diff --git a/platformio.ini b/platformio.ini
index 522db27..b0874a4 100644
--- a/platformio.ini
+++ b/platformio.ini
@@ -17,7 +17,7 @@ upload_protocol = stlink
 ; 1.80201.181220 ; 921kb/sec
 ; 1.90201.191206 ; 912kb/sec
 ; 1.90301.200702 ; default - 955kb/sec
-platform_packages = toolchain-gccarmnoneeabi@1.60301.0
+platform_packages = toolchain-gccarmnoneeabi
 
 build_unflags = 
     -Os
diff --git a/src/BlueSCSI.cpp b/src/BlueSCSI.cpp
index 3b5a5a7..bfce7b0 100644
--- a/src/BlueSCSI.cpp
+++ b/src/BlueSCSI.cpp
@@ -857,7 +857,7 @@ void readDataPhase(int len, byte* p)
 /*
  * See writeDataLoop for optimization info.
  */
-void readDataLoop(uint32_t blockSize) __attribute__ ((aligned(8)));
+void readDataLoop(uint32_t blockSize) __attribute__ ((aligned(16)));
 void readDataLoop(uint32_t blockSize)
 {
   register byte *dstptr= m_buf;
@@ -878,7 +878,7 @@ void readDataLoop(uint32_t blockSize)
     REQ_OFF();
     *dstptr++ = ~(ret >> 8);
     // Move wait loop in to a single 8 byte prefetch buffer
-    asm("nop.w");
+    asm("nop.w;nop");
     WAIT_ACK_INACTIVE();
     REQ_ON();
     // Extra 1 cycle delay