#include "freedreno_util.h"
#include "ir3.h"
+#include "ir3_shader.h"
/*
* Copy Propagate:
- *
*/
-static void block_cp(struct ir3_block *block);
-static struct ir3_instruction * instr_cp(struct ir3_instruction *instr, bool keep);
+struct ir3_cp_ctx {
+ struct ir3 *shader;
+ struct ir3_shader_variant *so;
+ unsigned immediate_idx;
+};
-/* XXX move this somewhere useful (and rename?) */
-static struct ir3_instruction *ssa(struct ir3_register *reg)
+/* is it a type preserving mov, with ok flags? */
+static bool is_eligible_mov(struct ir3_instruction *instr, bool allow_flags)
{
- if (reg->flags & IR3_REG_SSA)
- return reg->instr;
- return NULL;
-}
+ if (is_same_type_mov(instr)) {
+ struct ir3_register *dst = instr->regs[0];
+ struct ir3_register *src = instr->regs[1];
+ struct ir3_instruction *src_instr = ssa(src);
-static bool conflicts(struct ir3_instruction *a, struct ir3_instruction *b)
-{
- return (a && b) && (a != b);
-}
+ /* only if mov src is SSA (not const/immed): */
+ if (!src_instr)
+ return false;
-static void set_neighbors(struct ir3_instruction *instr,
- struct ir3_instruction *left, struct ir3_instruction *right)
-{
- debug_assert(!conflicts(instr->cp.left, left));
- if (left) {
- instr->cp.left_cnt++;
- instr->cp.left = left;
- }
- debug_assert(!conflicts(instr->cp.right, right));
- if (right) {
- instr->cp.right_cnt++;
- instr->cp.right = right;
+ /* no indirect: */
+ if (dst->flags & IR3_REG_RELATIV)
+ return false;
+ if (src->flags & IR3_REG_RELATIV)
+ return false;
+
+ if (!allow_flags)
+ if (src->flags & (IR3_REG_FABS | IR3_REG_FNEG |
+ IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT))
+ return false;
+
+ /* TODO: remove this hack: */
+ if (src_instr->opc == OPC_META_FO)
+ return false;
+ /* TODO: we currently don't handle left/right neighbors
+ * very well when inserting parallel-copies into phi..
+ * to avoid problems don't eliminate a mov coming out
+ * of phi..
+ */
+ if (src_instr->opc == OPC_META_PHI)
+ return false;
+ return true;
}
+ return false;
}
-/* remove neighbor reference, clearing left/right neighbor ptrs when
- * there are no more references:
- */
-static void remove_neighbors(struct ir3_instruction *instr)
+static unsigned cp_flags(unsigned flags)
{
- if (instr->cp.left) {
- if (--instr->cp.left_cnt == 0)
- instr->cp.left = NULL;
- }
- if (instr->cp.right) {
- if (--instr->cp.right_cnt == 0)
- instr->cp.right = NULL;
- }
+ /* 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;
}
-/* stop condition for iteration: */
-static bool check_stop(struct ir3_instruction *instr)
+static bool valid_flags(struct ir3_instruction *instr, unsigned n,
+ unsigned flags)
{
- if (ir3_instr_check_mark(instr))
- return true;
+ unsigned valid_flags;
+ flags = cp_flags(flags);
- /* stay within the block.. don't try to operate across
- * basic block boundaries or we'll have problems when
- * dealing with multiple basic blocks:
+ /* If destination is indirect, then source cannot be.. at least
+ * I don't think so..
*/
- if (is_meta(instr) && (instr->opc == OPC_META_INPUT))
- return true;
+ if ((instr->regs[0]->flags & IR3_REG_RELATIV) &&
+ (flags & IR3_REG_RELATIV))
+ return false;
+
+ /* TODO it seems to *mostly* work to cp RELATIV, except we get some
+ * intermittent piglit variable-indexing fails. Newer blob driver
+ * doesn't seem to cp these. Possibly this is hw workaround? Not
+ * sure, but until that is understood better, lets just switch off
+ * cp for indirect src's:
+ */
+ if (flags & IR3_REG_RELATIV)
+ return false;
- 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;
-static bool is_eligible_mov(struct ir3_instruction *instr)
-{
- if ((instr->category == 1) &&
- (instr->cat1.src_type == instr->cat1.dst_type)) {
- struct ir3_register *dst = instr->regs[0];
- struct ir3_register *src = instr->regs[1];
- struct ir3_instruction *src_instr = ssa(src);
- if (dst->flags & IR3_REG_ADDR)
+ if (ir3_cat2_int(instr->opc))
+ valid_flags |= IR3_REG_IMMED;
+
+ if (flags & ~valid_flags)
return false;
- /* TODO: propagate abs/neg modifiers if possible */
- if (src->flags & (IR3_REG_ABS | IR3_REG_NEGATE | IR3_REG_RELATIV))
+
+ 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 (src_instr) {
- /* check that eliminating the move won't result in
- * a neighbor conflict, ie. if an instruction feeds
- * into multiple fanins it can still only have at
- * most one left and one right neighbor:
+
+ 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 (conflicts(instr->cp.left, src_instr->cp.left))
+ if (is_store(instr) && (n == 1))
return false;
- if (conflicts(instr->cp.right, src_instr->cp.right))
+
+ /* disallow CP into anything but the SSBO slot argument for
+ * atomics:
+ */
+ if (is_atomic(instr->opc) && (n != 0))
return false;
- return true;
}
+
+ break;
}
- return false;
+
+ return true;
}
-static void walk_children(struct ir3_instruction *instr, bool keep)
+/* propagate register flags from src to dst.. negates need special
+ * handling to cancel each other out.
+ */
+static void combine_flags(unsigned *dstflags, struct ir3_instruction *src)
{
- unsigned i;
+ unsigned srcflags = src->regs[1]->flags;
- /* walk down the graph from each src: */
- for (i = 1; i < instr->regs_count; i++) {
- struct ir3_register *src = instr->regs[i];
- if (src->flags & IR3_REG_SSA)
- src->instr = instr_cp(src->instr, keep);
- }
+ /* if what we are combining into already has (abs) flags,
+ * we can drop (neg) from src:
+ */
+ if (*dstflags & IR3_REG_FABS)
+ srcflags &= ~IR3_REG_FNEG;
+ if (*dstflags & IR3_REG_SABS)
+ srcflags &= ~IR3_REG_SNEG;
+
+ if (srcflags & IR3_REG_FABS)
+ *dstflags |= IR3_REG_FABS;
+ if (srcflags & IR3_REG_SABS)
+ *dstflags |= IR3_REG_SABS;
+ if (srcflags & IR3_REG_FNEG)
+ *dstflags ^= IR3_REG_FNEG;
+ if (srcflags & IR3_REG_SNEG)
+ *dstflags ^= IR3_REG_SNEG;
+ if (srcflags & IR3_REG_BNOT)
+ *dstflags ^= IR3_REG_BNOT;
+
+ *dstflags &= ~IR3_REG_SSA;
+ *dstflags |= srcflags & IR3_REG_SSA;
+ *dstflags |= srcflags & IR3_REG_CONST;
+ *dstflags |= srcflags & IR3_REG_IMMED;
+ *dstflags |= srcflags & IR3_REG_RELATIV;
+ *dstflags |= srcflags & IR3_REG_ARRAY;
+
+ /* if src of the src is boolean we can drop the (abs) since we know
+ * the source value is already a postitive integer. This cleans
+ * up the absnegs that get inserted when converting between nir and
+ * native boolean (see ir3_b2n/n2b)
+ */
+ struct ir3_instruction *srcsrc = ssa(src->regs[1]);
+ if (srcsrc && is_bool(srcsrc))
+ *dstflags &= ~IR3_REG_SABS;
}
-static struct ir3_instruction *
-instr_cp_fanin(struct ir3_instruction *instr)
+static struct ir3_register *
+lower_immed(struct ir3_cp_ctx *ctx, struct ir3_register *reg, unsigned new_flags)
{
- unsigned i, j;
+ unsigned swiz, idx, i;
- /* we need to handle fanin specially, to detect cases
- * when we need to keep a mov
+ reg = ir3_reg_clone(ctx->shader, reg);
+
+ /* in some cases, there are restrictions on (abs)/(neg) plus const..
+ * so just evaluate those and clear the flags:
*/
+ if (new_flags & IR3_REG_SABS) {
+ reg->iim_val = abs(reg->iim_val);
+ new_flags &= ~IR3_REG_SABS;
+ }
- for (i = 1; i < instr->regs_count; i++) {
- struct ir3_register *src = instr->regs[i];
- if (src->flags & IR3_REG_SSA) {
- struct ir3_instruction *cand =
- instr_cp(src->instr, false);
+ if (new_flags & IR3_REG_FABS) {
+ reg->fim_val = fabs(reg->fim_val);
+ new_flags &= ~IR3_REG_FABS;
+ }
- /* if the candidate is a fanout, then keep
- * the move.
- *
- * This is a bit, um, fragile, but it should
- * catch the extra mov's that the front-end
- * puts in for us already in these cases.
- */
- if (is_meta(cand) && (cand->opc == OPC_META_FO))
- cand = instr_cp(src->instr, true);
+ if (new_flags & IR3_REG_SNEG) {
+ reg->iim_val = -reg->iim_val;
+ new_flags &= ~IR3_REG_SNEG;
+ }
- /* we can't have 2 registers referring to the same instruction, so
- * go through and check if any already refer to the candidate
- * instruction. if so, don't do the propagation.
- *
- * NOTE: we need to keep this, despite the neighbor
- * conflict checks, to avoid A<->B<->A..
- */
- for (j = 1; j < instr->regs_count; j++)
- if (instr->regs[j]->instr == cand)
- break;
- if (j == instr->regs_count)
- src->instr = cand;
- }
+ if (new_flags & IR3_REG_FNEG) {
+ reg->fim_val = -reg->fim_val;
+ new_flags &= ~IR3_REG_FNEG;
}
- walk_children(instr, false);
+ for (i = 0; i < ctx->immediate_idx; i++) {
+ swiz = i % 4;
+ idx = i / 4;
- return instr;
-}
+ if (ctx->so->immediates[idx].val[swiz] == reg->uim_val) {
+ break;
+ }
+ }
-static struct ir3_instruction *
-instr_cp(struct ir3_instruction *instr, bool keep)
-{
- /* if we've already visited this instruction, bail now: */
- if (check_stop(instr))
- return instr;
-
- if (is_meta(instr) && (instr->opc == OPC_META_FI))
- return instr_cp_fanin(instr);
-
- if (!keep && is_eligible_mov(instr)) {
- struct ir3_instruction *src_instr = ssa(instr->regs[1]);
- set_neighbors(src_instr, instr->cp.left, instr->cp.right);
- remove_neighbors(instr);
- return instr_cp(src_instr, false);
+ if (i == ctx->immediate_idx) {
+ /* need to generate a new immediate: */
+ swiz = i % 4;
+ idx = i / 4;
+ ctx->so->immediates[idx].val[swiz] = reg->uim_val;
+ ctx->so->immediates_count = idx + 1;
+ ctx->immediate_idx++;
}
- walk_children(instr, false);
+ new_flags &= ~IR3_REG_IMMED;
+ new_flags |= IR3_REG_CONST;
+ reg->flags = new_flags;
+ reg->num = i + (4 * ctx->so->constbase.immediate);
- return instr;
+ return reg;
}
-static void block_cp(struct ir3_block *block)
+/**
+ * Handle cp for a given src register. This additionally handles
+ * the cases of collapsing immedate/const (which replace the src
+ * register with a non-ssa src) or collapsing mov's from relative
+ * src (which needs to also fixup the address src reference by the
+ * instruction).
+ */
+static void
+reg_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr,
+ struct ir3_register *reg, unsigned n)
{
- unsigned i, j;
+ struct ir3_instruction *src = ssa(reg);
- for (i = 0; i < block->noutputs; i++) {
- if (block->outputs[i]) {
- struct ir3_instruction *out =
- instr_cp(block->outputs[i], false);
+ /* don't propagate copies into a PHI, since we don't know if the
+ * src block executed:
+ */
+ if (instr->opc == OPC_META_PHI)
+ return;
- /* To deal with things like this:
- *
- * 43: MOV OUT[2], TEMP[5]
- * 44: MOV OUT[0], TEMP[5]
+ if (is_eligible_mov(src, true)) {
+ /* simple case, no immed/const/relativ, only mov's w/ ssa src: */
+ struct ir3_register *src_reg = src->regs[1];
+ unsigned new_flags = reg->flags;
+
+ combine_flags(&new_flags, src);
+
+ if (valid_flags(instr, n, new_flags)) {
+ if (new_flags & IR3_REG_ARRAY) {
+ debug_assert(!(reg->flags & IR3_REG_ARRAY));
+ reg->array = src_reg->array;
+ }
+ reg->flags = new_flags;
+ reg->instr = ssa(src_reg);
+ }
+
+ src = ssa(reg); /* could be null for IR3_REG_ARRAY case */
+ if (!src)
+ return;
+ } else if (is_same_type_mov(src) &&
+ /* cannot collapse const/immed/etc into meta instrs: */
+ !is_meta(instr)) {
+ /* immed/const/etc cases, which require some special handling: */
+ struct ir3_register *src_reg = src->regs[1];
+ unsigned new_flags = reg->flags;
+
+ combine_flags(&new_flags, src);
+
+ if (!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)) {
+ debug_assert(new_flags & IR3_REG_IMMED);
+ instr->regs[n + 1] = lower_immed(ctx, src_reg, new_flags);
+ return;
+ }
+
+ /* special case for "normal" mad instructions, we can
+ * try swapping the first two args if that fits better.
*
- * we need to ensure that no two outputs point to
- * the same instruction
+ * the "plain" MAD's (ie. the ones that don't shift first
+ * src prior to multiply) can swap their first two srcs if
+ * src[0] is !CONST and src[1] is CONST:
*/
- for (j = 0; j < i; j++) {
- if (block->outputs[j] == out) {
- out = instr_cp(block->outputs[i], true);
- break;
- }
+ if ((n == 1) && is_mad(instr->opc) &&
+ !(instr->regs[0 + 1]->flags & (IR3_REG_CONST | IR3_REG_RELATIV)) &&
+ valid_flags(instr, 0, new_flags)) {
+ /* swap src[0] and src[1]: */
+ struct ir3_register *tmp;
+ tmp = instr->regs[0 + 1];
+ instr->regs[0 + 1] = instr->regs[1 + 1];
+ instr->regs[1 + 1] = tmp;
+ n = 0;
+ } else {
+ return;
}
-
- block->outputs[i] = out;
}
- }
-}
-/*
- * Find instruction neighbors:
- */
+ /* Here we handle the special case of mov from
+ * CONST and/or RELATIV. These need to be handled
+ * specially, because in the case of move from CONST
+ * there is no src ir3_instruction so we need to
+ * replace the ir3_register. And in the case of
+ * RELATIV we need to handle the address register
+ * dependency.
+ */
+ if (src_reg->flags & IR3_REG_CONST) {
+ /* an instruction cannot reference two different
+ * address registers:
+ */
+ if ((src_reg->flags & IR3_REG_RELATIV) &&
+ conflicts(instr->address, reg->instr->address))
+ return;
-static void instr_find_neighbors(struct ir3_instruction *instr)
-{
- unsigned i;
+ /* This seems to be a hw bug, or something where the timings
+ * just somehow don't work out. This restriction may only
+ * apply if the first src is also CONST.
+ */
+ if ((opc_cat(instr->opc) == 3) && (n == 2) &&
+ (src_reg->flags & IR3_REG_RELATIV) &&
+ (src_reg->array.offset == 0))
+ return;
- if (check_stop(instr))
- return;
+ src_reg = ir3_reg_clone(instr->block->shader, src_reg);
+ src_reg->flags = new_flags;
+ instr->regs[n+1] = src_reg;
- if (is_meta(instr) && (instr->opc == OPC_META_FI)) {
- unsigned n = instr->regs_count;
- for (i = 1; i < n; i++) {
- struct ir3_instruction *src_instr = ssa(instr->regs[i]);
- if (src_instr) {
- struct ir3_instruction *left = (i > 1) ?
- ssa(instr->regs[i-1]) : NULL;
- struct ir3_instruction *right = (i < (n - 1)) ?
- ssa(instr->regs[i+1]) : NULL;
- set_neighbors(src_instr, left, right);
- instr_find_neighbors(src_instr);
- }
+ if (src_reg->flags & IR3_REG_RELATIV)
+ ir3_instr_set_address(instr, reg->instr->address);
+
+ return;
}
- } else {
- for (i = 1; i < instr->regs_count; i++) {
- struct ir3_instruction *src_instr = ssa(instr->regs[i]);
- if (src_instr)
- instr_find_neighbors(src_instr);
+
+ if ((src_reg->flags & IR3_REG_RELATIV) &&
+ !conflicts(instr->address, reg->instr->address)) {
+ src_reg = ir3_reg_clone(instr->block->shader, src_reg);
+ src_reg->flags = new_flags;
+ instr->regs[n+1] = src_reg;
+ ir3_instr_set_address(instr, reg->instr->address);
+
+ return;
+ }
+
+ /* NOTE: seems we can only do immed integers, so don't
+ * need to care about float. But we do need to handle
+ * abs/neg *before* checking that the immediate requires
+ * few enough bits to encode:
+ *
+ * TODO: do we need to do something to avoid accidentally
+ * catching a float immed?
+ */
+ if (src_reg->flags & IR3_REG_IMMED) {
+ int32_t iim_val = src_reg->iim_val;
+
+ debug_assert((opc_cat(instr->opc) == 1) ||
+ (opc_cat(instr->opc) == 6) ||
+ ir3_cat2_int(instr->opc));
+
+ if (new_flags & IR3_REG_SABS)
+ iim_val = abs(iim_val);
+
+ if (new_flags & IR3_REG_SNEG)
+ iim_val = -iim_val;
+
+ if (new_flags & IR3_REG_BNOT)
+ iim_val = ~iim_val;
+
+ /* other than category 1 (mov) we can only encode up to 10 bits: */
+ if ((instr->opc == OPC_MOV) ||
+ !((iim_val & ~0x3ff) && (-iim_val & ~0x3ff))) {
+ new_flags &= ~(IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT);
+ src_reg = ir3_reg_clone(instr->block->shader, src_reg);
+ 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)) {
+ /* See if lowering an immediate to const would help. */
+ instr->regs[n+1] = lower_immed(ctx, src_reg, new_flags);
+ }
+
+ return;
}
}
}
-static void block_find_neighbors(struct ir3_block *block)
+/* Handle special case of eliminating output mov, and similar cases where
+ * there isn't a normal "consuming" instruction. In this case we cannot
+ * collapse flags (ie. output mov from const, or w/ abs/neg flags, cannot
+ * be eliminated)
+ */
+static struct ir3_instruction *
+eliminate_output_mov(struct ir3_instruction *instr)
{
- unsigned i;
-
- for (i = 0; i < block->noutputs; i++) {
- if (block->outputs[i]) {
- struct ir3_instruction *instr = block->outputs[i];
- instr_find_neighbors(instr);
+ if (is_eligible_mov(instr, 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);
+ return src_instr;
}
}
+ return instr;
}
-static void instr_clear_neighbors(struct ir3_instruction *instr)
+/**
+ * Find instruction src's which are mov's that can be collapsed, replacing
+ * the mov dst with the mov src
+ */
+static void
+instr_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
{
- unsigned i;
+ struct ir3_register *reg;
+
+ if (instr->regs_count == 0)
+ return;
- if (check_stop(instr))
+ if (ir3_instr_check_mark(instr))
return;
- instr->cp.left_cnt = 0;
- instr->cp.left = NULL;
- instr->cp.right_cnt = 0;
- instr->cp.right = NULL;
+ /* walk down the graph from each src: */
+ foreach_src_n(reg, n, instr) {
+ struct ir3_instruction *src = ssa(reg);
+
+ if (!src)
+ continue;
+
+ instr_cp(ctx, src);
- for (i = 1; i < instr->regs_count; i++) {
- struct ir3_instruction *src_instr = ssa(instr->regs[i]);
- if (src_instr)
- instr_clear_neighbors(src_instr);
+ /* TODO non-indirect access we could figure out which register
+ * we actually want and allow cp..
+ */
+ if (reg->flags & IR3_REG_ARRAY)
+ continue;
+
+ reg_cp(ctx, instr, reg, n);
}
-}
-static void block_clear_neighbors(struct ir3_block *block)
-{
- unsigned i;
+ if (instr->regs[0]->flags & IR3_REG_ARRAY) {
+ struct ir3_instruction *src = ssa(instr->regs[0]);
+ if (src)
+ instr_cp(ctx, src);
+ }
- for (i = 0; i < block->noutputs; i++) {
- if (block->outputs[i]) {
- struct ir3_instruction *instr = block->outputs[i];
- instr_clear_neighbors(instr);
+ if (instr->address) {
+ instr_cp(ctx, instr->address);
+ ir3_instr_set_address(instr, eliminate_output_mov(instr->address));
+ }
+
+ /* we can end up with extra cmps.s from frontend, which uses a
+ *
+ * cmps.s p0.x, cond, 0
+ *
+ * as a way to mov into the predicate register. But frequently 'cond'
+ * is itself a cmps.s/cmps.f/cmps.u. So detect this special case and
+ * just re-write the instruction writing predicate register to get rid
+ * of the double cmps.
+ */
+ if ((instr->opc == OPC_CMPS_S) &&
+ (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)) {
+ struct ir3_instruction *cond = ssa(instr->regs[1]);
+ switch (cond->opc) {
+ case OPC_CMPS_S:
+ case OPC_CMPS_F:
+ case OPC_CMPS_U:
+ instr->opc = cond->opc;
+ instr->flags = cond->flags;
+ instr->cat2 = cond->cat2;
+ instr->address = cond->address;
+ instr->regs[1] = cond->regs[1];
+ instr->regs[2] = cond->regs[2];
+ break;
+ default:
+ break;
}
}
}
-void ir3_block_cp(struct ir3_block *block)
+void
+ir3_cp(struct ir3 *ir, struct ir3_shader_variant *so)
{
- ir3_clear_mark(block->shader);
- block_clear_neighbors(block);
- ir3_clear_mark(block->shader);
- block_find_neighbors(block);
- ir3_clear_mark(block->shader);
- block_cp(block);
+ struct ir3_cp_ctx ctx = {
+ .shader = ir,
+ .so = so,
+ };
+
+ 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]);
+ }
+ }
+
+ list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
+ if (block->condition) {
+ instr_cp(&ctx, block->condition);
+ block->condition = eliminate_output_mov(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]);
+ }
+ }
}