1/**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 VMware, Inc.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29/**
30 * @file
31 * Code generate the whole fragment pipeline.
32 *
33 * The fragment pipeline consists of the following stages:
34 * - early depth test
35 * - fragment shader
36 * - alpha test
37 * - depth/stencil test
38 * - blending
39 *
40 * This file has only the glue to assemble the fragment pipeline.  The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
45 *
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
51 *
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
56 *
57 * @author Jose Fonseca <jfonseca@vmware.com>
58 */
59
60#include <limits.h>
61#include "pipe/p_defines.h"
62#include "util/u_inlines.h"
63#include "util/u_memory.h"
64#include "util/u_pointer.h"
65#include "util/format/u_format.h"
66#include "util/u_dump.h"
67#include "util/u_string.h"
68#include "util/simple_list.h"
69#include "util/u_dual_blend.h"
70#include "util/u_upload_mgr.h"
71#include "util/os_time.h"
72#include "pipe/p_shader_tokens.h"
73#include "draw/draw_context.h"
74#include "tgsi/tgsi_dump.h"
75#include "tgsi/tgsi_scan.h"
76#include "tgsi/tgsi_parse.h"
77#include "gallivm/lp_bld_type.h"
78#include "gallivm/lp_bld_const.h"
79#include "gallivm/lp_bld_conv.h"
80#include "gallivm/lp_bld_init.h"
81#include "gallivm/lp_bld_intr.h"
82#include "gallivm/lp_bld_logic.h"
83#include "gallivm/lp_bld_tgsi.h"
84#include "gallivm/lp_bld_nir.h"
85#include "gallivm/lp_bld_swizzle.h"
86#include "gallivm/lp_bld_flow.h"
87#include "gallivm/lp_bld_debug.h"
88#include "gallivm/lp_bld_arit.h"
89#include "gallivm/lp_bld_bitarit.h"
90#include "gallivm/lp_bld_pack.h"
91#include "gallivm/lp_bld_format.h"
92#include "gallivm/lp_bld_quad.h"
93#include "gallivm/lp_bld_gather.h"
94
95#include "lp_bld_alpha.h"
96#include "lp_bld_blend.h"
97#include "lp_bld_depth.h"
98#include "lp_bld_interp.h"
99#include "lp_context.h"
100#include "lp_debug.h"
101#include "lp_perf.h"
102#include "lp_setup.h"
103#include "lp_state.h"
104#include "lp_tex_sample.h"
105#include "lp_flush.h"
106#include "lp_state_fs.h"
107#include "lp_rast.h"
108#include "nir/nir_to_tgsi_info.h"
109
110#include "lp_screen.h"
111#include "compiler/nir/nir_serialize.h"
112#include "util/mesa-sha1.h"
113/** Fragment shader number (for debugging) */
114static unsigned fs_no = 0;
115
116static void
117load_unswizzled_block(struct gallivm_state *gallivm,
118                      LLVMValueRef base_ptr,
119                      LLVMValueRef stride,
120                      unsigned block_width,
121                      unsigned block_height,
122                      LLVMValueRef* dst,
123                      struct lp_type dst_type,
124                      unsigned dst_count,
125                      unsigned dst_alignment,
126                      LLVMValueRef x_offset,
127                      LLVMValueRef y_offset,
128                      bool fb_fetch_twiddle);
129/**
130 * Checks if a format description is an arithmetic format
131 *
132 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
133 */
134static inline boolean
135is_arithmetic_format(const struct util_format_description *format_desc)
136{
137   boolean arith = false;
138   unsigned i;
139
140   for (i = 0; i < format_desc->nr_channels; ++i) {
141      arith |= format_desc->channel[i].size != format_desc->channel[0].size;
142      arith |= (format_desc->channel[i].size % 8) != 0;
143   }
144
145   return arith;
146}
147
148/**
149 * Checks if this format requires special handling due to required expansion
150 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
151 * SoA conversion.
152 */
153static inline boolean
154format_expands_to_float_soa(const struct util_format_description *format_desc)
155{
156   if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
157       format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
158      return true;
159   }
160   return false;
161}
162
163
164/**
165 * Retrieves the type representing the memory layout for a format
166 *
167 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
168 */
169static inline void
170lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
171                             struct lp_type* type)
172{
173   unsigned i;
174   unsigned chan;
175
176   if (format_expands_to_float_soa(format_desc)) {
177      /* just make this a uint with width of block */
178      type->floating = false;
179      type->fixed = false;
180      type->sign = false;
181      type->norm = false;
182      type->width = format_desc->block.bits;
183      type->length = 1;
184      return;
185   }
186
187   for (i = 0; i < 4; i++)
188      if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
189         break;
190   chan = i;
191
192   memset(type, 0, sizeof(struct lp_type));
193   type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
194   type->fixed    = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
195   type->sign     = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
196   type->norm     = format_desc->channel[chan].normalized;
197
198   if (is_arithmetic_format(format_desc)) {
199      type->width = 0;
200      type->length = 1;
201
202      for (i = 0; i < format_desc->nr_channels; ++i) {
203         type->width += format_desc->channel[i].size;
204      }
205   } else {
206      type->width = format_desc->channel[chan].size;
207      type->length = format_desc->nr_channels;
208   }
209}
210
211/**
212 * Expand the relevant bits of mask_input to a n*4-dword mask for the
213 * n*four pixels in n 2x2 quads.  This will set the n*four elements of the
214 * quad mask vector to 0 or ~0.
215 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
216 * quad arguments with fs length 8.
217 *
218 * \param first_quad  which quad(s) of the quad group to test, in [0,3]
219 * \param mask_input  bitwise mask for the whole 4x4 stamp
220 */
221static LLVMValueRef
222generate_quad_mask(struct gallivm_state *gallivm,
223                   struct lp_type fs_type,
224                   unsigned first_quad,
225                   unsigned sample,
226                   LLVMValueRef mask_input) /* int64 */
227{
228   LLVMBuilderRef builder = gallivm->builder;
229   struct lp_type mask_type;
230   LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
231   LLVMValueRef bits[16];
232   LLVMValueRef mask, bits_vec;
233   int shift, i;
234
235   /*
236    * XXX: We'll need a different path for 16 x u8
237    */
238   assert(fs_type.width == 32);
239   assert(fs_type.length <= ARRAY_SIZE(bits));
240   mask_type = lp_int_type(fs_type);
241
242   /*
243    * mask_input >>= (quad * 4)
244    */
245   switch (first_quad) {
246   case 0:
247      shift = 0;
248      break;
249   case 1:
250      assert(fs_type.length == 4);
251      shift = 2;
252      break;
253   case 2:
254      shift = 8;
255      break;
256   case 3:
257      assert(fs_type.length == 4);
258      shift = 10;
259      break;
260   default:
261      assert(0);
262      shift = 0;
263   }
264
265   mask_input = LLVMBuildLShr(builder, mask_input, lp_build_const_int64(gallivm, 16 * sample), "");
266   mask_input = LLVMBuildTrunc(builder, mask_input,
267                               i32t, "");
268   mask_input = LLVMBuildAnd(builder, mask_input, lp_build_const_int32(gallivm, 0xffff), "");
269
270   mask_input = LLVMBuildLShr(builder,
271                              mask_input,
272                              LLVMConstInt(i32t, shift, 0),
273                              "");
274
275   /*
276    * mask = { mask_input & (1 << i), for i in [0,3] }
277    */
278   mask = lp_build_broadcast(gallivm,
279                             lp_build_vec_type(gallivm, mask_type),
280                             mask_input);
281
282   for (i = 0; i < fs_type.length / 4; i++) {
283      unsigned j = 2 * (i % 2) + (i / 2) * 8;
284      bits[4*i + 0] = LLVMConstInt(i32t, 1ULL << (j + 0), 0);
285      bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
286      bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
287      bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
288   }
289   bits_vec = LLVMConstVector(bits, fs_type.length);
290   mask = LLVMBuildAnd(builder, mask, bits_vec, "");
291
292   /*
293    * mask = mask == bits ? ~0 : 0
294    */
295   mask = lp_build_compare(gallivm,
296                           mask_type, PIPE_FUNC_EQUAL,
297                           mask, bits_vec);
298
299   return mask;
300}
301
302
303#define EARLY_DEPTH_TEST  0x1
304#define LATE_DEPTH_TEST   0x2
305#define EARLY_DEPTH_WRITE 0x4
306#define LATE_DEPTH_WRITE  0x8
307
308static int
309find_output_by_semantic( const struct tgsi_shader_info *info,
310			 unsigned semantic,
311			 unsigned index )
312{
313   int i;
314
315   for (i = 0; i < info->num_outputs; i++)
316      if (info->output_semantic_name[i] == semantic &&
317	  info->output_semantic_index[i] == index)
318	 return i;
319
320   return -1;
321}
322
323
324/**
325 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
326 */
327static LLVMValueRef
328lp_llvm_viewport(LLVMValueRef context_ptr,
329                 struct gallivm_state *gallivm,
330                 LLVMValueRef viewport_index)
331{
332   LLVMBuilderRef builder = gallivm->builder;
333   LLVMValueRef ptr;
334   LLVMValueRef res;
335   struct lp_type viewport_type =
336      lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
337
338   ptr = lp_jit_context_viewports(gallivm, context_ptr);
339   ptr = LLVMBuildPointerCast(builder, ptr,
340            LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
341
342   res = lp_build_pointer_get(builder, ptr, viewport_index);
343
344   return res;
345}
346
347
348static LLVMValueRef
349lp_build_depth_clamp(struct gallivm_state *gallivm,
350                     LLVMBuilderRef builder,
351                     struct lp_type type,
352                     LLVMValueRef context_ptr,
353                     LLVMValueRef thread_data_ptr,
354                     LLVMValueRef z)
355{
356   LLVMValueRef viewport, min_depth, max_depth;
357   LLVMValueRef viewport_index;
358   struct lp_build_context f32_bld;
359
360   assert(type.floating);
361   lp_build_context_init(&f32_bld, gallivm, type);
362
363   /*
364    * Assumes clamping of the viewport index will occur in setup/gs. Value
365    * is passed through the rasterization stage via lp_rast_shader_inputs.
366    *
367    * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
368    *      semantics.
369    */
370   viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
371                       thread_data_ptr);
372
373   /*
374    * Load the min and max depth from the lp_jit_context.viewports
375    * array of lp_jit_viewport structures.
376    */
377   viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
378
379   /* viewports[viewport_index].min_depth */
380   min_depth = LLVMBuildExtractElement(builder, viewport,
381                  lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH), "");
382   min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
383
384   /* viewports[viewport_index].max_depth */
385   max_depth = LLVMBuildExtractElement(builder, viewport,
386                  lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH), "");
387   max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
388
389   /*
390    * Clamp to the min and max depth values for the given viewport.
391    */
392   return lp_build_clamp(&f32_bld, z, min_depth, max_depth);
393}
394
395static void
396lp_build_sample_alpha_to_coverage(struct gallivm_state *gallivm,
397                                  struct lp_type type,
398                                  unsigned coverage_samples,
399                                  LLVMValueRef num_loop,
400                                  LLVMValueRef loop_counter,
401                                  LLVMValueRef coverage_mask_store,
402                                  LLVMValueRef alpha)
403{
404   struct lp_build_context bld;
405   LLVMBuilderRef builder = gallivm->builder;
406   float step = 1.0 / coverage_samples;
407
408   lp_build_context_init(&bld, gallivm, type);
409   for (unsigned s = 0; s < coverage_samples; s++) {
410      LLVMValueRef alpha_ref_value = lp_build_const_vec(gallivm, type, step * s);
411      LLVMValueRef test = lp_build_cmp(&bld, PIPE_FUNC_GREATER, alpha, alpha_ref_value);
412
413      LLVMValueRef s_mask_idx = LLVMBuildMul(builder, lp_build_const_int32(gallivm, s), num_loop, "");
414      s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_counter, "");
415      LLVMValueRef s_mask_ptr = LLVMBuildGEP(builder, coverage_mask_store, &s_mask_idx, 1, "");
416      LLVMValueRef s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
417      s_mask = LLVMBuildAnd(builder, s_mask, test, "");
418      LLVMBuildStore(builder, s_mask, s_mask_ptr);
419   }
420};
421
422struct lp_build_fs_llvm_iface {
423   struct lp_build_fs_iface base;
424   struct lp_build_interp_soa_context *interp;
425   struct lp_build_for_loop_state *loop_state;
426   LLVMValueRef mask_store;
427   LLVMValueRef sample_id;
428   LLVMValueRef color_ptr_ptr;
429   LLVMValueRef color_stride_ptr;
430   LLVMValueRef color_sample_stride_ptr;
431   const struct lp_fragment_shader_variant_key *key;
432};
433
434static LLVMValueRef fs_interp(const struct lp_build_fs_iface *iface,
435                              struct lp_build_context *bld,
436                              unsigned attrib, unsigned chan,
437                              bool centroid, bool sample,
438                              LLVMValueRef attrib_indir,
439                              LLVMValueRef offsets[2])
440{
441   struct lp_build_fs_llvm_iface *fs_iface = (struct lp_build_fs_llvm_iface *)iface;
442   struct lp_build_interp_soa_context *interp = fs_iface->interp;
443   unsigned loc = TGSI_INTERPOLATE_LOC_CENTER;
444   if (centroid)
445      loc = TGSI_INTERPOLATE_LOC_CENTROID;
446   if (sample)
447      loc = TGSI_INTERPOLATE_LOC_SAMPLE;
448
449   return lp_build_interp_soa(interp, bld->gallivm, fs_iface->loop_state->counter,
450                              fs_iface->mask_store,
451                              attrib, chan, loc, attrib_indir, offsets);
452}
453
454static void fs_fb_fetch(const struct lp_build_fs_iface *iface,
455                        struct lp_build_context *bld,
456                        int location,
457                        LLVMValueRef result[4])
458{
459   assert(location >= FRAG_RESULT_DATA0 && location <= FRAG_RESULT_DATA7);
460   const int cbuf = location - FRAG_RESULT_DATA0;
461
462   struct lp_build_fs_llvm_iface *fs_iface = (struct lp_build_fs_llvm_iface *)iface;
463   struct gallivm_state *gallivm = bld->gallivm;
464   LLVMBuilderRef builder = gallivm->builder;
465   const struct lp_fragment_shader_variant_key *key = fs_iface->key;
466   LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
467   LLVMValueRef color_ptr = LLVMBuildLoad(builder, LLVMBuildGEP(builder, fs_iface->color_ptr_ptr, &index, 1, ""), "");
468   LLVMValueRef stride = LLVMBuildLoad(builder, LLVMBuildGEP(builder, fs_iface->color_stride_ptr, &index, 1, ""), "");
469
470   LLVMValueRef dst[4 * 4];
471   enum pipe_format cbuf_format = key->cbuf_format[cbuf];
472   const struct util_format_description* out_format_desc = util_format_description(cbuf_format);
473   struct lp_type dst_type;
474   unsigned block_size = bld->type.length;
475   unsigned block_height = key->resource_1d ? 1 : 2;
476   unsigned block_width = block_size / block_height;
477
478   lp_mem_type_from_format_desc(out_format_desc, &dst_type);
479
480   struct lp_type blend_type;
481   memset(&blend_type, 0, sizeof blend_type);
482   blend_type.floating = FALSE; /* values are integers */
483   blend_type.sign = FALSE;     /* values are unsigned */
484   blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
485   blend_type.width = 8;        /* 8-bit ubyte values */
486   blend_type.length = 16;      /* 16 elements per vector */
487
488   uint32_t dst_alignment;
489   /*
490    * Compute the alignment of the destination pointer in bytes
491    * We fetch 1-4 pixels, if the format has pot alignment then those fetches
492    * are always aligned by MIN2(16, fetch_width) except for buffers (not
493    * 1d tex but can't distinguish here) so need to stick with per-pixel
494    * alignment in this case.
495    */
496   if (key->resource_1d) {
497      dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
498   }
499   else {
500      dst_alignment = dst_type.length * dst_type.width / 8;
501   }
502   /* Force power-of-two alignment by extracting only the least-significant-bit */
503   dst_alignment = 1 << (ffs(dst_alignment) - 1);
504   /*
505    * Resource base and stride pointers are aligned to 16 bytes, so that's
506    * the maximum alignment we can guarantee
507    */
508   dst_alignment = MIN2(16, dst_alignment);
509
510   LLVMTypeRef blend_vec_type = lp_build_vec_type(gallivm, blend_type);
511   color_ptr = LLVMBuildBitCast(builder, color_ptr, LLVMPointerType(blend_vec_type, 0), "");
512
513   if (key->multisample) {
514      LLVMValueRef sample_stride = LLVMBuildLoad(builder,
515                                                 LLVMBuildGEP(builder, fs_iface->color_sample_stride_ptr,
516                                                              &index, 1, ""), "");
517      LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_stride, fs_iface->sample_id, "");
518      color_ptr = LLVMBuildGEP(builder, color_ptr, &sample_offset, 1, "");
519   }
520   /* fragment shader executes on 4x4 blocks. depending on vector width it can execute 2 or 4 iterations.
521    * only move to the next row once the top row has completed 8 wide 1 iteration, 4 wide 2 iterations */
522   LLVMValueRef x_offset = NULL, y_offset = NULL;
523   if (!key->resource_1d) {
524      LLVMValueRef counter = fs_iface->loop_state->counter;
525
526      if (block_size == 4) {
527         x_offset = LLVMBuildShl(builder,
528                                 LLVMBuildAnd(builder, fs_iface->loop_state->counter, lp_build_const_int32(gallivm, 1), ""),
529                                 lp_build_const_int32(gallivm, 1), "");
530         counter = LLVMBuildLShr(builder, fs_iface->loop_state->counter, lp_build_const_int32(gallivm, 1), "");
531      }
532      y_offset = LLVMBuildMul(builder, counter, lp_build_const_int32(gallivm, 2), "");
533   }
534   load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height, dst, dst_type, block_size, dst_alignment, x_offset, y_offset, true);
535
536   for (unsigned i = 0; i < block_size; i++) {
537      dst[i] = LLVMBuildBitCast(builder, dst[i], LLVMInt32TypeInContext(gallivm->context), "");
538   }
539   LLVMValueRef packed = lp_build_gather_values(gallivm, dst, block_size);
540
541   struct lp_type texel_type = bld->type;
542   if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
543       out_format_desc->channel[0].pure_integer) {
544      if (out_format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
545         texel_type = lp_type_int_vec(bld->type.width, bld->type.width * bld->type.length);
546      }
547      else if (out_format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
548         texel_type = lp_type_uint_vec(bld->type.width, bld->type.width * bld->type.length);
549      }
550   }
551   lp_build_unpack_rgba_soa(gallivm, out_format_desc,
552                            texel_type,
553                            packed, result);
554}
555
556/**
557 * Generate the fragment shader, depth/stencil test, and alpha tests.
558 */
559static void
560generate_fs_loop(struct gallivm_state *gallivm,
561                 struct lp_fragment_shader *shader,
562                 const struct lp_fragment_shader_variant_key *key,
563                 LLVMBuilderRef builder,
564                 struct lp_type type,
565                 LLVMValueRef context_ptr,
566                 LLVMValueRef sample_pos_array,
567                 LLVMValueRef num_loop,
568                 struct lp_build_interp_soa_context *interp,
569                 const struct lp_build_sampler_soa *sampler,
570                 const struct lp_build_image_soa *image,
571                 LLVMValueRef mask_store,
572                 LLVMValueRef (*out_color)[4],
573                 LLVMValueRef depth_base_ptr,
574                 LLVMValueRef depth_stride,
575                 LLVMValueRef depth_sample_stride,
576                 LLVMValueRef color_ptr_ptr,
577                 LLVMValueRef color_stride_ptr,
578                 LLVMValueRef color_sample_stride_ptr,
579                 LLVMValueRef facing,
580                 LLVMValueRef thread_data_ptr)
581{
582   const struct util_format_description *zs_format_desc = NULL;
583   const struct tgsi_token *tokens = shader->base.tokens;
584   struct lp_type int_type = lp_int_type(type);
585   LLVMTypeRef vec_type, int_vec_type;
586   LLVMValueRef mask_ptr = NULL, mask_val = NULL;
587   LLVMValueRef consts_ptr, num_consts_ptr;
588   LLVMValueRef ssbo_ptr, num_ssbo_ptr;
589   LLVMValueRef z;
590   LLVMValueRef z_value, s_value;
591   LLVMValueRef z_fb, s_fb;
592   LLVMValueRef depth_ptr;
593   LLVMValueRef stencil_refs[2];
594   LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
595   LLVMValueRef zs_samples = lp_build_const_int32(gallivm, key->zsbuf_nr_samples);
596   LLVMValueRef z_out = NULL, s_out = NULL;
597   struct lp_build_for_loop_state loop_state, sample_loop_state = {0};
598   struct lp_build_mask_context mask;
599   /*
600    * TODO: figure out if simple_shader optimization is really worthwile to
601    * keep. Disabled because it may hide some real bugs in the (depth/stencil)
602    * code since tests tend to take another codepath than real shaders.
603    */
604   boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
605                            shader->info.base.num_inputs < 3 &&
606                            shader->info.base.num_instructions < 8) && 0;
607   const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
608                                     util_blend_state_is_dual(&key->blend, 0);
609   const bool post_depth_coverage = shader->info.base.properties[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE];
610   unsigned attrib;
611   unsigned chan;
612   unsigned cbuf;
613   unsigned depth_mode;
614
615   struct lp_bld_tgsi_system_values system_values;
616
617   memset(&system_values, 0, sizeof(system_values));
618
619   /* truncate then sign extend. */
620   system_values.front_facing = LLVMBuildTrunc(gallivm->builder, facing, LLVMInt1TypeInContext(gallivm->context), "");
621   system_values.front_facing = LLVMBuildSExt(gallivm->builder, system_values.front_facing, LLVMInt32TypeInContext(gallivm->context), "");
622   system_values.view_index = lp_jit_thread_data_raster_state_view_index(gallivm,
623                                                                         thread_data_ptr);
624   if (key->depth.enabled ||
625       key->stencil[0].enabled) {
626
627      zs_format_desc = util_format_description(key->zsbuf_format);
628      assert(zs_format_desc);
629
630      if (shader->info.base.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL])
631         depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
632      else if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
633         if (shader->info.base.writes_memory)
634            depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
635         else if (key->alpha.enabled ||
636             key->blend.alpha_to_coverage ||
637             shader->info.base.uses_kill ||
638             shader->info.base.writes_samplemask) {
639            /* With alpha test and kill, can do the depth test early
640             * and hopefully eliminate some quads.  But need to do a
641             * special deferred depth write once the final mask value
642             * is known. This only works though if there's either no
643             * stencil test or the stencil value isn't written.
644             */
645            if (key->stencil[0].enabled && (key->stencil[0].writemask ||
646                                            (key->stencil[1].enabled &&
647                                             key->stencil[1].writemask)))
648               depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
649            else
650               depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
651         }
652         else
653            depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
654      }
655      else {
656         depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
657      }
658
659      if (!(key->depth.enabled && key->depth.writemask) &&
660          !(key->stencil[0].enabled && (key->stencil[0].writemask ||
661                                        (key->stencil[1].enabled &&
662                                         key->stencil[1].writemask))))
663         depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
664   }
665   else {
666      depth_mode = 0;
667   }
668
669   vec_type = lp_build_vec_type(gallivm, type);
670   int_vec_type = lp_build_vec_type(gallivm, int_type);
671
672   stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
673   stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
674   /* convert scalar stencil refs into vectors */
675   stencil_refs[0] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[0]);
676   stencil_refs[1] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[1]);
677
678   consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
679   num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
680
681   ssbo_ptr = lp_jit_context_ssbos(gallivm, context_ptr);
682   num_ssbo_ptr = lp_jit_context_num_ssbos(gallivm, context_ptr);
683
684   memset(outputs, 0, sizeof outputs);
685
686   /* Allocate color storage for each fragment sample */
687   LLVMValueRef color_store_size = num_loop;
688   if (key->min_samples > 1)
689      color_store_size = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, key->min_samples), "");
690
691   for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
692      for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
693         out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
694                                                       lp_build_vec_type(gallivm,
695                                                                         type),
696                                                       color_store_size, "color");
697      }
698   }
699   if (dual_source_blend) {
700      assert(key->nr_cbufs <= 1);
701      for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
702         out_color[1][chan] = lp_build_array_alloca(gallivm,
703                                                    lp_build_vec_type(gallivm,
704                                                                      type),
705                                                    color_store_size, "color1");
706      }
707   }
708   if (shader->info.base.writes_z) {
709      z_out = lp_build_array_alloca(gallivm,
710                                    lp_build_vec_type(gallivm, type),
711                                    color_store_size, "depth");
712   }
713
714   if (shader->info.base.writes_stencil) {
715      s_out = lp_build_array_alloca(gallivm,
716                                    lp_build_vec_type(gallivm, type),
717                                    color_store_size, "depth");
718   }
719
720   lp_build_for_loop_begin(&loop_state, gallivm,
721                           lp_build_const_int32(gallivm, 0),
722                           LLVMIntULT,
723                           num_loop,
724                           lp_build_const_int32(gallivm, 1));
725
726   LLVMValueRef sample_mask_in;
727   if (key->multisample) {
728      sample_mask_in = lp_build_const_int_vec(gallivm, type, 0);
729      /* create shader execution mask by combining all sample masks. */
730      for (unsigned s = 0; s < key->coverage_samples; s++) {
731         LLVMValueRef s_mask_idx = LLVMBuildMul(builder, num_loop, lp_build_const_int32(gallivm, s), "");
732         s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
733         LLVMValueRef s_mask = lp_build_pointer_get(builder, mask_store, s_mask_idx);
734         if (s == 0)
735            mask_val = s_mask;
736         else
737            mask_val = LLVMBuildOr(builder, s_mask, mask_val, "");
738
739         LLVMValueRef mask_in = LLVMBuildAnd(builder, s_mask, lp_build_const_int_vec(gallivm, type, (1ll << s)), "");
740         sample_mask_in = LLVMBuildOr(builder, sample_mask_in, mask_in, "");
741      }
742   } else {
743      sample_mask_in = lp_build_const_int_vec(gallivm, type, 1);
744      mask_ptr = LLVMBuildGEP(builder, mask_store,
745                              &loop_state.counter, 1, "mask_ptr");
746      mask_val = LLVMBuildLoad(builder, mask_ptr, "");
747
748      LLVMValueRef mask_in = LLVMBuildAnd(builder, mask_val, lp_build_const_int_vec(gallivm, type, 1), "");
749      sample_mask_in = LLVMBuildOr(builder, sample_mask_in, mask_in, "");
750   }
751
752   /* 'mask' will control execution based on quad's pixel alive/killed state */
753   lp_build_mask_begin(&mask, gallivm, type, mask_val);
754
755   if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
756      lp_build_mask_check(&mask);
757
758   /* Create storage for recombining sample masks after early Z pass. */
759   LLVMValueRef s_mask_or = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, type), "cov_mask_early_depth");
760   LLVMBuildStore(builder, LLVMConstNull(lp_build_int_vec_type(gallivm, type)), s_mask_or);
761
762   /* Create storage for post depth sample mask */
763   LLVMValueRef post_depth_sample_mask_in = NULL;
764   if (post_depth_coverage)
765      post_depth_sample_mask_in = lp_build_alloca(gallivm, int_vec_type, "post_depth_sample_mask_in");
766
767   LLVMValueRef s_mask = NULL, s_mask_ptr = NULL;
768   LLVMValueRef z_sample_value_store = NULL, s_sample_value_store = NULL;
769   LLVMValueRef z_fb_store = NULL, s_fb_store = NULL;
770   LLVMTypeRef z_type = NULL, z_fb_type = NULL;
771
772   /* Run early depth once per sample */
773   if (key->multisample) {
774
775      if (zs_format_desc) {
776         struct lp_type zs_type = lp_depth_type(zs_format_desc, type.length);
777         struct lp_type z_type = zs_type;
778         struct lp_type s_type = zs_type;
779         if (zs_format_desc->block.bits < type.width)
780            z_type.width = type.width;
781         if (zs_format_desc->block.bits == 8)
782            s_type.width = type.width;
783
784         else if (zs_format_desc->block.bits > 32) {
785            z_type.width = z_type.width / 2;
786            s_type.width = s_type.width / 2;
787            s_type.floating = 0;
788         }
789         z_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
790                                                      zs_samples, "z_sample_store");
791         s_sample_value_store = lp_build_array_alloca(gallivm, lp_build_int_vec_type(gallivm, type),
792                                                      zs_samples, "s_sample_store");
793         z_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, z_type),
794                                            zs_samples, "z_fb_store");
795         s_fb_store = lp_build_array_alloca(gallivm, lp_build_vec_type(gallivm, s_type),
796                                            zs_samples, "s_fb_store");
797      }
798      lp_build_for_loop_begin(&sample_loop_state, gallivm,
799                              lp_build_const_int32(gallivm, 0),
800                              LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
801                              lp_build_const_int32(gallivm, 1));
802
803      LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
804      s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
805      s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
806
807      s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
808      s_mask = LLVMBuildAnd(builder, s_mask, mask_val, "");
809   }
810
811
812   /* for multisample Z needs to be interpolated at sample points for testing. */
813   lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, key->multisample ? sample_loop_state.counter : NULL);
814   z = interp->pos[2];
815
816   depth_ptr = depth_base_ptr;
817   if (key->multisample) {
818      LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
819      depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
820   }
821
822   if (depth_mode & EARLY_DEPTH_TEST) {
823      /*
824       * Clamp according to ARB_depth_clamp semantics.
825       */
826      if (key->depth_clamp) {
827         z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
828                                  thread_data_ptr, z);
829      }
830      lp_build_depth_stencil_load_swizzled(gallivm, type,
831                                           zs_format_desc, key->resource_1d,
832                                           depth_ptr, depth_stride,
833                                           &z_fb, &s_fb, loop_state.counter);
834      lp_build_depth_stencil_test(gallivm,
835                                  &key->depth,
836                                  key->stencil,
837                                  type,
838                                  zs_format_desc,
839                                  key->multisample ? NULL : &mask,
840                                  &s_mask,
841                                  stencil_refs,
842                                  z, z_fb, s_fb,
843                                  facing,
844                                  &z_value, &s_value,
845                                  !simple_shader && !key->multisample);
846
847      if (depth_mode & EARLY_DEPTH_WRITE) {
848         lp_build_depth_stencil_write_swizzled(gallivm, type,
849                                               zs_format_desc, key->resource_1d,
850                                               NULL, NULL, NULL, loop_state.counter,
851                                               depth_ptr, depth_stride,
852                                               z_value, s_value);
853      }
854      /*
855       * Note mask check if stencil is enabled must be after ds write not after
856       * stencil test otherwise new stencil values may not get written if all
857       * fragments got killed by depth/stencil test.
858       */
859      if (!simple_shader && key->stencil[0].enabled && !key->multisample)
860         lp_build_mask_check(&mask);
861
862      if (key->multisample) {
863         z_fb_type = LLVMTypeOf(z_fb);
864         z_type = LLVMTypeOf(z_value);
865         lp_build_pointer_set(builder, z_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, z_value, lp_build_int_vec_type(gallivm, type), ""));
866         lp_build_pointer_set(builder, s_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, s_value, lp_build_int_vec_type(gallivm, type), ""));
867         lp_build_pointer_set(builder, z_fb_store, sample_loop_state.counter, z_fb);
868         lp_build_pointer_set(builder, s_fb_store, sample_loop_state.counter, s_fb);
869      }
870   }
871
872   if (key->multisample) {
873      /*
874       * Store the post-early Z coverage mask.
875       * Recombine the resulting coverage masks post early Z into the fragment
876       * shader execution mask.
877       */
878      LLVMValueRef tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
879      tmp_s_mask_or = LLVMBuildOr(builder, tmp_s_mask_or, s_mask, "");
880      LLVMBuildStore(builder, tmp_s_mask_or, s_mask_or);
881
882      if (post_depth_coverage) {
883         LLVMValueRef mask_bit_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
884         LLVMValueRef post_depth_mask_in = LLVMBuildLoad(builder, post_depth_sample_mask_in, "");
885         mask_bit_idx = LLVMBuildAnd(builder, s_mask, lp_build_broadcast(gallivm, int_vec_type, mask_bit_idx), "");
886         post_depth_mask_in = LLVMBuildOr(builder, post_depth_mask_in, mask_bit_idx, "");
887         LLVMBuildStore(builder, post_depth_mask_in, post_depth_sample_mask_in);
888      }
889
890      LLVMBuildStore(builder, s_mask, s_mask_ptr);
891
892      lp_build_for_loop_end(&sample_loop_state);
893
894      /* recombined all the coverage masks in the shader exec mask. */
895      tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");
896      lp_build_mask_update(&mask, tmp_s_mask_or);
897
898      if (key->min_samples == 1) {
899         /* for multisample Z needs to be re interpolated at pixel center */
900         lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, NULL);
901         z = interp->pos[2];
902         lp_build_mask_update(&mask, tmp_s_mask_or);
903      }
904   } else {
905      if (post_depth_coverage) {
906         LLVMValueRef post_depth_mask_in = LLVMBuildAnd(builder, lp_build_mask_value(&mask), lp_build_const_int_vec(gallivm, type, 1), "");
907         LLVMBuildStore(builder, post_depth_mask_in, post_depth_sample_mask_in);
908      }
909   }
910
911   LLVMValueRef out_sample_mask_storage = NULL;
912   if (shader->info.base.writes_samplemask) {
913      out_sample_mask_storage = lp_build_alloca(gallivm, int_vec_type, "write_mask");
914      if (key->min_samples > 1)
915         LLVMBuildStore(builder, LLVMConstNull(int_vec_type), out_sample_mask_storage);
916   }
917
918   if (post_depth_coverage) {
919      system_values.sample_mask_in = LLVMBuildLoad(builder, post_depth_sample_mask_in, "");
920   }
921   else
922      system_values.sample_mask_in = sample_mask_in;
923   if (key->multisample && key->min_samples > 1) {
924      lp_build_for_loop_begin(&sample_loop_state, gallivm,
925                              lp_build_const_int32(gallivm, 0),
926                              LLVMIntULT,
927                              lp_build_const_int32(gallivm, key->min_samples),
928                              lp_build_const_int32(gallivm, 1));
929
930      LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
931      s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
932      s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
933      s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
934      lp_build_mask_force(&mask, s_mask);
935      lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, sample_loop_state.counter);
936      system_values.sample_id = sample_loop_state.counter;
937      system_values.sample_mask_in = LLVMBuildAnd(builder, system_values.sample_mask_in,
938                                                  lp_build_broadcast(gallivm, int_vec_type,
939                                                                     LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "")), "");
940   } else {
941      system_values.sample_id = lp_build_const_int32(gallivm, 0);
942
943   }
944   system_values.sample_pos = sample_pos_array;
945
946   lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter, mask_store, sample_loop_state.counter);
947
948   struct lp_build_fs_llvm_iface fs_iface = {
949     .base.interp_fn = fs_interp,
950     .base.fb_fetch = fs_fb_fetch,
951     .interp = interp,
952     .loop_state = &loop_state,
953     .sample_id = system_values.sample_id,
954     .mask_store = mask_store,
955     .color_ptr_ptr = color_ptr_ptr,
956     .color_stride_ptr = color_stride_ptr,
957     .color_sample_stride_ptr = color_sample_stride_ptr,
958     .key = key,
959   };
960
961   struct lp_build_tgsi_params params;
962   memset(&params, 0, sizeof(params));
963
964   params.type = type;
965   params.mask = &mask;
966   params.fs_iface = &fs_iface.base;
967   params.consts_ptr = consts_ptr;
968   params.const_sizes_ptr = num_consts_ptr;
969   params.system_values = &system_values;
970   params.inputs = interp->inputs;
971   params.context_ptr = context_ptr;
972   params.thread_data_ptr = thread_data_ptr;
973   params.sampler = sampler;
974   params.info = &shader->info.base;
975   params.ssbo_ptr = ssbo_ptr;
976   params.ssbo_sizes_ptr = num_ssbo_ptr;
977   params.image = image;
978   params.aniso_filter_table = lp_jit_context_aniso_filter_table(gallivm, context_ptr);
979
980   /* Build the actual shader */
981   if (shader->base.type == PIPE_SHADER_IR_TGSI)
982      lp_build_tgsi_soa(gallivm, tokens, &params,
983                        outputs);
984   else
985      lp_build_nir_soa(gallivm, shader->base.ir.nir, &params,
986                       outputs);
987
988   /* Alpha test */
989   if (key->alpha.enabled) {
990      int color0 = find_output_by_semantic(&shader->info.base,
991                                           TGSI_SEMANTIC_COLOR,
992                                           0);
993
994      if (color0 != -1 && outputs[color0][3]) {
995         const struct util_format_description *cbuf_format_desc;
996         LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
997         LLVMValueRef alpha_ref_value;
998
999         alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
1000         alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
1001
1002         cbuf_format_desc = util_format_description(key->cbuf_format[0]);
1003
1004         lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
1005                             &mask, alpha, alpha_ref_value,
1006                             (depth_mode & LATE_DEPTH_TEST) != 0);
1007      }
1008   }
1009
1010   /* Emulate Alpha to Coverage with Alpha test */
1011   if (key->blend.alpha_to_coverage) {
1012      int color0 = find_output_by_semantic(&shader->info.base,
1013                                           TGSI_SEMANTIC_COLOR,
1014                                           0);
1015
1016      if (color0 != -1 && outputs[color0][3]) {
1017         LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
1018
1019         if (!key->multisample) {
1020            lp_build_alpha_to_coverage(gallivm, type,
1021                                       &mask, alpha,
1022                                       (depth_mode & LATE_DEPTH_TEST) != 0);
1023         } else {
1024            lp_build_sample_alpha_to_coverage(gallivm, type, key->coverage_samples, num_loop,
1025                                              loop_state.counter,
1026                                              mask_store, alpha);
1027         }
1028      }
1029   }
1030   if (key->blend.alpha_to_one && key->multisample) {
1031      for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib) {
1032         unsigned cbuf = shader->info.base.output_semantic_index[attrib];
1033         if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
1034             ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
1035            if (outputs[cbuf][3]) {
1036               LLVMBuildStore(builder, lp_build_const_vec(gallivm, type, 1.0), outputs[cbuf][3]);
1037            }
1038      }
1039   }
1040   if (shader->info.base.writes_samplemask) {
1041      LLVMValueRef output_smask = NULL;
1042      int smaski = find_output_by_semantic(&shader->info.base,
1043                                           TGSI_SEMANTIC_SAMPLEMASK,
1044                                           0);
1045      struct lp_build_context smask_bld;
1046      lp_build_context_init(&smask_bld, gallivm, int_type);
1047
1048      assert(smaski >= 0);
1049      output_smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");
1050      output_smask = LLVMBuildBitCast(builder, output_smask, smask_bld.vec_type, "");
1051      if (!key->multisample && key->no_ms_sample_mask_out) {
1052         output_smask = lp_build_and(&smask_bld, output_smask, smask_bld.one);
1053         output_smask = lp_build_cmp(&smask_bld, PIPE_FUNC_NOTEQUAL, output_smask, smask_bld.zero);
1054         lp_build_mask_update(&mask, output_smask);
1055      }
1056
1057      if (key->min_samples > 1) {
1058         /* only the bit corresponding to this sample is to be used. */
1059         LLVMValueRef tmp_mask = LLVMBuildLoad(builder, out_sample_mask_storage, "tmp_mask");
1060         LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
1061         LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, lp_build_broadcast(gallivm, int_vec_type, out_smask_idx), "");
1062         output_smask = LLVMBuildOr(builder, tmp_mask, smask_bit, "");
1063      }
1064
1065      LLVMBuildStore(builder, output_smask, out_sample_mask_storage);
1066   }
1067
1068   if (shader->info.base.writes_z) {
1069      int pos0 = find_output_by_semantic(&shader->info.base,
1070                                         TGSI_SEMANTIC_POSITION,
1071                                         0);
1072      LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos0][2], "");
1073      LLVMValueRef idx = loop_state.counter;
1074      if (key->min_samples > 1)
1075         idx = LLVMBuildAdd(builder, idx,
1076                            LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");
1077      LLVMValueRef ptr = LLVMBuildGEP(builder, z_out, &idx, 1, "");
1078      LLVMBuildStore(builder, out, ptr);
1079   }
1080
1081   if (shader->info.base.writes_stencil) {
1082      int sten_out = find_output_by_semantic(&shader->info.base,
1083                                             TGSI_SEMANTIC_STENCIL,
1084                                             0);
1085      LLVMValueRef out = LLVMBuildLoad(builder, outputs[sten_out][1], "output.s");
1086      LLVMValueRef idx = loop_state.counter;
1087      if (key->min_samples > 1)
1088         idx = LLVMBuildAdd(builder, idx,
1089                            LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");
1090      LLVMValueRef ptr = LLVMBuildGEP(builder, s_out, &idx, 1, "");
1091      LLVMBuildStore(builder, out, ptr);
1092   }
1093
1094
1095   /* Color write - per fragment sample */
1096   for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
1097   {
1098      unsigned cbuf = shader->info.base.output_semantic_index[attrib];
1099      if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
1100           ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
1101      {
1102         for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1103            if(outputs[attrib][chan]) {
1104               /* XXX: just initialize outputs to point at colors[] and
1105                * skip this.
1106                */
1107               LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
1108               LLVMValueRef color_ptr;
1109               LLVMValueRef color_idx = loop_state.counter;
1110               if (key->min_samples > 1)
1111                  color_idx = LLVMBuildAdd(builder, color_idx,
1112                                           LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");
1113               color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
1114                                        &color_idx, 1, "");
1115               lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
1116               LLVMBuildStore(builder, out, color_ptr);
1117            }
1118         }
1119      }
1120   }
1121
1122   if (key->multisample && key->min_samples > 1) {
1123      LLVMBuildStore(builder, lp_build_mask_value(&mask), s_mask_ptr);
1124      lp_build_for_loop_end(&sample_loop_state);
1125   }
1126
1127   if (key->multisample) {
1128      /* execute depth test for each sample */
1129      lp_build_for_loop_begin(&sample_loop_state, gallivm,
1130                              lp_build_const_int32(gallivm, 0),
1131                              LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),
1132                              lp_build_const_int32(gallivm, 1));
1133
1134      /* load the per-sample coverage mask */
1135      LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");
1136      s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");
1137      s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");
1138
1139      /* combine the execution mask post fragment shader with the coverage mask. */
1140      s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");
1141      if (key->min_samples == 1)
1142         s_mask = LLVMBuildAnd(builder, s_mask, lp_build_mask_value(&mask), "");
1143
1144      /* if the shader writes sample mask use that */
1145      if (shader->info.base.writes_samplemask) {
1146         LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");
1147         out_smask_idx = lp_build_broadcast(gallivm, int_vec_type, out_smask_idx);
1148         LLVMValueRef output_smask = LLVMBuildLoad(builder, out_sample_mask_storage, "");
1149         LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, out_smask_idx, "");
1150         LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int_vec(gallivm, int_type, 0), "");
1151         smask_bit = LLVMBuildSExt(builder, cmp, int_vec_type, "");
1152
1153         s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
1154      }
1155   }
1156
1157   depth_ptr = depth_base_ptr;
1158   if (key->multisample) {
1159      LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");
1160      depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");
1161   }
1162
1163   /* Late Z test */
1164   if (depth_mode & LATE_DEPTH_TEST) {
1165      if (shader->info.base.writes_z) {
1166         LLVMValueRef idx = loop_state.counter;
1167         if (key->min_samples > 1)
1168            idx = LLVMBuildAdd(builder, idx,
1169                               LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");
1170         LLVMValueRef ptr = LLVMBuildGEP(builder, z_out, &idx, 1, "");
1171         z = LLVMBuildLoad(builder, ptr, "output.z");
1172      } else {
1173         if (key->multisample) {
1174            lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, key->multisample ? sample_loop_state.counter : NULL);
1175            z = interp->pos[2];
1176         }
1177      }
1178
1179      /*
1180       * Clamp according to ARB_depth_clamp semantics.
1181       */
1182      if (key->depth_clamp) {
1183         z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
1184                                  thread_data_ptr, z);
1185      } else {
1186         struct lp_build_context f32_bld;
1187         lp_build_context_init(&f32_bld, gallivm, type);
1188         z = lp_build_clamp(&f32_bld, z,
1189                            lp_build_const_vec(gallivm, type, 0.0),
1190                            lp_build_const_vec(gallivm, type, 1.0));
1191      }
1192
1193      if (shader->info.base.writes_stencil) {
1194         LLVMValueRef idx = loop_state.counter;
1195         if (key->min_samples > 1)
1196            idx = LLVMBuildAdd(builder, idx,
1197                               LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");
1198         LLVMValueRef ptr = LLVMBuildGEP(builder, s_out, &idx, 1, "");
1199         stencil_refs[0] = LLVMBuildLoad(builder, ptr, "output.s");
1200         /* there's only one value, and spec says to discard additional bits */
1201         LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
1202         stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
1203         stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
1204         stencil_refs[1] = stencil_refs[0];
1205      }
1206
1207      lp_build_depth_stencil_load_swizzled(gallivm, type,
1208                                           zs_format_desc, key->resource_1d,
1209                                           depth_ptr, depth_stride,
1210                                           &z_fb, &s_fb, loop_state.counter);
1211
1212      lp_build_depth_stencil_test(gallivm,
1213                                  &key->depth,
1214                                  key->stencil,
1215                                  type,
1216                                  zs_format_desc,
1217                                  key->multisample ? NULL : &mask,
1218                                  &s_mask,
1219                                  stencil_refs,
1220                                  z, z_fb, s_fb,
1221                                  facing,
1222                                  &z_value, &s_value,
1223                                  !simple_shader);
1224      /* Late Z write */
1225      if (depth_mode & LATE_DEPTH_WRITE) {
1226         lp_build_depth_stencil_write_swizzled(gallivm, type,
1227                                               zs_format_desc, key->resource_1d,
1228                                               NULL, NULL, NULL, loop_state.counter,
1229                                               depth_ptr, depth_stride,
1230                                               z_value, s_value);
1231      }
1232   }
1233   else if ((depth_mode & EARLY_DEPTH_TEST) &&
1234            (depth_mode & LATE_DEPTH_WRITE))
1235   {
1236      /* Need to apply a reduced mask to the depth write.  Reload the
1237       * depth value, update from zs_value with the new mask value and
1238       * write that out.
1239       */
1240      if (key->multisample) {
1241         z_value = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_sample_value_store, sample_loop_state.counter), z_type, "");;
1242         s_value = lp_build_pointer_get(builder, s_sample_value_store, sample_loop_state.counter);
1243         z_fb = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_fb_store, sample_loop_state.counter), z_fb_type, "");
1244         s_fb = lp_build_pointer_get(builder, s_fb_store, sample_loop_state.counter);
1245      }
1246      lp_build_depth_stencil_write_swizzled(gallivm, type,
1247                                            zs_format_desc, key->resource_1d,
1248                                            key->multisample ? s_mask : lp_build_mask_value(&mask), z_fb, s_fb, loop_state.counter,
1249                                            depth_ptr, depth_stride,
1250                                            z_value, s_value);
1251   }
1252
1253   if (key->occlusion_count) {
1254      LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
1255      lp_build_name(counter, "counter");
1256
1257      lp_build_occlusion_count(gallivm, type,
1258                               key->multisample ? s_mask : lp_build_mask_value(&mask), counter);
1259   }
1260
1261   if (key->multisample) {
1262      /* store the sample mask for this loop */
1263      LLVMBuildStore(builder, s_mask, s_mask_ptr);
1264      lp_build_for_loop_end(&sample_loop_state);
1265   }
1266
1267   mask_val = lp_build_mask_end(&mask);
1268   if (!key->multisample)
1269      LLVMBuildStore(builder, mask_val, mask_ptr);
1270   lp_build_for_loop_end(&loop_state);
1271}
1272
1273
1274/**
1275 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
1276 *
1277 * Fragment Shader outputs pixels in small 2x2 blocks
1278 *  e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
1279 *
1280 * However in memory pixels are stored in rows
1281 *  e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
1282 *
1283 * @param type            fragment shader type (4x or 8x float)
1284 * @param num_fs          number of fs_src
1285 * @param is_1d           whether we're outputting to a 1d resource
1286 * @param dst_channels    number of output channels
1287 * @param fs_src          output from fragment shader
1288 * @param dst             pointer to store result
1289 * @param pad_inline      is channel padding inline or at end of row
1290 * @return                the number of dsts
1291 */
1292static int
1293generate_fs_twiddle(struct gallivm_state *gallivm,
1294                    struct lp_type type,
1295                    unsigned num_fs,
1296                    unsigned dst_channels,
1297                    LLVMValueRef fs_src[][4],
1298                    LLVMValueRef* dst,
1299                    bool pad_inline)
1300{
1301   LLVMValueRef src[16];
1302
1303   bool swizzle_pad;
1304   bool twiddle;
1305   bool split;
1306
1307   unsigned pixels = type.length / 4;
1308   unsigned reorder_group;
1309   unsigned src_channels;
1310   unsigned src_count;
1311   unsigned i;
1312
1313   src_channels = dst_channels < 3 ? dst_channels : 4;
1314   src_count = num_fs * src_channels;
1315
1316   assert(pixels == 2 || pixels == 1);
1317   assert(num_fs * src_channels <= ARRAY_SIZE(src));
1318
1319   /*
1320    * Transpose from SoA -> AoS
1321    */
1322   for (i = 0; i < num_fs; ++i) {
1323      lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
1324   }
1325
1326   /*
1327    * Pick transformation options
1328    */
1329   swizzle_pad = false;
1330   twiddle = false;
1331   split = false;
1332   reorder_group = 0;
1333
1334   if (dst_channels == 1) {
1335      twiddle = true;
1336
1337      if (pixels == 2) {
1338         split = true;
1339      }
1340   } else if (dst_channels == 2) {
1341      if (pixels == 1) {
1342         reorder_group = 1;
1343      }
1344   } else if (dst_channels > 2) {
1345      if (pixels == 1) {
1346         reorder_group = 2;
1347      } else {
1348         twiddle = true;
1349      }
1350
1351      if (!pad_inline && dst_channels == 3 && pixels > 1) {
1352         swizzle_pad = true;
1353      }
1354   }
1355
1356   /*
1357    * Split the src in half
1358    */
1359   if (split) {
1360      for (i = num_fs; i > 0; --i) {
1361         src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
1362         src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
1363      }
1364
1365      src_count *= 2;
1366      type.length = 4;
1367   }
1368
1369   /*
1370    * Ensure pixels are in memory order
1371    */
1372   if (reorder_group) {
1373      /* Twiddle pixels by reordering the array, e.g.:
1374       *
1375       * src_count =  8 -> 0 2 1 3 4 6 5 7
1376       * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
1377       */
1378      const unsigned reorder_sw[] = { 0, 2, 1, 3 };
1379
1380      for (i = 0; i < src_count; ++i) {
1381         unsigned group = i / reorder_group;
1382         unsigned block = (group / 4) * 4 * reorder_group;
1383         unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
1384         dst[i] = src[j];
1385      }
1386   } else if (twiddle) {
1387      /* Twiddle pixels across elements of array */
1388      /*
1389       * XXX: we should avoid this in some cases, but would need to tell
1390       * lp_build_conv to reorder (or deal with it ourselves).
1391       */
1392      lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
1393   } else {
1394      /* Do nothing */
1395      memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
1396   }
1397
1398   /*
1399    * Moves any padding between pixels to the end
1400    * e.g. RGBXRGBX -> RGBRGBXX
1401    */
1402   if (swizzle_pad) {
1403      unsigned char swizzles[16];
1404      unsigned elems = pixels * dst_channels;
1405
1406      for (i = 0; i < type.length; ++i) {
1407         if (i < elems)
1408            swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
1409         else
1410            swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
1411      }
1412
1413      for (i = 0; i < src_count; ++i) {
1414         dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
1415      }
1416   }
1417
1418   return src_count;
1419}
1420
1421
1422/*
1423 * Untwiddle and transpose, much like the above.
1424 * However, this is after conversion, so we get packed vectors.
1425 * At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,
1426 * the vectors will look like:
1427 * r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may
1428 * be swizzled here). Extending to 16bit should be trivial.
1429 * Should also be extended to handle twice wide vectors with AVX2...
1430 */
1431static void
1432fs_twiddle_transpose(struct gallivm_state *gallivm,
1433                     struct lp_type type,
1434                     LLVMValueRef *src,
1435                     unsigned src_count,
1436                     LLVMValueRef *dst)
1437{
1438   unsigned i, j;
1439   struct lp_type type64, type16, type32;
1440   LLVMTypeRef type64_t, type8_t, type16_t, type32_t;
1441   LLVMBuilderRef builder = gallivm->builder;
1442   LLVMValueRef tmp[4], shuf[8];
1443   for (j = 0; j < 2; j++) {
1444      shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);
1445      shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);
1446      shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);
1447      shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);
1448   }
1449
1450   assert(src_count == 4 || src_count == 2 || src_count == 1);
1451   assert(type.width == 8);
1452   assert(type.length == 16);
1453
1454   type8_t = lp_build_vec_type(gallivm, type);
1455
1456   type64 = type;
1457   type64.length /= 8;
1458   type64.width *= 8;
1459   type64_t = lp_build_vec_type(gallivm, type64);
1460
1461   type16 = type;
1462   type16.length /= 2;
1463   type16.width *= 2;
1464   type16_t = lp_build_vec_type(gallivm, type16);
1465
1466   type32 = type;
1467   type32.length /= 4;
1468   type32.width *= 4;
1469   type32_t = lp_build_vec_type(gallivm, type32);
1470
1471   lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);
1472
1473   if (src_count == 1) {
1474      /* transpose was no-op, just untwiddle */
1475      LLVMValueRef shuf_vec;
1476      shuf_vec = LLVMConstVector(shuf, 8);
1477      tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");
1478      tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");
1479      dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");
1480   } else if (src_count == 2) {
1481      LLVMValueRef shuf_vec;
1482      shuf_vec = LLVMConstVector(shuf, 4);
1483
1484      for (i = 0; i < 2; i++) {
1485         tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");
1486         tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");
1487         dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");
1488      }
1489   } else {
1490      for (j = 0; j < 2; j++) {
1491         LLVMValueRef lo, hi, lo2, hi2;
1492          /*
1493          * Note that if we only really have 3 valid channels (rgb)
1494          * and we don't need alpha we could substitute a undef here
1495          * for the respective channel (causing llvm to drop conversion
1496          * for alpha).
1497          */
1498         /* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */
1499         lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");
1500         hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");
1501         lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);
1502         hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);
1503         dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");
1504         dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");
1505      }
1506   }
1507}
1508
1509
1510/**
1511 * Load an unswizzled block of pixels from memory
1512 */
1513static void
1514load_unswizzled_block(struct gallivm_state *gallivm,
1515                      LLVMValueRef base_ptr,
1516                      LLVMValueRef stride,
1517                      unsigned block_width,
1518                      unsigned block_height,
1519                      LLVMValueRef* dst,
1520                      struct lp_type dst_type,
1521                      unsigned dst_count,
1522                      unsigned dst_alignment,
1523                      LLVMValueRef x_offset,
1524                      LLVMValueRef y_offset,
1525                      bool fb_fetch_twiddle)
1526{
1527   LLVMBuilderRef builder = gallivm->builder;
1528   unsigned row_size = dst_count / block_height;
1529   unsigned i;
1530
1531   /* Ensure block exactly fits into dst */
1532   assert((block_width * block_height) % dst_count == 0);
1533
1534   for (i = 0; i < dst_count; ++i) {
1535      unsigned x = i % row_size;
1536      unsigned y = i / row_size;
1537
1538      if (block_height == 2 && dst_count == 8 && fb_fetch_twiddle) {
1539         /* remap the raw slots into the fragment shader execution mode. */
1540         /* this math took me way too long to work out, I'm sure it's overkill. */
1541         x = (i & 1) + ((i >> 2) << 1);
1542         y = (i & 2) >> 1;
1543      }
1544
1545      LLVMValueRef x_val;
1546      if (x_offset) {
1547         x_val = lp_build_const_int32(gallivm, x);
1548         if (x_offset)
1549            x_val = LLVMBuildAdd(builder, x_val, x_offset, "");
1550         x_val = LLVMBuildMul(builder, x_val, lp_build_const_int32(gallivm, (dst_type.width / 8) * dst_type.length), "");
1551      } else
1552         x_val = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
1553
1554      LLVMValueRef bx = x_val;
1555
1556      LLVMValueRef y_val = lp_build_const_int32(gallivm, y);
1557      if (y_offset)
1558         y_val = LLVMBuildAdd(builder, y_val, y_offset, "");
1559      LLVMValueRef by = LLVMBuildMul(builder, y_val, stride, "");
1560
1561      LLVMValueRef gep[2];
1562      LLVMValueRef dst_ptr;
1563
1564      gep[0] = lp_build_const_int32(gallivm, 0);
1565      gep[1] = LLVMBuildAdd(builder, bx, by, "");
1566
1567      dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1568      dst_ptr = LLVMBuildBitCast(builder, dst_ptr,
1569                                 LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
1570
1571      dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
1572
1573      LLVMSetAlignment(dst[i], dst_alignment);
1574   }
1575}
1576
1577
1578/**
1579 * Store an unswizzled block of pixels to memory
1580 */
1581static void
1582store_unswizzled_block(struct gallivm_state *gallivm,
1583                       LLVMValueRef base_ptr,
1584                       LLVMValueRef stride,
1585                       unsigned block_width,
1586                       unsigned block_height,
1587                       LLVMValueRef* src,
1588                       struct lp_type src_type,
1589                       unsigned src_count,
1590                       unsigned src_alignment)
1591{
1592   LLVMBuilderRef builder = gallivm->builder;
1593   unsigned row_size = src_count / block_height;
1594   unsigned i;
1595
1596   /* Ensure src exactly fits into block */
1597   assert((block_width * block_height) % src_count == 0);
1598
1599   for (i = 0; i < src_count; ++i) {
1600      unsigned x = i % row_size;
1601      unsigned y = i / row_size;
1602
1603      LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
1604      LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
1605
1606      LLVMValueRef gep[2];
1607      LLVMValueRef src_ptr;
1608
1609      gep[0] = lp_build_const_int32(gallivm, 0);
1610      gep[1] = LLVMBuildAdd(builder, bx, by, "");
1611
1612      src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
1613      src_ptr = LLVMBuildBitCast(builder, src_ptr,
1614                                 LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
1615
1616      src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
1617
1618      LLVMSetAlignment(src_ptr, src_alignment);
1619   }
1620}
1621
1622
1623
1624/**
1625 * Retrieves the type for a format which is usable in the blending code.
1626 *
1627 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
1628 */
1629static inline void
1630lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
1631                               struct lp_type* type)
1632{
1633   unsigned i;
1634   unsigned chan;
1635
1636   if (format_expands_to_float_soa(format_desc)) {
1637      /* always use ordinary floats for blending */
1638      type->floating = true;
1639      type->fixed = false;
1640      type->sign = true;
1641      type->norm = false;
1642      type->width = 32;
1643      type->length = 4;
1644      return;
1645   }
1646
1647   for (i = 0; i < 4; i++)
1648      if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
1649         break;
1650   chan = i;
1651
1652   memset(type, 0, sizeof(struct lp_type));
1653   type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
1654   type->fixed    = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
1655   type->sign     = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
1656   type->norm     = format_desc->channel[chan].normalized;
1657   type->width    = format_desc->channel[chan].size;
1658   type->length   = format_desc->nr_channels;
1659
1660   for (i = 1; i < format_desc->nr_channels; ++i) {
1661      if (format_desc->channel[i].size > type->width)
1662         type->width = format_desc->channel[i].size;
1663   }
1664
1665   if (type->floating) {
1666      type->width = 32;
1667   } else {
1668      if (type->width <= 8) {
1669         type->width = 8;
1670      } else if (type->width <= 16) {
1671         type->width = 16;
1672      } else {
1673         type->width = 32;
1674      }
1675   }
1676
1677   if (is_arithmetic_format(format_desc) && type->length == 3) {
1678      type->length = 4;
1679   }
1680}
1681
1682
1683/**
1684 * Scale a normalized value from src_bits to dst_bits.
1685 *
1686 * The exact calculation is
1687 *
1688 *    dst = iround(src * dst_mask / src_mask)
1689 *
1690 *  or with integer rounding
1691 *
1692 *    dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
1693 *
1694 *  where
1695 *
1696 *    src_mask = (1 << src_bits) - 1
1697 *    dst_mask = (1 << dst_bits) - 1
1698 *
1699 * but we try to avoid division and multiplication through shifts.
1700 */
1701static inline LLVMValueRef
1702scale_bits(struct gallivm_state *gallivm,
1703           int src_bits,
1704           int dst_bits,
1705           LLVMValueRef src,
1706           struct lp_type src_type)
1707{
1708   LLVMBuilderRef builder = gallivm->builder;
1709   LLVMValueRef result = src;
1710
1711   if (dst_bits < src_bits) {
1712      int delta_bits = src_bits - dst_bits;
1713
1714      if (delta_bits <= dst_bits) {
1715
1716         if (dst_bits == 4) {
1717            struct lp_type flt_type = lp_type_float_vec(32, src_type.length * 32);
1718
1719            result = lp_build_unsigned_norm_to_float(gallivm, src_bits, flt_type, src);
1720            result = lp_build_clamped_float_to_unsigned_norm(gallivm, flt_type, dst_bits, result);
1721            return result;
1722         }
1723
1724         /*
1725          * Approximate the rescaling with a single shift.
1726          *
1727          * This gives the wrong rounding.
1728          */
1729
1730         result = LLVMBuildLShr(builder,
1731                                src,
1732                                lp_build_const_int_vec(gallivm, src_type, delta_bits),
1733                                "");
1734
1735      } else {
1736         /*
1737          * Try more accurate rescaling.
1738          */
1739
1740         /*
1741          * Drop the least significant bits to make space for the multiplication.
1742          *
1743          * XXX: A better approach would be to use a wider integer type as intermediate.  But
1744          * this is enough to convert alpha from 16bits -> 2 when rendering to
1745          * PIPE_FORMAT_R10G10B10A2_UNORM.
1746          */
1747         result = LLVMBuildLShr(builder,
1748                                src,
1749                                lp_build_const_int_vec(gallivm, src_type, dst_bits),
1750                                "");
1751
1752
1753         result = LLVMBuildMul(builder,
1754                               result,
1755                               lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
1756                               "");
1757
1758         /*
1759          * Add a rounding term before the division.
1760          *
1761          * TODO: Handle signed integers too.
1762          */
1763         if (!src_type.sign) {
1764            result = LLVMBuildAdd(builder,
1765                                  result,
1766                                  lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
1767                                  "");
1768         }
1769
1770         /*
1771          * Approximate the division by src_mask with a src_bits shift.
1772          *
1773          * Given the src has already been shifted by dst_bits, all we need
1774          * to do is to shift by the difference.
1775          */
1776
1777         result = LLVMBuildLShr(builder,
1778                                result,
1779                                lp_build_const_int_vec(gallivm, src_type, delta_bits),
1780                                "");
1781      }
1782
1783   } else if (dst_bits > src_bits) {
1784      /* Scale up bits */
1785      int db = dst_bits - src_bits;
1786
1787      /* Shift left by difference in bits */
1788      result = LLVMBuildShl(builder,
1789                            src,
1790                            lp_build_const_int_vec(gallivm, src_type, db),
1791                            "");
1792
1793      if (db <= src_bits) {
1794         /* Enough bits in src to fill the remainder */
1795         LLVMValueRef lower = LLVMBuildLShr(builder,
1796                                            src,
1797                                            lp_build_const_int_vec(gallivm, src_type, src_bits - db),
1798                                            "");
1799
1800         result = LLVMBuildOr(builder, result, lower, "");
1801      } else if (db > src_bits) {
1802         /* Need to repeatedly copy src bits to fill remainder in dst */
1803         unsigned n;
1804
1805         for (n = src_bits; n < dst_bits; n *= 2) {
1806            LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
1807
1808            result = LLVMBuildOr(builder,
1809                                 result,
1810                                 LLVMBuildLShr(builder, result, shuv, ""),
1811                                 "");
1812         }
1813      }
1814   }
1815
1816   return result;
1817}
1818
1819/**
1820 * If RT is a smallfloat (needing denorms) format
1821 */
1822static inline int
1823have_smallfloat_format(struct lp_type dst_type,
1824                       enum pipe_format format)
1825{
1826   return ((dst_type.floating && dst_type.width != 32) ||
1827    /* due to format handling hacks this format doesn't have floating set
1828     * here (and actually has width set to 32 too) so special case this. */
1829    (format == PIPE_FORMAT_R11G11B10_FLOAT));
1830}
1831
1832
1833/**
1834 * Convert from memory format to blending format
1835 *
1836 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1837 */
1838static void
1839convert_to_blend_type(struct gallivm_state *gallivm,
1840                      unsigned block_size,
1841                      const struct util_format_description *src_fmt,
1842                      struct lp_type src_type,
1843                      struct lp_type dst_type,
1844                      LLVMValueRef* src, // and dst
1845                      unsigned num_srcs)
1846{
1847   LLVMValueRef *dst = src;
1848   LLVMBuilderRef builder = gallivm->builder;
1849   struct lp_type blend_type;
1850   struct lp_type mem_type;
1851   unsigned i, j;
1852   unsigned pixels = block_size / num_srcs;
1853   bool is_arith;
1854
1855   /*
1856    * full custom path for packed floats and srgb formats - none of the later
1857    * functions would do anything useful, and given the lp_type representation they
1858    * can't be fixed. Should really have some SoA blend path for these kind of
1859    * formats rather than hacking them in here.
1860    */
1861   if (format_expands_to_float_soa(src_fmt)) {
1862      LLVMValueRef tmpsrc[4];
1863      /*
1864       * This is pretty suboptimal for this case blending in SoA would be much
1865       * better, since conversion gets us SoA values so need to convert back.
1866       */
1867      assert(src_type.width == 32 || src_type.width == 16);
1868      assert(dst_type.floating);
1869      assert(dst_type.width == 32);
1870      assert(dst_type.length % 4 == 0);
1871      assert(num_srcs % 4 == 0);
1872
1873      if (src_type.width == 16) {
1874         /* expand 4x16bit values to 4x32bit */
1875         struct lp_type type32x4 = src_type;
1876         LLVMTypeRef ltype32x4;
1877         unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
1878         type32x4.width = 32;
1879         ltype32x4 = lp_build_vec_type(gallivm, type32x4);
1880         for (i = 0; i < num_fetch; i++) {
1881            src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
1882         }
1883         src_type.width = 32;
1884      }
1885      for (i = 0; i < 4; i++) {
1886         tmpsrc[i] = src[i];
1887      }
1888      for (i = 0; i < num_srcs / 4; i++) {
1889         LLVMValueRef tmpsoa[4];
1890         LLVMValueRef tmps = tmpsrc[i];
1891         if (dst_type.length == 8) {
1892            LLVMValueRef shuffles[8];
1893            unsigned j;
1894            /* fetch was 4 values but need 8-wide output values */
1895            tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
1896            /*
1897             * for 8-wide aos transpose would give us wrong order not matching
1898             * incoming converted fs values and mask. ARGH.
1899             */
1900            for (j = 0; j < 4; j++) {
1901               shuffles[j] = lp_build_const_int32(gallivm, j * 2);
1902               shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
1903            }
1904            tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
1905                                          LLVMConstVector(shuffles, 8), "");
1906         }
1907         if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
1908            lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
1909         }
1910         else {
1911            lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
1912         }
1913         lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
1914      }
1915      return;
1916   }
1917
1918   lp_mem_type_from_format_desc(src_fmt, &mem_type);
1919   lp_blend_type_from_format_desc(src_fmt, &blend_type);
1920
1921   /* Is the format arithmetic */
1922   is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
1923   is_arith &= !(mem_type.width == 16 && mem_type.floating);
1924
1925   /* Pad if necessary */
1926   if (!is_arith && src_type.length < dst_type.length) {
1927      for (i = 0; i < num_srcs; ++i) {
1928         dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
1929      }
1930
1931      src_type.length = dst_type.length;
1932   }
1933
1934   /* Special case for half-floats */
1935   if (mem_type.width == 16 && mem_type.floating) {
1936      assert(blend_type.width == 32 && blend_type.floating);
1937      lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
1938      is_arith = false;
1939   }
1940
1941   if (!is_arith) {
1942      return;
1943   }
1944
1945   src_type.width = blend_type.width * blend_type.length;
1946   blend_type.length *= pixels;
1947   src_type.length *= pixels / (src_type.length / mem_type.length);
1948
1949   for (i = 0; i < num_srcs; ++i) {
1950      LLVMValueRef chans[4];
1951      LLVMValueRef res = NULL;
1952
1953      dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
1954
1955      for (j = 0; j < src_fmt->nr_channels; ++j) {
1956         unsigned mask = 0;
1957         unsigned sa = src_fmt->channel[j].shift;
1958#if UTIL_ARCH_LITTLE_ENDIAN
1959         unsigned from_lsb = j;
1960#else
1961         unsigned from_lsb = src_fmt->nr_channels - j - 1;
1962#endif
1963
1964         mask = (1 << src_fmt->channel[j].size) - 1;
1965
1966         /* Extract bits from source */
1967         chans[j] = LLVMBuildLShr(builder,
1968                                  dst[i],
1969                                  lp_build_const_int_vec(gallivm, src_type, sa),
1970                                  "");
1971
1972         chans[j] = LLVMBuildAnd(builder,
1973                                 chans[j],
1974                                 lp_build_const_int_vec(gallivm, src_type, mask),
1975                                 "");
1976
1977         /* Scale bits */
1978         if (src_type.norm) {
1979            chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
1980                                  blend_type.width, chans[j], src_type);
1981         }
1982
1983         /* Insert bits into correct position */
1984         chans[j] = LLVMBuildShl(builder,
1985                                 chans[j],
1986                                 lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
1987                                 "");
1988
1989         if (j == 0) {
1990            res = chans[j];
1991         } else {
1992            res = LLVMBuildOr(builder, res, chans[j], "");
1993         }
1994      }
1995
1996      dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
1997   }
1998}
1999
2000
2001/**
2002 * Convert from blending format to memory format
2003 *
2004 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
2005 */
2006static void
2007convert_from_blend_type(struct gallivm_state *gallivm,
2008                        unsigned block_size,
2009                        const struct util_format_description *src_fmt,
2010                        struct lp_type src_type,
2011                        struct lp_type dst_type,
2012                        LLVMValueRef* src, // and dst
2013                        unsigned num_srcs)
2014{
2015   LLVMValueRef* dst = src;
2016   unsigned i, j, k;
2017   struct lp_type mem_type;
2018   struct lp_type blend_type;
2019   LLVMBuilderRef builder = gallivm->builder;
2020   unsigned pixels = block_size / num_srcs;
2021   bool is_arith;
2022
2023   /*
2024    * full custom path for packed floats and srgb formats - none of the later
2025    * functions would do anything useful, and given the lp_type representation they
2026    * can't be fixed. Should really have some SoA blend path for these kind of
2027    * formats rather than hacking them in here.
2028    */
2029   if (format_expands_to_float_soa(src_fmt)) {
2030      /*
2031       * This is pretty suboptimal for this case blending in SoA would be much
2032       * better - we need to transpose the AoS values back to SoA values for
2033       * conversion/packing.
2034       */
2035      assert(src_type.floating);
2036      assert(src_type.width == 32);
2037      assert(src_type.length % 4 == 0);
2038      assert(dst_type.width == 32 || dst_type.width == 16);
2039
2040      for (i = 0; i < num_srcs / 4; i++) {
2041         LLVMValueRef tmpsoa[4], tmpdst;
2042         lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
2043         /* really really need SoA here */
2044
2045         if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
2046            tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
2047         }
2048         else {
2049            tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
2050                                                   src_type, tmpsoa);
2051         }
2052
2053         if (src_type.length == 8) {
2054            LLVMValueRef tmpaos, shuffles[8];
2055            unsigned j;
2056            /*
2057             * for 8-wide aos transpose has given us wrong order not matching
2058             * output order. HMPF. Also need to split the output values manually.
2059             */
2060            for (j = 0; j < 4; j++) {
2061               shuffles[j * 2] = lp_build_const_int32(gallivm, j);
2062               shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
2063            }
2064            tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
2065                                            LLVMConstVector(shuffles, 8), "");
2066            src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
2067            src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
2068         }
2069         else {
2070            src[i] = tmpdst;
2071         }
2072      }
2073      if (dst_type.width == 16) {
2074         struct lp_type type16x8 = dst_type;
2075         struct lp_type type32x4 = dst_type;
2076         LLVMTypeRef ltype16x4, ltypei64, ltypei128;
2077         unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
2078         type16x8.length = 8;
2079         type32x4.width = 32;
2080         ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
2081         ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
2082         ltype16x4 = lp_build_vec_type(gallivm, dst_type);
2083         /* We could do vector truncation but it doesn't generate very good code */
2084         for (i = 0; i < num_fetch; i++) {
2085            src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
2086                                    src[i], lp_build_zero(gallivm, type32x4));
2087            src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
2088            src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
2089            src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
2090         }
2091      }
2092      return;
2093   }
2094
2095   lp_mem_type_from_format_desc(src_fmt, &mem_type);
2096   lp_blend_type_from_format_desc(src_fmt, &blend_type);
2097
2098   is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
2099
2100   /* Special case for half-floats */
2101   if (mem_type.width == 16 && mem_type.floating) {
2102      int length = dst_type.length;
2103      assert(blend_type.width == 32 && blend_type.floating);
2104
2105      dst_type.length = src_type.length;
2106
2107      lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
2108
2109      dst_type.length = length;
2110      is_arith = false;
2111   }
2112
2113   /* Remove any padding */
2114   if (!is_arith && (src_type.length % mem_type.length)) {
2115      src_type.length -= (src_type.length % mem_type.length);
2116
2117      for (i = 0; i < num_srcs; ++i) {
2118         dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
2119      }
2120   }
2121
2122   /* No bit arithmetic to do */
2123   if (!is_arith) {
2124      return;
2125   }
2126
2127   src_type.length = pixels;
2128   src_type.width = blend_type.length * blend_type.width;
2129   dst_type.length = pixels;
2130
2131   for (i = 0; i < num_srcs; ++i) {
2132      LLVMValueRef chans[4];
2133      LLVMValueRef res = NULL;
2134
2135      dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
2136
2137      for (j = 0; j < src_fmt->nr_channels; ++j) {
2138         unsigned mask = 0;
2139         unsigned sa = src_fmt->channel[j].shift;
2140         unsigned sz_a = src_fmt->channel[j].size;
2141#if UTIL_ARCH_LITTLE_ENDIAN
2142         unsigned from_lsb = j;
2143#else
2144         unsigned from_lsb = src_fmt->nr_channels - j - 1;
2145#endif
2146
2147         assert(blend_type.width > src_fmt->channel[j].size);
2148
2149         for (k = 0; k < blend_type.width; ++k) {
2150            mask |= 1 << k;
2151         }
2152
2153         /* Extract bits */
2154         chans[j] = LLVMBuildLShr(builder,
2155                                  dst[i],
2156                                  lp_build_const_int_vec(gallivm, src_type,
2157                                                         from_lsb * blend_type.width),
2158                                  "");
2159
2160         chans[j] = LLVMBuildAnd(builder,
2161                                 chans[j],
2162                                 lp_build_const_int_vec(gallivm, src_type, mask),
2163                                 "");
2164
2165         /* Scale down bits */
2166         if (src_type.norm) {
2167            chans[j] = scale_bits(gallivm, blend_type.width,
2168                                  src_fmt->channel[j].size, chans[j], src_type);
2169         } else if (!src_type.floating && sz_a < blend_type.width) {
2170            LLVMValueRef mask_val = lp_build_const_int_vec(gallivm, src_type, (1UL << sz_a) - 1);
2171            LLVMValueRef mask = LLVMBuildICmp(builder, LLVMIntUGT, chans[j], mask_val, "");
2172            chans[j] = LLVMBuildSelect(builder, mask, mask_val, chans[j], "");
2173         }
2174
2175         /* Insert bits */
2176         chans[j] = LLVMBuildShl(builder,
2177                                 chans[j],
2178                                 lp_build_const_int_vec(gallivm, src_type, sa),
2179                                 "");
2180
2181         sa += src_fmt->channel[j].size;
2182
2183         if (j == 0) {
2184            res = chans[j];
2185         } else {
2186            res = LLVMBuildOr(builder, res, chans[j], "");
2187         }
2188      }
2189
2190      assert (dst_type.width != 24);
2191
2192      dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
2193   }
2194}
2195
2196
2197/**
2198 * Convert alpha to same blend type as src
2199 */
2200static void
2201convert_alpha(struct gallivm_state *gallivm,
2202              struct lp_type row_type,
2203              struct lp_type alpha_type,
2204              const unsigned block_size,
2205              const unsigned block_height,
2206              const unsigned src_count,
2207              const unsigned dst_channels,
2208              const bool pad_inline,
2209              LLVMValueRef* src_alpha)
2210{
2211   LLVMBuilderRef builder = gallivm->builder;
2212   unsigned i, j;
2213   unsigned length = row_type.length;
2214   row_type.length = alpha_type.length;
2215
2216   /* Twiddle the alpha to match pixels */
2217   lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
2218
2219   /*
2220    * TODO this should use single lp_build_conv call for
2221    * src_count == 1 && dst_channels == 1 case (dropping the concat below)
2222    */
2223   for (i = 0; i < block_height; ++i) {
2224      lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
2225   }
2226
2227   alpha_type = row_type;
2228   row_type.length = length;
2229
2230   /* If only one channel we can only need the single alpha value per pixel */
2231   if (src_count == 1 && dst_channels == 1) {
2232
2233      lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
2234   } else {
2235      /* If there are more srcs than rows then we need to split alpha up */
2236      if (src_count > block_height) {
2237         for (i = src_count; i > 0; --i) {
2238            unsigned pixels = block_size / src_count;
2239            unsigned idx = i - 1;
2240
2241            src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
2242                                                    (idx * pixels) % 4, pixels);
2243         }
2244      }
2245
2246      /* If there is a src for each pixel broadcast the alpha across whole row */
2247      if (src_count == block_size) {
2248         for (i = 0; i < src_count; ++i) {
2249            src_alpha[i] = lp_build_broadcast(gallivm,
2250                              lp_build_vec_type(gallivm, row_type), src_alpha[i]);
2251         }
2252      } else {
2253         unsigned pixels = block_size / src_count;
2254         unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
2255         unsigned alpha_span = 1;
2256         LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
2257
2258         /* Check if we need 2 src_alphas for our shuffles */
2259         if (pixels > alpha_type.length) {
2260            alpha_span = 2;
2261         }
2262
2263         /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
2264         for (j = 0; j < row_type.length; ++j) {
2265            if (j < pixels * channels) {
2266               shuffles[j] = lp_build_const_int32(gallivm, j / channels);
2267            } else {
2268               shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
2269            }
2270         }
2271
2272         for (i = 0; i < src_count; ++i) {
2273            unsigned idx1 = i, idx2 = i;
2274
2275            if (alpha_span > 1){
2276               idx1 *= alpha_span;
2277               idx2 = idx1 + 1;
2278            }
2279
2280            src_alpha[i] = LLVMBuildShuffleVector(builder,
2281                                                  src_alpha[idx1],
2282                                                  src_alpha[idx2],
2283                                                  LLVMConstVector(shuffles, row_type.length),
2284                                                  "");
2285         }
2286      }
2287   }
2288}
2289
2290
2291/**
2292 * Generates the blend function for unswizzled colour buffers
2293 * Also generates the read & write from colour buffer
2294 */
2295static void
2296generate_unswizzled_blend(struct gallivm_state *gallivm,
2297                          unsigned rt,
2298                          struct lp_fragment_shader_variant *variant,
2299                          enum pipe_format out_format,
2300                          unsigned int num_fs,
2301                          struct lp_type fs_type,
2302                          LLVMValueRef* fs_mask,
2303                          LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
2304                          LLVMValueRef context_ptr,
2305                          LLVMValueRef color_ptr,
2306                          LLVMValueRef stride,
2307                          unsigned partial_mask,
2308                          boolean do_branch)
2309{
2310   const unsigned alpha_channel = 3;
2311   const unsigned block_width = LP_RASTER_BLOCK_SIZE;
2312   const unsigned block_height = LP_RASTER_BLOCK_SIZE;
2313   const unsigned block_size = block_width * block_height;
2314   const unsigned lp_integer_vector_width = 128;
2315
2316   LLVMBuilderRef builder = gallivm->builder;
2317   LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
2318   LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
2319   LLVMValueRef src_alpha[4 * 4];
2320   LLVMValueRef src1_alpha[4 * 4] = { NULL };
2321   LLVMValueRef src_mask[4 * 4];
2322   LLVMValueRef src[4 * 4];
2323   LLVMValueRef src1[4 * 4];
2324   LLVMValueRef dst[4 * 4];
2325   LLVMValueRef blend_color;
2326   LLVMValueRef blend_alpha;
2327   LLVMValueRef i32_zero;
2328   LLVMValueRef check_mask;
2329   LLVMValueRef undef_src_val;
2330
2331   struct lp_build_mask_context mask_ctx;
2332   struct lp_type mask_type;
2333   struct lp_type blend_type;
2334   struct lp_type row_type;
2335   struct lp_type dst_type;
2336   struct lp_type ls_type;
2337
2338   unsigned char swizzle[TGSI_NUM_CHANNELS];
2339   unsigned vector_width;
2340   unsigned src_channels = TGSI_NUM_CHANNELS;
2341   unsigned dst_channels;
2342   unsigned dst_count;
2343   unsigned src_count;
2344   unsigned i, j;
2345
2346   const struct util_format_description* out_format_desc = util_format_description(out_format);
2347
2348   unsigned dst_alignment;
2349
2350   bool pad_inline = is_arithmetic_format(out_format_desc);
2351   bool has_alpha = false;
2352   const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
2353                                     util_blend_state_is_dual(&variant->key.blend, 0);
2354
2355   const boolean is_1d = variant->key.resource_1d;
2356   boolean twiddle_after_convert = FALSE;
2357   unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
2358   LLVMValueRef fpstate = 0;
2359
2360   /* Get type from output format */
2361   lp_blend_type_from_format_desc(out_format_desc, &row_type);
2362   lp_mem_type_from_format_desc(out_format_desc, &dst_type);
2363
2364   /*
2365    * Technically this code should go into lp_build_smallfloat_to_float
2366    * and lp_build_float_to_smallfloat but due to the
2367    * http://llvm.org/bugs/show_bug.cgi?id=6393
2368    * llvm reorders the mxcsr intrinsics in a way that breaks the code.
2369    * So the ordering is important here and there shouldn't be any
2370    * llvm ir instrunctions in this function before
2371    * this, otherwise half-float format conversions won't work
2372    * (again due to llvm bug #6393).
2373    */
2374   if (have_smallfloat_format(dst_type, out_format)) {
2375      /* We need to make sure that denorms are ok for half float
2376         conversions */
2377      fpstate = lp_build_fpstate_get(gallivm);
2378      lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
2379   }
2380
2381   mask_type = lp_int32_vec4_type();
2382   mask_type.length = fs_type.length;
2383
2384   for (i = num_fs; i < num_fullblock_fs; i++) {
2385      fs_mask[i] = lp_build_zero(gallivm, mask_type);
2386   }
2387
2388   /* Do not bother executing code when mask is empty.. */
2389   if (do_branch) {
2390      check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
2391
2392      for (i = 0; i < num_fullblock_fs; ++i) {
2393         check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
2394      }
2395
2396      lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
2397      lp_build_mask_check(&mask_ctx);
2398   }
2399
2400   partial_mask |= !variant->opaque;
2401   i32_zero = lp_build_const_int32(gallivm, 0);
2402
2403   undef_src_val = lp_build_undef(gallivm, fs_type);
2404
2405   row_type.length = fs_type.length;
2406   vector_width    = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
2407
2408   /* Compute correct swizzle and count channels */
2409   memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
2410   dst_channels = 0;
2411
2412   for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
2413      /* Ensure channel is used */
2414      if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
2415         continue;
2416      }
2417
2418      /* Ensure not already written to (happens in case with GL_ALPHA) */
2419      if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
2420         continue;
2421      }
2422
2423      /* Ensure we haven't already found all channels */
2424      if (dst_channels >= out_format_desc->nr_channels) {
2425         continue;
2426      }
2427
2428      swizzle[out_format_desc->swizzle[i]] = i;
2429      ++dst_channels;
2430
2431      if (i == alpha_channel) {
2432         has_alpha = true;
2433      }
2434   }
2435
2436   if (format_expands_to_float_soa(out_format_desc)) {
2437      /*
2438       * the code above can't work for layout_other
2439       * for srgb it would sort of work but we short-circuit swizzles, etc.
2440       * as that is done as part of unpack / pack.
2441       */
2442      dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
2443      has_alpha = true;
2444      swizzle[0] = 0;
2445      swizzle[1] = 1;
2446      swizzle[2] = 2;
2447      swizzle[3] = 3;
2448      pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
2449   }
2450
2451   /* If 3 channels then pad to include alpha for 4 element transpose */
2452   if (dst_channels == 3) {
2453      assert (!has_alpha);
2454      for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
2455         if (swizzle[i] > TGSI_NUM_CHANNELS)
2456            swizzle[i] = 3;
2457      }
2458      if (out_format_desc->nr_channels == 4) {
2459         dst_channels = 4;
2460         /*
2461          * We use alpha from the color conversion, not separate one.
2462          * We had to include it for transpose, hence it will get converted
2463          * too (albeit when doing transpose after conversion, that would
2464          * no longer be the case necessarily).
2465          * (It works only with 4 channel dsts, e.g. rgbx formats, because
2466          * otherwise we really have padding, not alpha, included.)
2467          */
2468         has_alpha = true;
2469      }
2470   }
2471
2472   /*
2473    * Load shader output
2474    */
2475   for (i = 0; i < num_fullblock_fs; ++i) {
2476      /* Always load alpha for use in blending */
2477      LLVMValueRef alpha;
2478      if (i < num_fs) {
2479         alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
2480      }
2481      else {
2482         alpha = undef_src_val;
2483      }
2484
2485      /* Load each channel */
2486      for (j = 0; j < dst_channels; ++j) {
2487         assert(swizzle[j] < 4);
2488         if (i < num_fs) {
2489            fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
2490         }
2491         else {
2492            fs_src[i][j] = undef_src_val;
2493         }
2494      }
2495
2496      /* If 3 channels then pad to include alpha for 4 element transpose */
2497      /*
2498       * XXX If we include that here maybe could actually use it instead of
2499       * separate alpha for blending?
2500       * (Difficult though we actually convert pad channels, not alpha.)
2501       */
2502      if (dst_channels == 3 && !has_alpha) {
2503         fs_src[i][3] = alpha;
2504      }
2505
2506      /* We split the row_mask and row_alpha as we want 128bit interleave */
2507      if (fs_type.length == 8) {
2508         src_mask[i*2 + 0]  = lp_build_extract_range(gallivm, fs_mask[i],
2509                                                     0, src_channels);
2510         src_mask[i*2 + 1]  = lp_build_extract_range(gallivm, fs_mask[i],
2511                                                     src_channels, src_channels);
2512
2513         src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2514         src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2515                                                     src_channels, src_channels);
2516      } else {
2517         src_mask[i] = fs_mask[i];
2518         src_alpha[i] = alpha;
2519      }
2520   }
2521   if (dual_source_blend) {
2522      /* same as above except different src/dst, skip masks and comments... */
2523      for (i = 0; i < num_fullblock_fs; ++i) {
2524         LLVMValueRef alpha;
2525         if (i < num_fs) {
2526            alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
2527         }
2528         else {
2529            alpha = undef_src_val;
2530         }
2531
2532         for (j = 0; j < dst_channels; ++j) {
2533            assert(swizzle[j] < 4);
2534            if (i < num_fs) {
2535               fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
2536            }
2537            else {
2538               fs_src1[i][j] = undef_src_val;
2539            }
2540         }
2541         if (dst_channels == 3 && !has_alpha) {
2542            fs_src1[i][3] = alpha;
2543         }
2544         if (fs_type.length == 8) {
2545            src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
2546            src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,
2547                                                         src_channels, src_channels);
2548         } else {
2549            src1_alpha[i] = alpha;
2550         }
2551      }
2552   }
2553
2554   if (util_format_is_pure_integer(out_format)) {
2555      /*
2556       * In this case fs_type was really ints or uints disguised as floats,
2557       * fix that up now.
2558       */
2559      fs_type.floating = 0;
2560      fs_type.sign = dst_type.sign;
2561      for (i = 0; i < num_fullblock_fs; ++i) {
2562         for (j = 0; j < dst_channels; ++j) {
2563            fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
2564                                            lp_build_vec_type(gallivm, fs_type), "");
2565         }
2566         if (dst_channels == 3 && !has_alpha) {
2567            fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
2568                                            lp_build_vec_type(gallivm, fs_type), "");
2569         }
2570      }
2571   }
2572
2573   /*
2574    * We actually should generally do conversion first (for non-1d cases)
2575    * when the blend format is 8 or 16 bits. The reason is obvious,
2576    * there's 2 or 4 times less vectors to deal with for the interleave...
2577    * Albeit for the AVX (not AVX2) case there's no benefit with 16 bit
2578    * vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit
2579    * unpack only with 128bit vectors).
2580    * Note: for 16bit sizes really need matching pack conversion code
2581    */
2582   if (!is_1d && dst_channels != 3 && dst_type.width == 8) {
2583      twiddle_after_convert = TRUE;
2584   }
2585
2586   /*
2587    * Pixel twiddle from fragment shader order to memory order
2588    */
2589   if (!twiddle_after_convert) {
2590      src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
2591                                      dst_channels, fs_src, src, pad_inline);
2592      if (dual_source_blend) {
2593         generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
2594                             fs_src1, src1, pad_inline);
2595      }
2596   } else {
2597      src_count = num_fullblock_fs * dst_channels;
2598      /*
2599       * We reorder things a bit here, so the cases for 4-wide and 8-wide
2600       * (AVX) turn out the same later when untwiddling/transpose (albeit
2601       * for true AVX2 path untwiddle needs to be different).
2602       * For now just order by colors first (so we can use unpack later).
2603       */
2604      for (j = 0; j < num_fullblock_fs; j++) {
2605         for (i = 0; i < dst_channels; i++) {
2606            src[i*num_fullblock_fs + j] = fs_src[j][i];
2607            if (dual_source_blend) {
2608               src1[i*num_fullblock_fs + j] = fs_src1[j][i];
2609            }
2610         }
2611      }
2612   }
2613
2614   src_channels = dst_channels < 3 ? dst_channels : 4;
2615   if (src_count != num_fullblock_fs * src_channels) {
2616      unsigned ds = src_count / (num_fullblock_fs * src_channels);
2617      row_type.length /= ds;
2618      fs_type.length = row_type.length;
2619   }
2620
2621   blend_type = row_type;
2622   mask_type.length = 4;
2623
2624   /* Convert src to row_type */
2625   if (dual_source_blend) {
2626      struct lp_type old_row_type = row_type;
2627      lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2628      src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
2629   }
2630   else {
2631      src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
2632   }
2633
2634   /* If the rows are not an SSE vector, combine them to become SSE size! */
2635   if ((row_type.width * row_type.length) % 128) {
2636      unsigned bits = row_type.width * row_type.length;
2637      unsigned combined;
2638
2639      assert(src_count >= (vector_width / bits));
2640
2641      dst_count = src_count / (vector_width / bits);
2642
2643      combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
2644      if (dual_source_blend) {
2645         lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
2646      }
2647
2648      row_type.length *= combined;
2649      src_count /= combined;
2650
2651      bits = row_type.width * row_type.length;
2652      assert(bits == 128 || bits == 256);
2653   }
2654
2655   if (twiddle_after_convert) {
2656      fs_twiddle_transpose(gallivm, row_type, src, src_count, src);
2657      if (dual_source_blend) {
2658         fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);
2659      }
2660   }
2661
2662   /*
2663    * Blend Colour conversion
2664    */
2665   blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
2666   blend_color = LLVMBuildPointerCast(builder, blend_color,
2667                    LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
2668   blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,
2669                               &i32_zero, 1, ""), "");
2670
2671   /* Convert */
2672   lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
2673
2674   if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
2675      /*
2676       * since blending is done with floats, there was no conversion.
2677       * However, the rules according to fixed point renderbuffers still
2678       * apply, that is we must clamp inputs to 0.0/1.0.
2679       * (This would apply to separate alpha conversion too but we currently
2680       * force has_alpha to be true.)
2681       * TODO: should skip this with "fake" blend, since post-blend conversion
2682       * will clamp anyway.
2683       * TODO: could also skip this if fragment color clamping is enabled. We
2684       * don't support it natively so it gets baked into the shader however, so
2685       * can't really tell here.
2686       */
2687      struct lp_build_context f32_bld;
2688      assert(row_type.floating);
2689      lp_build_context_init(&f32_bld, gallivm, row_type);
2690      for (i = 0; i < src_count; i++) {
2691         src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
2692      }
2693      if (dual_source_blend) {
2694         for (i = 0; i < src_count; i++) {
2695            src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
2696         }
2697      }
2698      /* probably can't be different than row_type but better safe than sorry... */
2699      lp_build_context_init(&f32_bld, gallivm, blend_type);
2700      blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
2701   }
2702
2703   /* Extract alpha */
2704   blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
2705
2706   /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
2707   pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
2708   if (pad_inline) {
2709      /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
2710      blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
2711   } else {
2712      /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
2713      blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
2714   }
2715
2716   /*
2717    * Mask conversion
2718    */
2719   lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
2720
2721   if (src_count < block_height) {
2722      lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
2723   } else if (src_count > block_height) {
2724      for (i = src_count; i > 0; --i) {
2725         unsigned pixels = block_size / src_count;
2726         unsigned idx = i - 1;
2727
2728         src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
2729                                                (idx * pixels) % 4, pixels);
2730      }
2731   }
2732
2733   assert(mask_type.width == 32);
2734
2735   for (i = 0; i < src_count; ++i) {
2736      unsigned pixels = block_size / src_count;
2737      unsigned pixel_width = row_type.width * dst_channels;
2738
2739      if (pixel_width == 24) {
2740         mask_type.width = 8;
2741         mask_type.length = vector_width / mask_type.width;
2742      } else {
2743         mask_type.length = pixels;
2744         mask_type.width = row_type.width * dst_channels;
2745
2746         /*
2747          * If mask_type width is smaller than 32bit, this doesn't quite
2748          * generate the most efficient code (could use some pack).
2749          */
2750         src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],
2751                                        lp_build_int_vec_type(gallivm, mask_type), "");
2752
2753         mask_type.length *= dst_channels;
2754         mask_type.width /= dst_channels;
2755      }
2756
2757      src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],
2758                                     lp_build_int_vec_type(gallivm, mask_type), "");
2759      src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
2760   }
2761
2762   /*
2763    * Alpha conversion
2764    */
2765   if (!has_alpha) {
2766      struct lp_type alpha_type = fs_type;
2767      alpha_type.length = 4;
2768      convert_alpha(gallivm, row_type, alpha_type,
2769                    block_size, block_height,
2770                    src_count, dst_channels,
2771                    pad_inline, src_alpha);
2772      if (dual_source_blend) {
2773         convert_alpha(gallivm, row_type, alpha_type,
2774                       block_size, block_height,
2775                       src_count, dst_channels,
2776                       pad_inline, src1_alpha);
2777      }
2778   }
2779
2780
2781   /*
2782    * Load dst from memory
2783    */
2784   if (src_count < block_height) {
2785      dst_count = block_height;
2786   } else {
2787      dst_count = src_count;
2788   }
2789
2790   dst_type.length *= block_size / dst_count;
2791
2792   if (format_expands_to_float_soa(out_format_desc)) {
2793      /*
2794       * we need multiple values at once for the conversion, so can as well
2795       * load them vectorized here too instead of concatenating later.
2796       * (Still need concatenation later for 8-wide vectors).
2797       */
2798      dst_count = block_height;
2799      dst_type.length = block_width;
2800   }
2801
2802   /*
2803    * Compute the alignment of the destination pointer in bytes
2804    * We fetch 1-4 pixels, if the format has pot alignment then those fetches
2805    * are always aligned by MIN2(16, fetch_width) except for buffers (not
2806    * 1d tex but can't distinguish here) so need to stick with per-pixel
2807    * alignment in this case.
2808    */
2809   if (is_1d) {
2810      dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
2811   }
2812   else {
2813      dst_alignment = dst_type.length * dst_type.width / 8;
2814   }
2815   /* Force power-of-two alignment by extracting only the least-significant-bit */
2816   dst_alignment = 1 << (ffs(dst_alignment) - 1);
2817   /*
2818    * Resource base and stride pointers are aligned to 16 bytes, so that's
2819    * the maximum alignment we can guarantee
2820    */
2821   dst_alignment = MIN2(16, dst_alignment);
2822
2823   ls_type = dst_type;
2824
2825   if (dst_count > src_count) {
2826      if ((dst_type.width == 8 || dst_type.width == 16) &&
2827          util_is_power_of_two_or_zero(dst_type.length) &&
2828          dst_type.length * dst_type.width < 128) {
2829         /*
2830          * Never try to load values as 4xi8 which we will then
2831          * concatenate to larger vectors. This gives llvm a real
2832          * headache (the problem is the type legalizer (?) will
2833          * try to load that as 4xi8 zext to 4xi32 to fill the vector,
2834          * then the shuffles to concatenate are more or less impossible
2835          * - llvm is easily capable of generating a sequence of 32
2836          * pextrb/pinsrb instructions for that. Albeit it appears to
2837          * be fixed in llvm 4.0. So, load and concatenate with 32bit
2838          * width to avoid the trouble (16bit seems not as bad, llvm
2839          * probably recognizes the load+shuffle as only one shuffle
2840          * is necessary, but we can do just the same anyway).
2841          */
2842         ls_type.length = dst_type.length * dst_type.width / 32;
2843         ls_type.width = 32;
2844      }
2845   }
2846
2847   if (is_1d) {
2848      load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2849                            dst, ls_type, dst_count / 4, dst_alignment, NULL, NULL, false);
2850      for (i = dst_count / 4; i < dst_count; i++) {
2851         dst[i] = lp_build_undef(gallivm, ls_type);
2852      }
2853
2854   }
2855   else {
2856      load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2857                            dst, ls_type, dst_count, dst_alignment, NULL, NULL, false);
2858   }
2859
2860
2861   /*
2862    * Convert from dst/output format to src/blending format.
2863    *
2864    * This is necessary as we can only read 1 row from memory at a time,
2865    * so the minimum dst_count will ever be at this point is 4.
2866    *
2867    * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
2868    * this will take the 4 dsts and combine them into 1 src so we can perform blending
2869    * on all 16 pixels in that single vector at once.
2870    */
2871   if (dst_count > src_count) {
2872      if (ls_type.length != dst_type.length && ls_type.length == 1) {
2873         LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);
2874         LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);
2875         for (i = 0; i < dst_count; i++) {
2876            dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");
2877         }
2878      }
2879
2880      lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);
2881
2882      if (ls_type.length != dst_type.length) {
2883         struct lp_type tmp_type = dst_type;
2884         tmp_type.length = dst_type.length * 4 / src_count;
2885         for (i = 0; i < src_count; i++) {
2886            dst[i] = LLVMBuildBitCast(builder, dst[i],
2887                                      lp_build_vec_type(gallivm, tmp_type), "");
2888         }
2889      }
2890   }
2891
2892   /*
2893    * Blending
2894    */
2895   /* XXX this is broken for RGB8 formats -
2896    * they get expanded from 12 to 16 elements (to include alpha)
2897    * by convert_to_blend_type then reduced to 15 instead of 12
2898    * by convert_from_blend_type (a simple fix though breaks A8...).
2899    * R16G16B16 also crashes differently however something going wrong
2900    * inside llvm handling npot vector sizes seemingly.
2901    * It seems some cleanup could be done here (like skipping conversion/blend
2902    * when not needed).
2903    */
2904   convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,
2905                         row_type, dst, src_count);
2906
2907   /*
2908    * FIXME: Really should get logic ops / masks out of generic blend / row
2909    * format. Logic ops will definitely not work on the blend float format
2910    * used for SRGB here and I think OpenGL expects this to work as expected
2911    * (that is incoming values converted to srgb then logic op applied).
2912    */
2913   for (i = 0; i < src_count; ++i) {
2914      dst[i] = lp_build_blend_aos(gallivm,
2915                                  &variant->key.blend,
2916                                  out_format,
2917                                  row_type,
2918                                  rt,
2919                                  src[i],
2920                                  has_alpha ? NULL : src_alpha[i],
2921                                  src1[i],
2922                                  has_alpha ? NULL : src1_alpha[i],
2923                                  dst[i],
2924                                  partial_mask ? src_mask[i] : NULL,
2925                                  blend_color,
2926                                  has_alpha ? NULL : blend_alpha,
2927                                  swizzle,
2928                                  pad_inline ? 4 : dst_channels);
2929   }
2930
2931   convert_from_blend_type(gallivm, block_size, out_format_desc,
2932                           row_type, dst_type, dst, src_count);
2933
2934   /* Split the blend rows back to memory rows */
2935   if (dst_count > src_count) {
2936      row_type.length = dst_type.length * (dst_count / src_count);
2937
2938      if (src_count == 1) {
2939         dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2940         dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2941
2942         row_type.length /= 2;
2943         src_count *= 2;
2944      }
2945
2946      dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
2947      dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
2948      dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
2949      dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
2950
2951      row_type.length /= 2;
2952      src_count *= 2;
2953   }
2954
2955   /*
2956    * Store blend result to memory
2957    */
2958   if (is_1d) {
2959      store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
2960                             dst, dst_type, dst_count / 4, dst_alignment);
2961   }
2962   else {
2963      store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
2964                             dst, dst_type, dst_count, dst_alignment);
2965   }
2966
2967   if (have_smallfloat_format(dst_type, out_format)) {
2968      lp_build_fpstate_set(gallivm, fpstate);
2969   }
2970
2971   if (do_branch) {
2972      lp_build_mask_end(&mask_ctx);
2973   }
2974}
2975
2976
2977/**
2978 * Generate the runtime callable function for the whole fragment pipeline.
2979 * Note that the function which we generate operates on a block of 16
2980 * pixels at at time.  The block contains 2x2 quads.  Each quad contains
2981 * 2x2 pixels.
2982 */
2983static void
2984generate_fragment(struct llvmpipe_context *lp,
2985                  struct lp_fragment_shader *shader,
2986                  struct lp_fragment_shader_variant *variant,
2987                  unsigned partial_mask)
2988{
2989   struct gallivm_state *gallivm = variant->gallivm;
2990   struct lp_fragment_shader_variant_key *key = &variant->key;
2991   struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
2992   char func_name[64];
2993   struct lp_type fs_type;
2994   struct lp_type blend_type;
2995   LLVMTypeRef fs_elem_type;
2996   LLVMTypeRef blend_vec_type;
2997   LLVMTypeRef arg_types[15];
2998   LLVMTypeRef func_type;
2999   LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
3000   LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
3001   LLVMValueRef context_ptr;
3002   LLVMValueRef x;
3003   LLVMValueRef y;
3004   LLVMValueRef a0_ptr;
3005   LLVMValueRef dadx_ptr;
3006   LLVMValueRef dady_ptr;
3007   LLVMValueRef color_ptr_ptr;
3008   LLVMValueRef stride_ptr;
3009   LLVMValueRef color_sample_stride_ptr;
3010   LLVMValueRef depth_ptr;
3011   LLVMValueRef depth_stride;
3012   LLVMValueRef depth_sample_stride;
3013   LLVMValueRef mask_input;
3014   LLVMValueRef thread_data_ptr;
3015   LLVMBasicBlockRef block;
3016   LLVMBuilderRef builder;
3017   struct lp_build_sampler_soa *sampler;
3018   struct lp_build_image_soa *image;
3019   struct lp_build_interp_soa_context interp;
3020   LLVMValueRef fs_mask[(16 / 4) * LP_MAX_SAMPLES];
3021   LLVMValueRef fs_out_color[LP_MAX_SAMPLES][PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
3022   LLVMValueRef function;
3023   LLVMValueRef facing;
3024   unsigned num_fs;
3025   unsigned i;
3026   unsigned chan;
3027   unsigned cbuf;
3028   boolean cbuf0_write_all;
3029   const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
3030                                     util_blend_state_is_dual(&key->blend, 0);
3031
3032   assert(lp_native_vector_width / 32 >= 4);
3033
3034   /* Adjust color input interpolation according to flatshade state:
3035    */
3036   memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
3037   for (i = 0; i < shader->info.base.num_inputs; i++) {
3038      if (inputs[i].interp == LP_INTERP_COLOR) {
3039	 if (key->flatshade)
3040	    inputs[i].interp = LP_INTERP_CONSTANT;
3041	 else
3042	    inputs[i].interp = LP_INTERP_PERSPECTIVE;
3043      }
3044   }
3045
3046   /* check if writes to cbuf[0] are to be copied to all cbufs */
3047   cbuf0_write_all =
3048     shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
3049
3050   /* TODO: actually pick these based on the fs and color buffer
3051    * characteristics. */
3052
3053   memset(&fs_type, 0, sizeof fs_type);
3054   fs_type.floating = TRUE;      /* floating point values */
3055   fs_type.sign = TRUE;          /* values are signed */
3056   fs_type.norm = FALSE;         /* values are not limited to [0,1] or [-1,1] */
3057   fs_type.width = 32;           /* 32-bit float */
3058   fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
3059
3060   memset(&blend_type, 0, sizeof blend_type);
3061   blend_type.floating = FALSE; /* values are integers */
3062   blend_type.sign = FALSE;     /* values are unsigned */
3063   blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
3064   blend_type.width = 8;        /* 8-bit ubyte values */
3065   blend_type.length = 16;      /* 16 elements per vector */
3066
3067   /*
3068    * Generate the function prototype. Any change here must be reflected in
3069    * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
3070    */
3071
3072   fs_elem_type = lp_build_elem_type(gallivm, fs_type);
3073
3074   blend_vec_type = lp_build_vec_type(gallivm, blend_type);
3075
3076   snprintf(func_name, sizeof(func_name), "fs_variant_%s",
3077            partial_mask ? "partial" : "whole");
3078
3079   arg_types[0] = variant->jit_context_ptr_type;       /* context */
3080   arg_types[1] = int32_type;                          /* x */
3081   arg_types[2] = int32_type;                          /* y */
3082   arg_types[3] = int32_type;                          /* facing */
3083   arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
3084   arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
3085   arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
3086   arg_types[7] = LLVMPointerType(LLVMPointerType(int8_type, 0), 0);  /* color */
3087   arg_types[8] = LLVMPointerType(int8_type, 0);       /* depth */
3088   arg_types[9] = LLVMInt64TypeInContext(gallivm->context);  /* mask_input */
3089   arg_types[10] = variant->jit_thread_data_ptr_type;  /* per thread data */
3090   arg_types[11] = LLVMPointerType(int32_type, 0);     /* stride */
3091   arg_types[12] = int32_type;                         /* depth_stride */
3092   arg_types[13] = LLVMPointerType(int32_type, 0);     /* color sample strides */
3093   arg_types[14] = int32_type;                         /* depth sample stride */
3094
3095   func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
3096                                arg_types, ARRAY_SIZE(arg_types), 0);
3097
3098   function = LLVMAddFunction(gallivm->module, func_name, func_type);
3099   LLVMSetFunctionCallConv(function, LLVMCCallConv);
3100
3101   variant->function[partial_mask] = function;
3102
3103   /* XXX: need to propagate noalias down into color param now we are
3104    * passing a pointer-to-pointer?
3105    */
3106   for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
3107      if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
3108         lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
3109
3110   if (variant->gallivm->cache->data_size)
3111      return;
3112
3113   context_ptr  = LLVMGetParam(function, 0);
3114   x            = LLVMGetParam(function, 1);
3115   y            = LLVMGetParam(function, 2);
3116   facing       = LLVMGetParam(function, 3);
3117   a0_ptr       = LLVMGetParam(function, 4);
3118   dadx_ptr     = LLVMGetParam(function, 5);
3119   dady_ptr     = LLVMGetParam(function, 6);
3120   color_ptr_ptr = LLVMGetParam(function, 7);
3121   depth_ptr    = LLVMGetParam(function, 8);
3122   mask_input   = LLVMGetParam(function, 9);
3123   thread_data_ptr  = LLVMGetParam(function, 10);
3124   stride_ptr   = LLVMGetParam(function, 11);
3125   depth_stride = LLVMGetParam(function, 12);
3126   color_sample_stride_ptr = LLVMGetParam(function, 13);
3127   depth_sample_stride = LLVMGetParam(function, 14);
3128
3129   lp_build_name(context_ptr, "context");
3130   lp_build_name(x, "x");
3131   lp_build_name(y, "y");
3132   lp_build_name(a0_ptr, "a0");
3133   lp_build_name(dadx_ptr, "dadx");
3134   lp_build_name(dady_ptr, "dady");
3135   lp_build_name(color_ptr_ptr, "color_ptr_ptr");
3136   lp_build_name(depth_ptr, "depth");
3137   lp_build_name(mask_input, "mask_input");
3138   lp_build_name(thread_data_ptr, "thread_data");
3139   lp_build_name(stride_ptr, "stride_ptr");
3140   lp_build_name(depth_stride, "depth_stride");
3141   lp_build_name(color_sample_stride_ptr, "color_sample_stride_ptr");
3142   lp_build_name(depth_sample_stride, "depth_sample_stride");
3143
3144   /*
3145    * Function body
3146    */
3147
3148   block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
3149   builder = gallivm->builder;
3150   assert(builder);
3151   LLVMPositionBuilderAtEnd(builder, block);
3152
3153   /*
3154    * Must not count ps invocations if there's a null shader.
3155    * (It would be ok to count with null shader if there's d/s tests,
3156    * but only if there's d/s buffers too, which is different
3157    * to implicit rasterization disable which must not depend
3158    * on the d/s buffers.)
3159    * Could use popcount on mask, but pixel accuracy is not required.
3160    * Could disable if there's no stats query, but maybe not worth it.
3161    */
3162   if (shader->info.base.num_instructions > 1) {
3163      LLVMValueRef invocs, val;
3164      invocs = lp_jit_thread_data_invocations(gallivm, thread_data_ptr);
3165      val = LLVMBuildLoad(builder, invocs, "");
3166      val = LLVMBuildAdd(builder, val,
3167                         LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 1, 0),
3168                         "invoc_count");
3169      LLVMBuildStore(builder, val, invocs);
3170   }
3171
3172   /* code generated texture sampling */
3173   sampler = lp_llvm_sampler_soa_create(lp_fs_variant_key_samplers(key), key->nr_samplers);
3174   image = lp_llvm_image_soa_create(lp_fs_variant_key_images(key), key->nr_images);
3175
3176   num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
3177   /* for 1d resources only run "upper half" of stamp */
3178   if (key->resource_1d)
3179      num_fs /= 2;
3180
3181   {
3182      LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
3183      LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
3184      LLVMValueRef num_loop_samp = lp_build_const_int32(gallivm, num_fs * key->coverage_samples);
3185      LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
3186                                                      num_loop_samp, "mask_store");
3187
3188      LLVMTypeRef flt_type = LLVMFloatTypeInContext(gallivm->context);
3189      LLVMValueRef glob_sample_pos = LLVMAddGlobal(gallivm->module, LLVMArrayType(flt_type, key->coverage_samples * 2), "");
3190      LLVMValueRef sample_pos_array;
3191
3192      if (key->multisample && key->coverage_samples == 4) {
3193         LLVMValueRef sample_pos_arr[8];
3194         for (unsigned i = 0; i < 4; i++) {
3195            sample_pos_arr[i * 2] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][0]);
3196            sample_pos_arr[i * 2 + 1] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][1]);
3197         }
3198         sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 8);
3199      } else {
3200         LLVMValueRef sample_pos_arr[2];
3201         sample_pos_arr[0] = LLVMConstReal(flt_type, 0.5);
3202         sample_pos_arr[1] = LLVMConstReal(flt_type, 0.5);
3203         sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 2);
3204      }
3205      LLVMSetInitializer(glob_sample_pos, sample_pos_array);
3206
3207      LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
3208      boolean pixel_center_integer =
3209         shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
3210
3211      /*
3212       * The shader input interpolation info is not explicitely baked in the
3213       * shader key, but everything it derives from (TGSI, and flatshade) is
3214       * already included in the shader key.
3215       */
3216      lp_build_interp_soa_init(&interp,
3217                               gallivm,
3218                               shader->info.base.num_inputs,
3219                               inputs,
3220                               pixel_center_integer,
3221                               key->coverage_samples, glob_sample_pos,
3222                               num_loop,
3223                               key->depth_clamp,
3224                               builder, fs_type,
3225                               a0_ptr, dadx_ptr, dady_ptr,
3226                               x, y);
3227
3228      for (i = 0; i < num_fs; i++) {
3229         if (key->multisample) {
3230            LLVMValueRef smask_val = LLVMBuildLoad(builder, lp_jit_context_sample_mask(gallivm, context_ptr), "");
3231
3232            /*
3233             * For multisampling, extract the per-sample mask from the incoming 64-bit mask,
3234             * store to the per sample mask storage. Or all of them together to generate
3235             * the fragment shader mask. (sample shading TODO).
3236             * Take the incoming state coverage mask into account.
3237             */
3238            for (unsigned s = 0; s < key->coverage_samples; s++) {
3239               LLVMValueRef sindexi = lp_build_const_int32(gallivm, i + (s * num_fs));
3240               LLVMValueRef sample_mask_ptr = LLVMBuildGEP(builder, mask_store,
3241                                                           &sindexi, 1, "sample_mask_ptr");
3242               LLVMValueRef s_mask = generate_quad_mask(gallivm, fs_type,
3243                                                        i*fs_type.length/4, s, mask_input);
3244
3245               LLVMValueRef smask_bit = LLVMBuildAnd(builder, smask_val, lp_build_const_int32(gallivm, (1 << s)), "");
3246               LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int32(gallivm, 0), "");
3247               smask_bit = LLVMBuildSExt(builder, cmp, int32_type, "");
3248               smask_bit = lp_build_broadcast(gallivm, mask_type, smask_bit);
3249
3250               s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");
3251               LLVMBuildStore(builder, s_mask, sample_mask_ptr);
3252            }
3253         } else {
3254            LLVMValueRef mask;
3255            LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
3256            LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
3257                                                 &indexi, 1, "mask_ptr");
3258
3259            if (partial_mask) {
3260               mask = generate_quad_mask(gallivm, fs_type,
3261                                         i*fs_type.length/4, 0, mask_input);
3262            }
3263            else {
3264               mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
3265            }
3266            LLVMBuildStore(builder, mask, mask_ptr);
3267         }
3268      }
3269
3270      generate_fs_loop(gallivm,
3271                       shader, key,
3272                       builder,
3273                       fs_type,
3274                       context_ptr,
3275                       glob_sample_pos,
3276                       num_loop,
3277                       &interp,
3278                       sampler,
3279                       image,
3280                       mask_store, /* output */
3281                       color_store,
3282                       depth_ptr,
3283                       depth_stride,
3284                       depth_sample_stride,
3285                       color_ptr_ptr,
3286                       stride_ptr,
3287                       color_sample_stride_ptr,
3288                       facing,
3289                       thread_data_ptr);
3290
3291      for (i = 0; i < num_fs; i++) {
3292         LLVMValueRef ptr;
3293         for (unsigned s = 0; s < key->coverage_samples; s++) {
3294            int idx = (i + (s * num_fs));
3295            LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
3296            ptr = LLVMBuildGEP(builder, mask_store, &sindexi, 1, "");
3297
3298            fs_mask[idx] = LLVMBuildLoad(builder, ptr, "smask");
3299         }
3300
3301         for (unsigned s = 0; s < key->min_samples; s++) {
3302            /* This is fucked up need to reorganize things */
3303            int idx = s * num_fs + i;
3304            LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);
3305            for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
3306               for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3307                  ptr = LLVMBuildGEP(builder,
3308                                     color_store[cbuf * !cbuf0_write_all][chan],
3309                                     &sindexi, 1, "");
3310                  fs_out_color[s][cbuf][chan][i] = ptr;
3311               }
3312            }
3313            if (dual_source_blend) {
3314               /* only support one dual source blend target hence always use output 1 */
3315               for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3316                  ptr = LLVMBuildGEP(builder,
3317                                     color_store[1][chan],
3318                                     &sindexi, 1, "");
3319                  fs_out_color[s][1][chan][i] = ptr;
3320               }
3321            }
3322         }
3323      }
3324   }
3325
3326   sampler->destroy(sampler);
3327   image->destroy(image);
3328   /* Loop over color outputs / color buffers to do blending.
3329    */
3330   for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
3331      if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
3332         LLVMValueRef color_ptr;
3333         LLVMValueRef stride;
3334         LLVMValueRef sample_stride = NULL;
3335         LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
3336
3337         boolean do_branch = ((key->depth.enabled
3338                               || key->stencil[0].enabled
3339                               || key->alpha.enabled)
3340                              && !shader->info.base.uses_kill);
3341
3342         color_ptr = LLVMBuildLoad(builder,
3343                                   LLVMBuildGEP(builder, color_ptr_ptr,
3344                                                &index, 1, ""),
3345                                   "");
3346
3347         stride = LLVMBuildLoad(builder,
3348                                LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
3349                                "");
3350
3351         if (key->cbuf_nr_samples[cbuf] > 1)
3352            sample_stride = LLVMBuildLoad(builder,
3353                                          LLVMBuildGEP(builder, color_sample_stride_ptr,
3354                                                       &index, 1, ""), "");
3355
3356         for (unsigned s = 0; s < key->cbuf_nr_samples[cbuf]; s++) {
3357            unsigned mask_idx = num_fs * (key->multisample ? s : 0);
3358            unsigned out_idx = key->min_samples == 1 ? 0 : s;
3359            LLVMValueRef out_ptr = color_ptr;;
3360
3361            if (sample_stride) {
3362               LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_stride, lp_build_const_int32(gallivm, s), "");
3363               out_ptr = LLVMBuildGEP(builder, out_ptr, &sample_offset, 1, "");
3364            }
3365            out_ptr = LLVMBuildBitCast(builder, out_ptr, LLVMPointerType(blend_vec_type, 0), "");
3366
3367            lp_build_name(out_ptr, "color_ptr%d", cbuf);
3368
3369            generate_unswizzled_blend(gallivm, cbuf, variant,
3370                                      key->cbuf_format[cbuf],
3371                                      num_fs, fs_type, &fs_mask[mask_idx], fs_out_color[out_idx],
3372                                      context_ptr, out_ptr, stride,
3373                                      partial_mask, do_branch);
3374         }
3375      }
3376   }
3377
3378   LLVMBuildRetVoid(builder);
3379
3380   gallivm_verify_function(gallivm, function);
3381}
3382
3383
3384static void
3385dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)
3386{
3387   unsigned i;
3388
3389   debug_printf("fs variant %p:\n", (void *) key);
3390
3391   if (key->flatshade) {
3392      debug_printf("flatshade = 1\n");
3393   }
3394   if (key->multisample) {
3395      debug_printf("multisample = 1\n");
3396      debug_printf("coverage samples = %d\n", key->coverage_samples);
3397      debug_printf("min samples = %d\n", key->min_samples);
3398   }
3399   for (i = 0; i < key->nr_cbufs; ++i) {
3400      debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
3401      debug_printf("cbuf nr_samples[%u] = %d\n", i, key->cbuf_nr_samples[i]);
3402   }
3403   if (key->depth.enabled || key->stencil[0].enabled) {
3404      debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
3405      debug_printf("depth nr_samples = %d\n", key->zsbuf_nr_samples);
3406   }
3407   if (key->depth.enabled) {
3408      debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));
3409      debug_printf("depth.writemask = %u\n", key->depth.writemask);
3410   }
3411
3412   for (i = 0; i < 2; ++i) {
3413      if (key->stencil[i].enabled) {
3414         debug_printf("stencil[%u].func = %s\n", i, util_str_func(key->stencil[i].func, TRUE));
3415         debug_printf("stencil[%u].fail_op = %s\n", i, util_str_stencil_op(key->stencil[i].fail_op, TRUE));
3416         debug_printf("stencil[%u].zpass_op = %s\n", i, util_str_stencil_op(key->stencil[i].zpass_op, TRUE));
3417         debug_printf("stencil[%u].zfail_op = %s\n", i, util_str_stencil_op(key->stencil[i].zfail_op, TRUE));
3418         debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
3419         debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
3420      }
3421   }
3422
3423   if (key->alpha.enabled) {
3424      debug_printf("alpha.func = %s\n", util_str_func(key->alpha.func, TRUE));
3425   }
3426
3427   if (key->occlusion_count) {
3428      debug_printf("occlusion_count = 1\n");
3429   }
3430
3431   if (key->blend.logicop_enable) {
3432      debug_printf("blend.logicop_func = %s\n", util_str_logicop(key->blend.logicop_func, TRUE));
3433   }
3434   else if (key->blend.rt[0].blend_enable) {
3435      debug_printf("blend.rgb_func = %s\n",   util_str_blend_func  (key->blend.rt[0].rgb_func, TRUE));
3436      debug_printf("blend.rgb_src_factor = %s\n",   util_str_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
3437      debug_printf("blend.rgb_dst_factor = %s\n",   util_str_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
3438      debug_printf("blend.alpha_func = %s\n",       util_str_blend_func  (key->blend.rt[0].alpha_func, TRUE));
3439      debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
3440      debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
3441   }
3442   debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
3443   if (key->blend.alpha_to_coverage) {
3444      debug_printf("blend.alpha_to_coverage is enabled\n");
3445   }
3446   for (i = 0; i < key->nr_samplers; ++i) {
3447      const struct lp_sampler_static_state *samplers = lp_fs_variant_key_samplers(key);
3448      const struct lp_static_sampler_state *sampler = &samplers[i].sampler_state;
3449      debug_printf("sampler[%u] = \n", i);
3450      debug_printf("  .wrap = %s %s %s\n",
3451                   util_str_tex_wrap(sampler->wrap_s, TRUE),
3452                   util_str_tex_wrap(sampler->wrap_t, TRUE),
3453                   util_str_tex_wrap(sampler->wrap_r, TRUE));
3454      debug_printf("  .min_img_filter = %s\n",
3455                   util_str_tex_filter(sampler->min_img_filter, TRUE));
3456      debug_printf("  .min_mip_filter = %s\n",
3457                   util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));
3458      debug_printf("  .mag_img_filter = %s\n",
3459                   util_str_tex_filter(sampler->mag_img_filter, TRUE));
3460      if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
3461         debug_printf("  .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));
3462      debug_printf("  .normalized_coords = %u\n", sampler->normalized_coords);
3463      debug_printf("  .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
3464      debug_printf("  .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
3465      debug_printf("  .apply_min_lod = %u\n", sampler->apply_min_lod);
3466      debug_printf("  .apply_max_lod = %u\n", sampler->apply_max_lod);
3467      debug_printf("  .reduction_mode = %u\n", sampler->reduction_mode);
3468      debug_printf("  .aniso = %u\n", sampler->aniso);
3469   }
3470   for (i = 0; i < key->nr_sampler_views; ++i) {
3471      const struct lp_sampler_static_state *samplers = lp_fs_variant_key_samplers(key);
3472      const struct lp_static_texture_state *texture = &samplers[i].texture_state;
3473      debug_printf("texture[%u] = \n", i);
3474      debug_printf("  .format = %s\n",
3475                   util_format_name(texture->format));
3476      debug_printf("  .target = %s\n",
3477                   util_str_tex_target(texture->target, TRUE));
3478      debug_printf("  .level_zero_only = %u\n",
3479                   texture->level_zero_only);
3480      debug_printf("  .pot = %u %u %u\n",
3481                   texture->pot_width,
3482                   texture->pot_height,
3483                   texture->pot_depth);
3484   }
3485   struct lp_image_static_state *images = lp_fs_variant_key_images(key);
3486   for (i = 0; i < key->nr_images; ++i) {
3487      const struct lp_static_texture_state *image = &images[i].image_state;
3488      debug_printf("image[%u] = \n", i);
3489      debug_printf("  .format = %s\n",
3490                   util_format_name(image->format));
3491      debug_printf("  .target = %s\n",
3492                   util_str_tex_target(image->target, TRUE));
3493      debug_printf("  .level_zero_only = %u\n",
3494                   image->level_zero_only);
3495      debug_printf("  .pot = %u %u %u\n",
3496                   image->pot_width,
3497                   image->pot_height,
3498                   image->pot_depth);
3499   }
3500}
3501
3502const char *
3503lp_debug_fs_kind(enum lp_fs_kind kind)
3504{
3505   switch(kind) {
3506   case LP_FS_KIND_GENERAL:
3507      return "GENERAL";
3508   case LP_FS_KIND_BLIT_RGBA:
3509      return "BLIT_RGBA";
3510   case LP_FS_KIND_BLIT_RGB1:
3511      return "BLIT_RGB1";
3512   case LP_FS_KIND_AERO_MINIFICATION:
3513      return "AERO_MINIFICATION";
3514   case LP_FS_KIND_LLVM_LINEAR:
3515      return "LLVM_LINEAR";
3516   default:
3517      return "unknown";
3518   }
3519}
3520
3521void
3522lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
3523{
3524   debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
3525                variant->shader->no, variant->no);
3526   if (variant->shader->base.type == PIPE_SHADER_IR_TGSI)
3527      tgsi_dump(variant->shader->base.tokens, 0);
3528   else
3529      nir_print_shader(variant->shader->base.ir.nir, stderr);
3530   dump_fs_variant_key(&variant->key);
3531   debug_printf("variant->opaque = %u\n", variant->opaque);
3532   debug_printf("variant->potentially_opaque = %u\n", variant->potentially_opaque);
3533   debug_printf("variant->blit = %u\n", variant->blit);
3534   debug_printf("shader->kind = %s\n", lp_debug_fs_kind(variant->shader->kind));
3535   debug_printf("\n");
3536}
3537
3538static void
3539lp_fs_get_ir_cache_key(struct lp_fragment_shader_variant *variant,
3540                            unsigned char ir_sha1_cache_key[20])
3541{
3542   struct blob blob = { 0 };
3543   unsigned ir_size;
3544   void *ir_binary;
3545
3546   blob_init(&blob);
3547   nir_serialize(&blob, variant->shader->base.ir.nir, true);
3548   ir_binary = blob.data;
3549   ir_size = blob.size;
3550
3551   struct mesa_sha1 ctx;
3552   _mesa_sha1_init(&ctx);
3553   _mesa_sha1_update(&ctx, &variant->key, variant->shader->variant_key_size);
3554   _mesa_sha1_update(&ctx, ir_binary, ir_size);
3555   _mesa_sha1_final(&ctx, ir_sha1_cache_key);
3556
3557   blob_finish(&blob);
3558}
3559
3560/**
3561 * Generate a new fragment shader variant from the shader code and
3562 * other state indicated by the key.
3563 */
3564static struct lp_fragment_shader_variant *
3565generate_variant(struct llvmpipe_context *lp,
3566                 struct lp_fragment_shader *shader,
3567                 const struct lp_fragment_shader_variant_key *key)
3568{
3569   struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
3570   struct lp_fragment_shader_variant *variant;
3571   const struct util_format_description *cbuf0_format_desc = NULL;
3572   boolean fullcolormask;
3573   boolean no_kill;
3574   boolean linear;
3575   char module_name[64];
3576   unsigned char ir_sha1_cache_key[20];
3577   struct lp_cached_code cached = { 0 };
3578   bool needs_caching = false;
3579   variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
3580   if (!variant)
3581      return NULL;
3582
3583   memset(variant, 0, sizeof(*variant));
3584   snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
3585            shader->no, shader->variants_created);
3586
3587   pipe_reference_init(&variant->reference, 1);
3588   lp_fs_reference(lp, &variant->shader, shader);
3589
3590   memcpy(&variant->key, key, shader->variant_key_size);
3591
3592   if (shader->base.ir.nir) {
3593      lp_fs_get_ir_cache_key(variant, ir_sha1_cache_key);
3594
3595      lp_disk_cache_find_shader(screen, &cached, ir_sha1_cache_key);
3596      if (!cached.data_size)
3597         needs_caching = true;
3598   }
3599   variant->gallivm = gallivm_create(module_name, lp->context, &cached);
3600   if (!variant->gallivm) {
3601      FREE(variant);
3602      return NULL;
3603   }
3604
3605   variant->list_item_global.base = variant;
3606   variant->list_item_local.base = variant;
3607   variant->no = shader->variants_created++;
3608
3609
3610
3611   /*
3612    * Determine whether we are touching all channels in the color buffer.
3613    */
3614   fullcolormask = FALSE;
3615   if (key->nr_cbufs == 1) {
3616      cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
3617      fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
3618   }
3619
3620   /* The scissor is ignored here as only tiles inside the scissoring
3621    * rectangle will refer to this */
3622   no_kill =
3623         fullcolormask &&
3624         !key->stencil[0].enabled &&
3625         !key->alpha.enabled &&
3626         !key->multisample &&
3627         !key->blend.alpha_to_coverage &&
3628         !key->depth.enabled &&
3629         !shader->info.base.uses_kill &&
3630         !shader->info.base.writes_samplemask;
3631
3632   variant->opaque =
3633         no_kill &&
3634         !key->blend.logicop_enable &&
3635         !key->blend.rt[0].blend_enable
3636         ? TRUE : FALSE;
3637
3638   variant->potentially_opaque =
3639         no_kill &&
3640         !key->blend.logicop_enable &&
3641         key->blend.rt[0].blend_enable &&
3642         key->blend.rt[0].rgb_func == PIPE_BLEND_ADD &&
3643         key->blend.rt[0].rgb_dst_factor == PIPE_BLENDFACTOR_INV_SRC_ALPHA &&
3644         key->blend.rt[0].alpha_func == key->blend.rt[0].rgb_func &&
3645         key->blend.rt[0].alpha_dst_factor == key->blend.rt[0].rgb_dst_factor &&
3646         shader->base.type == PIPE_SHADER_IR_TGSI &&
3647         /*
3648          * FIXME: for NIR, all of the fields of info.xxx (except info.base)
3649          * are zeros, hence shader analysis (here and elsewhere) using these
3650          * bits cannot work and will silently fail (cbuf is the only pointer
3651          * field, hence causing a crash).
3652          */
3653         shader->info.cbuf[0][3].file != TGSI_FILE_NULL
3654         ? TRUE : FALSE;
3655
3656   /* We only care about opaque blits for now */
3657   if (variant->opaque &&
3658       (shader->kind == LP_FS_KIND_BLIT_RGBA ||
3659        shader->kind == LP_FS_KIND_BLIT_RGB1)) {
3660      unsigned target, min_img_filter, mag_img_filter, min_mip_filter;
3661      enum pipe_format texture_format;
3662      struct lp_sampler_static_state *samp0 = lp_fs_variant_key_sampler_idx(key, 0);
3663      assert(samp0);
3664      texture_format = samp0->texture_state.format;
3665      target = samp0->texture_state.target;
3666      min_img_filter = samp0->sampler_state.min_img_filter;
3667      mag_img_filter = samp0->sampler_state.mag_img_filter;
3668      if (samp0->texture_state.level_zero_only) {
3669         min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
3670      } else {
3671         min_mip_filter = samp0->sampler_state.min_mip_filter;
3672      }
3673
3674      if (target == PIPE_TEXTURE_2D &&
3675          min_img_filter == PIPE_TEX_FILTER_NEAREST &&
3676          mag_img_filter == PIPE_TEX_FILTER_NEAREST &&
3677          min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
3678          ((texture_format &&
3679            util_is_format_compatible(util_format_description(texture_format),
3680                                      cbuf0_format_desc)) ||
3681           (shader->kind == LP_FS_KIND_BLIT_RGB1 &&
3682            (texture_format == PIPE_FORMAT_B8G8R8A8_UNORM ||
3683             texture_format == PIPE_FORMAT_B8G8R8X8_UNORM) &&
3684            (key->cbuf_format[0] == PIPE_FORMAT_B8G8R8A8_UNORM ||
3685             key->cbuf_format[0] == PIPE_FORMAT_B8G8R8X8_UNORM))))
3686         variant->blit = 1;
3687   }
3688
3689
3690   /* Whether this is a candidate for the linear path */
3691   linear =
3692         !key->stencil[0].enabled &&
3693         !key->depth.enabled &&
3694         !shader->info.base.uses_kill &&
3695         !key->blend.logicop_enable &&
3696         (key->cbuf_format[0] == PIPE_FORMAT_B8G8R8A8_UNORM ||
3697          key->cbuf_format[0] == PIPE_FORMAT_B8G8R8X8_UNORM);
3698
3699   memcpy(&variant->key, key, sizeof *key);
3700
3701   if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3702      lp_debug_fs_variant(variant);
3703   }
3704
3705   llvmpipe_fs_variant_fastpath(variant);
3706
3707   lp_jit_init_types(variant);
3708
3709   if (variant->jit_function[RAST_EDGE_TEST] == NULL)
3710      generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
3711
3712   if (variant->jit_function[RAST_WHOLE] == NULL) {
3713      if (variant->opaque) {
3714         /* Specialized shader, which doesn't need to read the color buffer. */
3715         generate_fragment(lp, shader, variant, RAST_WHOLE);
3716      }
3717   }
3718
3719   if (linear) {
3720      /* Currently keeping both the old fastpaths and new linear path
3721       * active.  The older code is still somewhat faster for the cases
3722       * it covers.
3723       *
3724       * XXX: consider restricting this to aero-mode only.
3725       */
3726      if (fullcolormask &&
3727          !key->alpha.enabled &&
3728          !key->blend.alpha_to_coverage) {
3729         llvmpipe_fs_variant_linear_fastpath(variant);
3730      }
3731
3732      /* If the original fastpath doesn't cover this variant, try the new
3733       * code:
3734       */
3735      if (variant->jit_linear == NULL) {
3736         if (shader->kind == LP_FS_KIND_BLIT_RGBA ||
3737             shader->kind == LP_FS_KIND_BLIT_RGB1 ||
3738             shader->kind == LP_FS_KIND_LLVM_LINEAR) {
3739            llvmpipe_fs_variant_linear_llvm(lp, shader, variant);
3740         }
3741      }
3742   } else {
3743      if (LP_DEBUG & DEBUG_LINEAR) {
3744         lp_debug_fs_variant(variant);
3745         debug_printf("    ----> no linear path for this variant\n");
3746      }
3747   }
3748
3749   /*
3750    * Compile everything
3751    */
3752
3753   gallivm_compile_module(variant->gallivm);
3754
3755   variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
3756
3757   if (variant->function[RAST_EDGE_TEST]) {
3758      variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
3759            gallivm_jit_function(variant->gallivm,
3760                                 variant->function[RAST_EDGE_TEST]);
3761   }
3762
3763   if (variant->function[RAST_WHOLE]) {
3764         variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
3765               gallivm_jit_function(variant->gallivm,
3766                                    variant->function[RAST_WHOLE]);
3767   } else if (!variant->jit_function[RAST_WHOLE]) {
3768      variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
3769   }
3770
3771   if (linear) {
3772      if (variant->linear_function) {
3773         variant->jit_linear_llvm = (lp_jit_linear_llvm_func)
3774               gallivm_jit_function(variant->gallivm, variant->linear_function);
3775      }
3776
3777      /*
3778       * This must be done after LLVM compilation, as it will call the JIT'ed
3779       * code to determine active inputs.
3780       */
3781      lp_linear_check_variant(variant);
3782   }
3783
3784   if (needs_caching) {
3785      lp_disk_cache_insert_shader(screen, &cached, ir_sha1_cache_key);
3786   }
3787
3788   gallivm_free_ir(variant->gallivm);
3789
3790   return variant;
3791}
3792
3793
3794static void *
3795llvmpipe_create_fs_state(struct pipe_context *pipe,
3796                         const struct pipe_shader_state *templ)
3797{
3798   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3799   struct lp_fragment_shader *shader;
3800   int nr_samplers;
3801   int nr_sampler_views;
3802   int nr_images;
3803   int i;
3804
3805   shader = CALLOC_STRUCT(lp_fragment_shader);
3806   if (!shader)
3807      return NULL;
3808
3809   pipe_reference_init(&shader->reference, 1);
3810   shader->no = fs_no++;
3811   make_empty_list(&shader->variants);
3812
3813   shader->base.type = templ->type;
3814   if (templ->type == PIPE_SHADER_IR_TGSI) {
3815      /* get/save the summary info for this shader */
3816      lp_build_tgsi_info(templ->tokens, &shader->info);
3817
3818      /* we need to keep a local copy of the tokens */
3819      shader->base.tokens = tgsi_dup_tokens(templ->tokens);
3820   } else {
3821      shader->base.ir.nir = templ->ir.nir;
3822      nir_tgsi_scan_shader(templ->ir.nir, &shader->info.base, true);
3823   }
3824
3825   shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
3826   if (shader->draw_data == NULL) {
3827      FREE((void *) shader->base.tokens);
3828      FREE(shader);
3829      return NULL;
3830   }
3831
3832   nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
3833   nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
3834   nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
3835   shader->variant_key_size = lp_fs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);
3836
3837   for (i = 0; i < shader->info.base.num_inputs; i++) {
3838      shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
3839      shader->inputs[i].location = shader->info.base.input_interpolate_loc[i];
3840
3841      switch (shader->info.base.input_interpolate[i]) {
3842      case TGSI_INTERPOLATE_CONSTANT:
3843         shader->inputs[i].interp = LP_INTERP_CONSTANT;
3844         break;
3845      case TGSI_INTERPOLATE_LINEAR:
3846         shader->inputs[i].interp = LP_INTERP_LINEAR;
3847         break;
3848      case TGSI_INTERPOLATE_PERSPECTIVE:
3849         shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
3850         break;
3851      case TGSI_INTERPOLATE_COLOR:
3852         shader->inputs[i].interp = LP_INTERP_COLOR;
3853         break;
3854      default:
3855         assert(0);
3856         break;
3857      }
3858
3859      switch (shader->info.base.input_semantic_name[i]) {
3860      case TGSI_SEMANTIC_FACE:
3861         shader->inputs[i].interp = LP_INTERP_FACING;
3862         break;
3863      case TGSI_SEMANTIC_POSITION:
3864         /* Position was already emitted above
3865          */
3866         shader->inputs[i].interp = LP_INTERP_POSITION;
3867         shader->inputs[i].src_index = 0;
3868         continue;
3869      }
3870
3871      /* XXX this is a completely pointless index map... */
3872      shader->inputs[i].src_index = i+1;
3873   }
3874
3875   if (LP_DEBUG & DEBUG_TGSI && templ->type == PIPE_SHADER_IR_TGSI) {
3876      unsigned attrib;
3877      debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
3878                   shader->no, (void *) shader);
3879      tgsi_dump(templ->tokens, 0);
3880      debug_printf("usage masks:\n");
3881      for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
3882         unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
3883         debug_printf("  IN[%u].%s%s%s%s\n",
3884                      attrib,
3885                      usage_mask & TGSI_WRITEMASK_X ? "x" : "",
3886                      usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
3887                      usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
3888                      usage_mask & TGSI_WRITEMASK_W ? "w" : "");
3889      }
3890      debug_printf("\n");
3891   }
3892
3893   /* This will put a derived copy of the tokens into shader->base.tokens */
3894   if (templ->type == PIPE_SHADER_IR_TGSI)
3895     llvmpipe_fs_analyse(shader, templ->tokens);
3896   else
3897     shader->kind = LP_FS_KIND_GENERAL;
3898
3899   return shader;
3900}
3901
3902
3903static void
3904llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
3905{
3906   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3907   struct lp_fragment_shader *lp_fs = (struct lp_fragment_shader *)fs;
3908   if (llvmpipe->fs == lp_fs)
3909      return;
3910
3911   draw_bind_fragment_shader(llvmpipe->draw,
3912                             (lp_fs ? lp_fs->draw_data : NULL));
3913
3914   lp_fs_reference(llvmpipe, &llvmpipe->fs, lp_fs);
3915
3916   /* invalidate the setup link, NEW_FS will make it update */
3917   lp_setup_set_fs_variant(llvmpipe->setup, NULL);
3918   llvmpipe->dirty |= LP_NEW_FS;
3919}
3920
3921
3922/**
3923 * Remove shader variant from two lists: the shader's variant list
3924 * and the context's variant list.
3925 */
3926
3927static
3928void llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
3929                                    struct lp_fragment_shader_variant *variant)
3930{
3931   if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
3932      debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "
3933                   "v total cached %u inst %u total inst %u\n",
3934                   variant->shader->no, variant->no,
3935                   variant->shader->variants_created,
3936                   variant->shader->variants_cached,
3937                   lp->nr_fs_variants, variant->nr_instrs, lp->nr_fs_instrs);
3938   }
3939
3940   /* remove from shader's list */
3941   remove_from_list(&variant->list_item_local);
3942   variant->shader->variants_cached--;
3943
3944   /* remove from context's list */
3945   remove_from_list(&variant->list_item_global);
3946   lp->nr_fs_variants--;
3947   lp->nr_fs_instrs -= variant->nr_instrs;
3948}
3949
3950void
3951llvmpipe_destroy_shader_variant(struct llvmpipe_context *lp,
3952                               struct lp_fragment_shader_variant *variant)
3953{
3954   gallivm_destroy(variant->gallivm);
3955
3956   lp_fs_reference(lp, &variant->shader, NULL);
3957
3958   FREE(variant);
3959}
3960
3961void
3962llvmpipe_destroy_fs(struct llvmpipe_context *llvmpipe,
3963                    struct lp_fragment_shader *shader)
3964{
3965   /* Delete draw module's data */
3966   draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
3967
3968   if (shader->base.ir.nir)
3969      ralloc_free(shader->base.ir.nir);
3970   assert(shader->variants_cached == 0);
3971   FREE((void *) shader->base.tokens);
3972   FREE(shader);
3973}
3974
3975static void
3976llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
3977{
3978   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
3979   struct lp_fragment_shader *shader = fs;
3980   struct lp_fs_variant_list_item *li;
3981
3982   /* Delete all the variants */
3983   li = first_elem(&shader->variants);
3984   while(!at_end(&shader->variants, li)) {
3985      struct lp_fs_variant_list_item *next = next_elem(li);
3986      struct lp_fragment_shader_variant *variant;
3987      variant = li->base;
3988      llvmpipe_remove_shader_variant(llvmpipe, li->base);
3989      lp_fs_variant_reference(llvmpipe, &variant, NULL);
3990      li = next;
3991   }
3992
3993   lp_fs_reference(llvmpipe, &shader, NULL);
3994}
3995
3996static void
3997llvmpipe_set_constant_buffer(struct pipe_context *pipe,
3998                             enum pipe_shader_type shader, uint index,
3999                             bool take_ownership,
4000                             const struct pipe_constant_buffer *cb)
4001{
4002   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
4003   struct pipe_constant_buffer *constants = &llvmpipe->constants[shader][index];
4004
4005   assert(shader < PIPE_SHADER_TYPES);
4006   assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
4007
4008   /* note: reference counting */
4009   util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb,
4010                             take_ownership);
4011
4012   /* user_buffer is only valid until the next set_constant_buffer (at most,
4013    * possibly until shader deletion), so we need to upload it now to make sure
4014    * it doesn't get updated/freed out from under us.
4015    */
4016   if (constants->user_buffer) {
4017      u_upload_data(llvmpipe->pipe.const_uploader, 0, constants->buffer_size, 16,
4018                    constants->user_buffer, &constants->buffer_offset,
4019                    &constants->buffer);
4020   }
4021   if (constants->buffer) {
4022       if (!(constants->buffer->bind & PIPE_BIND_CONSTANT_BUFFER)) {
4023         debug_printf("Illegal set constant without bind flag\n");
4024         constants->buffer->bind |= PIPE_BIND_CONSTANT_BUFFER;
4025      }
4026   }
4027
4028   if (shader == PIPE_SHADER_VERTEX ||
4029       shader == PIPE_SHADER_GEOMETRY ||
4030       shader == PIPE_SHADER_TESS_CTRL ||
4031       shader == PIPE_SHADER_TESS_EVAL) {
4032      /* Pass the constants to the 'draw' module */
4033      const unsigned size = cb ? cb->buffer_size : 0;
4034
4035      const ubyte *data = NULL;
4036      if (constants->buffer)
4037         data = (ubyte *) llvmpipe_resource_data(constants->buffer) + constants->buffer_offset;
4038
4039      draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
4040                                      index, data, size);
4041   }
4042   else if (shader == PIPE_SHADER_COMPUTE)
4043      llvmpipe->cs_dirty |= LP_CSNEW_CONSTANTS;
4044   else
4045      llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
4046}
4047
4048static void
4049llvmpipe_set_shader_buffers(struct pipe_context *pipe,
4050                            enum pipe_shader_type shader, unsigned start_slot,
4051                            unsigned count, const struct pipe_shader_buffer *buffers,
4052                            unsigned writable_bitmask)
4053{
4054   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
4055   unsigned i, idx;
4056   for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
4057      const struct pipe_shader_buffer *buffer = buffers ? &buffers[idx] : NULL;
4058
4059      util_copy_shader_buffer(&llvmpipe->ssbos[shader][i], buffer);
4060
4061      if (shader == PIPE_SHADER_VERTEX ||
4062          shader == PIPE_SHADER_GEOMETRY ||
4063          shader == PIPE_SHADER_TESS_CTRL ||
4064          shader == PIPE_SHADER_TESS_EVAL) {
4065         const unsigned size = buffer ? buffer->buffer_size : 0;
4066         const ubyte *data = NULL;
4067         if (buffer && buffer->buffer)
4068            data = (ubyte *) llvmpipe_resource_data(buffer->buffer);
4069         if (data)
4070            data += buffer->buffer_offset;
4071         draw_set_mapped_shader_buffer(llvmpipe->draw, shader,
4072                                       i, data, size);
4073      } else if (shader == PIPE_SHADER_COMPUTE) {
4074	 llvmpipe->cs_dirty |= LP_CSNEW_SSBOS;
4075      } else if (shader == PIPE_SHADER_FRAGMENT) {
4076         llvmpipe->dirty |= LP_NEW_FS_SSBOS;
4077      }
4078   }
4079}
4080
4081static void
4082llvmpipe_set_shader_images(struct pipe_context *pipe,
4083                            enum pipe_shader_type shader, unsigned start_slot,
4084                           unsigned count, unsigned unbind_num_trailing_slots,
4085                           const struct pipe_image_view *images)
4086{
4087   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
4088   unsigned i, idx;
4089
4090   draw_flush(llvmpipe->draw);
4091   for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {
4092      const struct pipe_image_view *image = images ? &images[idx] : NULL;
4093
4094      util_copy_image_view(&llvmpipe->images[shader][i], image);
4095   }
4096
4097   llvmpipe->num_images[shader] = start_slot + count;
4098   if (shader == PIPE_SHADER_VERTEX ||
4099       shader == PIPE_SHADER_GEOMETRY ||
4100       shader == PIPE_SHADER_TESS_CTRL ||
4101       shader == PIPE_SHADER_TESS_EVAL) {
4102      draw_set_images(llvmpipe->draw,
4103                      shader,
4104                      llvmpipe->images[shader],
4105                      start_slot + count);
4106   } else if (shader == PIPE_SHADER_COMPUTE)
4107      llvmpipe->cs_dirty |= LP_CSNEW_IMAGES;
4108   else
4109      llvmpipe->dirty |= LP_NEW_FS_IMAGES;
4110
4111   if (unbind_num_trailing_slots) {
4112      llvmpipe_set_shader_images(pipe, shader, start_slot + count,
4113                                 unbind_num_trailing_slots, 0, NULL);
4114   }
4115}
4116
4117/**
4118 * Return the blend factor equivalent to a destination alpha of one.
4119 */
4120static inline unsigned
4121force_dst_alpha_one(unsigned factor, boolean clamped_zero)
4122{
4123   switch(factor) {
4124   case PIPE_BLENDFACTOR_DST_ALPHA:
4125      return PIPE_BLENDFACTOR_ONE;
4126   case PIPE_BLENDFACTOR_INV_DST_ALPHA:
4127      return PIPE_BLENDFACTOR_ZERO;
4128   case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
4129      if (clamped_zero)
4130         return PIPE_BLENDFACTOR_ZERO;
4131      else
4132         return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
4133   }
4134
4135   return factor;
4136}
4137
4138
4139/**
4140 * We need to generate several variants of the fragment pipeline to match
4141 * all the combinations of the contributing state atoms.
4142 *
4143 * TODO: there is actually no reason to tie this to context state -- the
4144 * generated code could be cached globally in the screen.
4145 */
4146static struct lp_fragment_shader_variant_key *
4147make_variant_key(struct llvmpipe_context *lp,
4148                 struct lp_fragment_shader *shader,
4149                 char *store)
4150{
4151   unsigned i;
4152   struct lp_fragment_shader_variant_key *key;
4153
4154   key = (struct lp_fragment_shader_variant_key *)store;
4155
4156   memset(key, 0, sizeof(*key));
4157
4158   if (lp->framebuffer.zsbuf) {
4159      enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
4160      const struct util_format_description *zsbuf_desc =
4161         util_format_description(zsbuf_format);
4162
4163      if (lp->depth_stencil->depth_enabled &&
4164          util_format_has_depth(zsbuf_desc)) {
4165         key->zsbuf_format = zsbuf_format;
4166         key->depth.enabled = lp->depth_stencil->depth_enabled;
4167         key->depth.writemask = lp->depth_stencil->depth_writemask;
4168         key->depth.func = lp->depth_stencil->depth_func;
4169      }
4170      if (lp->depth_stencil->stencil[0].enabled &&
4171          util_format_has_stencil(zsbuf_desc)) {
4172         key->zsbuf_format = zsbuf_format;
4173         memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
4174      }
4175      if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
4176         key->resource_1d = TRUE;
4177      }
4178      key->zsbuf_nr_samples = util_res_sample_count(lp->framebuffer.zsbuf->texture);
4179   }
4180
4181   /*
4182    * Propagate the depth clamp setting from the rasterizer state.
4183    */
4184   key->depth_clamp = lp->rasterizer->depth_clamp;
4185
4186   /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
4187   if (!lp->framebuffer.nr_cbufs ||
4188       !lp->framebuffer.cbufs[0] ||
4189       !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
4190      key->alpha.enabled = lp->depth_stencil->alpha_enabled;
4191   }
4192   if(key->alpha.enabled)
4193      key->alpha.func = lp->depth_stencil->alpha_func;
4194   /* alpha.ref_value is passed in jit_context */
4195
4196   key->flatshade = lp->rasterizer->flatshade;
4197   key->multisample = lp->rasterizer->multisample;
4198   key->no_ms_sample_mask_out = lp->rasterizer->no_ms_sample_mask_out;
4199   if (lp->active_occlusion_queries && !lp->queries_disabled) {
4200      key->occlusion_count = TRUE;
4201   }
4202
4203   memcpy(&key->blend, lp->blend, sizeof key->blend);
4204
4205   key->coverage_samples = 1;
4206   key->min_samples = 1;
4207   if (key->multisample) {
4208      key->coverage_samples = util_framebuffer_get_num_samples(&lp->framebuffer);
4209      key->min_samples = lp->min_samples == 1 ? 1 : key->coverage_samples;
4210   }
4211   key->nr_cbufs = lp->framebuffer.nr_cbufs;
4212
4213   if (!key->blend.independent_blend_enable) {
4214      /* we always need independent blend otherwise the fixups below won't work */
4215      for (i = 1; i < key->nr_cbufs; i++) {
4216         memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
4217      }
4218      key->blend.independent_blend_enable = 1;
4219   }
4220
4221   for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
4222      struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
4223
4224      if (lp->framebuffer.cbufs[i]) {
4225         enum pipe_format format = lp->framebuffer.cbufs[i]->format;
4226         const struct util_format_description *format_desc;
4227
4228         key->cbuf_format[i] = format;
4229         key->cbuf_nr_samples[i] = util_res_sample_count(lp->framebuffer.cbufs[i]->texture);
4230
4231         /*
4232          * Figure out if this is a 1d resource. Note that OpenGL allows crazy
4233          * mixing of 2d textures with height 1 and 1d textures, so make sure
4234          * we pick 1d if any cbuf or zsbuf is 1d.
4235          */
4236         if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
4237            key->resource_1d = TRUE;
4238         }
4239
4240         format_desc = util_format_description(format);
4241         assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
4242                format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
4243
4244         /*
4245          * Mask out color channels not present in the color buffer.
4246          */
4247         blend_rt->colormask &= util_format_colormask(format_desc);
4248
4249         /*
4250          * Disable blend for integer formats.
4251          */
4252         if (util_format_is_pure_integer(format)) {
4253            blend_rt->blend_enable = 0;
4254         }
4255
4256         /*
4257          * Our swizzled render tiles always have an alpha channel, but the
4258          * linear render target format often does not, so force here the dst
4259          * alpha to be one.
4260          *
4261          * This is not a mere optimization. Wrong results will be produced if
4262          * the dst alpha is used, the dst format does not have alpha, and the
4263          * previous rendering was not flushed from the swizzled to linear
4264          * buffer. For example, NonPowTwo DCT.
4265          *
4266          * TODO: This should be generalized to all channels for better
4267          * performance, but only alpha causes correctness issues.
4268          *
4269          * Also, force rgb/alpha func/factors match, to make AoS blending
4270          * easier.
4271          */
4272         if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
4273             format_desc->swizzle[3] == format_desc->swizzle[0]) {
4274            /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
4275            boolean clamped_zero = !util_format_is_float(format) &&
4276                                   !util_format_is_snorm(format);
4277            blend_rt->rgb_src_factor =
4278               force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
4279            blend_rt->rgb_dst_factor =
4280               force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
4281            blend_rt->alpha_func       = blend_rt->rgb_func;
4282            blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
4283            blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
4284         }
4285      }
4286      else {
4287         /* no color buffer for this fragment output */
4288         key->cbuf_format[i] = PIPE_FORMAT_NONE;
4289         key->cbuf_nr_samples[i] = 0;
4290         blend_rt->colormask = 0x0;
4291         blend_rt->blend_enable = 0;
4292      }
4293   }
4294
4295   /* This value will be the same for all the variants of a given shader:
4296    */
4297   key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
4298
4299   struct lp_sampler_static_state *fs_sampler;
4300
4301   fs_sampler = lp_fs_variant_key_samplers(key);
4302
4303   memset(fs_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *fs_sampler);
4304
4305   for(i = 0; i < key->nr_samplers; ++i) {
4306      if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
4307         lp_sampler_static_sampler_state(&fs_sampler[i].sampler_state,
4308                                         lp->samplers[PIPE_SHADER_FRAGMENT][i]);
4309      }
4310   }
4311
4312   /*
4313    * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
4314    * are dx10-style? Can't really have mixed opcodes, at least not
4315    * if we want to skip the holes here (without rescanning tgsi).
4316    */
4317   if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
4318      key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
4319      for(i = 0; i < key->nr_sampler_views; ++i) {
4320         /*
4321          * Note sview may exceed what's representable by file_mask.
4322          * This will still work, the only downside is that not actually
4323          * used views may be included in the shader key.
4324          */
4325         if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
4326            lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
4327                                            lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
4328         }
4329      }
4330   }
4331   else {
4332      key->nr_sampler_views = key->nr_samplers;
4333      for(i = 0; i < key->nr_sampler_views; ++i) {
4334         if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
4335            lp_sampler_static_texture_state(&fs_sampler[i].texture_state,
4336                                            lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
4337         }
4338      }
4339   }
4340
4341   struct lp_image_static_state *lp_image;
4342   lp_image = lp_fs_variant_key_images(key);
4343   key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
4344   for (i = 0; i < key->nr_images; ++i) {
4345      if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {
4346         lp_sampler_static_texture_state_image(&lp_image[i].image_state,
4347                                               &lp->images[PIPE_SHADER_FRAGMENT][i]);
4348      }
4349   }
4350
4351   if (shader->kind == LP_FS_KIND_AERO_MINIFICATION) {
4352      struct lp_sampler_static_state *samp0 = lp_fs_variant_key_sampler_idx(key, 0);
4353      assert(samp0);
4354      samp0->sampler_state.min_img_filter = PIPE_TEX_FILTER_NEAREST;
4355      samp0->sampler_state.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
4356   }
4357
4358   return key;
4359}
4360
4361
4362/**
4363 * Update fragment shader state.  This is called just prior to drawing
4364 * something when some fragment-related state has changed.
4365 */
4366void
4367llvmpipe_update_fs(struct llvmpipe_context *lp)
4368{
4369   struct lp_fragment_shader *shader = lp->fs;
4370   struct lp_fragment_shader_variant_key *key;
4371   struct lp_fragment_shader_variant *variant = NULL;
4372   struct lp_fs_variant_list_item *li;
4373   char store[LP_FS_MAX_VARIANT_KEY_SIZE];
4374
4375   key = make_variant_key(lp, shader, store);
4376
4377   /* Search the variants for one which matches the key */
4378   li = first_elem(&shader->variants);
4379   while(!at_end(&shader->variants, li)) {
4380      if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
4381         variant = li->base;
4382         break;
4383      }
4384      li = next_elem(li);
4385   }
4386
4387   if (variant) {
4388      /* Move this variant to the head of the list to implement LRU
4389       * deletion of shader's when we have too many.
4390       */
4391      move_to_head(&lp->fs_variants_list, &variant->list_item_global);
4392   }
4393   else {
4394      /* variant not found, create it now */
4395      int64_t t0, t1, dt;
4396      unsigned i;
4397      unsigned variants_to_cull;
4398
4399      if (LP_DEBUG & DEBUG_FS) {
4400         debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
4401                      lp->nr_fs_variants,
4402                      lp->nr_fs_instrs,
4403                      lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
4404      }
4405
4406      /* First, check if we've exceeded the max number of shader variants.
4407       * If so, free 6.25% of them (the least recently used ones).
4408       */
4409      variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 16 : 0;
4410
4411      if (variants_to_cull ||
4412          lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
4413         if (gallivm_debug & GALLIVM_DEBUG_PERF) {
4414            debug_printf("Evicting FS: %u fs variants,\t%u total variants,"
4415                         "\t%u instrs,\t%u instrs/variant\n",
4416                         shader->variants_cached,
4417                         lp->nr_fs_variants, lp->nr_fs_instrs,
4418                         lp->nr_fs_instrs / lp->nr_fs_variants);
4419         }
4420
4421         /*
4422          * We need to re-check lp->nr_fs_variants because an arbitrarliy large
4423          * number of shader variants (potentially all of them) could be
4424          * pending for destruction on flush.
4425          */
4426
4427         for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
4428            struct lp_fs_variant_list_item *item;
4429            if (is_empty_list(&lp->fs_variants_list)) {
4430               break;
4431            }
4432            item = last_elem(&lp->fs_variants_list);
4433            assert(item);
4434            assert(item->base);
4435            llvmpipe_remove_shader_variant(lp, item->base);
4436            struct lp_fragment_shader_variant *variant = item->base;
4437            lp_fs_variant_reference(lp, &variant, NULL);
4438         }
4439      }
4440
4441      /*
4442       * Generate the new variant.
4443       */
4444      t0 = os_time_get();
4445      variant = generate_variant(lp, shader, key);
4446      t1 = os_time_get();
4447      dt = t1 - t0;
4448      LP_COUNT_ADD(llvm_compile_time, dt);
4449      LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
4450
4451      /* Put the new variant into the list */
4452      if (variant) {
4453         insert_at_head(&shader->variants, &variant->list_item_local);
4454         insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
4455         lp->nr_fs_variants++;
4456         lp->nr_fs_instrs += variant->nr_instrs;
4457         shader->variants_cached++;
4458      }
4459   }
4460
4461   /* Bind this variant */
4462   lp_setup_set_fs_variant(lp->setup, variant);
4463}
4464
4465
4466
4467
4468
4469void
4470llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
4471{
4472   llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
4473   llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
4474   llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
4475
4476   llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
4477
4478   llvmpipe->pipe.set_shader_buffers = llvmpipe_set_shader_buffers;
4479   llvmpipe->pipe.set_shader_images = llvmpipe_set_shader_images;
4480}
4481
4482
4483