tcg/optimize: Split out fold_masks

Move all of the known-zero optimizations into the per-opcode
functions.  Use fold_masks when there is a possibility of the
result being determined, and simply set ctx->z_mask otherwise.

Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Luis Pires <luis.pires@eldorado.org.br>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This commit is contained in:
Richard Henderson 2021-08-25 22:42:19 -07:00
parent da48e27202
commit fae450ba47
1 changed files with 294 additions and 251 deletions

View File

@ -50,7 +50,8 @@ typedef struct OptContext {
TCGTempSet temps_used;
/* In flight values from optimization. */
uint64_t z_mask;
uint64_t a_mask; /* mask bit is 0 iff value identical to first input */
uint64_t z_mask; /* mask bit is 0 iff value bit is 0 */
TCGType type;
} OptContext;
@ -694,6 +695,31 @@ static bool fold_const2(OptContext *ctx, TCGOp *op)
return false;
}
static bool fold_masks(OptContext *ctx, TCGOp *op)
{
uint64_t a_mask = ctx->a_mask;
uint64_t z_mask = ctx->z_mask;
/*
* 32-bit ops generate 32-bit results. For the result is zero test
* below, we can ignore high bits, but for further optimizations we
* need to record that the high bits contain garbage.
*/
if (ctx->type == TCG_TYPE_I32) {
ctx->z_mask |= MAKE_64BIT_MASK(32, 32);
a_mask &= MAKE_64BIT_MASK(0, 32);
z_mask &= MAKE_64BIT_MASK(0, 32);
}
if (z_mask == 0) {
return tcg_opt_gen_movi(ctx, op, op->args[0], 0);
}
if (a_mask == 0) {
return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
}
return false;
}
/*
* Convert @op to NOT, if NOT is supported by the host.
* Return true f the conversion is successful, which will still
@ -847,24 +873,55 @@ static bool fold_add2_i32(OptContext *ctx, TCGOp *op)
static bool fold_and(OptContext *ctx, TCGOp *op)
{
uint64_t z1, z2;
if (fold_const2(ctx, op) ||
fold_xi_to_i(ctx, op, 0) ||
fold_xi_to_x(ctx, op, -1) ||
fold_xx_to_x(ctx, op)) {
return true;
}
return false;
z1 = arg_info(op->args[1])->z_mask;
z2 = arg_info(op->args[2])->z_mask;
ctx->z_mask = z1 & z2;
/*
* Known-zeros does not imply known-ones. Therefore unless
* arg2 is constant, we can't infer affected bits from it.
*/
if (arg_is_const(op->args[2])) {
ctx->a_mask = z1 & ~z2;
}
return fold_masks(ctx, op);
}
static bool fold_andc(OptContext *ctx, TCGOp *op)
{
uint64_t z1;
if (fold_const2(ctx, op) ||
fold_xx_to_i(ctx, op, 0) ||
fold_xi_to_x(ctx, op, 0) ||
fold_ix_to_not(ctx, op, -1)) {
return true;
}
return false;
z1 = arg_info(op->args[1])->z_mask;
/*
* Known-zeros does not imply known-ones. Therefore unless
* arg2 is constant, we can't infer anything from it.
*/
if (arg_is_const(op->args[2])) {
uint64_t z2 = ~arg_info(op->args[2])->z_mask;
ctx->a_mask = z1 & ~z2;
z1 &= z2;
}
ctx->z_mask = z1;
return fold_masks(ctx, op);
}
static bool fold_brcond(OptContext *ctx, TCGOp *op)
@ -963,13 +1020,52 @@ static bool fold_brcond2(OptContext *ctx, TCGOp *op)
static bool fold_bswap(OptContext *ctx, TCGOp *op)
{
uint64_t z_mask, sign;
if (arg_is_const(op->args[1])) {
uint64_t t = arg_info(op->args[1])->val;
t = do_constant_folding(op->opc, ctx->type, t, op->args[2]);
return tcg_opt_gen_movi(ctx, op, op->args[0], t);
}
return false;
z_mask = arg_info(op->args[1])->z_mask;
switch (op->opc) {
case INDEX_op_bswap16_i32:
case INDEX_op_bswap16_i64:
z_mask = bswap16(z_mask);
sign = INT16_MIN;
break;
case INDEX_op_bswap32_i32:
case INDEX_op_bswap32_i64:
z_mask = bswap32(z_mask);
sign = INT32_MIN;
break;
case INDEX_op_bswap64_i64:
z_mask = bswap64(z_mask);
sign = INT64_MIN;
break;
default:
g_assert_not_reached();
}
switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) {
case TCG_BSWAP_OZ:
break;
case TCG_BSWAP_OS:
/* If the sign bit may be 1, force all the bits above to 1. */
if (z_mask & sign) {
z_mask |= sign;
}
break;
default:
/* The high bits are undefined: force all bits above the sign to 1. */
z_mask |= sign << 1;
break;
}
ctx->z_mask = z_mask;
return fold_masks(ctx, op);
}
static bool fold_call(OptContext *ctx, TCGOp *op)
@ -1006,6 +1102,8 @@ static bool fold_call(OptContext *ctx, TCGOp *op)
static bool fold_count_zeros(OptContext *ctx, TCGOp *op)
{
uint64_t z_mask;
if (arg_is_const(op->args[1])) {
uint64_t t = arg_info(op->args[1])->val;
@ -1015,12 +1113,39 @@ static bool fold_count_zeros(OptContext *ctx, TCGOp *op)
}
return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
}
switch (ctx->type) {
case TCG_TYPE_I32:
z_mask = 31;
break;
case TCG_TYPE_I64:
z_mask = 63;
break;
default:
g_assert_not_reached();
}
ctx->z_mask = arg_info(op->args[2])->z_mask | z_mask;
return false;
}
static bool fold_ctpop(OptContext *ctx, TCGOp *op)
{
return fold_const1(ctx, op);
if (fold_const1(ctx, op)) {
return true;
}
switch (ctx->type) {
case TCG_TYPE_I32:
ctx->z_mask = 32 | 31;
break;
case TCG_TYPE_I64:
ctx->z_mask = 64 | 63;
break;
default:
g_assert_not_reached();
}
return false;
}
static bool fold_deposit(OptContext *ctx, TCGOp *op)
@ -1032,6 +1157,10 @@ static bool fold_deposit(OptContext *ctx, TCGOp *op)
t1 = deposit64(t1, op->args[3], op->args[4], t2);
return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
}
ctx->z_mask = deposit64(arg_info(op->args[1])->z_mask,
op->args[3], op->args[4],
arg_info(op->args[2])->z_mask);
return false;
}
@ -1077,6 +1206,8 @@ static bool fold_eqv(OptContext *ctx, TCGOp *op)
static bool fold_extract(OptContext *ctx, TCGOp *op)
{
uint64_t z_mask_old, z_mask;
if (arg_is_const(op->args[1])) {
uint64_t t;
@ -1084,7 +1215,15 @@ static bool fold_extract(OptContext *ctx, TCGOp *op)
t = extract64(t, op->args[2], op->args[3]);
return tcg_opt_gen_movi(ctx, op, op->args[0], t);
}
return false;
z_mask_old = arg_info(op->args[1])->z_mask;
z_mask = extract64(z_mask_old, op->args[2], op->args[3]);
if (op->args[2] == 0) {
ctx->a_mask = z_mask_old ^ z_mask;
}
ctx->z_mask = z_mask;
return fold_masks(ctx, op);
}
static bool fold_extract2(OptContext *ctx, TCGOp *op)
@ -1108,12 +1247,83 @@ static bool fold_extract2(OptContext *ctx, TCGOp *op)
static bool fold_exts(OptContext *ctx, TCGOp *op)
{
return fold_const1(ctx, op);
uint64_t z_mask_old, z_mask, sign;
bool type_change = false;
if (fold_const1(ctx, op)) {
return true;
}
z_mask_old = z_mask = arg_info(op->args[1])->z_mask;
switch (op->opc) {
CASE_OP_32_64(ext8s):
sign = INT8_MIN;
z_mask = (uint8_t)z_mask;
break;
CASE_OP_32_64(ext16s):
sign = INT16_MIN;
z_mask = (uint16_t)z_mask;
break;
case INDEX_op_ext_i32_i64:
type_change = true;
QEMU_FALLTHROUGH;
case INDEX_op_ext32s_i64:
sign = INT32_MIN;
z_mask = (uint32_t)z_mask;
break;
default:
g_assert_not_reached();
}
if (z_mask & sign) {
z_mask |= sign;
} else if (!type_change) {
ctx->a_mask = z_mask_old ^ z_mask;
}
ctx->z_mask = z_mask;
return fold_masks(ctx, op);
}
static bool fold_extu(OptContext *ctx, TCGOp *op)
{
return fold_const1(ctx, op);
uint64_t z_mask_old, z_mask;
bool type_change = false;
if (fold_const1(ctx, op)) {
return true;
}
z_mask_old = z_mask = arg_info(op->args[1])->z_mask;
switch (op->opc) {
CASE_OP_32_64(ext8u):
z_mask = (uint8_t)z_mask;
break;
CASE_OP_32_64(ext16u):
z_mask = (uint16_t)z_mask;
break;
case INDEX_op_extrl_i64_i32:
case INDEX_op_extu_i32_i64:
type_change = true;
QEMU_FALLTHROUGH;
case INDEX_op_ext32u_i64:
z_mask = (uint32_t)z_mask;
break;
case INDEX_op_extrh_i64_i32:
type_change = true;
z_mask >>= 32;
break;
default:
g_assert_not_reached();
}
ctx->z_mask = z_mask;
if (!type_change) {
ctx->a_mask = z_mask_old ^ z_mask;
}
return fold_masks(ctx, op);
}
static bool fold_mb(OptContext *ctx, TCGOp *op)
@ -1154,6 +1364,9 @@ static bool fold_movcond(OptContext *ctx, TCGOp *op)
return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]);
}
ctx->z_mask = arg_info(op->args[3])->z_mask
| arg_info(op->args[4])->z_mask;
if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) {
uint64_t tv = arg_info(op->args[3])->val;
uint64_t fv = arg_info(op->args[4])->val;
@ -1228,9 +1441,16 @@ static bool fold_nand(OptContext *ctx, TCGOp *op)
static bool fold_neg(OptContext *ctx, TCGOp *op)
{
uint64_t z_mask;
if (fold_const1(ctx, op)) {
return true;
}
/* Set to 1 all bits to the left of the rightmost. */
z_mask = arg_info(op->args[1])->z_mask;
ctx->z_mask = -(z_mask & -z_mask);
/*
* Because of fold_sub_to_neg, we want to always return true,
* via finish_folding.
@ -1266,7 +1486,10 @@ static bool fold_or(OptContext *ctx, TCGOp *op)
fold_xx_to_x(ctx, op)) {
return true;
}
return false;
ctx->z_mask = arg_info(op->args[1])->z_mask
| arg_info(op->args[2])->z_mask;
return fold_masks(ctx, op);
}
static bool fold_orc(OptContext *ctx, TCGOp *op)
@ -1281,6 +1504,15 @@ static bool fold_orc(OptContext *ctx, TCGOp *op)
static bool fold_qemu_ld(OptContext *ctx, TCGOp *op)
{
const TCGOpDef *def = &tcg_op_defs[op->opc];
MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs];
MemOp mop = get_memop(oi);
int width = 8 * memop_size(mop);
if (!(mop & MO_SIGN) && width < 64) {
ctx->z_mask = MAKE_64BIT_MASK(0, width);
}
/* Opcodes that touch guest memory stop the mb optimization. */
ctx->prev_mb = NULL;
return false;
@ -1306,6 +1538,8 @@ static bool fold_setcond(OptContext *ctx, TCGOp *op)
if (i >= 0) {
return tcg_opt_gen_movi(ctx, op, op->args[0], i);
}
ctx->z_mask = 1;
return false;
}
@ -1372,6 +1606,8 @@ static bool fold_setcond2(OptContext *ctx, TCGOp *op)
op->opc = INDEX_op_setcond_i32;
break;
}
ctx->z_mask = 1;
return false;
do_setcond_const:
@ -1380,6 +1616,8 @@ static bool fold_setcond2(OptContext *ctx, TCGOp *op)
static bool fold_sextract(OptContext *ctx, TCGOp *op)
{
int64_t z_mask_old, z_mask;
if (arg_is_const(op->args[1])) {
uint64_t t;
@ -1387,7 +1625,15 @@ static bool fold_sextract(OptContext *ctx, TCGOp *op)
t = sextract64(t, op->args[2], op->args[3]);
return tcg_opt_gen_movi(ctx, op, op->args[0], t);
}
return false;
z_mask_old = arg_info(op->args[1])->z_mask;
z_mask = sextract64(z_mask_old, op->args[2], op->args[3]);
if (op->args[2] == 0 && z_mask >= 0) {
ctx->a_mask = z_mask_old ^ z_mask;
}
ctx->z_mask = z_mask;
return fold_masks(ctx, op);
}
static bool fold_shift(OptContext *ctx, TCGOp *op)
@ -1397,6 +1643,13 @@ static bool fold_shift(OptContext *ctx, TCGOp *op)
fold_xi_to_x(ctx, op, 0)) {
return true;
}
if (arg_is_const(op->args[2])) {
ctx->z_mask = do_constant_folding(op->opc, ctx->type,
arg_info(op->args[1])->z_mask,
arg_info(op->args[2])->val);
return fold_masks(ctx, op);
}
return false;
}
@ -1452,6 +1705,25 @@ static bool fold_sub2_i32(OptContext *ctx, TCGOp *op)
return fold_addsub2_i32(ctx, op, false);
}
static bool fold_tcg_ld(OptContext *ctx, TCGOp *op)
{
/* We can't do any folding with a load, but we can record bits. */
switch (op->opc) {
CASE_OP_32_64(ld8u):
ctx->z_mask = MAKE_64BIT_MASK(0, 8);
break;
CASE_OP_32_64(ld16u):
ctx->z_mask = MAKE_64BIT_MASK(0, 16);
break;
case INDEX_op_ld32u_i64:
ctx->z_mask = MAKE_64BIT_MASK(0, 32);
break;
default:
g_assert_not_reached();
}
return false;
}
static bool fold_xor(OptContext *ctx, TCGOp *op)
{
if (fold_const2(ctx, op) ||
@ -1460,7 +1732,10 @@ static bool fold_xor(OptContext *ctx, TCGOp *op)
fold_xi_to_not(ctx, op, -1)) {
return true;
}
return false;
ctx->z_mask = arg_info(op->args[1])->z_mask
| arg_info(op->args[2])->z_mask;
return fold_masks(ctx, op);
}
/* Propagate constants and copies, fold constant expressions. */
@ -1481,7 +1756,6 @@ void tcg_optimize(TCGContext *s)
}
QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
uint64_t z_mask, partmask, affected, tmp;
TCGOpcode opc = op->opc;
const TCGOpDef *def;
bool done = false;
@ -1562,245 +1836,9 @@ void tcg_optimize(TCGContext *s)
break;
}
/* Simplify using known-zero bits. Currently only ops with a single
output argument is supported. */
z_mask = -1;
affected = -1;
switch (opc) {
CASE_OP_32_64(ext8s):
if ((arg_info(op->args[1])->z_mask & 0x80) != 0) {
break;
}
QEMU_FALLTHROUGH;
CASE_OP_32_64(ext8u):
z_mask = 0xff;
goto and_const;
CASE_OP_32_64(ext16s):
if ((arg_info(op->args[1])->z_mask & 0x8000) != 0) {
break;
}
QEMU_FALLTHROUGH;
CASE_OP_32_64(ext16u):
z_mask = 0xffff;
goto and_const;
case INDEX_op_ext32s_i64:
if ((arg_info(op->args[1])->z_mask & 0x80000000) != 0) {
break;
}
QEMU_FALLTHROUGH;
case INDEX_op_ext32u_i64:
z_mask = 0xffffffffU;
goto and_const;
CASE_OP_32_64(and):
z_mask = arg_info(op->args[2])->z_mask;
if (arg_is_const(op->args[2])) {
and_const:
affected = arg_info(op->args[1])->z_mask & ~z_mask;
}
z_mask = arg_info(op->args[1])->z_mask & z_mask;
break;
case INDEX_op_ext_i32_i64:
if ((arg_info(op->args[1])->z_mask & 0x80000000) != 0) {
break;
}
QEMU_FALLTHROUGH;
case INDEX_op_extu_i32_i64:
/* We do not compute affected as it is a size changing op. */
z_mask = (uint32_t)arg_info(op->args[1])->z_mask;
break;
CASE_OP_32_64(andc):
/* Known-zeros does not imply known-ones. Therefore unless
op->args[2] is constant, we can't infer anything from it. */
if (arg_is_const(op->args[2])) {
z_mask = ~arg_info(op->args[2])->z_mask;
goto and_const;
}
/* But we certainly know nothing outside args[1] may be set. */
z_mask = arg_info(op->args[1])->z_mask;
break;
case INDEX_op_sar_i32:
if (arg_is_const(op->args[2])) {
tmp = arg_info(op->args[2])->val & 31;
z_mask = (int32_t)arg_info(op->args[1])->z_mask >> tmp;
}
break;
case INDEX_op_sar_i64:
if (arg_is_const(op->args[2])) {
tmp = arg_info(op->args[2])->val & 63;
z_mask = (int64_t)arg_info(op->args[1])->z_mask >> tmp;
}
break;
case INDEX_op_shr_i32:
if (arg_is_const(op->args[2])) {
tmp = arg_info(op->args[2])->val & 31;
z_mask = (uint32_t)arg_info(op->args[1])->z_mask >> tmp;
}
break;
case INDEX_op_shr_i64:
if (arg_is_const(op->args[2])) {
tmp = arg_info(op->args[2])->val & 63;
z_mask = (uint64_t)arg_info(op->args[1])->z_mask >> tmp;
}
break;
case INDEX_op_extrl_i64_i32:
z_mask = (uint32_t)arg_info(op->args[1])->z_mask;
break;
case INDEX_op_extrh_i64_i32:
z_mask = (uint64_t)arg_info(op->args[1])->z_mask >> 32;
break;
CASE_OP_32_64(shl):
if (arg_is_const(op->args[2])) {
tmp = arg_info(op->args[2])->val & (TCG_TARGET_REG_BITS - 1);
z_mask = arg_info(op->args[1])->z_mask << tmp;
}
break;
CASE_OP_32_64(neg):
/* Set to 1 all bits to the left of the rightmost. */
z_mask = -(arg_info(op->args[1])->z_mask
& -arg_info(op->args[1])->z_mask);
break;
CASE_OP_32_64(deposit):
z_mask = deposit64(arg_info(op->args[1])->z_mask,
op->args[3], op->args[4],
arg_info(op->args[2])->z_mask);
break;
CASE_OP_32_64(extract):
z_mask = extract64(arg_info(op->args[1])->z_mask,
op->args[2], op->args[3]);
if (op->args[2] == 0) {
affected = arg_info(op->args[1])->z_mask & ~z_mask;
}
break;
CASE_OP_32_64(sextract):
z_mask = sextract64(arg_info(op->args[1])->z_mask,
op->args[2], op->args[3]);
if (op->args[2] == 0 && (tcg_target_long)z_mask >= 0) {
affected = arg_info(op->args[1])->z_mask & ~z_mask;
}
break;
CASE_OP_32_64(or):
CASE_OP_32_64(xor):
z_mask = arg_info(op->args[1])->z_mask
| arg_info(op->args[2])->z_mask;
break;
case INDEX_op_clz_i32:
case INDEX_op_ctz_i32:
z_mask = arg_info(op->args[2])->z_mask | 31;
break;
case INDEX_op_clz_i64:
case INDEX_op_ctz_i64:
z_mask = arg_info(op->args[2])->z_mask | 63;
break;
case INDEX_op_ctpop_i32:
z_mask = 32 | 31;
break;
case INDEX_op_ctpop_i64:
z_mask = 64 | 63;
break;
CASE_OP_32_64(setcond):
case INDEX_op_setcond2_i32:
z_mask = 1;
break;
CASE_OP_32_64(movcond):
z_mask = arg_info(op->args[3])->z_mask
| arg_info(op->args[4])->z_mask;
break;
CASE_OP_32_64(ld8u):
z_mask = 0xff;
break;
CASE_OP_32_64(ld16u):
z_mask = 0xffff;
break;
case INDEX_op_ld32u_i64:
z_mask = 0xffffffffu;
break;
CASE_OP_32_64(qemu_ld):
{
MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs];
MemOp mop = get_memop(oi);
if (!(mop & MO_SIGN)) {
z_mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1;
}
}
break;
CASE_OP_32_64(bswap16):
z_mask = arg_info(op->args[1])->z_mask;
if (z_mask <= 0xffff) {
op->args[2] |= TCG_BSWAP_IZ;
}
z_mask = bswap16(z_mask);
switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) {
case TCG_BSWAP_OZ:
break;
case TCG_BSWAP_OS:
z_mask = (int16_t)z_mask;
break;
default: /* undefined high bits */
z_mask |= MAKE_64BIT_MASK(16, 48);
break;
}
break;
case INDEX_op_bswap32_i64:
z_mask = arg_info(op->args[1])->z_mask;
if (z_mask <= 0xffffffffu) {
op->args[2] |= TCG_BSWAP_IZ;
}
z_mask = bswap32(z_mask);
switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) {
case TCG_BSWAP_OZ:
break;
case TCG_BSWAP_OS:
z_mask = (int32_t)z_mask;
break;
default: /* undefined high bits */
z_mask |= MAKE_64BIT_MASK(32, 32);
break;
}
break;
default:
break;
}
/* 32-bit ops generate 32-bit results. For the result is zero test
below, we can ignore high bits, but for further optimizations we
need to record that the high bits contain garbage. */
partmask = z_mask;
if (ctx.type == TCG_TYPE_I32) {
z_mask |= ~(tcg_target_ulong)0xffffffffu;
partmask &= 0xffffffffu;
affected &= 0xffffffffu;
}
ctx.z_mask = z_mask;
if (partmask == 0) {
tcg_opt_gen_movi(&ctx, op, op->args[0], 0);
continue;
}
if (affected == 0) {
tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]);
continue;
}
/* Assume all bits affected, and no bits known zero. */
ctx.a_mask = -1;
ctx.z_mask = -1;
/*
* Process each opcode.
@ -1873,6 +1911,11 @@ void tcg_optimize(TCGContext *s)
case INDEX_op_extrh_i64_i32:
done = fold_extu(&ctx, op);
break;
CASE_OP_32_64(ld8u):
CASE_OP_32_64(ld16u):
case INDEX_op_ld32u_i64:
done = fold_tcg_ld(&ctx, op);
break;
case INDEX_op_mb:
done = fold_mb(&ctx, op);
break;