)>
> > -  _1 = 1 << bit_13;
> > -  _2 = ~_1;
> > -  tmp_8 = _2 & tmp_11;
> > -  bit_9 = bit_13 + -3;
> > -  if (bit_9 != -3(OVF))
> > -    goto <bb 3>; [95.65%]
> > -  else
> > -    goto <bb 4>; [4.35%]
> > -
> > -  <bb 4> [local count: 46707768]:
> > -  return tmp_8;
> > +  tmp_12 = tmp_6(D) & 7905747460161236406;
> > +  return tmp_12;
> >
> >  }
> >
> >
> > Boostrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> > Ok for trunk?
> >
> > gcc/ChangeLog:
> >
> >         PR middle-end/103462
> >         * match.pd (bitwise_induction_p): New match.
> >         * tree-scalar-evolution.c (gimple_bitwise_induction_p):
> >         Declare.
> >         (analyze_and_compute_bitwise_induction_effect): New function.
> >         (enum bit_op_kind): New enum.
> >         (final_value_replacement_loop): Enhanced to handle bitwise
> >         induction.
> >
> > gcc/testsuite/ChangeLog:
> >
> >         * gcc.target/i386/pr103462-1.c: New test.
> >         * gcc.target/i386/pr103462-2.c: New test.
> >         * gcc.target/i386/pr103462-3.c: New test.
> >         * gcc.target/i386/pr103462-4.c: New test.
> >         * gcc.target/i386/pr103462-5.c: New test.
> >         * gcc.target/i386/pr103462-6.c: New test.
> > ---
> >  gcc/match.pd                               |   7 +
> >  gcc/testsuite/gcc.target/i386/pr103462-1.c | 111 +++++++++++++
> >  gcc/testsuite/gcc.target/i386/pr103462-2.c |  45 ++++++
> >  gcc/testsuite/gcc.target/i386/pr103462-3.c | 111 +++++++++++++
> >  gcc/testsuite/gcc.target/i386/pr103462-4.c |  46 ++++++
> >  gcc/testsuite/gcc.target/i386/pr103462-5.c | 111 +++++++++++++
> >  gcc/testsuite/gcc.target/i386/pr103462-6.c |  46 ++++++
> >  gcc/tree-scalar-evolution.cc               | 178 ++++++++++++++++++++-
> >  8 files changed, 654 insertions(+), 1 deletion(-)
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr103462-1.c
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr103462-2.c
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr103462-3.c
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr103462-4.c
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr103462-5.c
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr103462-6.c
> >
> > diff --git a/gcc/match.pd b/gcc/match.pd
> > index 6d691d302b3..24ff5f9e6a8 100644
> > --- a/gcc/match.pd
> > +++ b/gcc/match.pd
> > @@ -7746,3 +7746,10 @@ and,
> >                == TYPE_UNSIGNED (TREE_TYPE (@3))))
> >         && single_use (@4)
> >         && single_use (@5))))
> > +
> > +(for bit_op (bit_and bit_ior bit_xor)
> > + (match (bitwise_induction_p @0 @2 @3)
> > +   (bit_op:c (nop_convert1? (bit_not2?@0 (convert3? (lshift integer_onep@1 
> > @2)))) @3)))
> > +
> > +(match (bitwise_induction_p @0 @2 @3)
> > +  (bit_not (nop_convert1? (bit_xor@0 (convert2? (lshift integer_onep@1 
> > @2)) @3))))
> > diff --git a/gcc/testsuite/gcc.target/i386/pr103462-1.c 
> > b/gcc/testsuite/gcc.target/i386/pr103462-1.c
> > new file mode 100644
> > index 00000000000..1dc4c2acad6
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/pr103462-1.c
> > @@ -0,0 +1,111 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O1 -fdump-tree-sccp-details" } */
> > +/* { dg-final { scan-tree-dump-times {final value replacement} 12 "sccp" } 
> > } */
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo (unsigned long long tmp)
> > +{
> > +  for (int bit = 0; bit < 64; bit += 3)
> > +    tmp &= ~(1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo1 (unsigned long long tmp)
> > +{
> > +  for (int bit = 63; bit >= 0; bit -= 3)
> > +    tmp &= ~(1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo2 (unsigned long long tmp)
> > +{
> > +  for (int bit = 0; bit < 64; bit += 3)
> > +    tmp &= (1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo3 (unsigned long long tmp)
> > +{
> > +  for (int bit = 63; bit >= 0; bit -= 3)
> > +    tmp &= (1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo4 (unsigned long long tmp)
> > +{
> > +  for (int bit = 0; bit < 64; bit += 3)
> > +    tmp |= ~(1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo5 (unsigned long long tmp)
> > +{
> > +  for (int bit = 63; bit >= 0; bit -= 3)
> > +    tmp |= ~(1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo6 (unsigned long long tmp)
> > +{
> > +  for (int bit = 0; bit < 64; bit += 3)
> > +    tmp |= (1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo7 (unsigned long long tmp)
> > +{
> > +  for (int bit = 63; bit >= 0; bit -= 3)
> > +    tmp |= (1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo8 (unsigned long long tmp)
> > +{
> > +  for (int bit = 0; bit < 64; bit += 3)
> > +    tmp ^= ~(1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo9 (unsigned long long tmp)
> > +{
> > +  for (int bit = 63; bit >= 0; bit -= 3)
> > +    tmp ^= ~(1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo10 (unsigned long long tmp)
> > +{
> > +  for (int bit = 0; bit < 64; bit += 3)
> > +    tmp ^= (1ULL << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned long long
> > +__attribute__((noipa))
> > +foo11 (unsigned long long tmp)
> > +{
> > +  for (int bit = 63; bit >= 0; bit -= 3)
> > +    tmp ^= (1ULL << bit);
> > +  return tmp;
> > +}
> > diff --git a/gcc/testsuite/gcc.target/i386/pr103462-2.c 
> > b/gcc/testsuite/gcc.target/i386/pr103462-2.c
> > new file mode 100644
> > index 00000000000..bc375cb78d4
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/pr103462-2.c
> > @@ -0,0 +1,45 @@
> > +/* { dg-do run } */
> > +/* { dg-options "-O1" } */
> > +
> > +#include "pr103462-1.c"
> > +
> > +int main()
> > +{
> > +  unsigned long long tmp = 0x1111111111111111ULL;
> > +  if (foo (tmp) != 0x110110110110110ULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo1 (tmp) != 0x110110110110110ULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo2 (tmp) != 0x0ULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo3 (tmp) != 0x0ULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo4 (tmp) != 0xffffffffffffffffULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo5 (tmp) != 0xffffffffffffffffULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo6 (tmp) != 0x9359359359359359ULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo7 (tmp) != 0x9359359359359359ULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo8 (tmp) != 0x8358358358358358ULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo9 (tmp) != 0x8358358358358358ULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo10 (tmp) != 0x8358358358358358ULL)
> > +    __builtin_abort ();
> > +
> > +  if (foo11 (tmp) != 0x8358358358358358ULL)
> > +    __builtin_abort ();
> > +}
> > +
> > diff --git a/gcc/testsuite/gcc.target/i386/pr103462-3.c 
> > b/gcc/testsuite/gcc.target/i386/pr103462-3.c
> > new file mode 100644
> > index 00000000000..4ba248a4872
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/pr103462-3.c
> > @@ -0,0 +1,111 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O1 -fdump-tree-sccp-details" } */
> > +/* { dg-final { scan-tree-dump-times {final value replacement} 12 "sccp" } 
> > } */
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo (unsigned int tmp)
> > +{
> > +  for (int bit = 0; bit < 32; bit += 3)
> > +    tmp &= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo1 (unsigned int tmp)
> > +{
> > +  for (int bit = 31; bit >= 0; bit -= 3)
> > +    tmp &= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo2 (unsigned int tmp)
> > +{
> > +  for (int bit = 0; bit < 32; bit += 3)
> > +    tmp &= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo3 (unsigned int tmp)
> > +{
> > +  for (int bit = 31; bit >= 0; bit -= 3)
> > +    tmp &= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo4 (unsigned int tmp)
> > +{
> > +  for (int bit = 0; bit < 32; bit += 3)
> > +    tmp |= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo5 (unsigned int tmp)
> > +{
> > +  for (int bit = 31; bit >= 0; bit -= 3)
> > +    tmp |= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo6 (unsigned int tmp)
> > +{
> > +  for (int bit = 0; bit < 32; bit += 3)
> > +    tmp |= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo7 (unsigned int tmp)
> > +{
> > +  for (int bit = 31; bit >= 0; bit -= 3)
> > +    tmp |= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo8 (unsigned int tmp)
> > +{
> > +  for (int bit = 0; bit < 32; bit += 3)
> > +    tmp ^= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo9 (unsigned int tmp)
> > +{
> > +  for (int bit = 31; bit >= 0; bit -= 3)
> > +    tmp ^= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo10 (unsigned int tmp)
> > +{
> > +  for (int bit = 0; bit < 32; bit += 3)
> > +    tmp ^= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned int
> > +__attribute__((noipa))
> > +foo11 (unsigned int tmp)
> > +{
> > +  for (int bit = 31; bit >= 0; bit -= 3)
> > +    tmp ^= (1U << bit);
> > +  return tmp;
> > +}
> > diff --git a/gcc/testsuite/gcc.target/i386/pr103462-4.c 
> > b/gcc/testsuite/gcc.target/i386/pr103462-4.c
> > new file mode 100644
> > index 00000000000..e2f4056f525
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/pr103462-4.c
> > @@ -0,0 +1,46 @@
> > +/* { dg-do run } */
> > +/* { dg-options "-O1" } */
> > +
> > +#include "pr103462-3.c"
> > +
> > +int main()
> > +{
> > +  unsigned int tmp = 0x11111111U;
> > +
> > +  if (foo (tmp) != 0x10110110U)
> > +    __builtin_abort ();
> > +
> > +  if (foo1 (tmp) != 0x1101101U)
> > +    __builtin_abort ();
> > +
> > +  if (foo2 (tmp) != 0x0U)
> > +    __builtin_abort ();
> > +
> > +  if (foo3 (tmp) != 0x0U)
> > +    __builtin_abort ();
> > +
> > +  if (foo4 (tmp) != 0xffffffffU)
> > +    __builtin_abort ();
> > +
> > +  if (foo5 (tmp) != 0xffffffffU)
> > +    __builtin_abort ();
> > +
> > +  if (foo6 (tmp) != 0x59359359U)
> > +    __builtin_abort ();
> > +
> > +  if (foo7 (tmp) != 0x93593593U)
> > +    __builtin_abort ();
> > +
> > +  if (foo8 (tmp) != 0xa7ca7ca7U)
> > +    __builtin_abort ();
> > +
> > +  if (foo9 (tmp) != 0x7ca7ca7cU)
> > +    __builtin_abort ();
> > +
> > +  if (foo10 (tmp) != 0x58358358U)
> > +    __builtin_abort ();
> > +
> > +  if (foo11 (tmp) != 0x83583583U)
> > +    __builtin_abort ();
> > +}
> > +
> > diff --git a/gcc/testsuite/gcc.target/i386/pr103462-5.c 
> > b/gcc/testsuite/gcc.target/i386/pr103462-5.c
> > new file mode 100644
> > index 00000000000..1f4ffa34b48
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/pr103462-5.c
> > @@ -0,0 +1,111 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O1 -fdump-tree-sccp-details" } */
> > +/* { dg-final { scan-tree-dump-times {final value replacement} 12 "sccp" } 
> > } */
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo (unsigned short tmp)
> > +{
> > +  for (int bit = 0; bit < 16; bit += 3)
> > +    tmp &= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo1 (unsigned short tmp)
> > +{
> > +  for (int bit = 15; bit >= 0; bit -= 3)
> > +    tmp &= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo2 (unsigned short tmp)
> > +{
> > +  for (int bit = 0; bit < 16; bit += 3)
> > +    tmp &= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo3 (unsigned short tmp)
> > +{
> > +  for (int bit = 15; bit >= 0; bit -= 3)
> > +    tmp &= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo4 (unsigned short tmp)
> > +{
> > +  for (int bit = 0; bit < 16; bit += 3)
> > +    tmp |= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo5 (unsigned short tmp)
> > +{
> > +  for (int bit = 15; bit >= 0; bit -= 3)
> > +    tmp |= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo6 (unsigned short tmp)
> > +{
> > +  for (int bit = 0; bit < 16; bit += 3)
> > +    tmp |= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo7 (unsigned short tmp)
> > +{
> > +  for (int bit = 15; bit >= 0; bit -= 3)
> > +    tmp |= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo8 (unsigned short tmp)
> > +{
> > +  for (int bit = 0; bit < 16; bit += 3)
> > +    tmp ^= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo9 (unsigned short tmp)
> > +{
> > +  for (int bit = 15; bit >= 0; bit -= 3)
> > +    tmp ^= ~(1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo10 (unsigned short tmp)
> > +{
> > +  for (int bit = 0; bit < 16; bit += 3)
> > +    tmp ^= (1U << bit);
> > +  return tmp;
> > +}
> > +
> > +unsigned short
> > +__attribute__((noipa))
> > +foo11 (unsigned short tmp)
> > +{
> > +  for (int bit = 15; bit >= 0; bit -= 3)
> > +    tmp ^= (1U << bit);
> > +  return tmp;
> > +}
> > diff --git a/gcc/testsuite/gcc.target/i386/pr103462-6.c 
> > b/gcc/testsuite/gcc.target/i386/pr103462-6.c
> > new file mode 100644
> > index 00000000000..65426d71efe
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/pr103462-6.c
> > @@ -0,0 +1,46 @@
> > +/* { dg-do run } */
> > +/* { dg-options "-O1" } */
> > +
> > +#include "pr103462-5.c"
> > +
> > +int main()
> > +{
> > +  unsigned short tmp = 0x1111U;
> > +
> > +  if (foo (tmp) != 0x110)
> > +    __builtin_abort ();
> > +
> > +  if (foo1 (tmp) != 0x110)
> > +    __builtin_abort ();
> > +
> > +  if (foo2 (tmp) != 0x0)
> > +    __builtin_abort ();
> > +
> > +  if (foo3 (tmp) != 0x0)
> > +    __builtin_abort ();
> > +
> > +  if (foo4 (tmp) != 0xffff)
> > +    __builtin_abort ();
> > +
> > +  if (foo5 (tmp) != 0xffff)
> > +    __builtin_abort ();
> > +
> > +  if (foo6 (tmp) != 0x9359)
> > +    __builtin_abort ();
> > +
> > +  if (foo7 (tmp) != 0x9359)
> > +    __builtin_abort ();
> > +
> > +  if (foo8 (tmp) != 0x8358)
> > +    __builtin_abort ();
> > +
> > +  if (foo9 (tmp) != 0x8358)
> > +    __builtin_abort ();
> > +
> > +  if (foo10 (tmp) != 0x8358)
> > +    __builtin_abort ();
> > +
> > +  if (foo11 (tmp) != 0x8358)
> > +    __builtin_abort ();
> > +}
> > +
> > diff --git a/gcc/tree-scalar-evolution.cc b/gcc/tree-scalar-evolution.cc
> > index 72ceb4001e3..9b0aec4fd09 100644
> > --- a/gcc/tree-scalar-evolution.cc
> > +++ b/gcc/tree-scalar-evolution.cc
> > @@ -3487,6 +3487,164 @@ expression_expensive_p (tree expr)
> >           || expanded_size > cache.elements ());
> >  }
> >
> > +/* Match.pd function to match bitwise inductive expression.
> > +   .i.e.
> > +   _2 = 1 << _1;
> > +   _3 = ~_2;
> > +   tmp_9 = _3 & tmp_12;  */
> > +extern bool gimple_bitwise_induction_p (tree, tree *, tree (*)(tree));
> > +
> > +/* Return the inductive expression of bitwise operation if possible,
> > +   otherwise returns DEF.  */
> > +static tree
> > +analyze_and_compute_bitwise_induction_effect (class loop* ex_loop,
> > +                                             class loop* loop,
> > +                                             tree phidef,
> > +                                             tree def,
> > +                                             tree niter)
> > +{
> > +  tree match_op[3],inv, bitwise_scev, bitwise_res;
> > +  bool folded_casts = false;
> > +  tree type = TREE_TYPE (phidef);
> > +  gimple* header_phi = NULL;
>
> use a gphi *
>
Changed.
> > +
> > +  /* Match things like op2(MATCH_OP[2]), op1(MATCH_OP[1]), phidef(PHIDEF)
> > +
> > +     op2 = PHI <phidef, inv>
> > +     _1 = (int) bit_17;
> > +     _3 = 1 << _1;
> > +     op1 = ~_3;
> > +     phidef = op1 & op2;  */
> > +  if (!gimple_bitwise_induction_p (phidef, &match_op[0], NULL)
> > +      || TREE_CODE (match_op[2]) != SSA_NAME
> > +      || gimple_code (SSA_NAME_DEF_STMT (match_op[2])) != GIMPLE_PHI
>
>  || !(header_phi = dyn_cast <gphi *> (SSA_NAME_DEF_STMT (match_op[2])))
>
> > +      || gimple_phi_num_args (SSA_NAME_DEF_STMT (match_op[2])) != 2)
>
> and then use header_phi here, if you pass in the PHI to use you'll
> always have 2 args
changed.
>
> > +    return def;
> > +
> > +  header_phi = SSA_NAME_DEF_STMT (match_op[2]);
> > +  if (gimple_bb (header_phi) != loop->header)
> > +    return def;
>
> I think passing in the PHI node and the exit edge instead of phidef
> would make the above more clear
Remove the upper condition.
>
> > +  if (PHI_ARG_DEF_FROM_EDGE (header_phi, loop_latch_edge (loop)) != phidef)
> > +    return def;
> > +
> > +  inv = gimple_phi_arg_def (header_phi, 0);
> > +  if (inv == phidef)
> > +    inv = gimple_phi_arg_def (header_phi, 1);
>
> inv = gimple_phi_arg_def_from_edge (header_phi, loop_preheader_edge (loop));
>
> where is this used?
Move to the place where it's used.
>
> > +
> > +  bitwise_scev = analyze_scalar_evolution_in_loop (ex_loop, loop,
> > +                                                  match_op[1],
> > +                                                  &folded_casts);
>
> you want
>
>   bitwise_scev = analyze_scalar_evolution (loop, match_op[1]);
>   bitwise_scev = instantiate_parameters (loop, bitwise_scev);
>
> instead since you are interested in in-loop behavior?
Yes, changed.
>
> > +
> > +  /* Make sure bits is in range of type precision.  */
> > +  if (TREE_CODE (bitwise_scev) != POLYNOMIAL_CHREC
>
> I think you want to test || !INTEGRAL_TYPE_P (bitwise_scev) since you
> do not rule out pointers or vector types.
Changed.
>
> > +      || !tree_fits_uhwi_p (CHREC_LEFT (bitwise_scev))
> > +      || tree_to_uhwi (CHREC_LEFT (bitwise_scev)) >= TYPE_PRECISION (type)
> > +      || !tree_fits_shwi_p (CHREC_RIGHT (bitwise_scev)))
> > +    return def;
> > +
> > +  bitwise_res = compute_overall_effect_of_inner_loop (ex_loop, 
> > bitwise_scev);
> > +  if (!tree_fits_uhwi_p (bitwise_res)
> > +      || tree_to_uhwi (bitwise_res) >= TYPE_PRECISION (type))
> > +    return def;
>
> Hmm, so I think that bitwise_res is the final value of the 'bit' IV, correct?
> So name it bit_final?
Changed.
>
> > +
> > +enum bit_op_kind
> > +  {
> > +   INDUCTION_BIT_CLEAR,
> > +   INDUCTION_BIT_IOR,
> > +   INDUCTION_BIT_XOR,
> > +   INDUCTION_BIT_RESET,
> > +   INDUCTION_ZERO,
> > +   INDUCTION_ALL
> > +  };
> > +
> > +  enum bit_op_kind induction_kind;
> > +  enum tree_code code1
> > +    = gimple_assign_rhs_code (SSA_NAME_DEF_STMT (phidef));
> > +  enum tree_code code2
> > +    = gimple_assign_rhs_code (SSA_NAME_DEF_STMT (match_op[0]));
> > +
> > +  /* BIT_CLEAR: A &= ~(1 << bit)
> > +     BIT_RESET: A ^= (1 << bit).
> > +     BIT_IOR: A |= (1 << bit)
> > +     BIT_ZERO: A &= (1 << bit)
> > +     BIT_ALL: A |= ~(1 << bit)
> > +     BIT_XOR: A ^= ~(1 << bit).
> > +     bit is induction variable.  */
> > +  switch (code1)
> > +    {
> > +    case BIT_AND_EXPR:
> > +      induction_kind = code2 == BIT_NOT_EXPR
> > +       ? INDUCTION_BIT_CLEAR
> > +       : INDUCTION_ZERO;
> > +      break;
> > +    case BIT_IOR_EXPR:
> > +      induction_kind = code2 == BIT_NOT_EXPR
> > +       ? INDUCTION_ALL
> > +       : INDUCTION_BIT_IOR;
> > +      break;
> > +    case BIT_XOR_EXPR:
> > +      induction_kind = code2 == BIT_NOT_EXPR
> > +       ? INDUCTION_BIT_XOR
> > +       : INDUCTION_BIT_RESET;
> > +      break;
> > +      /* A ^ ~(1 << bit) is equal to ~(A ^ (1 << bit)).  */
> > +    case BIT_NOT_EXPR:
> > +      gcc_assert (code2 == BIT_XOR_EXPR);
> > +      induction_kind = INDUCTION_BIT_XOR;
> > +      break;
> > +    default:
> > +      gcc_unreachable();
> > +    }
> > +
> > +  if (induction_kind == INDUCTION_ZERO)
> > +    return build_zero_cst (type);
>
> is that true even when the loop doesn't iterate?
>
> > +  if (induction_kind == INDUCTION_ALL)
> > +    return build_all_ones_cst (type);
>
> Likewise?
the niter is known > 0 and < TYPE_PRECISION (TYPE_SIZE (type), so it
must be iterated.
>
> > +
> > +  wide_int bits = wi::zero (TYPE_PRECISION (type));
> > +  HOST_WIDE_INT start = tree_to_shwi (CHREC_LEFT (bitwise_scev));
>
> you checked fits_uhwi above, name it bit_start?
>
> > +  HOST_WIDE_INT step = tree_to_shwi (CHREC_RIGHT (bitwise_scev));
> > +  unsigned HOST_WIDE_INT num_iter = tree_to_uhwi (niter);
> > +
> > +  /* Loop tripcount should be num_iter + 1.  */
> > +  for (unsigned i = 0; i != num_iter + 1; i++)
> > +    {
> > +      bits = wi::bit_or (bits,
> > +                        wi::lshift (wi::one (TYPE_PRECISION (type)),
> > +                                    start));
>
> you can use bits = wi::set_bit (bits, start);
Changed.
>
> > +      start += step;
> > +    }
>
> Uh.  You don't limit 'num_iter' but you limit bit_final.  So supposed that
> the IV wraps and is unsigned char and you have a 256 bit integer the
> above could run quite a long time.  Please limit num_iter to TYPE_PRECISION
> as well (do you have a testcase with a wrapping IV?  try some != CST exit
> condition and  a large increment)
Limit to > 0 and < TYPE_SIZE.
>
> > +
> > +  bool inverted = false;
> > +  switch (induction_kind)
> > +    {
> > +    case INDUCTION_BIT_CLEAR:
> > +      code1 = BIT_AND_EXPR;
> > +      inverted = true;
> > +      break;
> > +    case INDUCTION_BIT_IOR:
> > +      code1 = BIT_IOR_EXPR;
> > +      break;
> > +    case INDUCTION_BIT_RESET:
> > +      code1 = BIT_XOR_EXPR;
> > +      break;
> > +    /* A ^= ~(1 << bit) is special, when loop tripcount is even,
> > +       it's equal to  A ^= bits, else A ^= ~bits.  */
> > +    case INDUCTION_BIT_XOR:
> > +      code1 = BIT_XOR_EXPR;
> > +      if (num_iter % 2 == 0)
> > +       inverted = true;
> > +      break;
> > +    default:
> > +      gcc_unreachable ();
> > +    }
> > +
> > +  if (inverted)
> > +    bits = wi::bit_not (bits);
> > +  return fold_build2 (code1, type, inv, wide_int_to_tree (type, bits));
> > +}
> > +
> >  /* Do final value replacement for LOOP, return true if we did anything.  */
> >
> >  bool
> > @@ -3519,7 +3677,8 @@ final_value_replacement_loop (class loop *loop)
> >      {
> >        gphi *phi = psi.phi ();
> >        tree rslt = PHI_RESULT (phi);
> > -      tree def = PHI_ARG_DEF_FROM_EDGE (phi, exit);
> > +      tree phidef = PHI_ARG_DEF_FROM_EDGE (phi, exit);
> > +      tree def = phidef;
> >        if (virtual_operand_p (def))
> >         {
> >           gsi_next (&psi);
> > @@ -3537,6 +3696,23 @@ final_value_replacement_loop (class loop *loop)
> >        def = analyze_scalar_evolution_in_loop (ex_loop, loop, def,
> >                                               &folded_casts);
> >        def = compute_overall_effect_of_inner_loop (ex_loop, def);
> > +
> > +      /* Handle bitwise induction expression.
> > +
> > +        .i.e.
> > +        for (int i = 0; i != 64; i+=3)
> > +          res &= ~(1UL << i);
> > +
> > +        RES can't be analyzed out by SCEV because it is not polynomially
> > +        expressible, but in fact final value of RES can be replaced by
> > +        RES & CONSTANT where CONSTANT all ones with bit {0,3,6,9,... ,63}
> > +        being cleared, similar for BIT_IOR_EXPR/BIT_XOR_EXPR.  */
> > +      if (tree_fits_uhwi_p (niter)
> > +         && tree_to_uhwi (niter)
>
> that's redundant - ah, it's the not iterating case that's excluded.
> please write it
> as tree_to_uhwi (niter) != 0, also consider assigning to an unsigned
> HOST_WIDE_INT
> and pass it that way to analyze_and_compute_bitwise_induction_effect
Changed.
>
> > +         && tree_to_uhwi (niter) != HOST_WIDE_INT_M1U)
> > +       def = analyze_and_compute_bitwise_induction_effect (ex_loop, loop,
> > +                                                           phidef, def, 
> > niter);
> > +
>
> it seems you overwrite the SCEV analysis result unconditionally here, please
> move this inside of the following if () so it is done as fallback only
> (or add it as && !(def = ...) to the condition)
original def will be returned when bitwise induction fails which means
rewriting is ok, but i think it's better to not pass original def and
use your suggestion.
Changed.
>
> >        if (!tree_does_not_contain_chrecs (def)
> >           || chrec_contains_symbols_defined_in_loop (def, ex_loop->num)
> >           /* Moving the computation from the loop may prolong life range
> > --
> > 2.18.1
> >

Here's an updated V2 patch.


-- 
BR,
Hongtao

Attachment: v2-0001-Middle-end-Enhance-final_value_replacement_loop-t.patch
Description: Binary data

Reply via email to