namespace brw {
+bool
+vec4_instruction::is_tex()
+{
+ return (opcode == SHADER_OPCODE_TEX ||
+ opcode == SHADER_OPCODE_TXD ||
+ opcode == SHADER_OPCODE_TXF ||
+ opcode == SHADER_OPCODE_TXL ||
+ opcode == SHADER_OPCODE_TXS);
+}
+
bool
vec4_instruction::is_math()
{
opcode == SHADER_OPCODE_LOG2 ||
opcode == SHADER_OPCODE_SIN ||
opcode == SHADER_OPCODE_COS ||
+ opcode == SHADER_OPCODE_INT_QUOTIENT ||
+ opcode == SHADER_OPCODE_INT_REMAINDER ||
opcode == SHADER_OPCODE_POW);
}
+/**
+ * Returns how many MRFs an opcode will write over.
+ *
+ * Note that this is not the 0 or 1 implied writes in an actual gen
+ * instruction -- the generate_* functions generate additional MOVs
+ * for setup.
+ */
+int
+vec4_visitor::implied_mrf_writes(vec4_instruction *inst)
+{
+ if (inst->mlen == 0)
+ return 0;
+
+ switch (inst->opcode) {
+ case SHADER_OPCODE_RCP:
+ case SHADER_OPCODE_RSQ:
+ case SHADER_OPCODE_SQRT:
+ case SHADER_OPCODE_EXP2:
+ case SHADER_OPCODE_LOG2:
+ case SHADER_OPCODE_SIN:
+ case SHADER_OPCODE_COS:
+ return 1;
+ case SHADER_OPCODE_POW:
+ return 2;
+ case VS_OPCODE_URB_WRITE:
+ return 1;
+ case VS_OPCODE_PULL_CONSTANT_LOAD:
+ return 2;
+ case VS_OPCODE_SCRATCH_READ:
+ return 2;
+ case VS_OPCODE_SCRATCH_WRITE:
+ return 3;
+ default:
+ assert(!"not reached");
+ return inst->mlen;
+ }
+}
bool
src_reg::equals(src_reg *r)
}
}
+bool
+src_reg::is_zero() const
+{
+ if (file != IMM)
+ return false;
+
+ if (type == BRW_REGISTER_TYPE_F) {
+ return imm.f == 0.0;
+ } else {
+ return imm.i == 0;
+ }
+}
+
+bool
+src_reg::is_one() const
+{
+ if (file != IMM)
+ return false;
+
+ if (type == BRW_REGISTER_TYPE_F) {
+ return imm.f == 1.0;
+ } else {
+ return imm.i == 1;
+ }
+}
+
+/**
+ * Does algebraic optimizations (0 * a = 0, 1 * a = a, a + 0 = a).
+ *
+ * While GLSL IR also performs this optimization, we end up with it in
+ * our instruction stream for a couple of reasons. One is that we
+ * sometimes generate silly instructions, for example in array access
+ * where we'll generate "ADD offset, index, base" even if base is 0.
+ * The other is that GLSL IR's constant propagation doesn't track the
+ * components of aggregates, so some VS patterns (initialize matrix to
+ * 0, accumulate in vertex blending factors) end up breaking down to
+ * instructions involving 0.
+ */
+bool
+vec4_visitor::opt_algebraic()
+{
+ bool progress = false;
+
+ foreach_list(node, &this->instructions) {
+ vec4_instruction *inst = (vec4_instruction *)node;
+
+ switch (inst->opcode) {
+ case BRW_OPCODE_ADD:
+ if (inst->src[1].is_zero()) {
+ inst->opcode = BRW_OPCODE_MOV;
+ inst->src[1] = src_reg();
+ progress = true;
+ }
+ break;
+
+ case BRW_OPCODE_MUL:
+ if (inst->src[1].is_zero()) {
+ inst->opcode = BRW_OPCODE_MOV;
+ switch (inst->src[0].type) {
+ case BRW_REGISTER_TYPE_F:
+ inst->src[0] = src_reg(0.0f);
+ break;
+ case BRW_REGISTER_TYPE_D:
+ inst->src[0] = src_reg(0);
+ break;
+ case BRW_REGISTER_TYPE_UD:
+ inst->src[0] = src_reg(0u);
+ break;
+ default:
+ assert(!"not reached");
+ inst->src[0] = src_reg(0.0f);
+ break;
+ }
+ inst->src[1] = src_reg();
+ progress = true;
+ } else if (inst->src[1].is_one()) {
+ inst->opcode = BRW_OPCODE_MOV;
+ inst->src[1] = src_reg();
+ progress = true;
+ }
+ break;
+ default:
+ break;
+ }
+ }
+
+ if (progress)
+ this->live_intervals_valid = false;
+
+ return progress;
+}
+
/**
* Only a limited number of hardware registers may be used for push
* constants, so this turns access to the overflowed constants into
pack_uniform_registers();
}
+/*
+ * Tries to reduce extra MOV instructions by taking GRFs that get just
+ * written and then MOVed into an MRF and making the original write of
+ * the GRF write directly to the MRF instead.
+ */
+bool
+vec4_visitor::opt_compute_to_mrf()
+{
+ bool progress = false;
+ int next_ip = 0;
+
+ calculate_live_intervals();
+
+ foreach_list_safe(node, &this->instructions) {
+ vec4_instruction *inst = (vec4_instruction *)node;
+
+ int ip = next_ip;
+ next_ip++;
+
+ if (inst->opcode != BRW_OPCODE_MOV ||
+ inst->predicate ||
+ inst->dst.file != MRF || inst->src[0].file != GRF ||
+ inst->dst.type != inst->src[0].type ||
+ inst->src[0].abs || inst->src[0].negate || inst->src[0].reladdr)
+ continue;
+
+ int mrf = inst->dst.reg;
+
+ /* Can't compute-to-MRF this GRF if someone else was going to
+ * read it later.
+ */
+ if (this->virtual_grf_use[inst->src[0].reg] > ip)
+ continue;
+
+ /* We need to check interference with the MRF between this
+ * instruction and the earliest instruction involved in writing
+ * the GRF we're eliminating. To do that, keep track of which
+ * of our source channels we've seen initialized.
+ */
+ bool chans_needed[4] = {false, false, false, false};
+ int chans_remaining = 0;
+ for (int i = 0; i < 4; i++) {
+ int chan = BRW_GET_SWZ(inst->src[0].swizzle, i);
+
+ if (!(inst->dst.writemask & (1 << i)))
+ continue;
+
+ /* We don't handle compute-to-MRF across a swizzle. We would
+ * need to be able to rewrite instructions above to output
+ * results to different channels.
+ */
+ if (chan != i)
+ chans_remaining = 5;
+
+ if (!chans_needed[chan]) {
+ chans_needed[chan] = true;
+ chans_remaining++;
+ }
+ }
+ if (chans_remaining > 4)
+ continue;
+
+ /* Now walk up the instruction stream trying to see if we can
+ * rewrite everything writing to the GRF into the MRF instead.
+ */
+ vec4_instruction *scan_inst;
+ for (scan_inst = (vec4_instruction *)inst->prev;
+ scan_inst->prev != NULL;
+ scan_inst = (vec4_instruction *)scan_inst->prev) {
+ if (scan_inst->dst.file == GRF &&
+ scan_inst->dst.reg == inst->src[0].reg &&
+ scan_inst->dst.reg_offset == inst->src[0].reg_offset) {
+ /* Found something writing to the reg we want to turn into
+ * a compute-to-MRF.
+ */
+
+ /* SEND instructions can't have MRF as a destination. */
+ if (scan_inst->mlen)
+ break;
+
+ if (intel->gen >= 6) {
+ /* gen6 math instructions must have the destination be
+ * GRF, so no compute-to-MRF for them.
+ */
+ if (scan_inst->is_math()) {
+ break;
+ }
+ }
+
+ /* Mark which channels we found unconditional writes for. */
+ if (!scan_inst->predicate) {
+ for (int i = 0; i < 4; i++) {
+ if (scan_inst->dst.writemask & (1 << i) &&
+ chans_needed[i]) {
+ chans_needed[i] = false;
+ chans_remaining--;
+ }
+ }
+ }
+
+ if (chans_remaining == 0)
+ break;
+ }
+
+ /* We don't handle flow control here. Most computation of
+ * values that end up in MRFs are shortly before the MRF
+ * write anyway.
+ */
+ if (scan_inst->opcode == BRW_OPCODE_DO ||
+ scan_inst->opcode == BRW_OPCODE_WHILE ||
+ scan_inst->opcode == BRW_OPCODE_ELSE ||
+ scan_inst->opcode == BRW_OPCODE_ENDIF) {
+ break;
+ }
+
+ /* You can't read from an MRF, so if someone else reads our
+ * MRF's source GRF that we wanted to rewrite, that stops us.
+ */
+ bool interfered = false;
+ for (int i = 0; i < 3; i++) {
+ if (scan_inst->src[i].file == GRF &&
+ scan_inst->src[i].reg == inst->src[0].reg &&
+ scan_inst->src[i].reg_offset == inst->src[0].reg_offset) {
+ interfered = true;
+ }
+ }
+ if (interfered)
+ break;
+
+ /* If somebody else writes our MRF here, we can't
+ * compute-to-MRF before that.
+ */
+ if (scan_inst->dst.file == MRF && mrf == scan_inst->dst.reg)
+ break;
+
+ if (scan_inst->mlen > 0) {
+ /* Found a SEND instruction, which means that there are
+ * live values in MRFs from base_mrf to base_mrf +
+ * scan_inst->mlen - 1. Don't go pushing our MRF write up
+ * above it.
+ */
+ if (mrf >= scan_inst->base_mrf &&
+ mrf < scan_inst->base_mrf + scan_inst->mlen) {
+ break;
+ }
+ }
+ }
+
+ if (chans_remaining == 0) {
+ /* If we've made it here, we have an inst we want to
+ * compute-to-MRF, and a scan_inst pointing to the earliest
+ * instruction involved in computing the value. Now go
+ * rewrite the instruction stream between the two.
+ */
+
+ while (scan_inst != inst) {
+ if (scan_inst->dst.file == GRF &&
+ scan_inst->dst.reg == inst->src[0].reg &&
+ scan_inst->dst.reg_offset == inst->src[0].reg_offset) {
+ scan_inst->dst.file = MRF;
+ scan_inst->dst.reg = mrf;
+ scan_inst->dst.reg_offset = 0;
+ scan_inst->saturate |= inst->saturate;
+ }
+ scan_inst = (vec4_instruction *)scan_inst->next;
+ }
+ inst->remove();
+ progress = true;
+ }
+ }
+
+ if (progress)
+ live_intervals_valid = false;
+
+ return progress;
+}
+
} /* namespace brw */