diff --git a/src/common/gsvector_sse.h b/src/common/gsvector_sse.h
index 77462d88b..e62246ace 100644
--- a/src/common/gsvector_sse.h
+++ b/src/common/gsvector_sse.h
@@ -2640,7 +2640,7 @@ public:
   template<s32 i>
   ALWAYS_INLINE GSVector8i srl() const
   {
-    return GSVector8i(_mm256_srli_si128(m, i));
+    return GSVector8i(_mm256_srli_si256(m, i));
   }
 
   template<s32 i>
@@ -2652,7 +2652,7 @@ public:
   template<s32 i>
   ALWAYS_INLINE GSVector8i sll() const
   {
-    return GSVector8i(_mm256_slli_si128(m, i));
+    return GSVector8i(_mm256_slli_si256(m, i));
   }
 
   template<s32 i>
diff --git a/src/core/gpu_sw_rasterizer.cpp b/src/core/gpu_sw_rasterizer.cpp
index 87fd65134..f2648dcd0 100644
--- a/src/core/gpu_sw_rasterizer.cpp
+++ b/src/core/gpu_sw_rasterizer.cpp
@@ -6,11 +6,19 @@
 
 #include "cpuinfo.h"
 
+#include "common/gsvector.h"
 #include "common/log.h"
 #include "common/string_util.h"
 
 LOG_CHANNEL(GPU_SW_Rasterizer);
 
+// Disable 256-bit. We emit that path in a separate file.
+// TODO: For those who are compiling with -march=native, probably only want to compile the 256-bit renderer.
+// Once it's done, anyway....
+#ifdef GSVECTOR_HAS_256
+#undef GSVECTOR_HAS_256
+#endif
+
 namespace GPU_SW_Rasterizer {
 constinit const DitherLUT g_dither_lut = []() constexpr {
   DitherLUT lut = {};