Force non inline of 16 bit memory poking functions

On the Raspberry Pi 4, the compiler was inlining mempoke16 and doing
some other optimizations on it, which I presume break the memory
alignment requirements on aarch64. By adding __attribute__ ((noinline))
to the uxn functions that deal with 16 bit operations, we can turn back
-O2 compiler optimizations and the code works properly again on
hardware.

4 files changed, 13 insertions, 13 deletions

diff --git a/Makefile b/Makefile
index 1fced4f..348d974 100644
--- a/Makefile
+++ b/Makefile
@@ -23,7 +23,7 @@ OBJ_START = $(BUILD_DIR)/start.o
 OBJ_MAIN = $(BUILD_DIR)/main.o
 SRC_LINK = $(SRC_DIR)/linker.ld
 
-CFLAGS := -Wall -ffreestanding -O1 -nostdlib -lgcc -mgeneral-regs-only -fstack-protector
+CFLAGS := -Wall -ffreestanding -O2 -nostdlib -lgcc -mgeneral-regs-only -fstack-protector
 CFLAGS += -DRPI_VERSION=$(RPI_VERSION)
 AFLAGS :=
 LDFLAGS :=
diff --git a/src/common.h b/src/common.h
index 94e499e..4c2c6b2 100644
--- a/src/common.h
+++ b/src/common.h
@@ -134,22 +134,22 @@ uart_init() {
 
 static inline void
 uart_putc(char c) {
-    while(!(AUX->mu_lsr & (1 << 5)));
+    while (!(AUX->mu_lsr & (1 << 5)));
     AUX->mu_io = c;
     return;
 }
 
 static inline char
 uart_getc() {
-    while(!(AUX->mu_lsr & 1));
+    while (!(AUX->mu_lsr & 1));
     u8 c = AUX->mu_io & 0xFF;
     return c == '\r' ? '\n' : c;
 }
 
 static inline void
 uart_puts(char *s) {
-    while(*s) {
-        if(*s == '\n') {
+    while (*s) {
+        if (*s == '\n') {
             uart_putc('\r');
         }
         uart_putc(*s++);
@@ -160,7 +160,7 @@ static inline void
 uart_hex(unsigned int d) {
     unsigned int n;
     uart_puts("0x");
-    for(int c = 28; c >= 0; c -= 4) {
+    for (int c = 28; c >= 0; c -= 4) {
         n = (d>>c) & 0xF;
         n += n> 9 ? 0x37 : 0x30;
         uart_putc(n);
@@ -238,14 +238,14 @@ typedef enum MboxChannels {
 
 static inline void
 mb_write(u8 channel, void *data) {
-    while(MBOX->status & MBOX_FULL);
+    while (MBOX->status & MBOX_FULL);
     MBOX->write = ((uintptr_t)data & ~0xF) | (channel & 0xF);
 }
 
 static inline u32
 mb_read(u8 channel) {
-    while(true) {
-        while(MBOX->status & MBOX_EMPTY);
+    while (true) {
+        while (MBOX->status & MBOX_EMPTY);
         u32 data = MBOX->read;
         if ((u8)(data & 0xF) == channel) {
             return data & 0xFFFFFFF0;
diff --git a/src/main.c b/src/main.c
index 7ac7037..c379ace 100644
--- a/src/main.c
+++ b/src/main.c
@@ -172,7 +172,7 @@ void main(void) {
 
     uxn_eval(&u, 0x0100);
     u64 current_ticks = timer_get_ticks();
-    while(1) {
+    while (true) {
         // Echo input to standard output.
         uxn_eval(&u, mempeek16(devscreen->dat, 0));
 
diff --git a/src/uxn-fast.c b/src/uxn-fast.c
index 5066376..7ee1d04 100644
--- a/src/uxn-fast.c
+++ b/src/uxn-fast.c
@@ -30,9 +30,9 @@ static void   mempoke8(Uint8 *m, Uint16 a, Uint8 b) { m[a] = b; }
 static Uint8  mempeek8(Uint8 *m, Uint16 a) { return m[a]; }
 static void   devpoke8(Device *d, Uint8 a, Uint8 b) { d->dat[a & 0xf] = b; d->talk(d, a & 0x0f, 1); }
 static Uint8  devpeek8(Device *d, Uint8 a) { d->talk(d, a & 0x0f, 0); return d->dat[a & 0xf];  }
-void   mempoke16(Uint8 *m, Uint16 a, Uint16 b) { mempoke8(m, a, b >> 8); mempoke8(m, a + 1, b); }
-Uint16 mempeek16(Uint8 *m, Uint16 a) { return (mempeek8(m, a) << 8) + mempeek8(m, a + 1); }
-static void   devpoke16(Device *d, Uint8 a, Uint16 b) { devpoke8(d, a, b >> 8); devpoke8(d, a + 1, b); }
+void   __attribute__ ((noinline)) mempoke16(Uint8 *m, Uint16 a, Uint16 b) { mempoke8(m, a, b >> 8); mempoke8(m, a + 1, b); }
+Uint16 __attribute__ ((noinline)) mempeek16(Uint8 *m, Uint16 a) { return (mempeek8(m, a) << 8) + mempeek8(m, a + 1); }
+static void  __attribute__ ((noinline)) devpoke16(Device *d, Uint8 a, Uint16 b) { devpoke8(d, a, b >> 8); devpoke8(d, a + 1, b); }
 
 /* clang-format on */