diff --git a/library/alignment.h b/library/alignment.h
index 248f29bc7..a17001dd9 100644
--- a/library/alignment.h
+++ b/library/alignment.h
@@ -53,28 +53,45 @@ typedef uint16_t __packed mbedtls_uint16_unaligned_t;
 typedef uint32_t __packed mbedtls_uint32_unaligned_t;
 typedef uint64_t __packed mbedtls_uint64_unaligned_t;
 #elif defined(MBEDTLS_COMPILER_IS_GCC) && (MBEDTLS_GCC_VERSION >= 40504) && \
-    ((MBEDTLS_GCC_VERSION < 90300) || (!defined(MBEDTLS_EFFICIENT_UNALIGNED_ACCESS)))
+    ((MBEDTLS_GCC_VERSION < 60300) || (!defined(MBEDTLS_EFFICIENT_UNALIGNED_ACCESS)))
 /*
- * Old versions of gcc, depending on how the target is specified, may generate a branch to memcpy
- * for calls like `memcpy(dest, src, 4)` rather than generating some LDR or LDRB instructions
- * (similar for stores).
- * Recent versions where unaligned access is not enabled also do this.
+ * gcc may generate a branch to memcpy for calls like `memcpy(dest, src, 4)` rather than
+ * generating some LDR or LDRB instructions (similar for stores).
+ *
+ * This is architecture dependent: x86-64 seems fine even with old gcc; 32-bit Arm
+ * is affected. To keep it simple, we enable for all architectures.
+ *
+ * For versions of gcc < 5.4.0 this issue always happens.
+ * For gcc < 6.3.0, this issue happens at -O0
+ * For all versions, this issue happens iff unaligned access is not supported.
+ *
+ * For gcc 4.x, this implementation will generate byte-by-byte loads even if unaligned access is
+ * supported, which is correct but not optimal.
  *
  * For performance (and code size, in some cases), we want to avoid the branch and just generate
  * some inline load/store instructions since the access is small and constant-size.
  *
  * The manual states:
- * "The aligned attribute specifies a minimum alignment for the variable or structure field,
- * measured in bytes."
- * https://gcc.gnu.org/onlinedocs/gcc/Common-Variable-Attributes.html
+ * "The packed attribute specifies that a variable or structure field should have the smallest
+ *  possible alignment—one byte for a variable"
+ * https://gcc.gnu.org/onlinedocs/gcc-4.5.4/gcc/Variable-Attributes.html
  *
- * Tested with several versions of GCC from 4.5.0 up to 9.3.0
+ * Previous implementations used __attribute__((__aligned__(1)), but had issues with a gcc bug:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94662
+ *
+ * Tested with several versions of GCC from 4.5.0 up to 13.2.0
  * We don't enable for older than 4.5.0 as this has not been tested.
  */
- #define UINT_UNALIGNED
-typedef uint16_t __attribute__((__aligned__(1))) mbedtls_uint16_unaligned_t;
-typedef uint32_t __attribute__((__aligned__(1))) mbedtls_uint32_unaligned_t;
-typedef uint64_t __attribute__((__aligned__(1))) mbedtls_uint64_unaligned_t;
+ #define UINT_UNALIGNED_STRUCT
+typedef struct {
+    uint16_t x;
+} __attribute__((packed)) mbedtls_uint16_unaligned_t;
+typedef struct {
+    uint32_t x;
+} __attribute__((packed)) mbedtls_uint32_unaligned_t;
+typedef struct {
+    uint64_t x;
+} __attribute__((packed)) mbedtls_uint64_unaligned_t;
  #endif
 
 /*
@@ -101,6 +118,9 @@ static inline uint16_t mbedtls_get_unaligned_uint16(const void *p)
 #if defined(UINT_UNALIGNED)
     mbedtls_uint16_unaligned_t *p16 = (mbedtls_uint16_unaligned_t *) p;
     r = *p16;
+#elif defined(UINT_UNALIGNED_STRUCT)
+    mbedtls_uint16_unaligned_t *p16 = (mbedtls_uint16_unaligned_t *) p;
+    r = p16->x;
 #else
     memcpy(&r, p, sizeof(r));
 #endif
@@ -124,6 +144,9 @@ static inline void mbedtls_put_unaligned_uint16(void *p, uint16_t x)
 #if defined(UINT_UNALIGNED)
     mbedtls_uint16_unaligned_t *p16 = (mbedtls_uint16_unaligned_t *) p;
     *p16 = x;
+#elif defined(UINT_UNALIGNED_STRUCT)
+    mbedtls_uint16_unaligned_t *p16 = (mbedtls_uint16_unaligned_t *) p;
+    p16->x = x;
 #else
     memcpy(p, &x, sizeof(x));
 #endif
@@ -147,6 +170,9 @@ static inline uint32_t mbedtls_get_unaligned_uint32(const void *p)
 #if defined(UINT_UNALIGNED)
     mbedtls_uint32_unaligned_t *p32 = (mbedtls_uint32_unaligned_t *) p;
     r = *p32;
+#elif defined(UINT_UNALIGNED_STRUCT)
+    mbedtls_uint32_unaligned_t *p32 = (mbedtls_uint32_unaligned_t *) p;
+    r = p32->x;
 #else
     memcpy(&r, p, sizeof(r));
 #endif
@@ -170,6 +196,9 @@ static inline void mbedtls_put_unaligned_uint32(void *p, uint32_t x)
 #if defined(UINT_UNALIGNED)
     mbedtls_uint32_unaligned_t *p32 = (mbedtls_uint32_unaligned_t *) p;
     *p32 = x;
+#elif defined(UINT_UNALIGNED_STRUCT)
+    mbedtls_uint32_unaligned_t *p32 = (mbedtls_uint32_unaligned_t *) p;
+    p32->x = x;
 #else
     memcpy(p, &x, sizeof(x));
 #endif
@@ -193,6 +222,9 @@ static inline uint64_t mbedtls_get_unaligned_uint64(const void *p)
 #if defined(UINT_UNALIGNED)
     mbedtls_uint64_unaligned_t *p64 = (mbedtls_uint64_unaligned_t *) p;
     r = *p64;
+#elif defined(UINT_UNALIGNED_STRUCT)
+    mbedtls_uint64_unaligned_t *p64 = (mbedtls_uint64_unaligned_t *) p;
+    r = p64->x;
 #else
     memcpy(&r, p, sizeof(r));
 #endif
@@ -216,6 +248,9 @@ static inline void mbedtls_put_unaligned_uint64(void *p, uint64_t x)
 #if defined(UINT_UNALIGNED)
     mbedtls_uint64_unaligned_t *p64 = (mbedtls_uint64_unaligned_t *) p;
     *p64 = x;
+#elif defined(UINT_UNALIGNED_STRUCT)
+    mbedtls_uint64_unaligned_t *p64 = (mbedtls_uint64_unaligned_t *) p;
+    p64->x = x;
 #else
     memcpy(p, &x, sizeof(x));
 #endif