[dev.power64] liblink: power64 fixes and ports of changes

Ports of platform-specific changes that happened on default
after dev.power64 forked (fixes for c2go, wrapper math fixes,
moved stackguard field, stackguard1 support, precise stacks).
Bug fixes (missing AMOVW in instruction table, correct
unsigned 32-bit remainder).

LGTM=rsc
R=rsc
CC=golang-codereviews
https://codereview.appspot.com/164920044

2 files changed, 102 insertions, 126 deletions

diff --git a/src/liblink/asm9.c b/src/liblink/asm9.c
index 47d515e05..c5a78e53c 100644
--- a/src/liblink/asm9.c
+++ b/src/liblink/asm9.c
@@ -232,6 +232,7 @@ static Optab	optab[] = {
 	{ AMOVBZ,	C_REG,	C_NONE, C_NONE, 	C_LAUTO,	35, 8, REGSP },
 	{ AMOVB,	C_REG,	C_NONE, C_NONE, 	C_LAUTO,	35, 8, REGSP },
 	{ AMOVD,	C_REG,	C_NONE, C_NONE, 	C_LOREG,	35, 8, REGZERO },
+	{ AMOVW,	C_REG,	C_NONE, C_NONE, 	C_LOREG,	35, 8, REGZERO },
 	{ AMOVWZ,	C_REG,	C_NONE, C_NONE, 	C_LOREG,	35, 8, REGZERO },
 	{ AMOVBZ,	C_REG,	C_NONE, C_NONE, 	C_LOREG,	35, 8, REGZERO },
 	{ AMOVB,	C_REG,	C_NONE, C_NONE, 	C_LOREG,	35, 8, REGZERO },
@@ -349,8 +350,12 @@ static Optab	optab[] = {
 
 	{ AREM,		C_REG,	C_NONE, C_NONE, 	C_REG,		50, 12, 0 },
 	{ AREM,		C_REG,	C_REG, C_NONE, 	C_REG,		50, 12, 0 },
+	{ AREMU,		C_REG,	C_NONE, C_NONE, 	C_REG,		50, 16, 0 },
+	{ AREMU,		C_REG,	C_REG, C_NONE, 	C_REG,		50, 16, 0 },
 	{ AREMD,		C_REG,	C_NONE, C_NONE, 	C_REG,		51, 12, 0 },
 	{ AREMD,		C_REG,	C_REG, C_NONE, 	C_REG,		51, 12, 0 },
+	{ AREMDU,		C_REG,	C_NONE, C_NONE, 	C_REG,		51, 12, 0 },
+	{ AREMDU,		C_REG,	C_REG, C_NONE, 	C_REG,		51, 12, 0 },
 
 	{ AMTFSB0,	C_SCON,	C_NONE, C_NONE, 	C_NONE,		52, 4, 0 },
 	{ AMOVFL, C_FPSCR, C_NONE, C_NONE,	C_FREG,		53, 4, 0 },
@@ -433,12 +438,11 @@ static Optab	optab[] = {
 	{ AUSEFIELD,	C_ADDR,	C_NONE,	C_NONE, C_NONE,	0, 0, 0 },
 	{ APCDATA,	C_LCON,	C_NONE,	C_NONE, C_LCON,	0, 0, 0 },
 	{ AFUNCDATA,	C_SCON,	C_NONE,	C_NONE, C_ADDR,	0, 0, 0 },
+	{ ANOP,		C_NONE, C_NONE, C_NONE, C_NONE, 0, 0, 0 },
 
 	{ ADUFFZERO,	C_NONE,	C_NONE, C_NONE,	C_LBRA,	11, 4, 0 },  // same as ABR/ABL
 	{ ADUFFCOPY,	C_NONE,	C_NONE, C_NONE,	C_LBRA,	11, 4, 0 },  // same as ABR/ABL
 
-	{ ANOP,		C_NONE, C_NONE, C_NONE, C_NONE, 0, 0, 0 },
-
 	{ AXXX,		C_NONE,	C_NONE, C_NONE, 	C_NONE,		 0, 4, 0 },
 };
 
@@ -472,7 +476,7 @@ static struct
 	Optab*	stop;
 } oprange[ALAST];
 
-static char	xcmp[C_NCLASS][C_NCLASS];
+static uchar	xcmp[C_NCLASS][C_NCLASS];
 
 
 void
@@ -747,7 +751,7 @@ static Optab*
 oplook(Link *ctxt, Prog *p)
 {
 	int a1, a2, a3, a4, r;
-	char *c1, *c3, *c4;
+	uchar *c1, *c3, *c4;
 	Optab *o, *e;
 
 	a1 = p->optab;
@@ -931,6 +935,8 @@ buildop(Link *ctxt)
 			oprange[AREMCC] = oprange[r];
 			oprange[AREMV] = oprange[r];
 			oprange[AREMVCC] = oprange[r];
+			break;
+		case AREMU:
 			oprange[AREMU] = oprange[r];
 			oprange[AREMUCC] = oprange[r];
 			oprange[AREMUV] = oprange[r];
@@ -940,6 +946,8 @@ buildop(Link *ctxt)
 			oprange[AREMDCC] = oprange[r];
 			oprange[AREMDV] = oprange[r];
 			oprange[AREMDVCC] = oprange[r];
+			break;
+		case AREMDU:
 			oprange[AREMDU] = oprange[r];
 			oprange[AREMDUCC] = oprange[r];
 			oprange[AREMDUV] = oprange[r];
@@ -2005,6 +2013,11 @@ asmout(Link *ctxt, Prog *p, Optab *o, int32 *out)
 		o1 = AOP_RRR(v&~t, REGTMP, r, p->from.reg);
 		o2 = AOP_RRR(OP_MULLW, REGTMP, REGTMP, p->from.reg);
 		o3 = AOP_RRR(OP_SUBF|t, p->to.reg, REGTMP, r);
+		if(p->as == AREMU) {
+			o4 = o3;
+			/* Clear top 32 bits */
+			o3 = OP_RLW(OP_RLDIC, REGTMP, REGTMP, 0, 0, 0) | (1<<5);
+		}
 		break;
 
 	case 51:	/* remd[u] r1[,r2],r3 */
@@ -2264,54 +2277,6 @@ asmout(Link *ctxt, Prog *p, Optab *o, int32 *out)
 	out[3] = o4;
 	out[4] = o5;
 	return;
-
-#if NOTDEF
-	v = p->pc;
-	switch(o->size) {
-	default:
-		if(debug['a'])
-			Bprint(&bso, " %.8lux:\t\t%P\n", v, p);
-		break;
-	case 4:
-		if(debug['a'])
-			Bprint(&bso, " %.8lux: %.8lux\t%P\n", v, o1, p);
-		lput(o1);
-		break;
-	case 8:
-		if(debug['a'])
-			Bprint(&bso, " %.8lux: %.8lux %.8lux%P\n", v, o1, o2, p);
-		lput(o1);
-		lput(o2);
-		break;
-	case 12:
-		if(debug['a'])
-			Bprint(&bso, " %.8lux: %.8lux %.8lux %.8lux%P\n", v, o1, o2, o3, p);
-		lput(o1);
-		lput(o2);
-		lput(o3);
-		break;
-	case 16:
-		if(debug['a'])
-			Bprint(&bso, " %.8lux: %.8lux %.8lux %.8lux %.8lux%P\n",
-				v, o1, o2, o3, o4, p);
-		lput(o1);
-		lput(o2);
-		lput(o3);
-		lput(o4);
-		break;
-	case 20:
-		if(debug['a'])
-			Bprint(&bso, " %.8lux: %.8lux %.8lux %.8lux %.8lux %.8lux%P\n",
-				v, o1, o2, o3, o4, o5, p);
-		lput(o1);
-		lput(o2);
-		lput(o3);
-		lput(o4);
-		lput(o5);
-		break;
-	}
-	return 0;
-#endif
 }
 
 static vlong
diff --git a/src/liblink/obj9.c b/src/liblink/obj9.c
index 3906181a3..fa8af5f83 100644
--- a/src/liblink/obj9.c
+++ b/src/liblink/obj9.c
@@ -199,7 +199,7 @@ parsetextconst(vlong arg, vlong *textstksiz, vlong *textarg)
 static void
 addstacksplit(Link *ctxt, LSym *cursym)
 {
-	Prog *p, *q, *q1;
+	Prog *p, *q, *p1, *p2, *q1;
 	int o, mov, aoffset;
 	vlong textstksiz, textarg;
 	int32 autoffset, autosize;
@@ -472,32 +472,92 @@ addstacksplit(Link *ctxt, LSym *cursym)
 				q->spadj = -aoffset;
 
 			if(cursym->text->reg & WRAPPER) {
-				// g->panicwrap += autosize;
-				// MOVWZ panicwrap_offset(g), R3
-				// ADD $autosize, R3
-				// MOVWZ R3, panicwrap_offset(g)
-				p = appendp(ctxt, q);
-				p->as = AMOVWZ;
+				// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
+				//
+				//	MOVD g_panic(g), R3
+				//	CMP R0, R3
+				//	BEQ end
+				//	MOVD panic_argp(R3), R4
+				//	ADD $(autosize+8), R1, R5
+				//	CMP R4, R5
+				//	BNE end
+				//	ADD $8, R1, R6
+				//	MOVD R6, panic_argp(R3)
+				// end:
+				//	NOP
+				//
+				// The NOP is needed to give the jumps somewhere to land.
+				// It is a liblink NOP, not a Power64 NOP: it encodes to 0 instruction bytes.
+
+
+				p = appendp(ctxt, p);
+				p->as = AMOVD;
 				p->from.type = D_OREG;
 				p->from.reg = REGG;
-				p->from.offset = 2*ctxt->arch->ptrsize;
+				p->from.offset = 4*ctxt->arch->ptrsize; // G.panic
 				p->to.type = D_REG;
 				p->to.reg = 3;
 
 				p = appendp(ctxt, p);
+				p->as = ACMP;
+				p->from.type = D_REG;
+				p->from.reg = 0;
+				p->to.type = D_REG;
+				p->to.reg = 3;
+
+				p = appendp(ctxt, p);
+				p->as = ABEQ;
+				p->to.type = D_BRANCH;
+				p1 = p;
+
+				p = appendp(ctxt, p);
+				p->as = AMOVD;
+				p->from.type = D_OREG;
+				p->from.reg = 3;
+				p->from.offset = 0; // Panic.argp
+				p->to.type = D_REG;
+				p->to.reg = 4;
+
+				p = appendp(ctxt, p);
 				p->as = AADD;
 				p->from.type = D_CONST;
-				p->from.offset = autosize;
+				p->from.offset = autosize+8;
+				p->reg = REGSP;
 				p->to.type = D_REG;
-				p->to.reg = 3;
+				p->to.reg = 5;
 
 				p = appendp(ctxt, p);
-				p->as = AMOVWZ;
+				p->as = ACMP;
 				p->from.type = D_REG;
-				p->from.reg = 3;
+				p->from.reg = 4;
+				p->to.type = D_REG;
+				p->to.reg = 5;
+
+				p = appendp(ctxt, p);
+				p->as = ABNE;
+				p->to.type = D_BRANCH;
+				p2 = p;
+
+				p = appendp(ctxt, p);
+				p->as = AADD;
+				p->from.type = D_CONST;
+				p->from.offset = 8;
+				p->reg = REGSP;
+				p->to.type = D_REG;
+				p->to.reg = 6;
+
+				p = appendp(ctxt, p);
+				p->as = AMOVD;
+				p->from.type = D_REG;
+				p->from.reg = 6;
 				p->to.type = D_OREG;
-				p->to.reg = REGG;
-				p->to.offset = 2*ctxt->arch->ptrsize;
+				p->to.reg = 3;
+				p->to.offset = 0; // Panic.argp
+
+				p = appendp(ctxt, p);
+				p->as = ANOP;
+				p1->pcond = p;
+				p2->pcond = p;
 			}
 
 			break;
@@ -512,36 +572,6 @@ addstacksplit(Link *ctxt, LSym *cursym)
 				p->to.type = D_BRANCH;
 				break;
 			}
-			if(cursym->text->reg & WRAPPER) {
-				// g->panicwrap -= autosize;
-				// MOVWZ panicwrap_offset(g), R3
-				// ADD $-autosize, R3
-				// MOVWZ R3, panicwrap_offset(g)
-				p->as = AMOVWZ;
-				p->from.type = D_OREG;
-				p->from.reg = REGG;
-				p->from.offset = 2*ctxt->arch->ptrsize;
-				p->to.type = D_REG;
-				p->to.reg = 3;
-				p = appendp(ctxt, p);
-
-				p->as = AADD;
-				p->from.type = D_CONST;
-				p->from.offset = -autosize;
-				p->to.type = D_REG;
-				p->to.reg = 3;
-				p = appendp(ctxt, p);
-
-				p->as = AMOVWZ;
-				p->from.type = D_REG;
-				p->from.reg = 3;
-				p->to.type = D_OREG;
-				p->to.reg = REGG;
-				p->to.offset = 2*ctxt->arch->ptrsize;
-				p = appendp(ctxt, p);
-
-				p->as = ARETURN;
-			}
 			if(cursym->text->mark & LEAF) {
 				if(!autosize) {
 					p->as = ABR;
@@ -673,18 +703,19 @@ addstacksplit(Link *ctxt, LSym *cursym)
 static Prog*
 stacksplit(Link *ctxt, Prog *p, int32 framesize, int noctxt)
 {
-	int32 arg;
-	Prog *q, *q1;
+	Prog *q1;
 
 	// MOVD	g_stackguard(g), R3
 	p = appendp(ctxt, p);
 	p->as = AMOVD;
 	p->from.type = D_OREG;
 	p->from.reg = REGG;
+	p->from.offset = 2*ctxt->arch->ptrsize;	// G.stackguard0
+	if(ctxt->cursym->cfunc)
+		p->from.offset = 3*ctxt->arch->ptrsize;	// G.stackguard1
 	p->to.type = D_REG;
 	p->to.reg = 3;
 
-	q = nil;
 	if(framesize <= StackSmall) {
 		// small stack: SP < stackguard
 		//	CMP	stackguard, SP
@@ -735,7 +766,7 @@ stacksplit(Link *ctxt, Prog *p, int32 framesize, int noctxt)
 		p->to.type = D_CONST;
 		p->to.offset = StackPreempt;
 
-		q = p = appendp(ctxt, p);
+		p = appendp(ctxt, p);
 		p->as = ABEQ;
 		p->to.type = D_BRANCH;
 
@@ -774,31 +805,6 @@ stacksplit(Link *ctxt, Prog *p, int32 framesize, int noctxt)
 	p->as = ABLT;
 	p->to.type = D_BRANCH;
 
-	// MOVD	$framesize, R3
-	p = appendp(ctxt, p);
-	p->as = AMOVD;
-	p->from.type = D_CONST;
-	p->from.offset = framesize;
-	p->to.type = D_REG;
-	p->to.reg = 3;
-	if(q)
-		q->pcond = p;
-
-	// MOVD	$args, R4
-	p = appendp(ctxt, p);
-	p->as = AMOVD;
-	p->from.type = D_CONST;
-	arg = (ctxt->cursym->text->to.offset >> 32) & 0xffffffffull;
-	if(arg == 1) // special marker for known 0
-		arg = 0;
-	else if(arg == ArgsSizeUnknown)
-		ctxt->diag("%s: arg size unknown, but split stack", ctxt->cursym->name);
-	if(arg&7)
-		ctxt->diag("misaligned argument size in stack split: %d", arg);
-	p->from.offset = arg;
-	p->to.type = D_REG;
-	p->to.reg = 4;
-
 	// MOVD	LR, R5
 	p = appendp(ctxt, p);
 	p->as = AMOVD;
@@ -811,7 +817,10 @@ stacksplit(Link *ctxt, Prog *p, int32 framesize, int noctxt)
 	p = appendp(ctxt, p);
 	p->as = ABL;
 	p->to.type = D_BRANCH;
-	p->to.sym = ctxt->symmorestack[noctxt];
+	if(ctxt->cursym->cfunc)
+		p->to.sym = linklookup(ctxt, "runtime.morestackc", 0);
+	else
+		p->to.sym = ctxt->symmorestack[noctxt];
 
 	// BR	start
 	p = appendp(ctxt, p);
@@ -1010,6 +1019,7 @@ LinkArch linkpower64 = {
 	.D_PARAM = D_PARAM,
 	.D_SCONST = D_SCONST,
 	.D_STATIC = D_STATIC,
+	.D_OREG = D_OREG,
 
 	.ACALL = ABL,
 	.ADATA = ADATA,
@@ -1056,6 +1066,7 @@ LinkArch linkpower64le = {
 	.D_PARAM = D_PARAM,
 	.D_SCONST = D_SCONST,
 	.D_STATIC = D_STATIC,
+	.D_OREG = D_OREG,
 
 	.ACALL = ABL,
 	.ADATA = ADATA,