From 4f32a167d53f7758c2b2a1befb06d7f504f161e0 Mon Sep 17 00:00:00 2001
From: Justin Viiret <justin.viiret@intel.com>
Date: Tue, 23 May 2017 14:40:04 +1000
Subject: [PATCH] teddy: align major structures to cachelines

---
 src/fdr/teddy_avx2.c           | 10 +++++-----
 src/fdr/teddy_compile.cpp      | 32 ++++++++++++++++++--------------
 src/fdr/teddy_runtime_common.h | 10 +++++-----
 3 files changed, 28 insertions(+), 24 deletions(-)

diff --git a/src/fdr/teddy_avx2.c b/src/fdr/teddy_avx2.c
index 299825cc..38ac3f72 100644
--- a/src/fdr/teddy_avx2.c
+++ b/src/fdr/teddy_avx2.c
@@ -196,14 +196,14 @@ m256 prep_conf_fat_teddy_m4(const m256 *maskBase, m256 *old_1, m256 *old_2,
 }
 
 static really_inline
-const m256 * getMaskBase_avx2(const struct Teddy *teddy) {
-    return (const m256 *)((const u8 *)teddy + sizeof(struct Teddy));
+const m256 *getMaskBase_avx2(const struct Teddy *teddy) {
+    return (const m256 *)((const u8 *)teddy + ROUNDUP_CL(sizeof(struct Teddy)));
 }
 
 static really_inline
-const u32 * getConfBase_avx2(const struct Teddy *teddy, u8 numMask) {
-    return (const u32 *)((const u8 *)teddy + sizeof(struct Teddy) +
-                         (numMask*32*2));
+const u32 *getConfBase_avx2(const struct Teddy *teddy, u8 numMask) {
+    return (const u32 *)((const u8 *)teddy + ROUNDUP_CL(sizeof(struct Teddy)) +
+                         ROUNDUP_CL((numMask * 32 * 2)));
 }
 
 hwlm_error_t fdr_exec_teddy_avx2_msks1_fat(const struct FDR *fdr,
diff --git a/src/fdr/teddy_compile.cpp b/src/fdr/teddy_compile.cpp
index 6f956e8c..19e595fb 100644
--- a/src/fdr/teddy_compile.cpp
+++ b/src/fdr/teddy_compile.cpp
@@ -313,35 +313,39 @@ bytecode_ptr<FDR> TeddyCompiler::build() {
     }
     u32 maskWidth = eng.getNumBuckets() / 8;
 
-    size_t maskLen = eng.numMasks * 16 * 2 * maskWidth;
+    size_t headerSize = ROUNDUP_CL(sizeof(Teddy));
+    size_t maskLen = ROUNDUP_CL(eng.numMasks * 16 * 2 * maskWidth);
 
-    auto floodControlTmp = setupFDRFloodControl(lits, eng, grey);
-    auto confirmTmp = setupFullConfs(lits, eng, bucketToLits, make_small);
+    auto floodTable = setupFDRFloodControl(lits, eng, grey);
+    auto confirmTable = setupFullConfs(lits, eng, bucketToLits, make_small);
 
-    size_t size = ROUNDUP_N(sizeof(Teddy) +
-                            maskLen +
-                            confirmTmp.size() +
-                            floodControlTmp.size(),
-                            16 * maskWidth);
+    size_t size = headerSize + maskLen + ROUNDUP_CL(confirmTable.size()) +
+                  floodTable.size();
 
     auto fdr = make_zeroed_bytecode_ptr<FDR>(size, 64);
     assert(fdr); // otherwise would have thrown std::bad_alloc
     Teddy *teddy = (Teddy *)fdr.get(); // ugly
     u8 *teddy_base = (u8 *)teddy;
 
+    // Write header.
     teddy->size = size;
     teddy->engineID = eng.getID();
     teddy->maxStringLen = verify_u32(maxLen(lits));
 
-    u8 *ptr = teddy_base + sizeof(Teddy) + maskLen;
-    memcpy(ptr, confirmTmp.get(), confirmTmp.size());
-    ptr += confirmTmp.size();
+    // Write confirm structures.
+    u8 *ptr = teddy_base + headerSize + maskLen;
+    assert(ISALIGNED_CL(ptr));
+    memcpy(ptr, confirmTable.get(), confirmTable.size());
+    ptr += ROUNDUP_CL(confirmTable.size());
 
+    // Write flood control structures.
+    assert(ISALIGNED_CL(ptr));
     teddy->floodOffset = verify_u32(ptr - teddy_base);
-    memcpy(ptr, floodControlTmp.get(), floodControlTmp.size());
-    ptr += floodControlTmp.size();
+    memcpy(ptr, floodTable.get(), floodTable.size());
+    ptr += floodTable.size();
 
-    u8 *baseMsk = teddy_base + sizeof(Teddy);
+    // Write teddy masks.
+    u8 *baseMsk = teddy_base + headerSize;
 
     for (const auto &b2l : bucketToLits) {
         const u32 &bucket_id = b2l.first;
diff --git a/src/fdr/teddy_runtime_common.h b/src/fdr/teddy_runtime_common.h
index c5f0885f..883a68fc 100644
--- a/src/fdr/teddy_runtime_common.h
+++ b/src/fdr/teddy_runtime_common.h
@@ -239,14 +239,14 @@ void do_confWithBitMany_teddy(TEDDY_CONF_TYPE *conf, u8 bucket, u8 offset,
 }
 
 static really_inline
-const m128 * getMaskBase(const struct Teddy *teddy) {
-    return (const m128 *)((const u8 *)teddy + sizeof(struct Teddy));
+const m128 *getMaskBase(const struct Teddy *teddy) {
+    return (const m128 *)((const u8 *)teddy + ROUNDUP_CL(sizeof(struct Teddy)));
 }
 
 static really_inline
-const u32 * getConfBase(const struct Teddy *teddy, u8 numMask) {
-    return (const u32 *)((const u8 *)teddy + sizeof(struct Teddy) +
-                         (numMask*32));
+const u32 *getConfBase(const struct Teddy *teddy, u8 numMask) {
+    return (const u32 *)((const u8 *)teddy + ROUNDUP_CL(sizeof(struct Teddy)) +
+                         ROUNDUP_CL(numMask * 32));
 }
 
 #endif /* TEDDY_RUNTIME_COMMON_H_ */