1
2/*
3 * Copyright 2012 Advanced Micro Devices, Inc.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 *
24 * Authors:
25 *	Tom Stellard <thomas.stellard@amd.com>
26 *	Michel Dänzer <michel.daenzer@amd.com>
27 *      Christian König <christian.koenig@amd.com>
28 */
29
30#include "gallivm/lp_bld_tgsi_action.h"
31#include "gallivm/lp_bld_const.h"
32#include "gallivm/lp_bld_gather.h"
33#include "gallivm/lp_bld_intr.h"
34#include "gallivm/lp_bld_tgsi.h"
35#include "radeon_llvm.h"
36#include "radeon_llvm_emit.h"
37#include "tgsi/tgsi_info.h"
38#include "tgsi/tgsi_parse.h"
39#include "tgsi/tgsi_scan.h"
40#include "tgsi/tgsi_dump.h"
41
42#include "radeonsi_pipe.h"
43#include "radeonsi_shader.h"
44#include "si_state.h"
45#include "sid.h"
46
47#include <assert.h>
48#include <errno.h>
49#include <stdio.h>
50
51/*
52static ps_remap_inputs(
53	struct tgsi_llvm_context * tl_ctx,
54	unsigned tgsi_index,
55	unsigned tgsi_chan)
56{
57	:
58}
59
60struct si_input
61{
62	struct list_head head;
63	unsigned tgsi_index;
64	unsigned tgsi_chan;
65	unsigned order;
66};
67*/
68
69
70struct si_shader_context
71{
72	struct radeon_llvm_context radeon_bld;
73	struct r600_context *rctx;
74	struct tgsi_parse_context parse;
75	struct tgsi_token * tokens;
76	struct si_pipe_shader *shader;
77	unsigned type; /* TGSI_PROCESSOR_* specifies the type of shader. */
78/*	unsigned num_inputs; */
79/*	struct list_head inputs; */
80/*	unsigned * input_mappings *//* From TGSI to SI hw */
81/*	struct tgsi_shader_info info;*/
82};
83
84static struct si_shader_context * si_shader_context(
85	struct lp_build_tgsi_context * bld_base)
86{
87	return (struct si_shader_context *)bld_base;
88}
89
90
91#define PERSPECTIVE_BASE 0
92#define LINEAR_BASE 9
93
94#define SAMPLE_OFFSET 0
95#define CENTER_OFFSET 2
96#define CENTROID_OFSET 4
97
98#define USE_SGPR_MAX_SUFFIX_LEN 5
99#define CONST_ADDR_SPACE 2
100#define USER_SGPR_ADDR_SPACE 8
101
102enum sgpr_type {
103	SGPR_CONST_PTR_F32,
104	SGPR_CONST_PTR_V4I32,
105	SGPR_CONST_PTR_V8I32,
106	SGPR_I32,
107	SGPR_I64
108};
109
110/**
111 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad
112 *
113 * @param offset The offset parameter specifies the number of
114 * elements to offset, not the number of bytes or dwords.  An element is the
115 * the type pointed to by the base_ptr parameter (e.g. int is the element of
116 * an int* pointer)
117 *
118 * When LLVM lowers the load instruction, it will convert the element offset
119 * into a dword offset automatically.
120 *
121 */
122static LLVMValueRef build_indexed_load(
123	struct gallivm_state * gallivm,
124	LLVMValueRef base_ptr,
125	LLVMValueRef offset)
126{
127	LLVMValueRef computed_ptr = LLVMBuildGEP(
128		gallivm->builder, base_ptr, &offset, 1, "");
129
130	return LLVMBuildLoad(gallivm->builder, computed_ptr, "");
131}
132
133/**
134 * Load a value stored in one of the user SGPRs
135 *
136 * @param sgpr This is the sgpr to load the value from.  If you need to load a
137 * value that is stored in consecutive SGPR registers (e.g. a 64-bit pointer),
138 * then you should pass the index of the first SGPR that holds the value.  For
139 * example, if you want to load a pointer that is stored in SGPRs 2 and 3, then
140 * use pass 2 for the sgpr parameter.
141 *
142 * The value of the sgpr parameter must also be aligned to the width of the type
143 * being loaded, so that the sgpr parameter is divisible by the dword width of the
144 * type.  For example, if the value being loaded is two dwords wide, then the sgpr
145 * parameter must be divisible by two.
146 */
147static LLVMValueRef use_sgpr(
148	struct gallivm_state * gallivm,
149	enum sgpr_type type,
150	unsigned sgpr)
151{
152	LLVMValueRef sgpr_index;
153	LLVMTypeRef ret_type;
154	LLVMValueRef ptr;
155
156	sgpr_index = lp_build_const_int32(gallivm, sgpr);
157
158	switch (type) {
159	case SGPR_CONST_PTR_F32:
160		assert(sgpr % 2 == 0);
161		ret_type = LLVMFloatTypeInContext(gallivm->context);
162		ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
163		break;
164
165	case SGPR_I32:
166		ret_type = LLVMInt32TypeInContext(gallivm->context);
167		break;
168
169	case SGPR_I64:
170		assert(sgpr % 2 == 0);
171		ret_type= LLVMInt64TypeInContext(gallivm->context);
172		break;
173
174	case SGPR_CONST_PTR_V4I32:
175		assert(sgpr % 2 == 0);
176		ret_type = LLVMInt32TypeInContext(gallivm->context);
177		ret_type = LLVMVectorType(ret_type, 4);
178		ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
179		break;
180
181	case SGPR_CONST_PTR_V8I32:
182		assert(sgpr % 2 == 0);
183		ret_type = LLVMInt32TypeInContext(gallivm->context);
184		ret_type = LLVMVectorType(ret_type, 8);
185		ret_type = LLVMPointerType(ret_type, CONST_ADDR_SPACE);
186		break;
187
188	default:
189		assert(!"Unsupported SGPR type in use_sgpr()");
190		return NULL;
191	}
192
193	ret_type = LLVMPointerType(ret_type, USER_SGPR_ADDR_SPACE);
194	ptr = LLVMBuildIntToPtr(gallivm->builder, sgpr_index, ret_type, "");
195	return LLVMBuildLoad(gallivm->builder, ptr, "");
196}
197
198static void declare_input_vs(
199	struct si_shader_context * si_shader_ctx,
200	unsigned input_index,
201	const struct tgsi_full_declaration *decl)
202{
203	LLVMValueRef t_list_ptr;
204	LLVMValueRef t_offset;
205	LLVMValueRef t_list;
206	LLVMValueRef attribute_offset;
207	LLVMValueRef buffer_index_reg;
208	LLVMValueRef args[3];
209	LLVMTypeRef vec4_type;
210	LLVMValueRef input;
211	struct lp_build_context * uint = &si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
212	struct lp_build_context * base = &si_shader_ctx->radeon_bld.soa.bld_base.base;
213	struct r600_context *rctx = si_shader_ctx->rctx;
214	//struct pipe_vertex_element *velem = &rctx->vertex_elements->elements[input_index];
215	unsigned chan;
216
217	/* Load the T list */
218	/* XXX: Communicate with the rest of the driver about which SGPR the T#
219	 * list pointer is going to be stored in.  Hard code to SGPR[6:7] for
220 	 * now */
221	t_list_ptr = use_sgpr(base->gallivm, SGPR_CONST_PTR_V4I32, 6);
222
223	t_offset = lp_build_const_int32(base->gallivm, input_index);
224
225	t_list = build_indexed_load(base->gallivm, t_list_ptr, t_offset);
226
227	/* Build the attribute offset */
228	attribute_offset = lp_build_const_int32(base->gallivm, 0);
229
230	/* Load the buffer index is always, which is always stored in VGPR0
231	 * for Vertex Shaders */
232	buffer_index_reg = build_intrinsic(base->gallivm->builder,
233		"llvm.SI.vs.load.buffer.index", uint->elem_type, NULL, 0,
234		LLVMReadNoneAttribute);
235
236	vec4_type = LLVMVectorType(base->elem_type, 4);
237	args[0] = t_list;
238	args[1] = attribute_offset;
239	args[2] = buffer_index_reg;
240	input = lp_build_intrinsic(base->gallivm->builder,
241		"llvm.SI.vs.load.input", vec4_type, args, 3);
242
243	/* Break up the vec4 into individual components */
244	for (chan = 0; chan < 4; chan++) {
245		LLVMValueRef llvm_chan = lp_build_const_int32(base->gallivm, chan);
246		/* XXX: Use a helper function for this.  There is one in
247 		 * tgsi_llvm.c. */
248		si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, chan)] =
249				LLVMBuildExtractElement(base->gallivm->builder,
250				input, llvm_chan, "");
251	}
252}
253
254static void declare_input_fs(
255	struct si_shader_context * si_shader_ctx,
256	unsigned input_index,
257	const struct tgsi_full_declaration *decl)
258{
259	const char * intr_name;
260	unsigned chan;
261	struct lp_build_context * base =
262				&si_shader_ctx->radeon_bld.soa.bld_base.base;
263	struct gallivm_state * gallivm = base->gallivm;
264
265	/* This value is:
266	 * [15:0] NewPrimMask (Bit mask for each quad.  It is set it the
267	 *                     quad begins a new primitive.  Bit 0 always needs
268	 *                     to be unset)
269	 * [32:16] ParamOffset
270	 *
271	 */
272	/* XXX: This register number must be identical to the S_00B02C_USER_SGPR
273	 * register field value
274	 */
275	LLVMValueRef params = use_sgpr(base->gallivm, SGPR_I32, 6);
276
277
278	/* XXX: Is this the input_index? */
279	LLVMValueRef attr_number = lp_build_const_int32(gallivm, input_index);
280
281	/* XXX: Handle all possible interpolation modes */
282	switch (decl->Interp.Interpolate) {
283	case TGSI_INTERPOLATE_COLOR:
284		/* XXX: Flat shading hangs the GPU */
285		if (si_shader_ctx->rctx->queued.named.rasterizer &&
286		    si_shader_ctx->rctx->queued.named.rasterizer->flatshade) {
287#if 0
288			intr_name = "llvm.SI.fs.interp.constant";
289#else
290			intr_name = "llvm.SI.fs.interp.linear.center";
291#endif
292		} else {
293			if (decl->Interp.Centroid)
294				intr_name = "llvm.SI.fs.interp.persp.centroid";
295			else
296				intr_name = "llvm.SI.fs.interp.persp.center";
297		}
298		break;
299	case TGSI_INTERPOLATE_CONSTANT:
300		/* XXX: Flat shading hangs the GPU */
301#if 0
302		intr_name = "llvm.SI.fs.interp.constant";
303		break;
304#endif
305	case TGSI_INTERPOLATE_LINEAR:
306		if (decl->Interp.Centroid)
307			intr_name = "llvm.SI.fs.interp.linear.centroid";
308		else
309			intr_name = "llvm.SI.fs.interp.linear.center";
310		break;
311	case TGSI_INTERPOLATE_PERSPECTIVE:
312		if (decl->Interp.Centroid)
313			intr_name = "llvm.SI.fs.interp.persp.centroid";
314		else
315			intr_name = "llvm.SI.fs.interp.persp.center";
316		break;
317	default:
318		fprintf(stderr, "Warning: Unhandled interpolation mode.\n");
319		return;
320	}
321
322	/* XXX: Could there be more than TGSI_NUM_CHANNELS (4) ? */
323	for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
324		LLVMValueRef args[3];
325		LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
326		unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
327		LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context);
328		args[0] = llvm_chan;
329		args[1] = attr_number;
330		args[2] = params;
331		si_shader_ctx->radeon_bld.inputs[soa_index] =
332			build_intrinsic(base->gallivm->builder, intr_name,
333				input_type, args, 3, LLVMReadOnlyAttribute);
334	}
335}
336
337static void declare_input(
338	struct radeon_llvm_context * radeon_bld,
339	unsigned input_index,
340	const struct tgsi_full_declaration *decl)
341{
342	struct si_shader_context * si_shader_ctx =
343				si_shader_context(&radeon_bld->soa.bld_base);
344	if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
345		declare_input_vs(si_shader_ctx, input_index, decl);
346	} else if (si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT) {
347		declare_input_fs(si_shader_ctx, input_index, decl);
348	} else {
349		fprintf(stderr, "Warning: Unsupported shader type,\n");
350	}
351}
352
353static LLVMValueRef fetch_constant(
354	struct lp_build_tgsi_context * bld_base,
355	const struct tgsi_full_src_register *reg,
356	enum tgsi_opcode_type type,
357	unsigned swizzle)
358{
359	struct lp_build_context * base = &bld_base->base;
360
361	LLVMValueRef const_ptr;
362	LLVMValueRef offset;
363	LLVMValueRef load;
364
365	/* XXX: Assume the pointer to the constant buffer is being stored in
366	 * SGPR[0:1] */
367	const_ptr = use_sgpr(base->gallivm, SGPR_CONST_PTR_F32, 0);
368
369	/* XXX: This assumes that the constant buffer is not packed, so
370	 * CONST[0].x will have an offset of 0 and CONST[1].x will have an
371	 * offset of 4. */
372	offset = lp_build_const_int32(base->gallivm,
373					(reg->Register.Index * 4) + swizzle);
374
375	load = build_indexed_load(base->gallivm, const_ptr, offset);
376	return bitcast(bld_base, type, load);
377}
378
379/* Initialize arguments for the shader export intrinsic */
380static void si_llvm_init_export_args(struct lp_build_tgsi_context *bld_base,
381				     struct tgsi_full_declaration *d,
382				     unsigned index,
383				     unsigned target,
384				     LLVMValueRef *args)
385{
386	struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
387	struct lp_build_context *uint =
388				&si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
389	struct lp_build_context *base = &bld_base->base;
390	unsigned compressed = 0;
391	unsigned chan;
392
393	if (si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT) {
394		int cbuf = target - V_008DFC_SQ_EXP_MRT;
395
396		if (cbuf >= 0 && cbuf < 8) {
397			struct r600_context *rctx = si_shader_ctx->rctx;
398			compressed = (rctx->export_16bpc >> cbuf) & 0x1;
399		}
400	}
401
402	if (compressed) {
403		/* Pixel shader needs to pack output values before export */
404		for (chan = 0; chan < 2; chan++ ) {
405			LLVMValueRef *out_ptr =
406				si_shader_ctx->radeon_bld.soa.outputs[index];
407			args[0] = LLVMBuildLoad(base->gallivm->builder,
408						out_ptr[2 * chan], "");
409			args[1] = LLVMBuildLoad(base->gallivm->builder,
410						out_ptr[2 * chan + 1], "");
411			args[chan + 5] =
412				build_intrinsic(base->gallivm->builder,
413						"llvm.SI.packf16",
414						LLVMInt32TypeInContext(base->gallivm->context),
415						args, 2,
416						LLVMReadNoneAttribute);
417			args[chan + 7] = args[chan + 5];
418		}
419
420		/* Set COMPR flag */
421		args[4] = uint->one;
422	} else {
423		for (chan = 0; chan < 4; chan++ ) {
424			LLVMValueRef out_ptr =
425				si_shader_ctx->radeon_bld.soa.outputs[index][chan];
426			/* +5 because the first output value will be
427			 * the 6th argument to the intrinsic. */
428			args[chan + 5] = LLVMBuildLoad(base->gallivm->builder,
429						       out_ptr, "");
430		}
431
432		/* Clear COMPR flag */
433		args[4] = uint->zero;
434	}
435
436	/* XXX: This controls which components of the output
437	 * registers actually get exported. (e.g bit 0 means export
438	 * X component, bit 1 means export Y component, etc.)  I'm
439	 * hard coding this to 0xf for now.  In the future, we might
440	 * want to do something else. */
441	args[0] = lp_build_const_int32(base->gallivm, 0xf);
442
443	/* Specify whether the EXEC mask represents the valid mask */
444	args[1] = uint->zero;
445
446	/* Specify whether this is the last export */
447	args[2] = uint->zero;
448
449	/* Specify the target we are exporting */
450	args[3] = lp_build_const_int32(base->gallivm, target);
451
452	/* XXX: We probably need to keep track of the output
453	 * values, so we know what we are passing to the next
454	 * stage. */
455}
456
457/* XXX: This is partially implemented for VS only at this point.  It is not complete */
458static void si_llvm_emit_epilogue(struct lp_build_tgsi_context * bld_base)
459{
460	struct si_shader_context * si_shader_ctx = si_shader_context(bld_base);
461	struct si_shader * shader = &si_shader_ctx->shader->shader;
462	struct lp_build_context * base = &bld_base->base;
463	struct lp_build_context * uint =
464				&si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
465	struct tgsi_parse_context *parse = &si_shader_ctx->parse;
466	LLVMValueRef last_args[9] = { 0 };
467	unsigned color_count = 0;
468	unsigned param_count = 0;
469
470	while (!tgsi_parse_end_of_tokens(parse)) {
471		struct tgsi_full_declaration *d =
472					&parse->FullToken.FullDeclaration;
473		LLVMValueRef args[9];
474		unsigned target;
475		unsigned index;
476		int i;
477
478		tgsi_parse_token(parse);
479		if (parse->FullToken.Token.Type != TGSI_TOKEN_TYPE_DECLARATION)
480			continue;
481
482		switch (d->Declaration.File) {
483		case TGSI_FILE_INPUT:
484			i = shader->ninput++;
485			shader->input[i].name = d->Semantic.Name;
486			shader->input[i].sid = d->Semantic.Index;
487			shader->input[i].interpolate = d->Interp.Interpolate;
488			shader->input[i].centroid = d->Interp.Centroid;
489			continue;
490
491		case TGSI_FILE_OUTPUT:
492			i = shader->noutput++;
493			shader->output[i].name = d->Semantic.Name;
494			shader->output[i].sid = d->Semantic.Index;
495			shader->output[i].interpolate = d->Interp.Interpolate;
496			break;
497
498		default:
499			continue;
500		}
501
502		for (index = d->Range.First; index <= d->Range.Last; index++) {
503			/* Select the correct target */
504			switch(d->Semantic.Name) {
505			case TGSI_SEMANTIC_POSITION:
506				target = V_008DFC_SQ_EXP_POS;
507				break;
508			case TGSI_SEMANTIC_COLOR:
509				if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
510					target = V_008DFC_SQ_EXP_PARAM + param_count;
511					shader->output[i].param_offset = param_count;
512					param_count++;
513				} else {
514					target = V_008DFC_SQ_EXP_MRT + color_count;
515					color_count++;
516				}
517				break;
518			case TGSI_SEMANTIC_GENERIC:
519				target = V_008DFC_SQ_EXP_PARAM + param_count;
520				shader->output[i].param_offset = param_count;
521				param_count++;
522				break;
523			default:
524				target = 0;
525				fprintf(stderr,
526					"Warning: SI unhandled output type:%d\n",
527					d->Semantic.Name);
528			}
529
530			si_llvm_init_export_args(bld_base, d, index, target, args);
531
532			if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX ?
533			    (d->Semantic.Name == TGSI_SEMANTIC_POSITION) :
534			    (d->Semantic.Name == TGSI_SEMANTIC_COLOR)) {
535				if (last_args[0]) {
536					lp_build_intrinsic(base->gallivm->builder,
537							   "llvm.SI.export",
538							   LLVMVoidTypeInContext(base->gallivm->context),
539							   last_args, 9);
540				}
541
542				memcpy(last_args, args, sizeof(args));
543			} else {
544				lp_build_intrinsic(base->gallivm->builder,
545						   "llvm.SI.export",
546						   LLVMVoidTypeInContext(base->gallivm->context),
547						   args, 9);
548			}
549
550		}
551	}
552
553	if (!last_args[0]) {
554		assert(si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT);
555
556		/* Specify which components to enable */
557		last_args[0] = lp_build_const_int32(base->gallivm, 0x0);
558
559		/* Specify the target we are exporting */
560		last_args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_MRT);
561
562		/* Set COMPR flag to zero to export data as 32-bit */
563		last_args[4] = uint->zero;
564
565		/* dummy bits */
566		last_args[5]= uint->zero;
567		last_args[6]= uint->zero;
568		last_args[7]= uint->zero;
569		last_args[8]= uint->zero;
570	}
571
572	/* Specify whether the EXEC mask represents the valid mask */
573	last_args[1] = lp_build_const_int32(base->gallivm,
574					    si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT);
575
576	/* Specify that this is the last export */
577	last_args[2] = lp_build_const_int32(base->gallivm, 1);
578
579	lp_build_intrinsic(base->gallivm->builder,
580			   "llvm.SI.export",
581			   LLVMVoidTypeInContext(base->gallivm->context),
582			   last_args, 9);
583
584/* XXX: Look up what this function does */
585/*		ctx->shader->output[i].spi_sid = r600_spi_sid(&ctx->shader->output[i]);*/
586}
587
588static void tex_fetch_args(
589	struct lp_build_tgsi_context * bld_base,
590	struct lp_build_emit_data * emit_data)
591{
592	const struct tgsi_full_instruction * inst = emit_data->inst;
593	LLVMValueRef ptr;
594	LLVMValueRef offset;
595
596	/* WriteMask */
597	/* XXX: should be optimized using emit_data->inst->Dst[0].Register.WriteMask*/
598	emit_data->args[0] = lp_build_const_int32(bld_base->base.gallivm, 0xf);
599
600	/* Coordinates */
601	/* XXX: Not all sample instructions need 4 address arguments. */
602	if (inst->Instruction.Opcode == TGSI_OPCODE_TXP) {
603		LLVMValueRef src_w;
604		unsigned chan;
605		LLVMValueRef coords[4];
606
607		emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
608		src_w = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
609
610		for (chan = 0; chan < 3; chan++ ) {
611			LLVMValueRef arg = lp_build_emit_fetch(bld_base,
612							       emit_data->inst, 0, chan);
613			coords[chan] = lp_build_emit_llvm_binary(bld_base,
614								 TGSI_OPCODE_DIV,
615								 arg, src_w);
616		}
617		coords[3] = bld_base->base.one;
618		emit_data->args[1] = lp_build_gather_values(bld_base->base.gallivm,
619							    coords, 4);
620	} else
621		emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
622							 0, LP_CHAN_ALL);
623
624	/* Resource */
625	ptr = use_sgpr(bld_base->base.gallivm, SGPR_CONST_PTR_V8I32, 4);
626	offset = lp_build_const_int32(bld_base->base.gallivm,
627				  emit_data->inst->Src[1].Register.Index);
628	emit_data->args[2] = build_indexed_load(bld_base->base.gallivm,
629						ptr, offset);
630
631	/* Sampler */
632	ptr = use_sgpr(bld_base->base.gallivm, SGPR_CONST_PTR_V4I32, 2);
633	offset = lp_build_const_int32(bld_base->base.gallivm,
634				  emit_data->inst->Src[1].Register.Index);
635	emit_data->args[3] = build_indexed_load(bld_base->base.gallivm,
636						ptr, offset);
637
638	/* Dimensions */
639	/* XXX: We might want to pass this information to the shader at some. */
640/*	emit_data->args[4] = lp_build_const_int32(bld_base->base.gallivm,
641					emit_data->inst->Texture.Texture);
642*/
643
644	emit_data->arg_count = 4;
645	/* XXX: To optimize, we could use a float or v2f32, if the last bits of
646	 * the writemask are clear */
647	emit_data->dst_type = LLVMVectorType(
648			LLVMFloatTypeInContext(bld_base->base.gallivm->context),
649			4);
650}
651
652static const struct lp_build_tgsi_action tex_action = {
653	.fetch_args = tex_fetch_args,
654	.emit = lp_build_tgsi_intrinsic,
655	.intr_name = "llvm.SI.sample"
656};
657
658
659int si_pipe_shader_create(
660	struct pipe_context *ctx,
661	struct si_pipe_shader *shader)
662{
663	struct r600_context *rctx = (struct r600_context*)ctx;
664	struct si_pipe_shader_selector *sel = shader->selector;
665	struct si_shader_context si_shader_ctx;
666	struct tgsi_shader_info shader_info;
667	struct lp_build_tgsi_context * bld_base;
668	LLVMModuleRef mod;
669	unsigned char * inst_bytes;
670	unsigned inst_byte_count;
671	unsigned i;
672	uint32_t *ptr;
673	bool dump;
674
675	dump = debug_get_bool_option("RADEON_DUMP_SHADERS", FALSE);
676
677	memset(&si_shader_ctx.radeon_bld, 0, sizeof(si_shader_ctx.radeon_bld));
678	radeon_llvm_context_init(&si_shader_ctx.radeon_bld);
679	bld_base = &si_shader_ctx.radeon_bld.soa.bld_base;
680
681	tgsi_scan_shader(sel->tokens, &shader_info);
682	bld_base->info = &shader_info;
683	bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
684	bld_base->emit_epilogue = si_llvm_emit_epilogue;
685
686	bld_base->op_actions[TGSI_OPCODE_TEX] = tex_action;
687	bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action;
688
689	si_shader_ctx.radeon_bld.load_input = declare_input;
690	si_shader_ctx.tokens = sel->tokens;
691	tgsi_parse_init(&si_shader_ctx.parse, si_shader_ctx.tokens);
692	si_shader_ctx.shader = shader;
693	si_shader_ctx.type = si_shader_ctx.parse.FullHeader.Processor.Processor;
694	si_shader_ctx.rctx = rctx;
695
696	shader->shader.nr_cbufs = rctx->framebuffer.nr_cbufs;
697
698	/* Dump TGSI code before doing TGSI->LLVM conversion in case the
699	 * conversion fails. */
700	if (dump) {
701		tgsi_dump(sel->tokens, 0);
702	}
703
704	if (!lp_build_tgsi_llvm(bld_base, sel->tokens)) {
705		fprintf(stderr, "Failed to translate shader from TGSI to LLVM\n");
706		return -EINVAL;
707	}
708
709	radeon_llvm_finalize_module(&si_shader_ctx.radeon_bld);
710
711	mod = bld_base->base.gallivm->module;
712	if (dump) {
713		LLVMDumpModule(mod);
714	}
715	radeon_llvm_compile(mod, &inst_bytes, &inst_byte_count, "SI", dump);
716	if (dump) {
717		fprintf(stderr, "SI CODE:\n");
718		for (i = 0; i < inst_byte_count; i+=4 ) {
719			fprintf(stderr, "%02x%02x%02x%02x\n", inst_bytes[i + 3],
720				inst_bytes[i + 2], inst_bytes[i + 1],
721				inst_bytes[i]);
722		}
723	}
724
725	shader->num_sgprs = util_le32_to_cpu(*(uint32_t*)inst_bytes);
726	shader->num_vgprs = util_le32_to_cpu(*(uint32_t*)(inst_bytes + 4));
727	shader->spi_ps_input_ena = util_le32_to_cpu(*(uint32_t*)(inst_bytes + 8));
728
729	radeon_llvm_dispose(&si_shader_ctx.radeon_bld);
730	tgsi_parse_free(&si_shader_ctx.parse);
731
732	/* copy new shader */
733	si_resource_reference(&shader->bo, NULL);
734	shader->bo = si_resource_create_custom(ctx->screen, PIPE_USAGE_IMMUTABLE,
735					       inst_byte_count - 12);
736	if (shader->bo == NULL) {
737		return -ENOMEM;
738	}
739
740	ptr = (uint32_t*)rctx->ws->buffer_map(shader->bo->cs_buf, rctx->cs, PIPE_TRANSFER_WRITE);
741	if (0 /*R600_BIG_ENDIAN*/) {
742		for (i = 0; i < (inst_byte_count-12)/4; ++i) {
743			ptr[i] = util_bswap32(*(uint32_t*)(inst_bytes+12 + i*4));
744		}
745	} else {
746		memcpy(ptr, inst_bytes + 12, inst_byte_count - 12);
747	}
748	rctx->ws->buffer_unmap(shader->bo->cs_buf);
749
750	free(inst_bytes);
751
752	return 0;
753}
754
755void si_pipe_shader_destroy(struct pipe_context *ctx, struct si_pipe_shader *shader)
756{
757	si_resource_reference(&shader->bo, NULL);
758}
759