5g: alignment fixes

Makes all.bash work after echo 4 >/proc/cpu/alignment,
which means kill the process on an unaligned access.

The default behavior on DreamPlug/GuruPlug/SheevaPlug
is to simulate an ARMv3 and just let the unaligned accesses
stop at the word boundary, resulting in all kinds of surprises.

Fixes #1240.

R=ken2
CC=golang-dev
https://golang.org/cl/4551064

src/cmd/5g/cgen.c

@@ -1190,7 +1190,8 @@ void
 sgen(Node *n, Node *res, int32 w)
 {
 	Node dst, src, tmp, nend;
-	int32 c, q, odst, osrc;
+	int32 c, odst, osrc;
+	int dir, align, op;
 	Prog *p, *ploop;
 
 	if(debug['g']) {
@@ -1200,17 +1201,39 @@ sgen(Node *n, Node *res, int32 w)
 	}
 	if(w == 0)
 		return;
-	if(n->ullman >= UINF && res->ullman >= UINF) {
-		fatal("sgen UINF");
-	}
-
 	if(w < 0)
 		fatal("sgen copy %d", w);
+	if(n->ullman >= UINF && res->ullman >= UINF)
+		fatal("sgen UINF");
+	if(n->type == T)
+		fatal("sgen: missing type");
+
+	// determine alignment.
+	// want to avoid unaligned access, so have to use
+	// smaller operations for less aligned types.
+	// for example moving [4]byte must use 4 MOVB not 1 MOVW.
+	align = n->type->align;
+	op = 0;
+	switch(align) {
+	default:
+		fatal("sgen: invalid alignment %d for %T", align, n->type);
+	case 1:
+		op = AMOVB;
+		break;
+	case 2:
+		op = AMOVH;
+		break;
+	case 4:
+		op = AMOVW;
+		break;
+	}
+	if(w%align)
+		fatal("sgen: unaligned size %d (align=%d) for %T", w, align, n->type);
+	c = w / align;
 
 	// offset on the stack
 	osrc = stkof(n);
 	odst = stkof(res);
-
 	if(osrc != -1000 && odst != -1000 && (osrc == 1000 || odst == 1000)) {
 		// osrc and odst both on stack, and at least one is in
 		// an unknown position.  Could generate code to test
@@ -1221,12 +1244,15 @@ sgen(Node *n, Node *res, int32 w)
 		sgen(&tmp, res, w);
 		return;
 	}
-
+	if(osrc%align != 0 || odst%align != 0)
-	if(osrc % 4 != 0 || odst %4 != 0)
+		fatal("sgen: unaligned offset src %d or dst %d (align %d)", osrc, odst, align);
-		fatal("sgen: non word(4) aligned offset src %d or dst %d", osrc, odst);
+	// if we are copying forward on the stack and
+	// the src and dst overlap, then reverse direction
+	dir = align;
+	if(osrc < odst && odst < osrc+w)
+		dir = -dir;
 
 	regalloc(&dst, types[tptr], res);
-
 	if(n->ullman >= res->ullman) {
 		agen(n, &dst);	// temporarily use dst
 		regalloc(&src, types[tptr], N);
@@ -1240,140 +1266,63 @@ sgen(Node *n, Node *res, int32 w)
 
 	regalloc(&tmp, types[TUINT32], N);
 
-	c = w % 4;	// bytes
+	// set up end marker
-	q = w / 4;	// quads
+	memset(&nend, 0, sizeof nend);
-
+	if(c >= 4) {
-	// if we are copying forward on the stack and
-	// the src and dst overlap, then reverse direction
-	if(osrc < odst && odst < osrc+w) {
-		if(c != 0)
-			fatal("sgen: reverse character copy not implemented");
-		if(q >= 4) {
 		regalloc(&nend, types[TUINT32], N);
-			// set up end marker to 4 bytes before source
+
 		p = gins(AMOVW, &src, &nend);
 		p->from.type = D_CONST;
-			p->from.offset = -4;
+		if(dir < 0)
+			p->from.offset = dir;
+		else
+			p->from.offset = w;
+	}
 
-			// move src and dest to the end of block
+	// move src and dest to the end of block if necessary
+	if(dir < 0) {
 		p = gins(AMOVW, &src, &src);
 		p->from.type = D_CONST;
-			p->from.offset = (q-1)*4;
+		p->from.offset = w + dir;
 
 		p = gins(AMOVW, &dst, &dst);
 		p->from.type = D_CONST;
-			p->from.offset = (q-1)*4;
+		p->from.offset = w + dir;
+	}
 	
-			p = gins(AMOVW, &src, &tmp);
+	// move
+	if(c >= 4) {
+		p = gins(op, &src, &tmp);
 		p->from.type = D_OREG;
-			p->from.offset = -4;
+		p->from.offset = dir;
 		p->scond |= C_PBIT;
 		ploop = p;
 
-			p = gins(AMOVW, &tmp, &dst);
+		p = gins(op, &tmp, &dst);
 		p->to.type = D_OREG;
-			p->to.offset = -4;
+		p->to.offset = dir;
 		p->scond |= C_PBIT;
 
 		p = gins(ACMP, &src, N);
 		raddr(&nend, p);
 
 		patch(gbranch(ABNE, T), ploop);
-
  		regfree(&nend);
 	} else {
-			// move src and dest to the end of block
+		while(c-- > 0) {
-			p = gins(AMOVW, &src, &src);
+			p = gins(op, &src, &tmp);
-			p->from.type = D_CONST;
-			p->from.offset = (q-1)*4;
-
-			p = gins(AMOVW, &dst, &dst);
-			p->from.type = D_CONST;
-			p->from.offset = (q-1)*4;
-
-			while(q > 0) {
-				p = gins(AMOVW, &src, &tmp);
 			p->from.type = D_OREG;
-				p->from.offset = -4;
+			p->from.offset = dir;
- 				p->scond |= C_PBIT;
-
-				p = gins(AMOVW, &tmp, &dst);
-				p->to.type = D_OREG;
-				p->to.offset = -4;
- 				p->scond |= C_PBIT;
-
-				q--;
-			}
-		}
-	} else {
-		// normal direction
-		if(q >= 4) {
-			regalloc(&nend, types[TUINT32], N);
-			p = gins(AMOVW, &src, &nend);
-			p->from.type = D_CONST;
-			p->from.offset = q*4;
-
-			p = gins(AMOVW, &src, &tmp);
-			p->from.type = D_OREG;
-			p->from.offset = 4;
 			p->scond |= C_PBIT;
 			ploop = p;
 	
-			p = gins(AMOVW, &tmp, &dst);
+			p = gins(op, &tmp, &dst);
 			p->to.type = D_OREG;
-			p->to.offset = 4;
+			p->to.offset = dir;
 			p->scond |= C_PBIT;
-
-			p = gins(ACMP, &src, N);
-			raddr(&nend, p);
-
-			patch(gbranch(ABNE, T), ploop);
-
- 			regfree(&nend);
-		} else
-		while(q > 0) {
-			p = gins(AMOVW, &src, &tmp);
-			p->from.type = D_OREG;
-			p->from.offset = 4;
- 			p->scond |= C_PBIT;
-
-			p = gins(AMOVW, &tmp, &dst);
-			p->to.type = D_OREG;
-			p->to.offset = 4;
- 			p->scond |= C_PBIT;
-
-			q--;
-		}
-
-		if (c != 0) {
-			//	MOVW	(src), tmp
-			p = gins(AMOVW, &src, &tmp);
-			p->from.type = D_OREG;
-
-			//	MOVW	tmp<<((4-c)*8),src
-			gshift(AMOVW, &tmp, SHIFT_LL, ((4-c)*8), &src);
-
-			//	MOVW	src>>((4-c)*8),src
-			gshift(AMOVW, &src, SHIFT_LR, ((4-c)*8), &src);
-
-			//	MOVW	(dst), tmp
-			p = gins(AMOVW, &dst, &tmp);
-			p->from.type = D_OREG;
-
-			//	MOVW	tmp>>(c*8),tmp
-			gshift(AMOVW, &tmp, SHIFT_LR, (c*8), &tmp);
-
-			//	MOVW	tmp<<(c*8),tmp
-			gshift(AMOVW, &tmp, SHIFT_LL, c*8, &tmp);
-
-			//	ORR		src, tmp
-			gins(AORR, &src, &tmp);
-
-			//	MOVW	tmp, (dst)
-			p = gins(AMOVW, &tmp, &dst);
-			p->to.type = D_OREG;
 		}
 	}
+
 	regfree(&dst);
 	regfree(&src);
 	regfree(&tmp);

src/cmd/gc/align.c

@@ -234,9 +234,11 @@ dowidth(Type *t)
 			if(t->bound > cap)
 				yyerror("type %lT larger than address space", t);
 			w = t->bound * t->type->width;
-			if(w == 0)
-				w = 1;
 			t->align = t->type->align;
+			if(w == 0) {
+				w = 1;
+				t->align = 1;
+			}
 		}
 		else if(t->bound == -1) {
 			w = sizeof_Array;
@@ -253,10 +255,10 @@ dowidth(Type *t)
 		if(t->funarg)
 			fatal("dowidth fn struct %T", t);
 		w = widstruct(t, 0, 1);
-		if(w == 0)
+		if(w == 0) {
 			w = 1;
-		//if(t->align < widthptr)
+			t->align = 1;
-		//	warn("align %d: %T\n", t->align, t);
+		}
 		break;
 
 	case TFUNC:

src/cmd/gc/gen.c

@@ -59,6 +59,8 @@ allocparams(void)
 			fatal("bad width");
 		stksize += w;
 		stksize = rnd(stksize, n->type->align);
+		if(thechar == '5')
+			stksize = rnd(stksize, widthptr);
 		n->xoffset = -stksize;
 	}
 	lineno = lno;
@@ -698,6 +700,8 @@ tempname(Node *n, Type *t)
 	w = t->width;
 	stksize += w;
 	stksize = rnd(stksize, t->align);
+	if(thechar == '5')
+		stksize = rnd(stksize, widthptr);
 	n->xoffset = -stksize;
 	n->pun = anyregalloc();
 }

src/cmd/gc/obj.c

@@ -284,6 +284,7 @@ stringsym(char *s, int len)
 		off = dsname(sym, off, s+n, m);
 	}
 	off = duint8(sym, off, 0);  // terminating NUL for runtime
+	off = (off+widthptr-1)&~(widthptr-1);  // round to pointer alignment
 	ggloblsym(sym, off, 1);
 	text();