1f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org/*
2f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Copyright 2009 Nicolai Hähnle <nhaehnle@gmail.com>
3f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org *
4f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Permission is hereby granted, free of charge, to any person obtaining a
5f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * copy of this software and associated documentation files (the "Software"),
6f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * to deal in the Software without restriction, including without limitation
7f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * on the rights to use, copy, modify, merge, publish, distribute, sub
8f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * license, and/or sell copies of the Software, and to permit persons to whom
9f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * the Software is furnished to do so, subject to the following conditions:
10f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org *
11f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * The above copyright notice and this permission notice (including the next
12f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * paragraph) shall be included in all copies or substantial portions of the
13f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Software.
14f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org *
15f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * USE OR OTHER DEALINGS IN THE SOFTWARE. */
22f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
23f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "radeon_compiler.h"
24f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
25f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include <stdio.h>
26f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
27f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "radeon_compiler_util.h"
28f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "radeon_dataflow.h"
29f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "radeon_emulate_branches.h"
30f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "radeon_emulate_loops.h"
31f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "radeon_program_alu.h"
32f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "radeon_program_tex.h"
33f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "radeon_rename_regs.h"
34f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "radeon_remove_constants.h"
35f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "r300_fragprog.h"
36f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "r300_fragprog_swizzle.h"
37f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "r500_fragprog.h"
38f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
39f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
40f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic void dataflow_outputs_mark_use(void * userdata, void * data,
41f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		void (*callback)(void *, unsigned int, unsigned int))
42f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{
43f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	struct r300_fragment_program_compiler * c = userdata;
44f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	callback(data, c->OutputColor[0], RC_MASK_XYZW);
45f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	callback(data, c->OutputColor[1], RC_MASK_XYZW);
46f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	callback(data, c->OutputColor[2], RC_MASK_XYZW);
47f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	callback(data, c->OutputColor[3], RC_MASK_XYZW);
48f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	callback(data, c->OutputDepth, RC_MASK_W);
49f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}
50f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
51f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic void rc_rewrite_depth_out(struct radeon_compiler *cc, void *user)
52f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{
53f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	struct r300_fragment_program_compiler *c = (struct r300_fragment_program_compiler*)cc;
54f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	struct rc_instruction *rci;
55f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
56f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	for (rci = c->Base.Program.Instructions.Next; rci != &c->Base.Program.Instructions; rci = rci->Next) {
57f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		struct rc_sub_instruction * inst = &rci->U.I;
58f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		unsigned i;
59f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		const struct rc_opcode_info *info = rc_get_opcode_info(inst->Opcode);
60f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
61f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		if (inst->DstReg.File != RC_FILE_OUTPUT || inst->DstReg.Index != c->OutputDepth)
62f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org			continue;
63f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
64f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		if (inst->DstReg.WriteMask & RC_MASK_Z) {
65f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org			inst->DstReg.WriteMask = RC_MASK_W;
66f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		} else {
67f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org			inst->DstReg.WriteMask = 0;
68f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org			continue;
69f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		}
70f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
71f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		if (!info->IsComponentwise) {
72f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org			continue;
73f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		}
74f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
75f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		for (i = 0; i < info->NumSrcRegs; i++) {
76f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org			inst->SrcReg[i] = lmul_swizzle(RC_SWIZZLE_ZZZZ, inst->SrcReg[i]);
77f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		}
78f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	}
79f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}
80f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
81f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgvoid r3xx_compile_fragment_program(struct r300_fragment_program_compiler* c)
82f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{
83f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	int is_r500 = c->Base.is_r500;
84f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	int opt = !c->Base.disable_optimizations;
85f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
86f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	/* Lists of instruction transformations. */
87f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	struct radeon_program_transformation rewrite_tex[] = {
88f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ &radeonTransformTEX, c },
89f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ 0, 0 }
90f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	};
91f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
92f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	struct radeon_program_transformation rewrite_if[] = {
93f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ &r500_transform_IF, 0 },
94f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{0, 0}
95f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	};
96f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
97f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	struct radeon_program_transformation native_rewrite_r500[] = {
98f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ &radeonTransformALU, 0 },
99f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ &radeonTransformDeriv, 0 },
100f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ &radeonTransformTrigScale, 0 },
101f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ 0, 0 }
102f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	};
103f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
104f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	struct radeon_program_transformation native_rewrite_r300[] = {
105f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ &radeonTransformALU, 0 },
106f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ &r300_transform_trig_simple, 0 },
107f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{ 0, 0 }
108f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	};
109f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
110f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	/* List of compiler passes. */
111f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	struct radeon_compiler_pass fs_list[] = {
112f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		/* NAME				DUMP PREDICATE	FUNCTION			PARAM */
113f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"rewrite depth out",		1, 1,		rc_rewrite_depth_out,		NULL},
114f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		/* This transformation needs to be done before any of the IF
115f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		 * instructions are modified. */
116f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"transform KILP",		1, 1,		rc_transform_KILP,		NULL},
117f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"unroll loops",		1, is_r500,	rc_unroll_loops,		NULL},
118f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"transform loops",		1, !is_r500,	rc_transform_loops,		NULL},
119f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"emulate branches",		1, !is_r500,	rc_emulate_branches,		NULL},
120f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"transform TEX",		1, 1,		rc_local_transform,		rewrite_tex},
121f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"transform IF",		1, is_r500,	rc_local_transform,		rewrite_if},
122f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"native rewrite",		1, is_r500,	rc_local_transform,		native_rewrite_r500},
123f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"native rewrite",		1, !is_r500,	rc_local_transform,		native_rewrite_r300},
124f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"deadcode",			1, opt,		rc_dataflow_deadcode,		dataflow_outputs_mark_use},
125f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"emulate loops",		1, !is_r500,	rc_emulate_loops,		NULL},
126f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"register rename",		1, !is_r500 || opt,		rc_rename_regs,			NULL},
127f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"dataflow optimize",		1, opt,		rc_optimize,			NULL},
128f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"inline literals",		1, is_r500 && opt,		rc_inline_literals,			NULL},
129f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"dataflow swizzles",		1, 1,		rc_dataflow_swizzles,		NULL},
130f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"dead constants",		1, 1,		rc_remove_unused_constants,	&c->code->constants_remap_table},
131f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"pair translate",		1, 1,		rc_pair_translate,		NULL},
132f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"pair scheduling",		1, 1,		rc_pair_schedule,		&opt},
133f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"dead sources",		1, 1,		rc_pair_remove_dead_sources, NULL},
134f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"register allocation",		1, 1,		rc_pair_regalloc,		&opt},
135f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"final code validation",	0, 1,		rc_validate_final_shader,	NULL},
136f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"machine code generation",	0, is_r500,	r500BuildFragmentProgramHwCode,	NULL},
137f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"machine code generation",	0, !is_r500,	r300BuildFragmentProgramHwCode,	NULL},
138f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"dump machine code",		0, is_r500  && (c->Base.Debug & RC_DBG_LOG), r500FragmentProgramDump, NULL},
139f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{"dump machine code",		0, !is_r500 && (c->Base.Debug & RC_DBG_LOG), r300FragmentProgramDump, NULL},
140f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org		{NULL, 0, 0, NULL, NULL}
141f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	};
142f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
143f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	c->Base.type = RC_FRAGMENT_PROGRAM;
144f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	c->Base.SwizzleCaps = c->Base.is_r500 ? &r500_swizzle_caps : &r300_swizzle_caps;
145f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
146f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	rc_run_compiler(&c->Base, fs_list);
147f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org
148f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org	rc_constants_copy(&c->code->constants, &c->Base.Program.Constants);
149f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}
150