qemu/target/hexagon/mmvec/decode_ext_mmvec.c
Marco Liebel 3fd49e2217 Hexagon (target/hexagon) Fix assignment to tmp registers
The order in which instructions are generated by gen_insn() influences
assignment to tmp registers. During generation, tmp instructions (e.g.
generate_V6_vassign_tmp) use vreg_src_off() to determine what kind of
register to use as source. If some instruction (e.g.
generate_V6_vmpyowh_64_acc) uses a tmp register but is generated prior
to the corresponding tmp instruction, the vregs_updated_tmp bit map
isn't updated in time.

Exmple:
    { v14.tmp = v16; v25 = v14 } This works properly because
    generate_V6_vassign_tmp is generated before generate_V6_vassign
    and the bit map is updated.

    { v15:14.tmp = vcombine(v21, v16); v25:24 += vmpyo(v18.w,v14.h) }
    This does not work properly because vmpyo is generated before
    vcombine and therefore the bit map does not yet know that there's
    a tmp register.

The parentheses in the decoding function were in the wrong place.
Moving them to the correct location makes shuffling of .tmp vector
registers work as expected.

Signed-off-by: Marco Liebel <quic_mliebel@quicinc.com>
Reviewed-by: Taylor Simpson <tsimpson@quicinc.com>
Tested-by: Taylor Simpson <tsimpson@quicinc.com>
Signed-off-by: Taylor Simpson <tsimpson@quicinc.com>
Reviewed-by: Brian Cain <bcain@quicinc.com>
Message-Id: <20230522174708.464197-1-quic_mliebel@quicinc.com>
2023-05-26 07:03:41 -07:00

237 lines
7.7 KiB
C

/*
* Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, see <http://www.gnu.org/licenses/>.
*/
#include "qemu/osdep.h"
#include "decode.h"
#include "opcodes.h"
#include "insn.h"
#include "iclass.h"
#include "mmvec/mmvec.h"
#include "mmvec/decode_ext_mmvec.h"
static void
check_new_value(Packet *pkt)
{
/* .new value for a MMVector store */
int i, j;
const char *reginfo;
const char *destletters;
const char *dststr = NULL;
uint16_t def_opcode;
char letter;
int def_regnum;
for (i = 1; i < pkt->num_insns; i++) {
uint16_t use_opcode = pkt->insn[i].opcode;
if (GET_ATTRIB(use_opcode, A_DOTNEWVALUE) &&
GET_ATTRIB(use_opcode, A_CVI) &&
GET_ATTRIB(use_opcode, A_STORE)) {
int use_regidx = strchr(opcode_reginfo[use_opcode], 's') -
opcode_reginfo[use_opcode];
/*
* What's encoded at the N-field is the offset to who's producing
* the value.
* Shift off the LSB which indicates odd/even register.
*/
int def_off = ((pkt->insn[i].regno[use_regidx]) >> 1);
int def_oreg = pkt->insn[i].regno[use_regidx] & 1;
int def_idx = -1;
for (j = i - 1; (j >= 0) && (def_off >= 0); j--) {
if (!GET_ATTRIB(pkt->insn[j].opcode, A_CVI)) {
continue;
}
def_off--;
if (def_off == 0) {
def_idx = j;
break;
}
}
/*
* Check for a badly encoded N-field which points to an instruction
* out-of-range
*/
g_assert(!((def_off != 0) || (def_idx < 0) ||
(def_idx > (pkt->num_insns - 1))));
/* def_idx is the index of the producer */
def_opcode = pkt->insn[def_idx].opcode;
reginfo = opcode_reginfo[def_opcode];
destletters = "dexy";
for (j = 0; (letter = destletters[j]) != 0; j++) {
dststr = strchr(reginfo, letter);
if (dststr != NULL) {
break;
}
}
if ((dststr == NULL) && GET_ATTRIB(def_opcode, A_CVI_GATHER)) {
def_regnum = 0;
pkt->insn[i].regno[use_regidx] = def_oreg;
pkt->insn[i].new_value_producer_slot = pkt->insn[def_idx].slot;
} else {
if (dststr == NULL) {
/* still not there, we have a bad packet */
g_assert_not_reached();
}
def_regnum = pkt->insn[def_idx].regno[dststr - reginfo];
/* Now patch up the consumer with the register number */
pkt->insn[i].regno[use_regidx] = def_regnum ^ def_oreg;
/* special case for (Vx,Vy) */
dststr = strchr(reginfo, 'y');
if (def_oreg && strchr(reginfo, 'x') && dststr) {
def_regnum = pkt->insn[def_idx].regno[dststr - reginfo];
pkt->insn[i].regno[use_regidx] = def_regnum;
}
/*
* We need to remember who produces this value to later
* check if it was dynamically cancelled
*/
pkt->insn[i].new_value_producer_slot = pkt->insn[def_idx].slot;
}
}
}
}
/*
* We don't want to reorder slot1/slot0 with respect to each other.
* So in our shuffling, we don't want to move the .cur / .tmp vmem earlier
* Instead, we should move the producing instruction later
* But the producing instruction might feed a .new store!
* So we may need to move that even later.
*/
static void
decode_mmvec_move_cvi_to_end(Packet *pkt, int max)
{
int i;
for (i = 0; i < max; i++) {
if (GET_ATTRIB(pkt->insn[i].opcode, A_CVI)) {
int last_inst = pkt->num_insns - 1;
uint16_t last_opcode = pkt->insn[last_inst].opcode;
/*
* If the last instruction is an endloop, move to the one before it
* Keep endloop as the last thing always
*/
if ((last_opcode == J2_endloop0) ||
(last_opcode == J2_endloop1) ||
(last_opcode == J2_endloop01)) {
last_inst--;
}
decode_send_insn_to(pkt, i, last_inst);
max--;
i--; /* Retry this index now that packet has rotated */
}
}
}
static void
decode_shuffle_for_execution_vops(Packet *pkt)
{
/*
* Sort for .new
*/
int i;
for (i = 0; i < pkt->num_insns; i++) {
uint16_t opcode = pkt->insn[i].opcode;
if ((GET_ATTRIB(opcode, A_LOAD) &&
GET_ATTRIB(opcode, A_CVI_NEW)) ||
GET_ATTRIB(opcode, A_CVI_TMP)) {
/*
* Find prior consuming vector instructions
* Move to end of packet
*/
decode_mmvec_move_cvi_to_end(pkt, i);
break;
}
}
/* Move HVX new value stores to the end of the packet */
for (i = 0; i < pkt->num_insns - 1; i++) {
uint16_t opcode = pkt->insn[i].opcode;
if (GET_ATTRIB(opcode, A_STORE) &&
GET_ATTRIB(opcode, A_CVI_NEW) &&
!GET_ATTRIB(opcode, A_CVI_SCATTER_RELEASE)) {
int last_inst = pkt->num_insns - 1;
uint16_t last_opcode = pkt->insn[last_inst].opcode;
/*
* If the last instruction is an endloop, move to the one before it
* Keep endloop as the last thing always
*/
if ((last_opcode == J2_endloop0) ||
(last_opcode == J2_endloop1) ||
(last_opcode == J2_endloop01)) {
last_inst--;
}
decode_send_insn_to(pkt, i, last_inst);
break;
}
}
}
static void
check_for_vhist(Packet *pkt)
{
pkt->vhist_insn = NULL;
for (int i = 0; i < pkt->num_insns; i++) {
Insn *insn = &pkt->insn[i];
int opcode = insn->opcode;
if (GET_ATTRIB(opcode, A_CVI) && GET_ATTRIB(opcode, A_CVI_4SLOT)) {
pkt->vhist_insn = insn;
return;
}
}
}
/*
* Public Functions
*/
SlotMask mmvec_ext_decode_find_iclass_slots(int opcode)
{
if (GET_ATTRIB(opcode, A_CVI_VM)) {
/* HVX memory instruction */
if (GET_ATTRIB(opcode, A_RESTRICT_SLOT0ONLY)) {
return SLOTS_0;
} else if (GET_ATTRIB(opcode, A_RESTRICT_SLOT1ONLY)) {
return SLOTS_1;
}
return SLOTS_01;
} else if (GET_ATTRIB(opcode, A_RESTRICT_SLOT2ONLY)) {
return SLOTS_2;
} else if (GET_ATTRIB(opcode, A_CVI_VX)) {
/* HVX multiply instruction */
return SLOTS_23;
} else if (GET_ATTRIB(opcode, A_CVI_VS_VX)) {
/* HVX permute/shift instruction */
return SLOTS_23;
} else {
return SLOTS_0123;
}
}
void mmvec_ext_decode_checks(Packet *pkt, bool disas_only)
{
check_new_value(pkt);
if (!disas_only) {
decode_shuffle_for_execution_vops(pkt);
}
check_for_vhist(pkt);
}