[Mesa-dev] [PATCH] nir: Add a loop analysis pass
Jason Ekstrand
jason at jlekstrand.net
Thu Dec 15 00:28:34 UTC 2016
On Wed, Dec 14, 2016 at 1:35 PM, Timothy Arceri <
timothy.arceri at collabora.com> wrote:
> From: Thomas Helland <thomashelland90 at gmail.com>
>
> This pass detects induction variables and calculates the
> trip count of loops to be used for loop unrolling.
>
> I've removed support for float induction values for now, for the
> simple reason that they don't appear in my shader-db collection,
> and so I don't see it as common enough that we want to pollute the
> pass with this in the initial version.
>
> V2: Rebase, adapt to removal of function overloads
>
> V3: (Timothy Arceri)
> - don't try to find trip count if loop terminator conditional is a phi
> - fix trip count for do-while loops
> - replace conditional type != alu assert with return
> - disable unrolling of loops with continues
> - multiple fixes to memory allocation, stop leaking and don't destroy
> structs we want to use for unrolling.
> - fix iteration count bugs when induction var not on RHS of condition
> - add FIXME for && conditions
> - calculate trip count for unsigned induction/limit vars
>
> V4: (Timothy Arceri)
> - count instructions in a loop
> - set the limiting_terminator even if we can't find the trip count for
> all terminators. This is needed for complex unrolling where we handle
> 2 terminators and the trip count is unknown for one of them.
> - restruct structs so we don't keep information not required after
> analysis and remove dead fields.
> - force unrolling in some cases as per the rules in the GLSL IR pass
>
> V5: (Timothy Arceri)
> - fix metadata mask value 0x10 vs 0x16
>
> V6: (Timothy Arceri)
> - merge loop_variable and nir_loop_variable structs and lists suggested by
> Jason
> - remove induction var hash table and store pointer to induction
> information in
> the loop_variable suggested by Jason.
> - use lowercase list_addtail() suggested by Jason.
> - tidy up init_loop_block() as per Jasons suggestions.
> - replace switch with nir_op_infos[alu->op].num_inputs == 2 in
> is_var_basic_induction_var() as suggested by Jason.
> - use nir_block_last_instr() in and rename foreach_cf_node_ex_loop() as
> suggested
> by Jason.
> - fix else check for is_trivial_loop_terminator() as per Connors
> suggetions.
> - simplify offset for induction valiables incremented before the exit
> conditions is
> checked.
> - replace nir_op_isub check with assert() as it should have been lowered
> away.
>
> V7: (Timothy Arceri)
> - use rzalloc() on nir_loop struct creation. Worked previously because
> ralloc()
> was broken and always zeroed the struct.
> - fix cf_node_find_loop_jumps() to find jumps when loops contain
> nested if statements. Code is tidier as a result.
>
> V8: (Timothy Arceri)
> - move is_trivial_loop_terminator() to nir.h so we can use it to assert is
> the loop unroll pass
> - fix analysis to not bail when looking for terminator when the break is
> in the else
> rather then the if
> - added new loop terminator fields: break_block, continue_from_block and
> continue_from_then so we don't have to gather these when doing unrolling.
> - get correct array length when forcing unrolling of variables
> indexed arrays that are the same size as the iteration count
> - add support for induction variables of type float
> - update trival loop terminator check to allow an if containing
> instructions as long as both branches contain only a single
> block.
> ---
> src/compiler/Makefile.sources | 2 +
> src/compiler/nir/nir.c | 2 +-
> src/compiler/nir/nir.h | 78 ++-
> src/compiler/nir/nir_loop_analyze.c | 992 ++++++++++++++++++++++++++++++
> ++++++
> src/compiler/nir/nir_metadata.c | 8 +-
> 5 files changed, 1079 insertions(+), 3 deletions(-)
> create mode 100644 src/compiler/nir/nir_loop_analyze.c
>
> diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
> index 17b15de..ca8a056 100644
> --- a/src/compiler/Makefile.sources
> +++ b/src/compiler/Makefile.sources
> @@ -193,6 +193,8 @@ NIR_FILES = \
> nir/nir_intrinsics.c \
> nir/nir_intrinsics.h \
> nir/nir_liveness.c \
> + nir/nir_loop_analyze.c \
> + nir/nir_loop_analyze.h \
> nir/nir_lower_alu_to_scalar.c \
> nir/nir_lower_atomics.c \
> nir/nir_lower_bitmap.c \
> diff --git a/src/compiler/nir/nir.c b/src/compiler/nir/nir.c
> index 2d882f7..2c3531c 100644
> --- a/src/compiler/nir/nir.c
> +++ b/src/compiler/nir/nir.c
> @@ -393,7 +393,7 @@ nir_if_create(nir_shader *shader)
> nir_loop *
> nir_loop_create(nir_shader *shader)
> {
> - nir_loop *loop = ralloc(shader, nir_loop);
> + nir_loop *loop = rzalloc(shader, nir_loop);
>
> cf_init(&loop->cf_node, nir_cf_node_loop);
>
> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> index ba88a24..0831025 100644
> --- a/src/compiler/nir/nir.h
> +++ b/src/compiler/nir/nir.h
> @@ -1506,9 +1506,41 @@ typedef struct nir_if {
> } nir_if;
>
> typedef struct {
> + nir_if *nif;
> +
> + nir_instr *conditional_instr;
> +
> + nir_block *break_block;
> + nir_block *continue_from_block;
> +
> + bool continue_from_then;
> +
> + struct list_head loop_terminator_link;
> +} nir_loop_terminator;
> +
> +typedef struct {
> + /* Number of instructions in the loop */
> + unsigned num_instructions;
> +
> + /* How many times the loop is run (if known) */
> + unsigned trip_count;
> + bool is_trip_count_known;
> +
> + /* Unroll the loop regardless of its size */
> + bool force_unroll;
> +
> + nir_loop_terminator *limiting_terminator;
> +
> + /* A list of loop_terminators terminating this loop. */
> + struct list_head loop_terminator_list;
> +} nir_loop_info;
> +
> +typedef struct {
> nir_cf_node cf_node;
>
> struct exec_list body; /** < list of nir_cf_node */
> +
> + nir_loop_info *info;
> } nir_loop;
>
> /**
> @@ -1521,6 +1553,7 @@ typedef enum {
> nir_metadata_dominance = 0x2,
> nir_metadata_live_ssa_defs = 0x4,
> nir_metadata_not_properly_reset = 0x8,
> + nir_metadata_loop_analysis = 0x10,
> } nir_metadata;
>
> typedef struct {
> @@ -1653,6 +1686,43 @@ nir_loop_last_block(nir_loop *loop)
> return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
> }
>
> +static inline bool
> +nir_if_is_trivial_loop_terminator(nir_if *nif, nir_block *break_block,
> + nir_block *continue_from_block,
> + bool contine_from_then)
> +{
> +
> + /* Both branches must only contain a single block */
> + if (contine_from_then) {
> + if (continue_from_block != nir_if_first_then_block(nif))
> + return false;
>
Is this really needed? Why can't the continue block contain some control
flow so long as it doesn't contain any breaks?
> + if (break_block != nir_if_first_else_block(nif))
> + return false;
> + } else {
> + if (continue_from_block != nir_if_first_else_block(nif))
> + return false;
> + if (break_block != nir_if_first_then_block(nif))
> + return false;
> + }
> +
> + /* If both the then and else branches contain a jump deaf cf should
> have
> + * eliminated the loop already.
> + */
> + nir_foreach_instr(instr, continue_from_block)
> + assert(instr->type != nir_instr_type_jump);
> +
> + /* dead_cf should have eliminated any instruction after the first
> break */
> + nir_instr *last_instr = nir_block_last_instr(break_block);
> + (void) last_instr;
> + assert(last_instr && last_instr->type == nir_instr_type_jump &&
> + nir_instr_as_jump(last_instr)->type == nir_jump_break);
> +
> + nir_foreach_instr(instr, break_block)
> + assert(instr->type != nir_instr_type_jump || instr == last_instr);
> +
> + return true;
> +}
> +
> typedef enum {
> nir_parameter_in,
> nir_parameter_out,
> @@ -1749,6 +1819,8 @@ typedef struct nir_shader_compiler_options {
> * information must be inferred from the list of input nir_variables.
> */
> bool use_interpolated_input_intrinsics;
> +
> + unsigned max_unroll_iterations;
> } nir_shader_compiler_options;
>
> typedef struct nir_shader {
> @@ -1859,7 +1931,7 @@ nir_loop *nir_loop_create(nir_shader *shader);
> nir_function_impl *nir_cf_node_get_function(nir_cf_node *node);
>
> /** requests that the given pieces of metadata be generated */
> -void nir_metadata_require(nir_function_impl *impl, nir_metadata
> required);
> +void nir_metadata_require(nir_function_impl *impl, nir_metadata
> required, ...);
> /** dirties all but the preserved metadata */
> void nir_metadata_preserve(nir_function_impl *impl, nir_metadata
> preserved);
>
> @@ -2479,6 +2551,10 @@ void nir_lower_double_pack(nir_shader *shader);
> bool nir_normalize_cubemap_coords(nir_shader *shader);
>
> void nir_live_ssa_defs_impl(nir_function_impl *impl);
> +
> +void nir_loop_analyze_impl(nir_function_impl *impl,
> + nir_variable_mode indirect_mask);
> +
> bool nir_ssa_defs_interfere(nir_ssa_def *a, nir_ssa_def *b);
>
> void nir_convert_to_ssa_impl(nir_function_impl *impl);
> diff --git a/src/compiler/nir/nir_loop_analyze.c
> b/src/compiler/nir/nir_loop_analyze.c
> new file mode 100644
> index 0000000..3591bd7
> --- /dev/null
> +++ b/src/compiler/nir/nir_loop_analyze.c
> @@ -0,0 +1,992 @@
> +/*
> + * Copyright © 2015 Thomas Helland
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the
> "Software"),
> + * to deal in the Software without restriction, including without
> limitation
> + * the rights to use, copy, modify, merge, publish, distribute,
> sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice (including the
> next
> + * paragraph) shall be included in all copies or substantial portions of
> the
> + * Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
> EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
> MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT
> SHALL
> + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
> OTHER
> + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
> DEALINGS
> + * IN THE SOFTWARE.
> + */
> +
> +#include "nir.h"
> +
> +typedef enum {
> + undefined,
> + invariant,
> + basic_induction
> +} nir_loop_variable_type;
> +
> +struct nir_basic_induction_var;
> +
> +typedef struct {
> + /* A link for the work list */
> + struct list_head process_link;
> +
> + bool in_loop;
> +
> + /* The ssa_def associated with this info */
> + nir_ssa_def *def;
> +
> + /* The type of this ssa_def */
> + nir_loop_variable_type type;
> +
> + /* If this is of type basic_induction */
> + struct nir_basic_induction_var *ind;
> +
> + /* If the ssa-def is constant */
> + bool is_constant;
>
Would you mind replacing this with a var_is_const helper similar to
var_is_alu and var_is_phi? It seems bit redundant to have it here.
> +
> + bool in_conditional_block;
> +
> + bool in_nested_loop;
> +
> +} nir_loop_variable;
> +
> +typedef struct nir_basic_induction_var {
> + nir_op alu_op; /* The type of alu-operation
> */
> + nir_loop_variable *alu_def; /* The def of the
> alu-operation */
> + nir_loop_variable *invariant; /* The invariant alu-operand
> */
> + nir_loop_variable *def_outside_loop; /* The phi-src outside the
> loop */
> +} nir_basic_induction_var;
> +
> +typedef struct {
> + bool contains_break;
> + bool contains_continue;
> +} loop_jumps;
> +
> +typedef struct {
> + /* The loop we store information for */
> + nir_loop *loop;
> +
> + /* Loop_variable for all ssa_defs in function */
> + nir_loop_variable *loop_vars;
> +
> + /* A list of the loop_vars to analyze */
> + struct list_head process_list;
> +
> + nir_variable_mode indirect_mask;
> +
> +} loop_info_state;
> +
> +static nir_loop_variable *
> +get_loop_var(nir_ssa_def *value, loop_info_state *state)
> +{
> + return &(state->loop_vars[value->index]);
> +}
> +
> +typedef struct {
> + loop_info_state *state;
> + bool mark_nested;
> + bool mark_in_conditional;
> +} init_loop_state;
> +
> +static bool
> +init_loop_def(nir_ssa_def *def, void *void_init_loop_state)
> +{
> + init_loop_state *loop_init_state = void_init_loop_state;
> + nir_loop_variable *var = get_loop_var(def, loop_init_state->state);
> +
> + /* Add to the tail of the list. That way we start at the beginning of
> the
> + * defs in the loop instead of the end when walking the list. This
> means
> + * less recursive calls. Only add defs that are not in nested loops or
> + * conditional blocks.
> + */
> + if (!(loop_init_state->mark_in_conditional ||
> + loop_init_state->mark_nested)) {
> + list_addtail(&(var->process_link),
> + &(loop_init_state->state->process_list));
> + }
> +
> + if (loop_init_state->mark_in_conditional)
> + var->in_conditional_block = true;
> +
> + if (loop_init_state->mark_nested)
> + var->in_nested_loop = true;
> +
> + var->in_loop = true;
> +
> + return true;
> +}
> +
> +static bool
> +init_loop_block(nir_block *block, loop_info_state *state,
> + bool in_conditional, bool nested)
> +{
> + init_loop_state init_state = {.mark_in_conditional = in_conditional,
> + .mark_nested = nested, .state = state };
> +
> + nir_foreach_instr(instr, block)
> + nir_foreach_ssa_def(instr, init_loop_def, &init_state);
> +
> + return true;
> +}
> +
> +static inline bool
> +is_var_alu(nir_loop_variable *var)
> +{
> + return (var->def->parent_instr->type == nir_instr_type_alu);
> +}
> +
> +static inline bool
> +is_var_phi(nir_loop_variable *var)
> +{
> + return (var->def->parent_instr->type == nir_instr_type_phi);
> +}
> +
> +static inline bool
> +is_ssa_def_invariant(nir_ssa_def *def, loop_info_state *state)
> +{
> + nir_loop_variable *var = get_loop_var(def, state);
>
+
> + if (var->type == invariant)
> + return true;
>
I think this would be much more efficient if we also had a case for "we
know this isn't invariant". Without that...
> +
> + if (!var->in_loop) {
> + var->type = invariant;
> + return true;
> + }
> +
> + if (var->type == basic_induction)
> + return false;
> +
> + if (is_var_alu(var)) {
> + nir_alu_instr *alu = nir_instr_as_alu(def->parent_instr);
> +
> + for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
> + if (!is_ssa_def_invariant(alu->src[i].src.ssa, state))
> + return false;
>
...this will get executed a lot of times when it doesn't need to be.
> + }
> + var->type = invariant;
> + return true;
> + }
> +
> + /* Phis shouldn't be invariant except if one operand is invariant, and
> the
> + * other is the phi itself. These should be removed by opt_remove_phis.
> + * load_consts are already set to invariant and constant during init,
> + * and so should return earlier. Remaining op_codes are set undefined.
> + */
> + var->type = undefined;
> + return false;
> +}
> +
> +static void
> +compute_invariance_information(loop_info_state *state)
> +{
> + /* An expression is invariant in a loop L if:
> + * (base cases)
> + * – it’s a constant
> + * – it’s a variable use, all of whose single defs are outside of L
> + * (inductive cases)
> + * – it’s a pure computation all of whose args are loop invariant
> + * – it’s a variable use whose single reaching def, and the
> + * rhs of that def is loop-invariant
> + */
> + bool changes;
> +
> + do {
> + changes = false;
> + list_for_each_entry_safe(nir_loop_variable, var,
> + &state->process_list, process_link) {
> +
> + if (var->in_conditional_block || var->in_nested_loop) {
>
This is a bit unfortunate. As I mentioned on IRC, SPIR-V tends to turn a
simple for loop such as
for (int i = 0; i < 5; i++) {
/* stuff */
}
int i = 0;
bool has_run = false;
loop {
if (has_run) {
i++;
}
has_run = true;
if (!(i < 5))
break;
/* stuff */
}
The reason for this is so that continues work and properly continue to the
loop counter increment rather than to the top of the loop without
incrementing. It's probably easier, for the moment, to change SPIR-V to
emit
int i = 0;
loop {
if (!(i < 5))
break;
/* stuff */
i++;
}
in the case where there are no continues. I'll look into that.
> + LIST_DEL(&var->process_link);
>
I'd rather we use the lowercase versions of these. The macros are mostly
there because I didn't want to break gallium when I moved it to util and
fleshed things out. A quick search for LIST should find them all.
> + continue;
> + }
> +
> + if (is_ssa_def_invariant(var->def, state)) {
> + LIST_DEL(&var->process_link);
> + changes = true;
> + }
> + }
> + } while (changes);
>
Sorry for making so many comments on this one function. They just seem to
be clumping...
I don't think this loop is actually needed. As long as the variables are
in order from top to bottom of the function (which they are), we should be
able to compute loop invariance with a single walk of the loop's
instructions. All non-ALU instructions are assumed to be not invariant if
they are inside the loop and ALU insstructions are invariant if all their
sources (which much occur either above the loop or earlier in the loop) are
invariant.
I think the loop may have been needed for the GLSL IR pass (not actually
sure about that), but I don't see how it's needed here.
> +}
> +
> +static inline bool
> +is_var_basic_induction_var(nir_loop_variable *var, loop_info_state
> *state)
> +{
> + if (var->type == basic_induction)
> + return true;
> +
> + /* We are only interested in checking phi's for the basic induction
> + * variable case as its simple to detect. All basic induction variables
> + * have a phi node
> + */
> + if (!is_var_phi(var))
> + return false;
> +
> + nir_phi_instr *phi = nir_instr_as_phi(var->def->parent_instr);
> + nir_basic_induction_var *biv = rzalloc(state, nir_basic_induction_var);
> +
> + nir_foreach_phi_src(src, phi) {
> + nir_loop_variable *src_var = get_loop_var(src->src.ssa, state);
> +
> + /* If one of the sources is in a conditional or nested block then
> panic.
> + */
> + if (src_var->in_conditional_block || src_var->in_nested_loop)
> + break;
> +
> + if (!src_var->in_loop)
> + biv->def_outside_loop = src_var;
> +
> + if (src_var->in_loop && is_var_alu(src_var)) {
> + nir_alu_instr *alu = nir_instr_as_alu(src_var->def-
> >parent_instr);
> +
> + if (nir_op_infos[alu->op].num_inputs == 2) {
> + biv->alu_def = src_var;
> + biv->alu_op = alu->op;
> +
> + for (unsigned i = 0; i < 2; i++) {
> + /* Is one of the operands invariant, and the other the phi
> */
> + if (is_ssa_def_invariant(alu->src[i].src.ssa, state) &&
> + alu->src[1-i].src.ssa->index == phi->dest.ssa.index)
>
I'm sure I've asked this before, but why are we comparing indices here
rather than pointers? "ssa == &phi->dest.ssa" seems more reliable.
Also, given that you check for is_constant below. Why isn't this just
/* Is one of the operands const, and the other the phi */
if (alu->src[i].src.ssa->parent_instr->type == nir_instr_type_load_const &&
alu->src[1-i].src.ssa == &phi->dest.ssa)
> + biv->invariant = get_loop_var(alu->src[i].src.ssa,
> state);
> + }
> + }
> + }
> + }
> +
> + if (biv->alu_def && biv->def_outside_loop && biv->invariant &&
> + biv->def_outside_loop->is_constant &&
> biv->invariant->is_constant) {
> + biv->alu_def->type = basic_induction;
> + biv->alu_def->ind = biv;
> + var->type = basic_induction;
> + var->ind = biv;
> + return true;
> + }
> +
> + /* The requirements for a basic induction variable are not fulfilled */
> + ralloc_free(biv);
> + return false;
> +}
> +
> +static bool
> +compute_induction_information(loop_info_state *state)
> +{
> + bool changes;
> + bool found_induction_var = false;
> +
> + do {
> + changes = false;
> + list_for_each_entry_safe(nir_loop_variable, var,
> + &state->process_list, process_link) {
> +
> + /* It can't be an induction variable if it is invariant. We don't
> + * want to deal with things in nested loops or conditionals.
> + */
> + if (var->type == invariant || var->in_conditional_block ||
> + var->in_nested_loop) {
> + LIST_DEL(&(var->process_link));
> + continue;
> + }
> +
> + if (is_var_basic_induction_var(var, state)) {
> + /* If a phi is marked basic_ind we also mark the alu_def
> basic_ind
> + * at the same time. It will then be detected as basic_ind
> here,
> + * on the second passing, and be removed from the list.
> + */
> + LIST_DEL(&(var->process_link));
> + changes = true;
> + found_induction_var = true;
> + }
>
I'm forced to ask the same question about this one. The
is_var_basic_induction_var only does anything on phi nodes and isn't
recursive. Why can't we simply do:
nir_foreach_instr(instr, header_block) {
if (!instr->type == nir_instr_type_phi)
break;
/* process phi->dest.ssa */
}
> + }
> + } while (changes);
> +
> + return found_induction_var;
> +}
> +
> +static bool
> +initialize_ssa_def(nir_ssa_def *def, void *void_state)
> +{
> + loop_info_state *state = void_state;
> + nir_loop_variable *var = get_loop_var(def, state);
> +
> + var->in_loop = false;
> + var->def = def;
> +
> + if (def->parent_instr->type == nir_instr_type_load_const) {
> + var->type = invariant;
> + var->is_constant = true;
> + } else {
> + var->type = undefined;
> + }
> +
> + return true;
> +}
> +
> +static void
> +cf_node_find_loop_jumps(nir_cf_node *node, loop_jumps *state)
> +{
> + switch (node->type) {
> + case nir_cf_node_block: {
> + nir_instr *instr = nir_block_last_instr(nir_cf_no
> de_as_block(node));
> + if (instr && instr->type == nir_instr_type_jump) {
> + if (nir_instr_as_jump(instr)->type == nir_jump_break) {
> + state->contains_break = true;
> + } else if (nir_instr_as_jump(instr)->type == nir_jump_continue)
> {
> + state->contains_continue = true;
>
else assert! We really don't want to hit a return...
Maybe a switch? Doesn't matter.
> + }
> + }
> + break;
> + }
> + case nir_cf_node_if: {
> + nir_if *if_stmt = nir_cf_node_as_if(node);
> +
> + foreach_list_typed_safe(nir_cf_node, node, node,
> &if_stmt->then_list)
> + cf_node_find_loop_jumps(node, state);
> +
> + foreach_list_typed_safe(nir_cf_node, node, node,
> &if_stmt->else_list)
> + cf_node_find_loop_jumps(node, state);
> +
> + break;
> + }
> + case nir_cf_node_loop:
> + /* Fall through */
> +
> + default:
> + break;
> + }
> +}
> +
> +static bool
> +ends_in_break(nir_block *block)
> +{
> + if (exec_list_is_empty(&block->instr_list))
> + return false;
> +
> + nir_instr *instr = nir_block_last_instr(block);
> + return instr->type == nir_instr_type_jump &&
> + nir_instr_as_jump(instr)->type == nir_jump_break;
> +}
> +
> +static bool
> +find_loop_terminators(loop_info_state *state)
> +{
> + bool success = false;
> + foreach_list_typed_safe(nir_cf_node, node, node, &state->loop->body) {
> + if (node->type == nir_cf_node_if) {
> + nir_if *nif = nir_cf_node_as_if(node);
> +
> + /* Don't check the nested loops if there are breaks */
> + loop_jumps lj;
> + lj.contains_break = false;
> + lj.contains_continue = false;
> +
> + cf_node_find_loop_jumps(&nif->cf_node, &lj);
> +
> + if (lj.contains_continue)
> + return false;
> +
> + if (!lj.contains_break)
> + continue;
> +
> + nir_block *break_blk = NULL;
> + nir_block *continue_from_blk = NULL;
> + bool continue_from_then = true;
> +
> + nir_block *last_then = nir_if_last_then_block(nif);
> + nir_block *last_else = nir_if_last_else_block(nif);
> + if (ends_in_break(last_then)) {
> + break_blk = last_then;
> + continue_from_blk = last_else;
> + continue_from_then = false;
> + } else {
> + if (ends_in_break(last_else)) {
>
Maybe just use "else if"?
> + break_blk = last_else;
> + continue_from_blk = last_then;
> + }
> +
> + /* The break is nested so bail out */
> + return false;
>
I don't think you intended to return false if the if above succeeds.
> + }
> +
> + /* If there is a break then we should find a terminator. If we
> can
> + * not find a loop terminator, but there is a break-statement
> then
> + * we should return false so that we do not try to find
> trip-count
> + */
> + if (!nir_if_is_trivial_loop_terminator(nif, break_blk,
> + continue_from_blk,
> + continue_from_then))
> + return false;
> +
> + if (nif->condition.ssa->parent_instr->type ==
> nir_instr_type_phi)
> + return false;
>
Why? I guess maybe the trip count code can't handle it yet but I think
we'd want to eventually handle something such as
bool run = true;
while (run) {
/* Do stuff */
run = false;
}
No idea if it comes up in practice but it seems like a case someone might
expect to get unrolled.
We can add support for that later if you'd rather.
> +
> + nir_loop_terminator *terminator =
> + rzalloc(state->loop->info, nir_loop_terminator);
> +
> + list_add(&terminator->loop_terminator_link,
> + &state->loop->info->loop_terminator_list);
> +
> + terminator->nif = nif;
> + terminator->break_block = break_blk;
> + terminator->continue_from_block = continue_from_blk;
> + terminator->continue_from_then = continue_from_then;
> + terminator->conditional_instr = nif->condition.ssa->parent_ins
> tr;
> +
> + success = true;
> + }
> + }
> +
> + return success;
> +}
> +
> +static int32_t
> +get_iteration(nir_op cond_op, nir_const_value *initial, nir_const_value
> *step,
> + nir_const_value *limit, nir_alu_instr *alu)
> +{
> + int32_t iter;
> +
> + switch (cond_op) {
> + case nir_op_ige:
> + case nir_op_ilt:
> + case nir_op_ieq:
> + case nir_op_ine: {
> + int32_t initial_val = initial->i32[0];
> + int32_t span = limit->i32[0] - initial_val;
> + iter = span / step->i32[0];
> + break;
> + }
> + case nir_op_uge:
> + case nir_op_ult: {
> + uint32_t initial_val = initial->u32[0];
> + uint32_t span = limit->u32[0] - initial_val;
> + iter = span / step->u32[0];
> + break;
> + }
> + case nir_op_fge:
> + case nir_op_flt:
> + case nir_op_feq:
> + case nir_op_fne: {
> + int32_t initial_val = initial->f32[0];
> + int32_t span = limit->f32[0] - initial_val;
> + iter = span / step->f32[0];
> + break;
> + }
> + default:
> + return -1;
> + }
> +
> + return iter;
> +}
> +
> +static bool
> +utest_interations(int32_t iter_int, nir_const_value *step,
> + nir_const_value *limit, nir_op cond_op,
> + uint32_t initial_val, bool limit_rhs)
> +{
> + bool valid_loop = false;
> + uint32_t mul = iter_int * step->u32[0];
> + uint32_t uadd = mul + initial_val;
> +
> + switch (cond_op) {
> + case nir_op_uge:
> + valid_loop = limit_rhs ? uadd >= limit->u32[0] : uadd <=
> limit->u32[0];
> + break;
> + case nir_op_ult:
> + valid_loop = limit_rhs ? uadd < limit->u32[0] : uadd >
> limit->u32[0];
> + break;
> + default:
> + unreachable("Unhandled loop condition!");
> + }
> +
> + return valid_loop;
> +}
> +
> +static bool
> +itest_interations(int32_t iter_int, nir_const_value *step,
> + nir_const_value *limit, nir_op cond_op,
> + int32_t initial_val, bool limit_rhs)
> +{
> + bool valid_loop = false;
> + int32_t mul = iter_int * step->i32[0];
> + int32_t iadd = mul + initial_val;
> +
> + switch (cond_op) {
> + case nir_op_ige:
> + valid_loop = limit_rhs ? iadd >= limit->i32[0] : iadd <=
> limit->i32[0];
> + break;
> + case nir_op_ilt:
> + valid_loop = limit_rhs ? iadd < limit->i32[0] : iadd >
> limit->i32[0];
> + break;
> + case nir_op_ieq:
> + valid_loop = iadd == limit->i32[0];
> + break;
> + case nir_op_ine:
> + valid_loop = iadd != limit->i32[0];
> + break;
> + default:
> + unreachable("Unhandled loop condition!");
> + }
> +
> + return valid_loop;
> +}
> +
> +static bool
> +ftest_interations(int32_t iter_int, nir_const_value *step,
> + nir_const_value *limit, nir_op cond_op,
> + float initial_val, bool limit_rhs)
> +{
> + bool valid_loop = false;
> + float mul = iter_int * step->f32[0];
> + float iadd = mul + initial_val;
> +
> + switch (cond_op) {
> + case nir_op_fge:
> + valid_loop = limit_rhs ? iadd >= limit->f32[0] : iadd <=
> limit->f32[0];
> + break;
> + case nir_op_flt:
> + valid_loop = limit_rhs ? iadd < limit->f32[0] : iadd >
> limit->f32[0];
> + break;
> + case nir_op_feq:
> + valid_loop = iadd == limit->f32[0];
> + break;
> + case nir_op_fne:
> + valid_loop = iadd != limit->f32[0];
> + break;
> + default:
> + unreachable("Unhandled loop condition!");
> + }
> +
> + return valid_loop;
> +}
>
All three of these test_iterations helpers could be replaced with a single
function that calls nir_eval_const_opcode in nir_constant_expressions.h
> +
> +static int
> +calculate_iterations(nir_const_value *initial, nir_const_value *step,
> + nir_const_value *limit, nir_op cond_op,
> + nir_loop_variable *alu_def, nir_alu_instr *cond_alu,
> + bool limit_rhs)
> +{
> + assert(initial != NULL && step != NULL && limit != NULL);
> +
> + nir_alu_instr *alu = nir_instr_as_alu(alu_def->def->parent_instr);
> +
> + /* nir_op_isub should have been lowered away by this point */
> + assert(alu->op != nir_op_isub);
> +
> + /* Check for nsupported alu operations */
> + if (alu->op != nir_op_iadd && alu->op != nir_op_fadd)
> + return -1;
>
We probably also want to check that the alu operation type matches the
comparison type. If you do a nir_op_fadd with a nir_op_ige, you're in
trouble. nir_op_infos has type information so it shouldn't be hard to
check.
> +
> + /* do-while loops can increment the starting value before the
> condition is
> + * checked. e.g.
> + *
> + * do {
> + * ndx++;
> + * } while (ndx < 3);
> + *
> + * Here we check if the induction variable is used directly by the loop
> + * condition and if so we assume we need to step the initial value.
> + */
> + unsigned trip_offset = 0;
> + if (cond_alu->src[0].src.ssa == alu_def->def ||
> + cond_alu->src[1].src.ssa == alu_def->def) {
> + trip_offset = 1;
> + }
> +
> + int iter_int = get_iteration(cond_op, initial, step, limit, alu);
> +
> + /* If iter_int is negative the loop is ill-formed or is the
> conditional is
> + * unsigned with a huge iteration count so don't bother going any
> further.
> + */
> + if (iter_int < 0)
> + return -1;
> +
> + /* An explanation from the GLSL unrolling pass:
> + *
> + * Make sure that the calculated number of iterations satisfies the
> exit
> + * condition. This is needed to catch off-by-one errors and some
> types of
> + * ill-formed loops. For example, we need to detect that the following
> + * loop does not have a maximum iteration count.
> + *
> + * for (float x = 0.0; x != 0.9; x += 0.2);
> + */
> + const int bias[] = { -1, 1, 1 };
> +
> + for (unsigned i = 0; i < ARRAY_SIZE(bias); i++) {
>
Can we some how reformulate this as
for (int bias = -1; bias <= 1; bias++)
The -1, 1, 1 is really non-obvious. What we're really trying to do here is
to try it at iter_int - 1, iter_int, and iter_int + 1 and return whichever
one passes.
> + iter_int = iter_int + bias[i];
> +
> + switch (cond_op) {
> + case nir_op_ige:
> + case nir_op_ilt:
> + case nir_op_ieq:
> + case nir_op_ine:
> + if (itest_interations(iter_int, step, limit, cond_op,
> + initial->i32[0], limit_rhs)) {
> + return iter_int > 0 ? iter_int - trip_offset : iter_int;
> + }
> + break;
> + case nir_op_uge:
> + case nir_op_ult:
> + if (utest_interations(iter_int, step, limit, cond_op,
> + initial->u32[0], limit_rhs)) {
> + return iter_int > 0 ? iter_int - trip_offset : iter_int;
> + }
> + break;
> + case nir_op_fge:
> + case nir_op_flt:
> + case nir_op_feq:
> + case nir_op_fne:
> + if (ftest_interations(iter_int, step, limit, cond_op,
> + initial->f32[0], limit_rhs)) {
> + return iter_int > 0 ? iter_int - trip_offset : iter_int;
> + }
> + break;
> + default:
> + return -1;
> + }
> + }
> +
> + return -1;
> +}
> +
> +/* Run through each of the terminators of the loop and try to infer a
> possible
> + * trip-count. We need to check them all, and set the lowest trip-count
> as the
> + * trip-count of our loop. If one of the terminators has an undecidable
> + * trip-count we can not safely assume anything about the duration of the
> + * loop.
> + */
> +static void
> +find_trip_count(loop_info_state *state)
> +{
> + bool trip_count_known = true;
> + nir_loop_terminator *limiting_terminator = NULL;
> + int min_trip_count = -1;
> +
> + list_for_each_entry(nir_loop_terminator, terminator,
> + &state->loop->info->loop_terminator_list,
> + loop_terminator_link) {
> +
> + if (terminator->conditional_instr->type != nir_instr_type_alu) {
> + /* If we get here the loop is possible the loop is dead and will
> get
> + * cleaned up by the nir_opt_dead_cf pass.
>
Please rephrase "the loop is possible the loop"?
> + */
> + trip_count_known = false;
> + continue;
> + }
> +
> + nir_alu_instr *alu = nir_instr_as_alu(terminator->c
> onditional_instr);
> + nir_loop_variable *basic_ind = NULL;
> + nir_loop_variable *limit = NULL;
> + bool limit_rhs = true;
> + nir_op cond_op;
> +
> + switch (alu->op) {
> + case nir_op_fge: case nir_op_ige: case nir_op_uge:
> + case nir_op_flt: case nir_op_ilt: case nir_op_ult:
> + case nir_op_feq: case nir_op_ieq:
> + case nir_op_fne: case nir_op_ine:
>
Might be worth adding nir_op_i2b to this list. It's effectively != 0.
> +
> + /* We assume that the limit is the "right" operand */
> + basic_ind = get_loop_var(alu->src[0].src.ssa, state);
> + limit = get_loop_var(alu->src[1].src.ssa, state);
> + cond_op = alu->op;
> +
> + if (basic_ind->type != basic_induction) {
> + /* We had it the wrong way, flip things around */
> + basic_ind = get_loop_var(alu->src[1].src.ssa, state);
> + limit = get_loop_var(alu->src[0].src.ssa, state);
> + limit_rhs = false;
> + }
> +
> + /* The comparison has to have a basic induction variable
> + * and a constant for us to be able to find trip counts
> + */
> + if (basic_ind->type != basic_induction || !limit->is_constant) {
> + trip_count_known = false;
> + continue;
> + }
> +
> + /* We have determined that we have the following constants:
> + * (With the typical int i = 0; i < x; i++; as an example)
> + * - Upper limit.
> + * - Starting value
> + * - Step / iteration size
> + * Thats all thats needed to calculate the trip-count
> + */
> +
> + nir_load_const_instr *initial_instr =
> + nir_instr_as_load_const(basic_ind->ind->def_outside_loop->
> + def->parent_instr);
> +
> + nir_const_value initial_val = initial_instr->value;
>
We never use initial_instr, these can be one statement.
> +
> + nir_load_const_instr *step_instr =
> + nir_instr_as_load_const(basic_ind->ind->invariant->def->
> + parent_instr);
> +
> + nir_const_value step_val = step_instr->value;
>
Same here.
> +
> + nir_load_const_instr *limit_instr =
> + nir_instr_as_load_const(limit->def->parent_instr);
> +
> + nir_const_value limit_val = limit_instr->value;
>
and here
> +
> + int iterations = calculate_iterations(&initial_val, &step_val,
> + &limit_val, cond_op,
> + basic_ind->ind->alu_def,
> alu,
> + limit_rhs);
> +
> + /* Where we not able to calculate the iteration count */
> + if (iterations == -1) {
> + trip_count_known = false;
> + continue;
> + }
> +
> + /* If this is the first run or we have found a smaller amount of
> + * iterations than previously (we have identified a more limiting
> + * terminator) set the trip count and limiting terminator.
> + */
> + if (min_trip_count == -1 || iterations < min_trip_count) {
> + min_trip_count = iterations;
> + limiting_terminator = terminator;
> + }
> + break;
> +
> + default:
> + trip_count_known = false;
> + }
> + }
> +
> + state->loop->info->is_trip_count_known = trip_count_known;
> + if (min_trip_count > -1)
> + state->loop->info->trip_count = min_trip_count;
> + state->loop->info->limiting_terminator = limiting_terminator;
> +}
> +
> +/* Checks if we should force the loop to be unrolled regardless of size */
> +static bool
> +force_unroll(loop_info_state *state, nir_shader *ns, nir_deref_var
> *variable)
> +{
> + nir_deref *tail = &variable->deref;
> +
> + while (tail->child != NULL) {
> + tail = tail->child;
> +
> + if (tail->deref_type == nir_deref_type_array) {
> +
> + nir_deref_array *deref_array = nir_deref_as_array(tail);
> + if (deref_array->deref_array_type !=
> nir_deref_array_type_indirect)
> + continue;
> +
> + nir_loop_variable *array_index =
> + get_loop_var(deref_array->indirect.ssa, state);
> +
> + if (array_index->type != basic_induction)
> + continue;
> +
> + /* If an array is indexed by a loop induction variable, and the
> + * array size is exactly the number of loop iterations, this is
> + * probably a simple for-loop trying to access each element in
> + * turn; the application may expect it to be unrolled.
> + */
> + if (glsl_get_length(variable->deref.type) ==
> + state->loop->info->trip_count) {
> + state->loop->info->force_unroll = true;
> + return state->loop->info->force_unroll;
> + }
> +
> + if (variable->var->data.mode & state->indirect_mask) {
> + state->loop->info->force_unroll = true;
> + return state->loop->info->force_unroll;
> + }
>
I'm not sure what I think of the force_unroll bool. I think we may, at
some point, want something more granular. That said, it does exactly what
we need now so good enough. (No need to change things. mostly making notes
to myself)
> + }
> + }
> +
> + return false;
> +}
> +
> +static bool
> +count_instructions(loop_info_state *state, nir_shader *ns, nir_block
> *block)
> +{
> + nir_foreach_instr(instr, block) {
> + if (instr->type == nir_instr_type_intrinsic ||
> + instr->type == nir_instr_type_alu) {
>
Should we add texture instructions to this list?
> + state->loop->info->num_instructions++;
> + }
> +
> + if (instr->type != nir_instr_type_intrinsic)
> + continue;
> +
> + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
> +
> + /* Check for arrays variably-indexed by a loop induction variable.
> + * Unrolling the loop may convert that access into
> constant-indexing.
> + */
> + if (intrin->intrinsic == nir_intrinsic_load_var ||
> + intrin->intrinsic == nir_intrinsic_store_var ||
> + intrin->intrinsic == nir_intrinsic_copy_var) {
> + unsigned num_vars =
> + nir_intrinsic_infos[intrin->intrinsic].num_variables;
> + for (unsigned i = 0; i < num_vars; i++) {
> + if (force_unroll(state, ns, intrin->variables[i]))
> + return true;
>
It's a bit deceptive that force_unroll being true means that the
instruction count is wrong. Maybe we should just count instructions
up-front as we initialize the SSA defs or something. It'd be easy enough
to put it there and then it would be true regardless of any unroll state.
> + }
> + }
> + }
> +
> + return false;
> +}
> +
> +static void
> +get_loop_info(loop_info_state *state, nir_function_impl *impl)
> +{
> + /* Initialize all variables to "outside_loop". This also marks defs
> + * invariant and constant if they are nir_instr_type_load_const's
> + */
> + nir_foreach_block(block, impl) {
> + nir_foreach_instr(instr, block)
> + nir_foreach_ssa_def(instr, initialize_ssa_def, state);
> + }
> +
> + /* Add all entries in the outermost part of the loop to the processing
> list
> + * Mark the entries in conditionals or in nested loops accordingly
> + */
> + foreach_list_typed_safe(nir_cf_node, node, node, &state->loop->body) {
> + switch (node->type) {
> +
> + case nir_cf_node_block:
> + init_loop_block(nir_cf_node_as_block(node), state, false,
> false);
> + break;
> +
> + case nir_cf_node_if:
> + nir_foreach_block_in_cf_node(block, node)
> + init_loop_block(block, state, true, false);
>
What about loops inside of ifs or ifs inside of loops? Maybe it doesn't
matter. If it doesn't, you should add some sort of comment to that
effect. Otherwise, we probably need something recursive.
Reading further... It doesn't matter. We always check them together and OR
the result. Maybe we should just have a single in_control_flow boolean
instead.
> + break;
> +
> + case nir_cf_node_loop:
> + nir_foreach_block_in_cf_node(block, node) {
> + init_loop_block(block, state, false, true);
> + }
> + break;
> +
> + case nir_cf_node_function:
> + break;
> + }
> + }
> +
> + /* Induction analysis needs invariance information so get that first */
> + compute_invariance_information(state);
> +
> + /* We may now have filled the process_list with instructions from
> inside
> + * the nested blocks in the loop. Remove all instructions from the list
> + * nir_foreach_block_in_cf_node before we start computing induction
> + * information.
> + */
> + list_inithead(&state->process_list);
> +
> + /* Add all entries in the outermost part of the loop to the processing
> list.
> + * Don't include defs in nested loops or in conditionals.
> + */
> + foreach_list_typed_safe(nir_cf_node, node, node, &state->loop->body) {
> + if (node->type == nir_cf_node_block)
> + init_loop_block(nir_cf_node_as_block(node), state, false,
> false);
> + }
> +
> + /* We have invariance information so try to find induction variables */
> + if (!compute_induction_information(state))
> + return;
> +
> + /* Try to find all simple terminators of the loop. If we can't find
> any,
> + * or we find possible terminators that have side effects then bail.
> + */
> + if (!find_loop_terminators(state))
> + return;
> +
> + /* Run through each of the terminators and try to compute a trip-count
> */
> + find_trip_count(state);
> +
> + nir_shader *ns = impl->function->shader;
> + foreach_list_typed_safe(nir_cf_node, node, node, &state->loop->body) {
> + if (node->type == nir_cf_node_block) {
> + if (count_instructions(state, ns, nir_cf_node_as_block(node)))
> + break;
> + } else {
> + nir_foreach_block_in_cf_node(block, node) {
> + if (count_instructions(state, ns, block))
> + break;
> + }
> + }
> + }
> +}
> +
> +static loop_info_state *
> +initialize_loop_info_state(nir_loop *loop, void *mem_ctx,
> + nir_function_impl *impl)
> +{
> + loop_info_state *state = rzalloc(mem_ctx, loop_info_state);
> + state->loop_vars = rzalloc_array(mem_ctx, nir_loop_variable,
> + impl->ssa_alloc);
> + state->loop = loop;
> +
> + LIST_INITHEAD(&state->process_list);
> +
> + if (loop->info)
> + ralloc_free(loop->info);
> +
> + loop->info = rzalloc(loop, nir_loop_info);
> +
> + LIST_INITHEAD(&loop->info->loop_terminator_list);
> +
> + return state;
> +}
> +
> +static void
> +process_loops(nir_cf_node *cf_node, nir_variable_mode indirect_mask)
> +{
> + switch (cf_node->type) {
> + case nir_cf_node_block:
> + return;
> + case nir_cf_node_if: {
> + nir_if *if_stmt = nir_cf_node_as_if(cf_node);
> + foreach_list_typed(nir_cf_node, nested_node, node,
> &if_stmt->then_list)
> + process_loops(nested_node, indirect_mask);
> + foreach_list_typed(nir_cf_node, nested_node, node,
> &if_stmt->else_list)
> + process_loops(nested_node, indirect_mask);
> + return;
> + }
> + case nir_cf_node_loop: {
> + nir_loop *loop = nir_cf_node_as_loop(cf_node);
> + foreach_list_typed(nir_cf_node, nested_node, node, &loop->body)
> + process_loops(nested_node, indirect_mask);
> + break;
> + }
> + default:
> + unreachable("unknown cf node type");
> + }
> +
> + nir_loop *loop = nir_cf_node_as_loop(cf_node);
> + nir_function_impl *impl = nir_cf_node_get_function(cf_node);
> + void *mem_ctx = ralloc_context(NULL);
> +
> + loop_info_state *state = initialize_loop_info_state(loop, mem_ctx,
> impl);
> + state->indirect_mask = indirect_mask;
> +
> + get_loop_info(state, impl);
> +
> + ralloc_free(mem_ctx);
> +}
> +
> +void
> +nir_loop_analyze_impl(nir_function_impl *impl,
> + nir_variable_mode indirect_mask)
> +{
> + if (impl->function->shader->options->max_unroll_iterations == 0)
> + return;
>
Now that loop analysis isn't making decisions, do we need this? If you
don't want loop analysis, just don't run it.
> +
> + nir_index_ssa_defs(impl);
> + foreach_list_typed(nir_cf_node, node, node, &impl->body)
> + process_loops(node, indirect_mask);
> +}
> diff --git a/src/compiler/nir/nir_metadata.c
> b/src/compiler/nir/nir_metadata.c
> index 9e1cff5..f71cf43 100644
> --- a/src/compiler/nir/nir_metadata.c
> +++ b/src/compiler/nir/nir_metadata.c
> @@ -31,7 +31,7 @@
> */
>
> void
> -nir_metadata_require(nir_function_impl *impl, nir_metadata required)
> +nir_metadata_require(nir_function_impl *impl, nir_metadata required, ...)
> {
> #define NEEDS_UPDATE(X) ((required & ~impl->valid_metadata) & (X))
>
> @@ -41,6 +41,12 @@ nir_metadata_require(nir_function_impl *impl,
> nir_metadata required)
> nir_calc_dominance_impl(impl);
> if (NEEDS_UPDATE(nir_metadata_live_ssa_defs))
> nir_live_ssa_defs_impl(impl);
> + if (NEEDS_UPDATE(nir_metadata_loop_analysis)) {
> + va_list ap;
> + va_start(ap, required);
> + nir_loop_analyze_impl(impl, va_arg(ap, nir_variable_mode));
> + va_end(ap);
> + }
>
> #undef NEEDS_UPDATE
>
> --
> 2.7.4
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20161214/440499d5/attachment-0001.html>
More information about the mesa-dev
mailing list