ply/global.c

#define _GNU_SOURCE 		/* asprintf */
#include <assert.h>
#include <errno.h>
#include <limits.h>
#include <stdlib.h>
#include <string.h>

#include "arch.h"
#include "func.h"
#include "node.h"
#include "ply.h"
#include "sym.h"
#include "type.h"

/* . */

static int global_dot_ir_pre(const struct func *func, struct node *n,
				struct prog *prog)
{
	struct node *sou = n->expr.args;

	if (node_is(sou, ":deref")) {
		/* (*ptr).member, if *ptr is not already loaded let it
		 * know that we're only interested in one member */
		sou->sym->irs.hint.dot = 1;

		/* this also means we need to put ourselves on the
		 * stack since data will be loaded via probe_read */
		n->sym->irs.hint.stack = 1;
	}
	return 0;
}

static int global_dot_ir_post(const struct func *func, struct node *n,
				struct prog *prog)
{
	struct node *sou, *member;
	struct irstate *dst;
	ssize_t offset;

	sou = n->expr.args;
	member = sou->next;
	dst = &n->sym->irs;
	
	ir_init_sym(prog->ir, n->sym);

	offset = type_offsetof(type_base(sou->sym->type), member->string.data);
	assert(offset >= 0);

	if (!sou->sym->irs.loc) {
		/* sou is a :deref which wasn't loaded by child, just
		 * read the member we're interested in. */
		struct node *ptr = sou->expr.args;

		ir_emit_sym_to_reg(prog->ir, BPF_REG_3, ptr->sym);
		ir_emit_insn(prog->ir, ALU_IMM(BPF_ADD, offset), BPF_REG_3, 0);
		goto probe_read;
	}

	offset += sou->sym->irs.stack;

	if (dst->loc == LOC_REG) {
		ir_emit_insn(prog->ir, LDX(bpf_width(dst->size), offset),
			     dst->reg, BPF_REG_BP);
		return 0;
	}

	ir_emit_insn(prog->ir, ALU_IMM(BPF_ADD, offset), BPF_REG_3, 0);
probe_read:
	ir_emit_insn(prog->ir, MOV_IMM((int32_t)dst->size), BPF_REG_2, 0);
	ir_emit_insn(prog->ir, MOV, BPF_REG_1, BPF_REG_BP);
	ir_emit_insn(prog->ir, ALU_IMM(BPF_ADD, dst->stack), BPF_REG_1, 0);
	ir_emit_insn(prog->ir, CALL(BPF_FUNC_probe_read), 0, 0);
	/* TODO if (r0) exit(r0); */
	return 0;
}

static int global_dot_type_infer(const struct func *func, struct node *n)
{
	struct node *sou, *member;
	struct type *t;
	struct tfield *f;

	if (n->sym->type)
		return 0;

	sou = n->expr.args;
	member = sou->next;
	if (!sou->sym->type)
		return 0;

	t = type_base(sou->sym->type);

	/* TODO: add union */
	if (t->ttype != T_STRUCT) {
		_e("%#N: %N is neither struct nor union (type '%T').\n",
		   n, sou, sou->sym->type);
		return -EINVAL;
	}

	f = tfields_get(t->sou.fields, member->string.data);
	if (!f) {
		_e("%#N: type '%T' has no member named %N.\n", n, t, member);
		return -EINVAL;
	}

	/* given `sou.member` where sou is a struct/union, infer that
	 * the expression's type is equal to member's type. */
	n->sym->type = f->type;
	return 0;
}


/* :deref */

static int global_deref_ir_post(const struct func *func, struct node *n,
				struct prog *prog)
{
	struct node *ptr = n->expr.args;
	struct irstate *dst;
	size_t size;

	dst = &n->sym->irs;
	if (dst->hint.dot)
		/* (*ptr).member, ptr points to a struct and our
		 * parent is only interested in one member. don't load
		 * the struct, let the dot operaton steal the address
		 * from our argument */
		return 0;

	ir_init_sym(prog->ir, n->sym);

	if (dst->hint.lval)
		/* *ptr = val, whatever is in our storage now it will
                    be overwritten, so skip the load. */
		return 0;

	ir_emit_insn(prog->ir, MOV, BPF_REG_1, BPF_REG_BP);
	ir_emit_insn(prog->ir, ALU_IMM(BPF_ADD, dst->stack), BPF_REG_1, 0);
	ir_emit_insn(prog->ir, MOV_IMM((int32_t)dst->size), BPF_REG_2, 0);
	ir_emit_sym_to_reg(prog->ir, BPF_REG_3, ptr->sym);
	ir_emit_insn(prog->ir, CALL(BPF_FUNC_probe_read), 0, 0);
	/* TODO if (r0) exit(r0); */
	return 0;
}


static int global_deref_type_infer(const struct func *func, struct node *n)
{
	struct node *ptr = n->expr.args;
	struct type *t;

	if (n->sym->type || !ptr->sym->type)
		return 0;

	t = type_base(ptr->sym->type);
	if (t->ttype != T_POINTER) {
		_e("%#N: can't dereference %N (type '%T').\n",
		   n, ptr, ptr->sym->type);
		return -EINVAL;
	}

	/* given `*p` where p is a pointer, infer that the
	 * expression's type is equal to p's concrete type. */
	n->sym->type = t->ptr.type;
	return 0;
}


/* :map */

static int global_map_ir_pre_key(struct node *n, struct prog *prog)
{
	struct node *map = n->expr.args, *arg;
	struct type *ktype = type_base(map->sym->type->map.ktype);
	ssize_t stack = map->sym->irs.stack;
	size_t offset, size, pad;
	struct tfield *f;

	arg = map->next;
	tfields_foreach(f, ktype->sou.fields) {
		offset = type_offsetof(ktype, f->name);
		size = type_sizeof(f->type);

		if (!arg->sym->irs.loc) {
			arg->sym->irs.hint.stack = 1;
			arg->sym->irs.stack = stack + offset;
		}

		if (arg->next) {
			pad = type_offsetof(ktype, f[1].name) - (offset + size);
			if (pad)
				ir_emit_bzero(prog->ir,
					      stack + offset + size, pad);
		}
		arg = arg->next;
	}

	pad = type_sizeof(ktype) - (offset + size);
	if (pad)
		ir_emit_bzero(prog->ir, stack + offset + size, pad);
	return 0;
}

static int global_map_ir_pre(const struct func *func, struct node *n,
			     struct prog *prog)
{
	struct irstate *kirs;
	struct node *map = n->expr.args;
	struct type *ktype = type_base(map->sym->type->map.ktype);

	map->sym->irs.hint.stack = 1;
	ir_init_irs(prog->ir, &map->sym->irs, ktype);


	if (ktype->ttype == T_STRUCT)
		return global_map_ir_pre_key(n, prog);

	kirs = &map->next->sym->irs;
	if (!kirs->loc) {
		kirs->hint.stack = 1;
		kirs->stack = map->sym->irs.stack;
	}
	return 0;
}

static int global_map_ir_post(const struct func *func, struct node *n,
			      struct prog *prog)
{
	struct node *map = n->expr.args, *arg;
	struct type *ktype = type_base(map->sym->type->map.ktype);
	ssize_t stack = map->sym->irs.stack;
	size_t offset;
	struct tfield *f;

	arg = map->next;
	tfields_foreach(f, ktype->sou.fields) {
		offset = type_offsetof(ktype, f->name);
		ir_emit_sym_to_stack(prog->ir, stack + offset, arg->sym);
		arg = arg->next;
	}
	return 0;
}

static struct type *global_map_ktype(struct node *n)
{
	struct node *map, *key;
	struct type *ktype;
	struct tfield *kfields, *f;
	int i, nargs = node_nargs(n);
	char *kname;

	map = n->expr.args;

	if (nargs == 2)
		return map->next->sym->type;

	ktype = calloc(1, sizeof(*ktype));
	assert(ktype);

	kfields = calloc(nargs, sizeof(*kfields));
	assert(kfields);

	for (key = map->next, f = kfields, i = 0; key; key = key->next, f++, i++) {
		asprintf(&f->name, "k%d", i);
		f->type = key->sym->type;
	}

	asprintf(&ktype->sou.name, ":%s_key", map->ident.name);
	ktype->ttype = T_STRUCT;
	ktype->sou.fields = kfields;

	type_add(ktype);
	return ktype;
}

static int global_map_type_infer(const struct func *func, struct node *n)
{
	struct node *map = n->expr.args;

	if (!map->sym->type)
		return 0;

	/* TODO validate key against known type */

	/* given `m[key]` where m's type is known, infer that the
	 * expression's type is equal to m's value type. */
	n->sym->type = map->sym->type->map.vtype;
	return 0;
}

static int global_map_static_validate(const struct func *func, struct node *n)
{
	if (n->expr.args->ntype != N_IDENT) {
		_e("%#N: can't lookup a key in %N, which is not a map.\n",
		   n, n);
		return -EINVAL;
	}

	return 0;
}

/* :assign */

static int global_assign_type_infer_map(struct node *n)
{
	struct node *map, *key;
	struct type *ktype;

	map = n->expr.args;

	for (key = map->next; key; key = key->next) {
		if (type_sizeof(key->sym->type) < 0)
			return 0;
	}

	map->sym->type = type_map_of(global_map_ktype(n), n->sym->type);
	return 0;
}

static int global_assign_type_infer(const struct func *func, struct node *n)
{
	struct node *lval, *rval;

	if (n->sym->type)
		return 0;

	lval = n->expr.args;
	rval = lval->next;

	if (!rval->sym->type)
		return 0;

	if (!lval->sym->type) {
		/* given `a = b` where b's type is known but not a's,
		 * infer that a's type must be equal to b's */
		lval->sym->type = rval->sym->type;

		/* TODO do we need assignment expressions? */
		n->sym->type = &t_void;
		
		if (node_is(lval, "{}"))
			return global_assign_type_infer_map(lval);

		return 0;
	}

	if (type_compatible(lval->sym->type, rval->sym->type))
		return 0;

	_e("%#N: can't assign %N (type '%T'), to %N (type '%T').\n",
	   n, rval, rval->sym->type, lval, lval->sym->type);

	return -EINVAL;
}

static int global_assign_static_validate(const struct func *func, struct node *n)
{
	struct node *lval;

	lval = n->expr.args;

	if (node_is(lval, "{}") || (lval->ntype == N_IDENT))
		return 0;

	_e("%#N: can't assign a value to %N.\n", n, lval);
	return -EINVAL;
}

/* :binop */

static int global_binop_type_infer(const struct func *func, struct node *n)
{
	struct node *lval, *rval;

	if (n->sym->type)
		return 0;

	lval = n->expr.args;
	rval = lval->next;

	if (!lval->sym->type || !rval->sym->type)
		return 0;

	if (type_equal(lval->sym->type, rval->sym->type)) {
		n->sym->type = lval->sym->type;
		return 0;
	}

	/* TODO handle integer promotion */
	return 0;
}

/* :binop */

static int global_quantize_type_infer(const struct func *func, struct node *n)
{
	struct node *arg;
	struct type *t;

	arg = n->expr.args;

	if (n->sym->type || !arg->sym->type)
		return 0;

	t = type_base(arg->sym->type);
	if (t->ttype != T_SCALAR) {
		_e("%#N: can't quantize non-scalar value %N (type '%T').\n",
		   n, arg, arg->sym->type);
		return -EINVAL;	
	}

	n->sym->type = type_array_of(arg->sym->type, type_sizeof(t) * 8);
	return 0;
}

/* pid */

static int global_pid_ir_post(const struct func *func, struct node *n,
				struct prog *prog)
{
	struct node *ptr = n->expr.args;

	ir_init_sym(prog->ir, n->sym);

	ir_emit_insn(prog->ir, CALL(BPF_FUNC_get_current_pid_tgid), 0, 0);
	ir_emit_insn(prog->ir, ALU64_IMM(BPF_RSH, 32), BPF_REG_0, 0);
	ir_emit_reg_to_sym(prog->ir, n->sym, BPF_REG_0);
	return 0;
}

struct type t_pid = {
	.ttype = T_TYPEDEF,

	.tdef = { .name = ":pid", .type = &t_u32 },
};

struct type t_pid_func = {
	.ttype = T_FUNC,

	.func = { .type = &t_pid },
};

/* time */

static int global_time_ir_post(const struct func *func, struct node *n,
			       struct prog *prog)
{
	struct node *ptr = n->expr.args;

	ir_init_sym(prog->ir, n->sym);

	ir_emit_insn(prog->ir, CALL(BPF_FUNC_ktime_get_ns), 0, 0);
	ir_emit_reg_to_sym(prog->ir, n->sym, BPF_REG_0);
	return 0;
}

struct type t_time = {
	.ttype = T_TYPEDEF,	/* TODO: should be a T_FUNC with a static
				 * signature */

	.tdef = { .name = ":time", .type = &t_s64 },
};

struct type t_time_func = {
	.ttype = T_FUNC,

	.func = { .type = &t_time },
};


/*  */

struct type t_block_func = {
	.ttype = T_FUNC,

	.func = { .type = &t_void, .vargs = 1 },
};

struct type t_string_array = {
	.ttype = T_ARRAY,

	.array = { .type = &t_char, .len = 64 }, /* TODO: tunable */
};

struct type t_string = {
	.ttype = T_TYPEDEF,

	.tdef = { .name = ":string", .type = &t_string_array },
};

struct tfield f_dot[] = {
	{ .type = &t_void },
	{ .type = &t_string },

	{ .type = NULL }
};

struct type t_dot_func = {
	.ttype = T_FUNC,

	.func = { .type = &t_void, .args = f_dot },
};

struct tfield f_2args[] = {
	{ .type = &t_void },
	{ .type = &t_void },

	{ .type = NULL }
};

struct type t_2args_func = {
	.ttype = T_FUNC,

	.func = { .type = &t_void, .args = f_2args },
};

struct tfield f_1arg[] = {
	{ .type = &t_void },

	{ .type = NULL }
};

struct type t_1arg_func = {
	.ttype = T_FUNC,

	.func = { .type = &t_void, .args = f_1arg },
};

static const struct func global_funcs[] = {
	{
		.name = ":block",
		.type = &t_block_func,
		.static_ret = 1,
	},

	{
		.name = ".",
		.type = &t_dot_func,
		.type_infer = global_dot_type_infer,

		.ir_pre  = global_dot_ir_pre,
		.ir_post = global_dot_ir_post,
	},
	{
		.name = ":deref",
		.type = &t_1arg_func,
		.type_infer = global_deref_type_infer,

		.ir_post = global_deref_ir_post,
	},

	{
		.name = "+",
		.type = &t_2args_func,
		.type_infer = global_binop_type_infer,
	},
	{
		.name = "-",
		.type = &t_2args_func,
		.type_infer = global_binop_type_infer,
	},
	
	{
		.name = "=",
		.type = &t_2args_func,
		.type_infer = global_assign_type_infer,
		.static_validate = global_assign_static_validate,
	},
	{
		.name = "{}",
		/* .type = t_map_func, */
		.type_infer = global_map_type_infer,
		.static_validate = global_map_static_validate,

		.ir_pre  = global_map_ir_pre,
		.ir_post = global_map_ir_post,
	},

	{
		.name = "pid",
		.type = &t_pid_func,
		.static_ret = 1,

		.ir_post = global_pid_ir_post,
	},
	{
		.name = "time",
		.type = &t_time_func,
		.static_ret = 1,

		.ir_post = global_time_ir_post,
	},

	{
		.name = "quantize",
		.type = &t_1arg_func,
		.type_infer = global_quantize_type_infer,
	},
	
	{ .name = NULL }
};

static struct type *global_num_type(struct node *n)
{
	if (n->num.unsignd) {
		if (n->num.u64 <= INT_MAX)
			return &t_int;
		else if (n->num.u64 <= UINT_MAX)
			return &t_uint;
		else if (n->num.u64 <= LONG_MAX)
			return &t_long;
		else if (n->num.u64 <= ULONG_MAX)
			return &t_ulong;
		else if (n->num.u64 <= LLONG_MAX)
			return &t_llong;
		else if (n->num.u64 <= ULLONG_MAX)
			return &t_ullong;
	} else {
		if (n->num.s64 >= INT_MIN && n->num.s64 <= INT_MAX)
			return &t_int;
		else if (n->num.s64 >= LONG_MIN && n->num.s64 <= LONG_MAX)
			return &t_long;
		else if (n->num.s64 >= LLONG_MIN && n->num.s64 <= LLONG_MAX)
			return &t_llong;
	}

	assert(0);
	return NULL;
}

static int global_num_ir_post(const struct func *func, struct node *n,
			      struct prog *prog)
{
	struct irstate *irs = &n->sym->irs;

	if ((n->num.unsignd && (n->num.u64 <= INT32_MAX)) ||
	    (n->num.s64 >= INT32_MIN && n->num.s64 <= INT32_MAX)) {
		irs->loc = LOC_IMM;
		irs->imm = n->num.s64;
		irs->size = 4;
		return 0;
	}

	/* we need to build the constant in a register, so ignore any
	 * advise about stack allocation. */
	irs->hint.stack = 0;

	ir_init_sym(prog->ir, n->sym);

	if (n->num.u64 > 0x3fffffffffffffff) {
		ir_emit_insn(prog->ir, MOV64_IMM(n->num.u64 >> 33), irs->reg, 0);
		ir_emit_insn(prog->ir, ALU64_IMM(BPF_LSH, 31), irs->reg, 0);

		if ((n->num.u64 >> 2) & 0x7fffffff)
			ir_emit_insn(prog->ir,
				     ALU64_IMM(BPF_OR, (n->num.u64 >> 2) & 0x7fffffff),
				     irs->reg, 0);
		ir_emit_insn(prog->ir, ALU64_IMM(BPF_LSH, 2), irs->reg, 0);

		if (n->num.u64 & 0x3)
			ir_emit_insn(prog->ir, ALU64_IMM(BPF_OR, n->num.u64 & 0x3),
				     irs->reg, 0);
	} else if (n->num.u64 > 0x7fffffff) {
		ir_emit_insn(prog->ir, MOV64_IMM(n->num.u64 >> 31), irs->reg, 0);
		ir_emit_insn(prog->ir, ALU64_IMM(BPF_LSH, 31), irs->reg, 0);

		if (n->num.u64 & 0x7fffffff)
			ir_emit_insn(prog->ir,
				     ALU64_IMM(BPF_OR, n->num.u64 & 0x7fffffff),
				     irs->reg, 0);
	}
	return 0;
}

static const struct func global_num_func = {
	.name = ":num",

	.ir_post = global_num_ir_post,
};

static const struct func global_string_func = {
	.name = ":string",
	.type = &t_string,
	.static_ret = 1,
};

static const struct func global_ident_func = {
	.name = ":ident",
};

static const struct func *global_sym_alloc_expr(struct node *n)
{
	const struct func *func;
	int err;

	for (func = global_funcs; func->name; func++) {
		if (strcmp(func->name, n->expr.func))
			continue;

		return func;
	}

	return NULL;
}

int global_sym_alloc(struct prog *prog, struct node *n)
{
	const struct func *func;
	struct symtab *st = prog->locals;
	int err;

	switch (n->ntype) {
	case N_EXPR:
		func = global_sym_alloc_expr(n);
		break;
	case N_IDENT:
		st = prog->globals;
		func = &global_ident_func;
		break;
	case N_NUM:
		func = &global_num_func;
		break;
	case N_STRING:
		func = &global_string_func;
		break;
	}

	if (!func)
		return -ENOENT;

	err = func_static_validate(func, n);
	if (err)
		return err;

	n->sym = sym_alloc(st, n, func);

	if (n->ntype == N_NUM)
		n->sym->type = global_num_type(n);
	else if (func->static_ret)
		n->sym->type = func_return_type(func);
	return 0;
}

int global_probe(struct prog *prog)
{
	return 0;
}

struct provider global = {
	.name = ":",

	.sym_alloc = global_sym_alloc,
	.probe = global_probe,
};

__attribute__((constructor))
static void global_init(void)
{
	provider_register(&global);
}