2 * Copyright 2011 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * Authors: Tom Stellard <thomas.stellard@amd.com>
26 #include "radeon_llvm.h"
28 #include "gallivm/lp_bld_const.h"
29 #include "gallivm/lp_bld_gather.h"
30 #include "gallivm/lp_bld_flow.h"
31 #include "gallivm/lp_bld_init.h"
32 #include "gallivm/lp_bld_intr.h"
33 #include "gallivm/lp_bld_swizzle.h"
34 #include "tgsi/tgsi_info.h"
35 #include "tgsi/tgsi_parse.h"
36 #include "util/u_math.h"
37 #include "util/u_memory.h"
38 #include "util/u_debug.h"
40 #include <llvm-c/Core.h>
41 #include <llvm-c/Transforms/Scalar.h>
43 static struct radeon_llvm_loop * get_current_loop(struct radeon_llvm_context * ctx)
45 return ctx->loop_depth > 0 ? ctx->loop + (ctx->loop_depth - 1) : NULL;
48 static struct radeon_llvm_branch * get_current_branch(
49 struct radeon_llvm_context * ctx)
51 return ctx->branch_depth > 0 ?
52 ctx->branch + (ctx->branch_depth - 1) : NULL;
55 unsigned radeon_llvm_reg_index_soa(unsigned index, unsigned chan)
57 return (index * 4) + chan;
60 static LLVMValueRef emit_swizzle(
61 struct lp_build_tgsi_context * bld_base,
68 LLVMValueRef swizzles[4];
70 LLVMInt32TypeInContext(bld_base->base.gallivm->context);
72 swizzles[0] = LLVMConstInt(i32t, swizzle_x, 0);
73 swizzles[1] = LLVMConstInt(i32t, swizzle_y, 0);
74 swizzles[2] = LLVMConstInt(i32t, swizzle_z, 0);
75 swizzles[3] = LLVMConstInt(i32t, swizzle_w, 0);
77 return LLVMBuildShuffleVector(bld_base->base.gallivm->builder,
79 LLVMGetUndef(LLVMTypeOf(value)),
80 LLVMConstVector(swizzles, 4), "");
83 static struct tgsi_declaration_range
84 get_array_range(struct lp_build_tgsi_context *bld_base,
85 unsigned File, const struct tgsi_ind_register *reg)
87 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
88 if (File != TGSI_FILE_TEMPORARY || reg->ArrayID == 0 ||
89 reg->ArrayID > RADEON_LLVM_MAX_ARRAYS) {
90 struct tgsi_declaration_range range;
92 range.Last = bld_base->info->file_max[File];
96 return ctx->arrays[reg->ArrayID - 1];
101 struct lp_build_tgsi_soa_context *bld,
102 const struct tgsi_ind_register *reg,
105 struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
107 LLVMValueRef addr = LLVMBuildLoad(gallivm->builder, bld->addr[reg->Index][reg->Swizzle], "");
108 return LLVMBuildAdd(gallivm->builder, addr, lp_build_const_int32(gallivm, offset), "");
113 struct lp_build_tgsi_context *bld_base,
114 const struct tgsi_full_src_register *reg,
115 enum tgsi_opcode_type type,
120 struct lp_build_tgsi_context *bld_base,
121 unsigned File, enum tgsi_opcode_type type,
122 struct tgsi_declaration_range range,
125 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
126 struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
127 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
129 unsigned i, size = range.Last - range.First + 1;
130 LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
131 LLVMValueRef result = LLVMGetUndef(vec);
133 struct tgsi_full_src_register tmp_reg = {};
134 tmp_reg.Register.File = File;
136 for (i = 0; i < size; ++i) {
137 tmp_reg.Register.Index = i + range.First;
138 LLVMValueRef temp = emit_fetch(bld_base, &tmp_reg, type, swizzle);
139 result = LLVMBuildInsertElement(builder, result, temp,
140 lp_build_const_int32(gallivm, i), "");
145 static bool uses_temp_indirect_addressing(
146 struct lp_build_tgsi_context *bld_base)
148 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
149 return (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY));
154 struct lp_build_tgsi_context *bld_base,
155 const struct tgsi_full_src_register *reg,
156 enum tgsi_opcode_type type,
159 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
160 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
161 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
162 LLVMValueRef result = NULL, ptr;
165 LLVMValueRef values[TGSI_NUM_CHANNELS];
167 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
168 values[chan] = emit_fetch(bld_base, reg, type, chan);
170 return lp_build_gather_values(bld_base->base.gallivm, values,
174 if (reg->Register.Indirect) {
175 struct tgsi_declaration_range range = get_array_range(bld_base,
176 reg->Register.File, ®->Indirect);
177 return LLVMBuildExtractElement(builder,
178 emit_array_fetch(bld_base, reg->Register.File, type, range, swizzle),
179 emit_array_index(bld, ®->Indirect, reg->Register.Index - range.First),
183 switch(reg->Register.File) {
184 case TGSI_FILE_IMMEDIATE: {
185 LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
186 return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
189 case TGSI_FILE_INPUT:
190 result = ctx->inputs[radeon_llvm_reg_index_soa(reg->Register.Index, swizzle)];
193 case TGSI_FILE_TEMPORARY:
194 if (uses_temp_indirect_addressing(bld_base)) {
195 ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
198 ptr = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle];
199 result = LLVMBuildLoad(builder, ptr, "");
202 case TGSI_FILE_OUTPUT:
203 ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
204 result = LLVMBuildLoad(builder, ptr, "");
208 return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
211 return bitcast(bld_base, type, result);
214 static LLVMValueRef fetch_system_value(
215 struct lp_build_tgsi_context * bld_base,
216 const struct tgsi_full_src_register *reg,
217 enum tgsi_opcode_type type,
220 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
221 struct gallivm_state *gallivm = bld_base->base.gallivm;
223 LLVMValueRef cval = ctx->system_values[reg->Register.Index];
224 if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
225 cval = LLVMBuildExtractElement(gallivm->builder, cval,
226 lp_build_const_int32(gallivm, swizzle), "");
228 return bitcast(bld_base, type, cval);
231 static void emit_declaration(
232 struct lp_build_tgsi_context * bld_base,
233 const struct tgsi_full_declaration *decl)
235 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
236 unsigned first, last, i, idx;
237 switch(decl->Declaration.File) {
238 case TGSI_FILE_ADDRESS:
241 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
243 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
244 ctx->soa.addr[idx][chan] = lp_build_alloca(
246 ctx->soa.bld_base.uint_bld.elem_type, "");
252 case TGSI_FILE_TEMPORARY:
253 if (decl->Declaration.Array && decl->Array.ArrayID <= RADEON_LLVM_MAX_ARRAYS)
254 ctx->arrays[decl->Array.ArrayID - 1] = decl->Range;
255 if (uses_temp_indirect_addressing(bld_base)) {
256 lp_emit_declaration_soa(bld_base, decl);
259 first = decl->Range.First;
260 last = decl->Range.Last;
261 if (!ctx->temps_count) {
262 ctx->temps_count = bld_base->info->file_max[TGSI_FILE_TEMPORARY] + 1;
263 ctx->temps = MALLOC(TGSI_NUM_CHANNELS * ctx->temps_count * sizeof(LLVMValueRef));
265 for (idx = first; idx <= last; idx++) {
266 for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
267 ctx->temps[idx * TGSI_NUM_CHANNELS + i] =
268 lp_build_alloca(bld_base->base.gallivm, bld_base->base.vec_type,
274 case TGSI_FILE_INPUT:
277 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
279 ctx->load_input(ctx, idx, decl);
284 case TGSI_FILE_SYSTEM_VALUE:
287 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
288 ctx->load_system_value(ctx, idx, decl);
293 case TGSI_FILE_OUTPUT:
296 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
298 assert(idx < RADEON_LLVM_MAX_OUTPUTS);
299 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
300 ctx->soa.outputs[idx][chan] = lp_build_alloca(&ctx->gallivm,
301 ctx->soa.bld_base.base.elem_type, "");
305 ctx->output_reg_count = MAX2(ctx->output_reg_count,
306 decl->Range.Last + 1);
317 struct lp_build_tgsi_context * bld_base,
318 const struct tgsi_full_instruction * inst,
319 const struct tgsi_opcode_info * info,
322 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
323 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
324 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
325 struct lp_build_context base = bld->bld_base.base;
326 const struct tgsi_full_dst_register *reg = &inst->Dst[0];
327 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
328 LLVMValueRef temp_ptr;
329 unsigned chan, chan_index;
330 boolean is_vec_store = FALSE;
333 LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
334 is_vec_store = (k == LLVMVectorTypeKind);
338 LLVMValueRef values[4] = {};
339 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan) {
340 LLVMValueRef index = lp_build_const_int32(gallivm, chan);
341 values[chan] = LLVMBuildExtractElement(gallivm->builder,
344 bld_base->emit_store(bld_base, inst, info, values);
348 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
349 LLVMValueRef value = dst[chan_index];
351 if (inst->Instruction.Saturate != TGSI_SAT_NONE) {
352 struct lp_build_emit_data clamp_emit_data;
354 memset(&clamp_emit_data, 0, sizeof(clamp_emit_data));
355 clamp_emit_data.arg_count = 3;
356 clamp_emit_data.args[0] = value;
357 clamp_emit_data.args[2] = base.one;
359 switch(inst->Instruction.Saturate) {
360 case TGSI_SAT_ZERO_ONE:
361 clamp_emit_data.args[1] = base.zero;
363 case TGSI_SAT_MINUS_PLUS_ONE:
364 clamp_emit_data.args[1] = LLVMConstReal(
365 base.elem_type, -1.0f);
370 value = lp_build_emit_llvm(bld_base, TGSI_OPCODE_CLAMP,
374 if (reg->Register.File == TGSI_FILE_ADDRESS) {
375 temp_ptr = bld->addr[reg->Register.Index][chan_index];
376 LLVMBuildStore(builder, value, temp_ptr);
380 value = bitcast(bld_base, TGSI_TYPE_FLOAT, value);
382 if (reg->Register.Indirect) {
383 struct tgsi_declaration_range range = get_array_range(bld_base,
384 reg->Register.File, ®->Indirect);
386 unsigned i, size = range.Last - range.First + 1;
387 LLVMValueRef array = LLVMBuildInsertElement(builder,
388 emit_array_fetch(bld_base, reg->Register.File, TGSI_TYPE_FLOAT, range, chan_index),
389 value, emit_array_index(bld, ®->Indirect, reg->Register.Index - range.First), "");
391 for (i = 0; i < size; ++i) {
392 switch(reg->Register.File) {
393 case TGSI_FILE_OUTPUT:
394 temp_ptr = bld->outputs[i + range.First][chan_index];
397 case TGSI_FILE_TEMPORARY:
398 if (uses_temp_indirect_addressing(bld_base))
399 temp_ptr = lp_get_temp_ptr_soa(bld, i + range.First, chan_index);
401 temp_ptr = ctx->temps[(i + range.First) * TGSI_NUM_CHANNELS + chan_index];
407 value = LLVMBuildExtractElement(builder, array,
408 lp_build_const_int32(gallivm, i), "");
409 LLVMBuildStore(builder, value, temp_ptr);
413 switch(reg->Register.File) {
414 case TGSI_FILE_OUTPUT:
415 temp_ptr = bld->outputs[reg->Register.Index][chan_index];
418 case TGSI_FILE_TEMPORARY:
419 if (uses_temp_indirect_addressing(bld_base)) {
423 temp_ptr = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index];
429 LLVMBuildStore(builder, value, temp_ptr);
434 static void bgnloop_emit(
435 const struct lp_build_tgsi_action * action,
436 struct lp_build_tgsi_context * bld_base,
437 struct lp_build_emit_data * emit_data)
439 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
440 struct gallivm_state * gallivm = bld_base->base.gallivm;
441 LLVMBasicBlockRef loop_block;
442 LLVMBasicBlockRef endloop_block;
443 endloop_block = LLVMAppendBasicBlockInContext(gallivm->context,
444 ctx->main_fn, "ENDLOOP");
445 loop_block = LLVMInsertBasicBlockInContext(gallivm->context,
446 endloop_block, "LOOP");
447 LLVMBuildBr(gallivm->builder, loop_block);
448 LLVMPositionBuilderAtEnd(gallivm->builder, loop_block);
450 if (++ctx->loop_depth > ctx->loop_depth_max) {
451 unsigned new_max = ctx->loop_depth_max << 1;
454 new_max = RADEON_LLVM_INITIAL_CF_DEPTH;
456 ctx->loop = REALLOC(ctx->loop, ctx->loop_depth_max *
457 sizeof(ctx->loop[0]),
458 new_max * sizeof(ctx->loop[0]));
459 ctx->loop_depth_max = new_max;
462 ctx->loop[ctx->loop_depth - 1].loop_block = loop_block;
463 ctx->loop[ctx->loop_depth - 1].endloop_block = endloop_block;
466 static void brk_emit(
467 const struct lp_build_tgsi_action * action,
468 struct lp_build_tgsi_context * bld_base,
469 struct lp_build_emit_data * emit_data)
471 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
472 struct gallivm_state * gallivm = bld_base->base.gallivm;
473 struct radeon_llvm_loop * current_loop = get_current_loop(ctx);
475 LLVMBuildBr(gallivm->builder, current_loop->endloop_block);
478 static void cont_emit(
479 const struct lp_build_tgsi_action * action,
480 struct lp_build_tgsi_context * bld_base,
481 struct lp_build_emit_data * emit_data)
483 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
484 struct gallivm_state * gallivm = bld_base->base.gallivm;
485 struct radeon_llvm_loop * current_loop = get_current_loop(ctx);
487 LLVMBuildBr(gallivm->builder, current_loop->loop_block);
490 static void else_emit(
491 const struct lp_build_tgsi_action * action,
492 struct lp_build_tgsi_context * bld_base,
493 struct lp_build_emit_data * emit_data)
495 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
496 struct gallivm_state * gallivm = bld_base->base.gallivm;
497 struct radeon_llvm_branch * current_branch = get_current_branch(ctx);
498 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(gallivm->builder);
500 /* We need to add a terminator to the current block if the previous
501 * instruction was an ENDIF.Example:
514 if (current_block != current_branch->if_block) {
515 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
517 if (!LLVMGetBasicBlockTerminator(current_branch->if_block)) {
518 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
520 current_branch->has_else = 1;
521 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->else_block);
524 static void endif_emit(
525 const struct lp_build_tgsi_action * action,
526 struct lp_build_tgsi_context * bld_base,
527 struct lp_build_emit_data * emit_data)
529 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
530 struct gallivm_state * gallivm = bld_base->base.gallivm;
531 struct radeon_llvm_branch * current_branch = get_current_branch(ctx);
532 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(gallivm->builder);
534 /* If we have consecutive ENDIF instructions, then the first ENDIF
535 * will not have a terminator, so we need to add one. */
536 if (current_block != current_branch->if_block
537 && current_block != current_branch->else_block
538 && !LLVMGetBasicBlockTerminator(current_block)) {
540 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
542 if (!LLVMGetBasicBlockTerminator(current_branch->else_block)) {
543 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->else_block);
544 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
547 if (!LLVMGetBasicBlockTerminator(current_branch->if_block)) {
548 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->if_block);
549 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
552 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->endif_block);
556 static void endloop_emit(
557 const struct lp_build_tgsi_action * action,
558 struct lp_build_tgsi_context * bld_base,
559 struct lp_build_emit_data * emit_data)
561 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
562 struct gallivm_state * gallivm = bld_base->base.gallivm;
563 struct radeon_llvm_loop * current_loop = get_current_loop(ctx);
565 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(gallivm->builder))) {
566 LLVMBuildBr(gallivm->builder, current_loop->loop_block);
569 LLVMPositionBuilderAtEnd(gallivm->builder, current_loop->endloop_block);
573 static void if_cond_emit(
574 const struct lp_build_tgsi_action * action,
575 struct lp_build_tgsi_context * bld_base,
576 struct lp_build_emit_data * emit_data,
579 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
580 struct gallivm_state * gallivm = bld_base->base.gallivm;
581 LLVMBasicBlockRef if_block, else_block, endif_block;
583 endif_block = LLVMAppendBasicBlockInContext(gallivm->context,
584 ctx->main_fn, "ENDIF");
585 if_block = LLVMInsertBasicBlockInContext(gallivm->context,
587 else_block = LLVMInsertBasicBlockInContext(gallivm->context,
588 endif_block, "ELSE");
589 LLVMBuildCondBr(gallivm->builder, cond, if_block, else_block);
590 LLVMPositionBuilderAtEnd(gallivm->builder, if_block);
592 if (++ctx->branch_depth > ctx->branch_depth_max) {
593 unsigned new_max = ctx->branch_depth_max << 1;
596 new_max = RADEON_LLVM_INITIAL_CF_DEPTH;
598 ctx->branch = REALLOC(ctx->branch, ctx->branch_depth_max *
599 sizeof(ctx->branch[0]),
600 new_max * sizeof(ctx->branch[0]));
601 ctx->branch_depth_max = new_max;
604 ctx->branch[ctx->branch_depth - 1].endif_block = endif_block;
605 ctx->branch[ctx->branch_depth - 1].if_block = if_block;
606 ctx->branch[ctx->branch_depth - 1].else_block = else_block;
607 ctx->branch[ctx->branch_depth - 1].has_else = 0;
611 const struct lp_build_tgsi_action * action,
612 struct lp_build_tgsi_context * bld_base,
613 struct lp_build_emit_data * emit_data)
615 struct gallivm_state * gallivm = bld_base->base.gallivm;
618 cond = LLVMBuildFCmp(gallivm->builder, LLVMRealUNE,
620 bld_base->base.zero, "");
622 if_cond_emit(action, bld_base, emit_data, cond);
625 static void uif_emit(
626 const struct lp_build_tgsi_action * action,
627 struct lp_build_tgsi_context * bld_base,
628 struct lp_build_emit_data * emit_data)
630 struct gallivm_state * gallivm = bld_base->base.gallivm;
633 cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
634 bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
635 bld_base->int_bld.zero, "");
637 if_cond_emit(action, bld_base, emit_data, cond);
640 static void kil_emit(
641 const struct lp_build_tgsi_action * action,
642 struct lp_build_tgsi_context * bld_base,
643 struct lp_build_emit_data * emit_data)
646 for (i = 0; i < emit_data->arg_count; i++) {
647 emit_data->output[i] = lp_build_intrinsic_unary(
648 bld_base->base.gallivm->builder,
650 emit_data->dst_type, emit_data->args[i]);
654 void radeon_llvm_emit_prepare_cube_coords(
655 struct lp_build_tgsi_context * bld_base,
656 struct lp_build_emit_data * emit_data,
657 LLVMValueRef *coords_arg)
660 unsigned target = emit_data->inst->Texture.Texture;
661 unsigned opcode = emit_data->inst->Instruction.Opcode;
662 struct gallivm_state * gallivm = bld_base->base.gallivm;
663 LLVMBuilderRef builder = gallivm->builder;
664 LLVMTypeRef type = bld_base->base.elem_type;
665 LLVMValueRef coords[4];
666 LLVMValueRef mad_args[3];
668 struct LLVMOpaqueValue *cube_vec;
672 cube_vec = lp_build_gather_values(bld_base->base.gallivm, coords_arg, 4);
673 v = build_intrinsic(builder, "llvm.AMDGPU.cube", LLVMVectorType(type, 4),
674 &cube_vec, 1, LLVMReadNoneAttribute);
676 for (i = 0; i < 4; ++i) {
677 idx = lp_build_const_int32(gallivm, i);
678 coords[i] = LLVMBuildExtractElement(builder, v, idx, "");
681 coords[2] = build_intrinsic(builder, "fabs",
682 type, &coords[2], 1, LLVMReadNoneAttribute);
683 coords[2] = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_RCP, coords[2]);
685 mad_args[1] = coords[2];
686 mad_args[2] = LLVMConstReal(type, 1.5);
688 mad_args[0] = coords[0];
689 coords[0] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
690 mad_args[0], mad_args[1], mad_args[2]);
692 mad_args[0] = coords[1];
693 coords[1] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
694 mad_args[0], mad_args[1], mad_args[2]);
696 /* apply xyz = yxw swizzle to cooords */
697 coords[2] = coords[3];
698 coords[3] = coords[1];
699 coords[1] = coords[0];
700 coords[0] = coords[3];
702 if (target == TGSI_TEXTURE_CUBE_ARRAY ||
703 target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
704 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
705 /* coords_arg.w component - array_index for cube arrays */
706 coords[2] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
707 coords_arg[3], lp_build_const_float(gallivm, 8.0), coords[2]);
710 /* Preserve compare/lod/bias. Put it in coords.w. */
711 if (opcode == TGSI_OPCODE_TEX2 ||
712 opcode == TGSI_OPCODE_TXB2 ||
713 opcode == TGSI_OPCODE_TXL2) {
714 coords[3] = coords_arg[4];
715 } else if (opcode == TGSI_OPCODE_TXB ||
716 opcode == TGSI_OPCODE_TXL ||
717 target == TGSI_TEXTURE_SHADOWCUBE) {
718 coords[3] = coords_arg[3];
721 memcpy(coords_arg, coords, sizeof(coords));
724 static void txd_fetch_args(
725 struct lp_build_tgsi_context * bld_base,
726 struct lp_build_emit_data * emit_data)
728 const struct tgsi_full_instruction * inst = emit_data->inst;
730 LLVMValueRef coords[4];
732 for (src = 0; src < 3; src++) {
733 for (chan = 0; chan < 4; chan++)
734 coords[chan] = lp_build_emit_fetch(bld_base, inst, src, chan);
736 emit_data->args[src] = lp_build_gather_values(bld_base->base.gallivm,
739 emit_data->arg_count = 3;
740 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
744 static void txp_fetch_args(
745 struct lp_build_tgsi_context * bld_base,
746 struct lp_build_emit_data * emit_data)
748 const struct tgsi_full_instruction * inst = emit_data->inst;
751 LLVMValueRef coords[4];
753 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
754 src_w = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
756 for (chan = 0; chan < 3; chan++ ) {
757 LLVMValueRef arg = lp_build_emit_fetch(bld_base,
758 emit_data->inst, 0, chan);
759 coords[chan] = lp_build_emit_llvm_binary(bld_base,
760 TGSI_OPCODE_DIV, arg, src_w);
762 coords[3] = bld_base->base.one;
764 if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
765 inst->Texture.Texture == TGSI_TEXTURE_CUBE_ARRAY ||
766 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
767 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE_ARRAY) &&
768 inst->Instruction.Opcode != TGSI_OPCODE_TXQ &&
769 inst->Instruction.Opcode != TGSI_OPCODE_TXQ_LZ) {
770 radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords);
773 emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
775 emit_data->arg_count = 1;
778 static void tex_fetch_args(
779 struct lp_build_tgsi_context * bld_base,
780 struct lp_build_emit_data * emit_data)
782 /* XXX: lp_build_swizzle_aos() was failing with wrong arg types,
783 * when we used CHAN_ALL. We should be able to get this to work,
784 * but for now we will swizzle it ourselves
785 emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
790 const struct tgsi_full_instruction * inst = emit_data->inst;
792 LLVMValueRef coords[5];
794 for (chan = 0; chan < 4; chan++) {
795 coords[chan] = lp_build_emit_fetch(bld_base, inst, 0, chan);
798 if (inst->Instruction.Opcode == TGSI_OPCODE_TEX2 ||
799 inst->Instruction.Opcode == TGSI_OPCODE_TXB2 ||
800 inst->Instruction.Opcode == TGSI_OPCODE_TXL2) {
801 /* These instructions have additional operand that should be packed
802 * into the cube coord vector by radeon_llvm_emit_prepare_cube_coords.
803 * That operand should be passed as a float value in the args array
804 * right after the coord vector. After packing it's not used anymore,
805 * that's why arg_count is not increased */
806 coords[4] = lp_build_emit_fetch(bld_base, inst, 1, 0);
809 if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
810 inst->Texture.Texture == TGSI_TEXTURE_CUBE_ARRAY ||
811 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
812 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE_ARRAY) &&
813 inst->Instruction.Opcode != TGSI_OPCODE_TXQ &&
814 inst->Instruction.Opcode != TGSI_OPCODE_TXQ_LZ) {
815 radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords);
818 emit_data->arg_count = 1;
819 emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
821 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
824 static void txf_fetch_args(
825 struct lp_build_tgsi_context * bld_base,
826 struct lp_build_emit_data * emit_data)
828 const struct tgsi_full_instruction * inst = emit_data->inst;
829 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
830 const struct tgsi_texture_offset * off = inst->TexOffsets;
831 LLVMTypeRef offset_type = bld_base->int_bld.elem_type;
833 /* fetch tex coords */
834 tex_fetch_args(bld_base, emit_data);
836 /* fetch tex offsets */
837 if (inst->Texture.NumOffsets) {
838 assert(inst->Texture.NumOffsets == 1);
840 emit_data->args[1] = LLVMConstBitCast(
841 bld->immediates[off->Index][off->SwizzleX],
843 emit_data->args[2] = LLVMConstBitCast(
844 bld->immediates[off->Index][off->SwizzleY],
846 emit_data->args[3] = LLVMConstBitCast(
847 bld->immediates[off->Index][off->SwizzleZ],
850 emit_data->args[1] = bld_base->int_bld.zero;
851 emit_data->args[2] = bld_base->int_bld.zero;
852 emit_data->args[3] = bld_base->int_bld.zero;
855 emit_data->arg_count = 4;
858 static void emit_icmp(
859 const struct lp_build_tgsi_action * action,
860 struct lp_build_tgsi_context * bld_base,
861 struct lp_build_emit_data * emit_data)
864 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
865 LLVMContextRef context = bld_base->base.gallivm->context;
867 switch (emit_data->inst->Instruction.Opcode) {
868 case TGSI_OPCODE_USEQ: pred = LLVMIntEQ; break;
869 case TGSI_OPCODE_USNE: pred = LLVMIntNE; break;
870 case TGSI_OPCODE_USGE: pred = LLVMIntUGE; break;
871 case TGSI_OPCODE_USLT: pred = LLVMIntULT; break;
872 case TGSI_OPCODE_ISGE: pred = LLVMIntSGE; break;
873 case TGSI_OPCODE_ISLT: pred = LLVMIntSLT; break;
875 assert(!"unknown instruction");
880 LLVMValueRef v = LLVMBuildICmp(builder, pred,
881 emit_data->args[0], emit_data->args[1],"");
883 v = LLVMBuildSExtOrBitCast(builder, v,
884 LLVMInt32TypeInContext(context), "");
886 emit_data->output[emit_data->chan] = v;
889 static void emit_ucmp(
890 const struct lp_build_tgsi_action * action,
891 struct lp_build_tgsi_context * bld_base,
892 struct lp_build_emit_data * emit_data)
894 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
896 LLVMValueRef arg0 = LLVMBuildBitCast(builder, emit_data->args[0],
897 bld_base->uint_bld.elem_type, "");
899 LLVMValueRef v = LLVMBuildICmp(builder, LLVMIntNE, arg0,
900 bld_base->uint_bld.zero, "");
902 emit_data->output[emit_data->chan] =
903 LLVMBuildSelect(builder, v, emit_data->args[1], emit_data->args[2], "");
906 static void emit_cmp(
907 const struct lp_build_tgsi_action *action,
908 struct lp_build_tgsi_context * bld_base,
909 struct lp_build_emit_data * emit_data)
911 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
912 LLVMRealPredicate pred;
915 /* Use ordered for everything but NE (which is usual for
918 switch (emit_data->inst->Instruction.Opcode) {
919 case TGSI_OPCODE_SGE: pred = LLVMRealOGE; break;
920 case TGSI_OPCODE_SEQ: pred = LLVMRealOEQ; break;
921 case TGSI_OPCODE_SLE: pred = LLVMRealOLE; break;
922 case TGSI_OPCODE_SLT: pred = LLVMRealOLT; break;
923 case TGSI_OPCODE_SNE: pred = LLVMRealUNE; break;
924 case TGSI_OPCODE_SGT: pred = LLVMRealOGT; break;
925 default: assert(!"unknown instruction"); pred = 0; break;
928 cond = LLVMBuildFCmp(builder,
929 pred, emit_data->args[0], emit_data->args[1], "");
931 emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
932 cond, bld_base->base.one, bld_base->base.zero, "");
935 static void emit_fcmp(
936 const struct lp_build_tgsi_action *action,
937 struct lp_build_tgsi_context * bld_base,
938 struct lp_build_emit_data * emit_data)
940 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
941 LLVMContextRef context = bld_base->base.gallivm->context;
942 LLVMRealPredicate pred;
944 /* Use ordered for everything but NE (which is usual for
947 switch (emit_data->inst->Instruction.Opcode) {
948 case TGSI_OPCODE_FSEQ: pred = LLVMRealOEQ; break;
949 case TGSI_OPCODE_FSGE: pred = LLVMRealOGE; break;
950 case TGSI_OPCODE_FSLT: pred = LLVMRealOLT; break;
951 case TGSI_OPCODE_FSNE: pred = LLVMRealUNE; break;
952 default: assert(!"unknown instruction"); pred = 0; break;
955 LLVMValueRef v = LLVMBuildFCmp(builder, pred,
956 emit_data->args[0], emit_data->args[1],"");
958 v = LLVMBuildSExtOrBitCast(builder, v,
959 LLVMInt32TypeInContext(context), "");
961 emit_data->output[emit_data->chan] = v;
964 static void emit_not(
965 const struct lp_build_tgsi_action * action,
966 struct lp_build_tgsi_context * bld_base,
967 struct lp_build_emit_data * emit_data)
969 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
970 LLVMValueRef v = bitcast(bld_base, TGSI_TYPE_UNSIGNED,
972 emit_data->output[emit_data->chan] = LLVMBuildNot(builder, v, "");
975 static void emit_arl(
976 const struct lp_build_tgsi_action * action,
977 struct lp_build_tgsi_context * bld_base,
978 struct lp_build_emit_data * emit_data)
980 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
981 LLVMValueRef floor_index = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_FLR, emit_data->args[0]);
982 emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
983 floor_index, bld_base->base.int_elem_type , "");
986 static void emit_and(
987 const struct lp_build_tgsi_action * action,
988 struct lp_build_tgsi_context * bld_base,
989 struct lp_build_emit_data * emit_data)
991 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
992 emit_data->output[emit_data->chan] = LLVMBuildAnd(builder,
993 emit_data->args[0], emit_data->args[1], "");
997 const struct lp_build_tgsi_action * action,
998 struct lp_build_tgsi_context * bld_base,
999 struct lp_build_emit_data * emit_data)
1001 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1002 emit_data->output[emit_data->chan] = LLVMBuildOr(builder,
1003 emit_data->args[0], emit_data->args[1], "");
1006 static void emit_uadd(
1007 const struct lp_build_tgsi_action * action,
1008 struct lp_build_tgsi_context * bld_base,
1009 struct lp_build_emit_data * emit_data)
1011 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1012 emit_data->output[emit_data->chan] = LLVMBuildAdd(builder,
1013 emit_data->args[0], emit_data->args[1], "");
1016 static void emit_udiv(
1017 const struct lp_build_tgsi_action * action,
1018 struct lp_build_tgsi_context * bld_base,
1019 struct lp_build_emit_data * emit_data)
1021 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1022 emit_data->output[emit_data->chan] = LLVMBuildUDiv(builder,
1023 emit_data->args[0], emit_data->args[1], "");
1026 static void emit_idiv(
1027 const struct lp_build_tgsi_action * action,
1028 struct lp_build_tgsi_context * bld_base,
1029 struct lp_build_emit_data * emit_data)
1031 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1032 emit_data->output[emit_data->chan] = LLVMBuildSDiv(builder,
1033 emit_data->args[0], emit_data->args[1], "");
1036 static void emit_mod(
1037 const struct lp_build_tgsi_action * action,
1038 struct lp_build_tgsi_context * bld_base,
1039 struct lp_build_emit_data * emit_data)
1041 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1042 emit_data->output[emit_data->chan] = LLVMBuildSRem(builder,
1043 emit_data->args[0], emit_data->args[1], "");
1046 static void emit_umod(
1047 const struct lp_build_tgsi_action * action,
1048 struct lp_build_tgsi_context * bld_base,
1049 struct lp_build_emit_data * emit_data)
1051 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1052 emit_data->output[emit_data->chan] = LLVMBuildURem(builder,
1053 emit_data->args[0], emit_data->args[1], "");
1056 static void emit_shl(
1057 const struct lp_build_tgsi_action * action,
1058 struct lp_build_tgsi_context * bld_base,
1059 struct lp_build_emit_data * emit_data)
1061 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1062 emit_data->output[emit_data->chan] = LLVMBuildShl(builder,
1063 emit_data->args[0], emit_data->args[1], "");
1066 static void emit_ushr(
1067 const struct lp_build_tgsi_action * action,
1068 struct lp_build_tgsi_context * bld_base,
1069 struct lp_build_emit_data * emit_data)
1071 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1072 emit_data->output[emit_data->chan] = LLVMBuildLShr(builder,
1073 emit_data->args[0], emit_data->args[1], "");
1075 static void emit_ishr(
1076 const struct lp_build_tgsi_action * action,
1077 struct lp_build_tgsi_context * bld_base,
1078 struct lp_build_emit_data * emit_data)
1080 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1081 emit_data->output[emit_data->chan] = LLVMBuildAShr(builder,
1082 emit_data->args[0], emit_data->args[1], "");
1085 static void emit_xor(
1086 const struct lp_build_tgsi_action * action,
1087 struct lp_build_tgsi_context * bld_base,
1088 struct lp_build_emit_data * emit_data)
1090 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1091 emit_data->output[emit_data->chan] = LLVMBuildXor(builder,
1092 emit_data->args[0], emit_data->args[1], "");
1095 static void emit_ssg(
1096 const struct lp_build_tgsi_action * action,
1097 struct lp_build_tgsi_context * bld_base,
1098 struct lp_build_emit_data * emit_data)
1100 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1102 LLVMValueRef cmp, val;
1104 if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_ISSG) {
1105 cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int_bld.zero, "");
1106 val = LLVMBuildSelect(builder, cmp, bld_base->int_bld.one, emit_data->args[0], "");
1107 cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
1108 val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
1109 } else { // float SSG
1110 cmp = LLVMBuildFCmp(builder, LLVMRealOGT, emit_data->args[0], bld_base->base.zero, "");
1111 val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
1112 cmp = LLVMBuildFCmp(builder, LLVMRealOGE, val, bld_base->base.zero, "");
1113 val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
1116 emit_data->output[emit_data->chan] = val;
1119 static void emit_ineg(
1120 const struct lp_build_tgsi_action * action,
1121 struct lp_build_tgsi_context * bld_base,
1122 struct lp_build_emit_data * emit_data)
1124 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1125 emit_data->output[emit_data->chan] = LLVMBuildNeg(builder,
1126 emit_data->args[0], "");
1129 static void emit_f2i(
1130 const struct lp_build_tgsi_action * action,
1131 struct lp_build_tgsi_context * bld_base,
1132 struct lp_build_emit_data * emit_data)
1134 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1135 emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
1136 emit_data->args[0], bld_base->int_bld.elem_type, "");
1139 static void emit_f2u(
1140 const struct lp_build_tgsi_action * action,
1141 struct lp_build_tgsi_context * bld_base,
1142 struct lp_build_emit_data * emit_data)
1144 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1145 emit_data->output[emit_data->chan] = LLVMBuildFPToUI(builder,
1146 emit_data->args[0], bld_base->uint_bld.elem_type, "");
1149 static void emit_i2f(
1150 const struct lp_build_tgsi_action * action,
1151 struct lp_build_tgsi_context * bld_base,
1152 struct lp_build_emit_data * emit_data)
1154 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1155 emit_data->output[emit_data->chan] = LLVMBuildSIToFP(builder,
1156 emit_data->args[0], bld_base->base.elem_type, "");
1159 static void emit_u2f(
1160 const struct lp_build_tgsi_action * action,
1161 struct lp_build_tgsi_context * bld_base,
1162 struct lp_build_emit_data * emit_data)
1164 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1165 emit_data->output[emit_data->chan] = LLVMBuildUIToFP(builder,
1166 emit_data->args[0], bld_base->base.elem_type, "");
1169 static void emit_immediate(struct lp_build_tgsi_context * bld_base,
1170 const struct tgsi_full_immediate *imm)
1173 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
1175 for (i = 0; i < 4; ++i) {
1176 ctx->soa.immediates[ctx->soa.num_immediates][i] =
1177 LLVMConstInt(bld_base->uint_bld.elem_type, imm->u[i].Uint, false );
1180 ctx->soa.num_immediates++;
1184 build_intrinsic(LLVMBuilderRef builder,
1186 LLVMTypeRef ret_type,
1191 LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
1192 LLVMValueRef function;
1194 function = LLVMGetNamedFunction(module, name);
1196 LLVMTypeRef arg_types[LP_MAX_FUNC_ARGS];
1199 assert(num_args <= LP_MAX_FUNC_ARGS);
1201 for(i = 0; i < num_args; ++i) {
1203 arg_types[i] = LLVMTypeOf(args[i]);
1206 function = lp_declare_intrinsic(module, name, ret_type, arg_types, num_args);
1209 LLVMAddFunctionAttr(function, attr);
1212 return LLVMBuildCall(builder, function, args, num_args, "");
1215 static void build_tgsi_intrinsic(
1216 const struct lp_build_tgsi_action * action,
1217 struct lp_build_tgsi_context * bld_base,
1218 struct lp_build_emit_data * emit_data,
1221 struct lp_build_context * base = &bld_base->base;
1222 emit_data->output[emit_data->chan] = build_intrinsic(
1223 base->gallivm->builder, action->intr_name,
1224 emit_data->dst_type, emit_data->args,
1225 emit_data->arg_count, attr);
1228 build_tgsi_intrinsic_nomem(
1229 const struct lp_build_tgsi_action * action,
1230 struct lp_build_tgsi_context * bld_base,
1231 struct lp_build_emit_data * emit_data)
1233 build_tgsi_intrinsic(action, bld_base, emit_data, LLVMReadNoneAttribute);
1236 static void build_tgsi_intrinsic_readonly(
1237 const struct lp_build_tgsi_action * action,
1238 struct lp_build_tgsi_context * bld_base,
1239 struct lp_build_emit_data * emit_data)
1241 build_tgsi_intrinsic(action, bld_base, emit_data, LLVMReadOnlyAttribute);
1244 void radeon_llvm_context_init(struct radeon_llvm_context * ctx)
1246 struct lp_type type;
1248 /* Initialize the gallivm object:
1249 * We are only using the module, context, and builder fields of this struct.
1250 * This should be enough for us to be able to pass our gallivm struct to the
1251 * helper functions in the gallivm module.
1253 memset(&ctx->gallivm, 0, sizeof (ctx->gallivm));
1254 memset(&ctx->soa, 0, sizeof(ctx->soa));
1255 ctx->gallivm.context = LLVMContextCreate();
1256 ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
1257 ctx->gallivm.context);
1258 ctx->gallivm.builder = LLVMCreateBuilderInContext(ctx->gallivm.context);
1260 ctx->store_output_intr = "llvm.AMDGPU.store.output.";
1261 ctx->swizzle_intr = "llvm.AMDGPU.swizzle";
1262 struct lp_build_tgsi_context * bld_base = &ctx->soa.bld_base;
1264 /* XXX: We need to revisit this.I think the correct way to do this is
1265 * to use length = 4 here and use the elem_bld for everything. */
1266 type.floating = TRUE;
1273 lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
1274 lp_build_context_init(&ctx->soa.bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
1275 lp_build_context_init(&ctx->soa.bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
1278 bld_base->emit_store = emit_store;
1279 bld_base->emit_swizzle = emit_swizzle;
1280 bld_base->emit_declaration = emit_declaration;
1281 bld_base->emit_immediate = emit_immediate;
1283 bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch;
1284 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch;
1285 bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch;
1286 bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = emit_fetch;
1287 bld_base->emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
1289 /* Allocate outputs */
1290 ctx->soa.outputs = ctx->outputs;
1292 ctx->num_arrays = 0;
1294 /* XXX: Is there a better way to initialize all this ? */
1296 lp_set_default_actions(bld_base);
1298 bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_readonly;
1299 bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "fabs";
1300 bld_base->op_actions[TGSI_OPCODE_ARL].emit = emit_arl;
1301 bld_base->op_actions[TGSI_OPCODE_AND].emit = emit_and;
1302 bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
1303 bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
1304 bld_base->op_actions[TGSI_OPCODE_CEIL].emit = build_tgsi_intrinsic_readonly;
1305 bld_base->op_actions[TGSI_OPCODE_CEIL].intr_name = "ceil";
1306 bld_base->op_actions[TGSI_OPCODE_CLAMP].emit = build_tgsi_intrinsic_nomem;
1307 bld_base->op_actions[TGSI_OPCODE_CLAMP].intr_name = "llvm.AMDIL.clamp.";
1308 bld_base->op_actions[TGSI_OPCODE_CMP].emit = build_tgsi_intrinsic_nomem;
1309 bld_base->op_actions[TGSI_OPCODE_CMP].intr_name = "llvm.AMDGPU.cndlt";
1310 bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
1311 bld_base->op_actions[TGSI_OPCODE_COS].emit = build_tgsi_intrinsic_readonly;
1312 bld_base->op_actions[TGSI_OPCODE_COS].intr_name = "llvm.cos.f32";
1313 bld_base->op_actions[TGSI_OPCODE_DDX].intr_name = "llvm.AMDGPU.ddx";
1314 bld_base->op_actions[TGSI_OPCODE_DDX].fetch_args = tex_fetch_args;
1315 bld_base->op_actions[TGSI_OPCODE_DDY].intr_name = "llvm.AMDGPU.ddy";
1316 bld_base->op_actions[TGSI_OPCODE_DDY].fetch_args = tex_fetch_args;
1317 bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
1318 bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
1319 bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
1320 bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
1321 bld_base->op_actions[TGSI_OPCODE_EX2].intr_name = "llvm.AMDIL.exp.";
1322 bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_readonly;
1323 bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "floor";
1324 bld_base->op_actions[TGSI_OPCODE_FRC].emit = build_tgsi_intrinsic_nomem;
1325 bld_base->op_actions[TGSI_OPCODE_FRC].intr_name = "llvm.AMDIL.fraction.";
1326 bld_base->op_actions[TGSI_OPCODE_F2I].emit = emit_f2i;
1327 bld_base->op_actions[TGSI_OPCODE_F2U].emit = emit_f2u;
1328 bld_base->op_actions[TGSI_OPCODE_FSEQ].emit = emit_fcmp;
1329 bld_base->op_actions[TGSI_OPCODE_FSGE].emit = emit_fcmp;
1330 bld_base->op_actions[TGSI_OPCODE_FSLT].emit = emit_fcmp;
1331 bld_base->op_actions[TGSI_OPCODE_FSNE].emit = emit_fcmp;
1332 bld_base->op_actions[TGSI_OPCODE_IABS].emit = build_tgsi_intrinsic_nomem;
1333 bld_base->op_actions[TGSI_OPCODE_IABS].intr_name = "llvm.AMDIL.abs.";
1334 bld_base->op_actions[TGSI_OPCODE_IDIV].emit = emit_idiv;
1335 bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
1336 bld_base->op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
1337 bld_base->op_actions[TGSI_OPCODE_IMAX].emit = build_tgsi_intrinsic_nomem;
1338 bld_base->op_actions[TGSI_OPCODE_IMAX].intr_name = "llvm.AMDGPU.imax";
1339 bld_base->op_actions[TGSI_OPCODE_IMIN].emit = build_tgsi_intrinsic_nomem;
1340 bld_base->op_actions[TGSI_OPCODE_IMIN].intr_name = "llvm.AMDGPU.imin";
1341 bld_base->op_actions[TGSI_OPCODE_INEG].emit = emit_ineg;
1342 bld_base->op_actions[TGSI_OPCODE_ISHR].emit = emit_ishr;
1343 bld_base->op_actions[TGSI_OPCODE_ISGE].emit = emit_icmp;
1344 bld_base->op_actions[TGSI_OPCODE_ISLT].emit = emit_icmp;
1345 bld_base->op_actions[TGSI_OPCODE_ISSG].emit = emit_ssg;
1346 bld_base->op_actions[TGSI_OPCODE_I2F].emit = emit_i2f;
1347 bld_base->op_actions[TGSI_OPCODE_KILL_IF].emit = kil_emit;
1348 bld_base->op_actions[TGSI_OPCODE_KILL_IF].intr_name = "llvm.AMDGPU.kill";
1349 bld_base->op_actions[TGSI_OPCODE_KILL].emit = lp_build_tgsi_intrinsic;
1350 bld_base->op_actions[TGSI_OPCODE_KILL].intr_name = "llvm.AMDGPU.kilp";
1351 bld_base->op_actions[TGSI_OPCODE_LG2].emit = build_tgsi_intrinsic_readonly;
1352 bld_base->op_actions[TGSI_OPCODE_LG2].intr_name = "llvm.log2.f32";
1353 bld_base->op_actions[TGSI_OPCODE_LRP].emit = build_tgsi_intrinsic_nomem;
1354 bld_base->op_actions[TGSI_OPCODE_LRP].intr_name = "llvm.AMDGPU.lrp";
1355 bld_base->op_actions[TGSI_OPCODE_MOD].emit = emit_mod;
1356 bld_base->op_actions[TGSI_OPCODE_NOT].emit = emit_not;
1357 bld_base->op_actions[TGSI_OPCODE_OR].emit = emit_or;
1358 bld_base->op_actions[TGSI_OPCODE_POW].emit = build_tgsi_intrinsic_readonly;
1359 bld_base->op_actions[TGSI_OPCODE_POW].intr_name = "llvm.pow.f32";
1360 bld_base->op_actions[TGSI_OPCODE_ROUND].emit = build_tgsi_intrinsic_nomem;
1361 bld_base->op_actions[TGSI_OPCODE_ROUND].intr_name = "llvm.AMDIL.round.nearest.";
1362 bld_base->op_actions[TGSI_OPCODE_SGE].emit = emit_cmp;
1363 bld_base->op_actions[TGSI_OPCODE_SEQ].emit = emit_cmp;
1364 bld_base->op_actions[TGSI_OPCODE_SHL].emit = emit_shl;
1365 bld_base->op_actions[TGSI_OPCODE_SLE].emit = emit_cmp;
1366 bld_base->op_actions[TGSI_OPCODE_SLT].emit = emit_cmp;
1367 bld_base->op_actions[TGSI_OPCODE_SNE].emit = emit_cmp;
1368 bld_base->op_actions[TGSI_OPCODE_SGT].emit = emit_cmp;
1369 bld_base->op_actions[TGSI_OPCODE_SIN].emit = build_tgsi_intrinsic_readonly;
1370 bld_base->op_actions[TGSI_OPCODE_SIN].intr_name = "llvm.sin.f32";
1371 bld_base->op_actions[TGSI_OPCODE_SSG].emit = emit_ssg;
1372 bld_base->op_actions[TGSI_OPCODE_TEX].fetch_args = tex_fetch_args;
1373 bld_base->op_actions[TGSI_OPCODE_TEX].intr_name = "llvm.AMDGPU.tex";
1374 bld_base->op_actions[TGSI_OPCODE_TEX2].fetch_args = tex_fetch_args;
1375 bld_base->op_actions[TGSI_OPCODE_TEX2].intr_name = "llvm.AMDGPU.tex";
1376 bld_base->op_actions[TGSI_OPCODE_TXB].fetch_args = tex_fetch_args;
1377 bld_base->op_actions[TGSI_OPCODE_TXB].intr_name = "llvm.AMDGPU.txb";
1378 bld_base->op_actions[TGSI_OPCODE_TXB2].fetch_args = tex_fetch_args;
1379 bld_base->op_actions[TGSI_OPCODE_TXB2].intr_name = "llvm.AMDGPU.txb";
1380 bld_base->op_actions[TGSI_OPCODE_TXD].fetch_args = txd_fetch_args;
1381 bld_base->op_actions[TGSI_OPCODE_TXD].intr_name = "llvm.AMDGPU.txd";
1382 bld_base->op_actions[TGSI_OPCODE_TXF].fetch_args = txf_fetch_args;
1383 bld_base->op_actions[TGSI_OPCODE_TXF].intr_name = "llvm.AMDGPU.txf";
1384 bld_base->op_actions[TGSI_OPCODE_TXL].fetch_args = tex_fetch_args;
1385 bld_base->op_actions[TGSI_OPCODE_TXL].intr_name = "llvm.AMDGPU.txl";
1386 bld_base->op_actions[TGSI_OPCODE_TXL2].fetch_args = tex_fetch_args;
1387 bld_base->op_actions[TGSI_OPCODE_TXL2].intr_name = "llvm.AMDGPU.txl";
1388 bld_base->op_actions[TGSI_OPCODE_TXP].fetch_args = txp_fetch_args;
1389 bld_base->op_actions[TGSI_OPCODE_TXP].intr_name = "llvm.AMDGPU.tex";
1390 bld_base->op_actions[TGSI_OPCODE_TXQ].fetch_args = tex_fetch_args;
1391 bld_base->op_actions[TGSI_OPCODE_TXQ].intr_name = "llvm.AMDGPU.txq";
1392 bld_base->op_actions[TGSI_OPCODE_TRUNC].emit = build_tgsi_intrinsic_nomem;
1393 bld_base->op_actions[TGSI_OPCODE_TRUNC].intr_name = "llvm.AMDGPU.trunc";
1394 bld_base->op_actions[TGSI_OPCODE_UADD].emit = emit_uadd;
1395 bld_base->op_actions[TGSI_OPCODE_UDIV].emit = emit_udiv;
1396 bld_base->op_actions[TGSI_OPCODE_UMAX].emit = build_tgsi_intrinsic_nomem;
1397 bld_base->op_actions[TGSI_OPCODE_UMAX].intr_name = "llvm.AMDGPU.umax";
1398 bld_base->op_actions[TGSI_OPCODE_UMIN].emit = build_tgsi_intrinsic_nomem;
1399 bld_base->op_actions[TGSI_OPCODE_UMIN].intr_name = "llvm.AMDGPU.umin";
1400 bld_base->op_actions[TGSI_OPCODE_UMOD].emit = emit_umod;
1401 bld_base->op_actions[TGSI_OPCODE_USEQ].emit = emit_icmp;
1402 bld_base->op_actions[TGSI_OPCODE_USGE].emit = emit_icmp;
1403 bld_base->op_actions[TGSI_OPCODE_USHR].emit = emit_ushr;
1404 bld_base->op_actions[TGSI_OPCODE_USLT].emit = emit_icmp;
1405 bld_base->op_actions[TGSI_OPCODE_USNE].emit = emit_icmp;
1406 bld_base->op_actions[TGSI_OPCODE_U2F].emit = emit_u2f;
1407 bld_base->op_actions[TGSI_OPCODE_XOR].emit = emit_xor;
1408 bld_base->op_actions[TGSI_OPCODE_UCMP].emit = emit_ucmp;
1410 bld_base->rsq_action.emit = build_tgsi_intrinsic_nomem;
1411 #if HAVE_LLVM >= 0x0305
1412 bld_base->rsq_action.intr_name = "llvm.AMDGPU.rsq.clamped.f32";
1414 bld_base->rsq_action.intr_name = "llvm.AMDGPU.rsq";
1418 void radeon_llvm_create_func(struct radeon_llvm_context * ctx,
1419 LLVMTypeRef *ParamTypes, unsigned ParamCount)
1421 LLVMTypeRef main_fn_type;
1422 LLVMBasicBlockRef main_fn_body;
1424 /* Setup the function */
1425 main_fn_type = LLVMFunctionType(LLVMVoidTypeInContext(ctx->gallivm.context),
1426 ParamTypes, ParamCount, 0);
1427 ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, "main", main_fn_type);
1428 main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
1429 ctx->main_fn, "main_body");
1430 LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
1433 void radeon_llvm_finalize_module(struct radeon_llvm_context * ctx)
1435 struct gallivm_state * gallivm = ctx->soa.bld_base.base.gallivm;
1436 /* End the main function with Return*/
1437 LLVMBuildRetVoid(gallivm->builder);
1439 /* Create the pass manager */
1440 ctx->gallivm.passmgr = LLVMCreateFunctionPassManagerForModule(
1443 /* This pass should eliminate all the load and store instructions */
1444 LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
1446 /* Add some optimization passes */
1447 LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
1448 LLVMAddLICMPass(gallivm->passmgr);
1449 LLVMAddAggressiveDCEPass(gallivm->passmgr);
1450 LLVMAddCFGSimplificationPass(gallivm->passmgr);
1453 LLVMRunFunctionPassManager(gallivm->passmgr, ctx->main_fn);
1455 LLVMDisposeBuilder(gallivm->builder);
1456 LLVMDisposePassManager(gallivm->passmgr);
1460 void radeon_llvm_dispose(struct radeon_llvm_context * ctx)
1462 LLVMDisposeModule(ctx->soa.bld_base.base.gallivm->module);
1463 LLVMContextDispose(ctx->soa.bld_base.base.gallivm->context);
1468 ctx->loop_depth_max = 0;
1471 ctx->branch_depth_max = 0;