static int emit_vs_consts( struct svga_context *svga,
unsigned dirty )
{
const struct svga_shader_result *result = svga->state.hw_draw.vs;
const struct svga_vs_compile_key *key = &result->key.vkey;
int ret = 0;
unsigned offset;
/* SVGA_NEW_VS_RESULT
*/
if (result == NULL)
return 0;
/* SVGA_NEW_VS_CONST_BUFFER
*/
ret = emit_consts( svga, 0, PIPE_SHADER_VERTEX );
if (ret)
return ret;
offset = result->shader->info.file_max[TGSI_FILE_CONSTANT] + 1;
/* SVGA_NEW_VS_RESULT
*/
if (key->need_prescale) {
ret = emit_const( svga, PIPE_SHADER_VERTEX, offset++,
svga->state.hw_clear.prescale.scale );
if (ret)
return ret;
ret = emit_const( svga, PIPE_SHADER_VERTEX, offset++,
svga->state.hw_clear.prescale.translate );
if (ret)
return ret;
}
/* SVGA_NEW_ZERO_STRIDE
*/
if (key->zero_stride_vertex_elements) {
unsigned i, curr_zero_stride = 0;
for (i = 0; i < PIPE_MAX_ATTRIBS; ++i) {
if (key->zero_stride_vertex_elements & (1 << i)) {
ret = emit_const( svga, PIPE_SHADER_VERTEX, offset++,
svga->curr.zero_stride_constants +
4 * curr_zero_stride );
if (ret)
return ret;
++curr_zero_stride;
}
}
}
return 0;
}
static enum pipe_error
emit_vs_consts(struct svga_context *svga, unsigned dirty)
{
const struct svga_shader_result *result = svga->state.hw_draw.vs;
const struct svga_vs_compile_key *key;
enum pipe_error ret = PIPE_OK;
unsigned offset;
/* SVGA_NEW_VS_RESULT
*/
if (result == NULL)
return PIPE_OK;
key = &result->key.vkey;
/* SVGA_NEW_VS_CONST_BUFFER
*/
ret = emit_consts( svga, PIPE_SHADER_VERTEX );
if (ret != PIPE_OK)
return ret;
/* offset = number of constants in the VS const buffer */
offset = result->shader->info.file_max[TGSI_FILE_CONSTANT] + 1;
/* SVGA_NEW_VS_PRESCALE
* Put the viewport pre-scale/translate values into the const buffer.
*/
if (key->need_prescale) {
ret = emit_const( svga, PIPE_SHADER_VERTEX, offset++,
svga->state.hw_clear.prescale.scale );
if (ret != PIPE_OK)
return ret;
ret = emit_const( svga, PIPE_SHADER_VERTEX, offset++,
svga->state.hw_clear.prescale.translate );
if (ret != PIPE_OK)
return ret;
}
return PIPE_OK;
}
/**
* Emit all the constants in a constant buffer for a shader stage.
*/
static enum pipe_error
emit_consts(struct svga_context *svga, unsigned shader)
{
struct svga_screen *ss = svga_screen(svga->pipe.screen);
struct pipe_transfer *transfer = NULL;
unsigned count;
const float (*data)[4] = NULL;
unsigned i;
enum pipe_error ret = PIPE_OK;
const unsigned offset = 0;
assert(shader < PIPE_SHADER_TYPES);
if (svga->curr.cb[shader] == NULL)
goto done;
count = svga->curr.cb[shader]->width0 / (4 * sizeof(float));
data = (const float (*)[4])pipe_buffer_map(&svga->pipe,
svga->curr.cb[shader],
PIPE_TRANSFER_READ,
&transfer);
if (data == NULL) {
ret = PIPE_ERROR_OUT_OF_MEMORY;
goto done;
}
if (ss->hw_version >= SVGA3D_HWVERSION_WS8_B1) {
ret = emit_const_range( svga, shader, offset, count, data );
if (ret != PIPE_OK) {
goto done;
}
} else {
for (i = 0; i < count; i++) {
ret = emit_const( svga, shader, offset + i, data[i] );
if (ret != PIPE_OK) {
goto done;
}
}
}
done:
if (data)
pipe_buffer_unmap(&svga->pipe, transfer);
return ret;
}
static int emit_fs_consts( struct svga_context *svga,
unsigned dirty )
{
const struct svga_shader_result *result = svga->state.hw_draw.fs;
const struct svga_fs_compile_key *key = &result->key.fkey;
int ret = 0;
ret = emit_consts( svga, 0, PIPE_SHADER_FRAGMENT );
if (ret)
return ret;
/* The internally generated fragment shader for xor blending
* doesn't have a 'result' struct. It should be fixed to avoid
* this special case, but work around it with a NULL check:
*/
if (result != NULL &&
key->num_unnormalized_coords)
{
unsigned offset = result->shader->info.file_max[TGSI_FILE_CONSTANT] + 1;
int i;
for (i = 0; i < key->num_textures; i++) {
if (key->tex[i].unnormalized) {
struct pipe_resource *tex = svga->curr.sampler_views[i]->texture;
float data[4];
data[0] = 1.0 / (float)tex->width0;
data[1] = 1.0 / (float)tex->height0;
data[2] = 1.0;
data[3] = 1.0;
ret = emit_const( svga,
PIPE_SHADER_FRAGMENT,
key->tex[i].width_height_idx + offset,
data );
if (ret)
return ret;
}
}
offset += key->num_unnormalized_coords;
}
return 0;
}
static int emit_consts( struct svga_context *svga,
int offset,
int unit )
{
struct pipe_transfer *transfer = NULL;
unsigned count;
const float (*data)[4] = NULL;
unsigned i;
int ret = PIPE_OK;
if (svga->curr.cb[unit] == NULL)
goto done;
count = svga->curr.cb[unit]->width0 / (4 * sizeof(float));
data = (const float (*)[4])pipe_buffer_map(&svga->pipe,
svga->curr.cb[unit],
PIPE_TRANSFER_READ,
&transfer);
if (data == NULL) {
ret = PIPE_ERROR_OUT_OF_MEMORY;
goto done;
}
for (i = 0; i < count; i++) {
ret = emit_const( svga, unit, offset + i, data[i] );
if (ret)
goto done;
}
done:
if (data)
pipe_buffer_unmap(&svga->pipe, svga->curr.cb[unit], transfer);
return ret;
}
enum eval_result_type
compile_bytecodes (struct agent_expr *aexpr)
{
int pc = 0;
int done = 0;
unsigned char op, next_op;
int arg;
/* This is only used to build 64-bit value for constants. */
ULONGEST top;
struct bytecode_address *aentry, *aentry2;
#define UNHANDLED \
do \
{ \
ax_debug ("Cannot compile op 0x%x\n", op); \
return expr_eval_unhandled_opcode; \
} while (0)
if (aexpr->length == 0)
{
ax_debug ("empty agent expression\n");
return expr_eval_empty_expression;
}
bytecode_address_table = NULL;
while (!done)
{
op = aexpr->bytes[pc];
ax_debug ("About to compile op 0x%x, pc=%d\n", op, pc);
/* Record the compiled-code address of the bytecode, for use by
jump instructions. */
aentry = XNEW (struct bytecode_address);
aentry->pc = pc;
aentry->address = current_insn_ptr;
aentry->goto_pc = -1;
aentry->from_offset = aentry->from_size = 0;
aentry->next = bytecode_address_table;
bytecode_address_table = aentry;
++pc;
emit_error = 0;
switch (op)
{
case gdb_agent_op_add:
emit_add ();
break;
case gdb_agent_op_sub:
emit_sub ();
break;
case gdb_agent_op_mul:
emit_mul ();
break;
case gdb_agent_op_div_signed:
UNHANDLED;
break;
case gdb_agent_op_div_unsigned:
UNHANDLED;
break;
case gdb_agent_op_rem_signed:
UNHANDLED;
break;
case gdb_agent_op_rem_unsigned:
UNHANDLED;
break;
case gdb_agent_op_lsh:
emit_lsh ();
break;
case gdb_agent_op_rsh_signed:
emit_rsh_signed ();
break;
case gdb_agent_op_rsh_unsigned:
emit_rsh_unsigned ();
break;
case gdb_agent_op_trace:
UNHANDLED;
break;
case gdb_agent_op_trace_quick:
UNHANDLED;
break;
case gdb_agent_op_log_not:
emit_log_not ();
break;
case gdb_agent_op_bit_and:
emit_bit_and ();
break;
case gdb_agent_op_bit_or:
emit_bit_or ();
break;
case gdb_agent_op_bit_xor:
emit_bit_xor ();
break;
case gdb_agent_op_bit_not:
emit_bit_not ();
break;
case gdb_agent_op_equal:
next_op = aexpr->bytes[pc];
if (next_op == gdb_agent_op_if_goto
&& !is_goto_target (aexpr, pc)
&& target_emit_ops ()->emit_eq_goto)
{
ax_debug ("Combining equal & if_goto");
pc += 1;
aentry->pc = pc;
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
aentry->goto_pc = arg;
emit_eq_goto (&(aentry->from_offset), &(aentry->from_size));
}
else if (next_op == gdb_agent_op_log_not
&& (aexpr->bytes[pc + 1] == gdb_agent_op_if_goto)
&& !is_goto_target (aexpr, pc + 1)
&& target_emit_ops ()->emit_ne_goto)
{
ax_debug ("Combining equal & log_not & if_goto");
pc += 2;
aentry->pc = pc;
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
aentry->goto_pc = arg;
emit_ne_goto (&(aentry->from_offset), &(aentry->from_size));
}
else
emit_equal ();
break;
case gdb_agent_op_less_signed:
next_op = aexpr->bytes[pc];
if (next_op == gdb_agent_op_if_goto
&& !is_goto_target (aexpr, pc))
{
ax_debug ("Combining less_signed & if_goto");
pc += 1;
aentry->pc = pc;
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
aentry->goto_pc = arg;
emit_lt_goto (&(aentry->from_offset), &(aentry->from_size));
}
else if (next_op == gdb_agent_op_log_not
&& !is_goto_target (aexpr, pc)
&& (aexpr->bytes[pc + 1] == gdb_agent_op_if_goto)
&& !is_goto_target (aexpr, pc + 1))
{
ax_debug ("Combining less_signed & log_not & if_goto");
pc += 2;
aentry->pc = pc;
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
aentry->goto_pc = arg;
emit_ge_goto (&(aentry->from_offset), &(aentry->from_size));
}
else
emit_less_signed ();
break;
case gdb_agent_op_less_unsigned:
emit_less_unsigned ();
break;
case gdb_agent_op_ext:
arg = aexpr->bytes[pc++];
if (arg < (sizeof (LONGEST) * 8))
emit_ext (arg);
break;
case gdb_agent_op_ref8:
emit_ref (1);
break;
case gdb_agent_op_ref16:
emit_ref (2);
break;
case gdb_agent_op_ref32:
emit_ref (4);
break;
case gdb_agent_op_ref64:
emit_ref (8);
break;
case gdb_agent_op_if_goto:
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
aentry->goto_pc = arg;
emit_if_goto (&(aentry->from_offset), &(aentry->from_size));
break;
case gdb_agent_op_goto:
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
aentry->goto_pc = arg;
emit_goto (&(aentry->from_offset), &(aentry->from_size));
break;
case gdb_agent_op_const8:
emit_stack_flush ();
top = aexpr->bytes[pc++];
emit_const (top);
break;
case gdb_agent_op_const16:
emit_stack_flush ();
top = aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
emit_const (top);
break;
case gdb_agent_op_const32:
emit_stack_flush ();
top = aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
emit_const (top);
break;
case gdb_agent_op_const64:
emit_stack_flush ();
top = aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
top = (top << 8) + aexpr->bytes[pc++];
emit_const (top);
break;
case gdb_agent_op_reg:
emit_stack_flush ();
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
emit_reg (arg);
break;
case gdb_agent_op_end:
ax_debug ("At end of expression\n");
/* Assume there is one stack element left, and that it is
cached in "top" where emit_epilogue can get to it. */
emit_stack_adjust (1);
done = 1;
break;
case gdb_agent_op_dup:
/* In our design, dup is equivalent to stack flushing. */
emit_stack_flush ();
break;
case gdb_agent_op_pop:
emit_pop ();
break;
case gdb_agent_op_zero_ext:
arg = aexpr->bytes[pc++];
if (arg < (sizeof (LONGEST) * 8))
emit_zero_ext (arg);
break;
case gdb_agent_op_swap:
next_op = aexpr->bytes[pc];
/* Detect greater-than comparison sequences. */
if (next_op == gdb_agent_op_less_signed
&& !is_goto_target (aexpr, pc)
&& (aexpr->bytes[pc + 1] == gdb_agent_op_if_goto)
&& !is_goto_target (aexpr, pc + 1))
{
ax_debug ("Combining swap & less_signed & if_goto");
pc += 2;
aentry->pc = pc;
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
aentry->goto_pc = arg;
emit_gt_goto (&(aentry->from_offset), &(aentry->from_size));
}
else if (next_op == gdb_agent_op_less_signed
&& !is_goto_target (aexpr, pc)
&& (aexpr->bytes[pc + 1] == gdb_agent_op_log_not)
&& !is_goto_target (aexpr, pc + 1)
&& (aexpr->bytes[pc + 2] == gdb_agent_op_if_goto)
&& !is_goto_target (aexpr, pc + 2))
{
ax_debug ("Combining swap & less_signed & log_not & if_goto");
pc += 3;
aentry->pc = pc;
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
aentry->goto_pc = arg;
emit_le_goto (&(aentry->from_offset), &(aentry->from_size));
}
else
emit_swap ();
break;
case gdb_agent_op_getv:
emit_stack_flush ();
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
emit_int_call_1 (get_get_tsv_func_addr (),
arg);
break;
case gdb_agent_op_setv:
arg = aexpr->bytes[pc++];
arg = (arg << 8) + aexpr->bytes[pc++];
emit_void_call_2 (get_set_tsv_func_addr (),
arg);
break;
case gdb_agent_op_tracev:
UNHANDLED;
break;
/* GDB never (currently) generates any of these ops. */
case gdb_agent_op_float:
case gdb_agent_op_ref_float:
case gdb_agent_op_ref_double:
case gdb_agent_op_ref_long_double:
case gdb_agent_op_l_to_d:
case gdb_agent_op_d_to_l:
case gdb_agent_op_trace16:
UNHANDLED;
break;
default:
ax_debug ("Agent expression op 0x%x not recognized\n", op);
/* Don't struggle on, things will just get worse. */
return expr_eval_unrecognized_opcode;
}
/* This catches errors that occur in target-specific code
emission. */
if (emit_error)
{
ax_debug ("Error %d while emitting code for %s\n",
emit_error, gdb_agent_op_name (op));
return expr_eval_unhandled_opcode;
}
ax_debug ("Op %s compiled\n", gdb_agent_op_name (op));
}
/* Now fill in real addresses as goto destinations. */
for (aentry = bytecode_address_table; aentry; aentry = aentry->next)
{
int written = 0;
if (aentry->goto_pc < 0)
continue;
/* Find the location that we are going to, and call back into
target-specific code to write the actual address or
displacement. */
for (aentry2 = bytecode_address_table; aentry2; aentry2 = aentry2->next)
{
if (aentry2->pc == aentry->goto_pc)
{
ax_debug ("Want to jump from %s to %s\n",
paddress (aentry->address),
paddress (aentry2->address));
write_goto_address (aentry->address + aentry->from_offset,
aentry2->address, aentry->from_size);
written = 1;
break;
}
}
/* Error out if we didn't find a destination. */
if (!written)
{
ax_debug ("Destination of goto %d not found\n",
aentry->goto_pc);
return expr_eval_invalid_goto;
}
}
return expr_eval_no_error;
}
/**
* Emit all the constants in a constant buffer for a shader stage.
* On VGPU10, emit_consts_vgpu10 is used instead.
*/
static enum pipe_error
emit_consts_vgpu9(struct svga_context *svga, unsigned shader)
{
const struct pipe_constant_buffer *cbuf;
struct svga_screen *ss = svga_screen(svga->pipe.screen);
struct pipe_transfer *transfer = NULL;
unsigned count;
const float (*data)[4] = NULL;
unsigned i;
enum pipe_error ret = PIPE_OK;
const unsigned offset = 0;
assert(shader < PIPE_SHADER_TYPES);
assert(!svga_have_vgpu10(svga));
/* Only one constant buffer per shader is supported before VGPU10.
* This is only an approximate check against that.
*/
assert(svga->curr.constbufs[shader][1].buffer == NULL);
cbuf = &svga->curr.constbufs[shader][0];
if (svga->curr.constbufs[shader][0].buffer) {
/* emit user-provided constants */
data = (const float (*)[4])
pipe_buffer_map(&svga->pipe, svga->curr.constbufs[shader][0].buffer,
PIPE_TRANSFER_READ, &transfer);
if (!data) {
return PIPE_ERROR_OUT_OF_MEMORY;
}
/* sanity check */
assert(cbuf->buffer->width0 >=
cbuf->buffer_size);
/* Use/apply the constant buffer size and offsets here */
count = cbuf->buffer_size / (4 * sizeof(float));
data += cbuf->buffer_offset / (4 * sizeof(float));
if (ss->hw_version >= SVGA3D_HWVERSION_WS8_B1) {
ret = emit_const_range( svga, shader, offset, count, data );
}
else {
for (i = 0; i < count; i++) {
ret = emit_const( svga, shader, offset + i, data[i] );
if (ret != PIPE_OK) {
break;
}
}
}
pipe_buffer_unmap(&svga->pipe, transfer);
if (ret != PIPE_OK) {
return ret;
}
}
/* emit extra shader constants */
{
const struct svga_shader_variant *variant = NULL;
unsigned offset;
float extras[MAX_EXTRA_CONSTS][4];
unsigned count, i;
switch (shader) {
case PIPE_SHADER_VERTEX:
variant = svga->state.hw_draw.vs;
count = svga_get_extra_vs_constants(svga, (float *) extras);
break;
case PIPE_SHADER_FRAGMENT:
variant = svga->state.hw_draw.fs;
count = svga_get_extra_fs_constants(svga, (float *) extras);
break;
default:
assert(!"Unexpected shader type");
count = 0;
}
assert(variant);
offset = variant->shader->info.file_max[TGSI_FILE_CONSTANT] + 1;
assert(count <= ARRAY_SIZE(extras));
if (count > 0) {
if (ss->hw_version >= SVGA3D_HWVERSION_WS8_B1) {
ret = emit_const_range(svga, shader, offset, count,
(const float (*) [4])extras);
}
else {
for (i = 0; i < count; i++) {
ret = emit_const(svga, shader, offset + i, extras[i]);
if (ret != PIPE_OK)
return ret;
}
}
}
}
return ret;
}
