struct ir3_cp_ctx {
struct ir3 *shader;
struct ir3_shader_variant *so;
+ bool progress;
};
/* is it a type preserving mov, with ok flags?
* We could possibly do a bit better, and copy-propagation if
* we can CP all components that are being fanned out.
*/
- if (src_instr->opc == OPC_META_FO) {
+ if (src_instr->opc == OPC_META_SPLIT) {
if (!dst_instr)
return false;
- if (dst_instr->opc == OPC_META_FI)
+ if (dst_instr->opc == OPC_META_COLLECT)
return false;
if (dst_instr->cp.left || dst_instr->cp.right)
return false;
return false;
}
-static unsigned cp_flags(unsigned flags)
-{
- /* only considering these flags (at least for now): */
- flags &= (IR3_REG_CONST | IR3_REG_IMMED |
- IR3_REG_FNEG | IR3_REG_FABS |
- IR3_REG_SNEG | IR3_REG_SABS |
- IR3_REG_BNOT | IR3_REG_RELATIV);
- return flags;
-}
-
-static bool valid_flags(struct ir3_instruction *instr, unsigned n,
- unsigned flags)
-{
- struct ir3_compiler *compiler = instr->block->shader->compiler;
- unsigned valid_flags;
-
- if ((flags & IR3_REG_HIGH) &&
- (opc_cat(instr->opc) > 1) &&
- (compiler->gpu_id >= 600))
- return false;
-
- flags = cp_flags(flags);
-
- /* If destination is indirect, then source cannot be.. at least
- * I don't think so..
- */
- if ((instr->regs[0]->flags & IR3_REG_RELATIV) &&
- (flags & IR3_REG_RELATIV))
- return false;
-
- if (flags & IR3_REG_RELATIV) {
- /* TODO need to test on earlier gens.. pretty sure the earlier
- * problem was just that we didn't check that the src was from
- * same block (since we can't propagate address register values
- * across blocks currently)
- */
- if (compiler->gpu_id < 600)
- return false;
-
- /* NOTE in the special try_swap_mad_two_srcs() case we can be
- * called on a src that has already had an indirect load folded
- * in, in which case ssa() returns NULL
- */
- struct ir3_instruction *src = ssa(instr->regs[n+1]);
- if (src && src->address->block != instr->block)
- return false;
- }
-
- switch (opc_cat(instr->opc)) {
- case 1:
- valid_flags = IR3_REG_IMMED | IR3_REG_CONST | IR3_REG_RELATIV;
- if (flags & ~valid_flags)
- return false;
- break;
- case 2:
- valid_flags = ir3_cat2_absneg(instr->opc) |
- IR3_REG_CONST | IR3_REG_RELATIV;
-
- if (ir3_cat2_int(instr->opc))
- valid_flags |= IR3_REG_IMMED;
-
- if (flags & ~valid_flags)
- return false;
-
- if (flags & (IR3_REG_CONST | IR3_REG_IMMED)) {
- unsigned m = (n ^ 1) + 1;
- /* cannot deal w/ const in both srcs:
- * (note that some cat2 actually only have a single src)
- */
- if (m < instr->regs_count) {
- struct ir3_register *reg = instr->regs[m];
- if ((flags & IR3_REG_CONST) && (reg->flags & IR3_REG_CONST))
- return false;
- if ((flags & IR3_REG_IMMED) && (reg->flags & IR3_REG_IMMED))
- return false;
- }
- /* cannot be const + ABS|NEG: */
- if (flags & (IR3_REG_FABS | IR3_REG_FNEG |
- IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT))
- return false;
- }
- break;
- case 3:
- valid_flags = ir3_cat3_absneg(instr->opc) |
- IR3_REG_CONST | IR3_REG_RELATIV;
-
- if (flags & ~valid_flags)
- return false;
-
- if (flags & (IR3_REG_CONST | IR3_REG_RELATIV)) {
- /* cannot deal w/ const/relativ in 2nd src: */
- if (n == 1)
- return false;
- }
-
- if (flags & IR3_REG_CONST) {
- /* cannot be const + ABS|NEG: */
- if (flags & (IR3_REG_FABS | IR3_REG_FNEG |
- IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT))
- return false;
- }
- break;
- case 4:
- /* seems like blob compiler avoids const as src.. */
- /* TODO double check if this is still the case on a4xx */
- if (flags & (IR3_REG_CONST | IR3_REG_IMMED))
- return false;
- if (flags & (IR3_REG_SABS | IR3_REG_SNEG))
- return false;
- break;
- case 5:
- /* no flags allowed */
- if (flags)
- return false;
- break;
- case 6:
- valid_flags = IR3_REG_IMMED;
- if (flags & ~valid_flags)
- return false;
-
- if (flags & IR3_REG_IMMED) {
- /* doesn't seem like we can have immediate src for store
- * instructions:
- *
- * TODO this restriction could also apply to load instructions,
- * but for load instructions this arg is the address (and not
- * really sure any good way to test a hard-coded immed addr src)
- */
- if (is_store(instr) && (n == 1))
- return false;
-
- if ((instr->opc == OPC_LDL) && (n == 0))
- return false;
-
- if ((instr->opc == OPC_STL) && (n != 2))
- return false;
-
- if (instr->opc == OPC_STLW && n == 0)
- return false;
-
- /* disallow CP into anything but the SSBO slot argument for
- * atomics:
- */
- if (is_atomic(instr->opc) && (n != 0))
- return false;
-
- if (is_atomic(instr->opc) && !(instr->flags & IR3_INSTR_G))
- return false;
-
- /* as with atomics, ldib on a6xx can only have immediate for
- * SSBO slot argument
- */
- if ((instr->opc == OPC_LDIB) && (n != 0))
- return false;
- }
-
- break;
- }
-
- return true;
-}
-
/* propagate register flags from src to dst.. negates need special
* handling to cancel each other out.
*/
*dstflags &= ~IR3_REG_SABS;
}
-static struct ir3_register *
-lower_immed(struct ir3_cp_ctx *ctx, struct ir3_register *reg, unsigned new_flags, bool f_opcode)
+/* Tries lowering an immediate register argument to a const buffer access by
+ * adding to the list of immediates to be pushed to the const buffer when
+ * switching to this shader.
+ */
+static bool
+lower_immed(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr, unsigned n,
+ struct ir3_register *reg, unsigned new_flags)
{
- unsigned swiz, idx, i;
+ if (!(new_flags & IR3_REG_IMMED))
+ return false;
+
+ new_flags &= ~IR3_REG_IMMED;
+ new_flags |= IR3_REG_CONST;
+
+ if (!ir3_valid_flags(instr, n, new_flags))
+ return false;
reg = ir3_reg_clone(ctx->shader, reg);
+ /* Half constant registers seems to handle only 32-bit values
+ * within floating-point opcodes. So convert back to 32-bit values.
+ */
+ bool f_opcode = (is_cat2_float(instr->opc) ||
+ is_cat3_float(instr->opc)) ? true : false;
+ if (f_opcode && (new_flags & IR3_REG_HALF))
+ reg->uim_val = fui(_mesa_half_to_float(reg->uim_val));
+
/* in some cases, there are restrictions on (abs)/(neg) plus const..
* so just evaluate those and clear the flags:
*/
new_flags &= ~IR3_REG_FNEG;
}
- /* Reallocate for 4 more elements whenever it's necessary */
- struct ir3_const_state *const_state = &ctx->so->shader->const_state;
- if (const_state->immediate_idx == const_state->immediates_size * 4) {
+ /* Reallocate for 4 more elements whenever it's necessary. Note that ir3
+ * printing relies on having groups of 4 dwords, so we fill the unused
+ * slots with a dummy value.
+ */
+ struct ir3_const_state *const_state = ir3_const_state(ctx->so);
+ if (const_state->immediates_count == const_state->immediates_size) {
+ const_state->immediates = rerzalloc(const_state,
+ const_state->immediates,
+ __typeof__(const_state->immediates[0]),
+ const_state->immediates_size,
+ const_state->immediates_size + 4);
const_state->immediates_size += 4;
- const_state->immediates = realloc (const_state->immediates,
- const_state->immediates_size * sizeof(const_state->immediates[0]));
- }
- for (i = 0; i < const_state->immediate_idx; i++) {
- swiz = i % 4;
- idx = i / 4;
+ for (int i = const_state->immediates_count; i < const_state->immediates_size; i++)
+ const_state->immediates[i] = 0xd0d0d0d0;
+ }
- if (const_state->immediates[idx].val[swiz] == reg->uim_val) {
+ int i;
+ for (i = 0; i < const_state->immediates_count; i++) {
+ if (const_state->immediates[i] == reg->uim_val)
break;
- }
}
- if (i == const_state->immediate_idx) {
- /* need to generate a new immediate: */
- swiz = i % 4;
- idx = i / 4;
-
- /* Half constant registers seems to handle only 32-bit values
- * within floating-point opcodes. So convert back to 32-bit values. */
- if (f_opcode && (new_flags & IR3_REG_HALF)) {
- reg->uim_val = fui(_mesa_half_to_float(reg->uim_val));
- }
+ if (i == const_state->immediates_count) {
+ /* Add on a new immediate to be pushed, if we have space left in the
+ * constbuf.
+ */
+ if (const_state->offsets.immediate + const_state->immediates_count / 4 >=
+ ir3_max_const(ctx->so))
+ return false;
- const_state->immediates[idx].val[swiz] = reg->uim_val;
- const_state->immediates_count = idx + 1;
- const_state->immediate_idx++;
+ const_state->immediates[i] = reg->uim_val;
+ const_state->immediates_count++;
}
- new_flags &= ~IR3_REG_IMMED;
- new_flags |= IR3_REG_CONST;
reg->flags = new_flags;
reg->num = i + (4 * const_state->offsets.immediate);
- return reg;
+ instr->regs[n + 1] = reg;
+
+ return true;
}
static void
bool valid_swap =
/* can we propagate mov if we move 2nd src to first? */
- valid_flags(instr, 0, new_flags) &&
+ ir3_valid_flags(instr, 0, new_flags) &&
/* and does first src fit in second slot? */
- valid_flags(instr, 1, instr->regs[1 + 1]->flags);
+ ir3_valid_flags(instr, 1, instr->regs[1 + 1]->flags);
if (!valid_swap) {
/* put things back the way they were: */
* src (which needs to also fixup the address src reference by the
* instruction).
*/
-static void
+static bool
reg_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr,
struct ir3_register *reg, unsigned n)
{
combine_flags(&new_flags, src);
- if (valid_flags(instr, n, new_flags)) {
+ if (ir3_valid_flags(instr, n, new_flags)) {
if (new_flags & IR3_REG_ARRAY) {
debug_assert(!(reg->flags & IR3_REG_ARRAY));
reg->array = src_reg->array;
unuse(src);
reg->instr->use_count++;
- }
- } else if (is_same_type_mov(src) &&
+ return true;
+ }
+ } else if ((is_same_type_mov(src) || is_const_mov(src)) &&
/* cannot collapse const/immed/etc into meta instrs: */
!is_meta(instr)) {
/* immed/const/etc cases, which require some special handling: */
combine_flags(&new_flags, src);
- if (!valid_flags(instr, n, new_flags)) {
+ if (!ir3_valid_flags(instr, n, new_flags)) {
/* See if lowering an immediate to const would help. */
- if (valid_flags(instr, n, (new_flags & ~IR3_REG_IMMED) | IR3_REG_CONST)) {
- bool f_opcode = (ir3_cat2_float(instr->opc) ||
- ir3_cat3_float(instr->opc)) ? true : false;
-
- debug_assert(new_flags & IR3_REG_IMMED);
-
- instr->regs[n + 1] = lower_immed(ctx, src_reg, new_flags, f_opcode);
- return;
- }
+ if (lower_immed(ctx, instr, n, src_reg, new_flags))
+ return true;
/* special case for "normal" mad instructions, we can
* try swapping the first two args if that fits better.
* src[0] is !CONST and src[1] is CONST:
*/
if ((n == 1) && try_swap_mad_two_srcs(instr, new_flags)) {
- /* we swapped, so now we are dealing with 1st src: */
- n = 0;
+ return true;
} else {
- return;
+ return false;
}
}
*/
if ((src_reg->flags & IR3_REG_RELATIV) &&
conflicts(instr->address, reg->instr->address))
- return;
+ return false;
/* This seems to be a hw bug, or something where the timings
* just somehow don't work out. This restriction may only
if ((opc_cat(instr->opc) == 3) && (n == 2) &&
(src_reg->flags & IR3_REG_RELATIV) &&
(src_reg->array.offset == 0))
- return;
+ return false;
+
+ /* When narrowing constant from 32b to 16b, it seems
+ * to work only for float. So we should do this only with
+ * float opcodes.
+ */
+ if (src->cat1.dst_type == TYPE_F16) {
+ if (instr->opc == OPC_MOV && !type_float(instr->cat1.src_type))
+ return false;
+ if (!is_cat2_float(instr->opc) && !is_cat3_float(instr->opc))
+ return false;
+ }
src_reg = ir3_reg_clone(instr->block->shader, src_reg);
src_reg->flags = new_flags;
if (src_reg->flags & IR3_REG_RELATIV)
ir3_instr_set_address(instr, reg->instr->address);
- return;
+ return true;
}
if ((src_reg->flags & IR3_REG_RELATIV) &&
instr->regs[n+1] = src_reg;
ir3_instr_set_address(instr, reg->instr->address);
- return;
+ return true;
}
/* NOTE: seems we can only do immed integers, so don't
iim_val = ~iim_val;
/* other than category 1 (mov) we can only encode up to 10 bits: */
- if (valid_flags(instr, n, new_flags) &&
+ if (ir3_valid_flags(instr, n, new_flags) &&
((instr->opc == OPC_MOV) ||
!((iim_val & ~0x3ff) && (-iim_val & ~0x3ff)))) {
new_flags &= ~(IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT);
src_reg->flags = new_flags;
src_reg->iim_val = iim_val;
instr->regs[n+1] = src_reg;
- } else if (valid_flags(instr, n, (new_flags & ~IR3_REG_IMMED) | IR3_REG_CONST)) {
- bool f_opcode = (ir3_cat2_float(instr->opc) ||
- ir3_cat3_float(instr->opc)) ? true : false;
- /* See if lowering an immediate to const would help. */
- instr->regs[n+1] = lower_immed(ctx, src_reg, new_flags, f_opcode);
+ return true;
+ } else if (lower_immed(ctx, instr, n, src_reg, new_flags)) {
+ /* Fell back to loading the immediate as a const */
+ return true;
}
- return;
}
}
+
+ return false;
}
/* Handle special case of eliminating output mov, and similar cases where
* be eliminated)
*/
static struct ir3_instruction *
-eliminate_output_mov(struct ir3_instruction *instr)
+eliminate_output_mov(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
{
if (is_eligible_mov(instr, NULL, false)) {
struct ir3_register *reg = instr->regs[1];
if (!(reg->flags & IR3_REG_ARRAY)) {
struct ir3_instruction *src_instr = ssa(reg);
debug_assert(src_instr);
+ ctx->progress = true;
return src_instr;
}
}
static void
instr_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
{
- struct ir3_register *reg;
-
if (instr->regs_count == 0)
return;
return;
/* walk down the graph from each src: */
- foreach_src_n(reg, n, instr) {
- struct ir3_instruction *src = ssa(reg);
+ bool progress;
+ do {
+ progress = false;
+ foreach_src_n (reg, n, instr) {
+ struct ir3_instruction *src = ssa(reg);
- if (!src)
- continue;
+ if (!src)
+ continue;
- instr_cp(ctx, src);
+ instr_cp(ctx, src);
- /* TODO non-indirect access we could figure out which register
- * we actually want and allow cp..
- */
- if (reg->flags & IR3_REG_ARRAY)
- continue;
+ /* TODO non-indirect access we could figure out which register
+ * we actually want and allow cp..
+ */
+ if (reg->flags & IR3_REG_ARRAY)
+ continue;
- /* Don't CP absneg into meta instructions, that won't end well: */
- if (is_meta(instr) && (src->opc != OPC_MOV))
- continue;
+ /* Don't CP absneg into meta instructions, that won't end well: */
+ if (is_meta(instr) && (src->opc != OPC_MOV))
+ continue;
- reg_cp(ctx, instr, reg, n);
- }
+ progress |= reg_cp(ctx, instr, reg, n);
+ ctx->progress |= progress;
+ }
+ } while (progress);
if (instr->regs[0]->flags & IR3_REG_ARRAY) {
struct ir3_instruction *src = ssa(instr->regs[0]);
if (instr->address) {
instr_cp(ctx, instr->address);
- ir3_instr_set_address(instr, eliminate_output_mov(instr->address));
+ ir3_instr_set_address(instr, eliminate_output_mov(ctx, instr->address));
}
/* we can end up with extra cmps.s from frontend, which uses a
(instr->regs[0]->num == regid(REG_P0, 0)) &&
ssa(instr->regs[1]) &&
(instr->regs[2]->flags & IR3_REG_IMMED) &&
- (instr->regs[2]->iim_val == 0)) {
+ (instr->regs[2]->iim_val == 0) &&
+ (instr->cat2.condition == IR3_COND_NE)) {
struct ir3_instruction *cond = ssa(instr->regs[1]);
switch (cond->opc) {
case OPC_CMPS_S:
instr->barrier_class |= cond->barrier_class;
instr->barrier_conflict |= cond->barrier_conflict;
unuse(cond);
+ ctx->progress = true;
break;
default:
break;
}
}
- /* Handle converting a sam.s2en (taking samp/tex idx params via
- * register) into a normal sam (encoding immediate samp/tex idx)
- * if they are immediate. This saves some instructions and regs
- * in the common case where we know samp/tex at compile time:
+ /* Handle converting a sam.s2en (taking samp/tex idx params via register)
+ * into a normal sam (encoding immediate samp/tex idx) if they are
+ * immediate. This saves some instructions and regs in the common case
+ * where we know samp/tex at compile time. This needs to be done in the
+ * frontend for bindless tex, though, so don't replicate it here.
*/
if (is_tex(instr) && (instr->flags & IR3_INSTR_S2EN) &&
+ !(instr->flags & IR3_INSTR_B) &&
!(ir3_shader_debug & IR3_DBG_FORCES2EN)) {
- /* The first src will be a fan-in (collect), if both of it's
+ /* The first src will be a collect, if both of it's
* two sources are mov from imm, then we can
*/
struct ir3_instruction *samp_tex = ssa(instr->regs[1]);
- debug_assert(samp_tex->opc == OPC_META_FI);
+ debug_assert(samp_tex->opc == OPC_META_COLLECT);
struct ir3_instruction *samp = ssa(samp_tex->regs[1]);
struct ir3_instruction *tex = ssa(samp_tex->regs[2]);
instr->flags &= ~IR3_INSTR_S2EN;
instr->cat5.samp = samp->regs[1]->iim_val;
instr->cat5.tex = tex->regs[1]->iim_val;
- instr->regs[1]->instr = NULL;
+
+ /* shuffle around the regs to remove the first src: */
+ instr->regs_count--;
+ for (unsigned i = 1; i < instr->regs_count; i++) {
+ instr->regs[i] = instr->regs[i + 1];
+ }
+
+ ctx->progress = true;
}
}
}
-void
+bool
ir3_cp(struct ir3 *ir, struct ir3_shader_variant *so)
{
struct ir3_cp_ctx ctx = {
* a mov, so we need to do a pass to first count consumers of a
* mov.
*/
- list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
- list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
- struct ir3_instruction *src;
+ foreach_block (block, &ir->block_list) {
+ foreach_instr (instr, &block->instr_list) {
/* by the way, we don't account for false-dep's, so the CP
* pass should always happen before false-dep's are inserted
*/
debug_assert(instr->deps_count == 0);
- foreach_ssa_src(src, instr) {
+ foreach_ssa_src (src, instr) {
src->use_count++;
}
}
ir3_clear_mark(ir);
- for (unsigned i = 0; i < ir->noutputs; i++) {
- if (ir->outputs[i]) {
- instr_cp(&ctx, ir->outputs[i]);
- ir->outputs[i] = eliminate_output_mov(ir->outputs[i]);
- }
+ foreach_output_n (out, n, ir) {
+ instr_cp(&ctx, out);
+ ir->outputs[n] = eliminate_output_mov(&ctx, out);
}
- list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
+ foreach_block (block, &ir->block_list) {
if (block->condition) {
instr_cp(&ctx, block->condition);
- block->condition = eliminate_output_mov(block->condition);
+ block->condition = eliminate_output_mov(&ctx, block->condition);
}
for (unsigned i = 0; i < block->keeps_count; i++) {
instr_cp(&ctx, block->keeps[i]);
- block->keeps[i] = eliminate_output_mov(block->keeps[i]);
+ block->keeps[i] = eliminate_output_mov(&ctx, block->keeps[i]);
}
}
+
+ return ctx.progress;
}