/*
* Copyright 2013 Vadim Girlin <vadimgirlin@gmail.com>
*
* 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
* on the rights to use, copy, modify, merge, publish, distribute, sub
* license, 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 (including the next
* paragraph) 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 NON-INFRINGEMENT. IN NO EVENT SHALL
* THE AUTHOR(S) AND/OR THEIR SUPPLIERS 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.
*
* Authors:
* Vadim Girlin
*/
#define BCP_DEBUG 0
#if BCP_DEBUG
#define BCP_DUMP(q) do { q } while (0)
#else
#define BCP_DUMP(q)
#endif
extern "C" {
#include "r600_pipe.h"
#include "r600_shader.h"
}
#include <stack>
#include "sb_bc.h"
#include "sb_shader.h"
#include "sb_pass.h"
namespace r600_sb {
int bc_parser::decode() {
dw = bc->bytecode;
bc_ndw = bc->ndw;
max_cf = 0;
dec = new bc_decoder(ctx, dw, bc_ndw);
shader_target t = TARGET_UNKNOWN;
if (pshader) {
switch (bc->type) {
case TGSI_PROCESSOR_FRAGMENT: t = TARGET_PS; break;
case TGSI_PROCESSOR_VERTEX: t = TARGET_VS; break;
case TGSI_PROCESSOR_COMPUTE: t = TARGET_COMPUTE; break;
default: assert(!"unknown shader target"); return -1; break;
}
} else {
if (bc->type == TGSI_PROCESSOR_COMPUTE)
t = TARGET_COMPUTE;
else
t = TARGET_FETCH;
}
sh = new shader(ctx, t, bc->debug_id);
sh->safe_math = sb_context::safe_math || (t == TARGET_COMPUTE);
int r = decode_shader();
delete dec;
sh->ngpr = bc->ngpr;
sh->nstack = bc->nstack;
return r;
}
int bc_parser::decode_shader() {
int r = 0;
unsigned i = 0;
bool eop = false;
sh->init();
do {
eop = false;
if ((r = decode_cf(i, eop)))
return r;
} while (!eop || (i >> 1) <= max_cf);
return 0;
}
int bc_parser::prepare() {
int r = 0;
if ((r = parse_decls()))
return r;
if ((r = prepare_ir()))
return r;
return 0;
}
int bc_parser::parse_decls() {
if (!pshader) {
if (gpr_reladdr)
sh->add_gpr_array(0, bc->ngpr, 0x0F);
// compute shaders have some values preloaded in R0, R1
sh->add_input(0 /* GPR */, true /* preloaded */, 0x0F /* mask */);
sh->add_input(1 /* GPR */, true /* preloaded */, 0x0F /* mask */);
return 0;
}
if (pshader->indirect_files & ~(1 << TGSI_FILE_CONSTANT)) {
assert(pshader->num_arrays);
if (pshader->num_arrays) {
for (unsigned i = 0; i < pshader->num_arrays; ++i) {
r600_shader_array &a = pshader->arrays[i];
sh->add_gpr_array(a.gpr_start, a.gpr_count, a.comp_mask);
}
} else {
sh->add_gpr_array(0, pshader->bc.ngpr, 0x0F);
}
}
if (sh->target == TARGET_VS)
sh->add_input(0, 1, 0x0F);
bool ps_interp = ctx.hw_class >= HW_CLASS_EVERGREEN
&& sh->target == TARGET_PS;
unsigned linear = 0, persp = 0, centroid = 1;
for (unsigned i = 0; i < pshader->ninput; ++i) {
r600_shader_io & in = pshader->input[i];
bool preloaded = sh->target == TARGET_PS && !(ps_interp && in.spi_sid);
sh->add_input(in.gpr, preloaded, /*in.write_mask*/ 0x0F);
if (ps_interp && in.spi_sid) {
if (in.interpolate == TGSI_INTERPOLATE_LINEAR ||
in.interpolate == TGSI_INTERPOLATE_COLOR)
linear = 1;
else if (in.interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
persp = 1;
if (in.centroid)
centroid = 2;
}
}
if (ps_interp) {
unsigned mask = (1 << (2 * (linear + persp) * centroid)) - 1;
unsigned gpr = 0;
while (mask) {
sh->add_input(gpr, true, mask & 0x0F);
++gpr;
mask >>= 4;
}
}
return 0;
}
int bc_parser::decode_cf(unsigned &i, bool &eop) {
int r;
cf_node *cf = sh->create_cf();
sh->root->push_back(cf);
unsigned id = i >> 1;
cf->bc.id = id;
if (cf_map.size() < id + 1)
cf_map.resize(id + 1);
cf_map[id] = cf;
if ((r = dec->decode_cf(i, cf->bc)))
return r;
cf_op_flags flags = (cf_op_flags)cf->bc.op_ptr->flags;
if (flags & CF_ALU) {
if ((r = decode_alu_clause(cf)))
return r;
} else if (flags & CF_FETCH) {
if ((r = decode_fetch_clause(cf)))
return r;;
} else if (flags & CF_EXP) {
if (cf->bc.rw_rel)
gpr_reladdr = true;
assert(!cf->bc.rw_rel);
} else if (flags & (CF_STRM | CF_RAT)) {
if (cf->bc.rw_rel)
gpr_reladdr = true;
assert(!cf->bc.rw_rel);
} else if (flags & CF_BRANCH) {
if (cf->bc.addr > max_cf)
max_cf = cf->bc.addr;
}
eop = cf->bc.end_of_program || cf->bc.op == CF_OP_CF_END ||
cf->bc.op == CF_OP_RET;
return 0;
}
int bc_parser::decode_alu_clause(cf_node* cf) {
unsigned i = cf->bc.addr << 1, cnt = cf->bc.count + 1, gcnt;
cf->subtype = NST_ALU_CLAUSE;
cgroup = 0;
memset(slots[0], 0, 5*sizeof(slots[0][0]));
unsigned ng = 0;
do {
decode_alu_group(cf, i, gcnt);
assert(gcnt <= cnt);
cnt -= gcnt;
ng++;
} while (cnt);
return 0;
}
int bc_parser::decode_alu_group(cf_node* cf, unsigned &i, unsigned &gcnt) {
int r;
alu_node *n;
alu_group_node *g = sh->create_alu_group();
cgroup = !cgroup;
memset(slots[cgroup], 0, 5*sizeof(slots[0][0]));
gcnt = 0;
unsigned literal_mask = 0;
do {
n = sh->create_alu();
g->push_back(n);
if ((r = dec->decode_alu(i, n->bc)))
return r;
if (!sh->assign_slot(n, slots[cgroup])) {
assert(!"alu slot assignment failed");
return -1;
}
gcnt++;
} while (gcnt <= 5 && !n->bc.last);
assert(n->bc.last);
for (node_iterator I = g->begin(), E = g->end(); I != E; ++I) {
n = static_cast<alu_node*>(*I);
if (n->bc.dst_rel)
gpr_reladdr = true;
for (int k = 0; k < n->bc.op_ptr->src_count; ++k) {
bc_alu_src &src = n->bc.src[k];
if (src.rel)
gpr_reladdr = true;
if (src.sel == ALU_SRC_LITERAL) {
literal_mask |= (1 << src.chan);
src.value.u = dw[i + src.chan];
}
}
}
unsigned literal_ndw = 0;
while (literal_mask) {
g->literals.push_back(dw[i + literal_ndw]);
literal_ndw += 1;
literal_mask >>= 1;
}
literal_ndw = (literal_ndw + 1) & ~1u;
i += literal_ndw;
gcnt += literal_ndw >> 1;
cf->push_back(g);
return 0;
}
int bc_parser::prepare_alu_clause(cf_node* cf) {
// loop over alu groups
for (node_iterator I = cf->begin(), E = cf->end(); I != E; ++I) {
assert(I->subtype == NST_ALU_GROUP);
alu_group_node *g = static_cast<alu_group_node*>(*I);
prepare_alu_group(cf, g);
}
return 0;
}
int bc_parser::prepare_alu_group(cf_node* cf, alu_group_node *g) {
alu_node *n;
cgroup = !cgroup;
memset(slots[cgroup], 0, 5*sizeof(slots[0][0]));
for (node_iterator I = g->begin(), E = g->end();
I != E; ++I) {
n = static_cast<alu_node*>(*I);
if (!sh->assign_slot(n, slots[cgroup])) {
assert(!"alu slot assignment failed");
return -1;
}
unsigned src_count = n->bc.op_ptr->src_count;
if (ctx.alu_slots(n->bc.op) & AF_4SLOT)
n->flags |= NF_ALU_4SLOT;
n->src.resize(src_count);
unsigned flags = n->bc.op_ptr->flags;
if (flags & AF_PRED) {
n->dst.resize(3);
if (n->bc.update_pred)
n->dst[1] = sh->get_special_value(SV_ALU_PRED);
if (n->bc.update_exec_mask)
n->dst[2] = sh->get_special_value(SV_EXEC_MASK);
n->flags |= NF_DONT_HOIST;
} else if (flags & AF_KILL) {
n->dst.resize(2);
n->dst[1] = sh->get_special_value(SV_VALID_MASK);
sh->set_uses_kill();
n->flags |= NF_DONT_HOIST | NF_DONT_MOVE |
NF_DONT_KILL | NF_SCHEDULE_EARLY;
} else {
n->dst.resize(1);
}
if (flags & AF_MOVA) {
n->dst[0] = sh->get_special_value(SV_AR_INDEX);
n->flags |= NF_DONT_HOIST;
} else if (n->bc.op_ptr->src_count == 3 || n->bc.write_mask) {
assert(!n->bc.dst_rel || n->bc.index_mode == INDEX_AR_X);
value *v = sh->get_gpr_value(false, n->bc.dst_gpr, n->bc.dst_chan,
n->bc.dst_rel);
n->dst[0] = v;
}
if (n->bc.pred_sel) {
sh->has_alu_predication = true;
n->pred = sh->get_special_value(SV_ALU_PRED);
}
for (unsigned s = 0; s < src_count; ++s) {
bc_alu_src &src = n->bc.src[s];
if (src.sel == ALU_SRC_LITERAL) {
n->src[s] = sh->get_const_value(src.value);
} else if (src.sel == ALU_SRC_PS || src.sel == ALU_SRC_PV) {
unsigned pgroup = !cgroup, prev_slot = src.sel == ALU_SRC_PS ?
SLOT_TRANS : src.chan;
// XXX shouldn't happen but llvm backend uses PS on cayman
if (prev_slot == SLOT_TRANS && ctx.is_cayman())
prev_slot = SLOT_X;
alu_node *prev_alu = slots[pgroup][prev_slot];
assert(prev_alu);
if (!prev_alu->dst[0]) {
value * t = sh->create_temp_value();
prev_alu->dst[0] = t;
}
value *d = prev_alu->dst[0];
if (d->is_rel()) {
d = sh->get_gpr_value(true, prev_alu->bc.dst_gpr,
prev_alu->bc.dst_chan,
prev_alu->bc.dst_rel);
}
n->src[s] = d;
} else if (ctx.is_kcache_sel(src.sel)) {
unsigned sel = src.sel, kc_addr;
unsigned kc_set = ((sel >> 7) & 2) + ((sel >> 5) & 1);
bc_kcache &kc = cf->bc.kc[kc_set];
kc_addr = (kc.addr << 4) + (sel & 0x1F);
n->src[s] = sh->get_kcache_value(kc.bank, kc_addr, src.chan);
} else if (src.sel < MAX_GPR) {
value *v = sh->get_gpr_value(true, src.sel, src.chan, src.rel);
n->src[s] = v;
} else if (src.sel >= ALU_SRC_PARAM_OFFSET) {
// using slot for value channel because in fact the slot
// determines the channel that is loaded by INTERP_LOAD_P0
// (and maybe some others).
// otherwise GVN will consider INTERP_LOAD_P0s with the same
// param index as equal instructions and leave only one of them
n->src[s] = sh->get_special_ro_value(sel_chan(src.sel,
n->bc.slot));
} else {
switch (src.sel) {
case ALU_SRC_0:
n->src[s] = sh->get_const_value(0);
break;
case ALU_SRC_0_5:
n->src[s] = sh->get_const_value(0.5f);
break;
case ALU_SRC_1:
n->src[s] = sh->get_const_value(1.0f);
break;
case ALU_SRC_1_INT:
n->src[s] = sh->get_const_value(1);
break;
case ALU_SRC_M_1_INT:
n->src[s] = sh->get_const_value(-1);
break;
default:
n->src[s] = sh->get_special_ro_value(src.sel);
break;
}
}
}
}
// pack multislot instructions into alu_packed_node
alu_packed_node *p = NULL;
for (node_iterator N, I = g->begin(), E = g->end(); I != E; I = N) {
N = I + 1;
alu_node *a = static_cast<alu_node*>(*I);
unsigned sflags = a->bc.slot_flags;
if (sflags == AF_4V || (ctx.is_cayman() && sflags == AF_S)) {
if (!p)
p = sh->create_alu_packed();
a->remove();
p->push_back(a);
}
}
if (p) {
g->push_front(p);
if (p->count() == 3 && ctx.is_cayman()) {
// cayman's scalar instruction that can use 3 or 4 slots
// FIXME for simplicity we'll always add 4th slot,
// but probably we might want to always remove 4th slot and make
// sure that regalloc won't choose 'w' component for dst
alu_node *f = static_cast<alu_node*>(p->first);
alu_node *a = sh->create_alu();
a->src = f->src;
a->dst.resize(f->dst.size());
a->bc = f->bc;
a->bc.slot = SLOT_W;
p->push_back(a);
}
}
return 0;
}
int bc_parser::decode_fetch_clause(cf_node* cf) {
int r;
unsigned i = cf->bc.addr << 1, cnt = cf->bc.count + 1;
cf->subtype = NST_TEX_CLAUSE;
while (cnt--) {
fetch_node *n = sh->create_fetch();
cf->push_back(n);
if ((r = dec->decode_fetch(i, n->bc)))
return r;
if (n->bc.src_rel || n->bc.dst_rel)
gpr_reladdr = true;
}
return 0;
}
int bc_parser::prepare_fetch_clause(cf_node *cf) {
vvec grad_v, grad_h;
for (node_iterator I = cf->begin(), E = cf->end(); I != E; ++I) {
fetch_node *n = static_cast<fetch_node*>(*I);
assert(n->is_valid());
unsigned flags = n->bc.op_ptr->flags;
unsigned vtx = flags & FF_VTX;
unsigned num_src = vtx ? ctx.vtx_src_num : 4;
n->dst.resize(4);
if (flags & (FF_SETGRAD | FF_USEGRAD | FF_GETGRAD)) {
sh->uses_gradients = true;
}
if (flags & FF_SETGRAD) {
vvec *grad = NULL;
switch (n->bc.op) {
case FETCH_OP_SET_GRADIENTS_V:
grad = &grad_v;
break;
case FETCH_OP_SET_GRADIENTS_H:
grad = &grad_h;
break;
default:
assert(!"unexpected SET_GRAD instruction");
return -1;
}
if (grad->empty())
grad->resize(4);
for(unsigned s = 0; s < 4; ++s) {
unsigned sw = n->bc.src_sel[s];
if (sw <= SEL_W)
(*grad)[s] = sh->get_gpr_value(true, n->bc.src_gpr,
sw, false);
else if (sw == SEL_0)
(*grad)[s] = sh->get_const_value(0.0f);
else if (sw == SEL_1)
(*grad)[s] = sh->get_const_value(1.0f);
}
} else {
if (flags & FF_USEGRAD) {
n->src.resize(12);
std::copy(grad_v.begin(), grad_v.end(), n->src.begin() + 4);
std::copy(grad_h.begin(), grad_h.end(), n->src.begin() + 8);
} else {
n->src.resize(4);
}
for(int s = 0; s < 4; ++s) {
if (n->bc.dst_sel[s] != SEL_MASK)
n->dst[s] = sh->get_gpr_value(false, n->bc.dst_gpr, s, false);
// NOTE: it doesn't matter here which components of the result we
// are using, but original n->bc.dst_sel should be taken into
// account when building the bytecode
}
for(unsigned s = 0; s < num_src; ++s) {
if (n->bc.src_sel[s] <= SEL_W)
n->src[s] = sh->get_gpr_value(true, n->bc.src_gpr,
n->bc.src_sel[s], false);
}
}
}
return 0;
}
int bc_parser::prepare_ir() {
for(id_cf_map::iterator I = cf_map.begin(), E = cf_map.end(); I != E; ++I) {
cf_node *c = *I;
if (!c)
continue;
unsigned flags = c->bc.op_ptr->flags;
if (flags & CF_ALU) {
prepare_alu_clause(c);
} else if (flags & CF_FETCH) {
prepare_fetch_clause(c);
} else if (c->bc.op == CF_OP_CALL_FS) {
sh->init_call_fs(c);
c->flags |= NF_SCHEDULE_EARLY | NF_DONT_MOVE;
} else if (flags & CF_LOOP_START) {
prepare_loop(c);
} else if (c->bc.op == CF_OP_JUMP) {
prepare_if(c);
} else if (c->bc.op == CF_OP_LOOP_END) {
loop_stack.pop();
} else if (c->bc.op == CF_OP_LOOP_CONTINUE) {
assert(!loop_stack.empty());
repeat_node *rep = sh->create_repeat(loop_stack.top());
if (c->parent->first != c)
rep->move(c->parent->first, c);
c->replace_with(rep);
sh->simplify_dep_rep(rep);
} else if (c->bc.op == CF_OP_LOOP_BREAK) {
assert(!loop_stack.empty());
depart_node *dep = sh->create_depart(loop_stack.top());
if (c->parent->first != c)
dep->move(c->parent->first, c);
c->replace_with(dep);
sh->simplify_dep_rep(dep);
} else if (flags & CF_EXP) {
// unroll burst exports
assert(c->bc.op == CF_OP_EXPORT || c->bc.op == CF_OP_EXPORT_DONE);
c->bc.set_op(CF_OP_EXPORT);
unsigned burst_count = c->bc.burst_count;
unsigned eop = c->bc.end_of_program;
c->bc.end_of_program = 0;
c->bc.burst_count = 0;
do {
c->src.resize(4);
for(int s = 0; s < 4; ++s) {
switch (c->bc.sel[s]) {
case SEL_0:
c->src[s] = sh->get_const_value(0.0f);
break;
case SEL_1:
c->src[s] = sh->get_const_value(1.0f);
break;
case SEL_MASK:
break;
default:
if (c->bc.sel[s] <= SEL_W)
c->src[s] = sh->get_gpr_value(true, c->bc.rw_gpr,
c->bc.sel[s], false);
else
assert(!"invalid src_sel for export");
}
}
if (!burst_count--)
break;
cf_node *cf_next = sh->create_cf();
cf_next->bc = c->bc;
++cf_next->bc.rw_gpr;
++cf_next->bc.array_base;
c->insert_after(cf_next);
c = cf_next;
} while (1);
c->bc.end_of_program = eop;
} else if (flags & (CF_STRM | CF_RAT)) {
unsigned burst_count = c->bc.burst_count;
unsigned eop = c->bc.end_of_program;
c->bc.end_of_program = 0;
c->bc.burst_count = 0;
do {
c->src.resize(4);
for(int s = 0; s < 4; ++s) {
if (c->bc.comp_mask & (1 << s))
c->src[s] =
sh->get_gpr_value(true, c->bc.rw_gpr, s, false);
}
if ((flags & CF_RAT) && (c->bc.type & 1)) { // indexed write
c->src.resize(8);
for(int s = 0; s < 3; ++s) {
c->src[4 + s] =
sh->get_gpr_value(true, c->bc.index_gpr, s, false);
}
// FIXME probably we can relax it a bit
c->flags |= NF_DONT_HOIST | NF_DONT_MOVE;
}
if (!burst_count--)
break;
cf_node *cf_next = sh->create_cf();
cf_next->bc = c->bc;
++cf_next->bc.rw_gpr;
// FIXME is it correct?
cf_next->bc.array_base += cf_next->bc.elem_size + 1;
c->insert_after(cf_next);
c = cf_next;
} while (1);
c->bc.end_of_program = eop;
}
}
assert(loop_stack.empty());
return 0;
}
int bc_parser::prepare_loop(cf_node* c) {
cf_node *end = cf_map[c->bc.addr - 1];
assert(end->bc.op == CF_OP_LOOP_END);
assert(c->parent == end->parent);
region_node *reg = sh->create_region();
repeat_node *rep = sh->create_repeat(reg);
reg->push_back(rep);
c->insert_before(reg);
rep->move(c, end->next);
loop_stack.push(reg);
return 0;
}
int bc_parser::prepare_if(cf_node* c) {
cf_node *c_else = NULL, *end = cf_map[c->bc.addr];
BCP_DUMP(
sblog << "parsing JUMP @" << c->bc.id;
sblog << "\n";
);
if (end->bc.op == CF_OP_ELSE) {
BCP_DUMP(
sblog << " found ELSE : ";
dump::dump_op(end);
sblog << "\n";
);
c_else = end;
end = cf_map[c_else->bc.addr];
} else {
BCP_DUMP(
sblog << " no else\n";
);
c_else = end;
}
if (c_else->parent != c->parent)
c_else = NULL;
if (end->parent != c->parent)
end = NULL;
region_node *reg = sh->create_region();
depart_node *dep2 = sh->create_depart(reg);
depart_node *dep = sh->create_depart(reg);
if_node *n_if = sh->create_if();
c->insert_before(reg);
if (c_else != end)
dep->move(c_else, end);
dep2->move(c, end);
reg->push_back(dep);
dep->push_front(n_if);
n_if->push_back(dep2);
n_if->cond = sh->get_special_value(SV_EXEC_MASK);
return 0;
}
} // namespace r600_sb