segmented stack

SVN=125151

src/cmd/6a/a.y

@@ -52,8 +52,8 @@
 %token	<dval>	LFCONST
 %token	<sval>	LSCONST LSP
 %token	<sym>	LNAME LLAB LVAR
-%type	<lval>	con expr pointer offset
-%type	<gen>	mem imm reg nam rel rem rim rom omem nmem
+%type	<lval>	con con3 expr pointer offset
+%type	<gen>	mem imm imm3 reg nam rel rem rim rom omem nmem
 %type	<gen2>	nonnon nonrel nonrem rimnon rimrem remrim spec10
 %type	<gen2>	spec1 spec2 spec3 spec4 spec5 spec6 spec7 spec8 spec9
 %%
@@ -177,12 +177,12 @@ spec1:	/* DATA */
 	}
 
 spec2:	/* TEXT */
-	mem ',' imm
+	mem ',' imm3
 	{
 		$$.from = $1;
 		$$.to = $3;
 	}
-|	mem ',' con ',' imm
+|	mem ',' con ',' imm3
 	{
 		$$.from = $1;
 		$$.from.scale = $3;
@@ -364,6 +364,14 @@ reg:
 		$$.type = $1;
 	}
 
+imm3:
+	'$' con3
+	{
+		$$ = nullgen;
+		$$.type = D_CONST;
+		$$.offset = $2;
+	}
+
 imm:
 	'$' con
 	{
@@ -548,6 +556,25 @@ con:
 		$$ = $2;
 	}
 
+con3:
+	LCONST
+|	'-' LCONST
+	{
+		$$ = -$2;
+	}
+|	LCONST '-' LCONST '-' LCONST
+	{
+		$$ = ($1 & 0xffffffffLL) +
+			(($3 & 0xffffLL) << 32) +
+			(($5 & 0xffffLL) << 48);
+	}
+|	'-' LCONST '-' LCONST '-' LCONST
+	{
+		$$ = (-$2 & 0xffffffffLL) +
+			(($4 & 0xffffLL) << 32) +
+			(($6 & 0xffffLL) << 48);
+	}
+
 expr:
 	con
 |	expr '+' expr

src/cmd/6l/6.out.h

@@ -32,6 +32,7 @@
 #define	NSNAME	8
 #define NOPROF	(1<<0)
 #define DUPOK	(1<<1)
+#define NOSPLIT	(1<<2)
 #define SOFmark	"\xa7\xf1\xd9\x2a\x82\xc8\xd8\xfe"
 
 /*

src/cmd/6l/l.h

@@ -345,6 +345,9 @@ EXTERN	int	exports, nexports;
 EXTERN	char*	EXPTAB;
 EXTERN	Prog	undefp;
 EXTERN	ulong	stroffset;
+EXTERN	vlong	textstksiz;
+EXTERN	vlong	textinarg;
+EXTERN	vlong	textoutarg;
 
 #define	UP	(&undefp)
 
@@ -407,6 +410,7 @@ void	objfile(char*);
 int	opsize(Prog*);
 void	patch(void);
 Prog*	prg(void);
+void	parsetextconst(vlong);
 void	readundefs(char*, int);
 int	relinv(int);
 long	reuse(Prog*, Sym*);

src/cmd/6l/list.c

@@ -56,10 +56,13 @@ Pconv(Fmt *fp)
 	switch(p->as) {
 	case ATEXT:
 		if(p->from.scale) {
-			sprint(str, "%-7s %-7A %D,%d,%D",
+			sprint(str, "%-7s %-7A %D,%d,%lD",
 				str1, p->as, &p->from, p->from.scale, &p->to);
 			break;
 		}
+		sprint(str, "%-7s %-7A %D,%lD",
+			str1, p->as, &p->from, &p->to);
+		break;
 
 	default:
 		sprint(str, "%-7s %-7A %D,%D",
@@ -95,6 +98,22 @@ Dconv(Fmt *fp)
 
 	a = va_arg(fp->args, Adr*);
 	i = a->type;
+
+	if(fp->flags & FmtLong) {
+		if(i != D_CONST) {
+			// ATEXT dst is not constant
+			sprint(str, "!!%D", a);
+			goto brk;
+		}
+		parsetextconst(a->offset);
+		if(textinarg == 0 && textoutarg == 0) {
+			sprint(str, "$%lld", textstksiz);
+			goto brk;
+		}
+		sprint(str, "$%lld-%lld-%lld", textstksiz, textinarg, textoutarg);
+		goto brk;
+	}
+
 	if(i >= D_INDIR) {
 		if(a->offset)
 			sprint(str, "%lld(%R)", a->offset, i-D_INDIR);
@@ -395,3 +414,26 @@ diag(char *fmt, ...)
 		errorexit();
 	}
 }
+
+void
+parsetextconst(vlong arg)
+{
+	textstksiz = arg & 0xffffffffLL;
+	if(textstksiz & 0x80000000LL)
+		textstksiz = -(-textstksiz & 0xffffffffLL);
+		
+
+	// the following throws away one bit
+	// of precision, but maintains compat
+	textinarg = (arg >> 32) & 0xffffLL;
+	if(textinarg & 0x8000LL)
+		textinarg = -(-textinarg & 0xffffLL);
+	if(textinarg <= 0)
+		textinarg = 100;
+
+	textoutarg = (arg >> 48) & 0xffffLL;
+	if(textoutarg & 0x8000LL)
+		textoutarg = -(-textoutarg & 0xffffLL);
+	if(textoutarg <= 0)
+		textoutarg = 0;
+}

src/cmd/6l/pass.c

@@ -568,6 +568,23 @@ dostkoff(void)
 	Prog *p, *q;
 	long autoffset, deltasp;
 	int a, f, curframe, curbecome, maxbecome, pcsize;
+	Prog *pmorestack;
+	Sym *symmorestack;
+
+	pmorestack = P;
+	symmorestack = lookup("_morestack", 0);
+
+	if(symmorestack->type == STEXT)
+	for(p = firstp; p != P; p = p->link) {
+		if(p->as == ATEXT) {
+			if(p->from.sym == symmorestack) {
+				pmorestack = p;
+				break;
+			}
+		}
+	}
+	if(pmorestack == P)
+		diag("_morestack not defined");
 
 	curframe = 0;
 	curbecome = 0;
@@ -643,15 +660,72 @@ dostkoff(void)
 	for(p = firstp; p != P; p = p->link) {
 		if(p->as == ATEXT) {
 			curtext = p;
-			autoffset = p->to.offset;
+			parsetextconst(p->to.offset);
+			autoffset = textstksiz;
 			if(autoffset < 0)
 				autoffset = 0;
+
+			q = P;
+			if(pmorestack != P)
+			if(!(p->from.scale & NOSPLIT)) {
+				if(autoffset <= 50) {
+					// small stack
+					p = appendp(p);
+					p->as = ACMPQ;
+					p->from.type = D_SP;
+					p->to.type = D_INDIR+D_R15;
+					
+				} else {
+					// large stack
+					p = appendp(p);
+					p->as = AMOVQ;
+					p->from.type = D_SP;
+					p->to.type = D_AX;
+
+					p = appendp(p);
+					p->as = ASUBQ;
+					p->from.type = D_CONST;
+					p->from.offset = autoffset-50;
+					p->to.type = D_AX;
+
+					p = appendp(p);
+					p->as = ACMPQ;
+					p->from.type = D_AX;
+					p->to.type = D_INDIR+D_R15;
+				}
+				// common
+				p = appendp(p);
+				p->as = AJHI;
+				p->to.type = D_BRANCH;
+				p->to.offset = 3;
+				q = p;
+
+				p = appendp(p);
+				p->as = AMOVQ;
+				p->from.type = D_CONST;
+				p->from.offset = curtext->to.offset;
+				p->to.type = D_AX;
+
+				p = appendp(p);
+				p->as = ACALL;
+				p->to.type = D_BRANCH;
+				p->pcond = pmorestack;
+				p->to.sym = symmorestack;
+
+			}
+
+			if(q != P)
+				q->pcond = p->link;
+
 			if(autoffset) {
 				p = appendp(p);
 				p->as = AADJSP;
 				p->from.type = D_CONST;
 				p->from.offset = autoffset;
+				if(q != P)
+					q->pcond = p;
 			}
+
 			deltasp = autoffset;
 		}
 		pcsize = p->mode/8;

src/runtime/rt0_amd64_darwin.s

@@ -36,10 +36,10 @@ done:
 	POPQ	AX
 	RET
 
-TEXT	FLUSH(SB),1,$-8
+TEXT	FLUSH(SB),7,$-8
 	RET
 
-TEXT	sys·exit(SB),1,$-8
+TEXT	sys·exit(SB),7,$-8
 	MOVL	8(SP), DI		// arg 1 exit status
 	MOVL	$(0x2000000+1), AX
 	SYSCALL

src/runtime/rt0_amd64_linux.s

@@ -3,48 +3,59 @@
 // license that can be found in the LICENSE file.
 
 
-TEXT	_rt0_amd64_linux(SB),1,$-8
-	PUSHQ	$0
-	MOVQ	SP, BP
-	ANDQ	$~15, SP
-	MOVQ	8(BP), DI	// argc
-	LEAQ	16(BP), SI	// argv
-	MOVL	DI, DX
-	ADDL	$1, DX
-	SHLL	$3, DX
-	ADDQ	SI, DX
-	MOVQ	DX, CX
-	CMPQ	(CX), $0
-	JEQ	done
-
-loop:
-	ADDQ	$8, CX
-	CMPQ	(CX), $0
-	JNE	loop
-
-done:
-	ADDQ	$8, CX
-	SUBQ	$16, SP
-	MOVL	DI, 0(SP)
-	MOVQ	SI, 8(SP)
-	CALL	args(SB)
-	ADDQ	$16, SP
+TEXT	_rt0_amd64_linux(SB),7,$-8
+
+// copy arguments forward on an even stack
+
+
+	MOVQ	0(SP), AX		// argc
+	LEAQ	8(SP), BX		// argv
+	ANDQ	$~7, SP
+	SUBQ	$32, SP
+	MOVQ	AX, 16(SP)
+	MOVQ	BX, 24(SP)
+
+// allocate the per-user block
+
+	LEAQ	peruser<>(SB), R15	// dedicated u. register
+	MOVQ	SP, AX
+	SUBQ	$4096, AX
+	MOVQ	AX, 0(R15)
+
 	CALL	check(SB)
+
+// process the arguments
+
+	MOVL	16(SP), AX
+	MOVL	AX, 0(SP)
+	MOVQ	24(SP), AX
+	MOVQ	AX, 8(SP)
+	CALL	args(SB)
+
 	CALL	main·main(SB)
+
+	MOVQ	$0, AX
+	MOVQ	AX, 0(SP)		// exit status
 	CALL	sys·exit(SB)
+
 	CALL	notok(SB)
-	POPQ	AX
+
+	ADDQ	$32, SP
+	RET
+
+TEXT	_morestack(SB), 7, $0
+	MOVQ	SP, AX
+	SUBQ	$1024, AX
+	MOVQ	AX, 0(R15)
 	RET
 
-TEXT	FLUSH(SB),1,$-8
+TEXT	FLUSH(SB),7,$-8
 	RET
 
 TEXT	sys·exit(SB),1,$-8
 	MOVL	8(SP), DI
 	MOVL	$60, AX
 	SYSCALL
-	JCC	2(PC)
-	CALL	notok(SB)
 	RET
 
 TEXT	sys·write(SB),1,$-8
@@ -53,8 +64,6 @@ TEXT	sys·write(SB),1,$-8
 	MOVL	24(SP), DX
 	MOVL	$1, AX			// syscall entry
 	SYSCALL
-	JCC	2(PC)
-	CALL	notok(SB)
 	RET
 
 TEXT	open(SB),1,$-8
@@ -93,8 +102,6 @@ TEXT	sys·rt_sigaction(SB),1,$-8
 	MOVL	CX, R10
 	MOVL	$13, AX			// syscall entry
 	SYSCALL
-	JCC	2(PC)
-	CALL	notok(SB)
 	RET
 
 TEXT	sigtramp(SB),1,$24
@@ -151,3 +158,5 @@ TEXT	sys·getcallerpc+0(SB),0,$0
 	MOVQ	x+0(FP),AX
 	MOVQ	-8(AX),AX
 	RET
+
+GLOBL	peruser<>(SB),$64