@@ -120,6 +120,7 @@ LIB_H= token.h parse.h lib.h symbol.h scope.h
expression.h target.h \
LIB_OBJS= target.o parse.o tokenize.o pre-process.o symbol.o lib.o scope.o \
expression.o show-parse.o evaluate.o expand.o inline.o linearize.o \
char.o sort.o allocate.o compat-$(OS).o ptrlist.o \
+ dominator.o \
builtin.o \
stats.o \
flow.o cse.o simplify.o memops.o liveness.o storage.o unssa.o dissect.o
new file mode 100644
@@ -0,0 +1,155 @@
+/*
+ * Dominator - walk the CFG, construct dominator tree.
+ *
+ * Copyright (C) 2017 Christopher Li
+ */
+
+#include "dominator.h"
+#include "flow.h"
+
+int dominator_valid = 0;
+
+static void postorder_walk(struct basic_block *b, int generation,
struct basic_block_list **po, int *nr)
+{
+ struct basic_block *child;
+
+ b->generation = generation;
+ FOR_EACH_PTR(b->children, child) {
+ if (child->generation != generation && child->ep)
+ postorder_walk(child, generation, po, nr);
+ } END_FOR_EACH_PTR(child);
+ add_bb(po, b);
+ b->nr = --*nr;
+ b->idom = NULL;
+ if (b->df)
+ free_ptr_list(&b->df);
+}
+
+static struct basic_block* intersect(struct basic_block *b1, struct
basic_block *b2)
+{
+ while (b1 != b2) {
+ /*
+ * In the paper the nodes are numbered by post visit order.
+ * The entry point has the largest number.
+ *
+ * In this implementation the node number are order in
+ * the REVERSE post visit order. E.g. The entry point
+ * has the smallest node number. There for the compare
+ * is inverted here as well.
+ */
+ while (b1->nr > b2->nr) {
+ b1 = b1->idom;
+ }
+ while (b2->nr > b1->nr) {
+ b2 = b2->idom;
+ }
+ }
+ return b1;
+}
+
+/*
+ * A Simple, Fast Dominance Algorithm
+ * Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy
+ */
+static void compute_dominator(struct basic_block *root, struct
basic_block_list *po)
+{
+ struct basic_block *b;
+ int changed;
+ unsigned long generation = bb_generation;
+
+ root->idom = root;
+ do {
+ changed = 0;
+ /* reverse postorder except root */
+ FOR_EACH_PTR_REVERSE(po, b) {
+ struct basic_block *new_idom = NULL;
+ struct basic_block *p;
+ new_idom = NULL;
+ FOR_EACH_PTR(b->parents, p) {
+ /* skip unprocessed or not reachable node */
+ if (!p->idom || p->generation != generation)
+ continue;
+
+ if (!new_idom) {
+ new_idom = p;
+ continue;
+ }
+ new_idom = intersect(p, new_idom);
+ } END_FOR_EACH_PTR(p);
+ if (b->idom != new_idom) {
+ b->idom = new_idom;
+ changed = 1;
+ }
+ } END_FOR_EACH_PTR_REVERSE(b);
+ } while (changed);
+ root->idom = NULL;
+}
+
+static int test_dominace(struct basic_block *begin, struct basic_block *end)
+{
+ if (!begin->idom)
+ return 1; /* begin is root */
+ while (begin != end) {
+ struct basic_block *idom = end->idom;
+ if (!idom)
+ return 0; /* end reach to root */
+ end = idom;
+ }
+ return 1;
+}
+
+static void add_new_bb(struct basic_block_list **list, struct
basic_block *new_bb)
+{
+ struct basic_block *b;
+ FOR_EACH_PTR(*list, b) {
+ if (b == new_bb)
+ return;
+ } END_FOR_EACH_PTR(b);
+ add_bb(list, new_bb);
+}
+
+
+static void compute_dominace_frontier(struct basic_block *root,
struct basic_block_list *po)
+{
+ struct basic_block *b;
+ FOR_EACH_PTR(po, b) {
+ struct basic_block *idom = b->idom;
+ struct basic_block *y, *w;
+
+ /* This loop compute DF_local(b) */
+ FOR_EACH_PTR(b->children, y) {
+ if (y->idom != b)
+ add_new_bb(&b->df, y);
+ } END_FOR_EACH_PTR(y);
+
+ /* This loop compute b branch part of DF_up(idom) */
+ FOR_EACH_PTR(b->df, w) {
+ if (idom == w || !test_dominace(idom, w))
+ add_new_bb(&idom->df, w);
+ } END_FOR_EACH_PTR(w);
+ } END_FOR_EACH_PTR(b);
+}
+
+void find_dominace_frontier(struct entrypoint *ep)
+{
+ struct basic_block_list *po = NULL;
+ struct basic_block *root = ep->entry->bb;
+ int nr = bb_list_size(ep->bbs);
+
+ dominator_valid = 1;
+ postorder_walk(root, ++bb_generation, &po, &nr);
+ ep->nr = root->nr;
+
+ /* last entry is root */
+ delete_last_basic_block(&po);
+
+ /* return if just one root node */
+ if (!po)
+ return;
+
+ compute_dominator(root, po);
+ compute_dominace_frontier(root, po);
+
+ free_ptr_list(&po);
+}
+
new file mode 100644
@@ -0,0 +1,8 @@
+#ifndef DOMINATOR_H
+#define DOMINATOR_H
+
+#include "linearize.h"
+void find_dominace_frontier(struct entrypoint *ep);
+extern int dominator_valid;
+
+#endif
@@ -17,6 +17,7 @@
#include "linearize.h"
#include "flow.h"
#include "target.h"
+#include "dominator.h"
unsigned long bb_generation;
@@ -36,6 +37,7 @@ static int rewrite_branch(struct basic_block *bb,
/* We might find new if-conversions or non-dominating CSEs */
/* we may also create new dead cycles */
repeat_phase |= REPEAT_CSE | REPEAT_CFG_CLEANUP;
+ dominator_valid = 0;
*ptr = new;
replace_bb_in_list(&bb->children, old, new, 1);
remove_bb_from_list(&old->parents, bb, 1);
@@ -1003,6 +1005,7 @@ out:
* Merge the two.
*/
repeat_phase |= REPEAT_CSE;
+ dominator_valid = 0;
parent->children = bb->children;
bb->children = NULL;
@@ -21,6 +21,7 @@
#include "linearize.h"
#include "flow.h"
#include "target.h"
+#include "dominator.h"
pseudo_t linearize_statement(struct entrypoint *ep, struct statement *stmt);
pseudo_t linearize_expression(struct entrypoint *ep, struct expression *expr);
@@ -509,7 +510,6 @@ void show_bb(struct basic_block *bb)
FOR_EACH_PTR(bb->defines, defines) {
printf(" **defines %s **\n", show_pseudo(defines));
} END_FOR_EACH_PTR(defines);
-
if (bb->parents) {
struct basic_block *from;
FOR_EACH_PTR(bb->parents, from) {
@@ -525,6 +525,16 @@ void show_bb(struct basic_block *bb)
stream_name(to->pos.stream), to->pos.line, to->pos.pos);
} END_FOR_EACH_PTR(to);
}
+ if (bb->idom)
+ printf(" **idom .L%u\n", bb->idom->nr);
+ if (bb->df) {
+ struct basic_block *df;
+ FOR_EACH_PTR(bb->df, df) {
+ printf(" **df .L%u (%s:%d:%d)**\n", df->nr,
+ stream_name(df->pos.stream), df->pos.line, df->pos.pos);
+ } END_FOR_EACH_PTR(df);
+ }
+
}
FOR_EACH_PTR(bb->insns, insn) {
@@ -674,6 +684,7 @@ void insert_branch(struct basic_block *bb, struct
instruction *jmp, struct basic
remove_parent(child, bb);
} END_FOR_EACH_PTR(child);
PACK_PTR_LIST(&bb->children);
+ dominator_valid = 0;
}
@@ -2226,12 +2237,15 @@ static struct entrypoint *linearize_fn(struct
symbol *sym, struct symbol *base_t
show_entry(ep);
}
+
/*
* Do trivial flow simplification - branches to
* branches, kill dead basicblocks etc
*/
kill_unreachable_bbs(ep);
+ dominator_valid = 0;
+
/*
* Turn symbols into pseudos
*/
@@ -227,6 +227,8 @@ struct basic_block {
unsigned long generation;
int context;
struct entrypoint *ep;
+ struct basic_block *idom; /* immediate dominator */
+ struct basic_block_list *df; /* dominance frontier */
struct basic_block_list *parents; /* sources */
struct basic_block_list *children; /* destinations */
struct instruction_list *insns; /* Linear list of instructions */
@@ -326,6 +328,7 @@ struct entrypoint {
struct basic_block_list *bbs;
struct basic_block *active;
struct instruction *entry;
+ int nr;
};
extern void insert_select(struct basic_block *bb, struct instruction
*br, struct instruction *phi, pseudo_t if_true, pseudo_t if_false);
@@ -15,6 +15,7 @@
#include "expression.h"
#include "linearize.h"
#include "flow.h"
+#include "dominator.h"
static int find_dominating_parents(pseudo_t pseudo, struct instruction *insn,
struct basic_block *bb, unsigned long generation, struct pseudo_list
Here is the git branch https://git.kernel.org/pub/scm/devel/sparse/chrisl/sparse.git/log/?h=dominator-flat This is what I have been working on for a while. I want to try some textbook way to fix up the SSA issue in the sparse. A big step towards SSA conversion is calculating the dominance frontier, that is where the phi node needs to be place on. This patch is actually not big and runs very fast. For the full kernel build, this is the base line: 1278.08user 510.61system 1:22.54elapsed 2166%CPU (0avgtext+0avgdata 238344maxresident)k 16inputs+13816outputs (0major+135777070minor)pagefaults 0swaps 1281.58user 511.58system 1:22.61elapsed 2170%CPU (0avgtext+0avgdata 238296maxresident)k 0inputs+13816outputs (0major+135776958minor)pagefaults 0swaps This is the dominator patch: 1291.94user 516.02system 1:23.24elapsed 2171%CPU (0avgtext+0avgdata 240832maxresident)k 8inputs+13824outputs (0major+136059495minor)pagefaults 0swaps 1289.89user 516.04system 1:23.15elapsed 2171%CPU (0avgtext+0avgdata 240820maxresident)k 0inputs+13816outputs (0major+136059689minor)pagefaults 0swaps So the performance slow down is about 0.76%. That is including rebuild dominator *tree* and DF when CFG changes. After this perform the proper SSA conversion on memory to register should be pretty straight forward. BTW, I really appreciate if some one can write some validation function to double check the output dominator tree and DF is correct and bug free. Chris Make memops use find_dominace_frontier When dominator_valid = 0, the DF need to be rebuild. --- Makefile | 1 + dominator.c | 155 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ dominator.h | 8 ++++ flow.c | 3 ++ linearize.c | 16 ++++++- linearize.h | 3 ++ memops.c | 4 ++ 7 files changed, 189 insertions(+), 1 deletion(-) create mode 100644 dominator.c create mode 100644 dominator.h **dominators, @@ -187,6 +188,9 @@ void simplify_memops(struct entrypoint *ep) { struct basic_block *bb; + if (!dominator_valid) + find_dominace_frontier(ep); + FOR_EACH_PTR_REVERSE(ep->bbs, bb) { simplify_loads(bb); } END_FOR_EACH_PTR_REVERSE(bb);