Add TLS (thread local storage) support

libctru/Makefile

@@ -40,11 +40,11 @@ INCLUDES	:=	include
 #---------------------------------------------------------------------------------
 # options for code generation
 #---------------------------------------------------------------------------------
-ARCH	:=	-march=armv6k -mtune=mpcore -mfloat-abi=hard
+ARCH	:=	-march=armv6k -mtune=mpcore -mfloat-abi=hard -mtp=soft
 
 CFLAGS	:=	-g -Wall -Werror -O2 -mword-relocations \
 			-ffunction-sections -fno-strict-aliasing \
-			-fomit-frame-pointer -ffast-math \
+			-fomit-frame-pointer \
 			$(ARCH)
 
 CFLAGS	+=	$(INCLUDE) -DARM11 -D_3DS

libctru/source/console.c

@@ -610,7 +610,6 @@ static void newRow() {
 		}
 
 		consoleClearLine('2');
-		gfxFlushBuffers();
 	}
 }
 //---------------------------------------------------------------------------------

libctru/source/internal.h

@@ -20,6 +20,9 @@ typedef struct
 	// Pointer to this thread's newlib state
 	struct _reent* reent;
 
+	// Pointer to this thread's thread-local segment
+	void* tls_tp; // !! Keep offset in sync inside __aeabi_read_tp !!
+
 	// FS session override
 	u32    fs_magic;
 	Handle fs_session;

libctru/source/system/readtp.s

@@ -0,0 +1,9 @@
+.arm
+.section .text.__aeabi_read_tp, "ax", %progbits
+.global __aeabi_read_tp
+.type __aeabi_read_tp, %function
+.align 2
+__aeabi_read_tp:
+	mrc p15, 0, r0, c13, c0, 3
+	ldr r0, [r0, #0xC] @ Read ThreadVars.tls_tp
+	bx lr

libctru/source/system/syscalls.c

@@ -13,6 +13,10 @@
 void __ctru_exit(int rc);
 int __libctru_gtod(struct _reent *ptr, struct timeval *tp, struct timezone *tz);
 
+extern const u8 __tdata_lma[];
+extern const u8 __tdata_lma_end[];
+extern u8 __tls_start[];
+
 static struct _reent* __ctru_get_reent()
 {
 	ThreadVars* tv = getThreadVars();
@@ -46,4 +50,9 @@ void __system_initSyscalls(void)
 	tv->magic = THREADVARS_MAGIC;
 	tv->reent = _impure_ptr;
 	tv->thread_ptr = NULL;
+	tv->tls_tp = __tls_start-8; // ARM ELF TLS ABI mandates an 8-byte header
+
+	u32 tls_size = __tdata_lma_end - __tdata_lma;
+	if (tls_size)
+		memcpy(__tls_start, __tdata_lma, tls_size);
 }

libctru/source/thread.c

@@ -2,6 +2,11 @@
 #include <stdlib.h>
 #include <string.h>
 
+extern const u8 __tdata_lma[];
+extern const u8 __tdata_lma_end[];
+extern u8 __tls_start[];
+extern u8 __tls_end[];
+
 struct Thread_tag
 {
 	Handle handle;
@@ -26,6 +31,7 @@ static void _thread_begin(void* arg)
 	tv->magic = THREADVARS_MAGIC;
 	tv->reent = &t->reent;
 	tv->thread_ptr = t;
+	tv->tls_tp = (u8*)t->stacktop-8; // ARM ELF TLS ABI mandates an 8-byte header
 	t->ep(t->arg);
 	threadExit(0);
 }
@@ -34,9 +40,16 @@ Thread threadCreate(ThreadFunc entrypoint, void* arg, size_t stack_size, int pri
 {
 	size_t stackoffset = (sizeof(struct Thread_tag)+7)&~7;
 	size_t allocsize   = stackoffset + ((stack_size+7)&~7);
-	if (allocsize < stackoffset) return NULL; // guard against overflow
+	size_t tlssize = __tls_end-__tls_start;
-	if ((allocsize-stackoffset) < stack_size) return NULL; // guard against overflow
+	size_t tlsloadsize = __tdata_lma_end-__tdata_lma;
-	Thread t = (Thread)malloc(allocsize);
+	size_t tbsssize = tlssize-tlsloadsize;
+
+	// Guard against overflow
+	if (allocsize < stackoffset) return NULL;
+	if ((allocsize-stackoffset) < stack_size) return NULL;
+	if ((allocsize+tlssize) < allocsize) return NULL;
+
+	Thread t = (Thread)malloc(allocsize+tlssize);
 	if (!t) return NULL;
 
 	t->ep       = entrypoint;
@@ -45,6 +58,11 @@ Thread threadCreate(ThreadFunc entrypoint, void* arg, size_t stack_size, int pri
 	t->finished = false;
 	t->stacktop = (u8*)t + allocsize;
 
+	if (tlsloadsize)
+		memcpy(t->stacktop, __tdata_lma, tlsloadsize);
+	if (tbsssize)
+		memset((u8*)t->stacktop+tlsloadsize, 0, tbsssize);
+
 	// Set up child thread's reent struct, inheriting standard file handles
 	_REENT_INIT_PTR(&t->reent);
 	struct _reent* cur = getThreadVars()->reent;