#include "config.h"#include "system.h"#include "coretypes.h"#include "tm.h"#include "rtl.h"#include "hard-reg-set.h"#include "obstack.h"#include "basic-block.h"#include "cfgloop.h"#include "expr.h"#include "intl.h"#include "diagnostic-core.h"#include "df.h"#include "hash-table.h"#include "dumpfile.h"
Data Structures | |
| struct | biv_entry |
| struct | biv_entry_hasher |
Macros | |
| #define | DF_REF_IV(REF) iv_ref_table[DF_REF_ID (REF)] |
| #define | DF_REF_IV_SET(REF, IV) iv_ref_table[DF_REF_ID (REF)] = (IV) |
Enumerations | |
| enum | iv_grd_result { GRD_INVALID, GRD_INVARIANT, GRD_MAYBE_BIV, GRD_SINGLE_DOM } |
Variables | |
| static bool | clean_slate = true |
| static unsigned int | iv_ref_table_size = 0 |
| static struct rtx_iv ** | iv_ref_table |
| static struct loop * | current_loop |
| static hash_table < biv_entry_hasher > | bivs |
Macro Definition Documentation
| #define DF_REF_IV | ( | REF | ) | iv_ref_table[DF_REF_ID (REF)] |
Induction variable stored at the reference.
| #define DF_REF_IV_SET | ( | REF, | |
| IV | |||
| ) | iv_ref_table[DF_REF_ID (REF)] = (IV) |
Enumeration Type Documentation
| enum iv_grd_result |
Rtl-level induction variable analysis. Copyright (C) 2004-2013 Free Software Foundation, Inc.
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/. This is a simple analysis of induction variables of the loop. The major use is for determining the number of iterations of a loop for loop unrolling, doloop optimization and branch prediction. The iv information is computed on demand.
Induction variables are analyzed by walking the use-def chains. When a basic induction variable (biv) is found, it is cached in the bivs hash table. When register is proved to be a biv, its description is stored to DF_REF_DATA of the def reference.
The analysis works always with one loop – you must call iv_analysis_loop_init (loop) for it. All the other functions then work with this loop. When you need to work with another loop, just call iv_analysis_loop_init for it. When you no longer need iv analysis, call iv_analysis_done () to clean up the memory.
The available functions are:
iv_analyze (insn, reg, iv): Stores the description of the induction variable corresponding to the use of register REG in INSN to IV. Returns true if REG is an induction variable in INSN. false otherwise. If use of REG is not found in INSN, following insns are scanned (so that we may call this function on insn returned by get_condition). iv_analyze_result (insn, def, iv): Stores to IV the description of the iv corresponding to DEF, which is a register defined in INSN. iv_analyze_expr (insn, rhs, mode, iv): Stores to IV the description of iv corresponding to expression EXPR evaluated at INSN. All registers used bu EXPR must also be used in INSN. Possible return values of iv_get_reaching_def.
- Enumerator:
Function Documentation
|
static |
Checks whether register *REG is in set ALT. Callback for for_each_rtx.
Referenced by canon_condition().
|
static |
If DEF was already analyzed for bivness, store the description of the biv to IV and return true. Otherwise return false.
| bool biv_p | ( | ) |
Checks whether definition of register REG in INSN is a basic induction variable. IV analysis must have been initialized (via a call to iv_analysis_loop_init) for this function to produce a result.
References clean_slate, clear_iv_info(), df_finish_pass(), hash_table< Descriptor, Allocator >::dispose(), iv_ref_table_size, and NULL.
| rtx canon_condition | ( | ) |
Canonicalizes COND so that
(1) Ensure that operands are ordered according to swap_commutative_operands_p. (2) (LE x const) will be replaced with (LT x <const+1>) and similarly for GE, GEU, and LEU.
When cross-compiling, const_val might be sign-extended from BITS_PER_WORD to HOST_BITS_PER_WIDE_INT
References altered_reg_used(), COMPARISON_P, const0_rtx, const_true_rtx, CONSTANT_P, exp(), for_each_rtx(), GET_CODE, implies_p(), REG_P, reversed_condition(), rtx_equal_p(), simplify_replace_rtx(), and XEXP.
|
static |
Transforms IV0 and IV1 compared by COND so that they are both compared as subregs of the same mode if possible (sometimes it is necessary to add some assumptions to DESC).
If the ivs behave specially in the first iteration, or are added/multiplied after extending, we ignore them.
If there is some extend, it must match signedness of the comparison.
Values of both variables should be computed in the same mode. These might indeed be different, if we have comparison like (compare (subreg:SI (iv0)) (subreg:SI (iv1))) and iv0 and iv1 are both ivs iterating in SI mode, but calculated in different modes. This does not seem impossible to handle, but it hardly ever occurs in practice. The only exception is the case when one of operands is invariant. For example pentium 3 generates comparisons like (lt (subreg:HI (reg:SI)) 100). Here we assign HImode to 100, but we definitely do not want this prevent the optimization.
Check that both ivs belong to a range of a single mode. If one of the operands is an invariant, we may need to shorten it into the common mode.
|
static |
References subreg_lowpart_p(), and SUBREG_REG.
|
static |
Checks whether E is a simple exit from LOOP and stores its description into DESC.
It must belong directly to the loop.
It must be tested (at least) once during any iteration.
It must end in a simple conditional jump.
Test whether the condition is suitable.
Check that we are able to determine number of iterations and fill in information about it.
References niter_desc::assumptions, find_simple_exit(), gettext, niter_desc::infinite, iv_analysis_loop_init(), N_, NULL_RTX, loop::simple_loop_desc, simple_loop_desc(), niter_desc::simple_p, and warning().
|
static |
Clears the information about ivs stored in df.
References BITMAP_ALLOC, clean_slate, hash_table< Descriptor, Allocator >::create(), DF_DEFER_INSN_RESCAN, DF_EQ_NOTES, df_set_flags(), get_loop_body_in_dom_order(), NULL, and loop::num_nodes.
Referenced by biv_p().
|
static |
Tries to estimate the maximum number of iterations in LOOP, and return the result. This function is called from iv_number_of_iterations with a number of fields in DESC already filled in. OLD_NITER is the original expression for the number of iterations, before we tried to simplify it.
We used to look for constant operand 0 of AND, but canonicalization should always make this impossible.
We could use a binary search here, but for now improving the upper bound by just one eliminates one important corner case.
| void dump_iv_info | ( | FILE * | , |
| struct rtx_iv * | |||
| ) |
Dumps information about IV to FILE.
| void dump_iv_info | ( | ) |
|
static |
Use relationship between A and *B to eventually eliminate *B. OP is the operation we consider.
If A implies *B, we may replace *B by true.
If *B implies A, we may replace *B by false.
|
static |
Eliminates the conditions in TAIL that are implied by HEAD. OP is the operation we consider.
| void find_simple_exit | ( | ) |
Finds a simple exit of LOOP and stores its description into DESC.
Prefer constant iterations; the less the better.
Also if the actual exit may be infinite, while the old one
not, prefer the old one.
References ggc_free(), NULL, loop::simple_loop_desc, and simple_loop_desc().
| void free_simple_loop_desc | ( | ) |
Releases simple loop description for LOOP.
|
static |
Gets the operation on register REG inside loop, in shape
OUTER_STEP + EXTEND_{OUTER_MODE} (SUBREG_{INNER_MODE} (REG + INNER_STEP))
If the operation cannot be described in this shape, return false. LAST_DEF is the definition of REG that dominates loop latch.
References hash_table< Descriptor, Allocator >::find_slot_with_hash(), gcc_assert, biv_entry::iv, biv_entry::regno, and REGNO.
|
static |
The recursive part of get_biv_step. Gets the value of the single value defined by DEF wrto initial value of REG inside loop, in shape described at get_biv_step.
ppc64 uses expressions like
(set x:SI (plus:SI (subreg:SI y:DI) 1)).
this is equivalent to
(set x':DI (plus:DI y:DI 1)) (set x:SI (subreg:SI (x':DI)).
See comment in previous switch.
References GET_CODE, GET_MODE, and SUBREG_REG.
| rtx get_iv_value | ( | ) |
Calculates value of IV at ITERATION-th iteration.
We would need to generate some if_then_else patterns, and so far it is not needed anywhere.
|
read |
Creates a simple loop description of LOOP if it was not computed already.
At least desc->infinite is not always initialized by find_simple_loop_exit.
Assume that no overflow happens and that the loop is finite.
We already warned at the tree level if we ran optimizations there.
|
static |
Checks whether A implies B.
A < B implies A + 1 <= B.
A < B or A > B imply A != B. TODO: Likewise A + n < B implies A != B + n if neither wraps.
For unsigned comparisons, A != 0 implies A > 0 and A >= 1.
A != N is equivalent to A - (N + 1) <u -1.
Avoid overflows.
Likewise, A != N implies A - N > 0.
Avoid overflows.
Avoid overflows.
A >s X, where X is positive, implies A <u Y, if Y is negative.
Referenced by canon_condition(), and simplify_using_condition().
|
static |
Computes inverse to X modulo (1 << MOD).
|
static |
Evaluates addition or subtraction (according to OP) of IV1 to IV0.
Extend the constant to extend_mode of the other operand if necessary.
Handle addition of constant.
| void iv_analysis_done | ( | void | ) |
Free the data for an induction variable analysis.
References function_invariant_p(), GET_CODE, HARD_REGISTER_P, REG_P, and XEXP.
| void iv_analysis_loop_init | ( | ) |
Prepare the data for an induction variable analysis of a LOOP.
Clear the information from the analysis of the previous loop.
Get rid of the ud chains before processing the rescans. Then add the problem back.
| bool iv_analyze | ( | ) |
Analyzes value VAL at INSN and stores the result to *IV.
We must find the insn in that val is used, so that we get to UD chains. Since the function is sometimes called on result of get_condition, this does not necessarily have to be directly INSN; scan also the following insns.
|
static |
Determines whether DEF is a biv and if so, stores its description to *IV.
Loop transforms base to es (base + inner_step) + outer_step, where es means extend of subreg between inner_mode and outer_mode. The corresponding induction variable is
es ((base - outer_step) + i * (inner_step + outer_step)) + outer_step
|
static |
Analyzes iv DEF and stores the result to *IV.
| bool iv_analyze_expr | ( | ) |
Analyzes expression RHS used at INSN and stores the result to *IV. The mode of the induction variable is MODE.
|
static |
Analyzes operand OP of INSN and stores the result to *IV.
| bool iv_analyze_result | ( | ) |
Analyzes definition of DEF in INSN and stores the result to IV.
|
static |
Sets IV to invariant CST in MODE. Always returns true (just for consistency with other iv manipulation functions that may fail).
|
static |
Evaluates application of EXTEND to MODE on IV.
If iv is invariant, just calculate the new value.
References rtx_iv::base, rtx_iv::delta, rtx_iv::extend, rtx_iv::extend_mode, IV_UNKNOWN_EXTEND, rtx_iv::mult, simplify_gen_unary(), and rtx_iv::step.
|
inlinestatic |
Return the RTX code corresponding to the IV extend code EXTEND.
|
static |
Gets definition of REG reaching its use in INSN and stores it to DEF.
More than one reaching def.
We do not handle setting only part of the register.
The definition does not dominate the use. This is still OK if this may be a use of a biv, i.e. if the def_bb dominates loop latch.
|
static |
Evaluates multiplication of IV by constant CST.
References rtx_iv::delta, rtx_iv::mult, and simplify_gen_binary().
|
static |
Evaluates negation of IV.
|
static |
Computes number of iterations of the CONDITION in INSN in LOOP and stores the result into DESC. Very similar to determine_number_of_iterations (basically its rtl version), complicated by things like subregs.
The meaning of these assumptions is this: if !assumptions then the rest of information does not have to be valid if noloop_assumptions then the loop does not roll if infinite then this exit is never used
The constant comparisons should be folded.
We only handle integers or pointers.
Check condition and normalize it.
Handle extends. This is relatively nontrivial, so we only try in some easy cases, when we can canonicalize the ivs (possibly by adding some assumptions) to shape subreg (base + i * step). This function also fills in desc->mode and desc->signed_p.
We can take care of the case of two induction variables chasing each other if the test is NE. I have never seen a loop using it, but still it is cool.
This is either infinite loop or the one that ends immediately, depending on initial values. Unswitching should remove this kind of conditions.
Ignore loops of while (i– < 10) type.
We do not care about whether the step is power of two in this
case. Some more condition normalization. We must record some assumptions due to overflows.
We want to take care only of non-sharp relationals; this is easy,
as in cases the overflow would make the transformation unsafe
the loop does not roll. Seemingly it would make more sense to want
to take care of sharp relationals instead, as NE is more similar to
them, but the problem is that here the transformation would be more
difficult due to possibly infinite loops. It will be useful to be able to tell the difference once more in
LE -> NE reduction. Take care of trivially infinite loops.
Fill in the remaining fields somehow.
Fill in the remaining fields somehow.
If we can we want to take care of NE conditions instead of size comparisons, as they are much more friendly (most importantly this takes care of special handling of loops with step 1). We can do it if we first check that upper bound is greater or equal to lower bound, their difference is constant c modulo step and that there is not an overflow.
A special case. We have transformed condition of type
for (i = 0; i < 4; i += 4)
into
for (i = 0; i <= 3; i += 4)
obviously if the test for overflow during that transformation
passed, we cannot overflow here. Most importantly any
loop with sharp end condition and step 1 falls into this
category, so handling this case specially is definitely
worth the troubles. We perform the transformation always provided that it is not
completely senseless. This is OK, as we would need this assumption
to determine the number of iterations anyway. If the step is a power of two and the final value we have
computed overflows, the cycle is infinite. Otherwise it
is nontrivial to compute the number of iterations. We are going to lose some information about upper bound on
number of iterations in this step, so record the information
here. Count the number of iterations.
Everything we do here is just arithmetics modulo size of mode. This
makes us able to do more involved computations of number of iterations
than in other cases. First transform the condition into shape
s * i <> c, with s positive. Let nsd (s, size of mode) = d. If d does not divide c, the loop
is infinite. Otherwise, the number of iterations is
(inverse(s/d) * (c/d)) mod (size of mode/d). Condition in shape a + s * i <= b
We must know that b + s does not overflow and a <= b + s and then we
can compute number of iterations as (b + s - a) / s. (It might
seem that we in fact could be more clever about testing the b + s
overflow condition using some information about b - a mod s,
but it was already taken into account during LE -> NE transform). If s is power of 2, we know that the loop is infinite if
a % s <= b % s and b + s overflows. Condition in shape a <= b - s * i
We must know that a - s does not overflow and a - s <= b and then
we can again compute number of iterations as (b - (a - s)) / s. If s is power of 2, we know that the loop is infinite if
a % s <= b % s and a - s overflows. Rerun the simplification. Consider code (created by copying loop headers)
i = 0;
if (0 < n)
{
do
{
i++;
} while (i < n);
}
The first pass determines that i = 0, the second pass uses it to eliminate
noloop assumption. simplify_using_initial_values does a copy propagation on the registers
in the expression for the number of iterations. This prolongs life
ranges of registers and increases register pressure, and usually
brings no gain (and if it happens to do, the cse pass will take care
of it anyway). So prevent this behavior, unless it enabled us to
derive that the number of iterations is a constant. Simplify the assumptions.
Fallthru.
|
static |
Evaluates shift of IV by constant CST.
|
static |
Evaluates application of subreg to MODE on IV.
If iv is invariant, just calculate the new value.
|
static |
Finds the definition of REG that dominates loop latch and stores it to DEF. Returns false if there is not a single definition dominating the latch. If REG has no definition in loop, DEF is set to NULL and true is returned.
More than one reaching definition.
Generates a subreg to get the least significant part of EXPR (in mode INNER_MODE) to OUTER_MODE.
|
static |
|
static |
References rtx_iv::base, dump_file, and NULL_RTX.
|
static |
Records information that DEF is induction variable IV.
References dump_file, and print_rtl().
|
static |
Using the data returned by suitable_set_for_replacement, replace DEST with SRC in *EXPR and return the new expression. Also call replace_single_def_regs if the replacement changed something.
|
static |
If REG has a single definition, replace it with its known value in EXPR. Callback for for_each_rtx.
|
static |
Transforms invariant IV into MODE. Adds assumptions based on the fact that IV occurs as left operands of comparison COND and its signedness is SIGNED_P to DESC.
|
static |
Checks whether REG is a well-behaved register.
|
static |
Checks whether RHS is simple enough to process.
Allow reg OP const and reg OP reg.
Allow reg OP const.
| void simplify_using_condition | ( | ) |
Tries to use the fact that COND holds to simplify EXPR. ALTERED is the set of altered regs.
If some register gets altered later, we do not really speak about its value at the time of comparison.
A proof by contradiction. If *EXPR implies (not cond), *EXPR must be false.
Similarly, If (not *EXPR) implies (not cond), *EXPR must be true.
We would like to have some other tests here. TODO.
References const0_rtx, const_true_rtx, gcc_unreachable, and implies_p().
|
static |
Simplifies *EXPR using initial values at the start of the LOOP. If *EXPR is a list, its elements are assumed to be combined using OP.
Kill all call clobbered registers.
We can no longer use a condition that has been simplified
to a constant, and simplify_using_condition will abort if
we try. Retry simplifications with this condition if either the
expression or the condition changed. If we did not use this insn to make a replacement, any overlap
between stores in this insn and our expression will cause the
expression to become invalid. Likewise for the conditions.
If the expression now contains regs that have been altered, we
can't return it to the caller. However, it is still valid for
further simplification, so keep searching to see if we can
eventually turn it into a constant.
|
static |
A subroutine of simplify_using_initial_values, this function examines INSN to see if it contains a suitable set that we can use to make a replacement. If it is suitable, return true and set DEST and SRC to the lhs and rhs of the set; return false otherwise.
References const_true_rtx, and simplify_replace_rtx().
Variable Documentation
|
static |
Bivs of the current loop.
Referenced by biv_p(), and clear_iv_info().
|
static |
The current loop.
|
static |
Table of rtx_ivs indexed by the df_ref uid field.
|
static |
Referenced by biv_p().
