From dbd373839d8843fb08f2f9d5f422c83bad36719b Mon Sep 17 00:00:00 2001
From: tevador <tevador@gmail.com>
Date: Mon, 25 Mar 2019 23:11:51 +0100
Subject: [PATCH] Native load from the cache

---
 src/blake2/endian.h |  6 ++++++
 src/dataset.cpp     | 16 ++++++++--------
 2 files changed, 14 insertions(+), 8 deletions(-)

diff --git a/src/blake2/endian.h b/src/blake2/endian.h
index fab1eed..631f9f6 100644
--- a/src/blake2/endian.h
+++ b/src/blake2/endian.h
@@ -60,6 +60,12 @@ static FORCE_INLINE uint64_t load64(const void *src) {
 #endif
 }
 
+static FORCE_INLINE uint64_t load64np(const void *src) {
+	uint64_t w;
+	memcpy(&w, src, sizeof w);
+	return w;
+}
+
 static FORCE_INLINE void store32(void *dst, uint32_t w) {
 #if defined(NATIVE_LITTLE_ENDIAN)
 	memcpy(dst, &w, sizeof w);
diff --git a/src/dataset.cpp b/src/dataset.cpp
index 40e72b1..c55c918 100644
--- a/src/dataset.cpp
+++ b/src/dataset.cpp
@@ -57,14 +57,14 @@ namespace RandomX {
 	}
 
 	static FORCE_INLINE void mixCache(uint8_t* mixBlock, uint64_t& c0, uint64_t& c1, uint64_t& c2, uint64_t& c3, uint64_t& c4, uint64_t& c5, uint64_t& c6, uint64_t& c7) {
-		c0 ^= load64(mixBlock + 0);
-		c1 ^= load64(mixBlock + 8);
-		c2 ^= load64(mixBlock + 16);
-		c3 ^= load64(mixBlock + 24);
-		c4 ^= load64(mixBlock + 32);
-		c5 ^= load64(mixBlock + 40);
-		c6 ^= load64(mixBlock + 48);
-		c7 ^= load64(mixBlock + 56);
+		c0 ^= load64np(mixBlock + 0);
+		c1 ^= load64np(mixBlock + 8);
+		c2 ^= load64np(mixBlock + 16);
+		c3 ^= load64np(mixBlock + 24);
+		c4 ^= load64np(mixBlock + 32);
+		c5 ^= load64np(mixBlock + 40);
+		c6 ^= load64np(mixBlock + 48);
+		c7 ^= load64np(mixBlock + 56);
 	}
 
 	void initBlock(const Cache& cache, uint8_t* out, uint64_t blockNumber, unsigned iterations) {