From 9669e0fe941318a729c363a570121a8d110ce27a Mon Sep 17 00:00:00 2001
From: Justin Viiret <justin.viiret@intel.com>
Date: Fri, 10 Jun 2016 11:49:08 +1000
Subject: [PATCH] eod: remove forced sparse iter optimization

---
 src/rose/rose.h                  | 39 +-------------------------------
 src/rose/rose_build_bytecode.cpp | 16 ++++---------
 src/rose/rose_dump.cpp           |  1 -
 src/rose/rose_internal.h         |  1 -
 4 files changed, 6 insertions(+), 51 deletions(-)

diff --git a/src/rose/rose.h b/src/rose/rose.h
index e90d2f21..d79c2f0c 100644
--- a/src/rose/rose.h
+++ b/src/rose/rose.h
@@ -43,39 +43,6 @@ void roseBlockEodExec(const struct RoseEngine *t, u64a offset,
                       struct hs_scratch *scratch);
 void roseBlockExec_i(const struct RoseEngine *t, struct hs_scratch *scratch);
 
-static really_inline
-int roseBlockHasEodWork(const struct RoseEngine *t,
-                        struct hs_scratch *scratch) {
-    if (t->ematcherOffset) {
-        DEBUG_PRINTF("eod matcher to run\n");
-        return 1;
-    }
-
-    if (t->eodProgramOffset) {
-        DEBUG_PRINTF("has eod program\n");
-        return 1;
-    }
-
-    void *state = scratch->core_info.state;
-    if (mmbit_any(getActiveLeafArray(t, state), t->activeArrayCount)) {
-        DEBUG_PRINTF("active outfix/suffix engines\n");
-        return 1;
-    }
-
-    if (t->eodIterOffset) {
-        u32 idx;
-        const struct mmbit_sparse_iter *it = getByOffset(t, t->eodIterOffset);
-        struct mmbit_sparse_state si_state[MAX_SPARSE_ITER_STATES];
-        if (mmbit_sparse_iter_begin(getRoleState(state), t->rolesWithStateCount,
-                                    &idx, it, si_state) != MMB_INVALID) {
-            DEBUG_PRINTF("eod iter has states on\n");
-            return 1;
-        }
-    }
-
-    return 0;
-}
-
 /* assumes core_info in scratch has been init to point to data */
 static really_inline
 void roseBlockExec(const struct RoseEngine *t, struct hs_scratch *scratch) {
@@ -102,6 +69,7 @@ void roseBlockExec(const struct RoseEngine *t, struct hs_scratch *scratch) {
     roseBlockExec_i(t, scratch);
 
     if (!t->requiresEodCheck) {
+        DEBUG_PRINTF("no eod check required\n");
         return;
     }
 
@@ -110,11 +78,6 @@ void roseBlockExec(const struct RoseEngine *t, struct hs_scratch *scratch) {
         return;
     }
 
-    if (!roseBlockHasEodWork(t, scratch)) {
-        DEBUG_PRINTF("no eod work\n");
-        return;
-    }
-
     roseBlockEodExec(t, length, scratch);
 }
 
diff --git a/src/rose/rose_build_bytecode.cpp b/src/rose/rose_build_bytecode.cpp
index b37b259a..5d235fca 100644
--- a/src/rose/rose_build_bytecode.cpp
+++ b/src/rose/rose_build_bytecode.cpp
@@ -3989,7 +3989,7 @@ vector<RoseInstruction> makeEodAnchorProgram(RoseBuildImpl &build,
  * Returns the pair (program offset, sparse iter offset).
  */
 static
-pair<u32, u32> buildEodAnchorProgram(RoseBuildImpl &build, build_context &bc) {
+u32 writeEodAnchorProgram(RoseBuildImpl &build, build_context &bc) {
     const RoseGraph &g = build.g;
 
     // pred state id -> list of programs
@@ -4030,18 +4030,15 @@ pair<u32, u32> buildEodAnchorProgram(RoseBuildImpl &build, build_context &bc) {
 
     if (predProgramLists.empty()) {
         DEBUG_PRINTF("no eod anchored roles\n");
-        return {0, 0};
+        return 0;
     }
 
     vector<RoseInstruction> program;
-
-    // Note: we force the use of a sparse iterator for the EOD program so we
-    // can easily guard EOD execution at runtime.
-    u32 iter_offset = addPredBlocks(bc, predProgramLists, program, true);
+    addPredBlocks(bc, predProgramLists, program, false);
 
     assert(program.size() > 1);
     applyFinalSpecialisation(program);
-    return {writeProgram(bc, program), iter_offset};
+    return writeProgram(bc, program);
 }
 
 static
@@ -4311,9 +4308,7 @@ aligned_unique_ptr<RoseEngine> RoseBuildImpl::buildFinalEngine(u32 minWidth) {
         buildLiteralPrograms(*this, bc);
 
     u32 eodProgramOffset = writeEodProgram(*this, bc, eodNfaIterOffset);
-    u32 eodIterProgramOffset;
-    u32 eodIterOffset;
-    tie(eodIterProgramOffset, eodIterOffset) = buildEodAnchorProgram(*this, bc);
+    u32 eodIterProgramOffset = writeEodAnchorProgram(*this, bc);
 
     vector<mmbit_sparse_iter> activeLeftIter;
     buildActiveLeftIter(leftInfoTable, activeLeftIter);
@@ -4511,7 +4506,6 @@ aligned_unique_ptr<RoseEngine> RoseBuildImpl::buildFinalEngine(u32 minWidth) {
 
     engine->eodProgramOffset = eodProgramOffset;
     engine->eodIterProgramOffset = eodIterProgramOffset;
-    engine->eodIterOffset = eodIterOffset;
 
     engine->lastByteHistoryIterOffset = lastByteOffset;
 
diff --git a/src/rose/rose_dump.cpp b/src/rose/rose_dump.cpp
index 59f7f751..8de33a44 100644
--- a/src/rose/rose_dump.cpp
+++ b/src/rose/rose_dump.cpp
@@ -1026,7 +1026,6 @@ void roseDumpStructRaw(const RoseEngine *t, FILE *f) {
     DUMP_U32(t, lookaroundReachOffset);
     DUMP_U32(t, eodProgramOffset);
     DUMP_U32(t, eodIterProgramOffset);
-    DUMP_U32(t, eodIterOffset);
     DUMP_U32(t, lastByteHistoryIterOffset);
     DUMP_U32(t, minWidth);
     DUMP_U32(t, minWidthExcludingBoundaries);
diff --git a/src/rose/rose_internal.h b/src/rose/rose_internal.h
index 2e921542..faab45f7 100644
--- a/src/rose/rose_internal.h
+++ b/src/rose/rose_internal.h
@@ -378,7 +378,6 @@ struct RoseEngine {
 
     u32 eodProgramOffset; //!< Unconditional EOD program, otherwise 0.
     u32 eodIterProgramOffset; // or 0 if no eod iterator program
-    u32 eodIterOffset; // offset to EOD sparse iter or 0 if none
 
     u32 lastByteHistoryIterOffset; // if non-zero