qemu/tcg/optimize.c

1125 lines
34 KiB
C
Raw Normal View History

/*
* Optimizations for Tiny Code Generator for QEMU
*
* Copyright (c) 2010 Samsung Electronics.
* Contributed by Kirill Batuzov <batuzovk@ispras.ru>
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include "config.h"
#include <stdlib.h>
#include <stdio.h>
#include "qemu-common.h"
#include "tcg-op.h"
#define CASE_OP_32_64(x) \
glue(glue(case INDEX_op_, x), _i32): \
glue(glue(case INDEX_op_, x), _i64)
typedef enum {
TCG_TEMP_UNDEF = 0,
TCG_TEMP_CONST,
TCG_TEMP_COPY,
} tcg_temp_state;
struct tcg_temp_info {
tcg_temp_state state;
uint16_t prev_copy;
uint16_t next_copy;
tcg_target_ulong val;
tcg_target_ulong mask;
};
static struct tcg_temp_info temps[TCG_MAX_TEMPS];
/* Reset TEMP's state to TCG_TEMP_UNDEF. If TEMP only had one copy, remove
the copy flag from the left temp. */
static void reset_temp(TCGArg temp)
{
if (temps[temp].state == TCG_TEMP_COPY) {
if (temps[temp].prev_copy == temps[temp].next_copy) {
temps[temps[temp].next_copy].state = TCG_TEMP_UNDEF;
} else {
temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy;
temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy;
}
}
temps[temp].state = TCG_TEMP_UNDEF;
temps[temp].mask = -1;
}
/* Reset all temporaries, given that there are NB_TEMPS of them. */
static void reset_all_temps(int nb_temps)
{
int i;
for (i = 0; i < nb_temps; i++) {
temps[i].state = TCG_TEMP_UNDEF;
temps[i].mask = -1;
}
}
static int op_bits(TCGOpcode op)
{
const TCGOpDef *def = &tcg_op_defs[op];
return def->flags & TCG_OPF_64BIT ? 64 : 32;
}
static TCGOpcode op_to_movi(TCGOpcode op)
{
switch (op_bits(op)) {
case 32:
return INDEX_op_movi_i32;
case 64:
return INDEX_op_movi_i64;
default:
fprintf(stderr, "op_to_movi: unexpected return value of "
"function op_bits.\n");
tcg_abort();
}
}
static TCGArg find_better_copy(TCGContext *s, TCGArg temp)
{
TCGArg i;
/* If this is already a global, we can't do better. */
if (temp < s->nb_globals) {
return temp;
}
/* Search for a global first. */
for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
if (i < s->nb_globals) {
return i;
}
}
/* If it is a temp, search for a temp local. */
if (!s->temps[temp].temp_local) {
for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
if (s->temps[i].temp_local) {
return i;
}
}
}
/* Failure to find a better representation, return the same temp. */
return temp;
}
static bool temps_are_copies(TCGArg arg1, TCGArg arg2)
{
TCGArg i;
if (arg1 == arg2) {
return true;
}
if (temps[arg1].state != TCG_TEMP_COPY
|| temps[arg2].state != TCG_TEMP_COPY) {
return false;
}
for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) {
if (i == arg2) {
return true;
}
}
return false;
}
static void tcg_opt_gen_mov(TCGContext *s, TCGArg *gen_args,
TCGArg dst, TCGArg src)
{
reset_temp(dst);
temps[dst].mask = temps[src].mask;
assert(temps[src].state != TCG_TEMP_CONST);
if (s->temps[src].type == s->temps[dst].type) {
if (temps[src].state != TCG_TEMP_COPY) {
temps[src].state = TCG_TEMP_COPY;
temps[src].next_copy = src;
temps[src].prev_copy = src;
}
temps[dst].state = TCG_TEMP_COPY;
temps[dst].next_copy = temps[src].next_copy;
temps[dst].prev_copy = src;
temps[temps[dst].next_copy].prev_copy = dst;
temps[src].next_copy = dst;
}
gen_args[0] = dst;
gen_args[1] = src;
}
static void tcg_opt_gen_movi(TCGArg *gen_args, TCGArg dst, TCGArg val)
{
reset_temp(dst);
temps[dst].state = TCG_TEMP_CONST;
temps[dst].val = val;
temps[dst].mask = val;
gen_args[0] = dst;
gen_args[1] = val;
}
static TCGOpcode op_to_mov(TCGOpcode op)
{
switch (op_bits(op)) {
case 32:
return INDEX_op_mov_i32;
case 64:
return INDEX_op_mov_i64;
default:
fprintf(stderr, "op_to_mov: unexpected return value of "
"function op_bits.\n");
tcg_abort();
}
}
static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
{
switch (op) {
CASE_OP_32_64(add):
return x + y;
CASE_OP_32_64(sub):
return x - y;
CASE_OP_32_64(mul):
return x * y;
CASE_OP_32_64(and):
return x & y;
CASE_OP_32_64(or):
return x | y;
CASE_OP_32_64(xor):
return x ^ y;
case INDEX_op_shl_i32:
return (uint32_t)x << (uint32_t)y;
case INDEX_op_shl_i64:
return (uint64_t)x << (uint64_t)y;
case INDEX_op_shr_i32:
return (uint32_t)x >> (uint32_t)y;
case INDEX_op_shr_i64:
return (uint64_t)x >> (uint64_t)y;
case INDEX_op_sar_i32:
return (int32_t)x >> (int32_t)y;
case INDEX_op_sar_i64:
return (int64_t)x >> (int64_t)y;
case INDEX_op_rotr_i32:
x = ((uint32_t)x << (32 - y)) | ((uint32_t)x >> y);
return x;
case INDEX_op_rotr_i64:
x = ((uint64_t)x << (64 - y)) | ((uint64_t)x >> y);
return x;
case INDEX_op_rotl_i32:
x = ((uint32_t)x << y) | ((uint32_t)x >> (32 - y));
return x;
case INDEX_op_rotl_i64:
x = ((uint64_t)x << y) | ((uint64_t)x >> (64 - y));
return x;
CASE_OP_32_64(not):
return ~x;
CASE_OP_32_64(neg):
return -x;
CASE_OP_32_64(andc):
return x & ~y;
CASE_OP_32_64(orc):
return x | ~y;
CASE_OP_32_64(eqv):
return ~(x ^ y);
CASE_OP_32_64(nand):
return ~(x & y);
CASE_OP_32_64(nor):
return ~(x | y);
CASE_OP_32_64(ext8s):
return (int8_t)x;
CASE_OP_32_64(ext16s):
return (int16_t)x;
CASE_OP_32_64(ext8u):
return (uint8_t)x;
CASE_OP_32_64(ext16u):
return (uint16_t)x;
case INDEX_op_ext32s_i64:
return (int32_t)x;
case INDEX_op_ext32u_i64:
return (uint32_t)x;
default:
fprintf(stderr,
"Unrecognized operation %d in do_constant_folding.\n", op);
tcg_abort();
}
}
static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
{
TCGArg res = do_constant_folding_2(op, x, y);
if (op_bits(op) == 32) {
res &= 0xffffffff;
}
return res;
}
static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
{
switch (c) {
case TCG_COND_EQ:
return x == y;
case TCG_COND_NE:
return x != y;
case TCG_COND_LT:
return (int32_t)x < (int32_t)y;
case TCG_COND_GE:
return (int32_t)x >= (int32_t)y;
case TCG_COND_LE:
return (int32_t)x <= (int32_t)y;
case TCG_COND_GT:
return (int32_t)x > (int32_t)y;
case TCG_COND_LTU:
return x < y;
case TCG_COND_GEU:
return x >= y;
case TCG_COND_LEU:
return x <= y;
case TCG_COND_GTU:
return x > y;
default:
tcg_abort();
}
}
static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
{
switch (c) {
case TCG_COND_EQ:
return x == y;
case TCG_COND_NE:
return x != y;
case TCG_COND_LT:
return (int64_t)x < (int64_t)y;
case TCG_COND_GE:
return (int64_t)x >= (int64_t)y;
case TCG_COND_LE:
return (int64_t)x <= (int64_t)y;
case TCG_COND_GT:
return (int64_t)x > (int64_t)y;
case TCG_COND_LTU:
return x < y;
case TCG_COND_GEU:
return x >= y;
case TCG_COND_LEU:
return x <= y;
case TCG_COND_GTU:
return x > y;
default:
tcg_abort();
}
}
static bool do_constant_folding_cond_eq(TCGCond c)
{
switch (c) {
case TCG_COND_GT:
case TCG_COND_LTU:
case TCG_COND_LT:
case TCG_COND_GTU:
case TCG_COND_NE:
return 0;
case TCG_COND_GE:
case TCG_COND_GEU:
case TCG_COND_LE:
case TCG_COND_LEU:
case TCG_COND_EQ:
return 1;
default:
tcg_abort();
}
}
/* Return 2 if the condition can't be simplified, and the result
of the condition (0 or 1) if it can */
static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
TCGArg y, TCGCond c)
{
if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) {
switch (op_bits(op)) {
case 32:
return do_constant_folding_cond_32(temps[x].val, temps[y].val, c);
case 64:
return do_constant_folding_cond_64(temps[x].val, temps[y].val, c);
default:
tcg_abort();
}
} else if (temps_are_copies(x, y)) {
return do_constant_folding_cond_eq(c);
} else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) {
switch (c) {
case TCG_COND_LTU:
return 0;
case TCG_COND_GEU:
return 1;
default:
return 2;
}
} else {
return 2;
}
}
/* Return 2 if the condition can't be simplified, and the result
of the condition (0 or 1) if it can */
static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c)
{
TCGArg al = p1[0], ah = p1[1];
TCGArg bl = p2[0], bh = p2[1];
if (temps[bl].state == TCG_TEMP_CONST
&& temps[bh].state == TCG_TEMP_CONST) {
uint64_t b = ((uint64_t)temps[bh].val << 32) | (uint32_t)temps[bl].val;
if (temps[al].state == TCG_TEMP_CONST
&& temps[ah].state == TCG_TEMP_CONST) {
uint64_t a;
a = ((uint64_t)temps[ah].val << 32) | (uint32_t)temps[al].val;
return do_constant_folding_cond_64(a, b, c);
}
if (b == 0) {
switch (c) {
case TCG_COND_LTU:
return 0;
case TCG_COND_GEU:
return 1;
default:
break;
}
}
}
if (temps_are_copies(al, bl) && temps_are_copies(ah, bh)) {
return do_constant_folding_cond_eq(c);
}
return 2;
}
static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
{
TCGArg a1 = *p1, a2 = *p2;
int sum = 0;
sum += temps[a1].state == TCG_TEMP_CONST;
sum -= temps[a2].state == TCG_TEMP_CONST;
/* Prefer the constant in second argument, and then the form
op a, a, b, which is better handled on non-RISC hosts. */
if (sum > 0 || (sum == 0 && dest == a2)) {
*p1 = a2;
*p2 = a1;
return true;
}
return false;
}
static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
{
int sum = 0;
sum += temps[p1[0]].state == TCG_TEMP_CONST;
sum += temps[p1[1]].state == TCG_TEMP_CONST;
sum -= temps[p2[0]].state == TCG_TEMP_CONST;
sum -= temps[p2[1]].state == TCG_TEMP_CONST;
if (sum > 0) {
TCGArg t;
t = p1[0], p1[0] = p2[0], p2[0] = t;
t = p1[1], p1[1] = p2[1], p2[1] = t;
return true;
}
return false;
}
/* Propagate constants and copies, fold constant expressions. */
static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
TCGArg *args, TCGOpDef *tcg_op_defs)
{
int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args;
optimize: optimize using nonzero bits This adds two optimizations using the non-zero bit mask. In some cases involving shifts or ANDs the value can become zero, and can thus be optimized to a move of zero. Second, useless zero-extension or an AND with constant can be detected that would only zero bits that are already zero. The main advantage of this optimization is that it turns zero-extensions into moves, thus enabling much better copy propagation (around 1% code reduction). Here is for example a "test $0xff0000,%ecx + je" before optimization: mov_i64 tmp0,rcx movi_i64 tmp1,$0xff0000 discard cc_src and_i64 cc_dst,tmp0,tmp1 movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 and after (without patch on the left, with on the right): movi_i64 tmp1,$0xff0000 movi_i64 tmp1,$0xff0000 discard cc_src discard cc_src and_i64 cc_dst,rcx,tmp1 and_i64 cc_dst,rcx,tmp1 movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0 Other similar cases: "test %eax, %eax + jne" where eax is already 32-bit (after optimization, without patch on the left, with on the right): discard cc_src discard cc_src mov_i64 cc_dst,rax mov_i64 cc_dst,rax movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,ne,$0x0 brcond_i64 rax,tmp12,ne,$0x0 "test $0x1, %dl + je": movi_i64 tmp1,$0x1 movi_i64 tmp1,$0x1 discard cc_src discard cc_src and_i64 cc_dst,rdx,tmp1 and_i64 cc_dst,rdx,tmp1 movi_i32 cc_op,$0x1a movi_i32 cc_op,$0x1a ext8u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0 In some cases TCG even outsmarts GCC. :) Here the input code has "and $0x2,%eax + movslq %eax,%rbx + test %rbx, %rbx" and the optimizer, thanks to copy propagation, does the following: movi_i64 tmp12,$0x2 movi_i64 tmp12,$0x2 and_i64 rax,rax,tmp12 and_i64 rax,rax,tmp12 mov_i64 cc_dst,rax mov_i64 cc_dst,rax ext32s_i64 tmp0,rax -> nop mov_i64 rbx,tmp0 -> mov_i64 rbx,cc_dst and_i64 cc_dst,rbx,rbx -> nop Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
2013-01-12 03:42:53 +04:00
tcg_target_ulong mask, affected;
TCGOpcode op;
const TCGOpDef *def;
TCGArg *gen_args;
TCGArg tmp;
/* Array VALS has an element for each temp.
If this temp holds a constant then its value is kept in VALS' element.
If this temp is a copy of other ones then the other copies are
available through the doubly linked circular list. */
nb_temps = s->nb_temps;
nb_globals = s->nb_globals;
reset_all_temps(nb_temps);
nb_ops = tcg_opc_ptr - s->gen_opc_buf;
gen_args = args;
for (op_index = 0; op_index < nb_ops; op_index++) {
op = s->gen_opc_buf[op_index];
def = &tcg_op_defs[op];
/* Do copy propagation */
if (op == INDEX_op_call) {
int nb_oargs = args[0] >> 16;
int nb_iargs = args[0] & 0xffff;
for (i = nb_oargs + 1; i < nb_oargs + nb_iargs + 1; i++) {
if (temps[args[i]].state == TCG_TEMP_COPY) {
args[i] = find_better_copy(s, args[i]);
}
}
} else {
for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) {
if (temps[args[i]].state == TCG_TEMP_COPY) {
args[i] = find_better_copy(s, args[i]);
}
}
}
/* For commutative operations make constant second argument */
switch (op) {
CASE_OP_32_64(add):
CASE_OP_32_64(mul):
CASE_OP_32_64(and):
CASE_OP_32_64(or):
CASE_OP_32_64(xor):
CASE_OP_32_64(eqv):
CASE_OP_32_64(nand):
CASE_OP_32_64(nor):
swap_commutative(args[0], &args[1], &args[2]);
break;
CASE_OP_32_64(brcond):
if (swap_commutative(-1, &args[0], &args[1])) {
args[2] = tcg_swap_cond(args[2]);
}
break;
CASE_OP_32_64(setcond):
if (swap_commutative(args[0], &args[1], &args[2])) {
args[3] = tcg_swap_cond(args[3]);
}
break;
CASE_OP_32_64(movcond):
if (swap_commutative(-1, &args[1], &args[2])) {
args[5] = tcg_swap_cond(args[5]);
}
/* For movcond, we canonicalize the "false" input reg to match
the destination reg so that the tcg backend can implement
a "move if true" operation. */
if (swap_commutative(args[0], &args[4], &args[3])) {
args[5] = tcg_invert_cond(args[5]);
}
break;
CASE_OP_32_64(add2):
swap_commutative(args[0], &args[2], &args[4]);
swap_commutative(args[1], &args[3], &args[5]);
break;
CASE_OP_32_64(mulu2):
CASE_OP_32_64(muls2):
swap_commutative(args[0], &args[2], &args[3]);
break;
case INDEX_op_brcond2_i32:
if (swap_commutative2(&args[0], &args[2])) {
args[4] = tcg_swap_cond(args[4]);
}
break;
case INDEX_op_setcond2_i32:
if (swap_commutative2(&args[1], &args[3])) {
args[5] = tcg_swap_cond(args[5]);
}
break;
default:
break;
}
/* Simplify expressions for "shift/rot r, 0, a => movi r, 0",
and "sub r, 0, a => neg r, a" case. */
switch (op) {
CASE_OP_32_64(shl):
CASE_OP_32_64(shr):
CASE_OP_32_64(sar):
CASE_OP_32_64(rotl):
CASE_OP_32_64(rotr):
if (temps[args[1]].state == TCG_TEMP_CONST
&& temps[args[1]].val == 0) {
s->gen_opc_buf[op_index] = op_to_movi(op);
tcg_opt_gen_movi(gen_args, args[0], 0);
args += 3;
gen_args += 2;
continue;
}
break;
CASE_OP_32_64(sub):
{
TCGOpcode neg_op;
bool have_neg;
if (temps[args[2]].state == TCG_TEMP_CONST) {
/* Proceed with possible constant folding. */
break;
}
if (op == INDEX_op_sub_i32) {
neg_op = INDEX_op_neg_i32;
have_neg = TCG_TARGET_HAS_neg_i32;
} else {
neg_op = INDEX_op_neg_i64;
have_neg = TCG_TARGET_HAS_neg_i64;
}
if (!have_neg) {
break;
}
if (temps[args[1]].state == TCG_TEMP_CONST
&& temps[args[1]].val == 0) {
s->gen_opc_buf[op_index] = neg_op;
reset_temp(args[0]);
gen_args[0] = args[0];
gen_args[1] = args[2];
args += 3;
gen_args += 2;
continue;
}
}
break;
default:
break;
}
/* Simplify expression for "op r, a, 0 => mov r, a" cases */
switch (op) {
CASE_OP_32_64(add):
CASE_OP_32_64(sub):
CASE_OP_32_64(shl):
CASE_OP_32_64(shr):
CASE_OP_32_64(sar):
CASE_OP_32_64(rotl):
CASE_OP_32_64(rotr):
CASE_OP_32_64(or):
CASE_OP_32_64(xor):
if (temps[args[1]].state == TCG_TEMP_CONST) {
/* Proceed with possible constant folding. */
break;
}
if (temps[args[2]].state == TCG_TEMP_CONST
&& temps[args[2]].val == 0) {
if (temps_are_copies(args[0], args[1])) {
s->gen_opc_buf[op_index] = INDEX_op_nop;
} else {
s->gen_opc_buf[op_index] = op_to_mov(op);
tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
gen_args += 2;
}
args += 3;
continue;
}
break;
default:
break;
}
/* Simplify using known-zero bits */
mask = -1;
optimize: optimize using nonzero bits This adds two optimizations using the non-zero bit mask. In some cases involving shifts or ANDs the value can become zero, and can thus be optimized to a move of zero. Second, useless zero-extension or an AND with constant can be detected that would only zero bits that are already zero. The main advantage of this optimization is that it turns zero-extensions into moves, thus enabling much better copy propagation (around 1% code reduction). Here is for example a "test $0xff0000,%ecx + je" before optimization: mov_i64 tmp0,rcx movi_i64 tmp1,$0xff0000 discard cc_src and_i64 cc_dst,tmp0,tmp1 movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 and after (without patch on the left, with on the right): movi_i64 tmp1,$0xff0000 movi_i64 tmp1,$0xff0000 discard cc_src discard cc_src and_i64 cc_dst,rcx,tmp1 and_i64 cc_dst,rcx,tmp1 movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0 Other similar cases: "test %eax, %eax + jne" where eax is already 32-bit (after optimization, without patch on the left, with on the right): discard cc_src discard cc_src mov_i64 cc_dst,rax mov_i64 cc_dst,rax movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,ne,$0x0 brcond_i64 rax,tmp12,ne,$0x0 "test $0x1, %dl + je": movi_i64 tmp1,$0x1 movi_i64 tmp1,$0x1 discard cc_src discard cc_src and_i64 cc_dst,rdx,tmp1 and_i64 cc_dst,rdx,tmp1 movi_i32 cc_op,$0x1a movi_i32 cc_op,$0x1a ext8u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0 In some cases TCG even outsmarts GCC. :) Here the input code has "and $0x2,%eax + movslq %eax,%rbx + test %rbx, %rbx" and the optimizer, thanks to copy propagation, does the following: movi_i64 tmp12,$0x2 movi_i64 tmp12,$0x2 and_i64 rax,rax,tmp12 and_i64 rax,rax,tmp12 mov_i64 cc_dst,rax mov_i64 cc_dst,rax ext32s_i64 tmp0,rax -> nop mov_i64 rbx,tmp0 -> mov_i64 rbx,cc_dst and_i64 cc_dst,rbx,rbx -> nop Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
2013-01-12 03:42:53 +04:00
affected = -1;
switch (op) {
CASE_OP_32_64(ext8s):
if ((temps[args[1]].mask & 0x80) != 0) {
break;
}
CASE_OP_32_64(ext8u):
mask = 0xff;
goto and_const;
CASE_OP_32_64(ext16s):
if ((temps[args[1]].mask & 0x8000) != 0) {
break;
}
CASE_OP_32_64(ext16u):
mask = 0xffff;
goto and_const;
case INDEX_op_ext32s_i64:
if ((temps[args[1]].mask & 0x80000000) != 0) {
break;
}
case INDEX_op_ext32u_i64:
mask = 0xffffffffU;
goto and_const;
CASE_OP_32_64(and):
mask = temps[args[2]].mask;
if (temps[args[2]].state == TCG_TEMP_CONST) {
and_const:
optimize: optimize using nonzero bits This adds two optimizations using the non-zero bit mask. In some cases involving shifts or ANDs the value can become zero, and can thus be optimized to a move of zero. Second, useless zero-extension or an AND with constant can be detected that would only zero bits that are already zero. The main advantage of this optimization is that it turns zero-extensions into moves, thus enabling much better copy propagation (around 1% code reduction). Here is for example a "test $0xff0000,%ecx + je" before optimization: mov_i64 tmp0,rcx movi_i64 tmp1,$0xff0000 discard cc_src and_i64 cc_dst,tmp0,tmp1 movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 and after (without patch on the left, with on the right): movi_i64 tmp1,$0xff0000 movi_i64 tmp1,$0xff0000 discard cc_src discard cc_src and_i64 cc_dst,rcx,tmp1 and_i64 cc_dst,rcx,tmp1 movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0 Other similar cases: "test %eax, %eax + jne" where eax is already 32-bit (after optimization, without patch on the left, with on the right): discard cc_src discard cc_src mov_i64 cc_dst,rax mov_i64 cc_dst,rax movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,ne,$0x0 brcond_i64 rax,tmp12,ne,$0x0 "test $0x1, %dl + je": movi_i64 tmp1,$0x1 movi_i64 tmp1,$0x1 discard cc_src discard cc_src and_i64 cc_dst,rdx,tmp1 and_i64 cc_dst,rdx,tmp1 movi_i32 cc_op,$0x1a movi_i32 cc_op,$0x1a ext8u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0 In some cases TCG even outsmarts GCC. :) Here the input code has "and $0x2,%eax + movslq %eax,%rbx + test %rbx, %rbx" and the optimizer, thanks to copy propagation, does the following: movi_i64 tmp12,$0x2 movi_i64 tmp12,$0x2 and_i64 rax,rax,tmp12 and_i64 rax,rax,tmp12 mov_i64 cc_dst,rax mov_i64 cc_dst,rax ext32s_i64 tmp0,rax -> nop mov_i64 rbx,tmp0 -> mov_i64 rbx,cc_dst and_i64 cc_dst,rbx,rbx -> nop Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
2013-01-12 03:42:53 +04:00
affected = temps[args[1]].mask & ~mask;
}
mask = temps[args[1]].mask & mask;
break;
CASE_OP_32_64(sar):
if (temps[args[2]].state == TCG_TEMP_CONST) {
mask = ((tcg_target_long)temps[args[1]].mask
>> temps[args[2]].val);
}
break;
CASE_OP_32_64(shr):
if (temps[args[2]].state == TCG_TEMP_CONST) {
mask = temps[args[1]].mask >> temps[args[2]].val;
}
break;
CASE_OP_32_64(shl):
if (temps[args[2]].state == TCG_TEMP_CONST) {
mask = temps[args[1]].mask << temps[args[2]].val;
}
break;
CASE_OP_32_64(neg):
/* Set to 1 all bits to the left of the rightmost. */
mask = -(temps[args[1]].mask & -temps[args[1]].mask);
break;
CASE_OP_32_64(deposit):
tmp = ((1ull << args[4]) - 1);
mask = ((temps[args[1]].mask & ~(tmp << args[3]))
| ((temps[args[2]].mask & tmp) << args[3]));
break;
CASE_OP_32_64(or):
CASE_OP_32_64(xor):
mask = temps[args[1]].mask | temps[args[2]].mask;
break;
CASE_OP_32_64(setcond):
mask = 1;
break;
CASE_OP_32_64(movcond):
mask = temps[args[3]].mask | temps[args[4]].mask;
break;
default:
break;
}
optimize: optimize using nonzero bits This adds two optimizations using the non-zero bit mask. In some cases involving shifts or ANDs the value can become zero, and can thus be optimized to a move of zero. Second, useless zero-extension or an AND with constant can be detected that would only zero bits that are already zero. The main advantage of this optimization is that it turns zero-extensions into moves, thus enabling much better copy propagation (around 1% code reduction). Here is for example a "test $0xff0000,%ecx + je" before optimization: mov_i64 tmp0,rcx movi_i64 tmp1,$0xff0000 discard cc_src and_i64 cc_dst,tmp0,tmp1 movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 and after (without patch on the left, with on the right): movi_i64 tmp1,$0xff0000 movi_i64 tmp1,$0xff0000 discard cc_src discard cc_src and_i64 cc_dst,rcx,tmp1 and_i64 cc_dst,rcx,tmp1 movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0 Other similar cases: "test %eax, %eax + jne" where eax is already 32-bit (after optimization, without patch on the left, with on the right): discard cc_src discard cc_src mov_i64 cc_dst,rax mov_i64 cc_dst,rax movi_i32 cc_op,$0x1c movi_i32 cc_op,$0x1c ext32u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,ne,$0x0 brcond_i64 rax,tmp12,ne,$0x0 "test $0x1, %dl + je": movi_i64 tmp1,$0x1 movi_i64 tmp1,$0x1 discard cc_src discard cc_src and_i64 cc_dst,rdx,tmp1 and_i64 cc_dst,rdx,tmp1 movi_i32 cc_op,$0x1a movi_i32 cc_op,$0x1a ext8u_i64 tmp0,cc_dst movi_i64 tmp12,$0x0 movi_i64 tmp12,$0x0 brcond_i64 tmp0,tmp12,eq,$0x0 brcond_i64 cc_dst,tmp12,eq,$0x0 In some cases TCG even outsmarts GCC. :) Here the input code has "and $0x2,%eax + movslq %eax,%rbx + test %rbx, %rbx" and the optimizer, thanks to copy propagation, does the following: movi_i64 tmp12,$0x2 movi_i64 tmp12,$0x2 and_i64 rax,rax,tmp12 and_i64 rax,rax,tmp12 mov_i64 cc_dst,rax mov_i64 cc_dst,rax ext32s_i64 tmp0,rax -> nop mov_i64 rbx,tmp0 -> mov_i64 rbx,cc_dst and_i64 cc_dst,rbx,rbx -> nop Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
2013-01-12 03:42:53 +04:00
if (mask == 0) {
assert(def->nb_oargs == 1);
s->gen_opc_buf[op_index] = op_to_movi(op);
tcg_opt_gen_movi(gen_args, args[0], 0);
args += def->nb_oargs + def->nb_iargs + def->nb_cargs;
gen_args += 2;
continue;
}
if (affected == 0) {
assert(def->nb_oargs == 1);
if (temps_are_copies(args[0], args[1])) {
s->gen_opc_buf[op_index] = INDEX_op_nop;
} else if (temps[args[1]].state != TCG_TEMP_CONST) {
s->gen_opc_buf[op_index] = op_to_mov(op);
tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
gen_args += 2;
} else {
s->gen_opc_buf[op_index] = op_to_movi(op);
tcg_opt_gen_movi(gen_args, args[0], temps[args[1]].val);
gen_args += 2;
}
args += def->nb_iargs + 1;
continue;
}
/* Simplify expression for "op r, a, 0 => movi r, 0" cases */
switch (op) {
CASE_OP_32_64(and):
CASE_OP_32_64(mul):
if ((temps[args[2]].state == TCG_TEMP_CONST
&& temps[args[2]].val == 0)) {
s->gen_opc_buf[op_index] = op_to_movi(op);
tcg_opt_gen_movi(gen_args, args[0], 0);
args += 3;
gen_args += 2;
continue;
}
break;
default:
break;
}
/* Simplify expression for "op r, a, a => mov r, a" cases */
switch (op) {
CASE_OP_32_64(or):
CASE_OP_32_64(and):
if (temps_are_copies(args[1], args[2])) {
if (temps_are_copies(args[0], args[1])) {
s->gen_opc_buf[op_index] = INDEX_op_nop;
} else {
s->gen_opc_buf[op_index] = op_to_mov(op);
tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
gen_args += 2;
}
args += 3;
continue;
}
break;
default:
break;
}
/* Simplify expression for "op r, a, a => movi r, 0" cases */
switch (op) {
CASE_OP_32_64(sub):
CASE_OP_32_64(xor):
if (temps_are_copies(args[1], args[2])) {
s->gen_opc_buf[op_index] = op_to_movi(op);
tcg_opt_gen_movi(gen_args, args[0], 0);
gen_args += 2;
args += 3;
continue;
}
break;
default:
break;
}
/* Propagate constants through copy operations and do constant
folding. Constants will be substituted to arguments by register
allocator where needed and possible. Also detect copies. */
switch (op) {
CASE_OP_32_64(mov):
if (temps_are_copies(args[0], args[1])) {
args += 2;
s->gen_opc_buf[op_index] = INDEX_op_nop;
break;
}
if (temps[args[1]].state != TCG_TEMP_CONST) {
tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
gen_args += 2;
args += 2;
break;
}
/* Source argument is constant. Rewrite the operation and
let movi case handle it. */
op = op_to_movi(op);
s->gen_opc_buf[op_index] = op;
args[1] = temps[args[1]].val;
/* fallthrough */
CASE_OP_32_64(movi):
tcg_opt_gen_movi(gen_args, args[0], args[1]);
gen_args += 2;
args += 2;
break;
CASE_OP_32_64(not):
CASE_OP_32_64(neg):
CASE_OP_32_64(ext8s):
CASE_OP_32_64(ext8u):
CASE_OP_32_64(ext16s):
CASE_OP_32_64(ext16u):
case INDEX_op_ext32s_i64:
case INDEX_op_ext32u_i64:
if (temps[args[1]].state == TCG_TEMP_CONST) {
s->gen_opc_buf[op_index] = op_to_movi(op);
tmp = do_constant_folding(op, temps[args[1]].val, 0);
tcg_opt_gen_movi(gen_args, args[0], tmp);
gen_args += 2;
args += 2;
break;
}
goto do_default;
CASE_OP_32_64(add):
CASE_OP_32_64(sub):
CASE_OP_32_64(mul):
CASE_OP_32_64(or):
CASE_OP_32_64(and):
CASE_OP_32_64(xor):
CASE_OP_32_64(shl):
CASE_OP_32_64(shr):
CASE_OP_32_64(sar):
CASE_OP_32_64(rotl):
CASE_OP_32_64(rotr):
CASE_OP_32_64(andc):
CASE_OP_32_64(orc):
CASE_OP_32_64(eqv):
CASE_OP_32_64(nand):
CASE_OP_32_64(nor):
if (temps[args[1]].state == TCG_TEMP_CONST
&& temps[args[2]].state == TCG_TEMP_CONST) {
s->gen_opc_buf[op_index] = op_to_movi(op);
tmp = do_constant_folding(op, temps[args[1]].val,
temps[args[2]].val);
tcg_opt_gen_movi(gen_args, args[0], tmp);
gen_args += 2;
args += 3;
break;
}
goto do_default;
CASE_OP_32_64(deposit):
if (temps[args[1]].state == TCG_TEMP_CONST
&& temps[args[2]].state == TCG_TEMP_CONST) {
s->gen_opc_buf[op_index] = op_to_movi(op);
tmp = ((1ull << args[4]) - 1);
tmp = (temps[args[1]].val & ~(tmp << args[3]))
| ((temps[args[2]].val & tmp) << args[3]);
tcg_opt_gen_movi(gen_args, args[0], tmp);
gen_args += 2;
args += 5;
break;
}
goto do_default;
CASE_OP_32_64(setcond):
tmp = do_constant_folding_cond(op, args[1], args[2], args[3]);
if (tmp != 2) {
s->gen_opc_buf[op_index] = op_to_movi(op);
tcg_opt_gen_movi(gen_args, args[0], tmp);
gen_args += 2;
args += 4;
break;
}
goto do_default;
CASE_OP_32_64(brcond):
tmp = do_constant_folding_cond(op, args[0], args[1], args[2]);
if (tmp != 2) {
if (tmp) {
reset_all_temps(nb_temps);
s->gen_opc_buf[op_index] = INDEX_op_br;
gen_args[0] = args[3];
gen_args += 1;
} else {
s->gen_opc_buf[op_index] = INDEX_op_nop;
}
args += 4;
break;
}
goto do_default;
CASE_OP_32_64(movcond):
tmp = do_constant_folding_cond(op, args[1], args[2], args[5]);
if (tmp != 2) {
if (temps_are_copies(args[0], args[4-tmp])) {
s->gen_opc_buf[op_index] = INDEX_op_nop;
} else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) {
s->gen_opc_buf[op_index] = op_to_movi(op);
tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val);
gen_args += 2;
} else {
s->gen_opc_buf[op_index] = op_to_mov(op);
tcg_opt_gen_mov(s, gen_args, args[0], args[4-tmp]);
gen_args += 2;
}
args += 6;
break;
}
goto do_default;
case INDEX_op_add2_i32:
case INDEX_op_sub2_i32:
if (temps[args[2]].state == TCG_TEMP_CONST
&& temps[args[3]].state == TCG_TEMP_CONST
&& temps[args[4]].state == TCG_TEMP_CONST
&& temps[args[5]].state == TCG_TEMP_CONST) {
uint32_t al = temps[args[2]].val;
uint32_t ah = temps[args[3]].val;
uint32_t bl = temps[args[4]].val;
uint32_t bh = temps[args[5]].val;
uint64_t a = ((uint64_t)ah << 32) | al;
uint64_t b = ((uint64_t)bh << 32) | bl;
TCGArg rl, rh;
if (op == INDEX_op_add2_i32) {
a += b;
} else {
a -= b;
}
/* We emit the extra nop when we emit the add2/sub2. */
assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop);
rl = args[0];
rh = args[1];
s->gen_opc_buf[op_index] = INDEX_op_movi_i32;
s->gen_opc_buf[++op_index] = INDEX_op_movi_i32;
tcg_opt_gen_movi(&gen_args[0], rl, (uint32_t)a);
tcg_opt_gen_movi(&gen_args[2], rh, (uint32_t)(a >> 32));
gen_args += 4;
args += 6;
break;
}
goto do_default;
case INDEX_op_mulu2_i32:
if (temps[args[2]].state == TCG_TEMP_CONST
&& temps[args[3]].state == TCG_TEMP_CONST) {
uint32_t a = temps[args[2]].val;
uint32_t b = temps[args[3]].val;
uint64_t r = (uint64_t)a * b;
TCGArg rl, rh;
/* We emit the extra nop when we emit the mulu2. */
assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop);
rl = args[0];
rh = args[1];
s->gen_opc_buf[op_index] = INDEX_op_movi_i32;
s->gen_opc_buf[++op_index] = INDEX_op_movi_i32;
tcg_opt_gen_movi(&gen_args[0], rl, (uint32_t)r);
tcg_opt_gen_movi(&gen_args[2], rh, (uint32_t)(r >> 32));
gen_args += 4;
args += 4;
break;
}
goto do_default;
case INDEX_op_brcond2_i32:
tmp = do_constant_folding_cond2(&args[0], &args[2], args[4]);
if (tmp != 2) {
if (tmp) {
reset_all_temps(nb_temps);
s->gen_opc_buf[op_index] = INDEX_op_br;
gen_args[0] = args[5];
gen_args += 1;
} else {
s->gen_opc_buf[op_index] = INDEX_op_nop;
}
} else if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE)
&& temps[args[2]].state == TCG_TEMP_CONST
&& temps[args[3]].state == TCG_TEMP_CONST
&& temps[args[2]].val == 0
&& temps[args[3]].val == 0) {
/* Simplify LT/GE comparisons vs zero to a single compare
vs the high word of the input. */
reset_all_temps(nb_temps);
s->gen_opc_buf[op_index] = INDEX_op_brcond_i32;
gen_args[0] = args[1];
gen_args[1] = args[3];
gen_args[2] = args[4];
gen_args[3] = args[5];
gen_args += 4;
} else {
goto do_default;
}
args += 6;
break;
case INDEX_op_setcond2_i32:
tmp = do_constant_folding_cond2(&args[1], &args[3], args[5]);
if (tmp != 2) {
s->gen_opc_buf[op_index] = INDEX_op_movi_i32;
tcg_opt_gen_movi(gen_args, args[0], tmp);
gen_args += 2;
} else if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE)
&& temps[args[3]].state == TCG_TEMP_CONST
&& temps[args[4]].state == TCG_TEMP_CONST
&& temps[args[3]].val == 0
&& temps[args[4]].val == 0) {
/* Simplify LT/GE comparisons vs zero to a single compare
vs the high word of the input. */
s->gen_opc_buf[op_index] = INDEX_op_setcond_i32;
reset_temp(args[0]);
gen_args[0] = args[0];
gen_args[1] = args[2];
gen_args[2] = args[4];
gen_args[3] = args[5];
gen_args += 4;
} else {
goto do_default;
}
args += 6;
break;
case INDEX_op_call:
nb_call_args = (args[0] >> 16) + (args[0] & 0xffff);
if (!(args[nb_call_args + 1] & (TCG_CALL_NO_READ_GLOBALS |
TCG_CALL_NO_WRITE_GLOBALS))) {
for (i = 0; i < nb_globals; i++) {
reset_temp(i);
}
}
for (i = 0; i < (args[0] >> 16); i++) {
reset_temp(args[i + 1]);
}
i = nb_call_args + 3;
while (i) {
*gen_args = *args;
args++;
gen_args++;
i--;
}
break;
default:
do_default:
/* Default case: we know nothing about operation (or were unable
to compute the operation result) so no propagation is done.
We trash everything if the operation is the end of a basic
block, otherwise we only trash the output args. "mask" is
the non-zero bits mask for the first output arg. */
if (def->flags & TCG_OPF_BB_END) {
reset_all_temps(nb_temps);
} else {
for (i = 0; i < def->nb_oargs; i++) {
reset_temp(args[i]);
}
}
for (i = 0; i < def->nb_args; i++) {
gen_args[i] = args[i];
}
args += def->nb_args;
gen_args += def->nb_args;
break;
}
}
return gen_args;
}
TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr,
TCGArg *args, TCGOpDef *tcg_op_defs)
{
TCGArg *res;
res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs);
return res;
}