Fix native Linux/ppc with recent enough glibc that supports TLS; r2 is used

in that case. Tell me if I broke other arches, e.g. r13 is no longer saved
in Video and Ethernet stubs, though it seems to be OK.

Colateral feature: SheepShaver should now run on Linux/ppc64 with relevant
32-bit runtime. Native Linux/ppc64 support is harder as low mem globals are
32-bit in mind and e.g. the TLS register there is %r13, %r2 is the TOC
(PowerOpen/AIX ABI)

SheepShaver/src/BeOS/sysdeps.h

@@ -36,6 +36,7 @@
 #ifdef __POWERPC__
 #define EMULATED_PPC 0
 #define WORDS_BIGENDIAN 1
+#define SYSTEM_CLOBBERS_R2 1
 #else
 #define EMULATED_PPC 1
 #undef  WORDS_BIGENDIAN

SheepShaver/src/Unix/Linux/paranoia.cpp

@@ -42,7 +42,9 @@ const uint32 SIG_STACK_SIZE = 0x10000;		// Size of signal stack
 
 // Prototypes
 extern "C" void *get_sp(void);
-extern "C" void set_r2(uint32 val);
+extern "C" void *get_r2(void);
+extern "C" void set_r2(void *);
+extern "C" void *get_r13(void);
 extern void paranoia_check(void);
 static void sigusr2_handler(int sig, sigcontext_struct *sc);
 
@@ -57,6 +59,12 @@ static void *sig_sc_regs = NULL;
 static uint32 sig_r2 = 0;
 
 
+int raise(int sig)
+{
+	// Reimplement to get rid of access to r2 (TLS pointer)
+	return kill(getpid(), sig);
+}
+
 void paranoia_check(void)
 {
 	char str[256];
@@ -95,8 +103,14 @@ void paranoia_check(void)
 	}
 
 	// Raise SIGUSR2
-	set_r2(0xaffebad5);
+	TOC = get_r2();
+	R13 = get_r13();
+	set_r2((void *)0xaffebad5);
 	raise(SIGUSR2);
+	if (TOC != get_r2())
+		err = 6;
+	if (R13 != get_r13())
+		err = 7;
 
 	// Check error code
 	switch (err) {
@@ -115,6 +129,11 @@ void paranoia_check(void)
 		case 5:
 			printf("FATAL: sc->regs->gpr[2] in signal handler (%08lx) doesn't have expected value (%08x)\n", (uint32)sig_r2, 0xaffebad5);
 			break;
+		case 6:
+			printf("FATAL: signal handler failed to restore initial r2 value (%08x, was %08x)\n", (uint32)get_r2(), (uint32)TOC);
+			break;
+		case 7:
+			printf("FATAL: signal handler failed to restore initial r13 value (%08x, was %08x)\n", get_r13(), (uint32)R13);
 	}
 	if (err) {
 		printf("Maybe you need a different kernel?\n");
@@ -135,34 +154,40 @@ static void sigusr2_handler(int sig, sigcontext_struct *sc)
 	sig_sp = get_sp();
 	if (sig_sp < sig_stack || sig_sp >= ((uint8 *)sig_stack + SIG_STACK_SIZE)) {
 		err = 1;
-		return;
+		goto ret;
 	}
 
 	// Check whether r4 points to info on the stack
 	sig_r4 = sc;
 	if (sig_r4 < sig_stack || sig_r4 >= ((uint8 *)sig_stack + SIG_STACK_SIZE)) {
 		err = 2;
-		return;
+		goto ret;
 	}
 
 	// Check whether r4 looks like a sigcontext
 	sig_sc_signal = sc->signal;
 	if (sig_sc_signal != SIGUSR2) {
 		err = 3;
-		return;
+		goto ret;
 	}
 
 	// Check whether sc->regs points to info on the stack
 	sig_sc_regs = sc->regs;
 	if (sig_sc_regs < sig_stack || sig_sc_regs >= ((uint8 *)sig_stack + SIG_STACK_SIZE)) {
 		err = 4;
-		return;
+		goto ret;
 	}
 
 	// Check whether r2 still holds the value we set it to
 	sig_r2 = sc->regs->gpr[2];
 	if (sig_r2 != 0xaffebad5) {
 		err = 5;
-		return;
+		goto ret;
 	}
+
+	// Restore pointer to Thread Local Storage
+  ret:
+#ifdef SYSTEM_CLOBBERS_R2
+	sc->regs->gpr[2] = (unsigned long)TOC;
+#endif
 }

SheepShaver/src/Unix/main_unix.cpp

@@ -27,12 +27,14 @@
  *  is slightly different from the SysV ABI used by Linux:
  *    - Stack frames are different (e.g. LR is stored in 8(r1) under
  *      MacOS, but in 4(r1) under Linux)
- *    - There is no TOC under Linux; r2 is free for the user
+ *    - There is a pointer to Thread Local Storage (TLS) under Linux with
+ *      recent enough glibc. This is r2 in 32-bit mode and r13 in
+ *      64-bit mode (PowerOpen/AIX ABI)
  *    - r13 is used as a small data pointer under Linux (but appearently
  *      it is not used this way? To be sure, we specify -msdata=none
  *      in the Makefile)
- *    - As there is no TOC, there are also no TVECTs under Linux;
+ *    - There are no TVECTs under Linux; function pointers point
- *      function pointers point directly to the function code
+ *      directly to the function code
  *  The Execute*() functions have to account for this. Additionally, we
  *  cannot simply call MacOS functions by getting their TVECT and jumping
  *  to it. Such calls are done via the call_macos*() functions in
@@ -284,7 +286,8 @@ static inline void sig_stack_release(void)
 
 // Global variables (exported)
 #if !EMULATED_PPC
-void *TOC;				// Small data pointer (r13)
+void *TOC;				// Pointer to Thread Local Storage (r2)
+void *R13;				// Pointer to .sdata section (r13 under Linux)
 #endif
 uint32 RAMBase;			// Base address of Mac RAM
 uint32 RAMSize;			// Size of Mac RAM
@@ -370,8 +373,11 @@ static void sigill_handler(int sig, siginfo_t *sip, void *scp);
 
 // From asm_linux.S
 #if !EMULATED_PPC
-extern "C" void *get_toc(void);
 extern "C" void *get_sp(void);
+extern "C" void *get_r2(void);
+extern "C" void set_r2(void *);
+extern "C" void *get_r13(void);
+extern "C" void set_r13(void *);
 extern "C" void flush_icache_range(uint32 start, uint32 end);
 extern "C" void jump_to_rom(uint32 entry, uint32 context);
 extern "C" void quit_emulator(void);
@@ -482,8 +488,14 @@ int main(int argc, char **argv)
 	printf(" %s\n", GetString(STR_ABOUT_TEXT2));
 
 #if !EMULATED_PPC
+#ifdef SYSTEM_CLOBBERS_R2
 	// Get TOC pointer
-	TOC = get_toc();
+	TOC = get_r2();
+#endif
+#ifdef SYSTEM_CLOBBERS_R13
+	// Get r13 register
+	R13 = get_r13();
+#endif
 #endif
 
 #ifdef ENABLE_GTK
@@ -1607,6 +1619,15 @@ static void sigusr2_handler(int sig, siginfo_t *sip, void *scp)
 	if (*(int32 *)XLM_IRQ_NEST > 0)
 		return;
 
+#ifdef SYSTEM_CLOBBERS_R2
+	// Restore pointer to Thread Local Storage
+	set_r2(TOC);
+#endif
+#ifdef SYSTEM_CLOBBERS_R13
+	// Restore pointer to .sdata section
+	set_r13(R13);
+#endif
+
 	// Disable MacOS stack sniffer
 	WriteMacInt32(0x110, 0);
 
@@ -1697,6 +1718,15 @@ static void sigsegv_handler(int sig, siginfo_t *sip, void *scp)
 	// Get effective address
 	uint32 addr = r->dar();
 	
+#ifdef SYSTEM_CLOBBERS_R2
+	// Restore pointer to Thread Local Storage
+	set_r2(TOC);
+#endif
+#ifdef SYSTEM_CLOBBERS_R13
+	// Restore pointer to .sdata section
+	set_r13(R13);
+#endif
+
 #if ENABLE_VOSF
 	// Handle screen fault.
 	extern bool Screen_fault_handler(sigsegv_address_t fault_address, sigsegv_address_t fault_instruction);
@@ -1946,6 +1976,15 @@ static void sigill_handler(int sig, siginfo_t *sip, void *scp)
 	machine_regs *r = MACHINE_REGISTERS(scp);
 	char str[256];
 
+#ifdef SYSTEM_CLOBBERS_R2
+	// Restore pointer to Thread Local Storage
+	set_r2(TOC);
+#endif
+#ifdef SYSTEM_CLOBBERS_R13
+	// Restore pointer to .sdata section
+	set_r13(R13);
+#endif
+
 	// Fault in Mac ROM or RAM?
 	bool mac_fault = (r->pc() >= ROM_BASE) && (r->pc() < (ROM_BASE + ROM_AREA_SIZE)) || (r->pc() >= RAMBase) && (r->pc() < (RAMBase + RAMSize));
 	if (mac_fault) {

SheepShaver/src/Unix/ppc_asm.S

@@ -24,16 +24,6 @@
 #define SAVE_FP_EXEC_68K 1
 
 
-/*
- *  void *get_toc(void) - Get TOC pointer (small data pointer r13 under Linux)
- */
-
-	ASM_GLOBAL_DIRECTIVE C_SYMBOL_NAME(get_toc)
-C_SYMBOL_NAME(get_toc):
-	mr	r3,r13
-	blr
-
-
 /*
  *  void *get_sp(void) - Get stack pointer
  */
@@ -45,7 +35,17 @@ C_SYMBOL_NAME(get_sp):
 
 
 /*
- *  void set_r2(uint32 val {r3}) - Set r2
+ *  void *get_r2(void) - Get r2
+ */
+
+	ASM_GLOBAL_DIRECTIVE C_SYMBOL_NAME(get_r2)
+C_SYMBOL_NAME(get_r2):
+	mr	r3,r2
+	blr
+
+
+/*
+ *  void set_r2(void *val {r3}) - Set r2
  */
 
 	ASM_GLOBAL_DIRECTIVE C_SYMBOL_NAME(set_r2)
@@ -54,6 +54,25 @@ C_SYMBOL_NAME(set_r2):
 	blr
 
 
+/*
+ *  void *get_r13(void) - Get r13 (small data pointer under Linux)
+ */
+
+	ASM_GLOBAL_DIRECTIVE C_SYMBOL_NAME(get_r13)
+C_SYMBOL_NAME(get_r13):
+	mr	r3,r13
+	blr
+
+/*
+ *  void set_r13(void *val {r3}) - Set r13 (small data pointer under Linux)
+ */
+
+	ASM_GLOBAL_DIRECTIVE C_SYMBOL_NAME(set_r13)
+C_SYMBOL_NAME(set_r13):
+	mr	r13,r3
+	blr
+
+
 /*
  *  void flush_icache_range(void *start {r3}, void *end {r3}) - Flush D and I cache
  */
@@ -225,6 +244,7 @@ C_SYMBOL_NAME(jump_to_rom):
 
 	// Restore PowerPC registers
 	lwz	r1,XLM_EMUL_RETURN_STACK(0)
+	RESTORE_SYSTEM_R2
 	lmw	r13,20(r1)
 	lfd	f14,20+19*4+0*8(r1)
 	lfd	f15,20+19*4+1*8(r1)
@@ -361,6 +381,7 @@ C_SYMBOL_NAME(jump_to_rom):
 	stw	r0,16(r1)
 	stw	r2,12(r1)
 	stwu	r1,-(24+16*4+15*8)(r1)
+	RESTORE_SYSTEM_R2
 
 	// Save 68k registers (M68kRegisters)
 	stw	r8,24+0*4(r1)	// d[0]..d[7]
@@ -397,7 +418,8 @@ C_SYMBOL_NAME(jump_to_rom):
 	stfd	f0,24+16*4+14*8(r1)
 
 	// Execute native routine
-	lwz	r13,XLM_TOC(0)
+	RESTORE_SYSTEM_R2
+	RESTORE_SYSTEM_R13
 	addi	r3,r1,24
 	mr	r4,r24
 	bl	C_SYMBOL_NAME(EmulOp)
@@ -573,7 +595,8 @@ ASM_MACRO_START	prolog
 ASM_MACRO_END
 
 ASM_MACRO_START	epilog
-	lwz	r13,XLM_TOC(0)
+	RESTORE_SYSTEM_R2
+	RESTORE_SYSTEM_R13
 	lwz	r0,64+4(r1)
 	mtlr	r0
 	addi	r1,r1,64
@@ -701,6 +724,7 @@ C_SYMBOL_NAME(get_resource):
 	lwz	r2,XLM_RES_LIB_TOC(0)
 	mtctr	r0
 	bctrl
+	RESTORE_SYSTEM_R2
 	stw	r3,56+8(r1)		// Save handle
 
 	// Call CheckLoad
@@ -732,6 +756,7 @@ C_SYMBOL_NAME(get_1_resource):
 	lwz	r2,XLM_RES_LIB_TOC(0)
 	mtctr	r0
 	bctrl
+	RESTORE_SYSTEM_R2
 	stw	r3,56+8(r1)		// Save handle
 
 	// Call CheckLoad
@@ -763,6 +788,7 @@ C_SYMBOL_NAME(get_ind_resource):
 	lwz	r2,XLM_RES_LIB_TOC(0)
 	mtctr	r0
 	bctrl
+	RESTORE_SYSTEM_R2
 	stw	r3,56+8(r1)		// Save handle
 
 	// Call CheckLoad
@@ -794,6 +820,7 @@ C_SYMBOL_NAME(get_1_ind_resource):
 	lwz	r2,XLM_RES_LIB_TOC(0)
 	mtctr	r0
 	bctrl
+	RESTORE_SYSTEM_R2
 	stw	r3,56+8(r1)		// Save handle
 
 	// Call CheckLoad
@@ -825,6 +852,7 @@ C_SYMBOL_NAME(r_get_resource):
 	lwz	r2,XLM_RES_LIB_TOC(0)
 	mtctr	r0
 	bctrl
+	RESTORE_SYSTEM_R2
 	stw	r3,56+8(r1)		// Save handle
 
 	// Call CheckLoad
@@ -888,3 +916,57 @@ C_SYMBOL_NAME(ppc_interrupt):
 	// Enter nanokernel
 	mtlr	r3
 	blr
+
+
+/*
+ *  Glue for glibc with TLS support which clobbers r2
+ *
+ *  The following is derived from dietlibc and only provides
+ *  pass-through to kernel syscalls. You also lose asynchronous
+ *  cancellation support.
+ */
+
+#if defined(__linux__)
+#define __NR_getpid		20
+#define __NR__newselect		142
+
+#define syscall_weak(name,wsym,sym) \
+.text; \
+.type wsym,@function; \
+.weak wsym; \
+wsym: ; \
+.type sym,@function; \
+.global sym; \
+sym: \
+	li	0,__NR_##name; \
+	b	__unified_syscall
+
+#define syscall(name,sym) \
+.text; \
+.type sym,@function; \
+.global sym; \
+sym: \
+	li	0,__NR_##name; \
+	b	__unified_syscall
+
+
+	.type __unified_syscall,@function
+__unified_syscall:
+	sc
+	bnslr+
+	stwu	1,-16(1)
+	mflr	0
+	stw	0,20(1)
+	stw	3,12(1)
+	bl	__errno_location
+	lwz	0,12(1)
+	stw	0,0(3)
+	lwz	0,20(1)
+	mtlr	0
+	addi	1,1,16
+	blr
+
+/*		name		weak symbol	symbol		*/
+syscall(	getpid,		getpid)
+syscall_weak(	_newselect,	select,		__select)
+#endif

SheepShaver/src/Unix/ppc_asm.tmpl

@@ -1,3 +1,27 @@
+/* Define usage of "reserved" registers */
+#if defined(__linux__)
+#define SYSTEM_CLOBBERS_R2	1	/* Pointer to Thread Local Storage */
+#define SYSTEM_CLOBBERS_R13	1	/* Pointer to .sdata section */
+#endif
+
+#ifdef __ASSEMBLY__
+/* Helper macros */
+#ifdef SYSTEM_CLOBBERS_R2
+#define RESTORE_SYSTEM_R2	lwz r2,XLM_TOC(0)
+#define SAVE_SYSTEM_R2		stw r2,XLM_TOC(0)
+#else
+#define RESTORE_SYSTEM_R2
+#define SAVE_SYSTEM_R2
+#endif
+#ifdef SYSTEM_CLOBBERS_R13
+#define RESTORE_SYSTEM_R13	lwz r13,XLM_R13(0)
+#define SAVE_SYSTEM_R13		stw r13,XLM_R13(0)
+#else
+#define RESTORE_SYSTEM_R13
+#define SAVE_SYSTEM_R13
+#endif
+
+/* Apple assembler perticularities */
 #if (defined(__APPLE__) && defined(__MACH__))
 #define C_SYMBOL_NAME(NAME)	_ ## NAME
 #define ASM_MACRO_END		.endmacro
@@ -87,3 +111,4 @@
 #define	f30	30
 #define	f31	31
 #endif
+#endif

SheepShaver/src/Unix/sysdeps.h

@@ -82,6 +82,7 @@
 // Otherwise, use Direct Addressing mode if NATMEM_OFFSET is set
 #if !defined(EMULATED_PPC)
 #define REAL_ADDRESSING 1
+#include "ppc_asm.tmpl"
 #elif defined(__CYGWIN__)
 #define DIRECT_ADDRESSING 1
 #define DIRECT_ADDRESSING_HACK 1

SheepShaver/src/include/main.h

@@ -23,6 +23,7 @@
 
 // Global variables
 extern void *TOC;				// TOC pointer
+extern void *R13;				// r13 register
 extern uint32 KernelDataAddr;	// Address of Kernel Data
 extern uint32 BootGlobsAddr;	// Address of BootGlobs structure at top of Mac RAM
 extern uint32 PVR;				// Theoretical PVR

SheepShaver/src/include/xlowmem.h

@@ -47,6 +47,7 @@
 #define XLM_R_GET_RESOURCE 0x2848		// Pointer to native RGetResource() routine
 #define XLM_EXEC_RETURN_OPCODE 0x284c	// EXEC_RETURN opcode for Execute68k()
 #define XLM_ZERO_PAGE 0x2850			// Pointer to read-only page with all bits set to 0
+#define XLM_R13 0x2854					// Pointer to .sdata section (Linux)
 
 #define XLM_ETHER_INIT 0x28c0			// Pointer to ethernet InitStreamModule() function
 #define XLM_ETHER_TERM 0x28c4			// Pointer to ethernet TerminateStreamModule() function

SheepShaver/src/main.cpp

@@ -232,7 +232,12 @@ bool InitAll(void)
 	WriteMacInt16(XLM_EXEC_RETURN_OPCODE, M68K_EXEC_RETURN);		// For Execute68k() (RTS from the executed 68k code will jump here and end 68k mode)
 	WriteMacInt32(XLM_ZERO_PAGE, SheepMem::ZeroPage());				// Pointer to read-only page with all bits set to 0
 #if !EMULATED_PPC
-	WriteMacInt32(XLM_TOC, (uint32)TOC);								// TOC pointer of emulator
+#ifdef SYSTEM_CLOBBERS_R2
+	WriteMacInt32(XLM_TOC, (uint32)TOC);							// TOC pointer of emulator
+#endif
+#ifdef SYSTEM_CLOBBERS_R13
+	WriteMacInt32(XLM_R13, (uint32)R13);							// TLS register
+#endif
 #endif
 	WriteMacInt32(XLM_ETHER_INIT, NativeFunction(NATIVE_ETHER_INIT));	// DLPI ethernet driver functions
 	WriteMacInt32(XLM_ETHER_TERM, NativeFunction(NATIVE_ETHER_TERM));