diff --git a/GL/platforms/sh4.h b/GL/platforms/sh4.h
index a2e3446..d84b25e 100644
--- a/GL/platforms/sh4.h
+++ b/GL/platforms/sh4.h
@@ -21,8 +21,8 @@
  * destination is properly aligned so we assert that. */
 #define FASTCPY(dst, src, bytes) \
     do { \
-        if(bytes % 32 == 0 && (uintptr_t) src % 32 == 0) { \
-            assert((uintptr_t) dst % 32 == 0); \
+        if(bytes % 32 == 0 && ((uintptr_t) src % 4) == 0) { \
+            assert(((uintptr_t) dst) % 32 == 0); \
             sq_cpy(dst, src, bytes); \
         } else { \
             PERF_WARNING(); \
diff --git a/GL/texture.c b/GL/texture.c
index 6c5260a..4deb9d5 100644
--- a/GL/texture.c
+++ b/GL/texture.c
@@ -381,6 +381,10 @@ GLubyte _glInitTextures() {
     size_t vram_free = GPUMemoryAvailable();
     YALLOC_SIZE = vram_free - PVR_MEM_BUFFER_SIZE; /* Take all but 64kb VRAM */
     YALLOC_BASE = GPUMemoryAlloc(YALLOC_SIZE);
+
+    /* Ensure memory is aligned */
+    assert((uintptr_t) YALLOC_BASE % 32 == 0);
+
     yalloc_init(YALLOC_BASE, YALLOC_SIZE);
     return 1;
 }
diff --git a/GL/yalloc/yalloc_internals.h b/GL/yalloc/yalloc_internals.h
index b0ec62a..ffb70cb 100644
--- a/GL/yalloc/yalloc_internals.h
+++ b/GL/yalloc/yalloc_internals.h
@@ -7,6 +7,10 @@ typedef struct
 {
   uint32_t prev; // low bit set if free
   uint32_t next; // for used blocks: low bit set if unused header at the end
+
+  /* We need user data to be 32-byte aligned, so the header needs
+   * to be 32 bytes in size (as user data follows the header) */
+  uint8_t padding[32 - (sizeof(uint32_t) * 2)];
 } Header;
 
 // NOTE: We have 32bit aligned data and 16bit offsets where the lowest bit is used as flag. So we remove the low bit and shift by 1 to address 128k bytes with the 15bit significant offset bits.