diff --git a/hal/stm32g0.c b/hal/stm32g0.c
index aea0fc8b..03d83380 100644
--- a/hal/stm32g0.c
+++ b/hal/stm32g0.c
@@ -142,11 +142,11 @@ int RAMFUNCTION hal_flash_write(uint32_t address, const uint8_t *data, int len)
         flash_clear_errors();
         if ((len - i >= 8) && ((((address + i) & 0x07) == 0) &&
                 ((((uint32_t)data) + i) & 0x07) == 0)) {
-            src = (uint32_t *)data;
-            dst = (uint32_t *)address;
+            src = (uint32_t *)(data + i);
+            dst = (uint32_t *)(address + i);
             flash_wait_complete();
-            dst[i >> 2] = src[i >> 2];
-            dst[(i >> 2) + 1] = src[(i >> 2) + 1];
+            dst[0] = src[0];
+            dst[1] = src[1];
             flash_wait_complete();
             i+=8;
         } else {
diff --git a/hal/stm32g4.c b/hal/stm32g4.c
index a880c0ad..3372607f 100644
--- a/hal/stm32g4.c
+++ b/hal/stm32g4.c
@@ -56,11 +56,11 @@ int RAMFUNCTION hal_flash_write(uint32_t address, const uint8_t *data, int len)
         flash_clear_errors();
         if ((len - i >= 8) && ((((address + i) & 0x07) == 0) &&
                 ((((uint32_t)data) + i) & 0x07) == 0)) {
-            src = (uint32_t *)data;
-            dst = (uint32_t *)address;
+            src = (uint32_t *)(data + i);
+            dst = (uint32_t *)(address + i);
             flash_wait_complete();
-            dst[i >> 2] = src[i >> 2];
-            dst[(i >> 2) + 1] = src[(i >> 2) + 1];
+            dst[0] = src[0];
+            dst[1] = src[1];
             flash_wait_complete();
             i += 8;
         } else {
diff --git a/tools/unit-tests/unit-stm32g4-write.c b/tools/unit-tests/unit-stm32g4-write.c
index c6eb00ae..2d51e6df 100644
--- a/tools/unit-tests/unit-stm32g4-write.c
+++ b/tools/unit-tests/unit-stm32g4-write.c
@@ -179,6 +179,25 @@ START_TEST(test_write_64_full_units)
 }
 END_TEST
 
+/* A write starting at an address with mod-8 residue 1: the byte-wise
+ * prefix leaves i = 7, where the fast path fires with i % 4 != 0.
+ * Pre-fix the double word was programmed 3 bytes below the intended
+ * address and the last 3 bytes of the unit were never programmed. */
+START_TEST(test_write_unaligned_start)
+{
+    int i;
+
+    ck_assert_int_eq(hal_flash_write((uint32_t)(uintptr_t)(g_flash_mem + 1),
+        g_data + 1, 16), 0);
+
+    ck_assert_int_eq(memcmp(g_flash_mem + 1, g_data + 1, 16), 0);
+    ck_assert_uint_eq(g_flash_mem[0], 0x12);
+    for (i = 17; i < FLASH_MEM_SZ; i++)
+        ck_assert_uint_eq(g_flash_mem[i], 0x12);
+    ck_assert_int_eq(canary_in_flash(), 0);
+}
+END_TEST
+
 Suite *stm32g4_write_suite(void)
 {
     Suite *s = suite_create("stm32g4-write");
@@ -189,6 +208,7 @@ Suite *stm32g4_write_suite(void)
     tcase_add_test(tc, test_write_58_partial_word_padded);
     tcase_add_test(tc, test_write_3_single_word_padded);
     tcase_add_test(tc, test_write_64_full_units);
+    tcase_add_test(tc, test_write_unaligned_start);
     suite_add_tcase(s, tc);
 
     return s;