From 8167ddea6cf12b364f59b8e55b8d14611557c1cd Mon Sep 17 00:00:00 2001
From: Glenn Anderson <github@mactcp.net>
Date: Wed, 1 Jun 2022 12:44:05 -0700
Subject: [PATCH 1/3] Move TRANSCEIVER_IO_SET out of writeDataLoop. Add
 matching TRANSCEIVER_IO_SET(vTR_DBP,TR_INPUT) to writeDataPhase() and
 writeDataPhaseSD()

---
 src/BlueSCSI.cpp | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/src/BlueSCSI.cpp b/src/BlueSCSI.cpp
index 1733ebd..c663f74 100644
--- a/src/BlueSCSI.cpp
+++ b/src/BlueSCSI.cpp
@@ -920,9 +920,6 @@ void writeDataLoop(uint32_t blocksize, const byte* srcptr)
   // Start the first bus cycle.
   FETCH_BSRR_DB();
   REQ_OFF_DB_SET(bsrr_val);
-#ifdef XCVR
-  TRANSCEIVER_IO_SET(vTR_DBP,TR_OUTPUT)
-#endif
   REQ_ON();
   FETCH_BSRR_DB();
   WAIT_ACK_ACTIVE();
@@ -959,9 +956,15 @@ void writeDataPhase(int len, const byte* p)
   LOGN("DATAIN PHASE");
   SCSI_PHASE_CHANGE(SCSI_PHASE_DATAIN);
   // Bus settle delay 400ns. Following code was measured at 800ns before REQ asserted. STM32F103.
+#ifdef XCVR
+  TRANSCEIVER_IO_SET(vTR_DBP,TR_OUTPUT)
+#endif
   SCSI_DB_OUTPUT()
   writeDataLoop(len, p);
   SCSI_DB_INPUT()
+#ifdef XCVR
+  TRANSCEIVER_IO_SET(vTR_DBP,TR_INPUT)
+#endif
 }
 
 /*
@@ -977,6 +980,9 @@ void writeDataPhaseSD(uint32_t adds, uint32_t len)
   uint64_t pos = (uint64_t)adds * m_img->m_blocksize;
   m_img->m_file.seekSet(pos);
 
+#ifdef XCVR
+  TRANSCEIVER_IO_SET(vTR_DBP,TR_OUTPUT)
+#endif
   SCSI_DB_OUTPUT()
   for(uint32_t i = 0; i < len; i++) {
       // Asynchronous reads will make it faster ...

From 1aecf323e1ac6e6545c3a326b2d7fc03cc7b225b Mon Sep 17 00:00:00 2001
From: Glenn Anderson <github@mactcp.net>
Date: Fri, 3 Jun 2022 16:11:53 -0700
Subject: [PATCH 2/3] Remove unnecessary SCSI_DB_INPUT() and
 TRANSCEIVER_IO_SET(vTR_DBP,TR_INPUT)

---
 src/BlueSCSI.cpp | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/src/BlueSCSI.cpp b/src/BlueSCSI.cpp
index c663f74..6d3ea4f 100644
--- a/src/BlueSCSI.cpp
+++ b/src/BlueSCSI.cpp
@@ -961,10 +961,6 @@ void writeDataPhase(int len, const byte* p)
 #endif
   SCSI_DB_OUTPUT()
   writeDataLoop(len, p);
-  SCSI_DB_INPUT()
-#ifdef XCVR
-  TRANSCEIVER_IO_SET(vTR_DBP,TR_INPUT)
-#endif
 }
 
 /*
@@ -992,10 +988,6 @@ void writeDataPhaseSD(uint32_t adds, uint32_t len)
 
     writeDataLoop(m_img->m_blocksize, m_buf);
   }
-  SCSI_DB_INPUT()
-#ifdef XCVR
-  TRANSCEIVER_IO_SET(vTR_DBP,TR_INPUT)
-#endif
 }
 
 #pragma GCC push_options

From dcbb7df80d26b10204289a970ff89a77a53bd8df Mon Sep 17 00:00:00 2001
From: Glenn Anderson <github@mactcp.net>
Date: Sat, 4 Jun 2022 23:33:19 -0700
Subject: [PATCH 3/3] Alignment adjustment to make XCVR perform the same as
 non-XCVR

---
 src/BlueSCSI.cpp | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/src/BlueSCSI.cpp b/src/BlueSCSI.cpp
index 6d3ea4f..dfae9e6 100644
--- a/src/BlueSCSI.cpp
+++ b/src/BlueSCSI.cpp
@@ -1768,5 +1768,11 @@ BusFree:
   SCSI_TARGET_INACTIVE() // Turn off BSY, REQ, MSG, CD, IO output
 #ifdef XCVR
   TRANSCEIVER_IO_SET(vTR_TARGET,TR_INPUT);
+  // Something in code linked after this function is performing better with a +4 alignment.
+  // Adding this nop is causing the next function (_GLOBAL__sub_I_SD) to have an address with a last digit of 0x4.
+  // Last digit of 0xc also works.
+  // This affects both with and without XCVR, currently without XCVR doesn't need any padding.
+  // Until the culprit can be tracked down and fixed, it may be necessary to do manual adjustment.
+  asm("nop.w");
 #endif
 }