From 840121e36926c797ca948d003aae8c4dc2e04e42 Mon Sep 17 00:00:00 2001
From: Volker Eckert <volker@maxfinite.com>
Date: Thu, 3 Oct 2024 19:22:25 +0100
Subject: [PATCH] cope with I$ miss on teensy

if the instruction cache is completely flushed *even before* the counter
is read, under specific situations (presumably when relevant parts of
show() are on different cache lines), the first bit would not be sent
(but discarded), often leading to a change in colour.

prevent compiler optimisation across counter reads,
order the delays so that they are easier to reason about.
---
 Adafruit_NeoPixel.cpp | 32 +++++++++++++++++++++++++++-----
 1 file changed, 27 insertions(+), 5 deletions(-)

diff --git a/Adafruit_NeoPixel.cpp b/Adafruit_NeoPixel.cpp
index 39f0b11..848edd8 100644
--- a/Adafruit_NeoPixel.cpp
+++ b/Adafruit_NeoPixel.cpp
@@ -45,6 +45,11 @@
 
 #include "Adafruit_NeoPixel.h"
 
+#undef REPRODUCE_TEENSY_TIMING
+#ifdef REPRODUCE_TEENSY_TIMING
+#include "imxrt.h"
+#endif
+
 #if defined(TARGET_LPC1768)
 #include <time.h>
 #endif
@@ -1912,16 +1917,32 @@ void Adafruit_NeoPixel::show(void) {
   ARM_DEMCR |= ARM_DEMCR_TRCENA;
   ARM_DWT_CTRL |= ARM_DWT_CTRL_CYCCNTENA;
 
+#ifdef REPRODUCE_TEENSY_TIMING
+	// reproducer for the timing issue
+	{
+		static uint8_t ct;
+
+		//if (!(ct & 1)) {
+		if (true) {
+			asm("dsb");
+			asm("isb");
+			SCB_CACHE_ICIALLU = 0UL;
+			asm("dsb");
+			asm("isb");
+		}
+		ct++;
+	}
+#endif
+
 #if defined(NEO_KHZ400) // 800 KHz check needed only if 400 KHz support enabled
   if (is800KHz) {
 #endif
-    cyc = ARM_DWT_CYCCNT + CYCLES_800;
+    cyc = ARM_DWT_CYCCNT;  // make sure it's initialised..
     while (p < end) {
       pix = *p++;
       for (mask = 0x80; mask; mask >>= 1) {
-        while (ARM_DWT_CYCCNT - cyc < CYCLES_800)
-          ;
         cyc = ARM_DWT_CYCCNT;
+        barrier();
         *set = msk;
         if (pix & mask) {
           while (ARM_DWT_CYCCNT - cyc < CYCLES_800_T1H)
@@ -1931,10 +1952,11 @@ void Adafruit_NeoPixel::show(void) {
             ;
         }
         *clr = msk;
+        barrier();
+        while (ARM_DWT_CYCCNT - cyc < CYCLES_800)
+            ;
       }
     }
-    while (ARM_DWT_CYCCNT - cyc < CYCLES_800)
-      ;
 #if defined(NEO_KHZ400)
   } else { // 400 kHz bitstream
     cyc = ARM_DWT_CYCCNT + CYCLES_400;