/** @file register_allocate.c
*
* Graph-coloring register allocator.
+ *
+ * The basic idea of graph coloring is to make a node in a graph for
+ * every thing that needs a register (color) number assigned, and make
+ * edges in the graph between nodes that interfere (can't be allocated
+ * to the same register at the same time).
+ *
+ * During the "simplify" process, any any node with fewer edges than
+ * there are registers means that that edge can get assigned a
+ * register regardless of what its neighbors choose, so that node is
+ * pushed on a stack and removed (with its edges) from the graph.
+ * That likely causes other nodes to become trivially colorable as well.
+ *
+ * Then during the "select" process, nodes are popped off of that
+ * stack, their edges restored, and assigned a color different from
+ * their neighbors. Because they were pushed on the stack only when
+ * they were trivially colorable, any color chosen won't interfere
+ * with the registers to be popped later.
+ *
+ * The downside to most graph coloring is that real hardware often has
+ * limitations, like registers that need to be allocated to a node in
+ * pairs, or aligned on some boundary. This implementation follows
+ * the paper "Retargetable Graph-Coloring Register Allocation for
+ * Irregular Architectures" by Johan Runeson and Sven-Olof Nyström.
+ *
+ * In this system, there are register classes each containing various
+ * registers, and registers may interfere with other registers. For
+ * example, one might have a class of base registers, and a class of
+ * aligned register pairs that would each interfere with their pair of
+ * the base registers. Each node has a register class it needs to be
+ * assigned to. Define p(B) to be the size of register class B, and
+ * q(B,C) to be the number of registers in B that the worst choice
+ * register in C could conflict with. Then, this system replaces the
+ * basic graph coloring test of "fewer edges from this node than there
+ * are registers" with "For this node of class B, the sum of q(B,C)
+ * for each neighbor node of class C is less than pB".
+ *
+ * A nice feature of the pq test is that q(B,C) can be computed once
+ * up front and stored in a 2-dimensional array, so that the cost of
+ * coloring a node is constant with the number of registers. We do
+ * this during ra_set_finalize().
*/
-#include <talloc.h>
+#include <stdbool.h>
+#include <ralloc.h>
#include "main/imports.h"
#include "main/macros.h"
#include "main/mtypes.h"
+#include "main/bitset.h"
#include "register_allocate.h"
+#define NO_REG ~0
+
struct ra_reg {
- char *name;
- GLboolean *conflicts;
+ bool *conflicts;
+ unsigned int *conflict_list;
+ unsigned int conflict_list_size;
+ unsigned int num_conflicts;
};
struct ra_regs {
struct ra_class **classes;
unsigned int class_count;
+
+ bool round_robin;
};
struct ra_class {
- GLboolean *regs;
+ bool *regs;
/**
- * p_B in Runeson/Nyström paper.
+ * p(B) in Runeson/Nyström paper.
*
* This is "how many regs are in the set."
*/
unsigned int p;
/**
- * q_B,C in Runeson/Nyström paper.
+ * q(B,C) (indexed by C, B is this register class) in
+ * Runeson/Nyström paper. This is "how many registers of B could
+ * the worst choice register from C conflict with".
*/
unsigned int *q;
};
struct ra_node {
- GLboolean *adjacency;
- unsigned int class;
+ /** @{
+ *
+ * List of which nodes this node interferes with. This should be
+ * symmetric with the other node.
+ */
+ BITSET_WORD *adjacency;
+ unsigned int *adjacency_list;
+ unsigned int adjacency_list_size;
unsigned int adjacency_count;
+ /** @} */
+
+ unsigned int class;
+
+ /* Register, if assigned, or NO_REG. */
unsigned int reg;
- GLboolean in_stack;
+
+ /**
+ * Set when the node is in the trivially colorable stack. When
+ * set, the adjacency to this node is ignored, to implement the
+ * "remove the edge from the graph" in simplification without
+ * having to actually modify the adjacency_list.
+ */
+ bool in_stack;
+
+ /* For an implementation that needs register spilling, this is the
+ * approximate cost of spilling this node.
+ */
+ float spill_cost;
};
struct ra_graph {
unsigned int *stack;
unsigned int stack_count;
+
+ /**
+ * Tracks the start of the set of optimistically-colored registers in the
+ * stack.
+ *
+ * Along with any registers not in the stack (if one called ra_simplify()
+ * and didn't do optimistic coloring), these need to be considered for
+ * spilling.
+ */
+ unsigned int stack_optimistic_start;
};
+/**
+ * Creates a set of registers for the allocator.
+ *
+ * mem_ctx is a ralloc context for the allocator. The reg set may be freed
+ * using ralloc_free().
+ */
struct ra_regs *
-ra_alloc_reg_set(unsigned int count)
+ra_alloc_reg_set(void *mem_ctx, unsigned int count)
{
unsigned int i;
struct ra_regs *regs;
- regs = talloc_zero(NULL, struct ra_regs);
+ regs = rzalloc(mem_ctx, struct ra_regs);
regs->count = count;
- regs->regs = talloc_zero_array(regs, struct ra_reg, count);
+ regs->regs = rzalloc_array(regs, struct ra_reg, count);
for (i = 0; i < count; i++) {
- regs->regs[i].conflicts = talloc_zero_array(regs->regs, GLboolean, count);
- regs->regs[i].conflicts[i] = GL_TRUE;
+ regs->regs[i].conflicts = rzalloc_array(regs->regs, bool, count);
+ regs->regs[i].conflicts[i] = true;
+
+ regs->regs[i].conflict_list = ralloc_array(regs->regs, unsigned int, 4);
+ regs->regs[i].conflict_list_size = 4;
+ regs->regs[i].conflict_list[0] = i;
+ regs->regs[i].num_conflicts = 1;
}
return regs;
}
+/**
+ * The register allocator by default prefers to allocate low register numbers,
+ * since it was written for hardware (gen4/5 Intel) that is limited in its
+ * multithreadedness by the number of registers used in a given shader.
+ *
+ * However, for hardware without that restriction, densely packed register
+ * allocation can put serious constraints on instruction scheduling. This
+ * function tells the allocator to rotate around the registers if possible as
+ * it allocates the nodes.
+ */
+void
+ra_set_allocate_round_robin(struct ra_regs *regs)
+{
+ regs->round_robin = true;
+}
+
+static void
+ra_add_conflict_list(struct ra_regs *regs, unsigned int r1, unsigned int r2)
+{
+ struct ra_reg *reg1 = ®s->regs[r1];
+
+ if (reg1->conflict_list_size == reg1->num_conflicts) {
+ reg1->conflict_list_size *= 2;
+ reg1->conflict_list = reralloc(regs->regs, reg1->conflict_list,
+ unsigned int, reg1->conflict_list_size);
+ }
+ reg1->conflict_list[reg1->num_conflicts++] = r2;
+ reg1->conflicts[r2] = true;
+}
+
void
ra_add_reg_conflict(struct ra_regs *regs, unsigned int r1, unsigned int r2)
{
- regs->regs[r1].conflicts[r2] = GL_TRUE;
- regs->regs[r2].conflicts[r1] = GL_TRUE;
+ if (!regs->regs[r1].conflicts[r2]) {
+ ra_add_conflict_list(regs, r1, r2);
+ ra_add_conflict_list(regs, r2, r1);
+ }
+}
+
+/**
+ * Adds a conflict between base_reg and reg, and also between reg and
+ * anything that base_reg conflicts with.
+ *
+ * This can simplify code for setting up multiple register classes
+ * which are aggregates of some base hardware registers, compared to
+ * explicitly using ra_add_reg_conflict.
+ */
+void
+ra_add_transitive_reg_conflict(struct ra_regs *regs,
+ unsigned int base_reg, unsigned int reg)
+{
+ int i;
+
+ ra_add_reg_conflict(regs, reg, base_reg);
+
+ for (i = 0; i < regs->regs[base_reg].num_conflicts; i++) {
+ ra_add_reg_conflict(regs, reg, regs->regs[base_reg].conflict_list[i]);
+ }
}
unsigned int
{
struct ra_class *class;
- regs->classes = talloc_realloc(regs, regs->classes,
- struct ra_class *,
- regs->class_count + 1);
+ regs->classes = reralloc(regs->regs, regs->classes, struct ra_class *,
+ regs->class_count + 1);
- class = talloc_zero(regs, struct ra_class);
+ class = rzalloc(regs, struct ra_class);
regs->classes[regs->class_count] = class;
- class->regs = talloc_zero_array(class, GLboolean, regs->count);
+ class->regs = rzalloc_array(class, bool, regs->count);
return regs->class_count++;
}
{
struct ra_class *class = regs->classes[c];
- class->regs[r] = GL_TRUE;
+ class->regs[r] = true;
class->p++;
}
/**
* Must be called after all conflicts and register classes have been
* set up and before the register set is used for allocation.
+ * To avoid costly q value computation, use the q_values paramater
+ * to pass precomputed q values to this function.
*/
void
-ra_set_finalize(struct ra_regs *regs)
+ra_set_finalize(struct ra_regs *regs, unsigned int **q_values)
{
unsigned int b, c;
for (b = 0; b < regs->class_count; b++) {
- regs->classes[b]->q = talloc_array(regs, unsigned int, regs->class_count);
+ regs->classes[b]->q = ralloc_array(regs, unsigned int, regs->class_count);
+ }
+
+ if (q_values) {
+ for (b = 0; b < regs->class_count; b++) {
+ for (c = 0; c < regs->class_count; c++) {
+ regs->classes[b]->q[c] = q_values[b][c];
+ }
+ }
+ return;
}
/* Compute, for each class B and C, how many regs of B an
int max_conflicts = 0;
for (rc = 0; rc < regs->count; rc++) {
- unsigned int rb;
int conflicts = 0;
+ int i;
if (!regs->classes[c]->regs[rc])
continue;
- for (rb = 0; rb < regs->count; rb++) {
- if (regs->classes[b]->regs[rb] &&
- regs->regs[rb].conflicts[rc])
+ for (i = 0; i < regs->regs[rc].num_conflicts; i++) {
+ unsigned int rb = regs->regs[rc].conflict_list[i];
+ if (regs->classes[b]->regs[rb])
conflicts++;
}
max_conflicts = MAX2(max_conflicts, conflicts);
}
}
+static void
+ra_add_node_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2)
+{
+ BITSET_SET(g->nodes[n1].adjacency, n2);
+
+ if (g->nodes[n1].adjacency_count >=
+ g->nodes[n1].adjacency_list_size) {
+ g->nodes[n1].adjacency_list_size *= 2;
+ g->nodes[n1].adjacency_list = reralloc(g, g->nodes[n1].adjacency_list,
+ unsigned int,
+ g->nodes[n1].adjacency_list_size);
+ }
+
+ g->nodes[n1].adjacency_list[g->nodes[n1].adjacency_count] = n2;
+ g->nodes[n1].adjacency_count++;
+}
+
struct ra_graph *
ra_alloc_interference_graph(struct ra_regs *regs, unsigned int count)
{
struct ra_graph *g;
unsigned int i;
- g = talloc_zero(regs, struct ra_graph);
+ g = rzalloc(regs, struct ra_graph);
g->regs = regs;
- g->nodes = talloc_zero_array(g, struct ra_node, count);
+ g->nodes = rzalloc_array(g, struct ra_node, count);
g->count = count;
- g->stack = talloc_zero_array(g, unsigned int, count);
+ g->stack = rzalloc_array(g, unsigned int, count);
for (i = 0; i < count; i++) {
- g->nodes[i].adjacency = talloc_zero_array(g, GLboolean, count);
- g->nodes[i].adjacency[i] = GL_TRUE;
- g->nodes[i].reg = ~0;
+ int bitset_count = BITSET_WORDS(count);
+ g->nodes[i].adjacency = rzalloc_array(g, BITSET_WORD, bitset_count);
+
+ g->nodes[i].adjacency_list_size = 4;
+ g->nodes[i].adjacency_list =
+ ralloc_array(g, unsigned int, g->nodes[i].adjacency_list_size);
+ g->nodes[i].adjacency_count = 0;
+
+ ra_add_node_adjacency(g, i, i);
+ g->nodes[i].reg = NO_REG;
}
return g;
ra_add_node_interference(struct ra_graph *g,
unsigned int n1, unsigned int n2)
{
- if (g->nodes[n1].adjacency[n2])
- return;
-
- g->nodes[n1].adjacency[n2] = GL_TRUE;
- g->nodes[n2].adjacency_count++;
- g->nodes[n2].adjacency[n1] = GL_TRUE;
- g->nodes[n2].adjacency_count++;
+ if (!BITSET_TEST(g->nodes[n1].adjacency, n2)) {
+ ra_add_node_adjacency(g, n1, n2);
+ ra_add_node_adjacency(g, n2, n1);
+ }
}
-static GLboolean pq_test(struct ra_graph *g, unsigned int n)
+static bool
+pq_test(struct ra_graph *g, unsigned int n)
{
unsigned int j;
unsigned int q = 0;
int n_class = g->nodes[n].class;
- for (j = 0; j < g->count; j++) {
- if (j == n || g->nodes[j].in_stack)
- continue;
+ for (j = 0; j < g->nodes[n].adjacency_count; j++) {
+ unsigned int n2 = g->nodes[n].adjacency_list[j];
+ unsigned int n2_class = g->nodes[n2].class;
- if (g->nodes[n].adjacency[j]) {
- unsigned int j_class = g->nodes[j].class;
- q += g->regs->classes[n_class]->q[j_class];
+ if (n != n2 && !g->nodes[n2].in_stack) {
+ q += g->regs->classes[n_class]->q[n2_class];
}
}
* trivially-colorable nodes into a stack of nodes to be colored,
* removing them from the graph, and rinsing and repeating.
*
- * Returns GL_TRUE if all nodes were removed from the graph. GL_FALSE
+ * Returns true if all nodes were removed from the graph. false
* means that either spilling will be required, or optimistic coloring
* should be applied.
*/
-GLboolean
+bool
ra_simplify(struct ra_graph *g)
{
- GLboolean progress = GL_TRUE;
+ bool progress = true;
int i;
while (progress) {
- progress = GL_FALSE;
+ progress = false;
for (i = g->count - 1; i >= 0; i--) {
- if (g->nodes[i].in_stack)
+ if (g->nodes[i].in_stack || g->nodes[i].reg != NO_REG)
continue;
if (pq_test(g, i)) {
g->stack[g->stack_count] = i;
g->stack_count++;
- g->nodes[i].in_stack = GL_TRUE;
- progress = GL_TRUE;
+ g->nodes[i].in_stack = true;
+ progress = true;
}
}
}
for (i = 0; i < g->count; i++) {
- if (!g->nodes[i].in_stack)
- return GL_FALSE;
+ if (!g->nodes[i].in_stack && g->nodes[i].reg == -1)
+ return false;
}
- return GL_TRUE;
+ return true;
}
/**
* registers as they go.
*
* If all nodes were trivially colorable, then this must succeed. If
- * not (optimistic coloring), then it may return GL_FALSE;
+ * not (optimistic coloring), then it may return false;
*/
-GLboolean
+bool
ra_select(struct ra_graph *g)
{
int i;
+ int start_search_reg = 0;
while (g->stack_count != 0) {
- unsigned int r;
+ unsigned int ri;
+ unsigned int r = -1;
int n = g->stack[g->stack_count - 1];
struct ra_class *c = g->regs->classes[g->nodes[n].class];
/* Find the lowest-numbered reg which is not used by a member
* of the graph adjacent to us.
*/
- for (r = 0; r < g->regs->count; r++) {
+ for (ri = 0; ri < g->regs->count; ri++) {
+ r = (start_search_reg + ri) % g->regs->count;
if (!c->regs[r])
continue;
/* Check if any of our neighbors conflict with this register choice. */
- for (i = 0; i < g->count; i++) {
- if (g->nodes[n].adjacency[i] &&
- !g->nodes[i].in_stack &&
- g->regs->regs[r].conflicts[g->nodes[i].reg]) {
+ for (i = 0; i < g->nodes[n].adjacency_count; i++) {
+ unsigned int n2 = g->nodes[n].adjacency_list[i];
+
+ if (!g->nodes[n2].in_stack &&
+ g->regs->regs[r].conflicts[g->nodes[n2].reg]) {
break;
}
}
- if (i == g->count)
+ if (i == g->nodes[n].adjacency_count)
break;
}
- if (r == g->regs->count)
- return GL_FALSE;
+ if (ri == g->regs->count)
+ return false;
g->nodes[n].reg = r;
- g->nodes[n].in_stack = GL_FALSE;
+ g->nodes[n].in_stack = false;
g->stack_count--;
+
+ if (g->regs->round_robin)
+ start_search_reg = r + 1;
}
- return GL_TRUE;
+ return true;
}
/**
{
unsigned int i;
+ g->stack_optimistic_start = g->stack_count;
for (i = 0; i < g->count; i++) {
- if (g->nodes[i].in_stack)
+ if (g->nodes[i].in_stack || g->nodes[i].reg != NO_REG)
continue;
g->stack[g->stack_count] = i;
g->stack_count++;
- g->nodes[i].in_stack = GL_TRUE;
+ g->nodes[i].in_stack = true;
}
}
-GLboolean
+bool
ra_allocate_no_spills(struct ra_graph *g)
{
if (!ra_simplify(g)) {
{
return g->nodes[n].reg;
}
+
+/**
+ * Forces a node to a specific register. This can be used to avoid
+ * creating a register class containing one node when handling data
+ * that must live in a fixed location and is known to not conflict
+ * with other forced register assignment (as is common with shader
+ * input data). These nodes do not end up in the stack during
+ * ra_simplify(), and thus at ra_select() time it is as if they were
+ * the first popped off the stack and assigned their fixed locations.
+ * Nodes that use this function do not need to be assigned a register
+ * class.
+ *
+ * Must be called before ra_simplify().
+ */
+void
+ra_set_node_reg(struct ra_graph *g, unsigned int n, unsigned int reg)
+{
+ g->nodes[n].reg = reg;
+ g->nodes[n].in_stack = false;
+}
+
+static float
+ra_get_spill_benefit(struct ra_graph *g, unsigned int n)
+{
+ int j;
+ float benefit = 0;
+ int n_class = g->nodes[n].class;
+
+ /* Define the benefit of eliminating an interference between n, n2
+ * through spilling as q(C, B) / p(C). This is similar to the
+ * "count number of edges" approach of traditional graph coloring,
+ * but takes classes into account.
+ */
+ for (j = 0; j < g->nodes[n].adjacency_count; j++) {
+ unsigned int n2 = g->nodes[n].adjacency_list[j];
+ if (n != n2) {
+ unsigned int n2_class = g->nodes[n2].class;
+ benefit += ((float)g->regs->classes[n_class]->q[n2_class] /
+ g->regs->classes[n_class]->p);
+ }
+ }
+
+ return benefit;
+}
+
+/**
+ * Returns a node number to be spilled according to the cost/benefit using
+ * the pq test, or -1 if there are no spillable nodes.
+ */
+int
+ra_get_best_spill_node(struct ra_graph *g)
+{
+ unsigned int best_node = -1;
+ float best_benefit = 0.0;
+ unsigned int n, i;
+
+ /* For any registers not in the stack to be colored, consider them for
+ * spilling. This will mostly collect nodes that were being optimistally
+ * colored as part of ra_allocate_no_spills() if we didn't successfully
+ * optimistically color.
+ *
+ * It also includes nodes not trivially colorable by ra_simplify() if it
+ * was used directly instead of as part of ra_allocate_no_spills().
+ */
+ for (n = 0; n < g->count; n++) {
+ float cost = g->nodes[n].spill_cost;
+ float benefit;
+
+ if (cost <= 0.0)
+ continue;
+
+ if (g->nodes[n].in_stack)
+ continue;
+
+ benefit = ra_get_spill_benefit(g, n);
+
+ if (benefit / cost > best_benefit) {
+ best_benefit = benefit / cost;
+ best_node = n;
+ }
+ }
+
+ /* Also consider spilling any nodes that were set up to be optimistically
+ * colored that we couldn't manage to color in ra_select().
+ */
+ for (i = g->stack_optimistic_start; i < g->stack_count; i++) {
+ float cost, benefit;
+
+ n = g->stack[i];
+ cost = g->nodes[n].spill_cost;
+
+ if (cost <= 0.0)
+ continue;
+
+ benefit = ra_get_spill_benefit(g, n);
+
+ if (benefit / cost > best_benefit) {
+ best_benefit = benefit / cost;
+ best_node = n;
+ }
+ }
+
+ return best_node;
+}
+
+/**
+ * Only nodes with a spill cost set (cost != 0.0) will be considered
+ * for register spilling.
+ */
+void
+ra_set_node_spill_cost(struct ra_graph *g, unsigned int n, float cost)
+{
+ g->nodes[n].spill_cost = cost;
+}