update 386 to new runtime (CL 30381)

R=r
DELTA=298  (119 added, 81 deleted, 98 changed)
OCL=30427
CL=30443

src/cmd/8l/obj.c

@@ -1038,7 +1038,7 @@ loop:
 		s = p->from.sym;
 		if(s != S && s->dupok) {
 			if(debug['v'])
-				Bprint(&bso, "skipping %s in %s: dupok", s->name, pn);
+				Bprint(&bso, "skipping %s in %s: dupok\n", s->name, pn);
 			goto loop;
 		}
 		if(s != S) {

src/cmd/8l/pass.c

@@ -30,7 +30,7 @@
 
 #include	"l.h"
 
-// see ../../runtime/proc.c:/StackGuard
+// see ../../pkg/runtime/proc.c:/StackGuard
 enum
 {
 	StackSmall = 128,
@@ -619,29 +619,17 @@ dostkoff(void)
 					q = p;
 				}
 
-				p = appendp(p);	// load m into DX
+				p = appendp(p);	// save frame size in DX
 				p->as = AMOVL;
-				p->from.type = D_INDIR+D_FS;
-				p->from.offset = 4;
 				p->to.type = D_DX;
-				if(q1) {
-					q1->pcond = p;
-					q1 = P;
-				}
-
-				p = appendp(p);	// save autoffset in 4(DX)
-				p->as = AMOVL;
-				p->to.type = D_INDIR+D_DX;
-				p->to.offset = 4;
 				/* 160 comes from 3 calls (3*8) 4 safes (4*8) and 104 guard */
 				p->from.type = D_CONST;
 				if(autoffset+160 > 4096)
 					p->from.offset = (autoffset+160) & ~7LL;
 
-				p = appendp(p);	// save textarg in 8(DX)
+				p = appendp(p);	// save arg size in AX
 				p->as = AMOVL;
-				p->to.type = D_INDIR+D_DX;
-				p->to.offset = 8;
+				p->to.type = D_AX;
 				p->from.type = D_CONST;
 				p->from.offset = curtext->to.offset2;
 

src/pkg/runtime/386/asm.s

@@ -2,6 +2,8 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
+#include "386/asm.h"
+
 TEXT _rt0_386(SB),7,$0
 	// copy arguments forward on an even stack
 	MOVL	0(SP), AX		// argc
@@ -11,17 +13,6 @@ TEXT _rt0_386(SB),7,$0
 	MOVL	AX, 120(SP)		// save argc, argv away
 	MOVL	BX, 124(SP)
 
-/*
-	// write "go386\n"
-	PUSHL	$6
-	PUSHL	$hello(SB)
-	PUSHL	$1
-	CALL	sys·write(SB)
-	POPL	AX
-	POPL	AX
-	POPL	AX
-*/
-
 	CALL	ldt0setup(SB)
 
 	// set up %fs to refer to that ldt entry
@@ -37,19 +28,18 @@ TEXT _rt0_386(SB),7,$0
 ok:
 
 	// set up m and g "registers"
-	// g is 0(FS), m is 4(FS)
 	LEAL	g0(SB), CX
-	MOVL	CX, 0(FS)
+	MOVL	CX, g
 	LEAL	m0(SB), AX
-	MOVL	AX, 4(FS)
+	MOVL	AX, m
 
 	// save m->g0 = g0
-	MOVL	CX, 0(AX)
+	MOVL	CX, m_g0(AX)
 
 	// create istack out of the OS stack
 	LEAL	(-8192+104)(SP), AX	// TODO: 104?
-	MOVL	AX, 0(CX)	// 8(g) is stack limit (w 104b guard)
-	MOVL	SP, 4(CX)	// 12(g) is base
+	MOVL	AX, g_stackguard(CX)
+	MOVL	SP, g_stackbase(CX)
 	CALL	emptyfunc(SB)	// fault if stack check is wrong
 
 	// convention is D is always cleared
@@ -68,7 +58,7 @@ ok:
 
 	// create a new goroutine to start program
 	PUSHL	$mainstart(SB)	// entry
-	PUSHL	$8	// arg size
+	PUSHL	$0	// arg size
 	CALL	sys·newproc(SB)
 	POPL	AX
 	POPL	AX
@@ -93,51 +83,105 @@ TEXT	breakpoint(SB),7,$0
 	BYTE $0xcc
 	RET
 
-// go-routine
-TEXT	gogo(SB), 7, $0
-	MOVL	4(SP), AX	// gobuf
-	MOVL	0(AX), SP	// restore SP
-	MOVL	4(AX), AX
-	MOVL	AX, 0(SP)	// put PC on the stack
-	MOVL	$1, AX
-	RET
+/*
+ *  go-routine
+ */
 
+// uintptr gosave(Gobuf*)
+// save state in Gobuf; setjmp
 TEXT gosave(SB), 7, $0
-	MOVL	4(SP), AX	// gobuf
-	MOVL	SP, 0(AX)	// save SP
-	MOVL	0(SP), BX
-	MOVL	BX, 4(AX)	// save PC
-	MOVL	$0, AX	// return 0
+	MOVL	4(SP), AX		// gobuf
+	LEAL	4(SP), BX		// caller's SP
+	MOVL	BX, gobuf_sp(AX)
+	MOVL	0(SP), BX		// caller's PC
+	MOVL	BX, gobuf_pc(AX)
+	MOVL	g, BX
+	MOVL	BX, gobuf_g(AX)
+	MOVL	$0, AX			// return 0
 	RET
 
-// support for morestack
-
-// return point when leaving new stack.
-// save AX, jmp to lesstack to switch back
-TEXT	retfromnewstack(SB),7,$0
-	MOVL	4(FS), BX	// m
-	MOVL	AX, 12(BX)	// save AX in m->cret
-	JMP	lessstack(SB)
-
-// gogo, returning 2nd arg instead of 1
-TEXT gogoret(SB), 7, $0
-	MOVL	8(SP), AX	// return 2nd arg
-	MOVL	4(SP), BX	// gobuf
-	MOVL	0(BX), SP	// restore SP
-	MOVL	4(BX), BX
-	MOVL	BX, 0(SP)	// put PC on the stack
-	RET
-
-TEXT setspgoto(SB), 7, $0
-	MOVL	4(SP), AX	// SP
-	MOVL	8(SP), BX	// fn to call
-	MOVL	12(SP), CX	// fn to return
-	MOVL	AX, SP
-	PUSHL	CX
+// void gogo(Gobuf*, uintptr)
+// restore state from Gobuf; longjmp
+TEXT gogo(SB), 7, $0
+	MOVL	8(SP), AX		// return 2nd arg
+	MOVL	4(SP), BX		// gobuf
+	MOVL	gobuf_g(BX), DX
+	MOVL	0(DX), CX		// make sure g != nil
+	MOVL	DX, g
+	MOVL	gobuf_sp(BX), SP	// restore SP
+	MOVL	gobuf_pc(BX), BX
 	JMP	BX
-	POPL	AX	// not reached
+
+// void gogocall(Gobuf*, void (*fn)(void))
+// restore state from Gobuf but then call fn.
+// (call fn, returning to state in Gobuf)
+TEXT gogocall(SB), 7, $0
+	MOVL	8(SP), AX		// fn
+	MOVL	4(SP), BX		// gobuf
+	MOVL	gobuf_g(BX), DX
+	MOVL	DX, g
+	MOVL	0(DX), CX		// make sure g != nil
+	MOVL	gobuf_sp(BX), SP	// restore SP
+	MOVL	gobuf_pc(BX), BX
+	PUSHL	BX
+	JMP	AX
+	POPL	BX	// not reached
+
+/*
+ * support for morestack
+ */
+
+// Called during function prolog when more stack is needed.
+TEXT sys·morestack(SB),7,$0
+	// Cannot grow scheduler stack (m->g0).
+	MOVL	m, BX
+	MOVL	m_g0(BX), SI
+	CMPL	g, SI
+	JNE	2(PC)
+	INT	$3
+
+	// frame size in DX
+	// arg size in AX
+	// Save in m.
+	MOVL	DX, m_moreframe(BX)
+	MOVL	AX, m_moreargs(BX)
+
+	// Called from f.
+	// Set m->morebuf to f's caller.
+	MOVL	4(SP), DI	// f's caller's PC
+	MOVL	DI, (m_morebuf+gobuf_pc)(BX)
+	LEAL	8(SP), CX	// f's caller's SP
+	MOVL	CX, (m_morebuf+gobuf_sp)(BX)
+	MOVL	g, SI
+	MOVL	SI, (m_morebuf+gobuf_g)(BX)
+
+	// Set m->morepc to f's PC.
+	MOVL	0(SP), AX
+	MOVL	AX, m_morepc(BX)
+
+	// Call newstack on m's scheduling stack.
+	MOVL	m_g0(BX), BP
+	MOVL	BP, g
+	MOVL	(m_sched+gobuf_sp)(BX), SP
+	CALL	newstack(SB)
+	MOVL	$0, 0x1003	// crash if newstack returns
 	RET
 
+// Return point when leaving stack.
+TEXT sys·lessstack(SB), 7, $0
+	// Save return value in m->cret
+	MOVL	m, BX
+	MOVL	AX, m_cret(BX)
+
+	// Call oldstack on m's scheduling stack.
+	MOVL	m_g0(BX), DX
+	MOVL	DX, g
+	MOVL	(m_sched+gobuf_sp)(BX), SP
+	CALL	oldstack(SB)
+	MOVL	$0, 0x1004	// crash if oldstack returns
+	RET
+
+
 // bool cas(int32 *val, int32 old, int32 new)
 // Atomically:
 //	if(*val == old){
@@ -212,6 +256,3 @@ TEXT emptyfunc(SB),0,$0
 TEXT	abort(SB),7,$0
 	INT $0x3
 
-DATA hello+0(SB)/8, $"go386\n\z\z"
-GLOBL hello+0(SB), $8
-

src/pkg/runtime/386/traceback.c

@@ -27,19 +27,19 @@ traceback(byte *pc0, byte *sp, G *g)
 
 	stk = (Stktop*)g->stackbase;
 	for(n=0; n<100; n++) {
-		while(pc == (uintptr)retfromnewstack) {
+		if(pc == (uint64)sys·lessstack) {
+			// printf("--\n");
 			// pop to earlier stack block
-			sp = stk->oldsp;
-			stk = (Stktop*)stk->oldbase;
-			pc = *(uintptr*)(sp+sizeof(uintptr));
-			sp += 2*sizeof(uintptr);	// two irrelevant calls on stack: morestack plus its call
+			pc = (uintptr)stk->gobuf.pc;
+			sp = stk->gobuf.sp;
+			stk = (Stktop*)stk->stackbase;
 		}
 		f = findfunc(pc);
 		if(f == nil) {
 			// dangerous, but poke around to see if it is a closure
 			p = (byte*)pc;
 			// ADDL $xxx, SP; RET
-			if(p[0] == 0x81 && p[1] == 0xc4 && p[6] == 0xc3) {
+			if(p != 0 && p[0] == 0x81 && p[1] == 0xc4 && p[6] == 0xc3) {
 				sp += *(uint32*)(p+2) + 8;
 				pc = *(uintptr*)(sp - 8);
 				if(pc <= 0x1000)
@@ -109,11 +109,10 @@ runtime·Caller(int32 n, uintptr retpc, String retfile, int32 retline, bool retb
 	// now unwind n levels
 	stk = (Stktop*)g->stackbase;
 	while(n-- > 0) {
-		while(pc == (uintptr)retfromnewstack) {
-			sp = stk->oldsp;
-			stk = (Stktop*)stk->oldbase;
-			pc = *((uintptr*)sp + 1);
-			sp += 2*sizeof(uintptr);
+		while(pc == (uintptr)sys·lessstack) {
+			pc = (uintptr)stk->gobuf.pc;
+			sp = stk->gobuf.sp;
+			stk = (Stktop*)stk->stackbase;
 		}
 
 		if(f->frame < sizeof(uintptr))	// assembly functions lie

src/pkg/runtime/darwin/386/sys.s

@@ -6,6 +6,8 @@
 // See http://fxr.watson.org/fxr/source/bsd/kern/syscalls.c?v=xnu-1228
 // or /usr/include/sys/syscall.h (on a Mac) for system call numbers.
 
+#include "386/asm.h"
+
 TEXT notok(SB),7,$0
 	MOVL	$0xf1, 0xf1
 	RET
@@ -26,7 +28,7 @@ TEXT exit1(SB),7,$0
 	CALL	notok(SB)
 	RET
 
-TEXT sys·write(SB),7,$0
+TEXT write(SB),7,$0
 	MOVL	$4, AX
 	INT	$0x80
 	JAE	2(PC)
@@ -56,9 +58,10 @@ TEXT sigaction(SB),7,$0
 //	16(FP)	siginfo
 //	20(FP)	context
 TEXT sigtramp(SB),7,$40
-	MOVL	4(FS), BP	// m
-	MOVL	28(BP), BP	// m->gsignal
-	MOVL	BP, 0(FS)	// g = m->gsignal
+	// g = m->gsignal
+	MOVL	m, BP
+	MOVL	m_gsignal(BP), BP
+	MOVL	BP, g
 
 	MOVL	handler+4(FP), DI
 	MOVL	signo+12(FP), AX
@@ -95,11 +98,11 @@ TEXT bsdthread_create(SB),7,$32
 	// 0(SP) is where the caller PC would be; kernel skips it
 	MOVL	func+12(FP), BX
 	MOVL	BX, 4(SP)	// func
-	MOVL	m+4(FP), BX
+	MOVL	mm+4(FP), BX
 	MOVL	BX, 8(SP)	// arg
 	MOVL	stk+0(FP), BX
 	MOVL	BX, 12(SP)	// stack
-	MOVL	g+8(FP), BX
+	MOVL	gg+8(FP), BX
 	MOVL	BX, 16(SP)	// pthread
 	MOVL	$0x1000000, 20(SP)	// flags = PTHREAD_START_CUSTOM
 	INT	$0x80
@@ -121,7 +124,7 @@ TEXT bsdthread_start(SB),7,$0
 	// set up ldt 7+id to point at m->tls.
 	// m->tls is at m+40.  newosproc left
 	// the m->id in tls[0].
-	LEAL	40(DX), BP
+	LEAL	m_tls(DX), BP
 	MOVL	0(BP), DI
 	ADDL	$7, DI	// m0 is LDT#7. count up.
 	// setldt(tls#, &tls, sizeof tls)
@@ -139,9 +142,9 @@ TEXT bsdthread_start(SB),7,$0
 	MOVW	DI, FS
 
 	// Now segment is established.  Initialize m, g.
-	MOVL	AX, 0(FS)	// g
-	MOVL	DX, 4(FS)	// m
-	MOVL	BX, 20(DX)	// m->procid = thread port (for debuggers)
+	MOVL	AX, g
+	MOVL	DX, m
+	MOVL	BX, m_procid(DX)	// m->procid = thread port (for debuggers)
 	CALL	CX	// fn()
 	CALL	exit1(SB)
 	RET

src/pkg/runtime/darwin/thread.c

@@ -151,8 +151,12 @@ osinit(void)
 void
 newosproc(M *m, G *g, void *stk, void (*fn)(void))
 {
-	// printf("newosproc m=%p g=%p stk=%p fn=%p\n", m, g, stk, fn);
 	m->tls[0] = m->id;	// so 386 asm can find it
+
+	if(0){
+		printf("newosproc stk=%p m=%p g=%p fn=%p id=%d/%d ostk=%p\n",
+			stk, m, g, fn, m->id, m->tls[0], &m);
+	}
 	bsdthread_create(stk, m, g, fn);
 }
 

src/pkg/runtime/linux/386/sys.s

@@ -6,6 +6,8 @@
 // System calls and other sys.stuff for 386, Linux
 //
 
+#include "386/asm.h"
+
 TEXT syscall(SB),7,$0
 	MOVL 4(SP), AX	// syscall number
 	MOVL 8(SP), BX	// arg1
@@ -42,26 +44,6 @@ TEXT write(SB),7,$0
 	INT	$0x80
 	RET
 
-TEXT getpid(SB),7,$0
-	MOVL	$20, AX
-	INT	$0x80
-	RET
-
-TEXT kill(SB),7,$0
-	MOVL	$37, AX
-	MOVL	4(SP), BX
-	MOVL	8(SP), CX
-	INT	$0x80
-	RET
-
-TEXT sys·write(SB),7,$0
-	MOVL	$4, AX		// syscall - write
-	MOVL	4(SP), BX
-	MOVL	8(SP), CX
-	MOVL	12(SP), DX
-	INT	$0x80
-	RET
-
 TEXT rt_sigaction(SB),7,$0
 	MOVL	$174, AX		// syscall - rt_sigaction
 	MOVL	4(SP), BX
@@ -72,18 +54,19 @@ TEXT rt_sigaction(SB),7,$0
 	RET
 
 TEXT sigtramp(SB),7,$0
-	MOVL	4(FS), BP	// m
-	MOVL	20(BP), AX	// m->gsignal
-	MOVL	AX, 0(FS)	// g = m->gsignal
+	MOVL	m, BP
+	MOVL	m_gsignal(BP), AX
+	MOVL	AX, g
 	JMP	sighandler(SB)
 
 TEXT sigignore(SB),7,$0
 	RET
 
 TEXT sigreturn(SB),7,$0
-	MOVL	4(FS), BP	// m
-	MOVL	32(BP), BP	// m->curg
-	MOVL	BP, 0(FS)	// g = m->curg
+	// g = m->curg
+	MOVL	m, BP
+	MOVL	m_curg(BP), BP
+	MOVL	BP, g
 	MOVL	$173, AX	// rt_sigreturn
 	INT $0x80
 	INT $3	// not reached
@@ -104,7 +87,7 @@ TEXT sys·mmap(SB),7,$0
 	INT	$3
 	RET
 
-// int64 futex(int32 *uaddr, int32 op, int32 val,
+// int32 futex(int32 *uaddr, int32 op, int32 val,
 //	struct timespec *timeout, int32 *uaddr2, int32 val2);
 TEXT futex(SB),7,$0
 	MOVL	$240, AX	// futex
@@ -117,22 +100,24 @@ TEXT futex(SB),7,$0
 	INT	$0x80
 	RET
 
-// int64 clone(int32 flags, void *stack, M *m, G *g, void (*fn)(void));
+// int32 clone(int32 flags, void *stack, M *m, G *g, void (*fn)(void));
 TEXT clone(SB),7,$0
 	MOVL	$120, AX	// clone
 	MOVL	flags+4(SP), BX
 	MOVL	stack+8(SP), CX
+	MOVL	$0, DX	// parent tid ptr
+	MOVL	$0, DI	// child tid ptr
 
 	// Copy m, g, fn off parent stack for use by child.
-	SUBL	$12, CX
-	MOVL	m+12(SP), DX
-	MOVL	DX, 0(CX)
-	MOVL	g+16(SP), DX
-	MOVL	DX, 4(CX)
-	MOVL	fn+20(SP), DX
-	MOVL	DX, 8(CX)
+	SUBL	$16, CX
+	MOVL	mm+12(SP), SI
+	MOVL	SI, 0(CX)
+	MOVL	gg+16(SP), SI
+	MOVL	SI, 4(CX)
+	MOVL	fn+20(SP), SI
+	MOVL	SI, 8(CX)
+	MOVL	$1234, 12(CX)
 
-	MOVL	$120, AX
 	INT	$0x80
 
 	// In parent, return.
@@ -140,29 +125,58 @@ TEXT clone(SB),7,$0
 	JEQ	2(PC)
 	RET
 
-	// In child, set up new stack, etc.
-	MOVL	0(CX), BX	// m
-	MOVL	12(AX), AX	// fs (= m->cret)
-	MOVW	AX, FS
-	MOVL	8(CX), DX	// fn
-	ADDL	$12, CX
-	MOVL	CX, SP
+	// Paranoia: check that SP is as we expect.
+	MOVL	12(SP), BP
+	CMPL	BP, $1234
+	JEQ	2(PC)
+	INT	$3
 
-	// fn is now on top of stack.
-
-	// initialize m->procid to Linux tid
+	// Initialize AX to Linux tid
 	MOVL	$224, AX
 	INT	$0x80
-	MOVL	AX, 20(BX)
 
-	// call fn
-	CALL	DX
+	// In child on new stack.  Reload registers (paranoia).
+	MOVL	0(SP), BX	// m
+	MOVL	4(SP), DX	// g
+	MOVL	8(SP), CX	// fn
 
-	// It shouldn't return; if it does, exit.
-	MOVL	$111, DI
-	MOVL	$1, AX
-	INT	$0x80
-	JMP	-3(PC)	// keep exiting
+	MOVL	AX, m_procid(BX)	// save tid as m->procid
+
+	// set up ldt 7+id to point at m->tls.
+	// m->tls is at m+40.  newosproc left the id in tls[0].
+	LEAL	m_tls(BX), BP
+	MOVL	0(BP), DI
+	ADDL	$7, DI	// m0 is LDT#7. count up.
+	// setldt(tls#, &tls, sizeof tls)
+	PUSHAL	// save registers
+	PUSHL	$32	// sizeof tls
+	PUSHL	BP	// &tls
+	PUSHL	DI	// tls #
+	CALL	setldt(SB)
+	POPL	AX
+	POPL	AX
+	POPL	AX
+	POPAL
+	SHLL	$3, DI	// segment# is ldt*8 + 7 (different 7 than above)
+	ADDL	$7, DI
+	MOVW	DI, FS
+
+	// Now segment is established.  Initialize m, g.
+	MOVL	DX, g
+	MOVL	BX, m
+
+	MOVL	0(DX), DX	// paranoia; check they are not nil
+	MOVL	0(BX), BX
+
+	// more paranoia; check that stack splitting code works
+	PUSHAL
+	CALL	emptyfunc(SB)
+	POPAL
+
+	CALL	CX	// fn()
+	CALL	exit1(SB)
+	MOVL	$0x1234, 0x1005
+	RET
 
 TEXT sigaltstack(SB),7,$-8
 	MOVL	$186, AX	// sigaltstack
@@ -174,10 +188,6 @@ TEXT sigaltstack(SB),7,$-8
 	INT	$3
 	RET
 
-//	// fake the per-goroutine and per-mach registers
-//	LEAL	m0(SB),
-
-// TODO(rsc): move to linux.s
 // <asm-i386/ldt.h>
 // struct user_desc {
 // 	unsigned int  entry_number;

src/pkg/runtime/linux/os.h

@@ -3,8 +3,8 @@
 // license that can be found in the LICENSE file.
 
 // Linux-specific system calls
-int64	futex(uint32*, int32, uint32, Timespec*, uint32*, uint32);
-int64	clone(int32, void*, M*, G*, void(*)(void));
+int32	futex(uint32*, int32, uint32, Timespec*, uint32*, uint32);
+int32	clone(int32, void*, M*, G*, void(*)(void));
 
 struct Sigaction;
 void	rt_sigaction(int64, struct Sigaction*, void*, uint64);

src/pkg/runtime/linux/thread.c

@@ -42,7 +42,7 @@ static Timespec longtime =
 static void
 futexsleep(uint32 *addr, uint32 val)
 {
-	int64 ret;
+	int32 ret;
 
 	ret = futex(addr, FUTEX_WAIT, val, &longtime, nil, 0);
 	if(ret >= 0 || ret == -EAGAIN || ret == -EINTR)
@@ -234,7 +234,7 @@ enum
 void
 newosproc(M *m, G *g, void *stk, void (*fn)(void))
 {
-	int64 ret;
+	int32 ret;
 	int32 flags;
 
 	/*
@@ -248,21 +248,14 @@ newosproc(M *m, G *g, void *stk, void (*fn)(void))
 		| CLONE_THREAD	/* revisit - okay for now */
 		;
 
+	m->tls[0] = m->id;	// so 386 asm can find it
 	if(0){
-		prints("newosproc stk=");
-		sys·printpointer(stk);
-		prints(" m=");
-		sys·printpointer(m);
-		prints(" g=");
-		sys·printpointer(g);
-		prints(" fn=");
-		sys·printpointer(fn);
-		prints(" clone=");
-		sys·printpointer(clone);
-		prints("\n");
+		printf("newosproc stk=%p m=%p g=%p fn=%p clone=%p id=%d/%d ostk=%p\n",
+			stk, m, g, fn, clone, m->id, m->tls[0], &m);
 	}
 
 	ret = clone(flags, stk, m, g, fn);
+
 	if(ret < 0)
 		*(int32*)123 = 123;
 }