+2011-01-25 Sebastian Pop <sebastian.pop@amd.com>
+
+ * MAINTAINERS (linear loop transforms): Removed.
+
2011-01-25 Jakub Jelinek <jakub@redhat.com>
* config/cloog.m4 (CLOOG_REQUESTED): Use $2 if --without-cloog.
tree browser/unparser Sebastian Pop sebastian.pop@amd.com
scev, data dependence Daniel Berlin dberlin@dberlin.org
scev, data dependence Sebastian Pop sebastian.pop@amd.com
-linear loop transforms Daniel Berlin dberlin@dberlin.org
profile feedback Jan Hubicka jh@suse.cz
type-safe vectors Nathan Sidwell nathan@codesourcery.com
alias analysis Daniel Berlin dberlin@dberlin.org
+2011-01-25 Sebastian Pop <sebastian.pop@amd.com>
+
+ * Makefile.in (LAMBDA_H): Removed.
+ (TREE_DATA_REF_H): Remove dependence on LAMBDA_H.
+ (OBJS-common): Remove dependence on lambda-code.o, lambda-mat.o,
+ lambda-trans.o, and tree-loop-linear.o.
+ (lto-symtab.o): Remove dependence on LAMBDA_H.
+ (tree-loop-linear.o): Remove rule.
+ (lambda-mat.o): Same.
+ (lambda-trans.o): Same.
+ (lambda-code.o): Same.
+ (tree-vect-loop.o): Add missing dependence on TREE_DATA_REF_H.
+ (tree-vect-slp.o): Same.
+ * hwint.h (gcd): Moved here.
+ (least_common_multiple): Same.
+ * lambda-code.c: Removed.
+ * lambda-mat.c: Removed.
+ * lambda-trans.c: Removed.
+ * lambda.h: Removed.
+ * tree-loop-linear.c: Removed.
+ * lto-symtab.c: Do not include lambda.h.
+ * omega.c (gcd): Removed.
+ * passes.c (init_optimization_passes): Remove pass_linear_transform.
+ * tree-data-ref.c (print_lambda_vector): Moved here.
+ (lambda_vector_copy): Same.
+ (lambda_matrix_copy): Same.
+ (lambda_matrix_id): Same.
+ (lambda_vector_first_nz): Same.
+ (lambda_matrix_row_add): Same.
+ (lambda_matrix_row_exchange): Same.
+ (lambda_vector_mult_const): Same.
+ (lambda_vector_negate): Same.
+ (lambda_matrix_row_negate): Same.
+ (lambda_vector_equal): Same.
+ (lambda_matrix_right_hermite): Same.
+ * tree-data-ref.h: Do not include lambda.h.
+ (lambda_vector): Moved here.
+ (lambda_matrix): Same.
+ (dependence_level): Same.
+ (lambda_transform_legal_p): Removed declaration.
+ (lambda_collect_parameters): Same.
+ (lambda_compute_access_matrices): Same.
+ (lambda_vector_gcd): Same.
+ (lambda_vector_new): Same.
+ (lambda_vector_clear): Same.
+ (lambda_vector_lexico_pos): Same.
+ (lambda_vector_zerop): Same.
+ (lambda_matrix_new): Same.
+ * tree-flow.h (least_common_multiple): Removed declaration.
+ * tree-parloops.c (lambda_trans_matrix): Moved here.
+ (LTM_MATRIX): Same.
+ (LTM_ROWSIZE): Same.
+ (LTM_COLSIZE): Same.
+ (LTM_DENOMINATOR): Same.
+ (lambda_trans_matrix_new): Same.
+ (lambda_matrix_vector_mult): Same.
+ (lambda_transform_legal_p): Same.
+ * tree-pass.h (pass_linear_transform): Removed declaration.
+ * tree-ssa-loop.c (tree_linear_transform): Removed.
+ (gate_tree_linear_transform): Removed.
+ (pass_linear_transform): Removed.
+ (gate_graphite_transforms): Make flag_tree_loop_linear an alias of
+ flag_loop_interchange.
+
2011-01-25 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/47265
C_PRETTY_PRINT_H = c-family/c-pretty-print.h $(PRETTY_PRINT_H) \
$(C_COMMON_H) $(TREE_H)
SCEV_H = tree-scalar-evolution.h $(GGC_H) tree-chrec.h $(PARAMS_H)
-LAMBDA_H = lambda.h $(TREE_H) $(VEC_H) $(GGC_H)
-TREE_DATA_REF_H = tree-data-ref.h $(LAMBDA_H) omega.h graphds.h $(SCEV_H)
+TREE_DATA_REF_H = tree-data-ref.h omega.h graphds.h $(SCEV_H)
TREE_INLINE_H = tree-inline.h vecir.h
REAL_H = real.h $(MACHMODE_H)
IRA_INT_H = ira.h ira-int.h $(CFGLOOP_H) alloc-pool.h
ira-emit.o \
ira-lives.o \
jump.o \
- lambda-code.o \
- lambda-mat.o \
- lambda-trans.o \
langhooks.o \
lcm.o \
lists.o \
tree-into-ssa.o \
tree-iterator.o \
tree-loop-distribution.o \
- tree-loop-linear.o \
tree-nested.o \
tree-nrv.o \
tree-object-size.o \
$(CGRAPH_H) $(FUNCTION_H) $(GGC_H) $(EXCEPT_H) pointer-set.h \
$(BITMAP_H) langhooks.h $(LTO_STREAMER_H) lto-compress.h
lto-symtab.o: lto-symtab.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \
- $(TREE_H) $(GIMPLE_H) $(GGC_H) $(LAMBDA_H) $(HASHTAB_H) \
+ $(TREE_H) $(GIMPLE_H) $(GGC_H) $(HASHTAB_H) \
$(LTO_STREAMER_H) $(LINKER_PLUGIN_API_H) gt-lto-symtab.h
lto-opts.o: lto-opts.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TREE_H) \
$(HASHTAB_H) $(GGC_H) $(BITMAP_H) $(FLAGS_H) $(OPTS_H) $(OPTIONS_H) \
$(TM_H) $(GGC_H) $(TREE_H) $(BASIC_BLOCK_H) $(DIAGNOSTIC_H) $(TREE_FLOW_H) \
$(TREE_DUMP_H) $(CFGLOOP_H) $(CFGLAYOUT_H) $(EXPR_H) $(RECOG_H) $(OPTABS_H) \
$(DIAGNOSTIC_CORE_H) $(SCEV_H) $(TREE_VECTORIZER_H) tree-pretty-print.h \
- gimple-pretty-print.h $(TARGET_H)
+ gimple-pretty-print.h $(TARGET_H) $(TREE_DATA_REF_H)
tree-vect-loop-manip.o: tree-vect-loop-manip.c $(CONFIG_H) $(SYSTEM_H) \
coretypes.h $(TM_H) $(GGC_H) $(TREE_H) $(BASIC_BLOCK_H) $(DIAGNOSTIC_H) \
$(TREE_FLOW_H) $(TREE_DUMP_H) $(CFGLOOP_H) $(CFGLAYOUT_H) $(EXPR_H) $(DIAGNOSTIC_CORE_H) \
coretypes.h $(TM_H) $(GGC_H) $(TREE_H) $(TARGET_H) $(BASIC_BLOCK_H) \
$(DIAGNOSTIC_H) $(TREE_FLOW_H) $(TREE_DUMP_H) $(CFGLOOP_H) $(CFGLAYOUT_H) \
$(EXPR_H) $(RECOG_H) $(OPTABS_H) $(TREE_VECTORIZER_H) tree-pretty-print.h \
- gimple-pretty-print.h
+ gimple-pretty-print.h $(TREE_DATA_REF_H)
tree-vect-stmts.o: tree-vect-stmts.c $(CONFIG_H) $(SYSTEM_H) \
coretypes.h $(TM_H) $(GGC_H) $(TREE_H) $(TARGET_H) $(BASIC_BLOCK_H) \
$(DIAGNOSTIC_H) $(TREE_FLOW_H) $(TREE_DUMP_H) $(CFGLOOP_H) $(CFGLAYOUT_H) \
$(TM_H) $(GGC_H) $(TREE_H) $(DIAGNOSTIC_H) $(TREE_FLOW_H) $(TREE_DUMP_H) \
$(CFGLOOP_H) $(TREE_PASS_H) $(TREE_VECTORIZER_H) $(TIMEVAR_H) \
tree-pretty-print.h
-tree-loop-linear.o: tree-loop-linear.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \
- $(TREE_FLOW_H) $(CFGLOOP_H) $(TREE_DATA_REF_H) $(TREE_PASS_H) $(LAMBDA_H)
tree-loop-distribution.o: tree-loop-distribution.c $(CONFIG_H) $(SYSTEM_H) \
coretypes.h $(TREE_FLOW_H) $(CFGLOOP_H) $(TREE_DATA_REF_H) $(TREE_PASS_H)
tree-parloops.o: tree-parloops.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \
$(TARGET_H) $(BASIC_BLOCK_H) $(EXPR_H) output.h $(EXCEPT_H) $(TM_P_H) \
$(OPTABS_H) $(CFGLOOP_H) hard-reg-set.h $(TIMEVAR_H) \
$(TREE_PASS_H) $(DF_H) $(DBGCNT_H)
-lambda-mat.o : lambda-mat.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TREE_FLOW_H) \
- $(LAMBDA_H)
-lambda-trans.o : lambda-trans.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \
- $(TREE_FLOW_H) $(LAMBDA_H)
-lambda-code.o : lambda-code.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \
- $(TREE_FLOW_H) $(CFGLOOP_H) $(TREE_DATA_REF_H) $(LAMBDA_H) $(TREE_PASS_H)
params.o : params.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(PARAMS_H) \
$(DIAGNOSTIC_CORE_H)
pointer-set.o: pointer-set.c pointer-set.h $(CONFIG_H) $(SYSTEM_H)
#endif /* GCC_VERSION >= 3004 */
+/* Compute the greatest common divisor of two numbers using
+ Euclid's algorithm. */
+
+static inline int
+gcd (int a, int b)
+{
+ int x, y, z;
+
+ x = abs (a);
+ y = abs (b);
+
+ while (x > 0)
+ {
+ z = y % x;
+ y = x;
+ x = z;
+ }
+
+ return y;
+}
+
+/* Compute the least common multiple of two numbers A and B . */
+
+static inline int
+least_common_multiple (int a, int b)
+{
+ return (abs (a) * abs (b) / gcd (a, b));
+}
+
#endif /* ! GCC_HWINT_H */
+++ /dev/null
-/* Loop transformation code generation
- Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
- Free Software Foundation, Inc.
- Contributed by Daniel Berlin <dberlin@dberlin.org>
-
- This file is part of GCC.
-
- GCC is free software; you can redistribute it and/or modify it under
- the terms of the GNU General Public License as published by the Free
- Software Foundation; either version 3, or (at your option) any later
- version.
-
- GCC is distributed in the hope that it will be useful, but WITHOUT ANY
- WARRANTY; without even the implied warranty of MERCHANTABILITY or
- FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
- for more details.
-
- 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/>. */
-
-#include "config.h"
-#include "system.h"
-#include "coretypes.h"
-#include "tree-flow.h"
-#include "cfgloop.h"
-#include "tree-chrec.h"
-#include "tree-data-ref.h"
-#include "tree-scalar-evolution.h"
-#include "lambda.h"
-#include "tree-pass.h"
-
-/* This loop nest code generation is based on non-singular matrix
- math.
-
- A little terminology and a general sketch of the algorithm. See "A singular
- loop transformation framework based on non-singular matrices" by Wei Li and
- Keshav Pingali for formal proofs that the various statements below are
- correct.
-
- A loop iteration space represents the points traversed by the loop. A point in the
- iteration space can be represented by a vector of size <loop depth>. You can
- therefore represent the iteration space as an integral combinations of a set
- of basis vectors.
-
- A loop iteration space is dense if every integer point between the loop
- bounds is a point in the iteration space. Every loop with a step of 1
- therefore has a dense iteration space.
-
- for i = 1 to 3, step 1 is a dense iteration space.
-
- A loop iteration space is sparse if it is not dense. That is, the iteration
- space skips integer points that are within the loop bounds.
-
- for i = 1 to 3, step 2 is a sparse iteration space, because the integer point
- 2 is skipped.
-
- Dense source spaces are easy to transform, because they don't skip any
- points to begin with. Thus we can compute the exact bounds of the target
- space using min/max and floor/ceil.
-
- For a dense source space, we take the transformation matrix, decompose it
- into a lower triangular part (H) and a unimodular part (U).
- We then compute the auxiliary space from the unimodular part (source loop
- nest . U = auxiliary space) , which has two important properties:
- 1. It traverses the iterations in the same lexicographic order as the source
- space.
- 2. It is a dense space when the source is a dense space (even if the target
- space is going to be sparse).
-
- Given the auxiliary space, we use the lower triangular part to compute the
- bounds in the target space by simple matrix multiplication.
- The gaps in the target space (IE the new loop step sizes) will be the
- diagonals of the H matrix.
-
- Sparse source spaces require another step, because you can't directly compute
- the exact bounds of the auxiliary and target space from the sparse space.
- Rather than try to come up with a separate algorithm to handle sparse source
- spaces directly, we just find a legal transformation matrix that gives you
- the sparse source space, from a dense space, and then transform the dense
- space.
-
- For a regular sparse space, you can represent the source space as an integer
- lattice, and the base space of that lattice will always be dense. Thus, we
- effectively use the lattice to figure out the transformation from the lattice
- base space, to the sparse iteration space (IE what transform was applied to
- the dense space to make it sparse). We then compose this transform with the
- transformation matrix specified by the user (since our matrix transformations
- are closed under composition, this is okay). We can then use the base space
- (which is dense) plus the composed transformation matrix, to compute the rest
- of the transform using the dense space algorithm above.
-
- In other words, our sparse source space (B) is decomposed into a dense base
- space (A), and a matrix (L) that transforms A into B, such that A.L = B.
- We then compute the composition of L and the user transformation matrix (T),
- so that T is now a transform from A to the result, instead of from B to the
- result.
- IE A.(LT) = result instead of B.T = result
- Since A is now a dense source space, we can use the dense source space
- algorithm above to compute the result of applying transform (LT) to A.
-
- Fourier-Motzkin elimination is used to compute the bounds of the base space
- of the lattice. */
-
-static bool perfect_nestify (struct loop *, VEC(tree,heap) *,
- VEC(tree,heap) *, VEC(int,heap) *,
- VEC(tree,heap) *);
-/* Lattice stuff that is internal to the code generation algorithm. */
-
-typedef struct lambda_lattice_s
-{
- /* Lattice base matrix. */
- lambda_matrix base;
- /* Lattice dimension. */
- int dimension;
- /* Origin vector for the coefficients. */
- lambda_vector origin;
- /* Origin matrix for the invariants. */
- lambda_matrix origin_invariants;
- /* Number of invariants. */
- int invariants;
-} *lambda_lattice;
-
-#define LATTICE_BASE(T) ((T)->base)
-#define LATTICE_DIMENSION(T) ((T)->dimension)
-#define LATTICE_ORIGIN(T) ((T)->origin)
-#define LATTICE_ORIGIN_INVARIANTS(T) ((T)->origin_invariants)
-#define LATTICE_INVARIANTS(T) ((T)->invariants)
-
-static bool lle_equal (lambda_linear_expression, lambda_linear_expression,
- int, int);
-static lambda_lattice lambda_lattice_new (int, int, struct obstack *);
-static lambda_lattice lambda_lattice_compute_base (lambda_loopnest,
- struct obstack *);
-
-static bool can_convert_to_perfect_nest (struct loop *);
-
-/* Create a new lambda loop in LAMBDA_OBSTACK. */
-
-static lambda_loop
-lambda_loop_new (struct obstack * lambda_obstack)
-{
- lambda_loop result = (lambda_loop)
- obstack_alloc (lambda_obstack, sizeof (struct lambda_loop_s));
- memset (result, 0, sizeof (struct lambda_loop_s));
- return result;
-}
-
-/* Create a new lambda body vector. */
-
-lambda_body_vector
-lambda_body_vector_new (int size, struct obstack * lambda_obstack)
-{
- lambda_body_vector ret;
-
- ret = (lambda_body_vector) obstack_alloc (lambda_obstack,
- sizeof (*ret));
- LBV_COEFFICIENTS (ret) = lambda_vector_new (size);
- LBV_SIZE (ret) = size;
- LBV_DENOMINATOR (ret) = 1;
- return ret;
-}
-
-/* Compute the new coefficients for the vector based on the
- *inverse* of the transformation matrix. */
-
-lambda_body_vector
-lambda_body_vector_compute_new (lambda_trans_matrix transform,
- lambda_body_vector vect,
- struct obstack * lambda_obstack)
-{
- lambda_body_vector temp;
- int depth;
-
- /* Make sure the matrix is square. */
- gcc_assert (LTM_ROWSIZE (transform) == LTM_COLSIZE (transform));
-
- depth = LTM_ROWSIZE (transform);
-
- temp = lambda_body_vector_new (depth, lambda_obstack);
- LBV_DENOMINATOR (temp) =
- LBV_DENOMINATOR (vect) * LTM_DENOMINATOR (transform);
- lambda_vector_matrix_mult (LBV_COEFFICIENTS (vect), depth,
- LTM_MATRIX (transform), depth,
- LBV_COEFFICIENTS (temp));
- LBV_SIZE (temp) = LBV_SIZE (vect);
- return temp;
-}
-
-/* Print out a lambda body vector. */
-
-void
-print_lambda_body_vector (FILE * outfile, lambda_body_vector body)
-{
- print_lambda_vector (outfile, LBV_COEFFICIENTS (body), LBV_SIZE (body));
-}
-
-/* Return TRUE if two linear expressions are equal. */
-
-static bool
-lle_equal (lambda_linear_expression lle1, lambda_linear_expression lle2,
- int depth, int invariants)
-{
- int i;
-
- if (lle1 == NULL || lle2 == NULL)
- return false;
- if (LLE_CONSTANT (lle1) != LLE_CONSTANT (lle2))
- return false;
- if (LLE_DENOMINATOR (lle1) != LLE_DENOMINATOR (lle2))
- return false;
- for (i = 0; i < depth; i++)
- if (LLE_COEFFICIENTS (lle1)[i] != LLE_COEFFICIENTS (lle2)[i])
- return false;
- for (i = 0; i < invariants; i++)
- if (LLE_INVARIANT_COEFFICIENTS (lle1)[i] !=
- LLE_INVARIANT_COEFFICIENTS (lle2)[i])
- return false;
- return true;
-}
-
-/* Create a new linear expression with dimension DIM, and total number
- of invariants INVARIANTS. */
-
-lambda_linear_expression
-lambda_linear_expression_new (int dim, int invariants,
- struct obstack * lambda_obstack)
-{
- lambda_linear_expression ret;
-
- ret = (lambda_linear_expression)obstack_alloc (lambda_obstack,
- sizeof (*ret));
- LLE_COEFFICIENTS (ret) = lambda_vector_new (dim);
- LLE_CONSTANT (ret) = 0;
- LLE_INVARIANT_COEFFICIENTS (ret) = lambda_vector_new (invariants);
- LLE_DENOMINATOR (ret) = 1;
- LLE_NEXT (ret) = NULL;
-
- return ret;
-}
-
-/* Print out a linear expression EXPR, with SIZE coefficients, to OUTFILE.
- The starting letter used for variable names is START. */
-
-static void
-print_linear_expression (FILE * outfile, lambda_vector expr, int size,
- char start)
-{
- int i;
- bool first = true;
- for (i = 0; i < size; i++)
- {
- if (expr[i] != 0)
- {
- if (first)
- {
- if (expr[i] < 0)
- fprintf (outfile, "-");
- first = false;
- }
- else if (expr[i] > 0)
- fprintf (outfile, " + ");
- else
- fprintf (outfile, " - ");
- if (abs (expr[i]) == 1)
- fprintf (outfile, "%c", start + i);
- else
- fprintf (outfile, "%d%c", abs (expr[i]), start + i);
- }
- }
-}
-
-/* Print out a lambda linear expression structure, EXPR, to OUTFILE. The
- depth/number of coefficients is given by DEPTH, the number of invariants is
- given by INVARIANTS, and the character to start variable names with is given
- by START. */
-
-void
-print_lambda_linear_expression (FILE * outfile,
- lambda_linear_expression expr,
- int depth, int invariants, char start)
-{
- fprintf (outfile, "\tLinear expression: ");
- print_linear_expression (outfile, LLE_COEFFICIENTS (expr), depth, start);
- fprintf (outfile, " constant: %d ", LLE_CONSTANT (expr));
- fprintf (outfile, " invariants: ");
- print_linear_expression (outfile, LLE_INVARIANT_COEFFICIENTS (expr),
- invariants, 'A');
- fprintf (outfile, " denominator: %d\n", LLE_DENOMINATOR (expr));
-}
-
-/* Print a lambda loop structure LOOP to OUTFILE. The depth/number of
- coefficients is given by DEPTH, the number of invariants is
- given by INVARIANTS, and the character to start variable names with is given
- by START. */
-
-void
-print_lambda_loop (FILE * outfile, lambda_loop loop, int depth,
- int invariants, char start)
-{
- int step;
- lambda_linear_expression expr;
-
- gcc_assert (loop);
-
- expr = LL_LINEAR_OFFSET (loop);
- step = LL_STEP (loop);
- fprintf (outfile, " step size = %d \n", step);
-
- if (expr)
- {
- fprintf (outfile, " linear offset: \n");
- print_lambda_linear_expression (outfile, expr, depth, invariants,
- start);
- }
-
- fprintf (outfile, " lower bound: \n");
- for (expr = LL_LOWER_BOUND (loop); expr != NULL; expr = LLE_NEXT (expr))
- print_lambda_linear_expression (outfile, expr, depth, invariants, start);
- fprintf (outfile, " upper bound: \n");
- for (expr = LL_UPPER_BOUND (loop); expr != NULL; expr = LLE_NEXT (expr))
- print_lambda_linear_expression (outfile, expr, depth, invariants, start);
-}
-
-/* Create a new loop nest structure with DEPTH loops, and INVARIANTS as the
- number of invariants. */
-
-lambda_loopnest
-lambda_loopnest_new (int depth, int invariants,
- struct obstack * lambda_obstack)
-{
- lambda_loopnest ret;
- ret = (lambda_loopnest)obstack_alloc (lambda_obstack, sizeof (*ret));
-
- LN_LOOPS (ret) = (lambda_loop *)
- obstack_alloc (lambda_obstack, depth * sizeof(LN_LOOPS(ret)));
- LN_DEPTH (ret) = depth;
- LN_INVARIANTS (ret) = invariants;
-
- return ret;
-}
-
-/* Print a lambda loopnest structure, NEST, to OUTFILE. The starting
- character to use for loop names is given by START. */
-
-void
-print_lambda_loopnest (FILE * outfile, lambda_loopnest nest, char start)
-{
- int i;
- for (i = 0; i < LN_DEPTH (nest); i++)
- {
- fprintf (outfile, "Loop %c\n", start + i);
- print_lambda_loop (outfile, LN_LOOPS (nest)[i], LN_DEPTH (nest),
- LN_INVARIANTS (nest), 'i');
- fprintf (outfile, "\n");
- }
-}
-
-/* Allocate a new lattice structure of DEPTH x DEPTH, with INVARIANTS number
- of invariants. */
-
-static lambda_lattice
-lambda_lattice_new (int depth, int invariants, struct obstack * lambda_obstack)
-{
- lambda_lattice ret
- = (lambda_lattice)obstack_alloc (lambda_obstack, sizeof (*ret));
- LATTICE_BASE (ret) = lambda_matrix_new (depth, depth, lambda_obstack);
- LATTICE_ORIGIN (ret) = lambda_vector_new (depth);
- LATTICE_ORIGIN_INVARIANTS (ret) = lambda_matrix_new (depth, invariants,
- lambda_obstack);
- LATTICE_DIMENSION (ret) = depth;
- LATTICE_INVARIANTS (ret) = invariants;
- return ret;
-}
-
-/* Compute the lattice base for NEST. The lattice base is essentially a
- non-singular transform from a dense base space to a sparse iteration space.
- We use it so that we don't have to specially handle the case of a sparse
- iteration space in other parts of the algorithm. As a result, this routine
- only does something interesting (IE produce a matrix that isn't the
- identity matrix) if NEST is a sparse space. */
-
-static lambda_lattice
-lambda_lattice_compute_base (lambda_loopnest nest,
- struct obstack * lambda_obstack)
-{
- lambda_lattice ret;
- int depth, invariants;
- lambda_matrix base;
-
- int i, j, step;
- lambda_loop loop;
- lambda_linear_expression expression;
-
- depth = LN_DEPTH (nest);
- invariants = LN_INVARIANTS (nest);
-
- ret = lambda_lattice_new (depth, invariants, lambda_obstack);
- base = LATTICE_BASE (ret);
- for (i = 0; i < depth; i++)
- {
- loop = LN_LOOPS (nest)[i];
- gcc_assert (loop);
- step = LL_STEP (loop);
- /* If we have a step of 1, then the base is one, and the
- origin and invariant coefficients are 0. */
- if (step == 1)
- {
- for (j = 0; j < depth; j++)
- base[i][j] = 0;
- base[i][i] = 1;
- LATTICE_ORIGIN (ret)[i] = 0;
- for (j = 0; j < invariants; j++)
- LATTICE_ORIGIN_INVARIANTS (ret)[i][j] = 0;
- }
- else
- {
- /* Otherwise, we need the lower bound expression (which must
- be an affine function) to determine the base. */
- expression = LL_LOWER_BOUND (loop);
- gcc_assert (expression && !LLE_NEXT (expression)
- && LLE_DENOMINATOR (expression) == 1);
-
- /* The lower triangular portion of the base is going to be the
- coefficient times the step */
- for (j = 0; j < i; j++)
- base[i][j] = LLE_COEFFICIENTS (expression)[j]
- * LL_STEP (LN_LOOPS (nest)[j]);
- base[i][i] = step;
- for (j = i + 1; j < depth; j++)
- base[i][j] = 0;
-
- /* Origin for this loop is the constant of the lower bound
- expression. */
- LATTICE_ORIGIN (ret)[i] = LLE_CONSTANT (expression);
-
- /* Coefficient for the invariants are equal to the invariant
- coefficients in the expression. */
- for (j = 0; j < invariants; j++)
- LATTICE_ORIGIN_INVARIANTS (ret)[i][j] =
- LLE_INVARIANT_COEFFICIENTS (expression)[j];
- }
- }
- return ret;
-}
-
-/* Compute the least common multiple of two numbers A and B . */
-
-int
-least_common_multiple (int a, int b)
-{
- return (abs (a) * abs (b) / gcd (a, b));
-}
-
-/* Perform Fourier-Motzkin elimination to calculate the bounds of the
- auxiliary nest.
- Fourier-Motzkin is a way of reducing systems of linear inequalities so that
- it is easy to calculate the answer and bounds.
- A sketch of how it works:
- Given a system of linear inequalities, ai * xj >= bk, you can always
- rewrite the constraints so they are all of the form
- a <= x, or x <= b, or x >= constant for some x in x1 ... xj (and some b
- in b1 ... bk, and some a in a1...ai)
- You can then eliminate this x from the non-constant inequalities by
- rewriting these as a <= b, x >= constant, and delete the x variable.
- You can then repeat this for any remaining x variables, and then we have
- an easy to use variable <= constant (or no variables at all) form that we
- can construct our bounds from.
-
- In our case, each time we eliminate, we construct part of the bound from
- the ith variable, then delete the ith variable.
-
- Remember the constant are in our vector a, our coefficient matrix is A,
- and our invariant coefficient matrix is B.
-
- SIZE is the size of the matrices being passed.
- DEPTH is the loop nest depth.
- INVARIANTS is the number of loop invariants.
- A, B, and a are the coefficient matrix, invariant coefficient, and a
- vector of constants, respectively. */
-
-static lambda_loopnest
-compute_nest_using_fourier_motzkin (int size,
- int depth,
- int invariants,
- lambda_matrix A,
- lambda_matrix B,
- lambda_vector a,
- struct obstack * lambda_obstack)
-{
-
- int multiple, f1, f2;
- int i, j, k;
- lambda_linear_expression expression;
- lambda_loop loop;
- lambda_loopnest auxillary_nest;
- lambda_matrix swapmatrix, A1, B1;
- lambda_vector swapvector, a1;
- int newsize;
-
- A1 = lambda_matrix_new (128, depth, lambda_obstack);
- B1 = lambda_matrix_new (128, invariants, lambda_obstack);
- a1 = lambda_vector_new (128);
-
- auxillary_nest = lambda_loopnest_new (depth, invariants, lambda_obstack);
-
- for (i = depth - 1; i >= 0; i--)
- {
- loop = lambda_loop_new (lambda_obstack);
- LN_LOOPS (auxillary_nest)[i] = loop;
- LL_STEP (loop) = 1;
-
- for (j = 0; j < size; j++)
- {
- if (A[j][i] < 0)
- {
- /* Any linear expression in the matrix with a coefficient less
- than 0 becomes part of the new lower bound. */
- expression = lambda_linear_expression_new (depth, invariants,
- lambda_obstack);
-
- for (k = 0; k < i; k++)
- LLE_COEFFICIENTS (expression)[k] = A[j][k];
-
- for (k = 0; k < invariants; k++)
- LLE_INVARIANT_COEFFICIENTS (expression)[k] = -1 * B[j][k];
-
- LLE_DENOMINATOR (expression) = -1 * A[j][i];
- LLE_CONSTANT (expression) = -1 * a[j];
-
- /* Ignore if identical to the existing lower bound. */
- if (!lle_equal (LL_LOWER_BOUND (loop),
- expression, depth, invariants))
- {
- LLE_NEXT (expression) = LL_LOWER_BOUND (loop);
- LL_LOWER_BOUND (loop) = expression;
- }
-
- }
- else if (A[j][i] > 0)
- {
- /* Any linear expression with a coefficient greater than 0
- becomes part of the new upper bound. */
- expression = lambda_linear_expression_new (depth, invariants,
- lambda_obstack);
- for (k = 0; k < i; k++)
- LLE_COEFFICIENTS (expression)[k] = -1 * A[j][k];
-
- for (k = 0; k < invariants; k++)
- LLE_INVARIANT_COEFFICIENTS (expression)[k] = B[j][k];
-
- LLE_DENOMINATOR (expression) = A[j][i];
- LLE_CONSTANT (expression) = a[j];
-
- /* Ignore if identical to the existing upper bound. */
- if (!lle_equal (LL_UPPER_BOUND (loop),
- expression, depth, invariants))
- {
- LLE_NEXT (expression) = LL_UPPER_BOUND (loop);
- LL_UPPER_BOUND (loop) = expression;
- }
-
- }
- }
-
- /* This portion creates a new system of linear inequalities by deleting
- the i'th variable, reducing the system by one variable. */
- newsize = 0;
- for (j = 0; j < size; j++)
- {
- /* If the coefficient for the i'th variable is 0, then we can just
- eliminate the variable straightaway. Otherwise, we have to
- multiply through by the coefficients we are eliminating. */
- if (A[j][i] == 0)
- {
- lambda_vector_copy (A[j], A1[newsize], depth);
- lambda_vector_copy (B[j], B1[newsize], invariants);
- a1[newsize] = a[j];
- newsize++;
- }
- else if (A[j][i] > 0)
- {
- for (k = 0; k < size; k++)
- {
- if (A[k][i] < 0)
- {
- multiple = least_common_multiple (A[j][i], A[k][i]);
- f1 = multiple / A[j][i];
- f2 = -1 * multiple / A[k][i];
-
- lambda_vector_add_mc (A[j], f1, A[k], f2,
- A1[newsize], depth);
- lambda_vector_add_mc (B[j], f1, B[k], f2,
- B1[newsize], invariants);
- a1[newsize] = f1 * a[j] + f2 * a[k];
- newsize++;
- }
- }
- }
- }
-
- swapmatrix = A;
- A = A1;
- A1 = swapmatrix;
-
- swapmatrix = B;
- B = B1;
- B1 = swapmatrix;
-
- swapvector = a;
- a = a1;
- a1 = swapvector;
-
- size = newsize;
- }
-
- return auxillary_nest;
-}
-
-/* Compute the loop bounds for the auxiliary space NEST.
- Input system used is Ax <= b. TRANS is the unimodular transformation.
- Given the original nest, this function will
- 1. Convert the nest into matrix form, which consists of a matrix for the
- coefficients, a matrix for the
- invariant coefficients, and a vector for the constants.
- 2. Use the matrix form to calculate the lattice base for the nest (which is
- a dense space)
- 3. Compose the dense space transform with the user specified transform, to
- get a transform we can easily calculate transformed bounds for.
- 4. Multiply the composed transformation matrix times the matrix form of the
- loop.
- 5. Transform the newly created matrix (from step 4) back into a loop nest
- using Fourier-Motzkin elimination to figure out the bounds. */
-
-static lambda_loopnest
-lambda_compute_auxillary_space (lambda_loopnest nest,
- lambda_trans_matrix trans,
- struct obstack * lambda_obstack)
-{
- lambda_matrix A, B, A1, B1;
- lambda_vector a, a1;
- lambda_matrix invertedtrans;
- int depth, invariants, size;
- int i, j;
- lambda_loop loop;
- lambda_linear_expression expression;
- lambda_lattice lattice;
-
- depth = LN_DEPTH (nest);
- invariants = LN_INVARIANTS (nest);
-
- /* Unfortunately, we can't know the number of constraints we'll have
- ahead of time, but this should be enough even in ridiculous loop nest
- cases. We must not go over this limit. */
- A = lambda_matrix_new (128, depth, lambda_obstack);
- B = lambda_matrix_new (128, invariants, lambda_obstack);
- a = lambda_vector_new (128);
-
- A1 = lambda_matrix_new (128, depth, lambda_obstack);
- B1 = lambda_matrix_new (128, invariants, lambda_obstack);
- a1 = lambda_vector_new (128);
-
- /* Store the bounds in the equation matrix A, constant vector a, and
- invariant matrix B, so that we have Ax <= a + B.
- This requires a little equation rearranging so that everything is on the
- correct side of the inequality. */
- size = 0;
- for (i = 0; i < depth; i++)
- {
- loop = LN_LOOPS (nest)[i];
-
- /* First we do the lower bound. */
- if (LL_STEP (loop) > 0)
- expression = LL_LOWER_BOUND (loop);
- else
- expression = LL_UPPER_BOUND (loop);
-
- for (; expression != NULL; expression = LLE_NEXT (expression))
- {
- /* Fill in the coefficient. */
- for (j = 0; j < i; j++)
- A[size][j] = LLE_COEFFICIENTS (expression)[j];
-
- /* And the invariant coefficient. */
- for (j = 0; j < invariants; j++)
- B[size][j] = LLE_INVARIANT_COEFFICIENTS (expression)[j];
-
- /* And the constant. */
- a[size] = LLE_CONSTANT (expression);
-
- /* Convert (2x+3y+2+b)/4 <= z to 2x+3y-4z <= -2-b. IE put all
- constants and single variables on */
- A[size][i] = -1 * LLE_DENOMINATOR (expression);
- a[size] *= -1;
- for (j = 0; j < invariants; j++)
- B[size][j] *= -1;
-
- size++;
- /* Need to increase matrix sizes above. */
- gcc_assert (size <= 127);
-
- }
-
- /* Then do the exact same thing for the upper bounds. */
- if (LL_STEP (loop) > 0)
- expression = LL_UPPER_BOUND (loop);
- else
- expression = LL_LOWER_BOUND (loop);
-
- for (; expression != NULL; expression = LLE_NEXT (expression))
- {
- /* Fill in the coefficient. */
- for (j = 0; j < i; j++)
- A[size][j] = LLE_COEFFICIENTS (expression)[j];
-
- /* And the invariant coefficient. */
- for (j = 0; j < invariants; j++)
- B[size][j] = LLE_INVARIANT_COEFFICIENTS (expression)[j];
-
- /* And the constant. */
- a[size] = LLE_CONSTANT (expression);
-
- /* Convert z <= (2x+3y+2+b)/4 to -2x-3y+4z <= 2+b. */
- for (j = 0; j < i; j++)
- A[size][j] *= -1;
- A[size][i] = LLE_DENOMINATOR (expression);
- size++;
- /* Need to increase matrix sizes above. */
- gcc_assert (size <= 127);
-
- }
- }
-
- /* Compute the lattice base x = base * y + origin, where y is the
- base space. */
- lattice = lambda_lattice_compute_base (nest, lambda_obstack);
-
- /* Ax <= a + B then becomes ALy <= a+B - A*origin. L is the lattice base */
-
- /* A1 = A * L */
- lambda_matrix_mult (A, LATTICE_BASE (lattice), A1, size, depth, depth);
-
- /* a1 = a - A * origin constant. */
- lambda_matrix_vector_mult (A, size, depth, LATTICE_ORIGIN (lattice), a1);
- lambda_vector_add_mc (a, 1, a1, -1, a1, size);
-
- /* B1 = B - A * origin invariant. */
- lambda_matrix_mult (A, LATTICE_ORIGIN_INVARIANTS (lattice), B1, size, depth,
- invariants);
- lambda_matrix_add_mc (B, 1, B1, -1, B1, size, invariants);
-
- /* Now compute the auxiliary space bounds by first inverting U, multiplying
- it by A1, then performing Fourier-Motzkin. */
-
- invertedtrans = lambda_matrix_new (depth, depth, lambda_obstack);
-
- /* Compute the inverse of U. */
- lambda_matrix_inverse (LTM_MATRIX (trans),
- invertedtrans, depth, lambda_obstack);
-
- /* A = A1 inv(U). */
- lambda_matrix_mult (A1, invertedtrans, A, size, depth, depth);
-
- return compute_nest_using_fourier_motzkin (size, depth, invariants,
- A, B1, a1, lambda_obstack);
-}
-
-/* Compute the loop bounds for the target space, using the bounds of
- the auxiliary nest AUXILLARY_NEST, and the triangular matrix H.
- The target space loop bounds are computed by multiplying the triangular
- matrix H by the auxiliary nest, to get the new loop bounds. The sign of
- the loop steps (positive or negative) is then used to swap the bounds if
- the loop counts downwards.
- Return the target loopnest. */
-
-static lambda_loopnest
-lambda_compute_target_space (lambda_loopnest auxillary_nest,
- lambda_trans_matrix H, lambda_vector stepsigns,
- struct obstack * lambda_obstack)
-{
- lambda_matrix inverse, H1;
- int determinant, i, j;
- int gcd1, gcd2;
- int factor;
-
- lambda_loopnest target_nest;
- int depth, invariants;
- lambda_matrix target;
-
- lambda_loop auxillary_loop, target_loop;
- lambda_linear_expression expression, auxillary_expr, target_expr, tmp_expr;
-
- depth = LN_DEPTH (auxillary_nest);
- invariants = LN_INVARIANTS (auxillary_nest);
-
- inverse = lambda_matrix_new (depth, depth, lambda_obstack);
- determinant = lambda_matrix_inverse (LTM_MATRIX (H), inverse, depth,
- lambda_obstack);
-
- /* H1 is H excluding its diagonal. */
- H1 = lambda_matrix_new (depth, depth, lambda_obstack);
- lambda_matrix_copy (LTM_MATRIX (H), H1, depth, depth);
-
- for (i = 0; i < depth; i++)
- H1[i][i] = 0;
-
- /* Computes the linear offsets of the loop bounds. */
- target = lambda_matrix_new (depth, depth, lambda_obstack);
- lambda_matrix_mult (H1, inverse, target, depth, depth, depth);
-
- target_nest = lambda_loopnest_new (depth, invariants, lambda_obstack);
-
- for (i = 0; i < depth; i++)
- {
-
- /* Get a new loop structure. */
- target_loop = lambda_loop_new (lambda_obstack);
- LN_LOOPS (target_nest)[i] = target_loop;
-
- /* Computes the gcd of the coefficients of the linear part. */
- gcd1 = lambda_vector_gcd (target[i], i);
-
- /* Include the denominator in the GCD. */
- gcd1 = gcd (gcd1, determinant);
-
- /* Now divide through by the gcd. */
- for (j = 0; j < i; j++)
- target[i][j] = target[i][j] / gcd1;
-
- expression = lambda_linear_expression_new (depth, invariants,
- lambda_obstack);
- lambda_vector_copy (target[i], LLE_COEFFICIENTS (expression), depth);
- LLE_DENOMINATOR (expression) = determinant / gcd1;
- LLE_CONSTANT (expression) = 0;
- lambda_vector_clear (LLE_INVARIANT_COEFFICIENTS (expression),
- invariants);
- LL_LINEAR_OFFSET (target_loop) = expression;
- }
-
- /* For each loop, compute the new bounds from H. */
- for (i = 0; i < depth; i++)
- {
- auxillary_loop = LN_LOOPS (auxillary_nest)[i];
- target_loop = LN_LOOPS (target_nest)[i];
- LL_STEP (target_loop) = LTM_MATRIX (H)[i][i];
- factor = LTM_MATRIX (H)[i][i];
-
- /* First we do the lower bound. */
- auxillary_expr = LL_LOWER_BOUND (auxillary_loop);
-
- for (; auxillary_expr != NULL;
- auxillary_expr = LLE_NEXT (auxillary_expr))
- {
- target_expr = lambda_linear_expression_new (depth, invariants,
- lambda_obstack);
- lambda_vector_matrix_mult (LLE_COEFFICIENTS (auxillary_expr),
- depth, inverse, depth,
- LLE_COEFFICIENTS (target_expr));
- lambda_vector_mult_const (LLE_COEFFICIENTS (target_expr),
- LLE_COEFFICIENTS (target_expr), depth,
- factor);
-
- LLE_CONSTANT (target_expr) = LLE_CONSTANT (auxillary_expr) * factor;
- lambda_vector_copy (LLE_INVARIANT_COEFFICIENTS (auxillary_expr),
- LLE_INVARIANT_COEFFICIENTS (target_expr),
- invariants);
- lambda_vector_mult_const (LLE_INVARIANT_COEFFICIENTS (target_expr),
- LLE_INVARIANT_COEFFICIENTS (target_expr),
- invariants, factor);
- LLE_DENOMINATOR (target_expr) = LLE_DENOMINATOR (auxillary_expr);
-
- if (!lambda_vector_zerop (LLE_COEFFICIENTS (target_expr), depth))
- {
- LLE_CONSTANT (target_expr) = LLE_CONSTANT (target_expr)
- * determinant;
- lambda_vector_mult_const (LLE_INVARIANT_COEFFICIENTS
- (target_expr),
- LLE_INVARIANT_COEFFICIENTS
- (target_expr), invariants,
- determinant);
- LLE_DENOMINATOR (target_expr) =
- LLE_DENOMINATOR (target_expr) * determinant;
- }
- /* Find the gcd and divide by it here, rather than doing it
- at the tree level. */
- gcd1 = lambda_vector_gcd (LLE_COEFFICIENTS (target_expr), depth);
- gcd2 = lambda_vector_gcd (LLE_INVARIANT_COEFFICIENTS (target_expr),
- invariants);
- gcd1 = gcd (gcd1, gcd2);
- gcd1 = gcd (gcd1, LLE_CONSTANT (target_expr));
- gcd1 = gcd (gcd1, LLE_DENOMINATOR (target_expr));
- for (j = 0; j < depth; j++)
- LLE_COEFFICIENTS (target_expr)[j] /= gcd1;
- for (j = 0; j < invariants; j++)
- LLE_INVARIANT_COEFFICIENTS (target_expr)[j] /= gcd1;
- LLE_CONSTANT (target_expr) /= gcd1;
- LLE_DENOMINATOR (target_expr) /= gcd1;
- /* Ignore if identical to existing bound. */
- if (!lle_equal (LL_LOWER_BOUND (target_loop), target_expr, depth,
- invariants))
- {
- LLE_NEXT (target_expr) = LL_LOWER_BOUND (target_loop);
- LL_LOWER_BOUND (target_loop) = target_expr;
- }
- }
- /* Now do the upper bound. */
- auxillary_expr = LL_UPPER_BOUND (auxillary_loop);
-
- for (; auxillary_expr != NULL;
- auxillary_expr = LLE_NEXT (auxillary_expr))
- {
- target_expr = lambda_linear_expression_new (depth, invariants,
- lambda_obstack);
- lambda_vector_matrix_mult (LLE_COEFFICIENTS (auxillary_expr),
- depth, inverse, depth,
- LLE_COEFFICIENTS (target_expr));
- lambda_vector_mult_const (LLE_COEFFICIENTS (target_expr),
- LLE_COEFFICIENTS (target_expr), depth,
- factor);
- LLE_CONSTANT (target_expr) = LLE_CONSTANT (auxillary_expr) * factor;
- lambda_vector_copy (LLE_INVARIANT_COEFFICIENTS (auxillary_expr),
- LLE_INVARIANT_COEFFICIENTS (target_expr),
- invariants);
- lambda_vector_mult_const (LLE_INVARIANT_COEFFICIENTS (target_expr),
- LLE_INVARIANT_COEFFICIENTS (target_expr),
- invariants, factor);
- LLE_DENOMINATOR (target_expr) = LLE_DENOMINATOR (auxillary_expr);
-
- if (!lambda_vector_zerop (LLE_COEFFICIENTS (target_expr), depth))
- {
- LLE_CONSTANT (target_expr) = LLE_CONSTANT (target_expr)
- * determinant;
- lambda_vector_mult_const (LLE_INVARIANT_COEFFICIENTS
- (target_expr),
- LLE_INVARIANT_COEFFICIENTS
- (target_expr), invariants,
- determinant);
- LLE_DENOMINATOR (target_expr) =
- LLE_DENOMINATOR (target_expr) * determinant;
- }
- /* Find the gcd and divide by it here, instead of at the
- tree level. */
- gcd1 = lambda_vector_gcd (LLE_COEFFICIENTS (target_expr), depth);
- gcd2 = lambda_vector_gcd (LLE_INVARIANT_COEFFICIENTS (target_expr),
- invariants);
- gcd1 = gcd (gcd1, gcd2);
- gcd1 = gcd (gcd1, LLE_CONSTANT (target_expr));
- gcd1 = gcd (gcd1, LLE_DENOMINATOR (target_expr));
- for (j = 0; j < depth; j++)
- LLE_COEFFICIENTS (target_expr)[j] /= gcd1;
- for (j = 0; j < invariants; j++)
- LLE_INVARIANT_COEFFICIENTS (target_expr)[j] /= gcd1;
- LLE_CONSTANT (target_expr) /= gcd1;
- LLE_DENOMINATOR (target_expr) /= gcd1;
- /* Ignore if equal to existing bound. */
- if (!lle_equal (LL_UPPER_BOUND (target_loop), target_expr, depth,
- invariants))
- {
- LLE_NEXT (target_expr) = LL_UPPER_BOUND (target_loop);
- LL_UPPER_BOUND (target_loop) = target_expr;
- }
- }
- }
- for (i = 0; i < depth; i++)
- {
- target_loop = LN_LOOPS (target_nest)[i];
- /* If necessary, exchange the upper and lower bounds and negate
- the step size. */
- if (stepsigns[i] < 0)
- {
- LL_STEP (target_loop) *= -1;
- tmp_expr = LL_LOWER_BOUND (target_loop);
- LL_LOWER_BOUND (target_loop) = LL_UPPER_BOUND (target_loop);
- LL_UPPER_BOUND (target_loop) = tmp_expr;
- }
- }
- return target_nest;
-}
-
-/* Compute the step signs of TRANS, using TRANS and stepsigns. Return the new
- result. */
-
-static lambda_vector
-lambda_compute_step_signs (lambda_trans_matrix trans,
- lambda_vector stepsigns,
- struct obstack * lambda_obstack)
-{
- lambda_matrix matrix, H;
- int size;
- lambda_vector newsteps;
- int i, j, factor, minimum_column;
- int temp;
-
- matrix = LTM_MATRIX (trans);
- size = LTM_ROWSIZE (trans);
- H = lambda_matrix_new (size, size, lambda_obstack);
-
- newsteps = lambda_vector_new (size);
- lambda_vector_copy (stepsigns, newsteps, size);
-
- lambda_matrix_copy (matrix, H, size, size);
-
- for (j = 0; j < size; j++)
- {
- lambda_vector row;
- row = H[j];
- for (i = j; i < size; i++)
- if (row[i] < 0)
- lambda_matrix_col_negate (H, size, i);
- while (lambda_vector_first_nz (row, size, j + 1) < size)
- {
- minimum_column = lambda_vector_min_nz (row, size, j);
- lambda_matrix_col_exchange (H, size, j, minimum_column);
-
- temp = newsteps[j];
- newsteps[j] = newsteps[minimum_column];
- newsteps[minimum_column] = temp;
-
- for (i = j + 1; i < size; i++)
- {
- factor = row[i] / row[j];
- lambda_matrix_col_add (H, size, j, i, -1 * factor);
- }
- }
- }
- return newsteps;
-}
-
-/* Transform NEST according to TRANS, and return the new loopnest.
- This involves
- 1. Computing a lattice base for the transformation
- 2. Composing the dense base with the specified transformation (TRANS)
- 3. Decomposing the combined transformation into a lower triangular portion,
- and a unimodular portion.
- 4. Computing the auxiliary nest using the unimodular portion.
- 5. Computing the target nest using the auxiliary nest and the lower
- triangular portion. */
-
-lambda_loopnest
-lambda_loopnest_transform (lambda_loopnest nest, lambda_trans_matrix trans,
- struct obstack * lambda_obstack)
-{
- lambda_loopnest auxillary_nest, target_nest;
-
- int depth, invariants;
- int i, j;
- lambda_lattice lattice;
- lambda_trans_matrix trans1, H, U;
- lambda_loop loop;
- lambda_linear_expression expression;
- lambda_vector origin;
- lambda_matrix origin_invariants;
- lambda_vector stepsigns;
- int f;
-
- depth = LN_DEPTH (nest);
- invariants = LN_INVARIANTS (nest);
-
- /* Keep track of the signs of the loop steps. */
- stepsigns = lambda_vector_new (depth);
- for (i = 0; i < depth; i++)
- {
- if (LL_STEP (LN_LOOPS (nest)[i]) > 0)
- stepsigns[i] = 1;
- else
- stepsigns[i] = -1;
- }
-
- /* Compute the lattice base. */
- lattice = lambda_lattice_compute_base (nest, lambda_obstack);
- trans1 = lambda_trans_matrix_new (depth, depth, lambda_obstack);
-
- /* Multiply the transformation matrix by the lattice base. */
-
- lambda_matrix_mult (LTM_MATRIX (trans), LATTICE_BASE (lattice),
- LTM_MATRIX (trans1), depth, depth, depth);
-
- /* Compute the Hermite normal form for the new transformation matrix. */
- H = lambda_trans_matrix_new (depth, depth, lambda_obstack);
- U = lambda_trans_matrix_new (depth, depth, lambda_obstack);
- lambda_matrix_hermite (LTM_MATRIX (trans1), depth, LTM_MATRIX (H),
- LTM_MATRIX (U));
-
- /* Compute the auxiliary loop nest's space from the unimodular
- portion. */
- auxillary_nest = lambda_compute_auxillary_space (nest, U,
- lambda_obstack);
-
- /* Compute the loop step signs from the old step signs and the
- transformation matrix. */
- stepsigns = lambda_compute_step_signs (trans1, stepsigns,
- lambda_obstack);
-
- /* Compute the target loop nest space from the auxiliary nest and
- the lower triangular matrix H. */
- target_nest = lambda_compute_target_space (auxillary_nest, H, stepsigns,
- lambda_obstack);
- origin = lambda_vector_new (depth);
- origin_invariants = lambda_matrix_new (depth, invariants, lambda_obstack);
- lambda_matrix_vector_mult (LTM_MATRIX (trans), depth, depth,
- LATTICE_ORIGIN (lattice), origin);
- lambda_matrix_mult (LTM_MATRIX (trans), LATTICE_ORIGIN_INVARIANTS (lattice),
- origin_invariants, depth, depth, invariants);
-
- for (i = 0; i < depth; i++)
- {
- loop = LN_LOOPS (target_nest)[i];
- expression = LL_LINEAR_OFFSET (loop);
- if (lambda_vector_zerop (LLE_COEFFICIENTS (expression), depth))
- f = 1;
- else
- f = LLE_DENOMINATOR (expression);
-
- LLE_CONSTANT (expression) += f * origin[i];
-
- for (j = 0; j < invariants; j++)
- LLE_INVARIANT_COEFFICIENTS (expression)[j] +=
- f * origin_invariants[i][j];
- }
-
- return target_nest;
-
-}
-
-/* Convert a gcc tree expression EXPR to a lambda linear expression, and
- return the new expression. DEPTH is the depth of the loopnest.
- OUTERINDUCTIONVARS is an array of the induction variables for outer loops
- in this nest. INVARIANTS is the array of invariants for the loop. EXTRA
- is the amount we have to add/subtract from the expression because of the
- type of comparison it is used in. */
-
-static lambda_linear_expression
-gcc_tree_to_linear_expression (int depth, tree expr,
- VEC(tree,heap) *outerinductionvars,
- VEC(tree,heap) *invariants, int extra,
- struct obstack * lambda_obstack)
-{
- lambda_linear_expression lle = NULL;
- switch (TREE_CODE (expr))
- {
- case INTEGER_CST:
- {
- lle = lambda_linear_expression_new (depth, 2 * depth, lambda_obstack);
- LLE_CONSTANT (lle) = TREE_INT_CST_LOW (expr);
- if (extra != 0)
- LLE_CONSTANT (lle) += extra;
-
- LLE_DENOMINATOR (lle) = 1;
- }
- break;
- case SSA_NAME:
- {
- tree iv, invar;
- size_t i;
- FOR_EACH_VEC_ELT (tree, outerinductionvars, i, iv)
- if (iv != NULL)
- {
- if (SSA_NAME_VAR (iv) == SSA_NAME_VAR (expr))
- {
- lle = lambda_linear_expression_new (depth, 2 * depth,
- lambda_obstack);
- LLE_COEFFICIENTS (lle)[i] = 1;
- if (extra != 0)
- LLE_CONSTANT (lle) = extra;
-
- LLE_DENOMINATOR (lle) = 1;
- }
- }
- FOR_EACH_VEC_ELT (tree, invariants, i, invar)
- if (invar != NULL)
- {
- if (SSA_NAME_VAR (invar) == SSA_NAME_VAR (expr))
- {
- lle = lambda_linear_expression_new (depth, 2 * depth,
- lambda_obstack);
- LLE_INVARIANT_COEFFICIENTS (lle)[i] = 1;
- if (extra != 0)
- LLE_CONSTANT (lle) = extra;
- LLE_DENOMINATOR (lle) = 1;
- }
- }
- }
- break;
- default:
- return NULL;
- }
-
- return lle;
-}
-
-/* Return the depth of the loopnest NEST */
-
-static int
-depth_of_nest (struct loop *nest)
-{
- size_t depth = 0;
- while (nest)
- {
- depth++;
- nest = nest->inner;
- }
- return depth;
-}
-
-
-/* Return true if OP is invariant in LOOP and all outer loops. */
-
-static bool
-invariant_in_loop_and_outer_loops (struct loop *loop, tree op)
-{
- if (is_gimple_min_invariant (op))
- return true;
- if (loop_depth (loop) == 0)
- return true;
- if (!expr_invariant_in_loop_p (loop, op))
- return false;
- if (!invariant_in_loop_and_outer_loops (loop_outer (loop), op))
- return false;
- return true;
-}
-
-/* Generate a lambda loop from a gcc loop LOOP. Return the new lambda loop,
- or NULL if it could not be converted.
- DEPTH is the depth of the loop.
- INVARIANTS is a pointer to the array of loop invariants.
- The induction variable for this loop should be stored in the parameter
- OURINDUCTIONVAR.
- OUTERINDUCTIONVARS is an array of induction variables for outer loops. */
-
-static lambda_loop
-gcc_loop_to_lambda_loop (struct loop *loop, int depth,
- VEC(tree,heap) ** invariants,
- tree * ourinductionvar,
- VEC(tree,heap) * outerinductionvars,
- VEC(tree,heap) ** lboundvars,
- VEC(tree,heap) ** uboundvars,
- VEC(int,heap) ** steps,
- struct obstack * lambda_obstack)
-{
- gimple phi;
- gimple exit_cond;
- tree access_fn, inductionvar;
- tree step;
- lambda_loop lloop = NULL;
- lambda_linear_expression lbound, ubound;
- tree test_lhs, test_rhs;
- int stepint;
- int extra = 0;
- tree lboundvar, uboundvar, uboundresult;
-
- /* Find out induction var and exit condition. */
- inductionvar = find_induction_var_from_exit_cond (loop);
- exit_cond = get_loop_exit_condition (loop);
-
- if (inductionvar == NULL || exit_cond == NULL)
- {
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: Cannot determine exit condition or induction variable for loop.\n");
- return NULL;
- }
-
- if (SSA_NAME_DEF_STMT (inductionvar) == NULL)
- {
-
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: Cannot find PHI node for induction variable\n");
-
- return NULL;
- }
-
- phi = SSA_NAME_DEF_STMT (inductionvar);
- if (gimple_code (phi) != GIMPLE_PHI)
- {
- tree op = SINGLE_SSA_TREE_OPERAND (phi, SSA_OP_USE);
- if (!op)
- {
-
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: Cannot find PHI node for induction variable\n");
-
- return NULL;
- }
-
- phi = SSA_NAME_DEF_STMT (op);
- if (gimple_code (phi) != GIMPLE_PHI)
- {
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: Cannot find PHI node for induction variable\n");
- return NULL;
- }
- }
-
- /* The induction variable name/version we want to put in the array is the
- result of the induction variable phi node. */
- *ourinductionvar = PHI_RESULT (phi);
- access_fn = instantiate_parameters
- (loop, analyze_scalar_evolution (loop, PHI_RESULT (phi)));
- if (access_fn == chrec_dont_know)
- {
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: Access function for induction variable phi is unknown\n");
-
- return NULL;
- }
-
- step = evolution_part_in_loop_num (access_fn, loop->num);
- if (!step || step == chrec_dont_know)
- {
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: Cannot determine step of loop.\n");
-
- return NULL;
- }
- if (TREE_CODE (step) != INTEGER_CST)
- {
-
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: Step of loop is not integer.\n");
- return NULL;
- }
-
- stepint = TREE_INT_CST_LOW (step);
-
- /* Only want phis for induction vars, which will have two
- arguments. */
- if (gimple_phi_num_args (phi) != 2)
- {
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: PHI node for induction variable has >2 arguments\n");
- return NULL;
- }
-
- /* Another induction variable check. One argument's source should be
- in the loop, one outside the loop. */
- if (flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, 0)->src)
- && flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, 1)->src))
- {
-
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: PHI edges both inside loop, or both outside loop.\n");
-
- return NULL;
- }
-
- if (flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, 0)->src))
- {
- lboundvar = PHI_ARG_DEF (phi, 1);
- lbound = gcc_tree_to_linear_expression (depth, lboundvar,
- outerinductionvars, *invariants,
- 0, lambda_obstack);
- }
- else
- {
- lboundvar = PHI_ARG_DEF (phi, 0);
- lbound = gcc_tree_to_linear_expression (depth, lboundvar,
- outerinductionvars, *invariants,
- 0, lambda_obstack);
- }
-
- if (!lbound)
- {
-
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: Cannot convert lower bound to linear expression\n");
-
- return NULL;
- }
- /* One part of the test may be a loop invariant tree. */
- VEC_reserve (tree, heap, *invariants, 1);
- test_lhs = gimple_cond_lhs (exit_cond);
- test_rhs = gimple_cond_rhs (exit_cond);
-
- if (TREE_CODE (test_rhs) == SSA_NAME
- && invariant_in_loop_and_outer_loops (loop, test_rhs))
- VEC_quick_push (tree, *invariants, test_rhs);
- else if (TREE_CODE (test_lhs) == SSA_NAME
- && invariant_in_loop_and_outer_loops (loop, test_lhs))
- VEC_quick_push (tree, *invariants, test_lhs);
-
- /* The non-induction variable part of the test is the upper bound variable.
- */
- if (test_lhs == inductionvar)
- uboundvar = test_rhs;
- else
- uboundvar = test_lhs;
-
- /* We only size the vectors assuming we have, at max, 2 times as many
- invariants as we do loops (one for each bound).
- This is just an arbitrary number, but it has to be matched against the
- code below. */
- gcc_assert (VEC_length (tree, *invariants) <= (unsigned int) (2 * depth));
-
-
- /* We might have some leftover. */
- if (gimple_cond_code (exit_cond) == LT_EXPR)
- extra = -1 * stepint;
- else if (gimple_cond_code (exit_cond) == NE_EXPR)
- extra = -1 * stepint;
- else if (gimple_cond_code (exit_cond) == GT_EXPR)
- extra = -1 * stepint;
- else if (gimple_cond_code (exit_cond) == EQ_EXPR)
- extra = 1 * stepint;
-
- ubound = gcc_tree_to_linear_expression (depth, uboundvar,
- outerinductionvars,
- *invariants, extra, lambda_obstack);
- uboundresult = build2 (PLUS_EXPR, TREE_TYPE (uboundvar), uboundvar,
- build_int_cst (TREE_TYPE (uboundvar), extra));
- VEC_safe_push (tree, heap, *uboundvars, uboundresult);
- VEC_safe_push (tree, heap, *lboundvars, lboundvar);
- VEC_safe_push (int, heap, *steps, stepint);
- if (!ubound)
- {
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file,
- "Unable to convert loop: Cannot convert upper bound to linear expression\n");
- return NULL;
- }
-
- lloop = lambda_loop_new (lambda_obstack);
- LL_STEP (lloop) = stepint;
- LL_LOWER_BOUND (lloop) = lbound;
- LL_UPPER_BOUND (lloop) = ubound;
- return lloop;
-}
-
-/* Given a LOOP, find the induction variable it is testing against in the exit
- condition. Return the induction variable if found, NULL otherwise. */
-
-tree
-find_induction_var_from_exit_cond (struct loop *loop)
-{
- gimple expr = get_loop_exit_condition (loop);
- tree ivarop;
- tree test_lhs, test_rhs;
- if (expr == NULL)
- return NULL_TREE;
- if (gimple_code (expr) != GIMPLE_COND)
- return NULL_TREE;
- test_lhs = gimple_cond_lhs (expr);
- test_rhs = gimple_cond_rhs (expr);
-
- /* Find the side that is invariant in this loop. The ivar must be the other
- side. */
-
- if (expr_invariant_in_loop_p (loop, test_lhs))
- ivarop = test_rhs;
- else if (expr_invariant_in_loop_p (loop, test_rhs))
- ivarop = test_lhs;
- else
- return NULL_TREE;
-
- if (TREE_CODE (ivarop) != SSA_NAME)
- return NULL_TREE;
- return ivarop;
-}
-
-DEF_VEC_P(lambda_loop);
-DEF_VEC_ALLOC_P(lambda_loop,heap);
-
-/* Generate a lambda loopnest from a gcc loopnest LOOP_NEST.
- Return the new loop nest.
- INDUCTIONVARS is a pointer to an array of induction variables for the
- loopnest that will be filled in during this process.
- INVARIANTS is a pointer to an array of invariants that will be filled in
- during this process. */
-
-lambda_loopnest
-gcc_loopnest_to_lambda_loopnest (struct loop *loop_nest,
- VEC(tree,heap) **inductionvars,
- VEC(tree,heap) **invariants,
- struct obstack * lambda_obstack)
-{
- lambda_loopnest ret = NULL;
- struct loop *temp = loop_nest;
- int depth = depth_of_nest (loop_nest);
- size_t i;
- VEC(lambda_loop,heap) *loops = NULL;
- VEC(tree,heap) *uboundvars = NULL;
- VEC(tree,heap) *lboundvars = NULL;
- VEC(int,heap) *steps = NULL;
- lambda_loop newloop;
- tree inductionvar = NULL;
- bool perfect_nest = perfect_nest_p (loop_nest);
-
- if (!perfect_nest && !can_convert_to_perfect_nest (loop_nest))
- goto fail;
-
- while (temp)
- {
- newloop = gcc_loop_to_lambda_loop (temp, depth, invariants,
- &inductionvar, *inductionvars,
- &lboundvars, &uboundvars,
- &steps, lambda_obstack);
- if (!newloop)
- goto fail;
-
- VEC_safe_push (tree, heap, *inductionvars, inductionvar);
- VEC_safe_push (lambda_loop, heap, loops, newloop);
- temp = temp->inner;
- }
-
- if (!perfect_nest)
- {
- if (!perfect_nestify (loop_nest, lboundvars, uboundvars, steps,
- *inductionvars))
- {
- if (dump_file)
- fprintf (dump_file,
- "Not a perfect loop nest and couldn't convert to one.\n");
- goto fail;
- }
- else if (dump_file)
- fprintf (dump_file,
- "Successfully converted loop nest to perfect loop nest.\n");
- }
-
- ret = lambda_loopnest_new (depth, 2 * depth, lambda_obstack);
-
- FOR_EACH_VEC_ELT (lambda_loop, loops, i, newloop)
- LN_LOOPS (ret)[i] = newloop;
-
- fail:
- VEC_free (lambda_loop, heap, loops);
- VEC_free (tree, heap, uboundvars);
- VEC_free (tree, heap, lboundvars);
- VEC_free (int, heap, steps);
-
- return ret;
-}
-
-/* Convert a lambda body vector LBV to a gcc tree, and return the new tree.
- STMTS_TO_INSERT is a pointer to a tree where the statements we need to be
- inserted for us are stored. INDUCTION_VARS is the array of induction
- variables for the loop this LBV is from. TYPE is the tree type to use for
- the variables and trees involved. */
-
-static tree
-lbv_to_gcc_expression (lambda_body_vector lbv,
- tree type, VEC(tree,heap) *induction_vars,
- gimple_seq *stmts_to_insert)
-{
- int k;
- tree resvar;
- tree expr = build_linear_expr (type, LBV_COEFFICIENTS (lbv), induction_vars);
-
- k = LBV_DENOMINATOR (lbv);
- gcc_assert (k != 0);
- if (k != 1)
- expr = fold_build2 (CEIL_DIV_EXPR, type, expr, build_int_cst (type, k));
-
- resvar = create_tmp_var (type, "lbvtmp");
- add_referenced_var (resvar);
- return force_gimple_operand (fold (expr), stmts_to_insert, true, resvar);
-}
-
-/* Convert a linear expression from coefficient and constant form to a
- gcc tree.
- Return the tree that represents the final value of the expression.
- LLE is the linear expression to convert.
- OFFSET is the linear offset to apply to the expression.
- TYPE is the tree type to use for the variables and math.
- INDUCTION_VARS is a vector of induction variables for the loops.
- INVARIANTS is a vector of the loop nest invariants.
- WRAP specifies what tree code to wrap the results in, if there is more than
- one (it is either MAX_EXPR, or MIN_EXPR).
- STMTS_TO_INSERT Is a pointer to the statement list we fill in with
- statements that need to be inserted for the linear expression. */
-
-static tree
-lle_to_gcc_expression (lambda_linear_expression lle,
- lambda_linear_expression offset,
- tree type,
- VEC(tree,heap) *induction_vars,
- VEC(tree,heap) *invariants,
- enum tree_code wrap, gimple_seq *stmts_to_insert)
-{
- int k;
- tree resvar;
- tree expr = NULL_TREE;
- VEC(tree,heap) *results = NULL;
-
- gcc_assert (wrap == MAX_EXPR || wrap == MIN_EXPR);
-
- /* Build up the linear expressions. */
- for (; lle != NULL; lle = LLE_NEXT (lle))
- {
- expr = build_linear_expr (type, LLE_COEFFICIENTS (lle), induction_vars);
- expr = fold_build2 (PLUS_EXPR, type, expr,
- build_linear_expr (type,
- LLE_INVARIANT_COEFFICIENTS (lle),
- invariants));
-
- k = LLE_CONSTANT (lle);
- if (k)
- expr = fold_build2 (PLUS_EXPR, type, expr, build_int_cst (type, k));
-
- k = LLE_CONSTANT (offset);
- if (k)
- expr = fold_build2 (PLUS_EXPR, type, expr, build_int_cst (type, k));
-
- k = LLE_DENOMINATOR (lle);
- if (k != 1)
- expr = fold_build2 (wrap == MAX_EXPR ? CEIL_DIV_EXPR : FLOOR_DIV_EXPR,
- type, expr, build_int_cst (type, k));
-
- expr = fold (expr);
- VEC_safe_push (tree, heap, results, expr);
- }
-
- gcc_assert (expr);
-
- /* We may need to wrap the results in a MAX_EXPR or MIN_EXPR. */
- if (VEC_length (tree, results) > 1)
- {
- size_t i;
- tree op;
-
- expr = VEC_index (tree, results, 0);
- for (i = 1; VEC_iterate (tree, results, i, op); i++)
- expr = fold_build2 (wrap, type, expr, op);
- }
-
- VEC_free (tree, heap, results);
-
- resvar = create_tmp_var (type, "lletmp");
- add_referenced_var (resvar);
- return force_gimple_operand (fold (expr), stmts_to_insert, true, resvar);
-}
-
-/* Remove the induction variable defined at IV_STMT. */
-
-void
-remove_iv (gimple iv_stmt)
-{
- gimple_stmt_iterator si = gsi_for_stmt (iv_stmt);
-
- if (gimple_code (iv_stmt) == GIMPLE_PHI)
- {
- unsigned i;
-
- for (i = 0; i < gimple_phi_num_args (iv_stmt); i++)
- {
- gimple stmt;
- imm_use_iterator imm_iter;
- tree arg = gimple_phi_arg_def (iv_stmt, i);
- bool used = false;
-
- if (TREE_CODE (arg) != SSA_NAME)
- continue;
-
- FOR_EACH_IMM_USE_STMT (stmt, imm_iter, arg)
- if (stmt != iv_stmt && !is_gimple_debug (stmt))
- used = true;
-
- if (!used)
- remove_iv (SSA_NAME_DEF_STMT (arg));
- }
-
- remove_phi_node (&si, true);
- }
- else
- {
- gsi_remove (&si, true);
- release_defs (iv_stmt);
- }
-}
-
-/* Transform a lambda loopnest NEW_LOOPNEST, which had TRANSFORM applied to
- it, back into gcc code. This changes the
- loops, their induction variables, and their bodies, so that they
- match the transformed loopnest.
- OLD_LOOPNEST is the loopnest before we've replaced it with the new
- loopnest.
- OLD_IVS is a vector of induction variables from the old loopnest.
- INVARIANTS is a vector of loop invariants from the old loopnest.
- NEW_LOOPNEST is the new lambda loopnest to replace OLD_LOOPNEST with.
- TRANSFORM is the matrix transform that was applied to OLD_LOOPNEST to get
- NEW_LOOPNEST. */
-
-void
-lambda_loopnest_to_gcc_loopnest (struct loop *old_loopnest,
- VEC(tree,heap) *old_ivs,
- VEC(tree,heap) *invariants,
- VEC(gimple,heap) **remove_ivs,
- lambda_loopnest new_loopnest,
- lambda_trans_matrix transform,
- struct obstack * lambda_obstack)
-{
- struct loop *temp;
- size_t i = 0;
- unsigned j;
- size_t depth = 0;
- VEC(tree,heap) *new_ivs = NULL;
- tree oldiv;
- gimple_stmt_iterator bsi;
-
- transform = lambda_trans_matrix_inverse (transform, lambda_obstack);
-
- if (dump_file)
- {
- fprintf (dump_file, "Inverse of transformation matrix:\n");
- print_lambda_trans_matrix (dump_file, transform);
- }
- depth = depth_of_nest (old_loopnest);
- temp = old_loopnest;
-
- while (temp)
- {
- lambda_loop newloop;
- basic_block bb;
- edge exit;
- tree ivvar, ivvarinced;
- gimple exitcond;
- gimple_seq stmts;
- enum tree_code testtype;
- tree newupperbound, newlowerbound;
- lambda_linear_expression offset;
- tree type;
- bool insert_after;
- gimple inc_stmt;
-
- oldiv = VEC_index (tree, old_ivs, i);
- type = TREE_TYPE (oldiv);
-
- /* First, build the new induction variable temporary */
-
- ivvar = create_tmp_var (type, "lnivtmp");
- add_referenced_var (ivvar);
-
- VEC_safe_push (tree, heap, new_ivs, ivvar);
-
- newloop = LN_LOOPS (new_loopnest)[i];
-
- /* Linear offset is a bit tricky to handle. Punt on the unhandled
- cases for now. */
- offset = LL_LINEAR_OFFSET (newloop);
-
- gcc_assert (LLE_DENOMINATOR (offset) == 1 &&
- lambda_vector_zerop (LLE_COEFFICIENTS (offset), depth));
-
- /* Now build the new lower bounds, and insert the statements
- necessary to generate it on the loop preheader. */
- stmts = NULL;
- newlowerbound = lle_to_gcc_expression (LL_LOWER_BOUND (newloop),
- LL_LINEAR_OFFSET (newloop),
- type,
- new_ivs,
- invariants, MAX_EXPR, &stmts);
-
- if (stmts)
- {
- gsi_insert_seq_on_edge (loop_preheader_edge (temp), stmts);
- gsi_commit_edge_inserts ();
- }
- /* Build the new upper bound and insert its statements in the
- basic block of the exit condition */
- stmts = NULL;
- newupperbound = lle_to_gcc_expression (LL_UPPER_BOUND (newloop),
- LL_LINEAR_OFFSET (newloop),
- type,
- new_ivs,
- invariants, MIN_EXPR, &stmts);
- exit = single_exit (temp);
- exitcond = get_loop_exit_condition (temp);
- bb = gimple_bb (exitcond);
- bsi = gsi_after_labels (bb);
- if (stmts)
- gsi_insert_seq_before (&bsi, stmts, GSI_NEW_STMT);
-
- /* Create the new iv. */
-
- standard_iv_increment_position (temp, &bsi, &insert_after);
- create_iv (newlowerbound,
- build_int_cst (type, LL_STEP (newloop)),
- ivvar, temp, &bsi, insert_after, &ivvar,
- NULL);
-
- /* Unfortunately, the incremented ivvar that create_iv inserted may not
- dominate the block containing the exit condition.
- So we simply create our own incremented iv to use in the new exit
- test, and let redundancy elimination sort it out. */
- inc_stmt = gimple_build_assign_with_ops (PLUS_EXPR, SSA_NAME_VAR (ivvar),
- ivvar,
- build_int_cst (type, LL_STEP (newloop)));
-
- ivvarinced = make_ssa_name (SSA_NAME_VAR (ivvar), inc_stmt);
- gimple_assign_set_lhs (inc_stmt, ivvarinced);
- bsi = gsi_for_stmt (exitcond);
- gsi_insert_before (&bsi, inc_stmt, GSI_SAME_STMT);
-
- /* Replace the exit condition with the new upper bound
- comparison. */
-
- testtype = LL_STEP (newloop) >= 0 ? LE_EXPR : GE_EXPR;
-
- /* We want to build a conditional where true means exit the loop, and
- false means continue the loop.
- So swap the testtype if this isn't the way things are.*/
-
- if (exit->flags & EDGE_FALSE_VALUE)
- testtype = swap_tree_comparison (testtype);
-
- gimple_cond_set_condition (exitcond, testtype, newupperbound, ivvarinced);
- update_stmt (exitcond);
- VEC_replace (tree, new_ivs, i, ivvar);
-
- i++;
- temp = temp->inner;
- }
-
- /* Rewrite uses of the old ivs so that they are now specified in terms of
- the new ivs. */
-
- FOR_EACH_VEC_ELT (tree, old_ivs, i, oldiv)
- {
- imm_use_iterator imm_iter;
- use_operand_p use_p;
- tree oldiv_def;
- gimple oldiv_stmt = SSA_NAME_DEF_STMT (oldiv);
- gimple stmt;
-
- if (gimple_code (oldiv_stmt) == GIMPLE_PHI)
- oldiv_def = PHI_RESULT (oldiv_stmt);
- else
- oldiv_def = SINGLE_SSA_TREE_OPERAND (oldiv_stmt, SSA_OP_DEF);
- gcc_assert (oldiv_def != NULL_TREE);
-
- FOR_EACH_IMM_USE_STMT (stmt, imm_iter, oldiv_def)
- {
- tree newiv;
- gimple_seq stmts;
- lambda_body_vector lbv, newlbv;
-
- if (is_gimple_debug (stmt))
- continue;
-
- /* Compute the new expression for the induction
- variable. */
- depth = VEC_length (tree, new_ivs);
- lbv = lambda_body_vector_new (depth, lambda_obstack);
- LBV_COEFFICIENTS (lbv)[i] = 1;
-
- newlbv = lambda_body_vector_compute_new (transform, lbv,
- lambda_obstack);
-
- stmts = NULL;
- newiv = lbv_to_gcc_expression (newlbv, TREE_TYPE (oldiv),
- new_ivs, &stmts);
-
- if (stmts && gimple_code (stmt) != GIMPLE_PHI)
- {
- bsi = gsi_for_stmt (stmt);
- gsi_insert_seq_before (&bsi, stmts, GSI_SAME_STMT);
- }
-
- FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
- propagate_value (use_p, newiv);
-
- if (stmts && gimple_code (stmt) == GIMPLE_PHI)
- for (j = 0; j < gimple_phi_num_args (stmt); j++)
- if (gimple_phi_arg_def (stmt, j) == newiv)
- gsi_insert_seq_on_edge (gimple_phi_arg_edge (stmt, j), stmts);
-
- update_stmt (stmt);
- }
-
- /* Remove the now unused induction variable. */
- VEC_safe_push (gimple, heap, *remove_ivs, oldiv_stmt);
- }
- VEC_free (tree, heap, new_ivs);
-}
-
-/* Return TRUE if this is not interesting statement from the perspective of
- determining if we have a perfect loop nest. */
-
-static bool
-not_interesting_stmt (gimple stmt)
-{
- /* Note that COND_EXPR's aren't interesting because if they were exiting the
- loop, we would have already failed the number of exits tests. */
- if (gimple_code (stmt) == GIMPLE_LABEL
- || gimple_code (stmt) == GIMPLE_GOTO
- || gimple_code (stmt) == GIMPLE_COND
- || is_gimple_debug (stmt))
- return true;
- return false;
-}
-
-/* Return TRUE if PHI uses DEF for it's in-the-loop edge for LOOP. */
-
-static bool
-phi_loop_edge_uses_def (struct loop *loop, gimple phi, tree def)
-{
- unsigned i;
- for (i = 0; i < gimple_phi_num_args (phi); i++)
- if (flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
- if (PHI_ARG_DEF (phi, i) == def)
- return true;
- return false;
-}
-
-/* Return TRUE if STMT is a use of PHI_RESULT. */
-
-static bool
-stmt_uses_phi_result (gimple stmt, tree phi_result)
-{
- tree use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
-
- /* This is conservatively true, because we only want SIMPLE bumpers
- of the form x +- constant for our pass. */
- return (use == phi_result);
-}
-
-/* STMT is a bumper stmt for LOOP if the version it defines is used in the
- in-loop-edge in a phi node, and the operand it uses is the result of that
- phi node.
- I.E. i_29 = i_3 + 1
- i_3 = PHI (0, i_29); */
-
-static bool
-stmt_is_bumper_for_loop (struct loop *loop, gimple stmt)
-{
- gimple use;
- tree def;
- imm_use_iterator iter;
- use_operand_p use_p;
-
- def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF);
- if (!def)
- return false;
-
- FOR_EACH_IMM_USE_FAST (use_p, iter, def)
- {
- use = USE_STMT (use_p);
- if (gimple_code (use) == GIMPLE_PHI)
- {
- if (phi_loop_edge_uses_def (loop, use, def))
- if (stmt_uses_phi_result (stmt, PHI_RESULT (use)))
- return true;
- }
- }
- return false;
-}
-
-
-/* Return true if LOOP is a perfect loop nest.
- Perfect loop nests are those loop nests where all code occurs in the
- innermost loop body.
- If S is a program statement, then
-
- i.e.
- DO I = 1, 20
- S1
- DO J = 1, 20
- ...
- END DO
- END DO
- is not a perfect loop nest because of S1.
-
- DO I = 1, 20
- DO J = 1, 20
- S1
- ...
- END DO
- END DO
- is a perfect loop nest.
-
- Since we don't have high level loops anymore, we basically have to walk our
- statements and ignore those that are there because the loop needs them (IE
- the induction variable increment, and jump back to the top of the loop). */
-
-bool
-perfect_nest_p (struct loop *loop)
-{
- basic_block *bbs;
- size_t i;
- gimple exit_cond;
-
- /* Loops at depth 0 are perfect nests. */
- if (!loop->inner)
- return true;
-
- bbs = get_loop_body (loop);
- exit_cond = get_loop_exit_condition (loop);
-
- for (i = 0; i < loop->num_nodes; i++)
- {
- if (bbs[i]->loop_father == loop)
- {
- gimple_stmt_iterator bsi;
-
- for (bsi = gsi_start_bb (bbs[i]); !gsi_end_p (bsi); gsi_next (&bsi))
- {
- gimple stmt = gsi_stmt (bsi);
-
- if (gimple_code (stmt) == GIMPLE_COND
- && exit_cond != stmt)
- goto non_perfectly_nested;
-
- if (stmt == exit_cond
- || not_interesting_stmt (stmt)
- || stmt_is_bumper_for_loop (loop, stmt))
- continue;
-
- non_perfectly_nested:
- free (bbs);
- return false;
- }
- }
- }
-
- free (bbs);
-
- return perfect_nest_p (loop->inner);
-}
-
-/* Replace the USES of X in STMT, or uses with the same step as X with Y.
- YINIT is the initial value of Y, REPLACEMENTS is a hash table to
- avoid creating duplicate temporaries and FIRSTBSI is statement
- iterator where new temporaries should be inserted at the beginning
- of body basic block. */
-
-static void
-replace_uses_equiv_to_x_with_y (struct loop *loop, gimple stmt, tree x,
- int xstep, tree y, tree yinit,
- htab_t replacements,
- gimple_stmt_iterator *firstbsi)
-{
- ssa_op_iter iter;
- use_operand_p use_p;
-
- FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
- {
- tree use = USE_FROM_PTR (use_p);
- tree step = NULL_TREE;
- tree scev, init, val, var;
- gimple setstmt;
- struct tree_map *h, in;
- void **loc;
-
- /* Replace uses of X with Y right away. */
- if (use == x)
- {
- SET_USE (use_p, y);
- continue;
- }
-
- scev = instantiate_parameters (loop,
- analyze_scalar_evolution (loop, use));
-
- if (scev == NULL || scev == chrec_dont_know)
- continue;
-
- step = evolution_part_in_loop_num (scev, loop->num);
- if (step == NULL
- || step == chrec_dont_know
- || TREE_CODE (step) != INTEGER_CST
- || int_cst_value (step) != xstep)
- continue;
-
- /* Use REPLACEMENTS hash table to cache already created
- temporaries. */
- in.hash = htab_hash_pointer (use);
- in.base.from = use;
- h = (struct tree_map *) htab_find_with_hash (replacements, &in, in.hash);
- if (h != NULL)
- {
- SET_USE (use_p, h->to);
- continue;
- }
-
- /* USE which has the same step as X should be replaced
- with a temporary set to Y + YINIT - INIT. */
- init = initial_condition_in_loop_num (scev, loop->num);
- gcc_assert (init != NULL && init != chrec_dont_know);
- if (TREE_TYPE (use) == TREE_TYPE (y))
- {
- val = fold_build2 (MINUS_EXPR, TREE_TYPE (y), init, yinit);
- val = fold_build2 (PLUS_EXPR, TREE_TYPE (y), y, val);
- if (val == y)
- {
- /* If X has the same type as USE, the same step
- and same initial value, it can be replaced by Y. */
- SET_USE (use_p, y);
- continue;
- }
- }
- else
- {
- val = fold_build2 (MINUS_EXPR, TREE_TYPE (y), y, yinit);
- val = fold_convert (TREE_TYPE (use), val);
- val = fold_build2 (PLUS_EXPR, TREE_TYPE (use), val, init);
- }
-
- /* Create a temporary variable and insert it at the beginning
- of the loop body basic block, right after the PHI node
- which sets Y. */
- var = create_tmp_var (TREE_TYPE (use), "perfecttmp");
- add_referenced_var (var);
- val = force_gimple_operand_gsi (firstbsi, val, false, NULL,
- true, GSI_SAME_STMT);
- setstmt = gimple_build_assign (var, val);
- var = make_ssa_name (var, setstmt);
- gimple_assign_set_lhs (setstmt, var);
- gsi_insert_before (firstbsi, setstmt, GSI_SAME_STMT);
- update_stmt (setstmt);
- SET_USE (use_p, var);
- h = ggc_alloc_tree_map ();
- h->hash = in.hash;
- h->base.from = use;
- h->to = var;
- loc = htab_find_slot_with_hash (replacements, h, in.hash, INSERT);
- gcc_assert ((*(struct tree_map **)loc) == NULL);
- *(struct tree_map **) loc = h;
- }
-}
-
-/* Return true if STMT is an exit PHI for LOOP */
-
-static bool
-exit_phi_for_loop_p (struct loop *loop, gimple stmt)
-{
- if (gimple_code (stmt) != GIMPLE_PHI
- || gimple_phi_num_args (stmt) != 1
- || gimple_bb (stmt) != single_exit (loop)->dest)
- return false;
-
- return true;
-}
-
-/* Return true if STMT can be put back into the loop INNER, by
- copying it to the beginning of that loop and changing the uses. */
-
-static bool
-can_put_in_inner_loop (struct loop *inner, gimple stmt)
-{
- imm_use_iterator imm_iter;
- use_operand_p use_p;
-
- gcc_assert (is_gimple_assign (stmt));
- if (gimple_vuse (stmt)
- || !stmt_invariant_in_loop_p (inner, stmt))
- return false;
-
- FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_assign_lhs (stmt))
- {
- if (!exit_phi_for_loop_p (inner, USE_STMT (use_p)))
- {
- basic_block immbb = gimple_bb (USE_STMT (use_p));
-
- if (!flow_bb_inside_loop_p (inner, immbb))
- return false;
- }
- }
- return true;
-}
-
-/* Return true if STMT can be put *after* the inner loop of LOOP. */
-
-static bool
-can_put_after_inner_loop (struct loop *loop, gimple stmt)
-{
- imm_use_iterator imm_iter;
- use_operand_p use_p;
-
- if (gimple_vuse (stmt))
- return false;
-
- FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_assign_lhs (stmt))
- {
- if (!exit_phi_for_loop_p (loop, USE_STMT (use_p)))
- {
- basic_block immbb = gimple_bb (USE_STMT (use_p));
-
- if (!dominated_by_p (CDI_DOMINATORS,
- immbb,
- loop->inner->header)
- && !can_put_in_inner_loop (loop->inner, stmt))
- return false;
- }
- }
- return true;
-}
-
-/* Return true when the induction variable IV is simple enough to be
- re-synthesized. */
-
-static bool
-can_duplicate_iv (tree iv, struct loop *loop)
-{
- tree scev = instantiate_parameters
- (loop, analyze_scalar_evolution (loop, iv));
-
- if (!automatically_generated_chrec_p (scev))
- {
- tree step = evolution_part_in_loop_num (scev, loop->num);
-
- if (step && step != chrec_dont_know && TREE_CODE (step) == INTEGER_CST)
- return true;
- }
-
- return false;
-}
-
-/* If this is a scalar operation that can be put back into the inner
- loop, or after the inner loop, through copying, then do so. This
- works on the theory that any amount of scalar code we have to
- reduplicate into or after the loops is less expensive that the win
- we get from rearranging the memory walk the loop is doing so that
- it has better cache behavior. */
-
-static bool
-cannot_convert_modify_to_perfect_nest (gimple stmt, struct loop *loop)
-{
- use_operand_p use_a, use_b;
- imm_use_iterator imm_iter;
- ssa_op_iter op_iter, op_iter1;
- tree op0 = gimple_assign_lhs (stmt);
-
- /* The statement should not define a variable used in the inner
- loop. */
- if (TREE_CODE (op0) == SSA_NAME
- && !can_duplicate_iv (op0, loop))
- FOR_EACH_IMM_USE_FAST (use_a, imm_iter, op0)
- if (gimple_bb (USE_STMT (use_a))->loop_father == loop->inner)
- return true;
-
- FOR_EACH_SSA_USE_OPERAND (use_a, stmt, op_iter, SSA_OP_USE)
- {
- gimple node;
- tree op = USE_FROM_PTR (use_a);
-
- /* The variables should not be used in both loops. */
- if (!can_duplicate_iv (op, loop))
- FOR_EACH_IMM_USE_FAST (use_b, imm_iter, op)
- if (gimple_bb (USE_STMT (use_b))->loop_father == loop->inner)
- return true;
-
- /* The statement should not use the value of a scalar that was
- modified in the loop. */
- node = SSA_NAME_DEF_STMT (op);
- if (gimple_code (node) == GIMPLE_PHI)
- FOR_EACH_PHI_ARG (use_b, node, op_iter1, SSA_OP_USE)
- {
- tree arg = USE_FROM_PTR (use_b);
-
- if (TREE_CODE (arg) == SSA_NAME)
- {
- gimple arg_stmt = SSA_NAME_DEF_STMT (arg);
-
- if (gimple_bb (arg_stmt)
- && (gimple_bb (arg_stmt)->loop_father == loop->inner))
- return true;
- }
- }
- }
-
- return false;
-}
-/* Return true when BB contains statements that can harm the transform
- to a perfect loop nest. */
-
-static bool
-cannot_convert_bb_to_perfect_nest (basic_block bb, struct loop *loop)
-{
- gimple_stmt_iterator bsi;
- gimple exit_condition = get_loop_exit_condition (loop);
-
- for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
- {
- gimple stmt = gsi_stmt (bsi);
-
- if (stmt == exit_condition
- || not_interesting_stmt (stmt)
- || stmt_is_bumper_for_loop (loop, stmt))
- continue;
-
- if (is_gimple_assign (stmt))
- {
- if (cannot_convert_modify_to_perfect_nest (stmt, loop))
- return true;
-
- if (can_duplicate_iv (gimple_assign_lhs (stmt), loop))
- continue;
-
- if (can_put_in_inner_loop (loop->inner, stmt)
- || can_put_after_inner_loop (loop, stmt))
- continue;
- }
-
- /* If the bb of a statement we care about isn't dominated by the
- header of the inner loop, then we can't handle this case
- right now. This test ensures that the statement comes
- completely *after* the inner loop. */
- if (!dominated_by_p (CDI_DOMINATORS,
- gimple_bb (stmt),
- loop->inner->header))
- return true;
- }
-
- return false;
-}
-
-
-/* Return TRUE if LOOP is an imperfect nest that we can convert to a
- perfect one. At the moment, we only handle imperfect nests of
- depth 2, where all of the statements occur after the inner loop. */
-
-static bool
-can_convert_to_perfect_nest (struct loop *loop)
-{
- basic_block *bbs;
- size_t i;
- gimple_stmt_iterator si;
-
- /* Can't handle triply nested+ loops yet. */
- if (!loop->inner || loop->inner->inner)
- return false;
-
- bbs = get_loop_body (loop);
- for (i = 0; i < loop->num_nodes; i++)
- if (bbs[i]->loop_father == loop
- && cannot_convert_bb_to_perfect_nest (bbs[i], loop))
- goto fail;
-
- /* We also need to make sure the loop exit only has simple copy phis in it,
- otherwise we don't know how to transform it into a perfect nest. */
- for (si = gsi_start_phis (single_exit (loop)->dest);
- !gsi_end_p (si);
- gsi_next (&si))
- if (gimple_phi_num_args (gsi_stmt (si)) != 1)
- goto fail;
-
- free (bbs);
- return true;
-
- fail:
- free (bbs);
- return false;
-}
-
-
-DEF_VEC_I(source_location);
-DEF_VEC_ALLOC_I(source_location,heap);
-
-/* Transform the loop nest into a perfect nest, if possible.
- LOOP is the loop nest to transform into a perfect nest
- LBOUNDS are the lower bounds for the loops to transform
- UBOUNDS are the upper bounds for the loops to transform
- STEPS is the STEPS for the loops to transform.
- LOOPIVS is the induction variables for the loops to transform.
-
- Basically, for the case of
-
- FOR (i = 0; i < 50; i++)
- {
- FOR (j =0; j < 50; j++)
- {
- <whatever>
- }
- <some code>
- }
-
- This function will transform it into a perfect loop nest by splitting the
- outer loop into two loops, like so:
-
- FOR (i = 0; i < 50; i++)
- {
- FOR (j = 0; j < 50; j++)
- {
- <whatever>
- }
- }
-
- FOR (i = 0; i < 50; i ++)
- {
- <some code>
- }
-
- Return FALSE if we can't make this loop into a perfect nest. */
-
-static bool
-perfect_nestify (struct loop *loop,
- VEC(tree,heap) *lbounds,
- VEC(tree,heap) *ubounds,
- VEC(int,heap) *steps,
- VEC(tree,heap) *loopivs)
-{
- basic_block *bbs;
- gimple exit_condition;
- gimple cond_stmt;
- basic_block preheaderbb, headerbb, bodybb, latchbb, olddest;
- int i;
- gimple_stmt_iterator bsi, firstbsi;
- bool insert_after;
- edge e;
- struct loop *newloop;
- gimple phi;
- tree uboundvar;
- gimple stmt;
- tree oldivvar, ivvar, ivvarinced;
- VEC(tree,heap) *phis = NULL;
- VEC(source_location,heap) *locations = NULL;
- htab_t replacements = NULL;
-
- /* Create the new loop. */
- olddest = single_exit (loop)->dest;
- preheaderbb = split_edge (single_exit (loop));
- headerbb = create_empty_bb (EXIT_BLOCK_PTR->prev_bb);
-
- /* Push the exit phi nodes that we are moving. */
- for (bsi = gsi_start_phis (olddest); !gsi_end_p (bsi); gsi_next (&bsi))
- {
- phi = gsi_stmt (bsi);
- VEC_reserve (tree, heap, phis, 2);
- VEC_reserve (source_location, heap, locations, 1);
- VEC_quick_push (tree, phis, PHI_RESULT (phi));
- VEC_quick_push (tree, phis, PHI_ARG_DEF (phi, 0));
- VEC_quick_push (source_location, locations,
- gimple_phi_arg_location (phi, 0));
- }
- e = redirect_edge_and_branch (single_succ_edge (preheaderbb), headerbb);
-
- /* Remove the exit phis from the old basic block. */
- for (bsi = gsi_start_phis (olddest); !gsi_end_p (bsi); )
- remove_phi_node (&bsi, false);
-
- /* and add them back to the new basic block. */
- while (VEC_length (tree, phis) != 0)
- {
- tree def;
- tree phiname;
- source_location locus;
- def = VEC_pop (tree, phis);
- phiname = VEC_pop (tree, phis);
- locus = VEC_pop (source_location, locations);
- phi = create_phi_node (phiname, preheaderbb);
- add_phi_arg (phi, def, single_pred_edge (preheaderbb), locus);
- }
- flush_pending_stmts (e);
- VEC_free (tree, heap, phis);
-
- bodybb = create_empty_bb (EXIT_BLOCK_PTR->prev_bb);
- latchbb = create_empty_bb (EXIT_BLOCK_PTR->prev_bb);
- make_edge (headerbb, bodybb, EDGE_FALLTHRU);
- cond_stmt = gimple_build_cond (NE_EXPR, integer_one_node, integer_zero_node,
- NULL_TREE, NULL_TREE);
- bsi = gsi_start_bb (bodybb);
- gsi_insert_after (&bsi, cond_stmt, GSI_NEW_STMT);
- e = make_edge (bodybb, olddest, EDGE_FALSE_VALUE);
- make_edge (bodybb, latchbb, EDGE_TRUE_VALUE);
- make_edge (latchbb, headerbb, EDGE_FALLTHRU);
-
- /* Update the loop structures. */
- newloop = duplicate_loop (loop, olddest->loop_father);
- newloop->header = headerbb;
- newloop->latch = latchbb;
- add_bb_to_loop (latchbb, newloop);
- add_bb_to_loop (bodybb, newloop);
- add_bb_to_loop (headerbb, newloop);
- set_immediate_dominator (CDI_DOMINATORS, bodybb, headerbb);
- set_immediate_dominator (CDI_DOMINATORS, headerbb, preheaderbb);
- set_immediate_dominator (CDI_DOMINATORS, preheaderbb,
- single_exit (loop)->src);
- set_immediate_dominator (CDI_DOMINATORS, latchbb, bodybb);
- set_immediate_dominator (CDI_DOMINATORS, olddest,
- recompute_dominator (CDI_DOMINATORS, olddest));
- /* Create the new iv. */
- oldivvar = VEC_index (tree, loopivs, 0);
- ivvar = create_tmp_var (TREE_TYPE (oldivvar), "perfectiv");
- add_referenced_var (ivvar);
- standard_iv_increment_position (newloop, &bsi, &insert_after);
- create_iv (VEC_index (tree, lbounds, 0),
- build_int_cst (TREE_TYPE (oldivvar), VEC_index (int, steps, 0)),
- ivvar, newloop, &bsi, insert_after, &ivvar, &ivvarinced);
-
- /* Create the new upper bound. This may be not just a variable, so we copy
- it to one just in case. */
-
- exit_condition = get_loop_exit_condition (newloop);
- uboundvar = create_tmp_var (TREE_TYPE (VEC_index (tree, ubounds, 0)),
- "uboundvar");
- add_referenced_var (uboundvar);
- stmt = gimple_build_assign (uboundvar, VEC_index (tree, ubounds, 0));
- uboundvar = make_ssa_name (uboundvar, stmt);
- gimple_assign_set_lhs (stmt, uboundvar);
-
- if (insert_after)
- gsi_insert_after (&bsi, stmt, GSI_SAME_STMT);
- else
- gsi_insert_before (&bsi, stmt, GSI_SAME_STMT);
- update_stmt (stmt);
- gimple_cond_set_condition (exit_condition, GE_EXPR, uboundvar, ivvarinced);
- update_stmt (exit_condition);
- replacements = htab_create_ggc (20, tree_map_hash,
- tree_map_eq, NULL);
- bbs = get_loop_body_in_dom_order (loop);
- /* Now move the statements, and replace the induction variable in the moved
- statements with the correct loop induction variable. */
- oldivvar = VEC_index (tree, loopivs, 0);
- firstbsi = gsi_start_bb (bodybb);
- for (i = loop->num_nodes - 1; i >= 0 ; i--)
- {
- gimple_stmt_iterator tobsi = gsi_last_bb (bodybb);
- if (bbs[i]->loop_father == loop)
- {
- /* If this is true, we are *before* the inner loop.
- If this isn't true, we are *after* it.
-
- The only time can_convert_to_perfect_nest returns true when we
- have statements before the inner loop is if they can be moved
- into the inner loop.
-
- The only time can_convert_to_perfect_nest returns true when we
- have statements after the inner loop is if they can be moved into
- the new split loop. */
-
- if (dominated_by_p (CDI_DOMINATORS, loop->inner->header, bbs[i]))
- {
- gimple_stmt_iterator header_bsi
- = gsi_after_labels (loop->inner->header);
-
- for (bsi = gsi_start_bb (bbs[i]); !gsi_end_p (bsi);)
- {
- gimple stmt = gsi_stmt (bsi);
-
- if (stmt == exit_condition
- || not_interesting_stmt (stmt)
- || stmt_is_bumper_for_loop (loop, stmt))
- {
- gsi_next (&bsi);
- continue;
- }
-
- gsi_move_before (&bsi, &header_bsi);
- }
- }
- else
- {
- /* Note that the bsi only needs to be explicitly incremented
- when we don't move something, since it is automatically
- incremented when we do. */
- for (bsi = gsi_start_bb (bbs[i]); !gsi_end_p (bsi);)
- {
- gimple stmt = gsi_stmt (bsi);
-
- if (stmt == exit_condition
- || not_interesting_stmt (stmt)
- || stmt_is_bumper_for_loop (loop, stmt))
- {
- gsi_next (&bsi);
- continue;
- }
-
- replace_uses_equiv_to_x_with_y
- (loop, stmt, oldivvar, VEC_index (int, steps, 0), ivvar,
- VEC_index (tree, lbounds, 0), replacements, &firstbsi);
-
- gsi_move_before (&bsi, &tobsi);
-
- /* If the statement has any virtual operands, they may
- need to be rewired because the original loop may
- still reference them. */
- if (gimple_vuse (stmt))
- mark_sym_for_renaming (gimple_vop (cfun));
- }
- }
-
- }
- }
-
- free (bbs);
- htab_delete (replacements);
- return perfect_nest_p (loop);
-}
-
-/* Return true if TRANS is a legal transformation matrix that respects
- the dependence vectors in DISTS and DIRS. The conservative answer
- is false.
-
- "Wolfe proves that a unimodular transformation represented by the
- matrix T is legal when applied to a loop nest with a set of
- lexicographically non-negative distance vectors RDG if and only if
- for each vector d in RDG, (T.d >= 0) is lexicographically positive.
- i.e.: if and only if it transforms the lexicographically positive
- distance vectors to lexicographically positive vectors. Note that
- a unimodular matrix must transform the zero vector (and only it) to
- the zero vector." S.Muchnick. */
-
-bool
-lambda_transform_legal_p (lambda_trans_matrix trans,
- int nb_loops,
- VEC (ddr_p, heap) *dependence_relations)
-{
- unsigned int i, j;
- lambda_vector distres;
- struct data_dependence_relation *ddr;
-
- gcc_assert (LTM_COLSIZE (trans) == nb_loops
- && LTM_ROWSIZE (trans) == nb_loops);
-
- /* When there are no dependences, the transformation is correct. */
- if (VEC_length (ddr_p, dependence_relations) == 0)
- return true;
-
- ddr = VEC_index (ddr_p, dependence_relations, 0);
- if (ddr == NULL)
- return true;
-
- /* When there is an unknown relation in the dependence_relations, we
- know that it is no worth looking at this loop nest: give up. */
- if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
- return false;
-
- distres = lambda_vector_new (nb_loops);
-
- /* For each distance vector in the dependence graph. */
- FOR_EACH_VEC_ELT (ddr_p, dependence_relations, i, ddr)
- {
- /* Don't care about relations for which we know that there is no
- dependence, nor about read-read (aka. output-dependences):
- these data accesses can happen in any order. */
- if (DDR_ARE_DEPENDENT (ddr) == chrec_known
- || (DR_IS_READ (DDR_A (ddr)) && DR_IS_READ (DDR_B (ddr))))
- continue;
-
- /* Conservatively answer: "this transformation is not valid". */
- if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
- return false;
-
- /* If the dependence could not be captured by a distance vector,
- conservatively answer that the transform is not valid. */
- if (DDR_NUM_DIST_VECTS (ddr) == 0)
- return false;
-
- /* Compute trans.dist_vect */
- for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++)
- {
- lambda_matrix_vector_mult (LTM_MATRIX (trans), nb_loops, nb_loops,
- DDR_DIST_VECT (ddr, j), distres);
-
- if (!lambda_vector_lexico_pos (distres, nb_loops))
- return false;
- }
- }
- return true;
-}
-
-
-/* Collects parameters from affine function ACCESS_FUNCTION, and push
- them in PARAMETERS. */
-
-static void
-lambda_collect_parameters_from_af (tree access_function,
- struct pointer_set_t *param_set,
- VEC (tree, heap) **parameters)
-{
- if (access_function == NULL)
- return;
-
- if (TREE_CODE (access_function) == SSA_NAME
- && pointer_set_contains (param_set, access_function) == 0)
- {
- pointer_set_insert (param_set, access_function);
- VEC_safe_push (tree, heap, *parameters, access_function);
- }
- else
- {
- int i, num_operands = tree_operand_length (access_function);
-
- for (i = 0; i < num_operands; i++)
- lambda_collect_parameters_from_af (TREE_OPERAND (access_function, i),
- param_set, parameters);
- }
-}
-
-/* Collects parameters from DATAREFS, and push them in PARAMETERS. */
-
-void
-lambda_collect_parameters (VEC (data_reference_p, heap) *datarefs,
- VEC (tree, heap) **parameters)
-{
- unsigned i, j;
- struct pointer_set_t *parameter_set = pointer_set_create ();
- data_reference_p data_reference;
-
- FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, data_reference)
- for (j = 0; j < DR_NUM_DIMENSIONS (data_reference); j++)
- lambda_collect_parameters_from_af (DR_ACCESS_FN (data_reference, j),
- parameter_set, parameters);
- pointer_set_destroy (parameter_set);
-}
-
-/* Translates BASE_EXPR to vector CY. AM is needed for inferring
- indexing positions in the data access vector. CST is the analyzed
- integer constant. */
-
-static bool
-av_for_af_base (tree base_expr, lambda_vector cy, struct access_matrix *am,
- int cst)
-{
- bool result = true;
-
- switch (TREE_CODE (base_expr))
- {
- case INTEGER_CST:
- /* Constant part. */
- cy[AM_CONST_COLUMN_INDEX (am)] += int_cst_value (base_expr) * cst;
- return true;
-
- case SSA_NAME:
- {
- int param_index =
- access_matrix_get_index_for_parameter (base_expr, am);
-
- if (param_index >= 0)
- {
- cy[param_index] = cst + cy[param_index];
- return true;
- }
-
- return false;
- }
-
- case PLUS_EXPR:
- return av_for_af_base (TREE_OPERAND (base_expr, 0), cy, am, cst)
- && av_for_af_base (TREE_OPERAND (base_expr, 1), cy, am, cst);
-
- case MINUS_EXPR:
- return av_for_af_base (TREE_OPERAND (base_expr, 0), cy, am, cst)
- && av_for_af_base (TREE_OPERAND (base_expr, 1), cy, am, -1 * cst);
-
- case MULT_EXPR:
- if (TREE_CODE (TREE_OPERAND (base_expr, 0)) == INTEGER_CST)
- result = av_for_af_base (TREE_OPERAND (base_expr, 1),
- cy, am, cst *
- int_cst_value (TREE_OPERAND (base_expr, 0)));
- else if (TREE_CODE (TREE_OPERAND (base_expr, 1)) == INTEGER_CST)
- result = av_for_af_base (TREE_OPERAND (base_expr, 0),
- cy, am, cst *
- int_cst_value (TREE_OPERAND (base_expr, 1)));
- else
- result = false;
-
- return result;
-
- case NEGATE_EXPR:
- return av_for_af_base (TREE_OPERAND (base_expr, 0), cy, am, -1 * cst);
-
- default:
- return false;
- }
-
- return result;
-}
-
-/* Translates ACCESS_FUN to vector CY. AM is needed for inferring
- indexing positions in the data access vector. */
-
-static bool
-av_for_af (tree access_fun, lambda_vector cy, struct access_matrix *am)
-{
- switch (TREE_CODE (access_fun))
- {
- case POLYNOMIAL_CHREC:
- {
- tree left = CHREC_LEFT (access_fun);
- tree right = CHREC_RIGHT (access_fun);
- unsigned var;
-
- if (TREE_CODE (right) != INTEGER_CST)
- return false;
-
- var = am_vector_index_for_loop (am, CHREC_VARIABLE (access_fun));
- cy[var] = int_cst_value (right);
-
- if (TREE_CODE (left) == POLYNOMIAL_CHREC)
- return av_for_af (left, cy, am);
- else
- return av_for_af_base (left, cy, am, 1);
- }
-
- case INTEGER_CST:
- /* Constant part. */
- return av_for_af_base (access_fun, cy, am, 1);
-
- default:
- return false;
- }
-}
-
-/* Initializes the access matrix for DATA_REFERENCE. */
-
-static bool
-build_access_matrix (data_reference_p data_reference,
- VEC (tree, heap) *parameters,
- VEC (loop_p, heap) *nest,
- struct obstack * lambda_obstack)
-{
- struct access_matrix *am = (struct access_matrix *)
- obstack_alloc(lambda_obstack, sizeof (struct access_matrix));
- unsigned i, ndim = DR_NUM_DIMENSIONS (data_reference);
- unsigned nivs = VEC_length (loop_p, nest);
- unsigned lambda_nb_columns;
-
- AM_LOOP_NEST (am) = nest;
- AM_NB_INDUCTION_VARS (am) = nivs;
- AM_PARAMETERS (am) = parameters;
-
- lambda_nb_columns = AM_NB_COLUMNS (am);
- AM_MATRIX (am) = VEC_alloc (lambda_vector, gc, ndim);
-
- for (i = 0; i < ndim; i++)
- {
- lambda_vector access_vector = lambda_vector_new (lambda_nb_columns);
- tree access_function = DR_ACCESS_FN (data_reference, i);
-
- if (!av_for_af (access_function, access_vector, am))
- return false;
-
- VEC_quick_push (lambda_vector, AM_MATRIX (am), access_vector);
- }
-
- DR_ACCESS_MATRIX (data_reference) = am;
- return true;
-}
-
-/* Returns false when one of the access matrices cannot be built. */
-
-bool
-lambda_compute_access_matrices (VEC (data_reference_p, heap) *datarefs,
- VEC (tree, heap) *parameters,
- VEC (loop_p, heap) *nest,
- struct obstack * lambda_obstack)
-{
- data_reference_p dataref;
- unsigned ix;
-
- FOR_EACH_VEC_ELT (data_reference_p, datarefs, ix, dataref)
- if (!build_access_matrix (dataref, parameters, nest, lambda_obstack))
- return false;
-
- return true;
-}
+++ /dev/null
-/* Integer matrix math routines
- Copyright (C) 2003, 2004, 2005, 2007, 2008, 2010
- Free Software Foundation, Inc.
- Contributed by Daniel Berlin <dberlin@dberlin.org>.
-
-This file is part of GCC.
-
-GCC is free software; you can redistribute it and/or modify it under
-the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 3, or (at your option) any later
-version.
-
-GCC is distributed in the hope that it will be useful, but WITHOUT ANY
-WARRANTY; without even the implied warranty of MERCHANTABILITY or
-FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
-for more details.
-
-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/>. */
-
-#include "config.h"
-#include "system.h"
-#include "coretypes.h"
-#include "tree-flow.h"
-#include "lambda.h"
-
-/* Allocate a matrix of M rows x N cols. */
-
-lambda_matrix
-lambda_matrix_new (int m, int n, struct obstack * lambda_obstack)
-{
- lambda_matrix mat;
- int i;
-
- mat = (lambda_matrix) obstack_alloc (lambda_obstack,
- sizeof (lambda_vector *) * m);
-
- for (i = 0; i < m; i++)
- mat[i] = lambda_vector_new (n);
-
- return mat;
-}
-
-/* Copy the elements of M x N matrix MAT1 to MAT2. */
-
-void
-lambda_matrix_copy (lambda_matrix mat1, lambda_matrix mat2,
- int m, int n)
-{
- int i;
-
- for (i = 0; i < m; i++)
- lambda_vector_copy (mat1[i], mat2[i], n);
-}
-
-/* Store the N x N identity matrix in MAT. */
-
-void
-lambda_matrix_id (lambda_matrix mat, int size)
-{
- int i, j;
-
- for (i = 0; i < size; i++)
- for (j = 0; j < size; j++)
- mat[i][j] = (i == j) ? 1 : 0;
-}
-
-/* Return true if MAT is the identity matrix of SIZE */
-
-bool
-lambda_matrix_id_p (lambda_matrix mat, int size)
-{
- int i, j;
- for (i = 0; i < size; i++)
- for (j = 0; j < size; j++)
- {
- if (i == j)
- {
- if (mat[i][j] != 1)
- return false;
- }
- else
- {
- if (mat[i][j] != 0)
- return false;
- }
- }
- return true;
-}
-
-/* Negate the elements of the M x N matrix MAT1 and store it in MAT2. */
-
-void
-lambda_matrix_negate (lambda_matrix mat1, lambda_matrix mat2, int m, int n)
-{
- int i;
-
- for (i = 0; i < m; i++)
- lambda_vector_negate (mat1[i], mat2[i], n);
-}
-
-/* Take the transpose of matrix MAT1 and store it in MAT2.
- MAT1 is an M x N matrix, so MAT2 must be N x M. */
-
-void
-lambda_matrix_transpose (lambda_matrix mat1, lambda_matrix mat2, int m, int n)
-{
- int i, j;
-
- for (i = 0; i < n; i++)
- for (j = 0; j < m; j++)
- mat2[i][j] = mat1[j][i];
-}
-
-
-/* Add two M x N matrices together: MAT3 = MAT1+MAT2. */
-
-void
-lambda_matrix_add (lambda_matrix mat1, lambda_matrix mat2,
- lambda_matrix mat3, int m, int n)
-{
- int i;
-
- for (i = 0; i < m; i++)
- lambda_vector_add (mat1[i], mat2[i], mat3[i], n);
-}
-
-/* MAT3 = CONST1 * MAT1 + CONST2 * MAT2. All matrices are M x N. */
-
-void
-lambda_matrix_add_mc (lambda_matrix mat1, int const1,
- lambda_matrix mat2, int const2,
- lambda_matrix mat3, int m, int n)
-{
- int i;
-
- for (i = 0; i < m; i++)
- lambda_vector_add_mc (mat1[i], const1, mat2[i], const2, mat3[i], n);
-}
-
-/* Multiply two matrices: MAT3 = MAT1 * MAT2.
- MAT1 is an M x R matrix, and MAT2 is R x N. The resulting MAT2
- must therefore be M x N. */
-
-void
-lambda_matrix_mult (lambda_matrix mat1, lambda_matrix mat2,
- lambda_matrix mat3, int m, int r, int n)
-{
-
- int i, j, k;
-
- for (i = 0; i < m; i++)
- {
- for (j = 0; j < n; j++)
- {
- mat3[i][j] = 0;
- for (k = 0; k < r; k++)
- mat3[i][j] += mat1[i][k] * mat2[k][j];
- }
- }
-}
-
-/* Delete rows r1 to r2 (not including r2). */
-
-void
-lambda_matrix_delete_rows (lambda_matrix mat, int rows, int from, int to)
-{
- int i;
- int dist;
- dist = to - from;
-
- for (i = to; i < rows; i++)
- mat[i - dist] = mat[i];
-
- for (i = rows - dist; i < rows; i++)
- mat[i] = NULL;
-}
-
-/* Swap rows R1 and R2 in matrix MAT. */
-
-void
-lambda_matrix_row_exchange (lambda_matrix mat, int r1, int r2)
-{
- lambda_vector row;
-
- row = mat[r1];
- mat[r1] = mat[r2];
- mat[r2] = row;
-}
-
-/* Add a multiple of row R1 of matrix MAT with N columns to row R2:
- R2 = R2 + CONST1 * R1. */
-
-void
-lambda_matrix_row_add (lambda_matrix mat, int n, int r1, int r2, int const1)
-{
- int i;
-
- if (const1 == 0)
- return;
-
- for (i = 0; i < n; i++)
- mat[r2][i] += const1 * mat[r1][i];
-}
-
-/* Negate row R1 of matrix MAT which has N columns. */
-
-void
-lambda_matrix_row_negate (lambda_matrix mat, int n, int r1)
-{
- lambda_vector_negate (mat[r1], mat[r1], n);
-}
-
-/* Multiply row R1 of matrix MAT with N columns by CONST1. */
-
-void
-lambda_matrix_row_mc (lambda_matrix mat, int n, int r1, int const1)
-{
- int i;
-
- for (i = 0; i < n; i++)
- mat[r1][i] *= const1;
-}
-
-/* Exchange COL1 and COL2 in matrix MAT. M is the number of rows. */
-
-void
-lambda_matrix_col_exchange (lambda_matrix mat, int m, int col1, int col2)
-{
- int i;
- int tmp;
- for (i = 0; i < m; i++)
- {
- tmp = mat[i][col1];
- mat[i][col1] = mat[i][col2];
- mat[i][col2] = tmp;
- }
-}
-
-/* Add a multiple of column C1 of matrix MAT with M rows to column C2:
- C2 = C2 + CONST1 * C1. */
-
-void
-lambda_matrix_col_add (lambda_matrix mat, int m, int c1, int c2, int const1)
-{
- int i;
-
- if (const1 == 0)
- return;
-
- for (i = 0; i < m; i++)
- mat[i][c2] += const1 * mat[i][c1];
-}
-
-/* Negate column C1 of matrix MAT which has M rows. */
-
-void
-lambda_matrix_col_negate (lambda_matrix mat, int m, int c1)
-{
- int i;
-
- for (i = 0; i < m; i++)
- mat[i][c1] *= -1;
-}
-
-/* Multiply column C1 of matrix MAT with M rows by CONST1. */
-
-void
-lambda_matrix_col_mc (lambda_matrix mat, int m, int c1, int const1)
-{
- int i;
-
- for (i = 0; i < m; i++)
- mat[i][c1] *= const1;
-}
-
-/* Compute the inverse of the N x N matrix MAT and store it in INV.
-
- We don't _really_ compute the inverse of MAT. Instead we compute
- det(MAT)*inv(MAT), and we return det(MAT) to the caller as the function
- result. This is necessary to preserve accuracy, because we are dealing
- with integer matrices here.
-
- The algorithm used here is a column based Gauss-Jordan elimination on MAT
- and the identity matrix in parallel. The inverse is the result of applying
- the same operations on the identity matrix that reduce MAT to the identity
- matrix.
-
- When MAT is a 2 x 2 matrix, we don't go through the whole process, because
- it is easily inverted by inspection and it is a very common case. */
-
-static int lambda_matrix_inverse_hard (lambda_matrix, lambda_matrix, int,
- struct obstack *);
-
-int
-lambda_matrix_inverse (lambda_matrix mat, lambda_matrix inv, int n,
- struct obstack * lambda_obstack)
-{
- if (n == 2)
- {
- int a, b, c, d, det;
- a = mat[0][0];
- b = mat[1][0];
- c = mat[0][1];
- d = mat[1][1];
- inv[0][0] = d;
- inv[0][1] = -c;
- inv[1][0] = -b;
- inv[1][1] = a;
- det = (a * d - b * c);
- if (det < 0)
- {
- det *= -1;
- inv[0][0] *= -1;
- inv[1][0] *= -1;
- inv[0][1] *= -1;
- inv[1][1] *= -1;
- }
- return det;
- }
- else
- return lambda_matrix_inverse_hard (mat, inv, n, lambda_obstack);
-}
-
-/* If MAT is not a special case, invert it the hard way. */
-
-static int
-lambda_matrix_inverse_hard (lambda_matrix mat, lambda_matrix inv, int n,
- struct obstack * lambda_obstack)
-{
- lambda_vector row;
- lambda_matrix temp;
- int i, j;
- int determinant;
-
- temp = lambda_matrix_new (n, n, lambda_obstack);
- lambda_matrix_copy (mat, temp, n, n);
- lambda_matrix_id (inv, n);
-
- /* Reduce TEMP to a lower triangular form, applying the same operations on
- INV which starts as the identity matrix. N is the number of rows and
- columns. */
- for (j = 0; j < n; j++)
- {
- row = temp[j];
-
- /* Make every element in the current row positive. */
- for (i = j; i < n; i++)
- if (row[i] < 0)
- {
- lambda_matrix_col_negate (temp, n, i);
- lambda_matrix_col_negate (inv, n, i);
- }
-
- /* Sweep the upper triangle. Stop when only the diagonal element in the
- current row is nonzero. */
- while (lambda_vector_first_nz (row, n, j + 1) < n)
- {
- int min_col = lambda_vector_min_nz (row, n, j);
- lambda_matrix_col_exchange (temp, n, j, min_col);
- lambda_matrix_col_exchange (inv, n, j, min_col);
-
- for (i = j + 1; i < n; i++)
- {
- int factor;
-
- factor = -1 * row[i];
- if (row[j] != 1)
- factor /= row[j];
-
- lambda_matrix_col_add (temp, n, j, i, factor);
- lambda_matrix_col_add (inv, n, j, i, factor);
- }
- }
- }
-
- /* Reduce TEMP from a lower triangular to the identity matrix. Also compute
- the determinant, which now is simply the product of the elements on the
- diagonal of TEMP. If one of these elements is 0, the matrix has 0 as an
- eigenvalue so it is singular and hence not invertible. */
- determinant = 1;
- for (j = n - 1; j >= 0; j--)
- {
- int diagonal;
-
- row = temp[j];
- diagonal = row[j];
-
- /* The matrix must not be singular. */
- gcc_assert (diagonal);
-
- determinant = determinant * diagonal;
-
- /* If the diagonal is not 1, then multiply the each row by the
- diagonal so that the middle number is now 1, rather than a
- rational. */
- if (diagonal != 1)
- {
- for (i = 0; i < j; i++)
- lambda_matrix_col_mc (inv, n, i, diagonal);
- for (i = j + 1; i < n; i++)
- lambda_matrix_col_mc (inv, n, i, diagonal);
-
- row[j] = diagonal = 1;
- }
-
- /* Sweep the lower triangle column wise. */
- for (i = j - 1; i >= 0; i--)
- {
- if (row[i])
- {
- int factor = -row[i];
- lambda_matrix_col_add (temp, n, j, i, factor);
- lambda_matrix_col_add (inv, n, j, i, factor);
- }
-
- }
- }
-
- return determinant;
-}
-
-/* Decompose a N x N matrix MAT to a product of a lower triangular H
- and a unimodular U matrix such that MAT = H.U. N is the size of
- the rows of MAT. */
-
-void
-lambda_matrix_hermite (lambda_matrix mat, int n,
- lambda_matrix H, lambda_matrix U)
-{
- lambda_vector row;
- int i, j, factor, minimum_col;
-
- lambda_matrix_copy (mat, H, n, n);
- lambda_matrix_id (U, n);
-
- for (j = 0; j < n; j++)
- {
- row = H[j];
-
- /* Make every element of H[j][j..n] positive. */
- for (i = j; i < n; i++)
- {
- if (row[i] < 0)
- {
- lambda_matrix_col_negate (H, n, i);
- lambda_vector_negate (U[i], U[i], n);
- }
- }
-
- /* Stop when only the diagonal element is nonzero. */
- while (lambda_vector_first_nz (row, n, j + 1) < n)
- {
- minimum_col = lambda_vector_min_nz (row, n, j);
- lambda_matrix_col_exchange (H, n, j, minimum_col);
- lambda_matrix_row_exchange (U, j, minimum_col);
-
- for (i = j + 1; i < n; i++)
- {
- factor = row[i] / row[j];
- lambda_matrix_col_add (H, n, j, i, -1 * factor);
- lambda_matrix_row_add (U, n, i, j, factor);
- }
- }
- }
-}
-
-/* Given an M x N integer matrix A, this function determines an M x
- M unimodular matrix U, and an M x N echelon matrix S such that
- "U.A = S". This decomposition is also known as "right Hermite".
-
- Ref: Algorithm 2.1 page 33 in "Loop Transformations for
- Restructuring Compilers" Utpal Banerjee. */
-
-void
-lambda_matrix_right_hermite (lambda_matrix A, int m, int n,
- lambda_matrix S, lambda_matrix U)
-{
- int i, j, i0 = 0;
-
- lambda_matrix_copy (A, S, m, n);
- lambda_matrix_id (U, m);
-
- for (j = 0; j < n; j++)
- {
- if (lambda_vector_first_nz (S[j], m, i0) < m)
- {
- ++i0;
- for (i = m - 1; i >= i0; i--)
- {
- while (S[i][j] != 0)
- {
- int sigma, factor, a, b;
-
- a = S[i-1][j];
- b = S[i][j];
- sigma = (a * b < 0) ? -1: 1;
- a = abs (a);
- b = abs (b);
- factor = sigma * (a / b);
-
- lambda_matrix_row_add (S, n, i, i-1, -factor);
- lambda_matrix_row_exchange (S, i, i-1);
-
- lambda_matrix_row_add (U, m, i, i-1, -factor);
- lambda_matrix_row_exchange (U, i, i-1);
- }
- }
- }
- }
-}
-
-/* Given an M x N integer matrix A, this function determines an M x M
- unimodular matrix V, and an M x N echelon matrix S such that "A =
- V.S". This decomposition is also known as "left Hermite".
-
- Ref: Algorithm 2.2 page 36 in "Loop Transformations for
- Restructuring Compilers" Utpal Banerjee. */
-
-void
-lambda_matrix_left_hermite (lambda_matrix A, int m, int n,
- lambda_matrix S, lambda_matrix V)
-{
- int i, j, i0 = 0;
-
- lambda_matrix_copy (A, S, m, n);
- lambda_matrix_id (V, m);
-
- for (j = 0; j < n; j++)
- {
- if (lambda_vector_first_nz (S[j], m, i0) < m)
- {
- ++i0;
- for (i = m - 1; i >= i0; i--)
- {
- while (S[i][j] != 0)
- {
- int sigma, factor, a, b;
-
- a = S[i-1][j];
- b = S[i][j];
- sigma = (a * b < 0) ? -1: 1;
- a = abs (a);
- b = abs (b);
- factor = sigma * (a / b);
-
- lambda_matrix_row_add (S, n, i, i-1, -factor);
- lambda_matrix_row_exchange (S, i, i-1);
-
- lambda_matrix_col_add (V, m, i-1, i, factor);
- lambda_matrix_col_exchange (V, m, i, i-1);
- }
- }
- }
- }
-}
-
-/* When it exists, return the first nonzero row in MAT after row
- STARTROW. Otherwise return rowsize. */
-
-int
-lambda_matrix_first_nz_vec (lambda_matrix mat, int rowsize, int colsize,
- int startrow)
-{
- int j;
- bool found = false;
-
- for (j = startrow; (j < rowsize) && !found; j++)
- {
- if ((mat[j] != NULL)
- && (lambda_vector_first_nz (mat[j], colsize, startrow) < colsize))
- found = true;
- }
-
- if (found)
- return j - 1;
- return rowsize;
-}
-
-/* Multiply a vector VEC by a matrix MAT.
- MAT is an M*N matrix, and VEC is a vector with length N. The result
- is stored in DEST which must be a vector of length M. */
-
-void
-lambda_matrix_vector_mult (lambda_matrix matrix, int m, int n,
- lambda_vector vec, lambda_vector dest)
-{
- int i, j;
-
- lambda_vector_clear (dest, m);
- for (i = 0; i < m; i++)
- for (j = 0; j < n; j++)
- dest[i] += matrix[i][j] * vec[j];
-}
-
-/* Print out an M x N matrix MAT to OUTFILE. */
-
-void
-print_lambda_matrix (FILE * outfile, lambda_matrix matrix, int m, int n)
-{
- int i;
-
- for (i = 0; i < m; i++)
- print_lambda_vector (outfile, matrix[i], n);
- fprintf (outfile, "\n");
-}
-
+++ /dev/null
-/* Lambda matrix transformations.
- Copyright (C) 2003, 2004, 2007, 2008, 2010 Free Software Foundation, Inc.
- Contributed by Daniel Berlin <dberlin@dberlin.org>.
-
-This file is part of GCC.
-
-GCC is free software; you can redistribute it and/or modify it under
-the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 3, or (at your option) any later
-version.
-
-GCC is distributed in the hope that it will be useful, but WITHOUT ANY
-WARRANTY; without even the implied warranty of MERCHANTABILITY or
-FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
-for more details.
-
-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/>. */
-
-#include "config.h"
-#include "system.h"
-#include "coretypes.h"
-#include "tree-flow.h"
-#include "lambda.h"
-
-/* Allocate a new transformation matrix. */
-
-lambda_trans_matrix
-lambda_trans_matrix_new (int colsize, int rowsize,
- struct obstack * lambda_obstack)
-{
- lambda_trans_matrix ret;
-
- ret = (lambda_trans_matrix)
- obstack_alloc (lambda_obstack, sizeof (struct lambda_trans_matrix_s));
- LTM_MATRIX (ret) = lambda_matrix_new (rowsize, colsize, lambda_obstack);
- LTM_ROWSIZE (ret) = rowsize;
- LTM_COLSIZE (ret) = colsize;
- LTM_DENOMINATOR (ret) = 1;
- return ret;
-}
-
-/* Return true if MAT is an identity matrix. */
-
-bool
-lambda_trans_matrix_id_p (lambda_trans_matrix mat)
-{
- if (LTM_ROWSIZE (mat) != LTM_COLSIZE (mat))
- return false;
- return lambda_matrix_id_p (LTM_MATRIX (mat), LTM_ROWSIZE (mat));
-}
-
-
-/* Compute the inverse of the transformation matrix MAT. */
-
-lambda_trans_matrix
-lambda_trans_matrix_inverse (lambda_trans_matrix mat,
- struct obstack * lambda_obstack)
-{
- lambda_trans_matrix inverse;
- int determinant;
-
- inverse = lambda_trans_matrix_new (LTM_ROWSIZE (mat), LTM_COLSIZE (mat),
- lambda_obstack);
- determinant = lambda_matrix_inverse (LTM_MATRIX (mat), LTM_MATRIX (inverse),
- LTM_ROWSIZE (mat), lambda_obstack);
- LTM_DENOMINATOR (inverse) = determinant;
- return inverse;
-}
-
-
-/* Print out a transformation matrix. */
-
-void
-print_lambda_trans_matrix (FILE *outfile, lambda_trans_matrix mat)
-{
- print_lambda_matrix (outfile, LTM_MATRIX (mat), LTM_ROWSIZE (mat),
- LTM_COLSIZE (mat));
-}
+++ /dev/null
-/* Lambda matrix and vector interface.
- Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
- Free Software Foundation, Inc.
- Contributed by Daniel Berlin <dberlin@dberlin.org>
-
-This file is part of GCC.
-
-GCC is free software; you can redistribute it and/or modify it under
-the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 3, or (at your option) any later
-version.
-
-GCC is distributed in the hope that it will be useful, but WITHOUT ANY
-WARRANTY; without even the implied warranty of MERCHANTABILITY or
-FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
-for more details.
-
-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/>. */
-
-#ifndef LAMBDA_H
-#define LAMBDA_H
-
-#include "vec.h"
-
-/* An integer vector. A vector formally consists of an element of a vector
- space. A vector space is a set that is closed under vector addition
- and scalar multiplication. In this vector space, an element is a list of
- integers. */
-typedef int *lambda_vector;
-DEF_VEC_P(lambda_vector);
-DEF_VEC_ALLOC_P(lambda_vector,heap);
-DEF_VEC_ALLOC_P(lambda_vector,gc);
-
-typedef VEC(lambda_vector, heap) *lambda_vector_vec_p;
-DEF_VEC_P (lambda_vector_vec_p);
-DEF_VEC_ALLOC_P (lambda_vector_vec_p, heap);
-
-/* An integer matrix. A matrix consists of m vectors of length n (IE
- all vectors are the same length). */
-typedef lambda_vector *lambda_matrix;
-
-DEF_VEC_P (lambda_matrix);
-DEF_VEC_ALLOC_P (lambda_matrix, heap);
-
-/* A transformation matrix, which is a self-contained ROWSIZE x COLSIZE
- matrix. Rather than use floats, we simply keep a single DENOMINATOR that
- represents the denominator for every element in the matrix. */
-typedef struct lambda_trans_matrix_s
-{
- lambda_matrix matrix;
- int rowsize;
- int colsize;
- int denominator;
-} *lambda_trans_matrix;
-#define LTM_MATRIX(T) ((T)->matrix)
-#define LTM_ROWSIZE(T) ((T)->rowsize)
-#define LTM_COLSIZE(T) ((T)->colsize)
-#define LTM_DENOMINATOR(T) ((T)->denominator)
-
-/* A vector representing a statement in the body of a loop.
- The COEFFICIENTS vector contains a coefficient for each induction variable
- in the loop nest containing the statement.
- The DENOMINATOR represents the denominator for each coefficient in the
- COEFFICIENT vector.
-
- This structure is used during code generation in order to rewrite the old
- induction variable uses in a statement in terms of the newly created
- induction variables. */
-typedef struct lambda_body_vector_s
-{
- lambda_vector coefficients;
- int size;
- int denominator;
-} *lambda_body_vector;
-#define LBV_COEFFICIENTS(T) ((T)->coefficients)
-#define LBV_SIZE(T) ((T)->size)
-#define LBV_DENOMINATOR(T) ((T)->denominator)
-
-/* Piecewise linear expression.
- This structure represents a linear expression with terms for the invariants
- and induction variables of a loop.
- COEFFICIENTS is a vector of coefficients for the induction variables, one
- per loop in the loop nest.
- CONSTANT is the constant portion of the linear expression
- INVARIANT_COEFFICIENTS is a vector of coefficients for the loop invariants,
- one per invariant.
- DENOMINATOR is the denominator for all of the coefficients and constants in
- the expression.
- The linear expressions can be linked together using the NEXT field, in
- order to represent MAX or MIN of a group of linear expressions. */
-typedef struct lambda_linear_expression_s
-{
- lambda_vector coefficients;
- int constant;
- lambda_vector invariant_coefficients;
- int denominator;
- struct lambda_linear_expression_s *next;
-} *lambda_linear_expression;
-
-#define LLE_COEFFICIENTS(T) ((T)->coefficients)
-#define LLE_CONSTANT(T) ((T)->constant)
-#define LLE_INVARIANT_COEFFICIENTS(T) ((T)->invariant_coefficients)
-#define LLE_DENOMINATOR(T) ((T)->denominator)
-#define LLE_NEXT(T) ((T)->next)
-
-struct obstack;
-
-lambda_linear_expression lambda_linear_expression_new (int, int,
- struct obstack *);
-void print_lambda_linear_expression (FILE *, lambda_linear_expression, int,
- int, char);
-
-/* Loop structure. Our loop structure consists of a constant representing the
- STEP of the loop, a set of linear expressions representing the LOWER_BOUND
- of the loop, a set of linear expressions representing the UPPER_BOUND of
- the loop, and a set of linear expressions representing the LINEAR_OFFSET of
- the loop. The linear offset is a set of linear expressions that are
- applied to *both* the lower bound, and the upper bound. */
-typedef struct lambda_loop_s
-{
- lambda_linear_expression lower_bound;
- lambda_linear_expression upper_bound;
- lambda_linear_expression linear_offset;
- int step;
-} *lambda_loop;
-
-#define LL_LOWER_BOUND(T) ((T)->lower_bound)
-#define LL_UPPER_BOUND(T) ((T)->upper_bound)
-#define LL_LINEAR_OFFSET(T) ((T)->linear_offset)
-#define LL_STEP(T) ((T)->step)
-
-/* Loop nest structure.
- The loop nest structure consists of a set of loop structures (defined
- above) in LOOPS, along with an integer representing the DEPTH of the loop,
- and an integer representing the number of INVARIANTS in the loop. Both of
- these integers are used to size the associated coefficient vectors in the
- linear expression structures. */
-typedef struct lambda_loopnest_s
-{
- lambda_loop *loops;
- int depth;
- int invariants;
-} *lambda_loopnest;
-
-#define LN_LOOPS(T) ((T)->loops)
-#define LN_DEPTH(T) ((T)->depth)
-#define LN_INVARIANTS(T) ((T)->invariants)
-
-lambda_loopnest lambda_loopnest_new (int, int, struct obstack *);
-lambda_loopnest lambda_loopnest_transform (lambda_loopnest,
- lambda_trans_matrix,
- struct obstack *);
-struct loop;
-bool perfect_nest_p (struct loop *);
-void print_lambda_loopnest (FILE *, lambda_loopnest, char);
-
-void print_lambda_loop (FILE *, lambda_loop, int, int, char);
-
-lambda_matrix lambda_matrix_new (int, int, struct obstack *);
-
-void lambda_matrix_id (lambda_matrix, int);
-bool lambda_matrix_id_p (lambda_matrix, int);
-void lambda_matrix_copy (lambda_matrix, lambda_matrix, int, int);
-void lambda_matrix_negate (lambda_matrix, lambda_matrix, int, int);
-void lambda_matrix_transpose (lambda_matrix, lambda_matrix, int, int);
-void lambda_matrix_add (lambda_matrix, lambda_matrix, lambda_matrix, int,
- int);
-void lambda_matrix_add_mc (lambda_matrix, int, lambda_matrix, int,
- lambda_matrix, int, int);
-void lambda_matrix_mult (lambda_matrix, lambda_matrix, lambda_matrix,
- int, int, int);
-void lambda_matrix_delete_rows (lambda_matrix, int, int, int);
-void lambda_matrix_row_exchange (lambda_matrix, int, int);
-void lambda_matrix_row_add (lambda_matrix, int, int, int, int);
-void lambda_matrix_row_negate (lambda_matrix mat, int, int);
-void lambda_matrix_row_mc (lambda_matrix, int, int, int);
-void lambda_matrix_col_exchange (lambda_matrix, int, int, int);
-void lambda_matrix_col_add (lambda_matrix, int, int, int, int);
-void lambda_matrix_col_negate (lambda_matrix, int, int);
-void lambda_matrix_col_mc (lambda_matrix, int, int, int);
-int lambda_matrix_inverse (lambda_matrix, lambda_matrix, int, struct obstack *);
-void lambda_matrix_hermite (lambda_matrix, int, lambda_matrix, lambda_matrix);
-void lambda_matrix_left_hermite (lambda_matrix, int, int, lambda_matrix, lambda_matrix);
-void lambda_matrix_right_hermite (lambda_matrix, int, int, lambda_matrix, lambda_matrix);
-int lambda_matrix_first_nz_vec (lambda_matrix, int, int, int);
-void lambda_matrix_project_to_null (lambda_matrix, int, int, int,
- lambda_vector);
-void print_lambda_matrix (FILE *, lambda_matrix, int, int);
-
-lambda_trans_matrix lambda_trans_matrix_new (int, int, struct obstack *);
-bool lambda_trans_matrix_nonsingular_p (lambda_trans_matrix);
-bool lambda_trans_matrix_fullrank_p (lambda_trans_matrix);
-int lambda_trans_matrix_rank (lambda_trans_matrix);
-lambda_trans_matrix lambda_trans_matrix_basis (lambda_trans_matrix);
-lambda_trans_matrix lambda_trans_matrix_padding (lambda_trans_matrix);
-lambda_trans_matrix lambda_trans_matrix_inverse (lambda_trans_matrix,
- struct obstack *);
-void print_lambda_trans_matrix (FILE *, lambda_trans_matrix);
-void lambda_matrix_vector_mult (lambda_matrix, int, int, lambda_vector,
- lambda_vector);
-bool lambda_trans_matrix_id_p (lambda_trans_matrix);
-
-lambda_body_vector lambda_body_vector_new (int, struct obstack *);
-lambda_body_vector lambda_body_vector_compute_new (lambda_trans_matrix,
- lambda_body_vector,
- struct obstack *);
-void print_lambda_body_vector (FILE *, lambda_body_vector);
-lambda_loopnest gcc_loopnest_to_lambda_loopnest (struct loop *,
- VEC(tree,heap) **,
- VEC(tree,heap) **,
- struct obstack *);
-void lambda_loopnest_to_gcc_loopnest (struct loop *,
- VEC(tree,heap) *, VEC(tree,heap) *,
- VEC(gimple,heap) **,
- lambda_loopnest, lambda_trans_matrix,
- struct obstack *);
-void remove_iv (gimple);
-tree find_induction_var_from_exit_cond (struct loop *);
-
-static inline void lambda_vector_negate (lambda_vector, lambda_vector, int);
-static inline void lambda_vector_mult_const (lambda_vector, lambda_vector, int, int);
-static inline void lambda_vector_add (lambda_vector, lambda_vector,
- lambda_vector, int);
-static inline void lambda_vector_add_mc (lambda_vector, int, lambda_vector, int,
- lambda_vector, int);
-static inline void lambda_vector_copy (lambda_vector, lambda_vector, int);
-static inline bool lambda_vector_zerop (lambda_vector, int);
-static inline void lambda_vector_clear (lambda_vector, int);
-static inline bool lambda_vector_equal (lambda_vector, lambda_vector, int);
-static inline int lambda_vector_min_nz (lambda_vector, int, int);
-static inline int lambda_vector_first_nz (lambda_vector, int, int);
-static inline void print_lambda_vector (FILE *, lambda_vector, int);
-
-/* Allocate a new vector of given SIZE. */
-
-static inline lambda_vector
-lambda_vector_new (int size)
-{
- return (lambda_vector) ggc_alloc_cleared_atomic (sizeof (int) * size);
-}
-
-
-
-/* Multiply vector VEC1 of length SIZE by a constant CONST1,
- and store the result in VEC2. */
-
-static inline void
-lambda_vector_mult_const (lambda_vector vec1, lambda_vector vec2,
- int size, int const1)
-{
- int i;
-
- if (const1 == 0)
- lambda_vector_clear (vec2, size);
- else
- for (i = 0; i < size; i++)
- vec2[i] = const1 * vec1[i];
-}
-
-/* Negate vector VEC1 with length SIZE and store it in VEC2. */
-
-static inline void
-lambda_vector_negate (lambda_vector vec1, lambda_vector vec2,
- int size)
-{
- lambda_vector_mult_const (vec1, vec2, size, -1);
-}
-
-/* VEC3 = VEC1+VEC2, where all three the vectors are of length SIZE. */
-
-static inline void
-lambda_vector_add (lambda_vector vec1, lambda_vector vec2,
- lambda_vector vec3, int size)
-{
- int i;
- for (i = 0; i < size; i++)
- vec3[i] = vec1[i] + vec2[i];
-}
-
-/* VEC3 = CONSTANT1*VEC1 + CONSTANT2*VEC2. All vectors have length SIZE. */
-
-static inline void
-lambda_vector_add_mc (lambda_vector vec1, int const1,
- lambda_vector vec2, int const2,
- lambda_vector vec3, int size)
-{
- int i;
- for (i = 0; i < size; i++)
- vec3[i] = const1 * vec1[i] + const2 * vec2[i];
-}
-
-/* Copy the elements of vector VEC1 with length SIZE to VEC2. */
-
-static inline void
-lambda_vector_copy (lambda_vector vec1, lambda_vector vec2,
- int size)
-{
- memcpy (vec2, vec1, size * sizeof (*vec1));
-}
-
-/* Return true if vector VEC1 of length SIZE is the zero vector. */
-
-static inline bool
-lambda_vector_zerop (lambda_vector vec1, int size)
-{
- int i;
- for (i = 0; i < size; i++)
- if (vec1[i] != 0)
- return false;
- return true;
-}
-
-/* Clear out vector VEC1 of length SIZE. */
-
-static inline void
-lambda_vector_clear (lambda_vector vec1, int size)
-{
- memset (vec1, 0, size * sizeof (*vec1));
-}
-
-/* Return true if two vectors are equal. */
-
-static inline bool
-lambda_vector_equal (lambda_vector vec1, lambda_vector vec2, int size)
-{
- int i;
- for (i = 0; i < size; i++)
- if (vec1[i] != vec2[i])
- return false;
- return true;
-}
-
-/* Return the minimum nonzero element in vector VEC1 between START and N.
- We must have START <= N. */
-
-static inline int
-lambda_vector_min_nz (lambda_vector vec1, int n, int start)
-{
- int j;
- int min = -1;
-
- gcc_assert (start <= n);
- for (j = start; j < n; j++)
- {
- if (vec1[j])
- if (min < 0 || vec1[j] < vec1[min])
- min = j;
- }
- gcc_assert (min >= 0);
-
- return min;
-}
-
-/* Return the first nonzero element of vector VEC1 between START and N.
- We must have START <= N. Returns N if VEC1 is the zero vector. */
-
-static inline int
-lambda_vector_first_nz (lambda_vector vec1, int n, int start)
-{
- int j = start;
- while (j < n && vec1[j] == 0)
- j++;
- return j;
-}
-
-
-/* Multiply a vector by a matrix. */
-
-static inline void
-lambda_vector_matrix_mult (lambda_vector vect, int m, lambda_matrix mat,
- int n, lambda_vector dest)
-{
- int i, j;
- lambda_vector_clear (dest, n);
- for (i = 0; i < n; i++)
- for (j = 0; j < m; j++)
- dest[i] += mat[j][i] * vect[j];
-}
-
-/* Compare two vectors returning an integer less than, equal to, or
- greater than zero if the first argument is considered to be respectively
- less than, equal to, or greater than the second.
- We use the lexicographic order. */
-
-static inline int
-lambda_vector_compare (lambda_vector vec1, int length1, lambda_vector vec2,
- int length2)
-{
- int min_length;
- int i;
-
- if (length1 < length2)
- min_length = length1;
- else
- min_length = length2;
-
- for (i = 0; i < min_length; i++)
- if (vec1[i] < vec2[i])
- return -1;
- else if (vec1[i] > vec2[i])
- return 1;
- else
- continue;
-
- return length1 - length2;
-}
-
-/* Print out a vector VEC of length N to OUTFILE. */
-
-static inline void
-print_lambda_vector (FILE * outfile, lambda_vector vector, int n)
-{
- int i;
-
- for (i = 0; i < n; i++)
- fprintf (outfile, "%3d ", vector[i]);
- fprintf (outfile, "\n");
-}
-
-/* Compute the greatest common divisor of two numbers using
- Euclid's algorithm. */
-
-static inline int
-gcd (int a, int b)
-{
- int x, y, z;
-
- x = abs (a);
- y = abs (b);
-
- while (x > 0)
- {
- z = y % x;
- y = x;
- x = z;
- }
-
- return y;
-}
-
-/* Compute the greatest common divisor of a VECTOR of SIZE numbers. */
-
-static inline int
-lambda_vector_gcd (lambda_vector vector, int size)
-{
- int i;
- int gcd1 = 0;
-
- if (size > 0)
- {
- gcd1 = vector[0];
- for (i = 1; i < size; i++)
- gcd1 = gcd (gcd1, vector[i]);
- }
- return gcd1;
-}
-
-/* Returns true when the vector V is lexicographically positive, in
- other words, when the first nonzero element is positive. */
-
-static inline bool
-lambda_vector_lexico_pos (lambda_vector v,
- unsigned n)
-{
- unsigned i;
- for (i = 0; i < n; i++)
- {
- if (v[i] == 0)
- continue;
- if (v[i] < 0)
- return false;
- if (v[i] > 0)
- return true;
- }
- return true;
-}
-
-/* Given a vector of induction variables IVS, and a vector of
- coefficients COEFS, build a tree that is a linear combination of
- the induction variables. */
-
-static inline tree
-build_linear_expr (tree type, lambda_vector coefs, VEC (tree, heap) *ivs)
-{
- unsigned i;
- tree iv;
- tree expr = build_zero_cst (type);
-
- for (i = 0; VEC_iterate (tree, ivs, i, iv); i++)
- {
- int k = coefs[i];
-
- if (k == 1)
- expr = fold_build2 (PLUS_EXPR, type, expr, iv);
-
- else if (k != 0)
- expr = fold_build2 (PLUS_EXPR, type, expr,
- fold_build2 (MULT_EXPR, type, iv,
- build_int_cst (type, k)));
- }
-
- return expr;
-}
-
-/* Returns the dependence level for a vector DIST of size LENGTH.
- LEVEL = 0 means a lexicographic dependence, i.e. a dependence due
- to the sequence of statements, not carried by any loop. */
-
-
-static inline unsigned
-dependence_level (lambda_vector dist_vect, int length)
-{
- int i;
-
- for (i = 0; i < length; i++)
- if (dist_vect[i] != 0)
- return i + 1;
-
- return 0;
-}
-
-#endif /* LAMBDA_H */
#include "tree.h"
#include "gimple.h"
#include "ggc.h"
-#include "lambda.h" /* gcd */
#include "hashtab.h"
#include "plugin-api.h"
#include "lto-streamer.h"
void (*omega_when_reduced) (omega_pb) = omega_no_procedure;
-/* Compute the greatest common divisor of A and B. */
-
-static inline int
-gcd (int b, int a)
-{
- if (b == 1)
- return 1;
-
- while (b != 0)
- {
- int t = b;
- b = a % b;
- a = t;
- }
-
- return a;
-}
-
/* Print to FILE from PB equation E with all its coefficients
multiplied by C. */
NEXT_PASS (pass_record_bounds);
NEXT_PASS (pass_check_data_deps);
NEXT_PASS (pass_loop_distribution);
- NEXT_PASS (pass_linear_transform);
NEXT_PASS (pass_copy_prop);
NEXT_PASS (pass_graphite);
{
+2011-01-25 Sebastian Pop <sebastian.pop@amd.com>
+
+ * gfortran.dg/graphite/interchange-4.f: New.
+ * gfortran.dg/graphite/interchange-5.f: New.
+
+ * gcc.dg/tree-ssa/ltrans-1.c: Removed.
+ * gcc.dg/tree-ssa/ltrans-2.c: Removed.
+ * gcc.dg/tree-ssa/ltrans-3.c: Removed.
+ * gcc.dg/tree-ssa/ltrans-4.c: Removed.
+ * gcc.dg/tree-ssa/ltrans-5.c: Removed.
+ * gcc.dg/tree-ssa/ltrans-6.c: Removed.
+ * gcc.dg/tree-ssa/ltrans-8.c: Removed.
+ * gfortran.dg/ltrans-7.f90: Removed.
+ * gcc.dg/tree-ssa/data-dep-1.c: Removed.
+
+ * gcc.dg/pr18792.c: -> gcc.dg/graphite/pr18792.c
+ * gcc.dg/pr19910.c: -> gcc.dg/graphite/pr19910.c
+ * gcc.dg/tree-ssa/20041110-1.c: -> gcc.dg/graphite/pr20041110-1.c
+ * gcc.dg/tree-ssa/pr20256.c: -> gcc.dg/graphite/pr20256.c
+ * gcc.dg/pr23625.c: -> gcc.dg/graphite/pr23625.c
+ * gcc.dg/tree-ssa/pr23820.c: -> gcc.dg/graphite/pr23820.c
+ * gcc.dg/tree-ssa/pr24309.c: -> gcc.dg/graphite/pr24309.c
+ * gcc.dg/tree-ssa/pr26435.c: -> gcc.dg/graphite/pr26435.c
+ * gcc.dg/pr29330.c: -> gcc.dg/graphite/pr29330.c
+ * gcc.dg/pr29581-1.c: -> gcc.dg/graphite/pr29581-1.c
+ * gcc.dg/pr29581-2.c: -> gcc.dg/graphite/pr29581-2.c
+ * gcc.dg/pr29581-3.c: -> gcc.dg/graphite/pr29581-3.c
+ * gcc.dg/pr29581-4.c: -> gcc.dg/graphite/pr29581-4.c
+ * gcc.dg/tree-ssa/loop-27.c: -> gcc.dg/graphite/pr30565.c
+ * gcc.dg/tree-ssa/pr31183.c: -> gcc.dg/graphite/pr31183.c
+ * gcc.dg/tree-ssa/pr33576.c: -> gcc.dg/graphite/pr33576.c
+ * gcc.dg/tree-ssa/pr33766.c: -> gcc.dg/graphite/pr33766.c
+ * gcc.dg/pr34016.c: -> gcc.dg/graphite/pr34016.c
+ * gcc.dg/tree-ssa/pr34017.c: -> gcc.dg/graphite/pr34017.c
+ * gcc.dg/tree-ssa/pr34123.c: -> gcc.dg/graphite/pr34123.c
+ * gcc.dg/tree-ssa/pr36287.c: -> gcc.dg/graphite/pr36287.c
+ * gcc.dg/tree-ssa/pr37686.c: -> gcc.dg/graphite/pr37686.c
+ * gcc.dg/pr42917.c: -> gcc.dg/graphite/pr42917.c
+ * gcc.dg/tree-ssa/data-dep-1.c
+ * gfortran.dg/loop_nest_1.f90: -> gfortran.dg/graphite/pr29290.f90
+ * gfortran.dg/pr29581.f90: -> gfortran.dg/graphite/pr29581.f90
+ * gfortran.dg/pr36286.f90: -> gfortran.dg/graphite/pr36286.f90
+ * gfortran.dg/pr36922.f: -> gfortran.dg/graphite/pr36922.f
+ * gfortran.dg/pr39516.f: -> gfortran.dg/graphite/pr39516.f
+
2011-01-25 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/47265
--- /dev/null
+/* PR tree-optimization/18792 */
+/* { dg-do compile } */
+/* { dg-options "-O1 -ftree-loop-linear" } */
+void put_atoms_in_triclinic_unitcell(float x[][3])
+{
+ int i=0,d;
+
+ while (x[i][3] < 0)
+ for (d=0; d<=3; d++)
+ x[i][d] = 0;
+
+ while (x[i][3] >= 0)
+ for (d=0; d<=3; d++)
+ x[i][d] = 0;
+
+}
--- /dev/null
+/* Contributed by Volker Reichelt <reichelt@gcc.gnu.org>. */
+
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+int a[3];
+
+void foo()
+{
+ int i, j;
+
+ for (i = 1; i >= 0; --i)
+ for (j = i; j >= 0; --j)
+ a[i+j] = 0;
+}
+
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+/* This testcase was causing an ICE in building distance vectors because
+ we weren't ignoring the fact that one of the induction variables
+ involved in the dependence was outside of the loop. */
+extern int foo (int, int);
+int
+main (void)
+{
+ int a[50];
+ int b[50];
+ int i, j, k;
+ for (i = 4; i < 30; i++)
+ {
+ for (j = 3; j < 40; j++)
+ {
+ for (k = 9; k < 50; k++)
+ {
+ b[j] = a[i];
+ a[k] = b[i];
+ }
+ }
+ }
+ foo (a[i], b[i]);
+}
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+/* { dg-require-effective-target size32plus } */
+
+int foo()
+{
+ int x[2][2], y[2];
+ int i, n, s;
+
+ /* This is a reduction: there is a scalar dependence that cannot be
+ removed by rewriting IVs. This code cannot and should not be
+ transformed into a perfect loop. */
+ for (n = 0; n < 2; n++)
+ {
+ s = 0;
+ for (i = 0; i < 2; i++)
+ s += x[n][i]*y[i];
+ s += 1;
+ }
+
+ return s;
+}
--- /dev/null
+/* Test case for PR23625 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-vectorize -ftree-loop-linear" } */
+
+typedef long INT32;
+void find_best_colors ()
+{
+int ic0, ic1, ic2;
+INT32 * bptr;
+INT32 dist1;
+INT32 dist2;
+INT32 xx1;
+for (ic0 = (1<<(5 -3))-1;ic0 >= 0;ic0--)
+{
+ for (ic1 = (1<<(6 -3))-1;ic1 >= 0;ic1--)
+ {
+ dist2 = dist1;
+ for (ic2 = (1<<(5 -3))-1;ic2 >= 0;ic2--)
+ {
+ *bptr = dist2;
+ bptr++;
+ }
+ dist1 += xx1;
+ }
+}
+}
+
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+int t [2][4];
+
+void foo (void)
+{
+ int i, j, k, v;
+ float e;
+ for (;;)
+ {
+ v = 0;
+ for (j = 0; j < 2; j ++)
+ {
+ for (k = 2; k < 4; k ++)
+ {
+ e = 0.0;
+ for (i = 0; i < 4; i ++)
+ e += t [j][i];
+ if (e)
+ v = j;
+ }
+ }
+ t [v][0] = 0;
+ }
+}
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+float weight[10];
+void lsp_weight_quant(float *x, char *cdbk)
+{
+ int i,j;
+ float dist;
+ int best_id=0;
+ for (i=0;i<16;i++)
+ {
+ for (j=0;j<10;j++)
+ dist=dist+weight[j];
+ if (dist<0)
+ best_id=i;
+ }
+ x[j] = cdbk[best_id*10+j];
+}
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+/* { dg-require-effective-target size32plus } */
+
+int foo(int *p, int n)
+{
+ int i, j, k = 0;
+
+ /* This is a reduction: there is a scalar dependence that cannot be
+ removed by rewriting IVs. This code cannot and should not be
+ transformed into a perfect loop. */
+ for (i = 0; i < 2; ++i, p += n)
+ for (j = 0; j < 2; ++j)
+ k += p[j];
+
+ return k;
+}
--- /dev/null
+/* PR tree-optimization/29330 */
+/* { dg-do compile } */
+/* { dg-options "-O -ftree-loop-linear -std=gnu99" } */
+
+int buf[2][2][2][2];
+
+void
+f (void)
+{
+ for (int a = 0; a < 2; ++a)
+ for (int b = 0; b < 2; ++b)
+ for (int c = 0; c < 2; ++c)
+ for (int d = 0; d < 2; ++d)
+ buf[a][b][c][d] = 0;
+}
--- /dev/null
+/* PR tree-optimization/29581 */
+/* Origin: gcc.dg/vect/vect-85.c */
+/* { dg-do run } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+extern void abort (void);
+
+#define N 16
+
+int main1 (int *a)
+{
+ int i, j, k;
+ int b[N];
+
+ for (i = 0; i < N; i++)
+ {
+ for (j = 0; j < N; j++)
+ {
+ k = i + N;
+ a[j] = k;
+ }
+ b[i] = k;
+ }
+
+
+ for (j = 0; j < N; j++)
+ if (a[j] != i + N - 1)
+ abort();
+
+ for (j = 0; j < N; j++)
+ if (b[j] != j + N)
+ abort();
+
+ return 0;
+}
+
+int main (void)
+{
+ int a[N] __attribute__ ((__aligned__(16)));
+
+ main1 (a);
+
+ return 0;
+}
--- /dev/null
+/* PR tree-optimization/29581 */
+/* Origin: gcc.dg/vect/vect-86.c */
+/* { dg-do run } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+extern void abort (void);
+
+#define N 16
+
+int main1 (int n)
+{
+ int i, j, k;
+ int a[N], b[N];
+
+ for (i = 0; i < n; i++)
+ {
+ for (j = 0; j < n; j++)
+ {
+ k = i + n;
+ a[j] = k;
+ }
+ b[i] = k;
+ }
+
+
+ for (j = 0; j < n; j++)
+ if (a[j] != i + n - 1)
+ abort();
+
+ for (i = 0; i < n; i++)
+ if (b[i] != i + n)
+ abort();
+
+ return 0;
+}
+
+int main (void)
+{
+ main1 (N);
+ main1 (0);
+ main1 (1);
+ main1 (2);
+ main1 (N-1);
+
+ return 0;
+}
--- /dev/null
+/* PR tree-optimization/29581 */
+/* Origin: gcc.dg/vect/vect-87.c */
+/* { dg-do run } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+extern void abort (void);
+
+#define N 16
+
+int main1 (int n, int *a)
+{
+ int i, j, k;
+ int b[N];
+
+ for (i = 0; i < n; i++)
+ {
+ for (j = 0; j < n; j++)
+ {
+ k = i + n;
+ a[j] = k;
+ }
+ b[i] = k;
+ }
+
+
+ for (j = 0; j < n; j++)
+ if (a[j] != i + n - 1)
+ abort();
+
+ for (j = 0; j < n; j++)
+ if (b[j] != j + n)
+ abort();
+
+ return 0;
+}
+
+int main (void)
+{
+ int a[N] __attribute__ ((__aligned__(16)));
+
+ main1 (N, a);
+ main1 (0, a);
+ main1 (1, a);
+ main1 (2, a);
+ main1 (N-1, a);
+
+ return 0;
+}
--- /dev/null
+/* PR tree-optimization/29581 */
+/* Origin: gcc.dg/vect/vect-88.c */
+/* { dg-do run } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+extern void abort (void);
+
+#define N 16
+
+int main1 (int n, int *a)
+{
+ int i, j, k;
+ int b[N];
+
+ for (i = 0; i < n; i++)
+ {
+ for (j = 0; j < n; j++)
+ {
+ k = i + n;
+ a[j] = k;
+ }
+ b[i] = k;
+ }
+
+
+ for (j = 0; j < n; j++)
+ if (a[j] != i + n - 1)
+ abort();
+
+ for (j = 0; j < n; j++)
+ if (b[j] != j + n)
+ abort();
+
+ return 0;
+}
+
+int main (void)
+{
+ int a[N+1] __attribute__ ((__aligned__(16)));
+
+ main1 (N, a+1);
+ main1 (0, a+1);
+ main1 (1, a+1);
+ main1 (2, a+1);
+ main1 (N-1, a+1);
+
+ return 0;
+}
--- /dev/null
+/* PR tree-optimization/30565 */
+
+/* { dg-do compile } */
+/* { dg-options "-O1 -ftree-pre -ftree-loop-linear" } */
+
+static double snrdef[32];
+void psycho_n1(double ltmin[2][32], int stereo)
+{
+ int i, k;
+
+ for (k = 0; k < stereo; k++)
+ for (i = 0; i < 32; i++)
+ ltmin[k][i] = snrdef[i];
+}
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+int buf[256 * 9];
+int f()
+{
+ int i, j;
+
+ for (i = 0; i < 256; ++i)
+ for (j = 0; j < 8; ++j)
+ buf[j + 1] = buf[j] + 1;
+
+ return buf[10];
+}
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+int a1[6][4][4];
+short b1[16];
+
+int c1;
+void CalculateQuantParam(void)
+{
+ int i, j, k, temp;
+
+ for(k=0; k<6; k++)
+ for(j=0; j<4; j++)
+ for(i=0; i<4; i++)
+ {
+ temp = (i<<2)+j;
+ a1[k][j][i] = c1/b1[temp];
+ }
+}
+
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+float
+fxt1_quantize_ALPHA1()
+{
+ int j1;
+ int i;
+ float *tv;
+ for (j1 = 1; j1; j1++) {
+ float e;
+ for (i = 1; i; i++)
+ e = tv[i];
+ if (e)
+ i = j1;
+ }
+ return tv[i];
+}
+
--- /dev/null
+/* PR tree-optimization/34016 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+extern void bar (double *);
+
+void foo (void)
+{
+ double gr[36];
+ int i, j;
+ for (i = 0; i <= 5; i++)
+ {
+ for (j = 0; j <= 5; j++)
+ gr[i + j * 6] = 0.0;
+ if (i <= 2)
+ gr[i + i * 6] = 1.0;
+ }
+ bar (gr);
+}
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+extern int s;
+
+void
+foo (int *x, int y, int z)
+{
+ int m, n;
+ int o;
+ int p = x[0];
+ o = s;
+ for (m = 0; m < s; m++)
+ for (n = 0; n < s; n++)
+ {
+ if (x[n] != p)
+ continue;
+ if (m > z)
+ z = m;
+ if (n < o)
+ o = n;
+ }
+ for (m = y; m <= z; m++)
+ {
+ }
+}
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-loop-linear" } */
+
+/* Testcase by Martin Michlmayr <tbm@cyrius.com> */
+
+static unsigned char sbox[256] = {
+};
+void MD2Transform (unsigned char state[16])
+{
+ unsigned char t = 0;
+ int i, j;
+ for (i = 0; i < 16; i++)
+ {
+ for (j = 0; j < 2; j++)
+ t = (state[j] ^= sbox[t]);
+ t += i;
+ }
+}
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O -ftree-loop-linear" } */
+
+int tab[2][2];
+
+int foo ()
+{
+ int i, j, k;
+
+ for (i = 0; i < 2; ++i)
+ for (j = 0; j < 2; ++j)
+ for (k = 0; k < 2; ++k)
+ {}
+
+ for (i = 0; i < 2; ++i)
+ for (j = 0; j < 2; ++j)
+ if (i == 0)
+ tab[i][j] = 0;
+
+ return tab[0][1];
+}
+
--- /dev/null
+/* { dg-do compile { target powerpc*-*-* } } */
+/* { dg-options "-O3 -ftree-loop-linear" } */
+
+unsigned char inUse[256];
+unsigned char len[6][258];
+int code[6][258];
+unsigned int crc32Table[256] = { };
+ unsigned int getGlobalCRC (void) { }
+ int bsLive;
+void bsW (int n, unsigned int v) {
+ while (bsLive >= 8) {}
+ }
+ void hbAssignCodes (int * code, unsigned char * length, int minLen,
+int maxLen, int alphaSize) {
+ int n, vec, i;
+ for (n = minLen;n <= maxLen;n++)
+ for (i = 0; i < alphaSize;i++)
+ code[i] = vec;
+ }
+ void sendMTFValues (void) {
+ int v, t, i, j, gs, ge, totc, bt, bc, iter;
+ int nSelectors, alphaSize, minLen, maxLen, selCtr;
+ int nGroups, nBytes;
+ {
+ while (1)
+ {
+ break;
+ }
+ hbAssignCodes (&code[t][0], &len[t][0], minLen, maxLen, alphaSize);
+ unsigned char inUse16[16];
+ for (i = 0;i < 16;i++)
+ if (inUse16[i])
+ {
+ for (j = 0;j < 16;j++)
+ if (inUse[i * 16 + j]) { }
+ }
+ }
+ for (i = 0; i < nSelectors;i++) { }
+ for (t = 0; t < nGroups;t++)
+ {
+ int curr = len[t][0];
+ for (i = 0; i < alphaSize;i++)
+ while (curr < len[t][i]) { }
+ }
+ while (1)
+ for (i = gs; i <= ge;i++) { }
+ }
+
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O1 -ftree-loop-linear -fcompare-debug" } */
+
+extern int A[];
+
+void
+foo ()
+{
+ int i, j;
+ for (i = 0; i < 4; i++)
+ for (j = 255; j >= 0; j--)
+ A[j] = 0;
+}
+++ /dev/null
-/* PR tree-optimization/18792 */
-/* { dg-do compile } */
-/* { dg-options "-O1 -ftree-loop-linear" } */
-void put_atoms_in_triclinic_unitcell(float x[][3])
-{
- int i=0,d;
-
- while (x[i][3] < 0)
- for (d=0; d<=3; d++)
- x[i][d] = 0;
-
- while (x[i][3] >= 0)
- for (d=0; d<=3; d++)
- x[i][d] = 0;
-
-}
+++ /dev/null
-/* Contributed by Volker Reichelt <reichelt@gcc.gnu.org>. */
-
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-int a[3];
-
-void foo()
-{
- int i, j;
-
- for (i = 1; i >= 0; --i)
- for (j = i; j >= 0; --j)
- a[i+j] = 0;
-}
-
+++ /dev/null
-/* Test case for PR23625 */
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-vectorize -ftree-loop-linear" } */
-
-typedef long INT32;
-void find_best_colors ()
-{
-int ic0, ic1, ic2;
-INT32 * bptr;
-INT32 dist1;
-INT32 dist2;
-INT32 xx1;
-for (ic0 = (1<<(5 -3))-1;ic0 >= 0;ic0--)
-{
- for (ic1 = (1<<(6 -3))-1;ic1 >= 0;ic1--)
- {
- dist2 = dist1;
- for (ic2 = (1<<(5 -3))-1;ic2 >= 0;ic2--)
- {
- *bptr = dist2;
- bptr++;
- }
- dist1 += xx1;
- }
-}
-}
-
+++ /dev/null
-/* PR tree-optimization/29330 */
-/* { dg-do compile } */
-/* { dg-options "-O -ftree-loop-linear -std=gnu99" } */
-
-int buf[2][2][2][2];
-
-void
-f (void)
-{
- for (int a = 0; a < 2; ++a)
- for (int b = 0; b < 2; ++b)
- for (int c = 0; c < 2; ++c)
- for (int d = 0; d < 2; ++d)
- buf[a][b][c][d] = 0;
-}
+++ /dev/null
-/* PR tree-optimization/29581 */
-/* Origin: gcc.dg/vect/vect-85.c */
-/* { dg-do run } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-extern void abort (void);
-
-#define N 16
-
-int main1 (int *a)
-{
- int i, j, k;
- int b[N];
-
- for (i = 0; i < N; i++)
- {
- for (j = 0; j < N; j++)
- {
- k = i + N;
- a[j] = k;
- }
- b[i] = k;
- }
-
-
- for (j = 0; j < N; j++)
- if (a[j] != i + N - 1)
- abort();
-
- for (j = 0; j < N; j++)
- if (b[j] != j + N)
- abort();
-
- return 0;
-}
-
-int main (void)
-{
- int a[N] __attribute__ ((__aligned__(16)));
-
- main1 (a);
-
- return 0;
-}
+++ /dev/null
-/* PR tree-optimization/29581 */
-/* Origin: gcc.dg/vect/vect-86.c */
-/* { dg-do run } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-extern void abort (void);
-
-#define N 16
-
-int main1 (int n)
-{
- int i, j, k;
- int a[N], b[N];
-
- for (i = 0; i < n; i++)
- {
- for (j = 0; j < n; j++)
- {
- k = i + n;
- a[j] = k;
- }
- b[i] = k;
- }
-
-
- for (j = 0; j < n; j++)
- if (a[j] != i + n - 1)
- abort();
-
- for (i = 0; i < n; i++)
- if (b[i] != i + n)
- abort();
-
- return 0;
-}
-
-int main (void)
-{
- main1 (N);
- main1 (0);
- main1 (1);
- main1 (2);
- main1 (N-1);
-
- return 0;
-}
+++ /dev/null
-/* PR tree-optimization/29581 */
-/* Origin: gcc.dg/vect/vect-87.c */
-/* { dg-do run } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-extern void abort (void);
-
-#define N 16
-
-int main1 (int n, int *a)
-{
- int i, j, k;
- int b[N];
-
- for (i = 0; i < n; i++)
- {
- for (j = 0; j < n; j++)
- {
- k = i + n;
- a[j] = k;
- }
- b[i] = k;
- }
-
-
- for (j = 0; j < n; j++)
- if (a[j] != i + n - 1)
- abort();
-
- for (j = 0; j < n; j++)
- if (b[j] != j + n)
- abort();
-
- return 0;
-}
-
-int main (void)
-{
- int a[N] __attribute__ ((__aligned__(16)));
-
- main1 (N, a);
- main1 (0, a);
- main1 (1, a);
- main1 (2, a);
- main1 (N-1, a);
-
- return 0;
-}
+++ /dev/null
-/* PR tree-optimization/29581 */
-/* Origin: gcc.dg/vect/vect-88.c */
-/* { dg-do run } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-extern void abort (void);
-
-#define N 16
-
-int main1 (int n, int *a)
-{
- int i, j, k;
- int b[N];
-
- for (i = 0; i < n; i++)
- {
- for (j = 0; j < n; j++)
- {
- k = i + n;
- a[j] = k;
- }
- b[i] = k;
- }
-
-
- for (j = 0; j < n; j++)
- if (a[j] != i + n - 1)
- abort();
-
- for (j = 0; j < n; j++)
- if (b[j] != j + n)
- abort();
-
- return 0;
-}
-
-int main (void)
-{
- int a[N+1] __attribute__ ((__aligned__(16)));
-
- main1 (N, a+1);
- main1 (0, a+1);
- main1 (1, a+1);
- main1 (2, a+1);
- main1 (N-1, a+1);
-
- return 0;
-}
+++ /dev/null
-/* PR tree-optimization/34016 */
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-extern void bar (double *);
-
-void foo (void)
-{
- double gr[36];
- int i, j;
- for (i = 0; i <= 5; i++)
- {
- for (j = 0; j <= 5; j++)
- gr[i + j * 6] = 0.0;
- if (i <= 2)
- gr[i + i * 6] = 1.0;
- }
- bar (gr);
-}
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O1 -ftree-loop-linear -fcompare-debug -fdump-tree-ltrans" } */
-
-extern int A[];
-
-void
-foo ()
-{
- int i, j;
- for (i = 0; i < 4; i++)
- for (j = 255; j >= 0; j--)
- A[j] = 0;
-}
-
-/* { dg-final { scan-tree-dump "Successfully transformed loop" "ltrans" } } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-/* This testcase was causing an ICE in building distance vectors because
- we weren't ignoring the fact that one of the induction variables
- involved in the dependence was outside of the loop. */
-extern int foo (int, int);
-int
-main (void)
-{
- int a[50];
- int b[50];
- int i, j, k;
- for (i = 4; i < 30; i++)
- {
- for (j = 3; j < 40; j++)
- {
- for (k = 9; k < 50; k++)
- {
- b[j] = a[i];
- a[k] = b[i];
- }
- }
- }
- foo (a[i], b[i]);
-}
+++ /dev/null
-/* { dg-do compile { target int32plus } } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-
-int foo (int n, int m)
-{
- int a[10000][10000];
- int i, j, k;
-
- for(k = 0; k < 1234; k++)
- for(j = 0; j < 5; j++)
- for(i = 0; i < 67; i++)
- {
- a[j+i-(-m+n+3)][i-k+4] = a[k+j][i];
- }
-
- return a[0][0];
-}
-
-
-/* For the data dependence analysis of the outermost loop, the
- evolution of "k+j" should be instantiated in the outermost loop "k"
- and the evolution should be taken in the innermost loop "i". The
- pattern below ensures that the evolution is not computed in the
- outermost "k" loop: the 4 comes from the instantiation of the
- number of iterations of loop "j". */
-
-/* { dg-final { scan-tree-dump-times "4, \\+, 1" 0 "ltrans" } } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* PR tree-optimization/30565 */
-
-/* { dg-do compile } */
-/* { dg-options "-O1 -ftree-pre -ftree-loop-linear" } */
-
-static double snrdef[32];
-void psycho_n1(double ltmin[2][32], int stereo)
-{
- int i, k;
-
- for (k = 0; k < stereo; k++)
- for (i = 0; i < 32; i++)
- ltmin[k][i] = snrdef[i];
-}
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all -march=i486" { target { i?86-*-* && ilp32} } } */
-/* { dg-require-effective-target size32plus } */
-
-double u[1782225];
-int foo(int N, int *res)
-{
- int i, j;
- double sum = 0.0;
- /* This loop should be converted to a perfect nest and
- interchanged. */
- for (i = 0; i < N; i++)
- {
- for (j = 0; j < N; j++)
- sum = sum + u[i + 1335 * j];
-
- u[1336 * i] *= 2;
- }
- *res = sum + N;
-}
-/* { dg-final { scan-tree-dump-times "converted loop nest to perfect loop nest" 1 "ltrans"} } */
-/* { dg-final { scan-tree-dump-times "transformed loop" 1 "ltrans"} } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-/* { dg-require-effective-target size32plus } */
-
-double u[1782225];
-int foo(int N, int *res)
-{
- unsigned int i, j;
- double sum = 0;
-
- /* This loop should be converted to a perfect nest and
- interchanged. */
- for (i = 0; i < N; i++)
- {
- for (j = 0; j < N; j++)
- {
- sum = sum + u[i + 1335 * j];
- if (j == N - 1)
- u[1336 * i] *= 2;
- }
- }
- *res = sum + N;
-}
-/* { dg-final { scan-tree-dump-times "transformed loop" 1 "ltrans"} {
- xfail *-*-*} } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all -march=i486" { target { i?86-*-* && ilp32} } } */
-/* { dg-require-effective-target size32plus } */
-
-double u[1782225];
-int foo(int N, int *res)
-{
- unsigned int i, j;
- double sum = 0;
- for (i = 0; i < N; i++)
- {
- for (j = 0; j < N; j++)
- {
- sum = sum + u[i + 1335 * j];
- }
- }
- *res = sum + N;
-}
-
-/* { dg-final { scan-tree-dump-times "transformed loop" 1 "ltrans" } } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all -march=i486" { target { i?86-*-* && ilp32} } } */
-/* { dg-require-effective-target size32plus } */
-
-double u[1782225];
-int foo(int N, int *res)
-{
- int i, j;
- double sum = 0;
- for (i = 0; i < N; i++)
- for (j = 0; j < N; j++)
- sum = sum + u[i + 1335 * j];
-
- for (i = 0; i < N; i++)
- u[1336 * i] *= 2;
- *res = sum + N;
-}
-
-/* { dg-final { scan-tree-dump-times "transformed loop" 1 "ltrans"} } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile { target { size32plus } } } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all -march=i486" { target { i?86-*-* && ilp32} } } */
-
-int foo ()
-{
- int A[100][1111];
- int i, j;
-
- for( i = 0; i < 1111; i++)
- for( j = 0; j < 100; j++)
- A[j][i] = 5 * A[j][i];
-
- return A[10][10];
-}
-
-/* { dg-final { scan-tree-dump-times "transformed loop" 1 "ltrans"} } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all -march=i486" { target { i?86-*-* && ilp32} } } */
-/* { dg-require-effective-target size32plus } */
-
-
-
-int medium_loop_interchange(int A[100][200])
-{
- int i,j;
-
- /* This loop should be interchanged. */
-
- for(j = 0; j < 200; j++)
- for(i = 0; i < 100; i++)
- A[i][j] = A[i][j] + A[i][j];
-
- return A[1][1];
-}
-
-/* { dg-final { scan-tree-dump-times "transformed loop" 1 "ltrans"} } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all -march=i486" { target { i?86-*-* && ilp32} } } */
-double foo(double *a)
-{
- int i,j;
- double r = 0.0;
- for (i=0; i<100; ++i)
- for (j=0; j<1000; ++j)
- r += a[j*100+i];
- return r;
-}
-
-/* { dg-final { scan-tree-dump-times "transformed loop" 1 "ltrans"} } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-/* { dg-require-effective-target size32plus } */
-
-int foo()
-{
- int x[2][2], y[2];
- int i, n, s;
-
- /* This is a reduction: there is a scalar dependence that cannot be
- removed by rewriting IVs. This code cannot and should not be
- transformed into a perfect loop. */
- for (n = 0; n < 2; n++)
- {
- s = 0;
- for (i = 0; i < 2; i++)
- s += x[n][i]*y[i];
- s += 1;
- }
-
- return s;
-}
-
-/* { dg-final { scan-tree-dump-times "converted loop nest to perfect loop nest" 0 "ltrans"} } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-int t [2][4];
-
-void foo (void)
-{
- int i, j, k, v;
- float e;
- for (;;)
- {
- v = 0;
- for (j = 0; j < 2; j ++)
- {
- for (k = 2; k < 4; k ++)
- {
- e = 0.0;
- for (i = 0; i < 4; i ++)
- e += t [j][i];
- if (e)
- v = j;
- }
- }
- t [v][0] = 0;
- }
-}
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-float weight[10];
-void lsp_weight_quant(float *x, char *cdbk)
-{
- int i,j;
- float dist;
- int best_id=0;
- for (i=0;i<16;i++)
- {
- for (j=0;j<10;j++)
- dist=dist+weight[j];
- if (dist<0)
- best_id=i;
- }
- x[j] = cdbk[best_id*10+j];
-}
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" } */
-/* { dg-require-effective-target size32plus } */
-
-int foo(int *p, int n)
-{
- int i, j, k = 0;
-
- /* This is a reduction: there is a scalar dependence that cannot be
- removed by rewriting IVs. This code cannot and should not be
- transformed into a perfect loop. */
- for (i = 0; i < 2; ++i, p += n)
- for (j = 0; j < 2; ++j)
- k += p[j];
-
- return k;
-}
-
-/* { dg-final { scan-tree-dump-times "converted loop nest to perfect loop nest" 0 "ltrans"} } */
-/* { dg-final { cleanup-tree-dump "ltrans" } } */
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-int buf[256 * 9];
-int f()
-{
- int i, j;
-
- for (i = 0; i < 256; ++i)
- for (j = 0; j < 8; ++j)
- buf[j + 1] = buf[j] + 1;
-
- return buf[10];
-}
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-int a1[6][4][4];
-short b1[16];
-
-int c1;
-void CalculateQuantParam(void)
-{
- int i, j, k, temp;
-
- for(k=0; k<6; k++)
- for(j=0; j<4; j++)
- for(i=0; i<4; i++)
- {
- temp = (i<<2)+j;
- a1[k][j][i] = c1/b1[temp];
- }
-}
-
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-float
-fxt1_quantize_ALPHA1()
-{
- int j1;
- int i;
- float *tv;
- for (j1 = 1; j1; j1++) {
- float e;
- for (i = 1; i; i++)
- e = tv[i];
- if (e)
- i = j1;
- }
- return tv[i];
-}
-
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-extern int s;
-
-void
-foo (int *x, int y, int z)
-{
- int m, n;
- int o;
- int p = x[0];
- o = s;
- for (m = 0; m < s; m++)
- for (n = 0; n < s; n++)
- {
- if (x[n] != p)
- continue;
- if (m > z)
- z = m;
- if (n < o)
- o = n;
- }
- for (m = y; m <= z; m++)
- {
- }
-}
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O2 -ftree-loop-linear" } */
-
-/* Testcase by Martin Michlmayr <tbm@cyrius.com> */
-
-static unsigned char sbox[256] = {
-};
-void MD2Transform (unsigned char state[16])
-{
- unsigned char t = 0;
- int i, j;
- for (i = 0; i < 16; i++)
- {
- for (j = 0; j < 2; j++)
- t = (state[j] ^= sbox[t]);
- t += i;
- }
-}
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O -ftree-loop-linear" } */
-
-int tab[2][2];
-
-int foo ()
-{
- int i, j, k;
-
- for (i = 0; i < 2; ++i)
- for (j = 0; j < 2; ++j)
- for (k = 0; k < 2; ++k)
- {}
-
- for (i = 0; i < 2; ++i)
- for (j = 0; j < 2; ++j)
- if (i == 0)
- tab[i][j] = 0;
-
- return tab[0][1];
-}
-
+++ /dev/null
-/* { dg-do compile { target powerpc*-*-* } } */
-/* { dg-options "-O3 -ftree-loop-linear" } */
-
-unsigned char inUse[256];
-unsigned char len[6][258];
-int code[6][258];
-unsigned int crc32Table[256] = { };
- unsigned int getGlobalCRC (void) { }
- int bsLive;
-void bsW (int n, unsigned int v) {
- while (bsLive >= 8) {}
- }
- void hbAssignCodes (int * code, unsigned char * length, int minLen,
-int maxLen, int alphaSize) {
- int n, vec, i;
- for (n = minLen;n <= maxLen;n++)
- for (i = 0; i < alphaSize;i++)
- code[i] = vec;
- }
- void sendMTFValues (void) {
- int v, t, i, j, gs, ge, totc, bt, bc, iter;
- int nSelectors, alphaSize, minLen, maxLen, selCtr;
- int nGroups, nBytes;
- {
- while (1)
- {
- break;
- }
- hbAssignCodes (&code[t][0], &len[t][0], minLen, maxLen, alphaSize);
- unsigned char inUse16[16];
- for (i = 0;i < 16;i++)
- if (inUse16[i])
- {
- for (j = 0;j < 16;j++)
- if (inUse[i * 16 + j]) { }
- }
- }
- for (i = 0; i < nSelectors;i++) { }
- for (t = 0; t < nGroups;t++)
- {
- int curr = len[t][0];
- for (i = 0; i < alphaSize;i++)
- while (curr < len[t][i]) { }
- }
- while (1)
- for (i = gs; i <= ge;i++) { }
- }
-
--- /dev/null
+ subroutine s231 (ntimes,ld,n,ctime,dtime,a,b,c,d,e,aa,bb,cc)
+c
+c loop interchange
+c loop with multiple dimension recursion
+c
+ integer ntimes, ld, n, i, nl, j
+ double precision a(n), b(n), c(n), d(n), e(n), aa(ld,n),
+ + bb(ld,n), cc(ld,n)
+ double precision chksum, cs2d
+ real t1, t2, second, ctime, dtime
+
+ call init(ld,n,a,b,c,d,e,aa,bb,cc,'s231 ')
+ t1 = second()
+ do 1 nl = 1,ntimes/n
+ do 10 i=1,n
+ do 20 j=2,n
+ aa(i,j) = aa(i,j-1) + bb(i,j)
+ 20 continue
+ 10 continue
+ call dummy(ld,n,a,b,c,d,e,aa,bb,cc,1.d0)
+ 1 continue
+ t2 = second() - t1 - ctime - ( dtime * float(ntimes/n) )
+ chksum = cs2d(n,aa)
+ call check (chksum,(ntimes/n)*n*(n-1),n,t2,'s231 ')
+ return
+ end
+
+! { dg-final { scan-tree-dump-times "will be interchanged" 1 "graphite" { xfail *-*-* } } }
+! { dg-final { cleanup-tree-dump "graphite" } }
--- /dev/null
+ subroutine s235 (ntimes,ld,n,ctime,dtime,a,b,c,d,e,aa,bb,cc)
+c
+c loop interchanging
+c imperfectly nested loops
+c
+ integer ntimes, ld, n, i, nl, j
+ double precision a(n), b(n), c(n), d(n), e(n), aa(ld,n),
+ + bb(ld,n), cc(ld,n)
+ double precision chksum, cs1d, cs2d
+ real t1, t2, second, ctime, dtime
+
+ call init(ld,n,a,b,c,d,e,aa,bb,cc,'s235 ')
+ t1 = second()
+ do 1 nl = 1,ntimes/n
+ do 10 i = 1,n
+ a(i) = a(i) + b(i) * c(i)
+ do 20 j = 2,n
+ aa(i,j) = aa(i,j-1) + bb(i,j) * a(i)
+ 20 continue
+ 10 continue
+ call dummy(ld,n,a,b,c,d,e,aa,bb,cc,1.d0)
+ 1 continue
+ t2 = second() - t1 - ctime - ( dtime * float(ntimes/n) )
+ chksum = cs2d(n,aa) + cs1d(n,a)
+ call check (chksum,(ntimes/n)*n*(n-1),n,t2,'s235 ')
+ return
+ end
+
+! { dg-final { scan-tree-dump-times "will be interchanged" 1 "graphite" { xfail *-*-* } } }
+! { dg-final { cleanup-tree-dump "graphite" } }
--- /dev/null
+! PR tree-optimization/29290
+! { dg-do compile }
+! { dg-options "-O3 -ftree-loop-linear" }
+
+subroutine pr29290 (a, b, c, d)
+ integer c, d
+ real*8 a(c,c), b(c,c)
+ a(1:d,1:d) = b(1:d,1:d)
+end
--- /dev/null
+! PR tree-optimization/29581
+! { dg-do run }
+! { dg-options "-O2 -ftree-loop-linear" }
+
+ SUBROUTINE FOO (K)
+ INTEGER I, J, K, A(5,5), B
+ COMMON A
+ A(1,1) = 1
+ 10 B = 0
+ DO 30 I = 1, K
+ DO 20 J = 1, K
+ B = B + A(I,J)
+ 20 CONTINUE
+ A(I,I) = A(I,I) * 2
+ 30 CONTINUE
+ IF (B.GE.3) RETURN
+ GO TO 10
+ END SUBROUTINE
+
+ PROGRAM BAR
+ INTEGER A(5,5)
+ COMMON A
+ CALL FOO (2)
+ IF (A(1,1).NE.8) CALL ABORT
+ A(1,1) = 0
+ IF (ANY(A.NE.0)) CALL ABORT
+ END
--- /dev/null
+! { dg-do compile }
+! { dg-options "-O1 -ftree-loop-linear" }
+! PR tree-optimization/36286
+
+program test_count
+ integer, dimension(2,3) :: a, b
+ a = reshape( (/ 1, 3, 5, 2, 4, 6 /), (/ 2, 3 /))
+ b = reshape( (/ 0, 3, 5, 7, 4, 8 /), (/ 2, 3 /))
+ print '(3l6)', a.ne.b
+ print *, a(1,:).ne.b(1,:)
+ print *, a(2,:).ne.b(2,:)
+ print *, count(a.ne.b)
+end program test_count
+
--- /dev/null
+C PR tree-optimization/36922
+C { dg-do compile }
+C { dg-options "-O2 -ftree-loop-linear" }
+ SUBROUTINE PR36922(N,F,Z,C)
+ IMPLICIT DOUBLE PRECISION(A-H,O-Z)
+ DIMENSION C(23821),Z(0:2*N+1),F(0:2*N)
+ I=0
+ DO L=0,N
+ DO M=0,L
+ DO M2=M,L
+ I=I+1
+ C(I)=F(L+M)*F(L-M)*Z(L-M2)/(F(M2+M)*F(M2-M)*F(L-M2)*F(L-M2))
+ ENDDO
+ ENDDO
+ ENDDO
+ END
--- /dev/null
+C PR tree-optimization/39516
+C { dg-do compile }
+C { dg-options "-O2 -ftree-loop-linear" }
+ SUBROUTINE SUB(A, B, M)
+ IMPLICIT NONE
+ DOUBLE PRECISION A(20,20), B(20)
+ INTEGER*8 I, J, K, M
+ DO I=1,M
+ DO J=1,M
+ A(I,J)=A(I,J)+1
+ END DO
+ END DO
+ DO K=1,20
+ DO I=1,M
+ DO J=1,M
+ B(I)=B(I)+A(I,J)
+ END DO
+ END DO
+ END DO
+ END SUBROUTINE
+++ /dev/null
-! PR tree-optimization/29290
-! { dg-do compile }
-! { dg-options "-O3 -ftree-loop-linear" }
-
-subroutine pr29290 (a, b, c, d)
- integer c, d
- real*8 a(c,c), b(c,c)
- a(1:d,1:d) = b(1:d,1:d)
-end
+++ /dev/null
-! { dg-do compile }
-! { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all" }
-! { dg-options "-O2 -ftree-loop-linear -fdump-tree-ltrans-all -march=i486" { target { i?86-*-* && ilp32 } } }
-
-Program FOO
- IMPLICIT INTEGER (I-N)
- IMPLICIT REAL*8 (A-H, O-Z)
- PARAMETER (N1=1335, N2=1335)
- COMMON U(N1,N2), V(N1,N2), P(N1,N2)
-
- PC = 0.0D0
- UC = 0.0D0
- VC = 0.0D0
-
- do I = 1, M
- do J = 1, M
- PC = PC + abs(P(I,J))
- UC = UC + abs(U(I,J))
- VC = VC + abs(V(I,J))
- end do
- U(I,I) = U(I,I) * ( mod (I, 100) /100.)
- end do
-
- write(6,366) PC, UC, VC
-366 format(/, ' PC = ',E12.4,/,' UC = ',E12.4,/,' VC = ',E12.4,/)
-
-end Program FOO
-
-! Please do not XFAIL.
-! { dg-final { scan-tree-dump-times "transformed loop" 1 "ltrans" } }
-! { dg-final { cleanup-tree-dump "ltrans" } }
+++ /dev/null
-! PR tree-optimization/29581
-! { dg-do run }
-! { dg-options "-O2 -ftree-loop-linear" }
-
- SUBROUTINE FOO (K)
- INTEGER I, J, K, A(5,5), B
- COMMON A
- A(1,1) = 1
- 10 B = 0
- DO 30 I = 1, K
- DO 20 J = 1, K
- B = B + A(I,J)
- 20 CONTINUE
- A(I,I) = A(I,I) * 2
- 30 CONTINUE
- IF (B.GE.3) RETURN
- GO TO 10
- END SUBROUTINE
-
- PROGRAM BAR
- INTEGER A(5,5)
- COMMON A
- CALL FOO (2)
- IF (A(1,1).NE.8) CALL ABORT
- A(1,1) = 0
- IF (ANY(A.NE.0)) CALL ABORT
- END
+++ /dev/null
-! { dg-do compile }
-! { dg-options "-O1 -ftree-loop-linear" }
-! PR tree-optimization/36286
-
-program test_count
- integer, dimension(2,3) :: a, b
- a = reshape( (/ 1, 3, 5, 2, 4, 6 /), (/ 2, 3 /))
- b = reshape( (/ 0, 3, 5, 7, 4, 8 /), (/ 2, 3 /))
- print '(3l6)', a.ne.b
- print *, a(1,:).ne.b(1,:)
- print *, a(2,:).ne.b(2,:)
- print *, count(a.ne.b)
-end program test_count
-
+++ /dev/null
-C PR tree-optimization/36922
-C { dg-do compile }
-C { dg-options "-O2 -ftree-loop-linear" }
- SUBROUTINE PR36922(N,F,Z,C)
- IMPLICIT DOUBLE PRECISION(A-H,O-Z)
- DIMENSION C(23821),Z(0:2*N+1),F(0:2*N)
- I=0
- DO L=0,N
- DO M=0,L
- DO M2=M,L
- I=I+1
- C(I)=F(L+M)*F(L-M)*Z(L-M2)/(F(M2+M)*F(M2-M)*F(L-M2)*F(L-M2))
- ENDDO
- ENDDO
- ENDDO
- END
+++ /dev/null
-C PR tree-optimization/39516
-C { dg-do compile }
-C { dg-options "-O2 -ftree-loop-linear" }
- SUBROUTINE SUB(A, B, M)
- IMPLICIT NONE
- DOUBLE PRECISION A(20,20), B(20)
- INTEGER*8 I, J, K, M
- DO I=1,M
- DO J=1,M
- A(I,J)=A(I,J)+1
- END DO
- END DO
- DO K=1,20
- DO I=1,M
- DO J=1,M
- B(I)=B(I)+A(I,J)
- END DO
- END DO
- END DO
- END SUBROUTINE
print_direction_vector (outf, v, length);
}
+/* Print out a vector VEC of length N to OUTFILE. */
+
+static inline void
+print_lambda_vector (FILE * outfile, lambda_vector vector, int n)
+{
+ int i;
+
+ for (i = 0; i < n; i++)
+ fprintf (outfile, "%3d ", vector[i]);
+ fprintf (outfile, "\n");
+}
+
/* Print a vector of distance vectors. */
void
affine_fn_free (overlaps_b_xyz);
}
+/* Copy the elements of vector VEC1 with length SIZE to VEC2. */
+
+static void
+lambda_vector_copy (lambda_vector vec1, lambda_vector vec2,
+ int size)
+{
+ memcpy (vec2, vec1, size * sizeof (*vec1));
+}
+
+/* Copy the elements of M x N matrix MAT1 to MAT2. */
+
+static void
+lambda_matrix_copy (lambda_matrix mat1, lambda_matrix mat2,
+ int m, int n)
+{
+ int i;
+
+ for (i = 0; i < m; i++)
+ lambda_vector_copy (mat1[i], mat2[i], n);
+}
+
+/* Store the N x N identity matrix in MAT. */
+
+static void
+lambda_matrix_id (lambda_matrix mat, int size)
+{
+ int i, j;
+
+ for (i = 0; i < size; i++)
+ for (j = 0; j < size; j++)
+ mat[i][j] = (i == j) ? 1 : 0;
+}
+
+/* Return the first nonzero element of vector VEC1 between START and N.
+ We must have START <= N. Returns N if VEC1 is the zero vector. */
+
+static int
+lambda_vector_first_nz (lambda_vector vec1, int n, int start)
+{
+ int j = start;
+ while (j < n && vec1[j] == 0)
+ j++;
+ return j;
+}
+
+/* Add a multiple of row R1 of matrix MAT with N columns to row R2:
+ R2 = R2 + CONST1 * R1. */
+
+static void
+lambda_matrix_row_add (lambda_matrix mat, int n, int r1, int r2, int const1)
+{
+ int i;
+
+ if (const1 == 0)
+ return;
+
+ for (i = 0; i < n; i++)
+ mat[r2][i] += const1 * mat[r1][i];
+}
+
+/* Swap rows R1 and R2 in matrix MAT. */
+
+static void
+lambda_matrix_row_exchange (lambda_matrix mat, int r1, int r2)
+{
+ lambda_vector row;
+
+ row = mat[r1];
+ mat[r1] = mat[r2];
+ mat[r2] = row;
+}
+
+/* Multiply vector VEC1 of length SIZE by a constant CONST1,
+ and store the result in VEC2. */
+
+static void
+lambda_vector_mult_const (lambda_vector vec1, lambda_vector vec2,
+ int size, int const1)
+{
+ int i;
+
+ if (const1 == 0)
+ lambda_vector_clear (vec2, size);
+ else
+ for (i = 0; i < size; i++)
+ vec2[i] = const1 * vec1[i];
+}
+
+/* Negate vector VEC1 with length SIZE and store it in VEC2. */
+
+static void
+lambda_vector_negate (lambda_vector vec1, lambda_vector vec2,
+ int size)
+{
+ lambda_vector_mult_const (vec1, vec2, size, -1);
+}
+
+/* Negate row R1 of matrix MAT which has N columns. */
+
+static void
+lambda_matrix_row_negate (lambda_matrix mat, int n, int r1)
+{
+ lambda_vector_negate (mat[r1], mat[r1], n);
+}
+
+/* Return true if two vectors are equal. */
+
+static bool
+lambda_vector_equal (lambda_vector vec1, lambda_vector vec2, int size)
+{
+ int i;
+ for (i = 0; i < size; i++)
+ if (vec1[i] != vec2[i])
+ return false;
+ return true;
+}
+
+/* Given an M x N integer matrix A, this function determines an M x
+ M unimodular matrix U, and an M x N echelon matrix S such that
+ "U.A = S". This decomposition is also known as "right Hermite".
+
+ Ref: Algorithm 2.1 page 33 in "Loop Transformations for
+ Restructuring Compilers" Utpal Banerjee. */
+
+static void
+lambda_matrix_right_hermite (lambda_matrix A, int m, int n,
+ lambda_matrix S, lambda_matrix U)
+{
+ int i, j, i0 = 0;
+
+ lambda_matrix_copy (A, S, m, n);
+ lambda_matrix_id (U, m);
+
+ for (j = 0; j < n; j++)
+ {
+ if (lambda_vector_first_nz (S[j], m, i0) < m)
+ {
+ ++i0;
+ for (i = m - 1; i >= i0; i--)
+ {
+ while (S[i][j] != 0)
+ {
+ int sigma, factor, a, b;
+
+ a = S[i-1][j];
+ b = S[i][j];
+ sigma = (a * b < 0) ? -1: 1;
+ a = abs (a);
+ b = abs (b);
+ factor = sigma * (a / b);
+
+ lambda_matrix_row_add (S, n, i, i-1, -factor);
+ lambda_matrix_row_exchange (S, i, i-1);
+
+ lambda_matrix_row_add (U, m, i, i-1, -factor);
+ lambda_matrix_row_exchange (U, i, i-1);
+ }
+ }
+ }
+ }
+}
+
/* Determines the overlapping elements due to accesses CHREC_A and
CHREC_B, that are affine functions. This function cannot handle
symbolic evolution functions, ie. when initial conditions are
#define GCC_TREE_DATA_REF_H
#include "graphds.h"
-#include "lambda.h"
#include "omega.h"
#include "tree-chrec.h"
bitmap vops;
};
+/* An integer vector. A vector formally consists of an element of a vector
+ space. A vector space is a set that is closed under vector addition
+ and scalar multiplication. In this vector space, an element is a list of
+ integers. */
+typedef int *lambda_vector;
+DEF_VEC_P(lambda_vector);
+DEF_VEC_ALLOC_P(lambda_vector,heap);
+DEF_VEC_ALLOC_P(lambda_vector,gc);
+
+/* An integer matrix. A matrix consists of m vectors of length n (IE
+ all vectors are the same length). */
+typedef lambda_vector *lambda_matrix;
+
/* Each vector of the access matrix represents a linear access
function for a subscript. First elements correspond to the
leftmost indices, ie. for a[i][j] the first vector corresponds to
return false;
}
+/* Returns the dependence level for a vector DIST of size LENGTH.
+ LEVEL = 0 means a lexicographic dependence, i.e. a dependence due
+ to the sequence of statements, not carried by any loop. */
+
+static inline unsigned
+dependence_level (lambda_vector dist_vect, int length)
+{
+ int i;
+
+ for (i = 0; i < length; i++)
+ if (dist_vect[i] != 0)
+ return i + 1;
+
+ return 0;
+}
+
/* Return the dependence level for the DDR relation. */
static inline unsigned
RDG_STMT (rdg, v2));
}
-/* In lambda-code.c */
-bool lambda_transform_legal_p (lambda_trans_matrix, int,
- VEC (ddr_p, heap) *);
-void lambda_collect_parameters (VEC (data_reference_p, heap) *,
- VEC (tree, heap) **);
-bool lambda_compute_access_matrices (VEC (data_reference_p, heap) *,
- VEC (tree, heap) *,
- VEC (loop_p, heap) *,
- struct obstack *);
-
/* In tree-data-ref.c */
void split_constant_offset (tree , tree *, tree *);
DEF_VEC_P (bitmap);
DEF_VEC_ALLOC_P (bitmap, heap);
+/* Compute the greatest common divisor of a VECTOR of SIZE numbers. */
+
+static inline int
+lambda_vector_gcd (lambda_vector vector, int size)
+{
+ int i;
+ int gcd1 = 0;
+
+ if (size > 0)
+ {
+ gcd1 = vector[0];
+ for (i = 1; i < size; i++)
+ gcd1 = gcd (gcd1, vector[i]);
+ }
+ return gcd1;
+}
+
+/* Allocate a new vector of given SIZE. */
+
+static inline lambda_vector
+lambda_vector_new (int size)
+{
+ return (lambda_vector) ggc_alloc_cleared_atomic (sizeof (int) * size);
+}
+
+/* Clear out vector VEC1 of length SIZE. */
+
+static inline void
+lambda_vector_clear (lambda_vector vec1, int size)
+{
+ memset (vec1, 0, size * sizeof (*vec1));
+}
+
+/* Returns true when the vector V is lexicographically positive, in
+ other words, when the first nonzero element is positive. */
+
+static inline bool
+lambda_vector_lexico_pos (lambda_vector v,
+ unsigned n)
+{
+ unsigned i;
+ for (i = 0; i < n; i++)
+ {
+ if (v[i] == 0)
+ continue;
+ if (v[i] < 0)
+ return false;
+ if (v[i] > 0)
+ return true;
+ }
+ return true;
+}
+
+/* Return true if vector VEC1 of length SIZE is the zero vector. */
+
+static inline bool
+lambda_vector_zerop (lambda_vector vec1, int size)
+{
+ int i;
+ for (i = 0; i < size; i++)
+ if (vec1[i] != 0)
+ return false;
+ return true;
+}
+
+/* Allocate a matrix of M rows x N cols. */
+
+static inline lambda_matrix
+lambda_matrix_new (int m, int n, struct obstack *lambda_obstack)
+{
+ lambda_matrix mat;
+ int i;
+
+ mat = (lambda_matrix) obstack_alloc (lambda_obstack,
+ sizeof (lambda_vector *) * m);
+
+ for (i = 0; i < m; i++)
+ mat[i] = lambda_vector_new (n);
+
+ return mat;
+}
+
#endif /* GCC_TREE_DATA_REF_H */
void swap_tree_operands (gimple, tree *, tree *);
-int least_common_multiple (int, int);
-
#endif /* _TREE_FLOW_H */
+++ /dev/null
-/* Linear Loop transforms
- Copyright (C) 2003, 2004, 2005, 2007, 2008, 2009, 2010
- Free Software Foundation, Inc.
- Contributed by Daniel Berlin <dberlin@dberlin.org>.
-
-This file is part of GCC.
-
-GCC is free software; you can redistribute it and/or modify it under
-the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 3, or (at your option) any later
-version.
-
-GCC is distributed in the hope that it will be useful, but WITHOUT ANY
-WARRANTY; without even the implied warranty of MERCHANTABILITY or
-FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
-for more details.
-
-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/>. */
-
-#include "config.h"
-#include "system.h"
-#include "coretypes.h"
-#include "tree-flow.h"
-#include "cfgloop.h"
-#include "tree-chrec.h"
-#include "tree-data-ref.h"
-#include "tree-scalar-evolution.h"
-#include "tree-pass.h"
-#include "lambda.h"
-
-/* Linear loop transforms include any composition of interchange,
- scaling, skewing, and reversal. They are used to change the
- iteration order of loop nests in order to optimize data locality of
- traversals, or remove dependences that prevent
- parallelization/vectorization/etc.
-
- TODO: Determine reuse vectors/matrix and use it to determine optimal
- transform matrix for locality purposes.
- TODO: Completion of partial transforms. */
-
-/* Gather statistics for loop interchange. LOOP is the loop being
- considered. The first loop in the considered loop nest is
- FIRST_LOOP, and consequently, the index of the considered loop is
- obtained by LOOP->DEPTH - FIRST_LOOP->DEPTH
-
- Initializes:
- - DEPENDENCE_STEPS the sum of all the data dependence distances
- carried by loop LOOP,
-
- - NB_DEPS_NOT_CARRIED_BY_LOOP the number of dependence relations
- for which the loop LOOP is not carrying any dependence,
-
- - ACCESS_STRIDES the sum of all the strides in LOOP.
-
- Example: for the following loop,
-
- | loop_1 runs 1335 times
- | loop_2 runs 1335 times
- | A[{{0, +, 1}_1, +, 1335}_2]
- | B[{{0, +, 1}_1, +, 1335}_2]
- | endloop_2
- | A[{0, +, 1336}_1]
- | endloop_1
-
- gather_interchange_stats (in loop_1) will return
- DEPENDENCE_STEPS = 3002
- NB_DEPS_NOT_CARRIED_BY_LOOP = 5
- ACCESS_STRIDES = 10694
-
- gather_interchange_stats (in loop_2) will return
- DEPENDENCE_STEPS = 3000
- NB_DEPS_NOT_CARRIED_BY_LOOP = 7
- ACCESS_STRIDES = 8010
-*/
-
-static void
-gather_interchange_stats (VEC (ddr_p, heap) *dependence_relations ATTRIBUTE_UNUSED,
- VEC (data_reference_p, heap) *datarefs ATTRIBUTE_UNUSED,
- struct loop *loop ATTRIBUTE_UNUSED,
- struct loop *first_loop ATTRIBUTE_UNUSED,
- unsigned int *dependence_steps ATTRIBUTE_UNUSED,
- unsigned int *nb_deps_not_carried_by_loop ATTRIBUTE_UNUSED,
- double_int *access_strides ATTRIBUTE_UNUSED)
-{
- unsigned int i, j;
- struct data_dependence_relation *ddr;
- struct data_reference *dr;
-
- *dependence_steps = 0;
- *nb_deps_not_carried_by_loop = 0;
- *access_strides = double_int_zero;
-
- FOR_EACH_VEC_ELT (ddr_p, dependence_relations, i, ddr)
- {
- /* If we don't know anything about this dependence, or the distance
- vector is NULL, or there is no dependence, then there is no reuse of
- data. */
- if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know
- || DDR_ARE_DEPENDENT (ddr) == chrec_known
- || DDR_NUM_DIST_VECTS (ddr) == 0)
- continue;
-
- for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++)
- {
- int dist = DDR_DIST_VECT (ddr, j)[loop_depth (loop) - loop_depth (first_loop)];
-
- if (dist == 0)
- (*nb_deps_not_carried_by_loop) += 1;
-
- else if (dist < 0)
- (*dependence_steps) += -dist;
-
- else
- (*dependence_steps) += dist;
- }
- }
-
- /* Compute the access strides. */
- FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
- {
- unsigned int it;
- tree ref = DR_REF (dr);
- gimple stmt = DR_STMT (dr);
- struct loop *stmt_loop = loop_containing_stmt (stmt);
- struct loop *inner_loop = first_loop->inner;
-
- if (inner_loop != stmt_loop
- && !flow_loop_nested_p (inner_loop, stmt_loop))
- continue;
-
- for (it = 0; it < DR_NUM_DIMENSIONS (dr);
- it++, ref = TREE_OPERAND (ref, 0))
- {
- int num = am_vector_index_for_loop (DR_ACCESS_MATRIX (dr), loop->num);
- int istride = AM_GET_ACCESS_MATRIX_ELEMENT (DR_ACCESS_MATRIX (dr), it, num);
- tree array_size = TYPE_SIZE (TREE_TYPE (ref));
- double_int dstride;
-
- if (array_size == NULL_TREE
- || TREE_CODE (array_size) != INTEGER_CST)
- continue;
-
- dstride = double_int_mul (tree_to_double_int (array_size),
- shwi_to_double_int (istride));
- (*access_strides) = double_int_add (*access_strides, dstride);
- }
- }
-}
-
-/* Attempt to apply interchange transformations to TRANS to maximize the
- spatial and temporal locality of the loop.
- Returns the new transform matrix. The smaller the reuse vector
- distances in the inner loops, the fewer the cache misses.
- FIRST_LOOP is the loop->num of the first loop in the analyzed loop
- nest. */
-
-
-static lambda_trans_matrix
-try_interchange_loops (lambda_trans_matrix trans,
- unsigned int depth,
- VEC (ddr_p, heap) *dependence_relations,
- VEC (data_reference_p, heap) *datarefs,
- struct loop *first_loop)
-{
- bool res;
- struct loop *loop_i;
- struct loop *loop_j;
- unsigned int dependence_steps_i, dependence_steps_j;
- double_int access_strides_i, access_strides_j;
- double_int small, large, nb_iter;
- double_int l1_cache_size, l2_cache_size;
- int cmp;
- unsigned int nb_deps_not_carried_by_i, nb_deps_not_carried_by_j;
- struct data_dependence_relation *ddr;
-
- if (VEC_length (ddr_p, dependence_relations) == 0)
- return trans;
-
- /* When there is an unknown relation in the dependence_relations, we
- know that it is no worth looking at this loop nest: give up. */
- ddr = VEC_index (ddr_p, dependence_relations, 0);
- if (ddr == NULL || DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
- return trans;
-
- l1_cache_size = uhwi_to_double_int (L1_CACHE_SIZE * 1024);
- l2_cache_size = uhwi_to_double_int (L2_CACHE_SIZE * 1024);
-
- /* LOOP_I is always the outer loop. */
- for (loop_j = first_loop->inner;
- loop_j;
- loop_j = loop_j->inner)
- for (loop_i = first_loop;
- loop_depth (loop_i) < loop_depth (loop_j);
- loop_i = loop_i->inner)
- {
- gather_interchange_stats (dependence_relations, datarefs,
- loop_i, first_loop,
- &dependence_steps_i,
- &nb_deps_not_carried_by_i,
- &access_strides_i);
- gather_interchange_stats (dependence_relations, datarefs,
- loop_j, first_loop,
- &dependence_steps_j,
- &nb_deps_not_carried_by_j,
- &access_strides_j);
-
- /* Heuristics for loop interchange profitability:
-
- 0. Don't transform if the smallest stride is larger than
- the L2 cache, or if the largest stride multiplied by the
- number of iterations is smaller than the L1 cache.
-
- 1. (spatial locality) Inner loops should have smallest
- dependence steps.
-
- 2. (spatial locality) Inner loops should contain more
- dependence relations not carried by the loop.
-
- 3. (temporal locality) Inner loops should have smallest
- array access strides.
- */
-
- cmp = double_int_ucmp (access_strides_i, access_strides_j);
- small = cmp < 0 ? access_strides_i : access_strides_j;
- large = cmp < 0 ? access_strides_j : access_strides_i;
-
- if (double_int_ucmp (small, l2_cache_size) > 0)
- continue;
-
- res = cmp < 0 ?
- estimated_loop_iterations (loop_j, false, &nb_iter):
- estimated_loop_iterations (loop_i, false, &nb_iter);
-
- if (res
- && double_int_ucmp (double_int_mul (large, nb_iter),
- l1_cache_size) < 0)
- continue;
-
- if (dependence_steps_i < dependence_steps_j
- || nb_deps_not_carried_by_i > nb_deps_not_carried_by_j
- || cmp < 0)
- {
- lambda_matrix_row_exchange (LTM_MATRIX (trans),
- loop_depth (loop_i) - loop_depth (first_loop),
- loop_depth (loop_j) - loop_depth (first_loop));
- /* Validate the resulting matrix. When the transformation
- is not valid, reverse to the previous transformation. */
- if (!lambda_transform_legal_p (trans, depth, dependence_relations))
- lambda_matrix_row_exchange (LTM_MATRIX (trans),
- loop_depth (loop_i) - loop_depth (first_loop),
- loop_depth (loop_j) - loop_depth (first_loop));
- }
- }
-
- return trans;
-}
-
-/* Return the number of nested loops in LOOP_NEST, or 0 if the loops
- are not perfectly nested. */
-
-unsigned int
-perfect_loop_nest_depth (struct loop *loop_nest)
-{
- struct loop *temp;
- unsigned int depth = 1;
-
- /* If it's not a loop nest, we don't want it. We also don't handle
- sibling loops properly, which are loops of the following form:
-
- | for (i = 0; i < 50; i++)
- | {
- | for (j = 0; j < 50; j++)
- | {
- | ...
- | }
- | for (j = 0; j < 50; j++)
- | {
- | ...
- | }
- | }
- */
-
- if (!loop_nest->inner || !single_exit (loop_nest))
- return 0;
-
- for (temp = loop_nest->inner; temp; temp = temp->inner)
- {
- /* If we have a sibling loop or multiple exit edges, jump ship. */
- if (temp->next || !single_exit (temp))
- return 0;
-
- depth++;
- }
-
- return depth;
-}
-
-/* Perform a set of linear transforms on loops. */
-
-void
-linear_transform_loops (void)
-{
- bool modified = false;
- loop_iterator li;
- VEC(tree,heap) *oldivs = NULL;
- VEC(tree,heap) *invariants = NULL;
- VEC(tree,heap) *lambda_parameters = NULL;
- VEC(gimple,heap) *remove_ivs = VEC_alloc (gimple, heap, 3);
- struct loop *loop_nest;
- gimple oldiv_stmt;
- unsigned i;
-
- FOR_EACH_LOOP (li, loop_nest, 0)
- {
- unsigned int depth = 0;
- VEC (ddr_p, heap) *dependence_relations;
- VEC (data_reference_p, heap) *datarefs;
-
- lambda_loopnest before, after;
- lambda_trans_matrix trans;
- struct obstack lambda_obstack;
- struct loop *loop;
- VEC (loop_p, heap) *nest;
- VEC (loop_p, heap) *ln;
-
- depth = perfect_loop_nest_depth (loop_nest);
- if (depth == 0)
- continue;
-
- nest = VEC_alloc (loop_p, heap, 3);
- for (loop = loop_nest; loop; loop = loop->inner)
- VEC_safe_push (loop_p, heap, nest, loop);
-
- gcc_obstack_init (&lambda_obstack);
- VEC_truncate (tree, oldivs, 0);
- VEC_truncate (tree, invariants, 0);
- VEC_truncate (tree, lambda_parameters, 0);
-
- datarefs = VEC_alloc (data_reference_p, heap, 10);
- dependence_relations = VEC_alloc (ddr_p, heap, 10 * 10);
- ln = VEC_alloc (loop_p, heap, 3);
- if (!compute_data_dependences_for_loop (loop_nest, true, &ln, &datarefs,
- &dependence_relations))
- goto free_and_continue;
-
- lambda_collect_parameters (datarefs, &lambda_parameters);
- if (!lambda_compute_access_matrices (datarefs, lambda_parameters,
- nest, &lambda_obstack))
- goto free_and_continue;
-
- if (dump_file && (dump_flags & TDF_DETAILS))
- dump_ddrs (dump_file, dependence_relations);
-
- /* Build the transformation matrix. */
- trans = lambda_trans_matrix_new (depth, depth, &lambda_obstack);
- lambda_matrix_id (LTM_MATRIX (trans), depth);
- trans = try_interchange_loops (trans, depth, dependence_relations,
- datarefs, loop_nest);
-
- if (lambda_trans_matrix_id_p (trans))
- {
- if (dump_file)
- fprintf (dump_file, "Won't transform loop. Optimal transform is the identity transform\n");
- goto free_and_continue;
- }
-
- /* Check whether the transformation is legal. */
- if (!lambda_transform_legal_p (trans, depth, dependence_relations))
- {
- if (dump_file)
- fprintf (dump_file, "Can't transform loop, transform is illegal:\n");
- goto free_and_continue;
- }
-
- before = gcc_loopnest_to_lambda_loopnest (loop_nest, &oldivs,
- &invariants, &lambda_obstack);
-
- if (!before)
- goto free_and_continue;
-
- if (dump_file)
- {
- fprintf (dump_file, "Before:\n");
- print_lambda_loopnest (dump_file, before, 'i');
- }
-
- after = lambda_loopnest_transform (before, trans, &lambda_obstack);
-
- if (dump_file)
- {
- fprintf (dump_file, "After:\n");
- print_lambda_loopnest (dump_file, after, 'u');
- }
-
- lambda_loopnest_to_gcc_loopnest (loop_nest, oldivs, invariants,
- &remove_ivs,
- after, trans, &lambda_obstack);
- modified = true;
-
- if (dump_file)
- fprintf (dump_file, "Successfully transformed loop.\n");
-
- free_and_continue:
- obstack_free (&lambda_obstack, NULL);
- free_dependence_relations (dependence_relations);
- free_data_refs (datarefs);
- VEC_free (loop_p, heap, nest);
- VEC_free (loop_p, heap, ln);
- }
-
- FOR_EACH_VEC_ELT (gimple, remove_ivs, i, oldiv_stmt)
- remove_iv (oldiv_stmt);
-
- VEC_free (tree, heap, oldivs);
- VEC_free (tree, heap, invariants);
- VEC_free (gimple, heap, remove_ivs);
- scev_reset ();
-
- if (modified)
- rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa_full_phi);
-}
return (hashval_t) a->version;
}
+/* A transformation matrix, which is a self-contained ROWSIZE x COLSIZE
+ matrix. Rather than use floats, we simply keep a single DENOMINATOR that
+ represents the denominator for every element in the matrix. */
+typedef struct lambda_trans_matrix_s
+{
+ lambda_matrix matrix;
+ int rowsize;
+ int colsize;
+ int denominator;
+} *lambda_trans_matrix;
+#define LTM_MATRIX(T) ((T)->matrix)
+#define LTM_ROWSIZE(T) ((T)->rowsize)
+#define LTM_COLSIZE(T) ((T)->colsize)
+#define LTM_DENOMINATOR(T) ((T)->denominator)
+
+/* Allocate a new transformation matrix. */
+
+static lambda_trans_matrix
+lambda_trans_matrix_new (int colsize, int rowsize,
+ struct obstack * lambda_obstack)
+{
+ lambda_trans_matrix ret;
+
+ ret = (lambda_trans_matrix)
+ obstack_alloc (lambda_obstack, sizeof (struct lambda_trans_matrix_s));
+ LTM_MATRIX (ret) = lambda_matrix_new (rowsize, colsize, lambda_obstack);
+ LTM_ROWSIZE (ret) = rowsize;
+ LTM_COLSIZE (ret) = colsize;
+ LTM_DENOMINATOR (ret) = 1;
+ return ret;
+}
+
+/* Multiply a vector VEC by a matrix MAT.
+ MAT is an M*N matrix, and VEC is a vector with length N. The result
+ is stored in DEST which must be a vector of length M. */
+
+static void
+lambda_matrix_vector_mult (lambda_matrix matrix, int m, int n,
+ lambda_vector vec, lambda_vector dest)
+{
+ int i, j;
+
+ lambda_vector_clear (dest, m);
+ for (i = 0; i < m; i++)
+ for (j = 0; j < n; j++)
+ dest[i] += matrix[i][j] * vec[j];
+}
+
+/* Return true if TRANS is a legal transformation matrix that respects
+ the dependence vectors in DISTS and DIRS. The conservative answer
+ is false.
+
+ "Wolfe proves that a unimodular transformation represented by the
+ matrix T is legal when applied to a loop nest with a set of
+ lexicographically non-negative distance vectors RDG if and only if
+ for each vector d in RDG, (T.d >= 0) is lexicographically positive.
+ i.e.: if and only if it transforms the lexicographically positive
+ distance vectors to lexicographically positive vectors. Note that
+ a unimodular matrix must transform the zero vector (and only it) to
+ the zero vector." S.Muchnick. */
+
+static bool
+lambda_transform_legal_p (lambda_trans_matrix trans,
+ int nb_loops,
+ VEC (ddr_p, heap) *dependence_relations)
+{
+ unsigned int i, j;
+ lambda_vector distres;
+ struct data_dependence_relation *ddr;
+
+ gcc_assert (LTM_COLSIZE (trans) == nb_loops
+ && LTM_ROWSIZE (trans) == nb_loops);
+
+ /* When there are no dependences, the transformation is correct. */
+ if (VEC_length (ddr_p, dependence_relations) == 0)
+ return true;
+
+ ddr = VEC_index (ddr_p, dependence_relations, 0);
+ if (ddr == NULL)
+ return true;
+
+ /* When there is an unknown relation in the dependence_relations, we
+ know that it is no worth looking at this loop nest: give up. */
+ if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
+ return false;
+
+ distres = lambda_vector_new (nb_loops);
+
+ /* For each distance vector in the dependence graph. */
+ FOR_EACH_VEC_ELT (ddr_p, dependence_relations, i, ddr)
+ {
+ /* Don't care about relations for which we know that there is no
+ dependence, nor about read-read (aka. output-dependences):
+ these data accesses can happen in any order. */
+ if (DDR_ARE_DEPENDENT (ddr) == chrec_known
+ || (DR_IS_READ (DDR_A (ddr)) && DR_IS_READ (DDR_B (ddr))))
+ continue;
+
+ /* Conservatively answer: "this transformation is not valid". */
+ if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
+ return false;
+
+ /* If the dependence could not be captured by a distance vector,
+ conservatively answer that the transform is not valid. */
+ if (DDR_NUM_DIST_VECTS (ddr) == 0)
+ return false;
+
+ /* Compute trans.dist_vect */
+ for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++)
+ {
+ lambda_matrix_vector_mult (LTM_MATRIX (trans), nb_loops, nb_loops,
+ DDR_DIST_VECT (ddr, j), distres);
+
+ if (!lambda_vector_lexico_pos (distres, nb_loops))
+ return false;
+ }
+ }
+ return true;
+}
/* Data dependency analysis. Returns true if the iterations of LOOP
are independent on each other (that is, if we can execute them
/* Insert PHI nodes everywhere they are needed. No pruning of the
IDF is done. This is used by passes that need the PHI nodes for
O_j even if it means that some arguments will come from the default
- definition of O_j's symbol (e.g., pass_linear_transform).
+ definition of O_j's symbol.
WARNING: If you need to use this flag, chances are that your pass
may be doing something wrong. Inserting PHI nodes for an old name
extern struct gimple_opt_pass pass_rest_of_compilation;
extern struct gimple_opt_pass pass_sink_code;
extern struct gimple_opt_pass pass_fre;
-extern struct gimple_opt_pass pass_linear_transform;
extern struct gimple_opt_pass pass_check_data_deps;
extern struct gimple_opt_pass pass_copy_prop;
extern struct gimple_opt_pass pass_vrp;
}
};
-/* Loop nest optimizations. */
-
-static unsigned int
-tree_linear_transform (void)
-{
- if (number_of_loops () <= 1)
- return 0;
-
- linear_transform_loops ();
- return 0;
-}
-
-static bool
-gate_tree_linear_transform (void)
-{
- return flag_tree_loop_linear != 0;
-}
-
-struct gimple_opt_pass pass_linear_transform =
-{
- {
- GIMPLE_PASS,
- "ltrans", /* name */
- gate_tree_linear_transform, /* gate */
- tree_linear_transform, /* execute */
- NULL, /* sub */
- NULL, /* next */
- 0, /* static_pass_number */
- TV_TREE_LINEAR_TRANSFORM, /* tv_id */
- PROP_cfg | PROP_ssa, /* properties_required */
- 0, /* properties_provided */
- 0, /* properties_destroyed */
- 0, /* todo_flags_start */
- TODO_dump_func
- | TODO_update_ssa_only_virtuals
- | TODO_ggc_collect /* todo_flags_finish */
- }
-};
-
/* GRAPHITE optimizations. */
static unsigned int
is turned on. */
if (flag_loop_block
|| flag_loop_interchange
+ || flag_tree_loop_linear
|| flag_loop_strip_mine
|| flag_graphite_identity
|| flag_loop_parallelize_all
|| flag_loop_flatten)
flag_graphite = 1;
+ /* Make flag_tree_loop_linear an alias of flag_loop_interchange. */
+ if (flag_tree_loop_linear)
+ flag_loop_interchange = flag_tree_loop_linear;
+
return flag_graphite != 0;
}
projects / platform / upstream / gcc.git / commitdiff