diff --git a/GL/platforms/sh4.c b/GL/platforms/sh4.c
index 33a5f8f..78e1772 100644
--- a/GL/platforms/sh4.c
+++ b/GL/platforms/sh4.c
@@ -100,7 +100,7 @@ GL_FORCE_INLINE void _glSubmitHeaderOrVertex(const Vertex* v) {
     d += 8;
 }
 
-static struct {
+static struct __attribute__((aligned(32))) {
     Vertex* v;
     int visible;
 } triangle[3];
@@ -182,6 +182,8 @@ GL_FORCE_INLINE void ShiftRotateTriangle() {
 #define SPAN_SORT_CFG 0x005F8030
 
 void SceneListSubmit(void* src, int n) {
+    Vertex __attribute__((aligned(32))) tmp;
+
     /* Do everything, everywhere, all at once */
     PVR_SET(SPAN_SORT_CFG, 0x0);
 
diff --git a/GL/private.h b/GL/private.h
index b61f29f..9fa45cf 100644
--- a/GL/private.h
+++ b/GL/private.h
@@ -251,7 +251,7 @@ typedef struct {
  * when a realloc could invalidate pointers. This structure holds all the information
  * we need on the target vertex array to allow passing around to the various stages (e.g. generate/clip etc.)
  */
-typedef struct {
+typedef struct __attribute__((aligned(32))) {
     PolyList* output;
     uint32_t header_offset; // The offset of the header in the output list
     uint32_t start_offset; // The offset into the output list
diff --git a/containers/aligned_vector.h b/containers/aligned_vector.h
index cf89ffe..12f3ae3 100644
--- a/containers/aligned_vector.h
+++ b/containers/aligned_vector.h
@@ -25,10 +25,10 @@ static inline void* memalign(size_t alignment, size_t size) {
 #endif
 
 typedef struct {
-    unsigned int size;
-    unsigned int capacity;
-    unsigned char* data;
-    unsigned int element_size;
+    uint8_t* __attribute__((aligned(32))) data;
+    uint32_t size;
+    uint32_t capacity;
+    uint32_t element_size;
 } AlignedVector;
 
 #define ALIGNED_VECTOR_CHUNK_SIZE 256u