2011-07-07 16:37:12 +04:00
|
|
|
/*
|
|
|
|
* 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.
|
|
|
|
*/
|
|
|
|
|
2016-01-26 21:17:08 +03:00
|
|
|
#include "qemu/osdep.h"
|
2011-07-07 16:37:12 +04:00
|
|
|
#include "qemu-common.h"
|
2016-03-15 15:16:36 +03:00
|
|
|
#include "exec/cpu-common.h"
|
2011-07-07 16:37:12 +04:00
|
|
|
#include "tcg-op.h"
|
|
|
|
|
|
|
|
#define CASE_OP_32_64(x) \
|
|
|
|
glue(glue(case INDEX_op_, x), _i32): \
|
|
|
|
glue(glue(case INDEX_op_, x), _i64)
|
|
|
|
|
2011-07-07 16:37:13 +04:00
|
|
|
struct tcg_temp_info {
|
2015-07-27 13:41:44 +03:00
|
|
|
bool is_const;
|
2017-06-20 23:43:15 +03:00
|
|
|
TCGTemp *prev_copy;
|
|
|
|
TCGTemp *next_copy;
|
2011-07-07 16:37:13 +04:00
|
|
|
tcg_target_ulong val;
|
2013-01-12 03:42:52 +04:00
|
|
|
tcg_target_ulong mask;
|
2011-07-07 16:37:13 +04:00
|
|
|
};
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
static inline struct tcg_temp_info *ts_info(TCGTemp *ts)
|
2015-07-27 13:41:44 +03:00
|
|
|
{
|
2017-06-20 23:43:15 +03:00
|
|
|
return ts->state_ptr;
|
2015-07-27 13:41:44 +03:00
|
|
|
}
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
static inline struct tcg_temp_info *arg_info(TCGArg arg)
|
2015-07-27 13:41:44 +03:00
|
|
|
{
|
2017-06-20 23:43:15 +03:00
|
|
|
return ts_info(arg_temp(arg));
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline bool ts_is_const(TCGTemp *ts)
|
|
|
|
{
|
|
|
|
return ts_info(ts)->is_const;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline bool arg_is_const(TCGArg arg)
|
|
|
|
{
|
|
|
|
return ts_is_const(arg_temp(arg));
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline bool ts_is_copy(TCGTemp *ts)
|
|
|
|
{
|
|
|
|
return ts_info(ts)->next_copy != ts;
|
2015-07-27 13:41:44 +03:00
|
|
|
}
|
|
|
|
|
2015-07-27 13:41:44 +03:00
|
|
|
/* Reset TEMP's state, possibly removing the temp for the list of copies. */
|
2017-06-20 23:43:15 +03:00
|
|
|
static void reset_ts(TCGTemp *ts)
|
|
|
|
{
|
|
|
|
struct tcg_temp_info *ti = ts_info(ts);
|
|
|
|
struct tcg_temp_info *pi = ts_info(ti->prev_copy);
|
|
|
|
struct tcg_temp_info *ni = ts_info(ti->next_copy);
|
|
|
|
|
|
|
|
ni->prev_copy = ti->prev_copy;
|
|
|
|
pi->next_copy = ti->next_copy;
|
|
|
|
ti->next_copy = ts;
|
|
|
|
ti->prev_copy = ts;
|
|
|
|
ti->is_const = false;
|
|
|
|
ti->mask = -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void reset_temp(TCGArg arg)
|
2011-07-07 16:37:13 +04:00
|
|
|
{
|
2017-06-20 23:43:15 +03:00
|
|
|
reset_ts(arg_temp(arg));
|
2011-07-07 16:37:13 +04:00
|
|
|
}
|
|
|
|
|
2015-07-27 13:41:44 +03:00
|
|
|
/* Initialize and activate a temporary. */
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
static void init_ts_info(struct tcg_temp_info *infos,
|
|
|
|
TCGTempSet *temps_used, TCGTemp *ts)
|
2015-07-27 13:41:44 +03:00
|
|
|
{
|
2017-06-20 23:43:15 +03:00
|
|
|
size_t idx = temp_idx(ts);
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
if (!test_bit(idx, temps_used->l)) {
|
|
|
|
struct tcg_temp_info *ti = &infos[idx];
|
2017-06-20 23:43:15 +03:00
|
|
|
|
|
|
|
ts->state_ptr = ti;
|
|
|
|
ti->next_copy = ts;
|
|
|
|
ti->prev_copy = ts;
|
|
|
|
ti->is_const = false;
|
|
|
|
ti->mask = -1;
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
set_bit(idx, temps_used->l);
|
2015-07-27 13:41:44 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
static void init_arg_info(struct tcg_temp_info *infos,
|
|
|
|
TCGTempSet *temps_used, TCGArg arg)
|
2017-06-20 23:43:15 +03:00
|
|
|
{
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
init_ts_info(infos, temps_used, arg_temp(arg));
|
2017-06-20 23:43:15 +03:00
|
|
|
}
|
|
|
|
|
2011-07-30 23:18:32 +04:00
|
|
|
static int op_bits(TCGOpcode op)
|
2011-07-07 16:37:13 +04:00
|
|
|
{
|
2011-08-18 01:11:45 +04:00
|
|
|
const TCGOpDef *def = &tcg_op_defs[op];
|
|
|
|
return def->flags & TCG_OPF_64BIT ? 64 : 32;
|
2011-07-07 16:37:13 +04:00
|
|
|
}
|
|
|
|
|
2014-05-22 21:59:12 +04:00
|
|
|
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();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-07-30 23:18:32 +04:00
|
|
|
static TCGOpcode op_to_movi(TCGOpcode op)
|
2011-07-07 16:37:13 +04:00
|
|
|
{
|
|
|
|
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();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
static TCGTemp *find_better_copy(TCGContext *s, TCGTemp *ts)
|
2012-09-11 14:31:21 +04:00
|
|
|
{
|
2017-06-20 23:43:15 +03:00
|
|
|
TCGTemp *i;
|
2012-09-11 14:31:21 +04:00
|
|
|
|
|
|
|
/* If this is already a global, we can't do better. */
|
2016-11-02 20:20:15 +03:00
|
|
|
if (ts->temp_global) {
|
2017-06-20 23:43:15 +03:00
|
|
|
return ts;
|
2012-09-11 14:31:21 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Search for a global first. */
|
2017-06-20 23:43:15 +03:00
|
|
|
for (i = ts_info(ts)->next_copy; i != ts; i = ts_info(i)->next_copy) {
|
|
|
|
if (i->temp_global) {
|
2012-09-11 14:31:21 +04:00
|
|
|
return i;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* If it is a temp, search for a temp local. */
|
2016-11-02 20:20:15 +03:00
|
|
|
if (!ts->temp_local) {
|
2017-06-20 23:43:15 +03:00
|
|
|
for (i = ts_info(ts)->next_copy; i != ts; i = ts_info(i)->next_copy) {
|
|
|
|
if (ts->temp_local) {
|
2012-09-11 14:31:21 +04:00
|
|
|
return i;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Failure to find a better representation, return the same temp. */
|
2017-06-20 23:43:15 +03:00
|
|
|
return ts;
|
2012-09-11 14:31:21 +04:00
|
|
|
}
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
static bool ts_are_copies(TCGTemp *ts1, TCGTemp *ts2)
|
2012-09-11 14:31:21 +04:00
|
|
|
{
|
2017-06-20 23:43:15 +03:00
|
|
|
TCGTemp *i;
|
2012-09-11 14:31:21 +04:00
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
if (ts1 == ts2) {
|
2012-09-11 14:31:21 +04:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
if (!ts_is_copy(ts1) || !ts_is_copy(ts2)) {
|
2012-09-11 14:31:21 +04:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
for (i = ts_info(ts1)->next_copy; i != ts1; i = ts_info(i)->next_copy) {
|
|
|
|
if (i == ts2) {
|
2012-09-11 14:31:21 +04:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
static bool args_are_copies(TCGArg arg1, TCGArg arg2)
|
|
|
|
{
|
|
|
|
return ts_are_copies(arg_temp(arg1), arg_temp(arg2));
|
|
|
|
}
|
|
|
|
|
2016-12-08 23:28:42 +03:00
|
|
|
static void tcg_opt_gen_movi(TCGContext *s, TCGOp *op, TCGArg dst, TCGArg val)
|
2015-06-05 12:19:18 +03:00
|
|
|
{
|
|
|
|
TCGOpcode new_op = op_to_movi(op->opc);
|
|
|
|
tcg_target_ulong mask;
|
2017-06-20 23:43:15 +03:00
|
|
|
struct tcg_temp_info *di = arg_info(dst);
|
2015-06-05 12:19:18 +03:00
|
|
|
|
|
|
|
op->opc = new_op;
|
|
|
|
|
|
|
|
reset_temp(dst);
|
2017-06-20 23:43:15 +03:00
|
|
|
di->is_const = true;
|
|
|
|
di->val = val;
|
2015-06-05 12:19:18 +03:00
|
|
|
mask = val;
|
2015-07-10 19:03:30 +03:00
|
|
|
if (TCG_TARGET_REG_BITS > 32 && new_op == INDEX_op_movi_i32) {
|
2015-06-05 12:19:18 +03:00
|
|
|
/* High bits of the destination are now garbage. */
|
|
|
|
mask |= ~0xffffffffull;
|
|
|
|
}
|
2017-06-20 23:43:15 +03:00
|
|
|
di->mask = mask;
|
2015-06-05 12:19:18 +03:00
|
|
|
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[0] = dst;
|
|
|
|
op->args[1] = val;
|
2015-06-05 12:19:18 +03:00
|
|
|
}
|
|
|
|
|
2016-12-08 23:28:42 +03:00
|
|
|
static void tcg_opt_gen_mov(TCGContext *s, TCGOp *op, TCGArg dst, TCGArg src)
|
2011-07-07 16:37:13 +04:00
|
|
|
{
|
2017-06-20 23:43:15 +03:00
|
|
|
TCGTemp *dst_ts = arg_temp(dst);
|
|
|
|
TCGTemp *src_ts = arg_temp(src);
|
|
|
|
struct tcg_temp_info *di;
|
|
|
|
struct tcg_temp_info *si;
|
|
|
|
tcg_target_ulong mask;
|
|
|
|
TCGOpcode new_op;
|
|
|
|
|
|
|
|
if (ts_are_copies(dst_ts, src_ts)) {
|
2015-06-04 22:53:25 +03:00
|
|
|
tcg_op_remove(s, op);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
reset_ts(dst_ts);
|
|
|
|
di = ts_info(dst_ts);
|
|
|
|
si = ts_info(src_ts);
|
|
|
|
new_op = op_to_mov(op->opc);
|
2014-05-22 21:59:12 +04:00
|
|
|
|
2014-09-20 00:49:15 +04:00
|
|
|
op->opc = new_op;
|
2017-06-20 23:43:15 +03:00
|
|
|
op->args[0] = dst;
|
|
|
|
op->args[1] = src;
|
2014-05-22 21:59:12 +04:00
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = si->mask;
|
2014-05-22 22:14:10 +04:00
|
|
|
if (TCG_TARGET_REG_BITS > 32 && new_op == INDEX_op_mov_i32) {
|
|
|
|
/* High bits of the destination are now garbage. */
|
|
|
|
mask |= ~0xffffffffull;
|
|
|
|
}
|
2017-06-20 23:43:15 +03:00
|
|
|
di->mask = mask;
|
2012-09-11 14:31:21 +04:00
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
if (src_ts->type == dst_ts->type) {
|
|
|
|
struct tcg_temp_info *ni = ts_info(si->next_copy);
|
|
|
|
|
|
|
|
di->next_copy = si->next_copy;
|
|
|
|
di->prev_copy = src_ts;
|
|
|
|
ni->prev_copy = dst_ts;
|
|
|
|
si->next_copy = dst_ts;
|
|
|
|
di->is_const = si->is_const;
|
|
|
|
di->val = si->val;
|
|
|
|
}
|
2011-07-07 16:37:13 +04:00
|
|
|
}
|
|
|
|
|
2011-07-30 23:18:32 +04:00
|
|
|
static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
|
2011-07-07 16:37:14 +04:00
|
|
|
{
|
2013-08-15 01:35:56 +04:00
|
|
|
uint64_t l64, h64;
|
|
|
|
|
2011-07-07 16:37:14 +04:00
|
|
|
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;
|
|
|
|
|
2011-07-07 16:37:15 +04:00
|
|
|
CASE_OP_32_64(and):
|
|
|
|
return x & y;
|
|
|
|
|
|
|
|
CASE_OP_32_64(or):
|
|
|
|
return x | y;
|
|
|
|
|
|
|
|
CASE_OP_32_64(xor):
|
|
|
|
return x ^ y;
|
|
|
|
|
2011-07-07 16:37:16 +04:00
|
|
|
case INDEX_op_shl_i32:
|
2014-03-18 18:45:39 +04:00
|
|
|
return (uint32_t)x << (y & 31);
|
2011-07-07 16:37:16 +04:00
|
|
|
|
|
|
|
case INDEX_op_shl_i64:
|
2014-03-18 18:45:39 +04:00
|
|
|
return (uint64_t)x << (y & 63);
|
2011-07-07 16:37:16 +04:00
|
|
|
|
|
|
|
case INDEX_op_shr_i32:
|
2014-03-18 18:45:39 +04:00
|
|
|
return (uint32_t)x >> (y & 31);
|
2011-07-07 16:37:16 +04:00
|
|
|
|
|
|
|
case INDEX_op_shr_i64:
|
2014-03-18 18:45:39 +04:00
|
|
|
return (uint64_t)x >> (y & 63);
|
2011-07-07 16:37:16 +04:00
|
|
|
|
|
|
|
case INDEX_op_sar_i32:
|
2014-03-18 18:45:39 +04:00
|
|
|
return (int32_t)x >> (y & 31);
|
2011-07-07 16:37:16 +04:00
|
|
|
|
|
|
|
case INDEX_op_sar_i64:
|
2014-03-18 18:45:39 +04:00
|
|
|
return (int64_t)x >> (y & 63);
|
2011-07-07 16:37:16 +04:00
|
|
|
|
|
|
|
case INDEX_op_rotr_i32:
|
2014-03-18 18:45:39 +04:00
|
|
|
return ror32(x, y & 31);
|
2011-07-07 16:37:16 +04:00
|
|
|
|
|
|
|
case INDEX_op_rotr_i64:
|
2014-03-18 18:45:39 +04:00
|
|
|
return ror64(x, y & 63);
|
2011-07-07 16:37:16 +04:00
|
|
|
|
|
|
|
case INDEX_op_rotl_i32:
|
2014-03-18 18:45:39 +04:00
|
|
|
return rol32(x, y & 31);
|
2011-07-07 16:37:16 +04:00
|
|
|
|
|
|
|
case INDEX_op_rotl_i64:
|
2014-03-18 18:45:39 +04:00
|
|
|
return rol64(x, y & 63);
|
2011-08-18 01:11:46 +04:00
|
|
|
|
|
|
|
CASE_OP_32_64(not):
|
2011-07-07 16:37:17 +04:00
|
|
|
return ~x;
|
2011-08-18 01:11:46 +04:00
|
|
|
|
2011-08-18 01:11:47 +04:00
|
|
|
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);
|
|
|
|
|
2016-11-16 11:23:28 +03:00
|
|
|
case INDEX_op_clz_i32:
|
|
|
|
return (uint32_t)x ? clz32(x) : y;
|
|
|
|
|
|
|
|
case INDEX_op_clz_i64:
|
|
|
|
return x ? clz64(x) : y;
|
|
|
|
|
|
|
|
case INDEX_op_ctz_i32:
|
|
|
|
return (uint32_t)x ? ctz32(x) : y;
|
|
|
|
|
|
|
|
case INDEX_op_ctz_i64:
|
|
|
|
return x ? ctz64(x) : y;
|
|
|
|
|
2016-11-21 13:13:39 +03:00
|
|
|
case INDEX_op_ctpop_i32:
|
|
|
|
return ctpop32(x);
|
|
|
|
|
|
|
|
case INDEX_op_ctpop_i64:
|
|
|
|
return ctpop64(x);
|
|
|
|
|
2011-08-18 01:11:46 +04:00
|
|
|
CASE_OP_32_64(ext8s):
|
2011-07-07 16:37:17 +04:00
|
|
|
return (int8_t)x;
|
2011-08-18 01:11:46 +04:00
|
|
|
|
|
|
|
CASE_OP_32_64(ext16s):
|
2011-07-07 16:37:17 +04:00
|
|
|
return (int16_t)x;
|
2011-08-18 01:11:46 +04:00
|
|
|
|
|
|
|
CASE_OP_32_64(ext8u):
|
2011-07-07 16:37:17 +04:00
|
|
|
return (uint8_t)x;
|
2011-08-18 01:11:46 +04:00
|
|
|
|
|
|
|
CASE_OP_32_64(ext16u):
|
2011-07-07 16:37:17 +04:00
|
|
|
return (uint16_t)x;
|
|
|
|
|
2015-07-27 13:41:45 +03:00
|
|
|
case INDEX_op_ext_i32_i64:
|
2011-07-07 16:37:17 +04:00
|
|
|
case INDEX_op_ext32s_i64:
|
|
|
|
return (int32_t)x;
|
|
|
|
|
2015-07-27 13:41:45 +03:00
|
|
|
case INDEX_op_extu_i32_i64:
|
2015-07-24 17:16:00 +03:00
|
|
|
case INDEX_op_extrl_i64_i32:
|
2011-07-07 16:37:17 +04:00
|
|
|
case INDEX_op_ext32u_i64:
|
|
|
|
return (uint32_t)x;
|
|
|
|
|
2015-07-24 17:16:00 +03:00
|
|
|
case INDEX_op_extrh_i64_i32:
|
|
|
|
return (uint64_t)x >> 32;
|
|
|
|
|
2013-08-15 01:35:56 +04:00
|
|
|
case INDEX_op_muluh_i32:
|
|
|
|
return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
|
|
|
|
case INDEX_op_mulsh_i32:
|
|
|
|
return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
|
|
|
|
|
|
|
|
case INDEX_op_muluh_i64:
|
|
|
|
mulu64(&l64, &h64, x, y);
|
|
|
|
return h64;
|
|
|
|
case INDEX_op_mulsh_i64:
|
|
|
|
muls64(&l64, &h64, x, y);
|
|
|
|
return h64;
|
|
|
|
|
2013-08-15 02:22:46 +04:00
|
|
|
case INDEX_op_div_i32:
|
|
|
|
/* Avoid crashing on divide by zero, otherwise undefined. */
|
|
|
|
return (int32_t)x / ((int32_t)y ? : 1);
|
|
|
|
case INDEX_op_divu_i32:
|
|
|
|
return (uint32_t)x / ((uint32_t)y ? : 1);
|
|
|
|
case INDEX_op_div_i64:
|
|
|
|
return (int64_t)x / ((int64_t)y ? : 1);
|
|
|
|
case INDEX_op_divu_i64:
|
|
|
|
return (uint64_t)x / ((uint64_t)y ? : 1);
|
|
|
|
|
|
|
|
case INDEX_op_rem_i32:
|
|
|
|
return (int32_t)x % ((int32_t)y ? : 1);
|
|
|
|
case INDEX_op_remu_i32:
|
|
|
|
return (uint32_t)x % ((uint32_t)y ? : 1);
|
|
|
|
case INDEX_op_rem_i64:
|
|
|
|
return (int64_t)x % ((int64_t)y ? : 1);
|
|
|
|
case INDEX_op_remu_i64:
|
|
|
|
return (uint64_t)x % ((uint64_t)y ? : 1);
|
|
|
|
|
2011-07-07 16:37:14 +04:00
|
|
|
default:
|
|
|
|
fprintf(stderr,
|
|
|
|
"Unrecognized operation %d in do_constant_folding.\n", op);
|
|
|
|
tcg_abort();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-07-30 23:18:32 +04:00
|
|
|
static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
|
2011-07-07 16:37:14 +04:00
|
|
|
{
|
|
|
|
TCGArg res = do_constant_folding_2(op, x, y);
|
|
|
|
if (op_bits(op) == 32) {
|
tcg/optimize: fix constant signedness
By convention, on a 64-bit host TCG internally stores 32-bit constants
as sign-extended. This is not the case in the optimizer when a 32-bit
constant is folded.
This doesn't seem to have more consequences than suboptimal code
generation. For instance the x86 backend assumes sign-extended constants,
and in some rare cases uses a 32-bit unsigned immediate 0xffffffff
instead of a 8-bit signed immediate 0xff for the constant -1. This is
with a ppc guest:
before
------
---- 0x9f29cc
movi_i32 tmp1,$0xffffffff
movi_i32 tmp2,$0x0
add2_i32 tmp0,CA,CA,tmp2,r6,tmp2
add2_i32 tmp0,CA,tmp0,CA,tmp1,tmp2
mov_i32 r10,tmp0
0x7fd8c7dfe90c: xor %ebp,%ebp
0x7fd8c7dfe90e: mov %ebp,%r11d
0x7fd8c7dfe911: mov 0x18(%r14),%r9d
0x7fd8c7dfe915: add %r9d,%r10d
0x7fd8c7dfe918: adc %ebp,%r11d
0x7fd8c7dfe91b: add $0xffffffff,%r10d
0x7fd8c7dfe922: adc %ebp,%r11d
0x7fd8c7dfe925: mov %r11d,0x134(%r14)
0x7fd8c7dfe92c: mov %r10d,0x28(%r14)
after
-----
---- 0x9f29cc
movi_i32 tmp1,$0xffffffffffffffff
movi_i32 tmp2,$0x0
add2_i32 tmp0,CA,CA,tmp2,r6,tmp2
add2_i32 tmp0,CA,tmp0,CA,tmp1,tmp2
mov_i32 r10,tmp0
0x7f37010d490c: xor %ebp,%ebp
0x7f37010d490e: mov %ebp,%r11d
0x7f37010d4911: mov 0x18(%r14),%r9d
0x7f37010d4915: add %r9d,%r10d
0x7f37010d4918: adc %ebp,%r11d
0x7f37010d491b: add $0xffffffffffffffff,%r10d
0x7f37010d491f: adc %ebp,%r11d
0x7f37010d4922: mov %r11d,0x134(%r14)
0x7f37010d4929: mov %r10d,0x28(%r14)
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Message-Id: <1436544211-2769-2-git-send-email-aurelien@aurel32.net>
Signed-off-by: Richard Henderson <rth@twiddle.net>
2015-07-10 19:03:31 +03:00
|
|
|
res = (int32_t)res;
|
2011-07-07 16:37:14 +04:00
|
|
|
}
|
|
|
|
return res;
|
|
|
|
}
|
|
|
|
|
2012-10-02 22:32:26 +04:00
|
|
|
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();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-09-18 21:37:00 +04:00
|
|
|
/* Return 2 if the condition can't be simplified, and the result
|
|
|
|
of the condition (0 or 1) if it can */
|
2012-09-06 18:47:14 +04:00
|
|
|
static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
|
|
|
|
TCGArg y, TCGCond c)
|
|
|
|
{
|
2017-06-20 23:43:15 +03:00
|
|
|
tcg_target_ulong xv = arg_info(x)->val;
|
|
|
|
tcg_target_ulong yv = arg_info(y)->val;
|
|
|
|
if (arg_is_const(x) && arg_is_const(y)) {
|
2012-09-18 21:37:00 +04:00
|
|
|
switch (op_bits(op)) {
|
|
|
|
case 32:
|
2017-06-20 23:43:15 +03:00
|
|
|
return do_constant_folding_cond_32(xv, yv, c);
|
2012-09-18 21:37:00 +04:00
|
|
|
case 64:
|
2017-06-20 23:43:15 +03:00
|
|
|
return do_constant_folding_cond_64(xv, yv, c);
|
2012-09-25 01:21:40 +04:00
|
|
|
default:
|
2012-10-02 22:32:26 +04:00
|
|
|
tcg_abort();
|
2012-09-18 21:37:00 +04:00
|
|
|
}
|
2017-06-20 23:43:15 +03:00
|
|
|
} else if (args_are_copies(x, y)) {
|
2012-10-02 22:32:26 +04:00
|
|
|
return do_constant_folding_cond_eq(c);
|
2017-06-20 23:43:15 +03:00
|
|
|
} else if (arg_is_const(y) && yv == 0) {
|
2012-09-18 21:37:00 +04:00
|
|
|
switch (c) {
|
2012-09-06 18:47:14 +04:00
|
|
|
case TCG_COND_LTU:
|
2012-09-18 21:37:00 +04:00
|
|
|
return 0;
|
2012-09-06 18:47:14 +04:00
|
|
|
case TCG_COND_GEU:
|
2012-09-18 21:37:00 +04:00
|
|
|
return 1;
|
|
|
|
default:
|
|
|
|
return 2;
|
2012-09-06 18:47:14 +04:00
|
|
|
}
|
|
|
|
}
|
2016-10-01 00:30:55 +03:00
|
|
|
return 2;
|
2012-09-06 18:47:14 +04:00
|
|
|
}
|
|
|
|
|
2012-10-02 22:32:27 +04:00
|
|
|
/* 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];
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(bl) && arg_is_const(bh)) {
|
|
|
|
tcg_target_ulong blv = arg_info(bl)->val;
|
|
|
|
tcg_target_ulong bhv = arg_info(bh)->val;
|
|
|
|
uint64_t b = deposit64(blv, 32, 32, bhv);
|
2012-10-02 22:32:27 +04:00
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(al) && arg_is_const(ah)) {
|
|
|
|
tcg_target_ulong alv = arg_info(al)->val;
|
|
|
|
tcg_target_ulong ahv = arg_info(ah)->val;
|
|
|
|
uint64_t a = deposit64(alv, 32, 32, ahv);
|
2012-10-02 22:32:27 +04:00
|
|
|
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;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2017-06-20 23:43:15 +03:00
|
|
|
if (args_are_copies(al, bl) && args_are_copies(ah, bh)) {
|
2012-10-02 22:32:27 +04:00
|
|
|
return do_constant_folding_cond_eq(c);
|
|
|
|
}
|
|
|
|
return 2;
|
|
|
|
}
|
|
|
|
|
tcg: Split out swap_commutative as a subroutine
Reduces code duplication and prefers
movcond d, c1, c2, const, s
to
movcond d, c1, c2, s, const
It also prefers
add r, r, c
over
add r, c, r
when both inputs are known constants. This doesn't matter for true add, as
we will fully constant fold that. But it matters for a follow-on patch using
this routine for add2 which may not be fully foldable.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
2012-10-02 22:32:21 +04:00
|
|
|
static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
|
|
|
|
{
|
|
|
|
TCGArg a1 = *p1, a2 = *p2;
|
|
|
|
int sum = 0;
|
2017-06-20 23:43:15 +03:00
|
|
|
sum += arg_is_const(a1);
|
|
|
|
sum -= arg_is_const(a2);
|
tcg: Split out swap_commutative as a subroutine
Reduces code duplication and prefers
movcond d, c1, c2, const, s
to
movcond d, c1, c2, s, const
It also prefers
add r, r, c
over
add r, c, r
when both inputs are known constants. This doesn't matter for true add, as
we will fully constant fold that. But it matters for a follow-on patch using
this routine for add2 which may not be fully foldable.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
2012-10-02 22:32:21 +04:00
|
|
|
|
|
|
|
/* 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;
|
|
|
|
}
|
|
|
|
|
2012-10-02 22:32:23 +04:00
|
|
|
static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
|
|
|
|
{
|
|
|
|
int sum = 0;
|
2017-06-20 23:43:15 +03:00
|
|
|
sum += arg_is_const(p1[0]);
|
|
|
|
sum += arg_is_const(p1[1]);
|
|
|
|
sum -= arg_is_const(p2[0]);
|
|
|
|
sum -= arg_is_const(p2[1]);
|
2012-10-02 22:32:23 +04:00
|
|
|
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;
|
|
|
|
}
|
|
|
|
|
2011-07-07 16:37:13 +04:00
|
|
|
/* Propagate constants and copies, fold constant expressions. */
|
2015-06-04 22:53:27 +03:00
|
|
|
void tcg_optimize(TCGContext *s)
|
2011-07-07 16:37:12 +04:00
|
|
|
{
|
2014-09-20 00:49:15 +04:00
|
|
|
int oi, oi_next, nb_temps, nb_globals;
|
2016-12-08 23:28:42 +03:00
|
|
|
TCGOp *prev_mb = NULL;
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
struct tcg_temp_info *infos;
|
|
|
|
TCGTempSet temps_used;
|
2012-09-21 21:13:38 +04:00
|
|
|
|
2011-07-07 16:37:13 +04:00
|
|
|
/* Array VALS has an element for each temp.
|
|
|
|
If this temp holds a constant then its value is kept in VALS' element.
|
2012-09-11 14:31:21 +04:00
|
|
|
If this temp is a copy of other ones then the other copies are
|
|
|
|
available through the doubly linked circular list. */
|
2011-07-07 16:37:12 +04:00
|
|
|
|
|
|
|
nb_temps = s->nb_temps;
|
|
|
|
nb_globals = s->nb_globals;
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
bitmap_zero(temps_used.l, nb_temps);
|
|
|
|
infos = tcg_malloc(sizeof(struct tcg_temp_info) * nb_temps);
|
2011-07-07 16:37:12 +04:00
|
|
|
|
2016-06-23 05:42:31 +03:00
|
|
|
for (oi = s->gen_op_buf[0].next; oi != 0; oi = oi_next) {
|
2014-05-22 22:14:10 +04:00
|
|
|
tcg_target_ulong mask, partmask, affected;
|
2014-09-20 00:49:15 +04:00
|
|
|
int nb_oargs, nb_iargs, i;
|
2014-03-23 07:06:52 +04:00
|
|
|
TCGArg tmp;
|
|
|
|
|
2014-09-20 00:49:15 +04:00
|
|
|
TCGOp * const op = &s->gen_op_buf[oi];
|
|
|
|
TCGOpcode opc = op->opc;
|
|
|
|
const TCGOpDef *def = &tcg_op_defs[opc];
|
|
|
|
|
|
|
|
oi_next = op->next;
|
2015-07-27 13:41:44 +03:00
|
|
|
|
|
|
|
/* Count the arguments, and initialize the temps that are
|
|
|
|
going to be used */
|
2014-09-20 00:49:15 +04:00
|
|
|
if (opc == INDEX_op_call) {
|
|
|
|
nb_oargs = op->callo;
|
|
|
|
nb_iargs = op->calli;
|
2015-07-27 13:41:44 +03:00
|
|
|
for (i = 0; i < nb_oargs + nb_iargs; i++) {
|
2017-06-20 23:43:15 +03:00
|
|
|
TCGTemp *ts = arg_temp(op->args[i]);
|
|
|
|
if (ts) {
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
init_ts_info(infos, &temps_used, ts);
|
2015-07-27 13:41:44 +03:00
|
|
|
}
|
|
|
|
}
|
2012-09-11 18:18:49 +04:00
|
|
|
} else {
|
2014-03-23 07:06:52 +04:00
|
|
|
nb_oargs = def->nb_oargs;
|
|
|
|
nb_iargs = def->nb_iargs;
|
2015-07-27 13:41:44 +03:00
|
|
|
for (i = 0; i < nb_oargs + nb_iargs; i++) {
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
init_arg_info(infos, &temps_used, op->args[i]);
|
2015-07-27 13:41:44 +03:00
|
|
|
}
|
2014-03-23 07:06:52 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Do copy propagation */
|
|
|
|
for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
|
2017-06-20 23:43:15 +03:00
|
|
|
TCGTemp *ts = arg_temp(op->args[i]);
|
|
|
|
if (ts && ts_is_copy(ts)) {
|
|
|
|
op->args[i] = temp_arg(find_better_copy(s, ts));
|
2011-07-07 16:37:13 +04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-07-07 16:37:14 +04:00
|
|
|
/* For commutative operations make constant second argument */
|
2014-09-20 00:49:15 +04:00
|
|
|
switch (opc) {
|
2011-07-07 16:37:14 +04:00
|
|
|
CASE_OP_32_64(add):
|
|
|
|
CASE_OP_32_64(mul):
|
2011-07-07 16:37:15 +04:00
|
|
|
CASE_OP_32_64(and):
|
|
|
|
CASE_OP_32_64(or):
|
|
|
|
CASE_OP_32_64(xor):
|
2011-08-18 01:11:47 +04:00
|
|
|
CASE_OP_32_64(eqv):
|
|
|
|
CASE_OP_32_64(nand):
|
|
|
|
CASE_OP_32_64(nor):
|
2013-08-15 01:35:56 +04:00
|
|
|
CASE_OP_32_64(muluh):
|
|
|
|
CASE_OP_32_64(mulsh):
|
2016-12-08 23:28:42 +03:00
|
|
|
swap_commutative(op->args[0], &op->args[1], &op->args[2]);
|
2011-07-07 16:37:14 +04:00
|
|
|
break;
|
2012-09-06 18:47:14 +04:00
|
|
|
CASE_OP_32_64(brcond):
|
2016-12-08 23:28:42 +03:00
|
|
|
if (swap_commutative(-1, &op->args[0], &op->args[1])) {
|
|
|
|
op->args[2] = tcg_swap_cond(op->args[2]);
|
2012-09-06 18:47:14 +04:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
CASE_OP_32_64(setcond):
|
2016-12-08 23:28:42 +03:00
|
|
|
if (swap_commutative(op->args[0], &op->args[1], &op->args[2])) {
|
|
|
|
op->args[3] = tcg_swap_cond(op->args[3]);
|
2012-09-06 18:47:14 +04:00
|
|
|
}
|
|
|
|
break;
|
2012-09-21 21:13:37 +04:00
|
|
|
CASE_OP_32_64(movcond):
|
2016-12-08 23:28:42 +03:00
|
|
|
if (swap_commutative(-1, &op->args[1], &op->args[2])) {
|
|
|
|
op->args[5] = tcg_swap_cond(op->args[5]);
|
2012-09-21 21:13:38 +04:00
|
|
|
}
|
|
|
|
/* 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. */
|
2016-12-08 23:28:42 +03:00
|
|
|
if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
|
|
|
|
op->args[5] = tcg_invert_cond(op->args[5]);
|
2012-09-21 21:13:37 +04:00
|
|
|
}
|
2012-10-02 22:32:22 +04:00
|
|
|
break;
|
2013-02-20 11:51:52 +04:00
|
|
|
CASE_OP_32_64(add2):
|
2016-12-08 23:28:42 +03:00
|
|
|
swap_commutative(op->args[0], &op->args[2], &op->args[4]);
|
|
|
|
swap_commutative(op->args[1], &op->args[3], &op->args[5]);
|
2012-10-02 22:32:22 +04:00
|
|
|
break;
|
2013-02-20 11:51:52 +04:00
|
|
|
CASE_OP_32_64(mulu2):
|
2013-02-20 11:51:53 +04:00
|
|
|
CASE_OP_32_64(muls2):
|
2016-12-08 23:28:42 +03:00
|
|
|
swap_commutative(op->args[0], &op->args[2], &op->args[3]);
|
2012-10-02 22:32:30 +04:00
|
|
|
break;
|
2012-10-02 22:32:23 +04:00
|
|
|
case INDEX_op_brcond2_i32:
|
2016-12-08 23:28:42 +03:00
|
|
|
if (swap_commutative2(&op->args[0], &op->args[2])) {
|
|
|
|
op->args[4] = tcg_swap_cond(op->args[4]);
|
2012-10-02 22:32:23 +04:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
case INDEX_op_setcond2_i32:
|
2016-12-08 23:28:42 +03:00
|
|
|
if (swap_commutative2(&op->args[1], &op->args[3])) {
|
|
|
|
op->args[5] = tcg_swap_cond(op->args[5]);
|
2012-10-02 22:32:23 +04:00
|
|
|
}
|
|
|
|
break;
|
2011-07-07 16:37:14 +04:00
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2013-03-21 20:13:33 +04:00
|
|
|
/* Simplify expressions for "shift/rot r, 0, a => movi r, 0",
|
|
|
|
and "sub r, 0, a => neg r, a" case. */
|
2014-09-20 00:49:15 +04:00
|
|
|
switch (opc) {
|
2012-09-06 18:47:14 +04:00
|
|
|
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):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[1])
|
|
|
|
&& arg_info(op->args[1])->val == 0) {
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], 0);
|
2012-09-06 18:47:14 +04:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
break;
|
2013-03-21 20:13:33 +04:00
|
|
|
CASE_OP_32_64(sub):
|
|
|
|
{
|
|
|
|
TCGOpcode neg_op;
|
|
|
|
bool have_neg;
|
|
|
|
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2])) {
|
2013-03-21 20:13:33 +04:00
|
|
|
/* Proceed with possible constant folding. */
|
|
|
|
break;
|
|
|
|
}
|
2014-09-20 00:49:15 +04:00
|
|
|
if (opc == INDEX_op_sub_i32) {
|
2013-03-21 20:13:33 +04:00
|
|
|
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;
|
|
|
|
}
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[1])
|
|
|
|
&& arg_info(op->args[1])->val == 0) {
|
2014-09-20 00:49:15 +04:00
|
|
|
op->opc = neg_op;
|
2016-12-08 23:28:42 +03:00
|
|
|
reset_temp(op->args[0]);
|
|
|
|
op->args[1] = op->args[2];
|
2013-03-21 20:13:33 +04:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
2014-01-29 01:15:38 +04:00
|
|
|
CASE_OP_32_64(xor):
|
|
|
|
CASE_OP_32_64(nand):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (!arg_is_const(op->args[1])
|
|
|
|
&& arg_is_const(op->args[2])
|
|
|
|
&& arg_info(op->args[2])->val == -1) {
|
2014-01-29 01:15:38 +04:00
|
|
|
i = 1;
|
|
|
|
goto try_not;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
CASE_OP_32_64(nor):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (!arg_is_const(op->args[1])
|
|
|
|
&& arg_is_const(op->args[2])
|
|
|
|
&& arg_info(op->args[2])->val == 0) {
|
2014-01-29 01:15:38 +04:00
|
|
|
i = 1;
|
|
|
|
goto try_not;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
CASE_OP_32_64(andc):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (!arg_is_const(op->args[2])
|
|
|
|
&& arg_is_const(op->args[1])
|
|
|
|
&& arg_info(op->args[1])->val == -1) {
|
2014-01-29 01:15:38 +04:00
|
|
|
i = 2;
|
|
|
|
goto try_not;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
CASE_OP_32_64(orc):
|
|
|
|
CASE_OP_32_64(eqv):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (!arg_is_const(op->args[2])
|
|
|
|
&& arg_is_const(op->args[1])
|
|
|
|
&& arg_info(op->args[1])->val == 0) {
|
2014-01-29 01:15:38 +04:00
|
|
|
i = 2;
|
|
|
|
goto try_not;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
try_not:
|
|
|
|
{
|
|
|
|
TCGOpcode not_op;
|
|
|
|
bool have_not;
|
|
|
|
|
|
|
|
if (def->flags & TCG_OPF_64BIT) {
|
|
|
|
not_op = INDEX_op_not_i64;
|
|
|
|
have_not = TCG_TARGET_HAS_not_i64;
|
|
|
|
} else {
|
|
|
|
not_op = INDEX_op_not_i32;
|
|
|
|
have_not = TCG_TARGET_HAS_not_i32;
|
|
|
|
}
|
|
|
|
if (!have_not) {
|
|
|
|
break;
|
|
|
|
}
|
2014-09-20 00:49:15 +04:00
|
|
|
op->opc = not_op;
|
2016-12-08 23:28:42 +03:00
|
|
|
reset_temp(op->args[0]);
|
|
|
|
op->args[1] = op->args[i];
|
2014-01-29 01:15:38 +04:00
|
|
|
continue;
|
|
|
|
}
|
2012-09-06 18:47:14 +04:00
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2014-01-31 17:42:11 +04:00
|
|
|
/* Simplify expression for "op r, a, const => mov r, a" cases */
|
2014-09-20 00:49:15 +04:00
|
|
|
switch (opc) {
|
2011-07-07 16:37:14 +04:00
|
|
|
CASE_OP_32_64(add):
|
|
|
|
CASE_OP_32_64(sub):
|
2011-07-07 16:37:16 +04:00
|
|
|
CASE_OP_32_64(shl):
|
|
|
|
CASE_OP_32_64(shr):
|
|
|
|
CASE_OP_32_64(sar):
|
2011-08-18 01:11:46 +04:00
|
|
|
CASE_OP_32_64(rotl):
|
|
|
|
CASE_OP_32_64(rotr):
|
2012-09-06 18:47:14 +04:00
|
|
|
CASE_OP_32_64(or):
|
|
|
|
CASE_OP_32_64(xor):
|
2014-01-31 17:42:11 +04:00
|
|
|
CASE_OP_32_64(andc):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (!arg_is_const(op->args[1])
|
|
|
|
&& arg_is_const(op->args[2])
|
|
|
|
&& arg_info(op->args[2])->val == 0) {
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_mov(s, op, op->args[0], op->args[1]);
|
2015-06-05 12:19:18 +03:00
|
|
|
continue;
|
2011-07-07 16:37:14 +04:00
|
|
|
}
|
|
|
|
break;
|
2014-01-31 17:42:11 +04:00
|
|
|
CASE_OP_32_64(and):
|
|
|
|
CASE_OP_32_64(orc):
|
|
|
|
CASE_OP_32_64(eqv):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (!arg_is_const(op->args[1])
|
|
|
|
&& arg_is_const(op->args[2])
|
|
|
|
&& arg_info(op->args[2])->val == -1) {
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_mov(s, op, op->args[0], op->args[1]);
|
2015-06-05 12:19:18 +03:00
|
|
|
continue;
|
2014-01-31 17:42:11 +04:00
|
|
|
}
|
|
|
|
break;
|
2012-09-06 18:47:13 +04:00
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2013-09-03 10:27:38 +04:00
|
|
|
/* Simplify using known-zero bits. Currently only ops with a single
|
|
|
|
output argument is supported. */
|
2013-01-12 03:42:52 +04:00
|
|
|
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;
|
2014-09-20 00:49:15 +04:00
|
|
|
switch (opc) {
|
2013-01-12 03:42:52 +04:00
|
|
|
CASE_OP_32_64(ext8s):
|
2017-06-20 23:43:15 +03:00
|
|
|
if ((arg_info(op->args[1])->mask & 0x80) != 0) {
|
2013-01-12 03:42:52 +04:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
CASE_OP_32_64(ext8u):
|
|
|
|
mask = 0xff;
|
|
|
|
goto and_const;
|
|
|
|
CASE_OP_32_64(ext16s):
|
2017-06-20 23:43:15 +03:00
|
|
|
if ((arg_info(op->args[1])->mask & 0x8000) != 0) {
|
2013-01-12 03:42:52 +04:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
CASE_OP_32_64(ext16u):
|
|
|
|
mask = 0xffff;
|
|
|
|
goto and_const;
|
|
|
|
case INDEX_op_ext32s_i64:
|
2017-06-20 23:43:15 +03:00
|
|
|
if ((arg_info(op->args[1])->mask & 0x80000000) != 0) {
|
2013-01-12 03:42:52 +04:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
case INDEX_op_ext32u_i64:
|
|
|
|
mask = 0xffffffffU;
|
|
|
|
goto and_const;
|
|
|
|
|
|
|
|
CASE_OP_32_64(and):
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = arg_info(op->args[2])->mask;
|
|
|
|
if (arg_is_const(op->args[2])) {
|
2013-01-12 03:42:52 +04:00
|
|
|
and_const:
|
2017-06-20 23:43:15 +03:00
|
|
|
affected = arg_info(op->args[1])->mask & ~mask;
|
2013-01-12 03:42:52 +04:00
|
|
|
}
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = arg_info(op->args[1])->mask & mask;
|
2013-01-12 03:42:52 +04:00
|
|
|
break;
|
|
|
|
|
2015-07-27 13:41:45 +03:00
|
|
|
case INDEX_op_ext_i32_i64:
|
2017-06-20 23:43:15 +03:00
|
|
|
if ((arg_info(op->args[1])->mask & 0x80000000) != 0) {
|
2015-07-27 13:41:45 +03:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
case INDEX_op_extu_i32_i64:
|
|
|
|
/* We do not compute affected as it is a size changing op. */
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = (uint32_t)arg_info(op->args[1])->mask;
|
2015-07-27 13:41:45 +03:00
|
|
|
break;
|
|
|
|
|
2014-01-29 00:03:24 +04:00
|
|
|
CASE_OP_32_64(andc):
|
|
|
|
/* Known-zeros does not imply known-ones. Therefore unless
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[2] is constant, we can't infer anything from it. */
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2])) {
|
|
|
|
mask = ~arg_info(op->args[2])->mask;
|
2014-01-29 00:03:24 +04:00
|
|
|
goto and_const;
|
|
|
|
}
|
2017-06-20 23:43:15 +03:00
|
|
|
/* But we certainly know nothing outside args[1] may be set. */
|
|
|
|
mask = arg_info(op->args[1])->mask;
|
2014-01-29 00:03:24 +04:00
|
|
|
break;
|
|
|
|
|
2013-09-03 10:27:38 +04:00
|
|
|
case INDEX_op_sar_i32:
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2])) {
|
|
|
|
tmp = arg_info(op->args[2])->val & 31;
|
|
|
|
mask = (int32_t)arg_info(op->args[1])->mask >> tmp;
|
2013-09-03 10:27:38 +04:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
case INDEX_op_sar_i64:
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2])) {
|
|
|
|
tmp = arg_info(op->args[2])->val & 63;
|
|
|
|
mask = (int64_t)arg_info(op->args[1])->mask >> tmp;
|
2013-01-12 03:42:52 +04:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
2013-09-03 10:27:38 +04:00
|
|
|
case INDEX_op_shr_i32:
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2])) {
|
|
|
|
tmp = arg_info(op->args[2])->val & 31;
|
|
|
|
mask = (uint32_t)arg_info(op->args[1])->mask >> tmp;
|
2013-09-03 10:27:38 +04:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
case INDEX_op_shr_i64:
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2])) {
|
|
|
|
tmp = arg_info(op->args[2])->val & 63;
|
|
|
|
mask = (uint64_t)arg_info(op->args[1])->mask >> tmp;
|
2013-01-12 03:42:52 +04:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
2015-07-24 17:16:00 +03:00
|
|
|
case INDEX_op_extrl_i64_i32:
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = (uint32_t)arg_info(op->args[1])->mask;
|
2015-07-24 17:16:00 +03:00
|
|
|
break;
|
|
|
|
case INDEX_op_extrh_i64_i32:
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = (uint64_t)arg_info(op->args[1])->mask >> 32;
|
2013-09-10 04:03:24 +04:00
|
|
|
break;
|
|
|
|
|
2013-01-12 03:42:52 +04:00
|
|
|
CASE_OP_32_64(shl):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2])) {
|
|
|
|
tmp = arg_info(op->args[2])->val & (TCG_TARGET_REG_BITS - 1);
|
|
|
|
mask = arg_info(op->args[1])->mask << tmp;
|
2013-01-12 03:42:52 +04:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
CASE_OP_32_64(neg):
|
|
|
|
/* Set to 1 all bits to the left of the rightmost. */
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = -(arg_info(op->args[1])->mask
|
|
|
|
& -arg_info(op->args[1])->mask);
|
2013-01-12 03:42:52 +04:00
|
|
|
break;
|
|
|
|
|
|
|
|
CASE_OP_32_64(deposit):
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = deposit64(arg_info(op->args[1])->mask,
|
|
|
|
op->args[3], op->args[4],
|
|
|
|
arg_info(op->args[2])->mask);
|
2013-01-12 03:42:52 +04:00
|
|
|
break;
|
|
|
|
|
2016-10-14 20:04:32 +03:00
|
|
|
CASE_OP_32_64(extract):
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = extract64(arg_info(op->args[1])->mask,
|
|
|
|
op->args[2], op->args[3]);
|
2016-12-08 23:28:42 +03:00
|
|
|
if (op->args[2] == 0) {
|
2017-06-20 23:43:15 +03:00
|
|
|
affected = arg_info(op->args[1])->mask & ~mask;
|
2016-10-14 20:04:32 +03:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
CASE_OP_32_64(sextract):
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = sextract64(arg_info(op->args[1])->mask,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[2], op->args[3]);
|
|
|
|
if (op->args[2] == 0 && (tcg_target_long)mask >= 0) {
|
2017-06-20 23:43:15 +03:00
|
|
|
affected = arg_info(op->args[1])->mask & ~mask;
|
2016-10-14 20:04:32 +03:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
2013-01-12 03:42:52 +04:00
|
|
|
CASE_OP_32_64(or):
|
|
|
|
CASE_OP_32_64(xor):
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = arg_info(op->args[1])->mask | arg_info(op->args[2])->mask;
|
2013-01-12 03:42:52 +04:00
|
|
|
break;
|
|
|
|
|
2016-11-16 11:23:28 +03:00
|
|
|
case INDEX_op_clz_i32:
|
|
|
|
case INDEX_op_ctz_i32:
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = arg_info(op->args[2])->mask | 31;
|
2016-11-16 11:23:28 +03:00
|
|
|
break;
|
|
|
|
|
|
|
|
case INDEX_op_clz_i64:
|
|
|
|
case INDEX_op_ctz_i64:
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = arg_info(op->args[2])->mask | 63;
|
2016-11-16 11:23:28 +03:00
|
|
|
break;
|
|
|
|
|
2016-11-21 13:13:39 +03:00
|
|
|
case INDEX_op_ctpop_i32:
|
|
|
|
mask = 32 | 31;
|
|
|
|
break;
|
|
|
|
case INDEX_op_ctpop_i64:
|
|
|
|
mask = 64 | 63;
|
|
|
|
break;
|
|
|
|
|
2013-01-12 03:42:52 +04:00
|
|
|
CASE_OP_32_64(setcond):
|
2014-04-24 09:18:30 +04:00
|
|
|
case INDEX_op_setcond2_i32:
|
2013-01-12 03:42:52 +04:00
|
|
|
mask = 1;
|
|
|
|
break;
|
|
|
|
|
|
|
|
CASE_OP_32_64(movcond):
|
2017-06-20 23:43:15 +03:00
|
|
|
mask = arg_info(op->args[3])->mask | arg_info(op->args[4])->mask;
|
2013-01-12 03:42:52 +04:00
|
|
|
break;
|
|
|
|
|
2013-09-03 10:27:39 +04:00
|
|
|
CASE_OP_32_64(ld8u):
|
|
|
|
mask = 0xff;
|
|
|
|
break;
|
|
|
|
CASE_OP_32_64(ld16u):
|
|
|
|
mask = 0xffff;
|
|
|
|
break;
|
|
|
|
case INDEX_op_ld32u_i64:
|
|
|
|
mask = 0xffffffffu;
|
|
|
|
break;
|
|
|
|
|
|
|
|
CASE_OP_32_64(qemu_ld):
|
|
|
|
{
|
2016-12-08 23:28:42 +03:00
|
|
|
TCGMemOpIdx oi = op->args[nb_oargs + nb_iargs];
|
2015-05-12 21:51:44 +03:00
|
|
|
TCGMemOp mop = get_memop(oi);
|
2013-09-03 10:27:39 +04:00
|
|
|
if (!(mop & MO_SIGN)) {
|
|
|
|
mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
2013-01-12 03:42:52 +04:00
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2014-06-09 05:24:14 +04:00
|
|
|
/* 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. */
|
2014-05-22 22:14:10 +04:00
|
|
|
partmask = mask;
|
2014-06-09 05:24:14 +04:00
|
|
|
if (!(def->flags & TCG_OPF_64BIT)) {
|
2014-05-22 22:14:10 +04:00
|
|
|
mask |= ~(tcg_target_ulong)0xffffffffu;
|
|
|
|
partmask &= 0xffffffffu;
|
|
|
|
affected &= 0xffffffffu;
|
2013-09-03 10:27:38 +04:00
|
|
|
}
|
|
|
|
|
2014-05-22 22:14:10 +04:00
|
|
|
if (partmask == 0) {
|
2016-04-21 11:48:49 +03:00
|
|
|
tcg_debug_assert(nb_oargs == 1);
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], 0);
|
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
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (affected == 0) {
|
2016-04-21 11:48:49 +03:00
|
|
|
tcg_debug_assert(nb_oargs == 1);
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_mov(s, op, op->args[0], op->args[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
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2012-09-06 18:47:13 +04:00
|
|
|
/* Simplify expression for "op r, a, 0 => movi r, 0" cases */
|
2014-09-20 00:49:15 +04:00
|
|
|
switch (opc) {
|
2012-09-06 18:47:14 +04:00
|
|
|
CASE_OP_32_64(and):
|
2011-07-07 16:37:14 +04:00
|
|
|
CASE_OP_32_64(mul):
|
2013-08-15 01:35:56 +04:00
|
|
|
CASE_OP_32_64(muluh):
|
|
|
|
CASE_OP_32_64(mulsh):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2])
|
|
|
|
&& arg_info(op->args[2])->val == 0) {
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], 0);
|
2011-07-07 16:37:14 +04:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
break;
|
2012-09-06 18:47:13 +04:00
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Simplify expression for "op r, a, a => mov r, a" cases */
|
2014-09-20 00:49:15 +04:00
|
|
|
switch (opc) {
|
2011-07-07 16:37:15 +04:00
|
|
|
CASE_OP_32_64(or):
|
|
|
|
CASE_OP_32_64(and):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (args_are_copies(op->args[1], op->args[2])) {
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_mov(s, op, op->args[0], op->args[1]);
|
2011-07-07 16:37:15 +04:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
break;
|
2011-07-30 23:18:32 +04:00
|
|
|
default:
|
|
|
|
break;
|
2011-07-07 16:37:14 +04:00
|
|
|
}
|
|
|
|
|
2012-09-18 21:12:36 +04:00
|
|
|
/* Simplify expression for "op r, a, a => movi r, 0" cases */
|
2014-09-20 00:49:15 +04:00
|
|
|
switch (opc) {
|
2014-01-29 01:26:17 +04:00
|
|
|
CASE_OP_32_64(andc):
|
2012-09-18 21:12:36 +04:00
|
|
|
CASE_OP_32_64(sub):
|
|
|
|
CASE_OP_32_64(xor):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (args_are_copies(op->args[1], op->args[2])) {
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], 0);
|
2012-09-18 21:12:36 +04:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2011-07-07 16:37:13 +04:00
|
|
|
/* 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. */
|
2014-09-20 00:49:15 +04:00
|
|
|
switch (opc) {
|
2011-07-07 16:37:13 +04:00
|
|
|
CASE_OP_32_64(mov):
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_mov(s, op, op->args[0], op->args[1]);
|
2015-06-05 12:19:18 +03:00
|
|
|
break;
|
2011-07-07 16:37:13 +04:00
|
|
|
CASE_OP_32_64(movi):
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], op->args[1]);
|
2011-07-07 16:37:13 +04:00
|
|
|
break;
|
2012-10-02 22:32:24 +04:00
|
|
|
|
2011-07-07 16:37:17 +04:00
|
|
|
CASE_OP_32_64(not):
|
2011-08-18 01:11:47 +04:00
|
|
|
CASE_OP_32_64(neg):
|
2011-08-18 01:11:46 +04:00
|
|
|
CASE_OP_32_64(ext8s):
|
|
|
|
CASE_OP_32_64(ext8u):
|
|
|
|
CASE_OP_32_64(ext16s):
|
|
|
|
CASE_OP_32_64(ext16u):
|
2016-11-21 13:13:39 +03:00
|
|
|
CASE_OP_32_64(ctpop):
|
2011-07-07 16:37:17 +04:00
|
|
|
case INDEX_op_ext32s_i64:
|
|
|
|
case INDEX_op_ext32u_i64:
|
2015-07-27 13:41:45 +03:00
|
|
|
case INDEX_op_ext_i32_i64:
|
|
|
|
case INDEX_op_extu_i32_i64:
|
2015-07-24 17:16:00 +03:00
|
|
|
case INDEX_op_extrl_i64_i32:
|
|
|
|
case INDEX_op_extrh_i64_i32:
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[1])) {
|
|
|
|
tmp = do_constant_folding(opc, arg_info(op->args[1])->val, 0);
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], tmp);
|
2012-10-02 22:32:24 +04:00
|
|
|
break;
|
2011-07-07 16:37:17 +04:00
|
|
|
}
|
2012-10-02 22:32:24 +04:00
|
|
|
goto do_default;
|
|
|
|
|
2011-07-07 16:37:14 +04:00
|
|
|
CASE_OP_32_64(add):
|
|
|
|
CASE_OP_32_64(sub):
|
|
|
|
CASE_OP_32_64(mul):
|
2011-07-07 16:37:15 +04:00
|
|
|
CASE_OP_32_64(or):
|
|
|
|
CASE_OP_32_64(and):
|
|
|
|
CASE_OP_32_64(xor):
|
2011-07-07 16:37:16 +04:00
|
|
|
CASE_OP_32_64(shl):
|
|
|
|
CASE_OP_32_64(shr):
|
|
|
|
CASE_OP_32_64(sar):
|
2011-08-18 01:11:46 +04:00
|
|
|
CASE_OP_32_64(rotl):
|
|
|
|
CASE_OP_32_64(rotr):
|
2011-08-18 01:11:47 +04:00
|
|
|
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):
|
2013-08-15 01:35:56 +04:00
|
|
|
CASE_OP_32_64(muluh):
|
|
|
|
CASE_OP_32_64(mulsh):
|
2013-08-15 02:22:46 +04:00
|
|
|
CASE_OP_32_64(div):
|
|
|
|
CASE_OP_32_64(divu):
|
|
|
|
CASE_OP_32_64(rem):
|
|
|
|
CASE_OP_32_64(remu):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
|
|
|
|
tmp = do_constant_folding(opc, arg_info(op->args[1])->val,
|
|
|
|
arg_info(op->args[2])->val);
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], tmp);
|
2012-10-02 22:32:24 +04:00
|
|
|
break;
|
2011-07-07 16:37:14 +04:00
|
|
|
}
|
2012-10-02 22:32:24 +04:00
|
|
|
goto do_default;
|
|
|
|
|
2016-11-16 11:23:28 +03:00
|
|
|
CASE_OP_32_64(clz):
|
|
|
|
CASE_OP_32_64(ctz):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[1])) {
|
|
|
|
TCGArg v = arg_info(op->args[1])->val;
|
2016-11-16 11:23:28 +03:00
|
|
|
if (v != 0) {
|
|
|
|
tmp = do_constant_folding(opc, v, 0);
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], tmp);
|
2016-11-16 11:23:28 +03:00
|
|
|
} else {
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_mov(s, op, op->args[0], op->args[2]);
|
2016-11-16 11:23:28 +03:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
goto do_default;
|
|
|
|
|
2012-09-21 13:07:29 +04:00
|
|
|
CASE_OP_32_64(deposit):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
|
|
|
|
tmp = deposit64(arg_info(op->args[1])->val,
|
|
|
|
op->args[3], op->args[4],
|
|
|
|
arg_info(op->args[2])->val);
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], tmp);
|
2012-10-02 22:32:24 +04:00
|
|
|
break;
|
2012-09-21 13:07:29 +04:00
|
|
|
}
|
2012-10-02 22:32:24 +04:00
|
|
|
goto do_default;
|
|
|
|
|
2016-10-14 20:04:32 +03:00
|
|
|
CASE_OP_32_64(extract):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[1])) {
|
|
|
|
tmp = extract64(arg_info(op->args[1])->val,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[2], op->args[3]);
|
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], tmp);
|
2016-10-14 20:04:32 +03:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
goto do_default;
|
|
|
|
|
|
|
|
CASE_OP_32_64(sextract):
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[1])) {
|
|
|
|
tmp = sextract64(arg_info(op->args[1])->val,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[2], op->args[3]);
|
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], tmp);
|
2016-10-14 20:04:32 +03:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
goto do_default;
|
|
|
|
|
2012-09-06 18:47:14 +04:00
|
|
|
CASE_OP_32_64(setcond):
|
2016-12-08 23:28:42 +03:00
|
|
|
tmp = do_constant_folding_cond(opc, op->args[1],
|
|
|
|
op->args[2], op->args[3]);
|
2012-09-18 21:37:00 +04:00
|
|
|
if (tmp != 2) {
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], tmp);
|
2012-10-02 22:32:24 +04:00
|
|
|
break;
|
2012-09-06 18:47:14 +04:00
|
|
|
}
|
2012-10-02 22:32:24 +04:00
|
|
|
goto do_default;
|
|
|
|
|
2012-09-06 18:47:14 +04:00
|
|
|
CASE_OP_32_64(brcond):
|
2016-12-08 23:28:42 +03:00
|
|
|
tmp = do_constant_folding_cond(opc, op->args[0],
|
|
|
|
op->args[1], op->args[2]);
|
2012-09-18 21:37:00 +04:00
|
|
|
if (tmp != 2) {
|
|
|
|
if (tmp) {
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
bitmap_zero(temps_used.l, nb_temps);
|
2014-09-20 00:49:15 +04:00
|
|
|
op->opc = INDEX_op_br;
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[0] = op->args[3];
|
2012-09-06 18:47:14 +04:00
|
|
|
} else {
|
2014-03-31 03:51:54 +04:00
|
|
|
tcg_op_remove(s, op);
|
2012-09-06 18:47:14 +04:00
|
|
|
}
|
2012-10-02 22:32:24 +04:00
|
|
|
break;
|
2012-09-06 18:47:14 +04:00
|
|
|
}
|
2012-10-02 22:32:24 +04:00
|
|
|
goto do_default;
|
|
|
|
|
2012-09-21 21:13:37 +04:00
|
|
|
CASE_OP_32_64(movcond):
|
2016-12-08 23:28:42 +03:00
|
|
|
tmp = do_constant_folding_cond(opc, op->args[1],
|
|
|
|
op->args[2], op->args[5]);
|
2012-09-18 21:37:00 +04:00
|
|
|
if (tmp != 2) {
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_mov(s, op, op->args[0], op->args[4-tmp]);
|
2012-10-02 22:32:24 +04:00
|
|
|
break;
|
2012-09-21 21:13:37 +04:00
|
|
|
}
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) {
|
|
|
|
tcg_target_ulong tv = arg_info(op->args[3])->val;
|
|
|
|
tcg_target_ulong fv = arg_info(op->args[4])->val;
|
2016-12-08 23:28:42 +03:00
|
|
|
TCGCond cond = op->args[5];
|
2016-10-24 06:44:32 +03:00
|
|
|
if (fv == 1 && tv == 0) {
|
|
|
|
cond = tcg_invert_cond(cond);
|
|
|
|
} else if (!(tv == 1 && fv == 0)) {
|
|
|
|
goto do_default;
|
|
|
|
}
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[3] = cond;
|
2016-10-24 06:44:32 +03:00
|
|
|
op->opc = opc = (opc == INDEX_op_movcond_i32
|
|
|
|
? INDEX_op_setcond_i32
|
|
|
|
: INDEX_op_setcond_i64);
|
|
|
|
nb_iargs = 2;
|
|
|
|
}
|
2012-10-02 22:32:24 +04:00
|
|
|
goto do_default;
|
2012-10-02 22:32:28 +04:00
|
|
|
|
|
|
|
case INDEX_op_add2_i32:
|
|
|
|
case INDEX_op_sub2_i32:
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])
|
|
|
|
&& arg_is_const(op->args[4]) && arg_is_const(op->args[5])) {
|
|
|
|
uint32_t al = arg_info(op->args[2])->val;
|
|
|
|
uint32_t ah = arg_info(op->args[3])->val;
|
|
|
|
uint32_t bl = arg_info(op->args[4])->val;
|
|
|
|
uint32_t bh = arg_info(op->args[5])->val;
|
2012-10-02 22:32:28 +04:00
|
|
|
uint64_t a = ((uint64_t)ah << 32) | al;
|
|
|
|
uint64_t b = ((uint64_t)bh << 32) | bl;
|
|
|
|
TCGArg rl, rh;
|
2016-06-24 06:34:33 +03:00
|
|
|
TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_movi_i32, 2);
|
2012-10-02 22:32:28 +04:00
|
|
|
|
2014-09-20 00:49:15 +04:00
|
|
|
if (opc == INDEX_op_add2_i32) {
|
2012-10-02 22:32:28 +04:00
|
|
|
a += b;
|
|
|
|
} else {
|
|
|
|
a -= b;
|
|
|
|
}
|
|
|
|
|
2016-12-08 23:28:42 +03:00
|
|
|
rl = op->args[0];
|
|
|
|
rh = op->args[1];
|
|
|
|
tcg_opt_gen_movi(s, op, rl, (int32_t)a);
|
|
|
|
tcg_opt_gen_movi(s, op2, rh, (int32_t)(a >> 32));
|
2014-09-20 00:49:15 +04:00
|
|
|
|
|
|
|
/* We've done all we need to do with the movi. Skip it. */
|
|
|
|
oi_next = op2->next;
|
2012-10-02 22:32:28 +04:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
goto do_default;
|
2012-10-02 22:32:30 +04:00
|
|
|
|
|
|
|
case INDEX_op_mulu2_i32:
|
2017-06-20 23:43:15 +03:00
|
|
|
if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
|
|
|
|
uint32_t a = arg_info(op->args[2])->val;
|
|
|
|
uint32_t b = arg_info(op->args[3])->val;
|
2012-10-02 22:32:30 +04:00
|
|
|
uint64_t r = (uint64_t)a * b;
|
|
|
|
TCGArg rl, rh;
|
2016-06-24 06:34:33 +03:00
|
|
|
TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_movi_i32, 2);
|
2012-10-02 22:32:30 +04:00
|
|
|
|
2016-12-08 23:28:42 +03:00
|
|
|
rl = op->args[0];
|
|
|
|
rh = op->args[1];
|
|
|
|
tcg_opt_gen_movi(s, op, rl, (int32_t)r);
|
|
|
|
tcg_opt_gen_movi(s, op2, rh, (int32_t)(r >> 32));
|
2014-09-20 00:49:15 +04:00
|
|
|
|
|
|
|
/* We've done all we need to do with the movi. Skip it. */
|
|
|
|
oi_next = op2->next;
|
2012-10-02 22:32:30 +04:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
goto do_default;
|
2012-10-02 22:32:24 +04:00
|
|
|
|
2012-10-02 22:32:25 +04:00
|
|
|
case INDEX_op_brcond2_i32:
|
2016-12-08 23:28:42 +03:00
|
|
|
tmp = do_constant_folding_cond2(&op->args[0], &op->args[2],
|
|
|
|
op->args[4]);
|
2012-10-02 22:32:27 +04:00
|
|
|
if (tmp != 2) {
|
|
|
|
if (tmp) {
|
2014-04-24 09:18:30 +04:00
|
|
|
do_brcond_true:
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
bitmap_zero(temps_used.l, nb_temps);
|
2014-09-20 00:49:15 +04:00
|
|
|
op->opc = INDEX_op_br;
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[0] = op->args[5];
|
2012-10-02 22:32:27 +04:00
|
|
|
} else {
|
2014-04-24 09:18:30 +04:00
|
|
|
do_brcond_false:
|
2014-03-31 03:51:54 +04:00
|
|
|
tcg_op_remove(s, op);
|
2012-10-02 22:32:27 +04:00
|
|
|
}
|
2016-12-08 23:28:42 +03:00
|
|
|
} else if ((op->args[4] == TCG_COND_LT
|
|
|
|
|| op->args[4] == TCG_COND_GE)
|
2017-06-20 23:43:15 +03:00
|
|
|
&& arg_is_const(op->args[2])
|
|
|
|
&& arg_info(op->args[2])->val == 0
|
|
|
|
&& arg_is_const(op->args[3])
|
|
|
|
&& arg_info(op->args[3])->val == 0) {
|
2012-10-02 22:32:27 +04:00
|
|
|
/* Simplify LT/GE comparisons vs zero to a single compare
|
|
|
|
vs the high word of the input. */
|
2014-04-24 09:18:30 +04:00
|
|
|
do_brcond_high:
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
bitmap_zero(temps_used.l, nb_temps);
|
2014-09-20 00:49:15 +04:00
|
|
|
op->opc = INDEX_op_brcond_i32;
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[0] = op->args[1];
|
|
|
|
op->args[1] = op->args[3];
|
|
|
|
op->args[2] = op->args[4];
|
|
|
|
op->args[3] = op->args[5];
|
|
|
|
} else if (op->args[4] == TCG_COND_EQ) {
|
2014-04-24 09:18:30 +04:00
|
|
|
/* Simplify EQ comparisons where one of the pairs
|
|
|
|
can be simplified. */
|
|
|
|
tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[0], op->args[2],
|
|
|
|
TCG_COND_EQ);
|
2014-04-24 09:18:30 +04:00
|
|
|
if (tmp == 0) {
|
|
|
|
goto do_brcond_false;
|
|
|
|
} else if (tmp == 1) {
|
|
|
|
goto do_brcond_high;
|
|
|
|
}
|
|
|
|
tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[1], op->args[3],
|
|
|
|
TCG_COND_EQ);
|
2014-04-24 09:18:30 +04:00
|
|
|
if (tmp == 0) {
|
|
|
|
goto do_brcond_false;
|
|
|
|
} else if (tmp != 1) {
|
|
|
|
goto do_default;
|
|
|
|
}
|
|
|
|
do_brcond_low:
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
bitmap_zero(temps_used.l, nb_temps);
|
2014-09-20 00:49:15 +04:00
|
|
|
op->opc = INDEX_op_brcond_i32;
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[1] = op->args[2];
|
|
|
|
op->args[2] = op->args[4];
|
|
|
|
op->args[3] = op->args[5];
|
|
|
|
} else if (op->args[4] == TCG_COND_NE) {
|
2014-04-24 09:18:30 +04:00
|
|
|
/* Simplify NE comparisons where one of the pairs
|
|
|
|
can be simplified. */
|
|
|
|
tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[0], op->args[2],
|
|
|
|
TCG_COND_NE);
|
2014-04-24 09:18:30 +04:00
|
|
|
if (tmp == 0) {
|
|
|
|
goto do_brcond_high;
|
|
|
|
} else if (tmp == 1) {
|
|
|
|
goto do_brcond_true;
|
|
|
|
}
|
|
|
|
tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[1], op->args[3],
|
|
|
|
TCG_COND_NE);
|
2014-04-24 09:18:30 +04:00
|
|
|
if (tmp == 0) {
|
|
|
|
goto do_brcond_low;
|
|
|
|
} else if (tmp == 1) {
|
|
|
|
goto do_brcond_true;
|
|
|
|
}
|
|
|
|
goto do_default;
|
2012-10-02 22:32:27 +04:00
|
|
|
} else {
|
|
|
|
goto do_default;
|
2012-10-02 22:32:25 +04:00
|
|
|
}
|
2012-10-02 22:32:27 +04:00
|
|
|
break;
|
2012-10-02 22:32:25 +04:00
|
|
|
|
|
|
|
case INDEX_op_setcond2_i32:
|
2016-12-08 23:28:42 +03:00
|
|
|
tmp = do_constant_folding_cond2(&op->args[1], &op->args[3],
|
|
|
|
op->args[5]);
|
2012-10-02 22:32:27 +04:00
|
|
|
if (tmp != 2) {
|
2014-04-24 09:18:30 +04:00
|
|
|
do_setcond_const:
|
2016-12-08 23:28:42 +03:00
|
|
|
tcg_opt_gen_movi(s, op, op->args[0], tmp);
|
|
|
|
} else if ((op->args[5] == TCG_COND_LT
|
|
|
|
|| op->args[5] == TCG_COND_GE)
|
2017-06-20 23:43:15 +03:00
|
|
|
&& arg_is_const(op->args[3])
|
|
|
|
&& arg_info(op->args[3])->val == 0
|
|
|
|
&& arg_is_const(op->args[4])
|
|
|
|
&& arg_info(op->args[4])->val == 0) {
|
2012-10-02 22:32:27 +04:00
|
|
|
/* Simplify LT/GE comparisons vs zero to a single compare
|
|
|
|
vs the high word of the input. */
|
2014-04-24 09:18:30 +04:00
|
|
|
do_setcond_high:
|
2016-12-08 23:28:42 +03:00
|
|
|
reset_temp(op->args[0]);
|
2017-06-20 23:43:15 +03:00
|
|
|
arg_info(op->args[0])->mask = 1;
|
2014-09-20 00:49:15 +04:00
|
|
|
op->opc = INDEX_op_setcond_i32;
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[1] = op->args[2];
|
|
|
|
op->args[2] = op->args[4];
|
|
|
|
op->args[3] = op->args[5];
|
|
|
|
} else if (op->args[5] == TCG_COND_EQ) {
|
2014-04-24 09:18:30 +04:00
|
|
|
/* Simplify EQ comparisons where one of the pairs
|
|
|
|
can be simplified. */
|
|
|
|
tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[1], op->args[3],
|
|
|
|
TCG_COND_EQ);
|
2014-04-24 09:18:30 +04:00
|
|
|
if (tmp == 0) {
|
|
|
|
goto do_setcond_const;
|
|
|
|
} else if (tmp == 1) {
|
|
|
|
goto do_setcond_high;
|
|
|
|
}
|
|
|
|
tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[2], op->args[4],
|
|
|
|
TCG_COND_EQ);
|
2014-04-24 09:18:30 +04:00
|
|
|
if (tmp == 0) {
|
|
|
|
goto do_setcond_high;
|
|
|
|
} else if (tmp != 1) {
|
|
|
|
goto do_default;
|
|
|
|
}
|
|
|
|
do_setcond_low:
|
2016-12-08 23:28:42 +03:00
|
|
|
reset_temp(op->args[0]);
|
2017-06-20 23:43:15 +03:00
|
|
|
arg_info(op->args[0])->mask = 1;
|
2014-09-20 00:49:15 +04:00
|
|
|
op->opc = INDEX_op_setcond_i32;
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[2] = op->args[3];
|
|
|
|
op->args[3] = op->args[5];
|
|
|
|
} else if (op->args[5] == TCG_COND_NE) {
|
2014-04-24 09:18:30 +04:00
|
|
|
/* Simplify NE comparisons where one of the pairs
|
|
|
|
can be simplified. */
|
|
|
|
tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[1], op->args[3],
|
|
|
|
TCG_COND_NE);
|
2014-04-24 09:18:30 +04:00
|
|
|
if (tmp == 0) {
|
|
|
|
goto do_setcond_high;
|
|
|
|
} else if (tmp == 1) {
|
|
|
|
goto do_setcond_const;
|
|
|
|
}
|
|
|
|
tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
|
2016-12-08 23:28:42 +03:00
|
|
|
op->args[2], op->args[4],
|
|
|
|
TCG_COND_NE);
|
2014-04-24 09:18:30 +04:00
|
|
|
if (tmp == 0) {
|
|
|
|
goto do_setcond_low;
|
|
|
|
} else if (tmp == 1) {
|
|
|
|
goto do_setcond_const;
|
|
|
|
}
|
|
|
|
goto do_default;
|
2012-10-02 22:32:27 +04:00
|
|
|
} else {
|
|
|
|
goto do_default;
|
2012-10-02 22:32:25 +04:00
|
|
|
}
|
2012-10-02 22:32:27 +04:00
|
|
|
break;
|
2012-10-02 22:32:25 +04:00
|
|
|
|
2011-07-07 16:37:12 +04:00
|
|
|
case INDEX_op_call:
|
2016-12-08 23:28:42 +03:00
|
|
|
if (!(op->args[nb_oargs + nb_iargs + 1]
|
2014-03-23 07:06:52 +04:00
|
|
|
& (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
|
2011-07-07 16:37:13 +04:00
|
|
|
for (i = 0; i < nb_globals; i++) {
|
2015-07-27 13:41:44 +03:00
|
|
|
if (test_bit(i, temps_used.l)) {
|
2017-06-20 23:43:15 +03:00
|
|
|
reset_ts(&s->temps[i]);
|
2015-07-27 13:41:44 +03:00
|
|
|
}
|
2011-07-07 16:37:13 +04:00
|
|
|
}
|
|
|
|
}
|
2014-03-23 07:06:52 +04:00
|
|
|
goto do_reset_output;
|
2012-10-02 22:32:24 +04:00
|
|
|
|
2011-07-07 16:37:12 +04:00
|
|
|
default:
|
2012-10-02 22:32:24 +04:00
|
|
|
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
|
2013-01-12 03:42:52 +04:00
|
|
|
block, otherwise we only trash the output args. "mask" is
|
|
|
|
the non-zero bits mask for the first output arg. */
|
2012-09-19 23:40:30 +04:00
|
|
|
if (def->flags & TCG_OPF_BB_END) {
|
tcg: allocate optimizer temps with tcg_malloc
Groundwork for supporting multiple TCG contexts.
While at it, also allocate temps_used directly as a bitmap of the
required size, instead of using a bitmap of TCG_MAX_TEMPS via
TCGTempSet.
Performance-wise we lose about 1.12% in a translation-heavy workload
such as booting+shutting down debian-arm:
Performance counter stats for 'taskset -c 0 arm-softmmu/qemu-system-arm \
-machine type=virt -nographic -smp 1 -m 4096 \
-netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=die-on-boot.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel kernel.img -append console=ttyAMA0 root=/dev/vda1 \
-name arm,debug-threads=on -smp 1' (10 runs):
exec time (s) Relative slowdown wrt original (%)
---------------------------------------------------------------
original 20.213321616 0.
tcg_malloc 20.441130078 1.1270214
TCGContext 20.477846517 1.3086662
g_malloc 20.780527895 2.8061013
The other two alternatives shown in the table are:
- TCGContext: embed temps[TCG_MAX_TEMPS] and TCGTempSet used_temps
in TCGContext. This is simple enough but it isn't faster than using
tcg_malloc; moreover, it wastes memory.
- g_malloc: allocate/deallocate both temps and used_temps every time
tcg_optimize is executed.
Suggested-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-19 21:32:24 +03:00
|
|
|
bitmap_zero(temps_used.l, nb_temps);
|
2012-09-19 23:40:30 +04:00
|
|
|
} else {
|
2014-03-23 07:06:52 +04:00
|
|
|
do_reset_output:
|
|
|
|
for (i = 0; i < nb_oargs; i++) {
|
2016-12-08 23:28:42 +03:00
|
|
|
reset_temp(op->args[i]);
|
2013-09-03 10:27:38 +04:00
|
|
|
/* Save the corresponding known-zero bits mask for the
|
|
|
|
first output argument (only one supported so far). */
|
|
|
|
if (i == 0) {
|
2017-06-20 23:43:15 +03:00
|
|
|
arg_info(op->args[i])->mask = mask;
|
2013-09-03 10:27:38 +04:00
|
|
|
}
|
2012-09-19 23:40:30 +04:00
|
|
|
}
|
2011-07-07 16:37:13 +04:00
|
|
|
}
|
2011-07-07 16:37:12 +04:00
|
|
|
break;
|
|
|
|
}
|
2016-08-23 16:48:25 +03:00
|
|
|
|
|
|
|
/* Eliminate duplicate and redundant fence instructions. */
|
2016-12-08 23:28:42 +03:00
|
|
|
if (prev_mb) {
|
2016-08-23 16:48:25 +03:00
|
|
|
switch (opc) {
|
|
|
|
case INDEX_op_mb:
|
|
|
|
/* Merge two barriers of the same type into one,
|
|
|
|
* or a weaker barrier into a stronger one,
|
|
|
|
* or two weaker barriers into a stronger one.
|
|
|
|
* mb X; mb Y => mb X|Y
|
|
|
|
* mb; strl => mb; st
|
|
|
|
* ldaq; mb => ld; mb
|
|
|
|
* ldaq; strl => ld; mb; st
|
|
|
|
* Other combinations are also merged into a strong
|
|
|
|
* barrier. This is stricter than specified but for
|
|
|
|
* the purposes of TCG is better than not optimizing.
|
|
|
|
*/
|
2016-12-08 23:28:42 +03:00
|
|
|
prev_mb->args[0] |= op->args[0];
|
2016-08-23 16:48:25 +03:00
|
|
|
tcg_op_remove(s, op);
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
/* Opcodes that end the block stop the optimization. */
|
|
|
|
if ((def->flags & TCG_OPF_BB_END) == 0) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
/* fallthru */
|
|
|
|
case INDEX_op_qemu_ld_i32:
|
|
|
|
case INDEX_op_qemu_ld_i64:
|
|
|
|
case INDEX_op_qemu_st_i32:
|
|
|
|
case INDEX_op_qemu_st_i64:
|
|
|
|
case INDEX_op_call:
|
|
|
|
/* Opcodes that touch guest memory stop the optimization. */
|
2016-12-08 23:28:42 +03:00
|
|
|
prev_mb = NULL;
|
2016-08-23 16:48:25 +03:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
} else if (opc == INDEX_op_mb) {
|
2016-12-08 23:28:42 +03:00
|
|
|
prev_mb = op;
|
2016-08-23 16:48:25 +03:00
|
|
|
}
|
2011-07-07 16:37:12 +04:00
|
|
|
}
|
|
|
|
}
|