Use GCC's may_alias attribute for unaligned memory access

25 files changed, 197 insertions, 150 deletions

diff --git a/arch/arm/chunkset_neon.c b/arch/arm/chunkset_neon.c
index da9d7f95b0..68c9fef699 100644
--- a/arch/arm/chunkset_neon.c
+++ b/arch/arm/chunkset_neon.c
@@ -5,6 +5,7 @@
 #ifdef ARM_NEON
 #include "neon_intrins.h"
 #include "zbuild.h"
+#include "zmemory.h"
 #include "arch/generic/chunk_permute_table.h"
 
 typedef uint8x16_t chunk_t;
@@ -31,21 +32,15 @@ static const lut_rem_pair perm_idx_lut[13] = {
 };
 
 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
-    uint16_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = vreinterpretq_u8_u16(vdupq_n_u16(tmp));
+    *chunk = vreinterpretq_u8_u16(vdupq_n_u16(zng_memread_2(from)));
 }
 
 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
-    uint32_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = vreinterpretq_u8_u32(vdupq_n_u32(tmp));
+    *chunk = vreinterpretq_u8_u32(vdupq_n_u32(zng_memread_4(from)));
 }
 
 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
-    uint64_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = vreinterpretq_u8_u64(vdupq_n_u64(tmp));
+    *chunk = vreinterpretq_u8_u64(vdupq_n_u64(zng_memread_8(from)));
 }
 
 #define CHUNKSIZE        chunksize_neon
diff --git a/arch/arm/compare256_neon.c b/arch/arm/compare256_neon.c
index 87d14c89c0..3d05152f34 100644
--- a/arch/arm/compare256_neon.c
+++ b/arch/arm/compare256_neon.c
@@ -4,7 +4,7 @@
  */
 
 #include "zbuild.h"
-#include "zutil_p.h"
+#include "zmemory.h"
 #include "deflate.h"
 #include "fallback_builtins.h"
 
diff --git a/arch/generic/Makefile.in b/arch/generic/Makefile.in
index 32c8242d02..15d51d3135 100644
--- a/arch/generic/Makefile.in
+++ b/arch/generic/Makefile.in
@@ -40,10 +40,10 @@ chunkset_c.o: $(SRCDIR)/chunkset_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/chunkset_tpl.
 chunkset_c.lo: $(SRCDIR)/chunkset_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/chunkset_tpl.h $(SRCTOP)/inffast_tpl.h
 	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/chunkset_c.c
 
-compare256_c.o: $(SRCDIR)/compare256_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/zutil_p.h $(SRCTOP)/deflate.h $(SRCTOP)/fallback_builtins.h
+compare256_c.o: $(SRCDIR)/compare256_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/zmemory.h $(SRCTOP)/deflate.h $(SRCTOP)/fallback_builtins.h
 	$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/compare256_c.c
 
-compare256_c.lo: $(SRCDIR)/compare256_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/zutil_p.h $(SRCTOP)/deflate.h $(SRCTOP)/fallback_builtins.h
+compare256_c.lo: $(SRCDIR)/compare256_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/zmemory.h $(SRCTOP)/deflate.h $(SRCTOP)/fallback_builtins.h
 	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/compare256_c.c
 
 crc32_braid_c.o: $(SRCDIR)/crc32_braid_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/crc32_braid_p.h $(SRCTOP)/crc32_braid_tbl.h
diff --git a/arch/generic/chunkset_c.c b/arch/generic/chunkset_c.c
index 7b2bb7ba36..0a585e6caa 100644
--- a/arch/generic/chunkset_c.c
+++ b/arch/generic/chunkset_c.c
@@ -3,6 +3,7 @@
  */
 
 #include "zbuild.h"
+#include "zmemory.h"
 
 typedef uint64_t chunk_t;
 
@@ -12,21 +13,20 @@ typedef uint64_t chunk_t;
 #define HAVE_CHUNKMEMSET_8
 
 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
-    uint8_t *dest = (uint8_t *)chunk;
-    memcpy(dest, from, sizeof(uint32_t));
-    memcpy(dest+4, from, sizeof(uint32_t));
+    uint32_t tmp = zng_memread_4(from);
+    *chunk = tmp | ((chunk_t)tmp << 32);
 }
 
 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
-    memcpy(chunk, from, sizeof(uint64_t));
+    *chunk = zng_memread_8(from);
 }
 
 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
-    memcpy(chunk, (uint8_t *)s, sizeof(uint64_t));
+    *chunk = zng_memread_8(s);
 }
 
 static inline void storechunk(uint8_t *out, chunk_t *chunk) {
-    memcpy(out, chunk, sizeof(uint64_t));
+    zng_memwrite_8(out, *chunk);
 }
 
 #define CHUNKSIZE        chunksize_c
diff --git a/arch/generic/compare256_c.c b/arch/generic/compare256_c.c
index d20c74ce84..ae0e717969 100644
--- a/arch/generic/compare256_c.c
+++ b/arch/generic/compare256_c.c
@@ -4,7 +4,7 @@
  */
 
 #include "zbuild.h"
-#include "zutil_p.h"
+#include "zmemory.h"
 #include "deflate.h"
 #include "fallback_builtins.h"
 
@@ -107,8 +107,8 @@ static inline uint32_t compare256_unaligned_32_static(const uint8_t *src0, const
     do {
         uint32_t sv, mv, diff;
 
-        memcpy(&sv, src0, sizeof(sv));
-        memcpy(&mv, src1, sizeof(mv));
+        sv = zng_memread_4(src0);
+        mv = zng_memread_4(src1);
 
         diff = sv ^ mv;
         if (diff) {
@@ -151,8 +151,8 @@ static inline uint32_t compare256_unaligned_64_static(const uint8_t *src0, const
     do {
         uint64_t sv, mv, diff;
 
-        memcpy(&sv, src0, sizeof(sv));
-        memcpy(&mv, src1, sizeof(mv));
+        sv = zng_memread_8(src0);
+        mv = zng_memread_8(src1);
 
         diff = sv ^ mv;
         if (diff) {
diff --git a/arch/power/chunkset_power8.c b/arch/power/chunkset_power8.c
index aef1973273..673fe0e112 100644
--- a/arch/power/chunkset_power8.c
+++ b/arch/power/chunkset_power8.c
@@ -5,6 +5,7 @@
 #ifdef POWER8_VSX
 #include <altivec.h>
 #include "zbuild.h"
+#include "zmemory.h"
 
 typedef vector unsigned char chunk_t;
 
@@ -15,21 +16,15 @@ typedef vector unsigned char chunk_t;
 #define HAVE_CHUNKMEMSET_8
 
 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
-    uint16_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = (vector unsigned char)vec_splats(tmp);
+    *chunk = (vector unsigned char)vec_splats(zng_memread_2(from));
 }
 
 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
-    uint32_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = (vector unsigned char)vec_splats(tmp);
+    *chunk = (vector unsigned char)vec_splats(zng_memread_4(from));
 }
 
 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
-    uint64_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = (vector unsigned char)vec_splats((unsigned long long)tmp);
+    *chunk = (vector unsigned char)vec_splats((unsigned long long)zng_memread_8(from));
 }
 
 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
diff --git a/arch/power/compare256_power9.c b/arch/power/compare256_power9.c
index c8be498e4f..2875719c47 100644
--- a/arch/power/compare256_power9.c
+++ b/arch/power/compare256_power9.c
@@ -6,7 +6,7 @@
 #ifdef POWER9
 #include <altivec.h>
 #include "zbuild.h"
-#include "zutil_p.h"
+#include "zmemory.h"
 #include "deflate.h"
 #include "zendian.h"
 
diff --git a/arch/riscv/compare256_rvv.c b/arch/riscv/compare256_rvv.c
index 3d6c3e3aa5..3ddb4db080 100644
--- a/arch/riscv/compare256_rvv.c
+++ b/arch/riscv/compare256_rvv.c
@@ -7,7 +7,7 @@
 #ifdef RISCV_RVV
 
 #include "zbuild.h"
-#include "zutil_p.h"
+#include "zmemory.h"
 #include "deflate.h"
 #include "fallback_builtins.h"
 
diff --git a/arch/x86/chunkset_avx2.c b/arch/x86/chunkset_avx2.c
index bbcdcaea93..b9051bb90e 100644
--- a/arch/x86/chunkset_avx2.c
+++ b/arch/x86/chunkset_avx2.c
@@ -2,6 +2,7 @@
  * For conditions of distribution and use, see copyright notice in zlib.h
  */
 #include "zbuild.h"
+#include "zmemory.h"
 
 #ifdef X86_AVX2
 #include "avx2_tables.h"
@@ -19,21 +20,15 @@ typedef __m128i halfchunk_t;
 #define HAVE_HALF_CHUNK
 
 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
-    int16_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm256_set1_epi16(tmp);
+    *chunk = _mm256_set1_epi16(zng_memread_2(from));
 }
 
 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
-    int32_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm256_set1_epi32(tmp);
+    *chunk = _mm256_set1_epi32(zng_memread_4(from));
 }
 
 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
-    int64_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm256_set1_epi64x(tmp);
+    *chunk = _mm256_set1_epi64x(zng_memread_8(from));
 }
 
 static inline void chunkmemset_16(uint8_t *from, chunk_t *chunk) {
diff --git a/arch/x86/chunkset_avx512.c b/arch/x86/chunkset_avx512.c
index 3d51ad1d9b..929d04cdd9 100644
--- a/arch/x86/chunkset_avx512.c
+++ b/arch/x86/chunkset_avx512.c
@@ -2,6 +2,7 @@
  * For conditions of distribution and use, see copyright notice in zlib.h
  */
 #include "zbuild.h"
+#include "zmemory.h"
 
 #ifdef X86_AVX512
 
@@ -33,21 +34,15 @@ static inline mask_t gen_mask(unsigned len) {
 }
 
 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
-    int16_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm256_set1_epi16(tmp);
+    *chunk = _mm256_set1_epi16(zng_memread_2(from));
 }
 
 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
-    int32_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm256_set1_epi32(tmp);
+    *chunk = _mm256_set1_epi32(zng_memread_4(from));
 }
 
 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
-    int64_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm256_set1_epi64x(tmp);
+    *chunk = _mm256_set1_epi64x(zng_memread_8(from));
 }
 
 static inline void chunkmemset_16(uint8_t *from, chunk_t *chunk) {
diff --git a/arch/x86/chunkset_sse2.c b/arch/x86/chunkset_sse2.c
index 0b1593b5b5..dcfe2c70d2 100644
--- a/arch/x86/chunkset_sse2.c
+++ b/arch/x86/chunkset_sse2.c
@@ -3,6 +3,7 @@
  */
 
 #include "zbuild.h"
+#include "zmemory.h"
 
 #ifdef X86_SSE2
 #include <immintrin.h>
@@ -14,21 +15,15 @@ typedef __m128i chunk_t;
 #define HAVE_CHUNKMEMSET_8
 
 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
-    int16_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm_set1_epi16(tmp);
+    *chunk = _mm_set1_epi16(zng_memread_2(from));
 }
 
 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
-    int32_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm_set1_epi32(tmp);
+    *chunk = _mm_set1_epi32(zng_memread_4(from));
 }
 
 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
-    int64_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm_set1_epi64x(tmp);
+    *chunk = _mm_set1_epi64x(zng_memread_8(from));
 }
 
 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
diff --git a/arch/x86/chunkset_ssse3.c b/arch/x86/chunkset_ssse3.c
index deedb6ce3c..75b698c61b 100644
--- a/arch/x86/chunkset_ssse3.c
+++ b/arch/x86/chunkset_ssse3.c
@@ -3,6 +3,7 @@
  */
 
 #include "zbuild.h"
+#include "zmemory.h"
 
 #if defined(X86_SSSE3)
 #include <immintrin.h>
@@ -33,21 +34,15 @@ static const lut_rem_pair perm_idx_lut[13] = {
 
 
 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
-    int16_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm_set1_epi16(tmp);
+    *chunk = _mm_set1_epi16(zng_memread_2(from));
 }
 
 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
-    int32_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm_set1_epi32(tmp);
+    *chunk = _mm_set1_epi32(zng_memread_4(from));
 }
 
 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
-    int64_t tmp;
-    memcpy(&tmp, from, sizeof(tmp));
-    *chunk = _mm_set1_epi64x(tmp);
+    *chunk = _mm_set1_epi64x(zng_memread_8(from));
 }
 
 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
diff --git a/arch/x86/compare256_avx2.c b/arch/x86/compare256_avx2.c
index d2c835e4ee..8a0213c3a6 100644
--- a/arch/x86/compare256_avx2.c
+++ b/arch/x86/compare256_avx2.c
@@ -4,7 +4,7 @@
  */
 
 #include "zbuild.h"
-#include "zutil_p.h"
+#include "zmemory.h"
 #include "deflate.h"
 #include "fallback_builtins.h"
 
diff --git a/arch/x86/compare256_sse2.c b/arch/x86/compare256_sse2.c
index 216bb3a705..25b65316a8 100644
--- a/arch/x86/compare256_sse2.c
+++ b/arch/x86/compare256_sse2.c
@@ -4,7 +4,7 @@
  */
 
 #include "zbuild.h"
-#include "zutil_p.h"
+#include "zmemory.h"
 #include "deflate.h"
 #include "fallback_builtins.h"
 
diff --git a/compare256_rle.h b/compare256_rle.h
index c7a4086af6..9940a28495 100644
--- a/compare256_rle.h
+++ b/compare256_rle.h
@@ -4,6 +4,7 @@
  */
 
 #include "zbuild.h"
+#include "zmemory.h"
 #include "fallback_builtins.h"
 #include "zendian.h"
 
@@ -47,25 +48,21 @@ static inline uint32_t compare256_rle_c(const uint8_t *src0, const uint8_t *src1
 /* 16-bit unaligned integer comparison */
 static inline uint32_t compare256_rle_unaligned_16(const uint8_t *src0, const uint8_t *src1) {
     uint32_t len = 0;
-    uint16_t src0_cmp, src1_cmp;
+    uint16_t src0_cmp;
 
-    memcpy(&src0_cmp, src0, sizeof(src0_cmp));
+    src0_cmp = zng_memread_2(src0);
 
     do {
-        memcpy(&src1_cmp, src1, sizeof(src1_cmp));
-        if (src0_cmp != src1_cmp)
+        if (src0_cmp != zng_memread_2(src1))
             return len + (*src0 == *src1);
         src1 += 2, len += 2;
-        memcpy(&src1_cmp, src1, sizeof(src1_cmp));
-        if (src0_cmp != src1_cmp)
+        if (src0_cmp != zng_memread_2(src1))
             return len + (*src0 == *src1);
         src1 += 2, len += 2;
-        memcpy(&src1_cmp, src1, sizeof(src1_cmp));
-        if (src0_cmp != src1_cmp)
+        if (src0_cmp != zng_memread_2(src1))
             return len + (*src0 == *src1);
         src1 += 2, len += 2;
-        memcpy(&src1_cmp, src1, sizeof(src1_cmp));
-        if (src0_cmp != src1_cmp)
+        if (src0_cmp != zng_memread_2(src1))
             return len + (*src0 == *src1);
         src1 += 2, len += 2;
     } while (len < 256);
@@ -79,13 +76,13 @@ static inline uint32_t compare256_rle_unaligned_32(const uint8_t *src0, const ui
     uint32_t sv, len = 0;
     uint16_t src0_cmp;
 
-    memcpy(&src0_cmp, src0, sizeof(src0_cmp));
+    src0_cmp = zng_memread_2(src0);
     sv = ((uint32_t)src0_cmp << 16) | src0_cmp;
 
     do {
         uint32_t mv, diff;
 
-        memcpy(&mv, src1, sizeof(mv));
+        mv = zng_memread_4(src1);
 
         diff = sv ^ mv;
         if (diff) {
@@ -112,14 +109,14 @@ static inline uint32_t compare256_rle_unaligned_64(const uint8_t *src0, const ui
     uint16_t src0_cmp;
     uint64_t sv;
 
-    memcpy(&src0_cmp, src0, sizeof(src0_cmp));
+    src0_cmp = zng_memread_2(src0);
     src0_cmp32 = ((uint32_t)src0_cmp << 16) | src0_cmp;
     sv = ((uint64_t)src0_cmp32 << 32) | src0_cmp32;
 
     do {
         uint64_t mv, diff;
 
-        memcpy(&mv, src1, sizeof(mv));
+        mv = zng_memread_8(src1);
 
         diff = sv ^ mv;
         if (diff) {
diff --git a/deflate.h b/deflate.h
index 2077f068e7..4b79f8f43b 100644
--- a/deflate.h
+++ b/deflate.h
@@ -12,6 +12,7 @@
 
 #include "zutil.h"
 #include "zendian.h"
+#include "zmemory.h"
 #include "crc32.h"
 
 #ifdef S390_DFLTCC_DEFLATE
@@ -355,7 +356,7 @@ static inline void put_short(deflate_state *s, uint16_t w) {
 #if BYTE_ORDER == BIG_ENDIAN
     w = ZSWAP16(w);
 #endif
-    memcpy(&s->pending_buf[s->pending], &w, sizeof(w));
+    zng_memwrite_2(&s->pending_buf[s->pending], w);
     s->pending += 2;
 }
 
@@ -367,7 +368,7 @@ static inline void put_short_msb(deflate_state *s, uint16_t w) {
 #if BYTE_ORDER == LITTLE_ENDIAN
     w = ZSWAP16(w);
 #endif
-    memcpy(&s->pending_buf[s->pending], &w, sizeof(w));
+    zng_memwrite_2(&s->pending_buf[s->pending], w);
     s->pending += 2;
 }
 
@@ -379,7 +380,7 @@ static inline void put_uint32(deflate_state *s, uint32_t dw) {
 #if BYTE_ORDER == BIG_ENDIAN
     dw = ZSWAP32(dw);
 #endif
-    memcpy(&s->pending_buf[s->pending], &dw, sizeof(dw));
+    zng_memwrite_4(&s->pending_buf[s->pending], dw);
     s->pending += 4;
 }
 
@@ -391,7 +392,7 @@ static inline void put_uint32_msb(deflate_state *s, uint32_t dw) {
 #if BYTE_ORDER == LITTLE_ENDIAN
     dw = ZSWAP32(dw);
 #endif
-    memcpy(&s->pending_buf[s->pending], &dw, sizeof(dw));
+    zng_memwrite_4(&s->pending_buf[s->pending], dw);
     s->pending += 4;
 }
 
@@ -403,7 +404,7 @@ static inline void put_uint64(deflate_state *s, uint64_t lld) {
 #if BYTE_ORDER == BIG_ENDIAN
     lld = ZSWAP64(lld);
 #endif
-    memcpy(&s->pending_buf[s->pending], &lld, sizeof(lld));
+    zng_memwrite_8(&s->pending_buf[s->pending], lld);
     s->pending += 8;
 }
 
diff --git a/deflate_quick.c b/deflate_quick.c
index 91c96ac52e..d5fd986d7a 100644
--- a/deflate_quick.c
+++ b/deflate_quick.c
@@ -18,7 +18,7 @@
  */
 
 #include "zbuild.h"
-#include "zutil_p.h"
+#include "zmemory.h"
 #include "deflate.h"
 #include "deflate_p.h"
 #include "functable.h"
diff --git a/inflate_p.h b/inflate_p.h
index 54c8dec900..179f76f734 100644
--- a/inflate_p.h
+++ b/inflate_p.h
@@ -6,6 +6,7 @@
 #define INFLATE_P_H
 
 #include <stdlib.h>
+#include "zmemory.h"
 
 /* Architecture-specific hooks. */
 #ifdef S390_DFLTCC_INFLATE
@@ -138,8 +139,7 @@
 
 /* Load 64 bits from IN and place the bytes at offset BITS in the result. */
 static inline uint64_t load_64_bits(const unsigned char *in, unsigned bits) {
-    uint64_t chunk;
-    memcpy(&chunk, in, sizeof(chunk));
+    uint64_t chunk = zng_memread_8(in);
 
 #if BYTE_ORDER == LITTLE_ENDIAN
     return chunk << bits;
@@ -175,7 +175,10 @@ static inline uint8_t* chunkcopy_safe(uint8_t *out, uint8_t *from, uint64_t len,
     uint64_t non_olap_size = llabs(from - out); // llabs vs labs for compatibility with windows
 
     /* So this doesn't give use a worst case scenario of function calls in a loop,
-     * we want to instead break this down into copy blocks of fixed lengths */
+     * we want to instead break this down into copy blocks of fixed lengths
+     *
+     * TODO: The memcpy calls aren't inlined on architectures with strict memory alignment
+     */
     while (len) {
         tocopy = MIN(non_olap_size, len);
         len -= tocopy;
diff --git a/insert_string_tpl.h b/insert_string_tpl.h
index e7037c04e6..1548ca741a 100644
--- a/insert_string_tpl.h
+++ b/insert_string_tpl.h
@@ -22,6 +22,8 @@
  *
  */
 
+#include "zmemory.h"
+
 #ifndef HASH_CALC_OFFSET
 #  define HASH_CALC_OFFSET 0
 #endif
@@ -31,11 +33,10 @@
 #ifndef HASH_CALC_READ
 #  if BYTE_ORDER == LITTLE_ENDIAN
 #    define HASH_CALC_READ \
-        memcpy(&val, strstart, sizeof(val));
+        val = zng_memread_4(strstart);
 #  else
 #    define HASH_CALC_READ \
-        memcpy(&val, strstart, sizeof(val)); \
-        val = ZSWAP32(val);
+        val = ZSWAP32(zng_memread_4(strstart));
 #  endif
 #endif
 
diff --git a/match_tpl.h b/match_tpl.h
index f44da750fb..5d00bd0139 100644
--- a/match_tpl.h
+++ b/match_tpl.h
@@ -40,10 +40,15 @@ Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, Pos cur_match) {
     uint32_t chain_length, nice_match, best_len, offset;
     uint32_t lookahead = s->lookahead;
     Pos match_offset = 0;
-#if OPTIMAL_CMP >= 32
-    uint8_t scan_start[8];
-#endif
+#if OPTIMAL_CMP >= 64
+    uint64_t scan_start;
+    uint64_t scan_end;
+#elif OPTIMAL_CMP >= 32
+    uint32_t scan_start;
+    uint32_t scan_end;
+#else
     uint8_t scan_end[8];
+#endif
 
 #define GOTO_NEXT_CHAIN \
     if (--chain_length && (cur_match = prev[cur_match & wmask]) > limit) \
@@ -70,11 +75,11 @@ Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, Pos cur_match) {
 #endif
 
 #if OPTIMAL_CMP >= 64
-    memcpy(scan_start, scan, sizeof(uint64_t));
-    memcpy(scan_end, scan+offset, sizeof(uint64_t));
+    scan_start = zng_memread_8(scan);
+    scan_end = zng_memread_8(scan+offset);
 #elif OPTIMAL_CMP >= 32
-    memcpy(scan_start, scan, sizeof(uint32_t));
-    memcpy(scan_end, scan+offset, sizeof(uint32_t));
+    scan_start = zng_memread_4(scan);
+    scan_end = zng_memread_4(scan+offset);
 #else
     scan_end[0] = *(scan+offset);
     scan_end[1] = *(scan+offset+1);
@@ -141,24 +146,24 @@ Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, Pos cur_match) {
 #if OPTIMAL_CMP >= 32
         if (best_len < sizeof(uint32_t)) {
             for (;;) {
-                if (zng_memcmp_2(mbase_end+cur_match, scan_end) == 0 &&
-                    zng_memcmp_2(mbase_start+cur_match, scan_start) == 0)
+                if (zng_memcmp_2(mbase_end+cur_match, &scan_end) == 0 &&
+                    zng_memcmp_2(mbase_start+cur_match, &scan_start) == 0)
                     break;
                 GOTO_NEXT_CHAIN;
             }
 #  if OPTIMAL_CMP >= 64
         } else if (best_len >= sizeof(uint64_t)) {
             for (;;) {
-                if (zng_memcmp_8(mbase_end+cur_match, scan_end) == 0 &&
-                    zng_memcmp_8(mbase_start+cur_match, scan_start) == 0)
+                if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 &&
+                    zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0)
                     break;
                 GOTO_NEXT_CHAIN;
             }
 #  endif
         } else {
             for (;;) {
-                if (zng_memcmp_4(mbase_end+cur_match, scan_end) == 0 &&
-                    zng_memcmp_4(mbase_start+cur_match, scan_start) == 0)
+                if (zng_memcmp_4(mbase_end+cur_match, &scan_end) == 0 &&
+                    zng_memcmp_4(mbase_start+cur_match, &scan_start) == 0)
                     break;
                 GOTO_NEXT_CHAIN;
             }
@@ -197,9 +202,9 @@ Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, Pos cur_match) {
 #endif
 
 #if OPTIMAL_CMP >= 64
-            memcpy(scan_end, scan+offset, sizeof(uint64_t));
+            scan_end = zng_memread_8(scan+offset);
 #elif OPTIMAL_CMP >= 32
-            memcpy(scan_end, scan+offset, sizeof(uint32_t));
+            scan_end = zng_memread_4(scan+offset);
 #else
             scan_end[0] = *(scan+offset);
             scan_end[1] = *(scan+offset+1);
diff --git a/win32/Makefile.a64 b/win32/Makefile.a64
index 9f8d6fb7fa..3209f6a305 100644
--- a/win32/Makefile.a64
+++ b/win32/Makefile.a64
@@ -183,7 +183,7 @@ adler32.obj: $(TOP)/adler32.c $(TOP)/zbuild.h $(TOP)/functable.h $(TOP)/adler32_
 adler32_c.obj: $(TOP)/arch/generic/adler32_c.c $(TOP)/zbuild.h $(TOP)/functable.h $(TOP)/adler32_p.h
 adler32_fold_c.obj: $(TOP)/arch/generic/adler32_fold_c.c $(TOP)/zbuild.h $(TOP)/functable.h
 chunkset_c.obj: $(TOP)/arch/generic/chunkset_c.c $(TOP)/zbuild.h $(TOP)/chunkset_tpl.h $(TOP)/inffast_tpl.h
-compare256_c.obj: $(TOP)/arch/generic/compare256_c.c $(TOP)/zbuild.h $(TOP)/zutil_p.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
+compare256_c.obj: $(TOP)/arch/generic/compare256_c.c $(TOP)/zbuild.h $(TOP)/zmemory.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
 compress.obj: $(TOP)/compress.c $(TOP)/zbuild.h $(TOP)/zutil.h
 cpu_features.obj: $(TOP)/cpu_features.c $(TOP)/cpu_features.h $(TOP)/zbuild.h
 crc32.obj: $(TOP)/crc32.c $(TOP)/zbuild.h $(TOP)/functable.h $(TOP)/crc32_braid_tbl.h
@@ -194,7 +194,7 @@ deflate.obj: $(TOP)/deflate.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.
 deflate_fast.obj: $(TOP)/deflate_fast.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
 deflate_huff.obj: $(TOP)/deflate_huff.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
 deflate_medium.obj: $(TOP)/deflate_medium.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
-deflate_quick.obj: $(TOP)/deflate_quick.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h $(TOP)/trees_emit.h $(TOP)/zutil_p.h
+deflate_quick.obj: $(TOP)/deflate_quick.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h $(TOP)/trees_emit.h $(TOP)/zmemory.h
 deflate_rle.obj: $(TOP)/deflate_rle.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h $(TOP)/compare256_rle.h
 deflate_slow.obj: $(TOP)/deflate_slow.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
 deflate_stored.obj: $(TOP)/deflate_stored.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
diff --git a/win32/Makefile.arm b/win32/Makefile.arm
index cab999dfe0..54da045ffd 100644
--- a/win32/Makefile.arm
+++ b/win32/Makefile.arm
@@ -204,7 +204,7 @@ adler32.obj: $(TOP)/adler32.c $(TOP)/zbuild.h $(TOP)/functable.h $(TOP)/adler32_
 adler32_c.obj: $(TOP)/arch/generic/adler32_c.c $(TOP)/zbuild.h $(TOP)/functable.h $(TOP)/adler32_p.h
 adler32_fold_c.obj: $(TOP)/arch/generic/adler32_fold_c.c $(TOP)/zbuild.h $(TOP)/functable.h
 chunkset_c.obj: $(TOP)/arch/generic/chunkset_c.c $(TOP)/zbuild.h $(TOP)/chunkset_tpl.h $(TOP)/inffast_tpl.h
-compare256_c.obj: $(TOP)/arch/generic/compare256_c.c $(TOP)/zbuild.h $(TOP)/zutil_p.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
+compare256_c.obj: $(TOP)/arch/generic/compare256_c.c $(TOP)/zbuild.h $(TOP)/zmemory.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
 compress.obj: $(TOP)/compress.c $(TOP)/zbuild.h $(TOP)/zutil.h
 cpu_features.obj: $(TOP)/cpu_features.c $(TOP)/cpu_features.h $(TOP)/zbuild.h
 crc32.obj: $(TOP)/crc32.c $(TOP)/zbuild.h $(TOP)/functable.h $(TOP)/crc32_braid_tbl.h
@@ -215,7 +215,7 @@ deflate.obj: $(TOP)/deflate.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.
 deflate_fast.obj: $(TOP)/deflate_fast.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
 deflate_huff.obj: $(TOP)/deflate_huff.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
 deflate_medium.obj: $(TOP)/deflate_medium.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
-deflate_quick.obj: $(TOP)/deflate_quick.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h $(TOP)/trees_emit.h $(TOP)/zutil_p.h
+deflate_quick.obj: $(TOP)/deflate_quick.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h $(TOP)/trees_emit.h $(TOP)/zmemory.h
 deflate_rle.obj: $(TOP)/deflate_rle.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h $(TOP)/compare256_rle.h
 deflate_slow.obj: $(TOP)/deflate_slow.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
 deflate_stored.obj: $(TOP)/deflate_stored.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
diff --git a/win32/Makefile.msc b/win32/Makefile.msc
index 8392fe46e7..62ca621aef 100644
--- a/win32/Makefile.msc
+++ b/win32/Makefile.msc
@@ -212,9 +212,9 @@ chunkset_c.obj: $(TOP)/arch/generic/chunkset_c.c $(TOP)/zbuild.h $(TOP)/chunkset
 chunkset_avx2.obj: $(TOP)/arch/x86/chunkset_avx2.c $(TOP)/zbuild.h $(TOP)/chunkset_tpl.h $(TOP)/inffast_tpl.h $(TOP)/arch/generic/chunk_permute_table.h
 chunkset_sse2.obj: $(TOP)/arch/x86/chunkset_sse2.c $(TOP)/zbuild.h $(TOP)/chunkset_tpl.h $(TOP)/inffast_tpl.h
 chunkset_ssse3.obj: $(TOP)/arch/x86/chunkset_ssse3.c $(TOP)/zbuild.h $(TOP)/chunkset_tpl.h $(TOP)/inffast_tpl.h $(TOP)/arch/generic/chunk_permute_table.h
-compare256_c.obj: $(TOP)/arch/generic/compare256_c.c $(TOP)/zbuild.h $(TOP)/zutil_p.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
-compare256_avx2.obj: $(TOP)/arch/x86/compare256_avx2.c $(TOP)/zbuild.h $(TOP)/zutil_p.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
-compare256_sse2.obj: $(TOP)/arch/x86/compare256_sse2.c $(TOP)/zbuild.h $(TOP)/zutil_p.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
+compare256_c.obj: $(TOP)/arch/generic/compare256_c.c $(TOP)/zbuild.h $(TOP)/zmemory.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
+compare256_avx2.obj: $(TOP)/arch/x86/compare256_avx2.c $(TOP)/zbuild.h $(TOP)/zmemory.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
+compare256_sse2.obj: $(TOP)/arch/x86/compare256_sse2.c $(TOP)/zbuild.h $(TOP)/zmemory.h $(TOP)/deflate.h $(TOP)/fallback_builtins.h $(TOP)/match_tpl.h
 compress.obj: $(TOP)/compress.c $(TOP)/zbuild.h $(TOP)/zutil.h
 cpu_features.obj: $(TOP)/cpu_features.c $(TOP)/cpu_features.h $(TOP)/zbuild.h
 crc32.obj: $(TOP)/crc32.c $(TOP)/zbuild.h $(TOP)/functable.h $(TOP)/crc32_braid_tbl.h
@@ -226,7 +226,7 @@ deflate.obj: $(TOP)/deflate.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.
 deflate_fast.obj: $(TOP)/deflate_fast.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
 deflate_huff.obj: $(TOP)/deflate_huff.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
 deflate_medium.obj: $(TOP)/deflate_medium.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
-deflate_quick.obj: $(TOP)/deflate_quick.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h $(TOP)/trees_emit.h $(TOP)/zutil_p.h
+deflate_quick.obj: $(TOP)/deflate_quick.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h $(TOP)/trees_emit.h $(TOP)/zmemory.h
 deflate_rle.obj: $(TOP)/deflate_rle.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h $(TOP)/compare256_rle.h
 deflate_slow.obj: $(TOP)/deflate_slow.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
 deflate_stored.obj: $(TOP)/deflate_stored.c $(TOP)/zbuild.h $(TOP)/deflate.h $(TOP)/deflate_p.h $(TOP)/functable.h
diff --git a/zmemory.h b/zmemory.h
new file mode 100644
index 0000000000..99ffd9e633
--- /dev/null
+++ b/zmemory.h
@@ -0,0 +1,99 @@
+/* zmemory.h -- Private inline functions used internally in zlib-ng
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#ifndef _ZMEMORY_H
+#define _ZMEMORY_H
+
+#if defined(__GNUC__) && (__GNUC__ >= 4)
+#  define HAVE_MAY_ALIAS
+#endif
+
+static inline uint16_t zng_memread_2(const void *ptr) {
+#if defined(HAVE_MAY_ALIAS)
+    typedef struct { uint16_t val; } __attribute__ ((__packed__, __may_alias__)) unaligned_uint16_t;
+    return ((const unaligned_uint16_t *)ptr)->val;
+#else
+    uint16_t val;
+    memcpy(&val, ptr, sizeof(val));
+    return val;
+#endif
+}
+
+static inline uint32_t zng_memread_4(const void *ptr) {
+#if defined(HAVE_MAY_ALIAS)
+    typedef struct { uint32_t val; } __attribute__ ((__packed__, __may_alias__)) unaligned_uint32_t;
+    return ((const unaligned_uint32_t *)ptr)->val;
+#else
+    uint32_t val;
+    memcpy(&val, ptr, sizeof(val));
+    return val;
+#endif
+}
+
+static inline uint64_t zng_memread_8(const void *ptr) {
+#if defined(HAVE_MAY_ALIAS)
+    typedef struct { uint64_t val; } __attribute__ ((__packed__, __may_alias__)) unaligned_uint64_t;
+    return ((const unaligned_uint64_t *)ptr)->val;
+#else
+    uint64_t val;
+    memcpy(&val, ptr, sizeof(val));
+    return val;
+#endif
+}
+
+static inline void zng_memwrite_2(void *ptr, uint16_t val) {
+#if defined(HAVE_MAY_ALIAS)
+    typedef struct { uint16_t val; } __attribute__ ((__packed__, __may_alias__)) unaligned_uint16_t;
+    ((unaligned_uint16_t *)ptr)->val = val;
+#else
+    memcpy(ptr, &val, sizeof(val));
+#endif
+}
+
+static inline void zng_memwrite_4(void *ptr, uint32_t val) {
+#if defined(HAVE_MAY_ALIAS)
+    typedef struct { uint32_t val; } __attribute__ ((__packed__, __may_alias__)) unaligned_uint32_t;
+    ((unaligned_uint32_t *)ptr)->val = val;
+#else
+    memcpy(ptr, &val, sizeof(val));
+#endif
+}
+
+static inline void zng_memwrite_8(void *ptr, uint64_t val) {
+#if defined(HAVE_MAY_ALIAS)
+    typedef struct { uint64_t val; } __attribute__ ((__packed__, __may_alias__)) unaligned_uint64_t;
+    ((unaligned_uint64_t *)ptr)->val = val;
+#else
+    memcpy(ptr, &val, sizeof(val));
+#endif
+}
+
+/* Use zng_memread_* instead of memcmp to avoid older compilers not converting memcmp
+   calls to unaligned comparisons when unaligned access is supported. Use memcmp only when
+   unaligned support is not available to avoid an extra call to memcpy. */
+static inline int32_t zng_memcmp_2(const void *src0, const void *src1) {
+#if defined(HAVE_MAY_ALIAS) || OPTIMAL_CMP >= 16
+    return zng_memread_2(src0) != zng_memread_2(src1);
+#else
+    return memcmp(src0, src1, 2);
+#endif
+}
+
+static inline int32_t zng_memcmp_4(const void *src0, const void *src1) {
+#if defined(HAVE_MAY_ALIAS) || OPTIMAL_CMP >= 32
+    return zng_memread_4(src0) != zng_memread_4(src1);
+#else
+    return memcmp(src0, src1, 4);
+#endif
+}
+
+static inline int32_t zng_memcmp_8(const void *src0, const void *src1) {
+#if defined(HAVE_MAY_ALIAS) || OPTIMAL_CMP >= 64
+    return zng_memread_8(src0) != zng_memread_8(src1);
+#else
+    return memcmp(src0, src1, 8);
+#endif
+}
+
+#endif
diff --git a/zutil_p.h b/zutil_p.h
index 97799f0ce3..835e12f4de 100644
--- a/zutil_p.h
+++ b/zutil_p.h
@@ -43,33 +43,4 @@ static inline void zng_free(void *ptr) {
 #endif
 }
 
-/* Use memcpy instead of memcmp to avoid older compilers not converting memcmp calls to
-   unaligned comparisons when unaligned access is supported. */
-static inline int32_t zng_memcmp_2(const void *src0, const void *src1) {
-    uint16_t src0_cmp, src1_cmp;
-
-    memcpy(&src0_cmp, src0, sizeof(src0_cmp));
-    memcpy(&src1_cmp, src1, sizeof(src1_cmp));
-
-    return src0_cmp != src1_cmp;
-}
-
-static inline int32_t zng_memcmp_4(const void *src0, const void *src1) {
-    uint32_t src0_cmp, src1_cmp;
-
-    memcpy(&src0_cmp, src0, sizeof(src0_cmp));
-    memcpy(&src1_cmp, src1, sizeof(src1_cmp));
-
-    return src0_cmp != src1_cmp;
-}
-
-static inline int32_t zng_memcmp_8(const void *src0, const void *src1) {
-    uint64_t src0_cmp, src1_cmp;
-
-    memcpy(&src0_cmp, src0, sizeof(src0_cmp));
-    memcpy(&src1_cmp, src1, sizeof(src1_cmp));
-
-    return src0_cmp != src1_cmp;
-}
-
 #endif