ir3/nir: Add a new pass 'ir3_nir_lower_io_offsets'

This NIR->NIR pass implements offset computations that are currently
done on the IR3 backend compiler, to give NIR a better chance of
optimizing them.

For now, it supports lowering the dword-offset computation for SSBO
instructions. It will take an SSBO intrinsic and replace it with the
new ir3-specific version that adds an extra source. That source will
hold the SSA value resulting from inserting a division by 4 (an SHR op)
of the original byte-offset source already provided by NIR in one of
the intrinsic sources.

Note that on a6xx the original byte-offset is not needed, so we could
potentially replace that source instead of adding a new one. But to
keep things simple and consistent we always add the new source and
a6xx will just ignore the original one.

Reviewed-by: Rob Clark <[email protected]>

4 files changed, 217 insertions, 0 deletions

diff --git a/src/freedreno/Makefile.sources b/src/freedreno/Makefile.sources
index d4c2103708d..91fb74a8564 100644
--- a/src/freedreno/Makefile.sources
+++ b/src/freedreno/Makefile.sources
@@ -35,6 +35,7 @@ ir3_SOURCES := \
 	ir3/ir3_legalize.c \
 	ir3/ir3_nir.c \
 	ir3/ir3_nir.h \
+	ir3/ir3_nir_lower_io_offsets.c \
 	ir3/ir3_nir_lower_tg4_to_tex.c \
 	ir3/ir3_print.c \
 	ir3/ir3_ra.c \
diff --git a/src/freedreno/ir3/ir3_nir.h b/src/freedreno/ir3/ir3_nir.h
index 74201d34160..7983b74af2c 100644
--- a/src/freedreno/ir3/ir3_nir.h
+++ b/src/freedreno/ir3/ir3_nir.h
@@ -36,6 +36,7 @@ void ir3_nir_scan_driver_consts(nir_shader *shader, struct ir3_driver_const_layo
 
 bool ir3_nir_apply_trig_workarounds(nir_shader *shader);
 bool ir3_nir_lower_tg4_to_tex(nir_shader *shader);
+bool ir3_nir_lower_io_offsets(nir_shader *shader);
 
 const nir_shader_compiler_options * ir3_get_compiler_options(struct ir3_compiler *compiler);
 bool ir3_key_lowers_nir(const struct ir3_shader_key *key);
diff --git a/src/freedreno/ir3/ir3_nir_lower_io_offsets.c b/src/freedreno/ir3/ir3_nir_lower_io_offsets.c
new file mode 100644
index 00000000000..264f1064545
--- /dev/null
+++ b/src/freedreno/ir3/ir3_nir_lower_io_offsets.c
@@ -0,0 +1,214 @@
+/*
+ * Copyright © 2018-2019 Igalia S.L.
+ *
+ * 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
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * 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 NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS 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.
+ */
+
+#include "ir3_nir.h"
+#include "compiler/nir/nir_builder.h"
+
+/**
+ * This pass moves to NIR certain offset computations for different I/O
+ * ops that are currently implemented on the IR3 backend compiler, to
+ * give NIR a chance to optimize them:
+ *
+ * - Dword-offset for SSBO load, store and atomics: A new, similar intrinsic
+ *   is emitted that replaces the original one, adding a new source that
+ *   holds the result of the original byte-offset source divided by 4.
+ */
+
+
+/* Returns the ir3-specific intrinsic opcode corresponding to an SSBO
+ * instruction that is handled by this pass. It also conveniently returns
+ * the offset source index in @offset_src_idx.
+ *
+ * If @intrinsic is not SSBO, or it is not handled by the pass, -1 is
+ * returned.
+ */
+static int
+get_ir3_intrinsic_for_ssbo_intrinsic(unsigned intrinsic,
+									 uint8_t *offset_src_idx)
+{
+	debug_assert(offset_src_idx);
+
+	*offset_src_idx = 1;
+
+	switch (intrinsic) {
+	case nir_intrinsic_store_ssbo:
+		*offset_src_idx = 2;
+		return nir_intrinsic_store_ssbo_ir3;
+	case nir_intrinsic_load_ssbo:
+		return nir_intrinsic_load_ssbo_ir3;
+	case nir_intrinsic_ssbo_atomic_add:
+		return nir_intrinsic_ssbo_atomic_add_ir3;
+	case nir_intrinsic_ssbo_atomic_imin:
+		return nir_intrinsic_ssbo_atomic_imin_ir3;
+	case nir_intrinsic_ssbo_atomic_umin:
+		return nir_intrinsic_ssbo_atomic_umin_ir3;
+	case nir_intrinsic_ssbo_atomic_imax:
+		return nir_intrinsic_ssbo_atomic_imax_ir3;
+	case nir_intrinsic_ssbo_atomic_umax:
+		return nir_intrinsic_ssbo_atomic_umax_ir3;
+	case nir_intrinsic_ssbo_atomic_and:
+		return nir_intrinsic_ssbo_atomic_and_ir3;
+	case nir_intrinsic_ssbo_atomic_or:
+		return nir_intrinsic_ssbo_atomic_or_ir3;
+	case nir_intrinsic_ssbo_atomic_xor:
+		return nir_intrinsic_ssbo_atomic_xor_ir3;
+	case nir_intrinsic_ssbo_atomic_exchange:
+		return nir_intrinsic_ssbo_atomic_exchange_ir3;
+	case nir_intrinsic_ssbo_atomic_comp_swap:
+		return nir_intrinsic_ssbo_atomic_comp_swap_ir3;
+	default:
+		break;
+	}
+
+	return -1;
+}
+
+static bool
+lower_offset_for_ssbo(nir_intrinsic_instr *intrinsic, nir_builder *b,
+					  unsigned ir3_ssbo_opcode, uint8_t offset_src_idx)
+{
+	unsigned num_srcs = nir_intrinsic_infos[intrinsic->intrinsic].num_srcs;
+
+	bool has_dest = nir_intrinsic_infos[intrinsic->intrinsic].has_dest;
+	nir_ssa_def *new_dest = NULL;
+
+	/* Here we create a new intrinsic and copy over all contents from the old one. */
+
+	nir_intrinsic_instr *new_intrinsic;
+	nir_src *target_src;
+
+	/* 'offset_src_idx' holds the index of the source that represent the offset. */
+	new_intrinsic =
+		nir_intrinsic_instr_create(b->shader, ir3_ssbo_opcode);
+
+	debug_assert(intrinsic->src[offset_src_idx].is_ssa);
+	nir_ssa_def *offset = intrinsic->src[offset_src_idx].ssa;
+
+	/* The new source that will hold the dword-offset is always the last
+	 * one for every intrinsic.
+	 */
+	target_src = &new_intrinsic->src[num_srcs];
+	*target_src = nir_src_for_ssa(offset);
+
+	if (has_dest) {
+		debug_assert(intrinsic->dest.is_ssa);
+		nir_ssa_def *dest = &intrinsic->dest.ssa;
+		nir_ssa_dest_init(&new_intrinsic->instr, &new_intrinsic->dest,
+						  dest->num_components, dest->bit_size, NULL);
+		new_dest = &new_intrinsic->dest.ssa;
+	}
+
+	for (unsigned i = 0; i < num_srcs; i++)
+		new_intrinsic->src[i] = nir_src_for_ssa(intrinsic->src[i].ssa);
+
+	for (unsigned i = 0; i < NIR_INTRINSIC_MAX_CONST_INDEX; i++)
+		new_intrinsic->const_index[i] = intrinsic->const_index[i];
+
+	new_intrinsic->num_components = intrinsic->num_components;
+
+	b->cursor = nir_before_instr(&intrinsic->instr);
+	nir_ssa_def *offset_div_4 = nir_ushr(b, offset, nir_imm_int(b, 2));
+	debug_assert(offset_div_4);
+
+	/* Insert the new intrinsic right before the old one. */
+	b->cursor = nir_before_instr(&intrinsic->instr);
+	nir_builder_instr_insert(b, &new_intrinsic->instr);
+
+	/* Replace the last source of the new intrinsic by the result of
+	 * the offset divided by 4.
+	 */
+	nir_instr_rewrite_src(&new_intrinsic->instr,
+						  target_src,
+						  nir_src_for_ssa(offset_div_4));
+
+	if (has_dest) {
+		/* Replace the uses of the original destination by that
+		 * of the new intrinsic.
+		 */
+		nir_ssa_def_rewrite_uses(&intrinsic->dest.ssa,
+								 nir_src_for_ssa(new_dest));
+	}
+
+	/* Finally remove the original intrinsic. */
+	nir_instr_remove(&intrinsic->instr);
+
+	return true;
+}
+
+static bool
+lower_io_offsets_block(nir_block *block, nir_builder *b, void *mem_ctx)
+{
+	bool progress = false;
+
+	nir_foreach_instr_safe(instr, block) {
+		if (instr->type != nir_instr_type_intrinsic)
+			continue;
+
+		nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+
+		/* SSBO */
+		int ir3_intrinsic;
+		uint8_t offset_src_idx;
+		ir3_intrinsic = get_ir3_intrinsic_for_ssbo_intrinsic(intr->intrinsic,
+															 &offset_src_idx);
+		if (ir3_intrinsic != -1) {
+			progress |= lower_offset_for_ssbo(intr, b, (unsigned) ir3_intrinsic,
+											  offset_src_idx);
+		}
+	}
+
+	return progress;
+}
+
+static bool
+lower_io_offsets_func(nir_function_impl *impl)
+{
+	void *mem_ctx = ralloc_parent(impl);
+	nir_builder b;
+	nir_builder_init(&b, impl);
+
+	bool progress = false;
+	nir_foreach_block_safe(block, impl) {
+		progress |= lower_io_offsets_block(block, &b, mem_ctx);
+	}
+
+	if (progress) {
+		nir_metadata_preserve(impl, nir_metadata_block_index |
+									nir_metadata_dominance);
+	}
+
+	return progress;
+}
+
+bool
+ir3_nir_lower_io_offsets(nir_shader *shader)
+{
+	bool progress = false;
+
+	nir_foreach_function(function, shader) {
+		if (function->impl)
+			progress |= lower_io_offsets_func(function->impl);
+	}
+
+	return progress;
+}
diff --git a/src/freedreno/ir3/meson.build b/src/freedreno/ir3/meson.build
index 5f87ca6185b..d210211cc97 100644
--- a/src/freedreno/ir3/meson.build
+++ b/src/freedreno/ir3/meson.build
@@ -50,6 +50,7 @@ libfreedreno_ir3_files = files(
   'ir3_legalize.c',
   'ir3_nir.c',
   'ir3_nir.h',
+  'ir3_nir_lower_io_offsets.c',
   'ir3_nir_lower_tg4_to_tex.c',
   'ir3_print.c',
   'ir3_ra.c',