From 369b5cb21e63159181ef7730a4641191f36f683d Mon Sep 17 00:00:00 2001
From: tmk <hasu@tmk-kbd.com>
Date: Tue, 6 Jul 2021 22:29:26 +0900
Subject: [PATCH] ibmpc: ringbuf optimization for cpp

---
 tmk_core/protocol/ibmpc.cpp | 31 +++++++++++++-------------
 tmk_core/protocol/ibmpc.hpp | 44 ++++++++++++++++++++++++++++++++-----
 2 files changed, 55 insertions(+), 20 deletions(-)

diff --git a/tmk_core/protocol/ibmpc.cpp b/tmk_core/protocol/ibmpc.cpp
index 7d6dfcbd..34b8458d 100644
--- a/tmk_core/protocol/ibmpc.cpp
+++ b/tmk_core/protocol/ibmpc.cpp
@@ -69,7 +69,6 @@ void IBMPC::host_init(void)
     inhibit();
     int_init();
     int_off();
-    ringbuf_init(&rb, rbuf, RINGBUF_SIZE);
     host_isr_clear();
 }
 
@@ -174,14 +173,14 @@ int16_t IBMPC::host_recv(void)
     int16_t ret = -1;
 
     // Enable ISR if buffer was full
-    if (ringbuf_is_full(&rb)) {
+    if (ringbuf_is_full()) {
         host_isr_clear();
         int_on();
         idle();
     }
 
     ATOMIC_BLOCK(ATOMIC_RESTORESTATE) {
-        ret = ringbuf_get(&rb);
+        ret = ringbuf_get();
     }
     if (ret != -1) dprintf("r%02X ", ret&0xFF);
     return ret;
@@ -204,7 +203,7 @@ void IBMPC::host_isr_clear(void)
     protocol = 0;
     error = 0;
     isr_state = 0x8000;
-    ringbuf_reset(&rb);
+    ringbuf_reset();
 }
 
 inline void IBMPC::isr(void)
@@ -281,8 +280,8 @@ inline void IBMPC::isr(void)
                 uint8_t us = 100;
                 // wait for rising and falling edge of b7 of XT_IBM
                 if (!protocol) {
-                    while (!(IBMPC_CLOCK_PIN&(1<<clock_bit)) && us) { wait_us(1); us--; }
-                    while (  IBMPC_CLOCK_PIN&(1<<clock_bit)  && us) { wait_us(1); us--; }
+                    while (!(IBMPC_CLOCK_PIN & clock_mask) && us) { wait_us(1); us--; }
+                    while ( (IBMPC_CLOCK_PIN & clock_mask) && us) { wait_us(1); us--; }
                 } else if (protocol == IBMPC_PROTOCOL_XT_CLONE) {
                     us = 0;
                 }
@@ -311,8 +310,8 @@ inline void IBMPC::isr(void)
                 uint8_t us = 100;
                 // wait for rising and falling edge of AT stop bit to discriminate between XT and AT
                 if (!protocol) {
-                    while (!(IBMPC_CLOCK_PIN&(1<<clock_bit)) && us) { wait_us(1); us--; }
-                    while (  IBMPC_CLOCK_PIN&(1<<clock_bit)  && us) { wait_us(1); us--; }
+                    while (!(IBMPC_CLOCK_PIN & clock_mask) && us) { wait_us(1); us--; }
+                    while ( (IBMPC_CLOCK_PIN & clock_mask) && us) { wait_us(1); us--; }
                 } else if (protocol == IBMPC_PROTOCOL_XT_IBM) {
                     us = 0;
                 }
@@ -363,15 +362,17 @@ inline void IBMPC::isr(void)
 
 DONE:
     // store data
-    if (!ringbuf_put(&rb, isr_state & 0xFF)) {
-        // buffer overflow
-        error = IBMPC_ERR_FULL;
-
+    ringbuf_put(isr_state & 0xFF);
+    if (ringbuf_is_full()) {
         // Disable ISR if buffer is full
         int_off();
         // inhibit: clock_lo() instead of inhibit() for ISR optimization
         clock_lo();
     }
+    if (ringbuf_is_empty()) {
+        // buffer overflow
+        error = IBMPC_ERR_FULL;
+    }
 ERROR:
     // clear for next data
     isr_state = 0x8000;
@@ -388,9 +389,9 @@ void IBMPC::host_set_led(uint8_t led)
 }
 
 
-// NOTE: With this ISR data line can be read within 2us after clock falling edge.
-// To read data line early as possible:
-// write naked ISR with asembly code to read the line and call C func to do other job?
+// NOTE: With this ISR data line should be read within 5us after clock falling edge.
+// Confirmed that ATmega32u4 can read data line in 2.5us from interrupt after
+// ISR prologue pushs r18, r19, r20, r21, r24, r25 r30 and r31 with GCC 5.4.0
 ISR(IBMPC_INT_VECT)
 {
     IBMPC::interface0.isr();
diff --git a/tmk_core/protocol/ibmpc.hpp b/tmk_core/protocol/ibmpc.hpp
index 28506d52..e13f3ea0 100644
--- a/tmk_core/protocol/ibmpc.hpp
+++ b/tmk_core/protocol/ibmpc.hpp
@@ -40,7 +40,6 @@ POSSIBILITY OF SUCH DAMAGE.
 
 #include <stdbool.h>
 #include "wait.h"
-#include "ringbuf.h"
 
 /*
  * IBM PC keyboard protocol
@@ -91,8 +90,6 @@ POSSIBILITY OF SUCH DAMAGE.
 #define IBMPC_LED_NUM_LOCK    1
 #define IBMPC_LED_CAPS_LOCK   2
 
-#define RINGBUF_SIZE    16
-
 
 class IBMPC
 {
@@ -127,8 +124,14 @@ class IBMPC
     private:
     volatile uint16_t isr_state;
     uint8_t timer_start;
-    ringbuf_t rb;
-    uint8_t rbuf[RINGBUF_SIZE];
+
+    /* ring buffer */
+    // Size should be power of 2
+    #define RINGBUF_SIZE    16
+    uint8_t rb_head;
+    uint8_t rb_tail;
+    uint8_t rb_buffer[RINGBUF_SIZE];
+
     const uint8_t clock_bit, data_bit;
     const uint8_t clock_mask, data_mask;
 
@@ -237,6 +240,37 @@ class IBMPC
     {
         EIMSK &= ~clock_mask;
     }
+
+    /*
+     * ring buffer
+     */
+    inline int16_t ringbuf_get(void) __attribute__((__always_inline__))
+    {
+        if (ringbuf_is_empty()) return -1;
+        uint8_t data = rb_buffer[rb_tail];
+        rb_tail++;
+        rb_tail &= (RINGBUF_SIZE - 1);
+        return  data;
+    }
+    inline void ringbuf_put(uint8_t data) __attribute__((__always_inline__))
+    {
+        rb_buffer[rb_head] = data;
+        rb_head++;
+        rb_head &= (RINGBUF_SIZE - 1);
+    }
+    inline bool ringbuf_is_empty(void) __attribute__((__always_inline__))
+    {
+        return (rb_head == rb_tail);
+    }
+    inline bool ringbuf_is_full(void) __attribute__((__always_inline__))
+    {
+        return (((rb_head + 1) & (RINGBUF_SIZE - 1)) == rb_tail);
+    }
+    inline void ringbuf_reset(void) __attribute__((__always_inline__))
+    {
+        rb_head = 0;
+        rb_tail = 0;
+    }
 };
 
 #endif