w6c: keep nested unwrap destinations distinct

cmd/w6c/cgen.c

@@ -39,11 +39,6 @@ static int *cg_frame;
  * semantics for synthetic scratches). 0 means "not yet allocated";
  * negative offsets returned by local_alloc are the live value. */
 static int cg_retscr;
-/* The aggregate-unwrap receiver must survive its producer call without
- * perturbing SP: an outstanding PUSHQ at CALL violates SysV alignment.
- * One BP-relative address spill per function mirrors wwstage's @-prefix
- * scratch reuse. */
-static int cg_unwrapdst;
 /* Per-fn @tupfscr offset (single-slot SSoT). A multi-float tuple return
  * (#164/#107) spills each float out of X0 to this scratch as the L→R
  * element walk clobbers X0, then reloads X0/X1 by SSE index after the
@@ -5367,14 +5362,12 @@ cgexpr(Cg *c, Node *n, Local *locals)
 					fatal("#16: global/chained aggregate "
 					    "unwrap field dest unresolved "
 					    "(cgplaceaddr)");
-				int dstscr;
+				/* Fresh per lowering: cgexpr(rhs) may recursively
-				if (cg_unwrapdst != 0) {
+				 * emit another #16 assignment while this address is
-					dstscr = cg_unwrapdst;
+				 * live. Sharing one per-fn slot would let the inner
-				} else {
+				 * assignment overwrite the outer receiver. */
-					dstscr = local_alloc(c, &locals, "@unwrapdst",
+				int dstscr = local_alloc(c, &locals, "@unwrapdst",
-					    8, cg_frame);
+				    8, cg_frame);
-					cg_unwrapdst = dstscr;
-				}
 				ins2(c, A_MOVQ, areg(D_BX), amem(D_BP, dstscr));
 				cgexpr(c, n->rhs, locals);
 				ins2(c, A_MOVQ, amem(D_BP, dstscr), areg(D_BX));
@@ -15803,7 +15796,6 @@ cgfn(Cg *c, FILE *out, Node *fn)
 	nloops = 0;
 	cg_ret_type = fn->type ? fn->type->ret : NULL;
 	cg_retscr = 0;
-	cg_unwrapdst = 0;
 	cg_tupfscr = 0;
 	cg_tupargscr = 0;
 	cg_tupargscr_sz = 0;

selfhost/cmd/wcc/cgenexpr.ww

@@ -8780,7 +8780,11 @@ fn cgassign(c: *cgen, n: *syntax.node) void = {
 						os.write(2, m16.ptr, m16.len: u64);
 						os.exit(1);
 					};
-					let dstscr: i32 = localadd(c, "@unwrapdst", 8, nil);
+					// Fresh per lowering: cgexpr(rhs) may recursively
+					// emit another #16 assignment while this address is
+					// live. Sharing one per-fn slot would let the inner
+					// assignment overwrite the outer receiver.
+					let dstscr: i32 = localalloc(c, "@unwrapdst", 8, nil);
 					emitline("\tMOVQ\tBX, ");
 					emitoff(dstscr: i64);
 					emitline("(BP)\n");

test/asm/chain_test.ww

@@ -299,6 +299,36 @@ fn unwrapaligncheck(label: str, stage: str, s: str) void = {
 	};
 };
 
+fn unwrapnestedcheck(label: str, stage: str, s: str) void = {
+	let w: str = probewindow(label, stage, s);
+	if (!testenv.has(w, "TEXT main.probe,$48")) {
+		fail(label, strings.concat(stage,
+			": nested spill frame is not $48"));
+	};
+	if (testenv.has(w, "\tPUSHQ\tBX\n")) {
+		fail(label, strings.concat(stage,
+			": destination PUSHQ remains in nested window"));
+	};
+	let osave: i32 = testenv.pos(w, "\tMOVQ\tBX, -8(BP)\n");
+	let choice: i32 = posafter(w, osave + 1,
+		"\tCALL\tmain.choice(SB)\n");
+	let isave: i32 = posafter(w, choice + 1,
+		"\tMOVQ\tBX, -48(BP)\n");
+	let make: i32 = posafter(w, isave + 1,
+		"\tCALL\tmain.make(SB)\n");
+	let iload: i32 = posafter(w, make + 1,
+		"\tMOVQ\t-48(BP), BX\n");
+	let pass: i32 = posafter(w, iload + 1,
+		"\tCALL\tmain.pass(SB)\n");
+	let oload: i32 = posafter(w, pass + 1,
+		"\tMOVQ\t-8(BP), BX\n");
+	if (osave < 0 || choice < 0 || isave < 0 || make < 0
+	    || iload < 0 || pass < 0 || oload < 0) {
+		fail(label, strings.concat(stage,
+			": nested destination spills are not distinct and scoped"));
+	};
+};
+
 fn commandexit(dir: str, name: str, av: []str) i32 = {
 	let co: testenv.commandout;
 	testenv.runcommand(dir, dir, name, av, tmo(), &co);
@@ -349,3 +379,65 @@ fn commandexit(dir: str, name: str, av: []str) i32 = {
 	};
 	testenv.clean(td);
 };
+
+// The outer destination stays live while its call argument lowers a
+// match arm containing another #16 assignment. A per-fn shared spill
+// redirected both stores to inner.f and left outer.f unchanged.
+@test fn unwrapnestedlive() void = {
+	let label: str = "nested_global_unwrap_destination";
+	let td: str = testenv.fresh();
+	let src: str = strings.concat(
+		"package main;\n",
+		"type e = !i32;\n",
+		"type pair = struct { a: i64, b: i64 };\n",
+		"type box = struct { f: pair, guard: i64 };\n",
+		"let outer: box = box {\n",
+		"    f = pair { a = 1i64, b = 2i64 }, guard = 7i64\n",
+		"};\n",
+		"let inner: box = box {\n",
+		"    f = pair { a = 3i64, b = 4i64 }, guard = 8i64\n",
+		"};\n",
+		"fn make() (pair | e) = {\n",
+		"    return pair { a = 111i64, b = 222i64 };\n",
+		"};\n",
+		"fn choice() (i32 | str) = { return 1i32; };\n",
+		"fn pass(x: i64) (pair | e) = {\n",
+		"    return pair { a = 111i64 + x, b = 222i64 };\n",
+		"};\n",
+		"fn probe() i32 = {\n",
+		"    outer.f = pass(match (choice()) {\n",
+		"    case i32 => {\n",
+		"        inner.f = make()!;\n",
+		"        yield 0i64;\n",
+		"    };\n",
+		"    case str => yield 0i64;\n",
+		"    })!;\n",
+		"    if (outer.f.a != 111i64 || outer.f.b != 222i64",
+		" || outer.guard != 7i64) { return 91; };\n",
+		"    if (inner.f.a != 111i64 || inner.f.b != 222i64",
+		" || inner.guard != 8i64) { return 92; };\n",
+		"    return 37;\n",
+		"};\n",
+		"export fn main() i32 = { return probe(); };\n");
+	testenv.writefile(strings.concat(td, "/src.ww"), src);
+	emitstage(td, label, "cstage", testenv.driver("w6c"), "cs.s");
+	emitstage(td, label, "wwstage", testenv.driver("w6c_ww"), "ws.s");
+	let cs: str = testenv.readfile(strings.concat(td, "/cs.s"));
+	let ws: str = testenv.readfile(strings.concat(td, "/ws.s"));
+	unwrapnestedcheck(label, "cstage", cs);
+	unwrapnestedcheck(label, "wwstage", ws);
+	if (!testenv.same(cs, ws)) {
+		fail(label, "cstage vs wwstage asm differs");
+	};
+	let out: str = strings.concat(td, "/unwrapnested");
+	let bav: []str = [testenv.driver("ww"), "build", "-o", out,
+		"src.ww"];
+	if (commandexit(td, "build", bav) != 0) {
+		fail(label, "nested runtime pin build failed");
+	};
+	let rav: []str = [out];
+	if (commandexit(td, "run", rav) != 37) {
+		fail(label, "nested payload/neighbour runtime exit != 37");
+	};
+	testenv.clean(td);
+};