tinycc/riscv64-gen.c

1147 lines
37 KiB
C
Raw Normal View History

2019-06-22 12:13:10 +08:00
#ifdef TARGET_DEFS_ONLY
// Number of registers available to allocator:
#define NB_REGS 19 // x10-x17 aka a0-a7, f10-f17 aka fa0-fa7, xxx, ra, sp
2019-06-22 12:13:10 +08:00
#define TREG_R(x) (x) // x = 0..7
#define TREG_F(x) (x + 8) // x = 0..7
2019-06-22 12:13:10 +08:00
// Register classes sorted from more general to more precise:
#define RC_INT (1 << 0)
#define RC_FLOAT (1 << 1)
#define RC_R(x) (1 << (2 + (x))) // x = 0..7
#define RC_F(x) (1 << (10 + (x))) // x = 0..7
#define RC_IRET (RC_R(0)) // int return register class
#define RC_FRET (RC_F(0)) // float return register class
#define REG_IRET (TREG_R(0)) // int return register number
#define REG_FRET (TREG_F(0)) // float return register number
#define PTR_SIZE 8
#define LDOUBLE_SIZE 16
#define LDOUBLE_ALIGN 16
#define MAX_ALIGN 16
#define CHAR_IS_UNSIGNED
#else
#include "tcc.h"
#include <assert.h>
#define XLEN 8
2019-06-22 12:13:10 +08:00
#define TREG_RA 17
#define TREG_SP 18
2019-06-22 12:13:10 +08:00
ST_DATA const int reg_classes[NB_REGS] = {
RC_INT | RC_R(0),
RC_INT | RC_R(1),
RC_INT | RC_R(2),
RC_INT | RC_R(3),
RC_INT | RC_R(4),
RC_INT | RC_R(5),
RC_INT | RC_R(6),
RC_INT | RC_R(7),
RC_FLOAT | RC_F(0),
RC_FLOAT | RC_F(1),
RC_FLOAT | RC_F(2),
RC_FLOAT | RC_F(3),
RC_FLOAT | RC_F(4),
RC_FLOAT | RC_F(5),
RC_FLOAT | RC_F(6),
RC_FLOAT | RC_F(7),
0,
1 << TREG_RA,
1 << TREG_SP
2019-06-22 12:13:10 +08:00
};
static int ireg(int r)
{
if (r == TREG_RA)
return 1; // ra
if (r == TREG_SP)
return 2; // sp
assert(r >= 0 && r < 8);
return r + 10; // tccrX --> aX == x(10+X)
}
static int is_ireg(int r)
{
return r < 8 || r == TREG_RA || r == TREG_SP;
}
static int freg(int r)
{
assert(r >= 8 && r < 16);
return r - 8 + 10; // tccfX --> faX == f(10+X)
}
static int is_freg(int r)
{
return r >= 8 && r < 16;
}
ST_FUNC void o(unsigned int c)
{
int ind1 = ind + 4;
if (nocode_wanted)
return;
if (ind1 > cur_text_section->data_allocated)
section_realloc(cur_text_section, ind1);
write32le(cur_text_section->data + ind, c);
ind = ind1;
}
static void EIu(uint32_t opcode, uint32_t func3,
uint32_t rd, uint32_t rs1, uint32_t imm)
{
o(opcode | (func3 << 12) | (rd << 7) | (rs1 << 15) | (imm << 20));
}
static void EI(uint32_t opcode, uint32_t func3,
uint32_t rd, uint32_t rs1, uint32_t imm)
{
assert(! ((imm + (1 << 11)) >> 12));
EIu(opcode, func3, rd, rs1, imm);
}
static void ES(uint32_t opcode, uint32_t func3,
uint32_t rs1, uint32_t rs2, uint32_t imm)
{
assert(! ((imm + (1 << 11)) >> 12));
o(opcode | (func3 << 12) | ((imm & 0x1f) << 7) | (rs1 << 15)
| (rs2 << 20) | ((imm >> 5) << 25));
}
2019-06-22 12:13:10 +08:00
// Patch all branches in list pointed to by t to branch to a:
ST_FUNC void gsym_addr(int t_, int a_)
{
uint32_t t = t_;
uint32_t a = a_;
while (t) {
unsigned char *ptr = cur_text_section->data + t;
uint32_t next = read32le(ptr);
uint32_t r = a - t, imm;
if ((r + (1 << 21)) & ~((1U << 22) - 2))
tcc_error("out-of-range branch chain");
imm = (((r >> 12) & 0xff) << 12)
| (((r >> 11) & 1) << 20)
| (((r >> 1) & 0x3ff) << 21)
| (((r >> 20) & 1) << 31);
write32le(ptr, r == 4 ? 0x33 : 0x6f | imm); // nop || j imm
2019-06-22 12:13:10 +08:00
t = next;
}
}
ST_FUNC void load(int r, SValue *sv)
{
int fr = sv->r;
int v = fr & VT_VALMASK;
int rr = is_ireg(r) ? ireg(r) : freg(r);
int fc = sv->c.i;
int bt = sv->type.t & VT_BTYPE;
int align, size = type_size(&sv->type, &align);
if (fr & VT_LVAL) {
int func3, opcode = 0x03, doload = 0;
if (is_freg(r)) {
assert(bt == VT_DOUBLE || bt == VT_FLOAT);
opcode = 0x07;
func3 = bt == VT_DOUBLE ? 3 : 2;
} else {
assert(is_ireg(r));
if (bt == VT_FUNC)
size = PTR_SIZE;
func3 = size == 1 ? 0 : size == 2 ? 1 : size == 4 ? 2 : 3;
if (size < 4 && !is_float(sv->type.t) && (sv->type.t & VT_UNSIGNED))
func3 |= 4;
}
if (v == VT_LOCAL) {
int br = 8; // s0
if (fc != sv->c.i)
tcc_error("unimp: load1(giant local ofs) (0x%llx)", (long long)sv->c.i);
if (((unsigned)fc + (1 << 11)) >> 12) {
br = is_ireg(r) ? rr : 5;
o(0x37 | (br << 7) | ((0x800 + fc) & 0xfffff000)); //lui BR, upper(fc)
o(0x33 | (br << 7) | (br << 15) | (8 << 20)); // add BR, BR, s0
fc = fc << 20 >> 20;
}
EI(opcode, func3, rr, br, fc); // l[bhwd][u]/fl[wd] RR, fc(BR)
} else if (v < VT_CONST) {
/*if (((unsigned)fc + (1 << 11)) >> 12)
tcc_error("unimp: load(large addend) (0x%x)", fc);*/
fc = 0; // XXX store ofs in LVAL(reg)
EI(opcode, func3, rr, ireg(v), fc); // l[bhwd][u] RR, 0(V)
} else if (v == VT_CONST && (fr & VT_SYM)) {
static Sym label;
int tempr;
if (sv->sym->type.t & VT_STATIC) { // XXX do this per linker relax
greloca(cur_text_section, sv->sym, ind,
R_RISCV_PCREL_HI20, sv->c.i);
fc = 0;
sv->c.i = 0;
} else {
if (((unsigned)fc + (1 << 11)) >> 12)
tcc_error("unimp: large addend for global address (0x%llx)", sv->c.i);
greloca(cur_text_section, sv->sym, ind,
R_RISCV_GOT_HI20, 0);
doload = 1;
}
if (!label.v) {
label.v = tok_alloc(".L0 ", 4)->tok;
label.type.t = VT_VOID | VT_STATIC;
}
label.c = 0; /* force new local ELF symbol */
put_extern_sym(&label, cur_text_section, ind, 0);
tempr = is_ireg(r) ? rr : 5;
o(0x17 | (tempr << 7)); // auipc TR, 0 %pcrel_hi(sym)+addend
greloca(cur_text_section, &label, ind,
R_RISCV_PCREL_LO12_I, 0);
if (doload) {
EI(0x03, 3, tempr, tempr, 0); // ld TR, 0(TR)
if (fc)
EI(0x13, 0, tempr, tempr, fc << 20 >> 20); // addi TR, TR, FC
fc = 0;
}
EI(opcode, func3, rr, tempr, fc); // l[bhwd][u] RR, fc(TR)
} else if (v == VT_LLOCAL) {
int br = 8, tempr = is_ireg(r) ? rr : 5;
if (fc != sv->c.i)
tcc_error("unimp: load2(giant local ofs) (0x%llx)", (long long)sv->c.i);
if (((unsigned)fc + (1 << 11)) >> 12) {
br = tempr;
o(0x37 | (br << 7) | ((0x800 + fc) & 0xfffff000)); //lui BR, upper(fc)
o(0x33 | (br << 7) | (br << 15) | (8 << 20)); // add BR, BR, s0
fc = fc << 20 >> 20;
}
EI(0x03, 3, tempr, br, fc); // ld TEMPR, fc(BR)
EI(opcode, func3, rr, tempr, 0); // l[bhwd][u] RR, 0(TEMPR)
} else {
tcc_error("unimp: load(non-local lval)");
}
} else if (v == VT_CONST) {
int rb = 0, do32bit = 8, doload = 0, zext = 0;
assert((!is_float(sv->type.t) && is_ireg(r)) || bt == VT_LDOUBLE);
if (fr & VT_SYM) {
static Sym label;
if (sv->sym->type.t & VT_STATIC) { // XXX do this per linker relax
greloca(cur_text_section, sv->sym, ind,
R_RISCV_PCREL_HI20, sv->c.i);
fc = 0;
sv->c.i = 0;
} else {
if (((unsigned)fc + (1 << 11)) >> 12)
tcc_error("unimp: large addend for global address (0x%llx)", sv->c.i);
greloca(cur_text_section, sv->sym, ind,
R_RISCV_GOT_HI20, 0);
doload = 1;
}
if (!label.v) {
label.v = tok_alloc(".L0 ", 4)->tok;
label.type.t = VT_VOID | VT_STATIC;
}
label.c = 0; /* force new local ELF symbol */
put_extern_sym(&label, cur_text_section, ind, 0);
o(0x17 | (rr << 7)); // auipc RR, 0 %call(func)
greloca(cur_text_section, &label, ind,
R_RISCV_PCREL_LO12_I, 0);
rb = rr;
do32bit = 0;
}
if (is_float(sv->type.t) && bt != VT_LDOUBLE)
tcc_error("unimp: load(float)");
if (fc != sv->c.i) {
int64_t si = sv->c.i;
uint32_t pi;
si >>= 32;
if (si != 0) {
pi = si;
if (fc < 0)
pi++;
o(0x37 | (rr << 7) | (((pi + 0x800) & 0xfffff000))); // lui RR, up(up(fc))
EI(0x13, 0, rr, rr, (int)pi << 20 >> 20); // addi RR, RR, lo(up(fc))
EI(0x13, 1, rr, rr, 12); // slli RR, RR, 12
EI(0x13, 0, rr, rr, (fc + (1 << 19)) >> 20); // addi RR, RR, up(lo(fc))
EI(0x13, 1, rr, rr, 12); // slli RR, RR, 12
fc = fc << 12 >> 12;
EI(0x13, 0, rr, rr, fc >> 8); // addi RR, RR, lo1(lo(fc))
EI(0x13, 1, rr, rr, 8); // slli RR, RR, 8
fc &= 0xff;
rb = rr;
do32bit = 0;
} else if (bt == VT_LLONG) {
/* A 32bit unsigned constant for a 64bit type.
lui always sign extends, so we need to do an explicit zext.*/
zext = 1;
}
}
if (((unsigned)fc + (1 << 11)) >> 12)
o(0x37 | (rr << 7) | ((0x800 + fc) & 0xfffff000)), rb = rr; //lui RR, upper(fc)
if (doload) {
EI(0x03, 3, rr, rr, 0); // ld RR, 0(RR)
if (fc)
EI(0x13 | do32bit, 0, rr, rr, fc << 20 >> 20); // addi[w] R, x0|R, FC
} else
EI(0x13 | do32bit, 0, rr, rb, fc << 20 >> 20); // addi[w] R, x0|R, FC
if (zext) {
EI(0x13, 1, rr, rr, 32); // slli RR, RR, 32
EI(0x13, 5, rr, rr, 32); // srli RR, RR, 32
}
} else if (v == VT_LOCAL) {
int br = 8; // s0
assert(is_ireg(r));
if (fc != sv->c.i)
tcc_error("unimp: load(addr giant local ofs) (0xll%x)", (long long)sv->c.i);
if (((unsigned)fc + (1 << 11)) >> 12) {
o(0x37 | (rr << 7) | ((0x800 + fc) & 0xfffff000)); //lui RR, upper(fc)
o(0x33 | (rr << 7) | (rr << 15) | (8 << 20)); // add RR, RR, s0
fc = fc << 20 >> 20;
br = rr;
}
EI(0x13, 0, rr, br, fc); // addi R, s0, FC
} else if (v < VT_CONST) {
/* reg-reg */
//assert(!fc); XXX support offseted regs
if (is_freg(r) && is_freg(v))
o(0x53 | (rr << 7) | (freg(v) << 15) | (freg(v) << 20) | ((bt == VT_DOUBLE ? 0x11 : 0x10) << 25)); //fsgnj.[sd] RR, V, V == fmv.[sd] RR, V
else if (is_ireg(r) && is_ireg(v))
EI(0x13, 0, rr, ireg(v), 0); // addi RR, V, 0 == mv RR, V
else {
int func7 = is_ireg(r) ? 0x70 : 0x78;
if (size == 8)
func7 |= 1;
assert(size == 4 || size == 8);
o(0x53 | (rr << 7) | ((is_freg(v) ? freg(v) : ireg(v)) << 15)
| (func7 << 25)); // fmv.{w.x, x.w, d.x, x.d} RR, VR
}
} else if (v == VT_CMP) { // we rely on cmp_r to be the correct result
EI(0x13, 0, rr, vtop->cmp_r, 0); // mv RR, CMP_R
} else if ((v & ~1) == VT_JMP) {
int t = v & 1;
assert(is_ireg(r));
EI(0x13, 0, rr, 0, t); // addi RR, x0, t
gjmp_addr(ind + 8);
gsym(fc);
EI(0x13, 0, rr, 0, t ^ 1); // addi RR, x0, !t
} else
tcc_error("unimp: load(non-const)");
2019-06-22 12:13:10 +08:00
}
ST_FUNC void store(int r, SValue *sv)
{
int fr = sv->r & VT_VALMASK;
int rr = is_ireg(r) ? ireg(r) : freg(r);
int fc = sv->c.i;
int ft = sv->type.t;
int bt = ft & VT_BTYPE;
int align, size = type_size(&sv->type, &align);
assert(!is_float(bt) || is_freg(r) || bt == VT_LDOUBLE);
/* long doubles are in two integer registers, but the load/store
primitives only deal with one, so do as if it's one reg. */
if (bt == VT_LDOUBLE)
size = align = 8;
if (bt == VT_STRUCT)
tcc_error("unimp: store(struct)");
if (size > 8)
tcc_error("unimp: large sized store");
assert(sv->r & VT_LVAL);
if (fr == VT_LOCAL) {
int br = 8; // s0
if (fc != sv->c.i)
tcc_error("unimp: store(giant local off) (0x%llx)", (long long)sv->c.i);
if (((unsigned)fc + (1 << 11)) >> 12) {
br = 5; // t0
o(0x37 | (br << 7) | ((0x800 + fc) & 0xfffff000)); //lui BR, upper(fc)
o(0x33 | (br << 7) | (br << 15) | (8 << 20)); // add BR, BR, s0
fc = fc << 20 >> 20;
}
if (is_freg(r))
ES(0x27, size == 4 ? 2 : 3, br, rr, fc); // fs[wd] RR, fc(base)
else
ES(0x23, size == 1 ? 0 : size == 2 ? 1 : size == 4 ? 2 : 3,
br, rr, fc); // s[bhwd] RR, fc(base)
} else if (fr < VT_CONST) {
int ptrreg = ireg(fr);
/*if (((unsigned)fc + (1 << 11)) >> 12)
tcc_error("unimp: store(large addend) (0x%x)", fc);*/
fc = 0; // XXX support offsets regs
if (is_freg(r))
ES(0x27, size == 4 ? 2 : 3, ptrreg, rr, fc); // fs[wd] RR, fc(PTRREG)
else
ES(0x23, size == 1 ? 0 : size == 2 ? 1 : size == 4 ? 2 : 3,
ptrreg, rr, fc); // s[bhwd] RR, fc(PTRREG)
} else if ((sv->r & ~VT_LVAL_TYPE) == (VT_CONST | VT_SYM | VT_LVAL)) {
static Sym label;
int tempr, doload = 0;
tempr = 5; // t0
if (sv->sym->type.t & VT_STATIC) { // XXX do this per linker relax
greloca(cur_text_section, sv->sym, ind,
R_RISCV_PCREL_HI20, sv->c.i);
fc = 0;
sv->c.i = 0;
} else {
if (((unsigned)fc + (1 << 11)) >> 12)
tcc_error("unimp: large addend for global address (0x%llx)", sv->c.i);
greloca(cur_text_section, sv->sym, ind,
R_RISCV_GOT_HI20, 0);
doload = 1;
}
if (!label.v) {
label.v = tok_alloc(".L0 ", 4)->tok;
label.type.t = VT_VOID | VT_STATIC;
}
label.c = 0; /* force new local ELF symbol */
put_extern_sym(&label, cur_text_section, ind, 0);
o(0x17 | (tempr << 7)); // auipc TEMPR, 0 %pcrel_hi(sym)+addend
greloca(cur_text_section, &label, ind,
doload ? R_RISCV_PCREL_LO12_I : R_RISCV_PCREL_LO12_S, 0);
if (doload) {
EI(0x03, 3, tempr, tempr, 0); // ld TR, 0(TR)
if (fc)
EI(0x13, 0, tempr, tempr, fc << 20 >> 20); // addi TR, TR, FC
fc = 0;
}
if (is_freg(r))
ES(0x27, size == 4 ? 2 : 3, tempr, rr, fc); // fs[wd] RR, fc(TEMPR)
else
ES(0x23, size == 1 ? 0 : size == 2 ? 1 : size == 4 ? 2 : 3,
tempr, rr, fc); // s[bhwd] RR, fc(TEMPR)
} else
tcc_error("implement me: %s(!local)", __FUNCTION__);
2019-06-22 12:13:10 +08:00
}
static void gcall_or_jmp(int docall)
{
int tr = docall ? 1 : 5; // ra or t0
if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST &&
((vtop->r & VT_SYM) && vtop->c.i == (int)vtop->c.i)) {
/* constant symbolic case -> simple relocation */
greloca(cur_text_section, vtop->sym, ind,
R_RISCV_CALL_PLT, (int)vtop->c.i);
o(0x17 | (tr << 7)); // auipc TR, 0 %call(func)
EI(0x67, 0, tr, tr, 0);// jalr TR, r(TR)
} else if (vtop->r < VT_CONST) {
int r = ireg(vtop->r);
EI(0x67, 0, tr, r, 0); // jalr TR, 0(R)
} else {
int r = TREG_RA;
load(r, vtop);
r = ireg(r);
EI(0x67, 0, tr, r, 0); // jalr TR, 0(R)
}
}
2019-06-22 12:13:10 +08:00
ST_FUNC void gfunc_call(int nb_args)
{
int i, align, size, aireg, afreg;
int info[nb_args ? nb_args : 1];
int stack_adj = 0, tempspace = 0, ofs, splitofs = 0;
int force_stack = 0;
SValue *sv;
Sym *sa;
aireg = afreg = 0;
sa = vtop[-nb_args].type.ref->next;
for (i = 0; i < nb_args; i++) {
int *pareg, nregs, infreg = 0, byref = 0, tempofs;
sv = &vtop[1 + i - nb_args];
sv->type.t &= ~VT_ARRAY; // XXX this should be done in tccgen.c
size = type_size(&sv->type, &align);
if (size > 16) {
if (align < XLEN)
align = XLEN;
tempspace = (tempspace + align - 1) & -align;
tempofs = tempspace;
tempspace += size;
size = align = 8;
byref = 1;
}
if (size > 8)
nregs = 2;
else
nregs = 1;
if ((sv->type.t & VT_BTYPE) == VT_LDOUBLE) {
infreg = 0;
} else
infreg = sa && is_float(sv->type.t);
if (!infreg && !sa && align == 2*XLEN && size <= 2*XLEN)
aireg = (aireg + 1) & ~1;
pareg = infreg ? &afreg : &aireg;
if ((*pareg < 8) && !force_stack) {
info[i] = *pareg + (infreg ? 8 : 0);
(*pareg)++;
if (nregs == 1)
;
else if (*pareg < 8)
(*pareg)++;
else {
info[i] |= 16;
stack_adj += 8;
}
} else {
info[i] = 32;
if (align < XLEN)
align = XLEN;
stack_adj += (size + align - 1) & -align;
if (!sa)
force_stack = 1;
}
if (byref)
info[i] |= 64 | (tempofs << 7);
if (sa)
sa = sa->next;
}
stack_adj = (stack_adj + 15) & -16;
tempspace = (tempspace + 15) & -16;
if (stack_adj + tempspace) {
EI(0x13, 0, 2, 2, -(stack_adj + tempspace)); // addi sp, sp, -adj
for (i = ofs = 0; i < nb_args; i++) {
if (info[i] >= 32) {
vrotb(nb_args - i);
size = type_size(&vtop->type, &align);
if (info[i] & 64) {
vset(&char_pointer_type, TREG_SP, 0);
vpushi(stack_adj + (info[i] >> 7));
gen_op('+');
vpushv(vtop); // this replaces the old argument
vrott(3);
indir();
vtop->type = vtop[-1].type;
vswap();
vstore();
vpop();
size = align = 8;
}
if (info[i] & 32) {
if (align < XLEN)
align = XLEN;
/* Once we support offseted regs we can do this:
vset(&vtop->type, TREG_SP | VT_LVAL, ofs);
to construct the lvalue for the outgoing stack slot,
until then we have to jump through hoops. */
vset(&char_pointer_type, TREG_SP, 0);
ofs = (ofs + align - 1) & -align;
vpushi(ofs);
gen_op('+');
indir();
vtop->type = vtop[-1].type;
vswap();
vstore();
vtop->r = vtop->r2 = VT_CONST; // this arg is done
ofs += size;
}
vrott(nb_args - i);
} else if (info[i] & 16) {
assert(!splitofs);
splitofs = ofs;
ofs += 8;
}
}
}
for (i = 0; i < nb_args; i++) {
int r = info[nb_args - 1 - i];
if (!(r & 32)) {
CType origtype;
r &= 15;
vrotb(i+1);
origtype = vtop->type;
size = type_size(&vtop->type, &align);
if (size > 8 && (vtop->type.t & VT_BTYPE) == VT_STRUCT)
vtop->type.t = VT_LDOUBLE; // force loading a pair of regs
gv(r < 8 ? RC_R(r) : RC_F(r - 8));
vtop->type = origtype;
if (size > 8) {
assert((vtop->type.t & VT_BTYPE) == VT_LDOUBLE
|| (vtop->type.t & VT_BTYPE) == VT_STRUCT);
assert(vtop->r2 < VT_CONST);
if (info[nb_args - 1 - i] & 16) {
ES(0x23, 3, 2, ireg(vtop->r2), splitofs); // sd t0, ofs(sp)
} else if (vtop->r2 != 1 + vtop->r) {
assert(vtop->r < 7);
/* XXX we'd like to have 'gv' move directly into
the right class instead of us fixing it up. */
EI(0x13, 0, ireg(vtop->r) + 1, ireg(vtop->r2), 0); // mv Ra+1, RR2
vtop->r2 = 1 + vtop->r;
}
}
vrott(i+1);
}
}
vrotb(nb_args + 1);
save_regs(nb_args + 1);
gcall_or_jmp(1);
vtop -= nb_args + 1;
if (stack_adj + tempspace)
EI(0x13, 0, 2, 2, stack_adj + tempspace); // addi sp, sp, adj
2019-06-22 12:13:10 +08:00
}
static int func_sub_sp_offset, num_va_regs;
2019-06-22 12:13:10 +08:00
ST_FUNC void gfunc_prolog(CType *func_type)
{
int i, addr, align, size;
int param_addr = 0;
int aireg, afreg;
Sym *sym;
CType *type;
sym = func_type->ref;
func_vt = sym->type;
loc = -16; // for ra and s0
func_sub_sp_offset = ind;
ind += 5 * 4;
aireg = afreg = 0;
addr = 0; // XXX not correct
/* if the function returns by reference, then add an
implicit pointer parameter */
size = type_size(&func_vt, &align);
if (size > 2 * XLEN) {
loc -= 8;
func_vc = loc;
ES(0x23, 3, 8, 10 + aireg, loc); // sd a0, loc(s0)
aireg++;
}
/* define parameters */
while ((sym = sym->next) != NULL) {
int byref = 0;
type = &sym->type;
size = type_size(type, &align);
if (size > 2 * XLEN) {
type = &char_pointer_type;
size = align = byref = 8;
}
if (size > 2 * XLEN) {
from_stack:
if (align < XLEN)
align = XLEN;
addr = (addr + align - 1) & -align;
param_addr = addr;
addr += size;
} else {
int regcount = 1, *pareg = &aireg;
if (is_float(type->t) && (type->t & VT_BTYPE) != VT_LDOUBLE)
pareg = &afreg;
if (regcount + *pareg > 8)
goto from_stack;
if (size > XLEN)
regcount++;
loc -= regcount * 8; // XXX could reserve only 'size' bytes
param_addr = loc;
for (i = 0; i < regcount; i++) {
if (*pareg >= 8) {
assert(i == 1 && regcount == 2 && !(addr & 7));
EI(0x03, 3, 5, 8, addr); // ld t0, addr(s0)
addr += 8;
ES(0x23, 3, 8, 5, loc + i*8); // sd t0, loc(s0)
continue;
}
if (pareg == &afreg) {
assert(type->t == VT_FLOAT || type->t == VT_DOUBLE);
ES(0x27, size == 4 ? 2 : 3, 8, 10 + *pareg, loc + i*8); // fs[wd] FAi, loc(s0)
} else {
ES(0x23, 3, 8, 10 + *pareg, loc + i*8); // sd aX, loc(s0) // XXX
}
(*pareg)++;
}
}
sym_push(sym->v & ~SYM_FIELD, &sym->type,
(byref ? VT_LLOCAL : VT_LOCAL) | lvalue_type(sym->type.t),
param_addr);
}
num_va_regs = 0;
if (func_type->ref->f.func_type == FUNC_ELLIPSIS) {
for (; aireg < 8; aireg++) {
num_va_regs++;
ES(0x23, 3, 8, 10 + aireg, -8 + num_va_regs * 8); // sd aX, loc(s0)
}
}
2019-06-22 12:13:10 +08:00
}
ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret,
int *ret_align, int *regsize)
2019-06-22 12:13:10 +08:00
{
/* generic code can only deal with structs of pow(2) sizes
(it always deals with whole registers), so go through our own
code. */
int align, size = type_size(vt, &align);
*ret_align = 1;
*regsize = 8;
if (size > 16)
return 0;
if (size > 8)
ret->t = VT_LLONG;
else if (size > 4)
ret->t = VT_LLONG;
else if (size > 2)
ret->t = VT_INT;
else if (size > 1)
ret->t = VT_SHORT;
else
ret->t = VT_BYTE;
return (size + 7) / 8;
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gfunc_return(CType *func_type)
2019-06-22 12:13:10 +08:00
{
int align, size = type_size(func_type, &align), nregs;
CType type = *func_type;
if (size > 2 * XLEN) {
mk_pointer(&type);
vset(&type, VT_LOCAL | VT_LVAL, func_vc);
indir();
vswap();
vstore();
vpop();
return;
}
nregs = (size + 7) / 8;
if (nregs == 2)
vtop->type.t = VT_LDOUBLE;
if (is_float(func_type->t) && (vtop->type.t & VT_BTYPE) != VT_LDOUBLE)
gv(RC_FRET);
else
gv(RC_IRET);
vtop--;
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gfunc_epilog(void)
2019-06-22 12:13:10 +08:00
{
int v, saved_ind, d, large_ofs_ind;
loc = (loc - num_va_regs * 8);
d = v = (-loc + 15) & -16;
if (v >= (1 << 11)) {
d = 16;
o(0x37 | (5 << 7) | ((0x800 + (v-16)) & 0xfffff000)); //lui t0, upper(v)
EI(0x13, 0, 5, 5, (v-16) << 20 >> 20); // addi t0, t0, lo(v)
o(0x33 | (2 << 7) | (2 << 15) | (5 << 20)); //add sp, sp, t0
}
EI(0x03, 3, 1, 2, d - 8 - num_va_regs * 8); // ld ra, v-8(sp)
EI(0x03, 3, 8, 2, d - 16 - num_va_regs * 8); // ld s0, v-16(sp)
EI(0x13, 0, 2, 2, d); // addi sp, sp, v
EI(0x67, 0, 0, 1, 0); // jalr x0, 0(x1), aka ret
if (v >= (1 << 11)) {
large_ofs_ind = ind;
EI(0x13, 0, 8, 2, d - num_va_regs * 8); // addi s0, sp, d
o(0x37 | (5 << 7) | ((0x800 + (v-16)) & 0xfffff000)); //lui t0, upper(v)
EI(0x13, 0, 5, 5, (v-16) << 20 >> 20); // addi t0, t0, lo(v)
o(0x33 | (2 << 7) | (2 << 15) | (5 << 20) | (0x20 << 25)); //sub sp, sp, t0
gjmp_addr(func_sub_sp_offset + 5*4);
}
saved_ind = ind;
ind = func_sub_sp_offset;
EI(0x13, 0, 2, 2, -d); // addi sp, sp, -d
ES(0x23, 3, 2, 1, d - 8 - num_va_regs * 8); // sd ra, d-8(sp)
ES(0x23, 3, 2, 8, d - 16 - num_va_regs * 8); // sd s0, d-16(sp)
if (v < (1 << 11))
EI(0x13, 0, 8, 2, d - num_va_regs * 8); // addi s0, sp, d
else
gjmp_addr(large_ofs_ind);
if ((ind - func_sub_sp_offset) != 5*4)
EI(0x13, 0, 0, 0, 0); // addi x0, x0, 0 == nop
ind = saved_ind;
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gen_va_start(void)
2019-06-22 12:13:10 +08:00
{
tcc_error("implement me: %s", __FUNCTION__);
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gen_va_arg(CType *t)
2019-06-22 12:13:10 +08:00
{
tcc_error("implement me: %s", __FUNCTION__);
2019-06-22 12:13:10 +08:00
}
2019-06-22 12:13:10 +08:00
ST_FUNC void gen_fill_nops(int bytes)
{
if ((bytes & 3))
tcc_error("alignment of code section not multiple of 4");
while (bytes > 0) {
EI(0x13, 0, 0, 0, 0); // addi x0, x0, 0 == nop
bytes -= 4;
}
2019-06-22 12:13:10 +08:00
}
// Generate forward branch to label:
ST_FUNC int gjmp(int t)
{
if (nocode_wanted)
return t;
o(t);
return ind - 4;
2019-06-22 12:13:10 +08:00
}
// Generate branch to known address:
ST_FUNC void gjmp_addr(int a)
{
uint32_t r = a - ind, imm;
if ((r + (1 << 21)) & ~((1U << 22) - 2)) {
o(0x17 | (5 << 7) | (((r + 0x800) & 0xfffff000))); // lui RR, up(r)
r = (int)r << 20 >> 20;
EI(0x67, 0, 0, 5, r); // jalr x0, r(t0)
} else {
imm = (((r >> 12) & 0xff) << 12)
| (((r >> 11) & 1) << 20)
| (((r >> 1) & 0x3ff) << 21)
| (((r >> 20) & 1) << 31);
o(0x6f | imm); // jal x0, imm == j imm
}
2019-06-22 12:13:10 +08:00
}
ST_FUNC int gjmp_cond(int op, int t)
{
int inv = op & 1;
assert(op == TOK_EQ || op == TOK_NE);
assert(vtop->cmp_r >= 10 && vtop->cmp_r < 18);
o(0x63 | (!inv << 12) | (vtop->cmp_r << 15) | (8 << 7)); // bne/beq x0,r,+4
return gjmp(t);
2019-06-22 12:13:10 +08:00
}
ST_FUNC int gjmp_append(int n, int t)
{
void *p;
/* insert jump list n into t */
if (n) {
uint32_t n1 = n, n2;
while ((n2 = read32le(p = cur_text_section->data + n1)))
n1 = n2;
write32le(p, t);
t = n;
}
return t;
}
static void gen_opil(int op, int ll)
{
int a, b, d;
int inv = 0;
int func3 = 0, func7 = 0;
/* XXX We could special-case some constant args. */
gv2(RC_INT, RC_INT);
a = ireg(vtop[-1].r);
b = ireg(vtop[0].r);
vtop -= 2;
d = get_reg(RC_INT);
vtop++;
vtop[0].r = d;
d = ireg(d);
ll = ll ? 0 : 8;
switch (op) {
default:
tcc_error("implement me: %s(%s)", __FUNCTION__, get_tok_str(op, NULL));
case '+':
o(0x33 | (d << 7) | (a << 15) | (b << 20)); // add d, a, b
break;
case '-':
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (0x20 << 25)); //sub d, a, b
break;
case TOK_SAR:
o(0x33 | ll | (d << 7) | (a << 15) | (b << 20) | (5 << 12) | (1 << 30)); //sra d, a, b
break;
case TOK_SHR:
o(0x33 | ll | (d << 7) | (a << 15) | (b << 20) | (5 << 12)); //srl d, a, b
break;
case TOK_SHL:
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (1 << 12)); //sll d, a, b
break;
case '*':
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (0x01 << 25)); //mul d, a, b
break;
case '/':
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (0x01 << 25) | (4 << 12)); //div d, a, b
break;
case '&':
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (7 << 12)); // and d, a, b
break;
case '^':
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (4 << 12)); // xor d, a, b
break;
case '|':
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (6 << 12)); // or d, a, b
break;
case '%':
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (0x01 << 25) | (6 << 12)); //rem d, a, b
break;
case TOK_UMOD:
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (0x01 << 25) | (7 << 12)); //remu d, a, b
break;
case TOK_PDIV:
case TOK_UDIV:
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (0x01 << 25) | (5 << 12)); //divu d, a, b
break;
case TOK_ULT:
case TOK_UGE:
case TOK_ULE:
case TOK_UGT:
case TOK_LT:
case TOK_GE:
case TOK_LE:
case TOK_GT:
if (op & 1) { // remove [U]GE,GT
inv = 1;
op--;
}
if ((op & 7) == 6) { // [U]LE
int t = a; a = b; b = t;
inv ^= 1;
}
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (((op > TOK_UGT) ? 2 : 3) << 12)); // slt[u] d, a, b
if (inv)
EI(0x13, 4, d, d, 1); // xori d, d, 1
vset_VT_CMP(TOK_NE);
vtop->cmp_r = d;
break;
case TOK_NE:
case TOK_EQ:
o(0x33 | (d << 7) | (a << 15) | (b << 20) | (0x20 << 25)); // sub d, a, b
if (op == TOK_NE)
o(0x33 | (3 << 12) | (d << 7) | (0 << 15) | (d << 20)); // sltu d, x0, d == snez d,d
else
EI(0x13, 3, d, d, 1); // sltiu d, d, 1 == seqz d,d
vset_VT_CMP(TOK_NE);
vtop->cmp_r = d;
break;
}
2019-06-22 12:13:10 +08:00
}
2019-06-22 12:13:10 +08:00
ST_FUNC void gen_opi(int op)
{
gen_opil(op, 0);
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gen_opl(int op)
{
gen_opil(op, 1);
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gen_opf(int op)
{
int rs1, rs2, rd, dbl, invert;
if (vtop[0].type.t == VT_LDOUBLE) {
CType type = vtop[0].type;
int func = 0;
int cond = -1;
switch (op) {
case '*': func = TOK___multf3; break;
case '+': func = TOK___addtf3; break;
case '-': func = TOK___subtf3; break;
case '/': func = TOK___divtf3; break;
case TOK_EQ: func = TOK___eqtf2; cond = 1; break;
case TOK_NE: func = TOK___netf2; cond = 0; break;
case TOK_LT: func = TOK___lttf2; cond = 10; break;
case TOK_GE: func = TOK___getf2; cond = 11; break;
case TOK_LE: func = TOK___letf2; cond = 12; break;
case TOK_GT: func = TOK___gttf2; cond = 13; break;
default: assert(0); break;
}
vpush_global_sym(&func_old_type, func);
vrott(3);
gfunc_call(2);
vpushi(0);
vtop->r = REG_IRET;
vtop->r2 = cond < 0 ? TREG_R(1) : VT_CONST;
if (cond < 0)
vtop->type = type;
else {
vpushi(0);
gen_opil(op, 1);
}
return;
}
gv2(RC_FLOAT, RC_FLOAT);
assert(vtop->type.t == VT_DOUBLE || vtop->type.t == VT_FLOAT);
dbl = vtop->type.t == VT_DOUBLE;
rs1 = freg(vtop[-1].r);
rs2 = freg(vtop->r);
vtop--;
invert = 0;
switch(op) {
default:
assert(0);
case '+':
op = 0; // fadd
arithop:
rd = get_reg(RC_FLOAT);
vtop->r = rd;
rd = freg(rd);
o(0x53 | (rd << 7) | (rs1 << 15) | (rs2 << 20) | (7 << 12) | (dbl << 25) | (op << 27)); // fop.[sd] RD, RS1, RS2 (dyn rm)
break;
case '-':
op = 1; // fsub
goto arithop;
case '*':
op = 2; // fmul
goto arithop;
case '/':
op = 3; // fdiv
goto arithop;
case TOK_EQ:
op = 2; // EQ
cmpop:
rd = get_reg(RC_INT);
vtop->r = rd;
rd = ireg(rd);
o(0x53 | (rd << 7) | (rs1 << 15) | (rs2 << 20) | (op << 12) | (dbl << 25) | (0x14 << 27)); // fcmp.[sd] RD, RS1, RS2 (op == eq/lt/le)
if (invert)
EI(0x13, 4, rd, rd, 1); // xori RD, 1
break;
case TOK_NE:
invert = 1;
op = 2; // EQ
goto cmpop;
case TOK_LT:
op = 1; // LT
goto cmpop;
case TOK_LE:
op = 0; // LE
goto cmpop;
case TOK_GT:
op = 1; // LT
rd = rs1, rs1 = rs2, rs2 = rd;
goto cmpop;
case TOK_GE:
op = 0; // LE
rd = rs1, rs1 = rs2, rs2 = rd;
goto cmpop;
}
2019-06-22 12:13:10 +08:00
}
2019-06-22 12:13:10 +08:00
ST_FUNC void gen_cvt_sxtw(void)
{
/* XXX on risc-v the registers are usually sign-extended already.
Let's try to not do anything here. */
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gen_cvt_itof(int t)
{
int rr = ireg(gv(RC_INT)), dr;
int u = vtop->type.t & VT_UNSIGNED;
int l = (vtop->type.t & VT_BTYPE) == VT_LLONG;
if (t == VT_LDOUBLE) {
int func = l ?
(u ? TOK___floatunditf : TOK___floatditf) :
(u ? TOK___floatunsitf : TOK___floatsitf);
vpush_global_sym(&func_old_type, func);
vrott(2);
gfunc_call(1);
vpushi(0);
vtop->type.t = t;
vtop->r = REG_IRET;
vtop->r2 = TREG_R(1);
} else {
vtop--;
dr = get_reg(RC_FLOAT);
vtop++;
vtop->r = dr;
dr = freg(dr);
EIu(0x53, 7, dr, rr, ((0x68 | (t == VT_DOUBLE ? 1 : 0)) << 5) | (u ? 1 : 0) | (l ? 2 : 0)); // fcvt.[sd].[wl][u]
}
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gen_cvt_ftoi(int t)
{
int ft = vtop->type.t & VT_BTYPE;
int l = (t & VT_BTYPE) == VT_LLONG;
int u = t & VT_UNSIGNED;
if (ft == VT_LDOUBLE) {
int func = l ?
(u ? TOK___fixunstfdi : TOK___fixtfdi) :
(u ? TOK___fixunstfsi : TOK___fixtfsi);
vpush_global_sym(&func_old_type, func);
vrott(2);
gfunc_call(1);
vpushi(0);
vtop->type.t = t;
vtop->r = REG_IRET;
} else {
int rr = freg(gv(RC_FLOAT)), dr;
vtop--;
dr = get_reg(RC_INT);
vtop++;
vtop->r = dr;
dr = ireg(dr);
EIu(0x53, 1, dr, rr, ((0x60 | (ft == VT_DOUBLE ? 1 : 0)) << 5) | (u ? 1 : 0) | (l ? 2 : 0)); // fcvt.[wl][u].[sd] rtz
}
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gen_cvt_ftof(int dt)
2019-06-22 12:13:10 +08:00
{
int st = vtop->type.t & VT_BTYPE, rs, rd;
dt &= VT_BTYPE;
if (st == dt)
return;
if (dt == VT_LDOUBLE || st == VT_LDOUBLE) {
int func = (dt == VT_LDOUBLE) ?
(st == VT_FLOAT ? TOK___extendsftf2 : TOK___extenddftf2) :
(dt == VT_FLOAT ? TOK___trunctfsf2 : TOK___trunctfdf2);
/* We can't use gfunc_call, as func_old_type works like vararg
functions, and on riscv unnamed float args are passed like
integers. But we really need them in the float argument registers
for extendsftf2/extenddftf2. So, do it explicitely. */
save_regs(1);
if (dt == VT_LDOUBLE)
gv(RC_F(0));
else {
gv(RC_R(0));
assert(vtop->r2 < 7);
if (vtop->r2 != 1 + vtop->r) {
EI(0x13, 0, ireg(vtop->r) + 1, ireg(vtop->r2), 0); // mv Ra+1, RR2
vtop->r2 = 1 + vtop->r;
}
}
vpush_global_sym(&func_old_type, func);
gcall_or_jmp(1);
vtop -= 2;
vpushi(0);
vtop->type.t = dt;
if (dt == VT_LDOUBLE)
vtop->r = REG_IRET, vtop->r2 = REG_IRET+1;
else
vtop->r = REG_FRET;
} else {
assert (dt == VT_FLOAT || dt == VT_DOUBLE);
assert (st == VT_FLOAT || st == VT_DOUBLE);
rs = gv(RC_FLOAT);
rd = get_reg(RC_FLOAT);
if (dt == VT_DOUBLE)
EI(0x53, 7, freg(rd), freg(rs), 0x21 << 5); // fcvt.d.s RD, RS (dyn rm)
else
EI(0x53, 7, freg(rd), freg(rs), (0x20 << 5) | 1); // fcvt.s.d RD, RS
vtop->r = rd;
}
2019-06-22 12:13:10 +08:00
}
ST_FUNC void ggoto(void)
{
gcall_or_jmp(0);
vtop--;
2019-06-22 12:13:10 +08:00
}
2019-06-22 12:13:10 +08:00
ST_FUNC void gen_vla_sp_save(int addr)
{
ES(0x23, 3, 8, 2, addr); // sd sp, fc(s0)
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gen_vla_sp_restore(int addr)
{
EI(0x03, 3, 2, 8, addr); // ld sp, fc(s0)
2019-06-22 12:13:10 +08:00
}
ST_FUNC void gen_vla_alloc(CType *type, int align)
{
int rr = ireg(gv(RC_INT));
EI(0x13, 0, rr, rr, 15); // addi RR, RR, 15
EI(0x13, 7, rr, rr, -16); // andi, RR, RR, -16
o(0x33 | (2 << 7) | (2 << 15) | (rr << 20) | (0x20 << 25)); //sub sp, sp, rr
vpop();
2019-06-22 12:13:10 +08:00
}
#endif