2 * Copyright 2016 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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 #include "si_shader_internal.h"
27 #include "gallivm/lp_bld_const.h"
28 #include "gallivm/lp_bld_gather.h"
29 #include "gallivm/lp_bld_flow.h"
30 #include "gallivm/lp_bld_init.h"
31 #include "gallivm/lp_bld_intr.h"
32 #include "gallivm/lp_bld_misc.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"
41 #include <llvm-c/Transforms/IPO.h>
42 #include <llvm-c/Transforms/Scalar.h>
43 #include <llvm-c/Support.h>
45 /* Data for if/else/endif and bgnloop/endloop control flow structures.
48 /* Loop exit or next part of if/else/endif. */
49 LLVMBasicBlockRef next_block
;
50 LLVMBasicBlockRef loop_entry_block
;
53 #define CPU_STRING_LEN 30
54 #define FS_STRING_LEN 30
55 #define TRIPLE_STRING_LEN 7
58 * Shader types for the LLVM backend.
60 enum si_llvm_shader_type
{
61 RADEON_LLVM_SHADER_PS
= 0,
62 RADEON_LLVM_SHADER_VS
= 1,
63 RADEON_LLVM_SHADER_GS
= 2,
64 RADEON_LLVM_SHADER_CS
= 3,
67 enum si_llvm_calling_convention
{
68 RADEON_LLVM_AMDGPU_VS
= 87,
69 RADEON_LLVM_AMDGPU_GS
= 88,
70 RADEON_LLVM_AMDGPU_PS
= 89,
71 RADEON_LLVM_AMDGPU_CS
= 90,
74 void si_llvm_add_attribute(LLVMValueRef F
, const char *name
, int value
)
78 snprintf(str
, sizeof(str
), "%i", value
);
79 LLVMAddTargetDependentFunctionAttr(F
, name
, str
);
83 * Set the shader type we want to compile
85 * @param type shader type to set
87 void si_llvm_shader_type(LLVMValueRef F
, unsigned type
)
89 enum si_llvm_shader_type llvm_type
;
90 enum si_llvm_calling_convention calling_conv
;
93 case PIPE_SHADER_VERTEX
:
94 case PIPE_SHADER_TESS_CTRL
:
95 case PIPE_SHADER_TESS_EVAL
:
96 llvm_type
= RADEON_LLVM_SHADER_VS
;
97 calling_conv
= RADEON_LLVM_AMDGPU_VS
;
99 case PIPE_SHADER_GEOMETRY
:
100 llvm_type
= RADEON_LLVM_SHADER_GS
;
101 calling_conv
= RADEON_LLVM_AMDGPU_GS
;
103 case PIPE_SHADER_FRAGMENT
:
104 llvm_type
= RADEON_LLVM_SHADER_PS
;
105 calling_conv
= RADEON_LLVM_AMDGPU_PS
;
107 case PIPE_SHADER_COMPUTE
:
108 llvm_type
= RADEON_LLVM_SHADER_CS
;
109 calling_conv
= RADEON_LLVM_AMDGPU_CS
;
112 unreachable("Unhandle shader type");
115 if (HAVE_LLVM
>= 0x309)
116 LLVMSetFunctionCallConv(F
, calling_conv
);
118 si_llvm_add_attribute(F
, "ShaderType", llvm_type
);
121 static void init_amdgpu_target()
123 gallivm_init_llvm_targets();
124 LLVMInitializeAMDGPUTargetInfo();
125 LLVMInitializeAMDGPUTarget();
126 LLVMInitializeAMDGPUTargetMC();
127 LLVMInitializeAMDGPUAsmPrinter();
129 /* For inline assembly. */
130 LLVMInitializeAMDGPUAsmParser();
132 if (HAVE_LLVM
>= 0x0400) {
134 * Workaround for bug in llvm 4.0 that causes image intrinsics
136 * https://reviews.llvm.org/D26348
138 const char *argv
[2] = {"mesa", "-simplifycfg-sink-common=false"};
139 LLVMParseCommandLineOptions(2, argv
, NULL
);
143 static once_flag init_amdgpu_target_once_flag
= ONCE_FLAG_INIT
;
145 LLVMTargetRef
si_llvm_get_amdgpu_target(const char *triple
)
147 LLVMTargetRef target
= NULL
;
148 char *err_message
= NULL
;
150 call_once(&init_amdgpu_target_once_flag
, init_amdgpu_target
);
152 if (LLVMGetTargetFromTriple(triple
, &target
, &err_message
)) {
153 fprintf(stderr
, "Cannot find target for triple %s ", triple
);
155 fprintf(stderr
, "%s\n", err_message
);
157 LLVMDisposeMessage(err_message
);
163 struct si_llvm_diagnostics
{
164 struct pipe_debug_callback
*debug
;
168 static void si_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
170 struct si_llvm_diagnostics
*diag
= (struct si_llvm_diagnostics
*)context
;
171 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
172 char *description
= LLVMGetDiagInfoDescription(di
);
173 const char *severity_str
= NULL
;
177 severity_str
= "error";
180 severity_str
= "warning";
183 severity_str
= "remark";
186 severity_str
= "note";
189 severity_str
= "unknown";
192 pipe_debug_message(diag
->debug
, SHADER_INFO
,
193 "LLVM diagnostic (%s): %s", severity_str
, description
);
195 if (severity
== LLVMDSError
) {
197 fprintf(stderr
,"LLVM triggered Diagnostic Handler: %s\n", description
);
200 LLVMDisposeMessage(description
);
204 * Compile an LLVM module to machine code.
206 * @returns 0 for success, 1 for failure
208 unsigned si_llvm_compile(LLVMModuleRef M
, struct ac_shader_binary
*binary
,
209 LLVMTargetMachineRef tm
,
210 struct pipe_debug_callback
*debug
)
212 struct si_llvm_diagnostics diag
;
214 LLVMContextRef llvm_ctx
;
215 LLVMMemoryBufferRef out_buffer
;
216 unsigned buffer_size
;
217 const char *buffer_data
;
223 /* Setup Diagnostic Handler*/
224 llvm_ctx
= LLVMGetModuleContext(M
);
226 LLVMContextSetDiagnosticHandler(llvm_ctx
, si_diagnostic_handler
, &diag
);
229 mem_err
= LLVMTargetMachineEmitToMemoryBuffer(tm
, M
, LLVMObjectFile
, &err
,
232 /* Process Errors/Warnings */
234 fprintf(stderr
, "%s: %s", __FUNCTION__
, err
);
235 pipe_debug_message(debug
, SHADER_INFO
,
236 "LLVM emit error: %s", err
);
242 /* Extract Shader Code*/
243 buffer_size
= LLVMGetBufferSize(out_buffer
);
244 buffer_data
= LLVMGetBufferStart(out_buffer
);
246 ac_elf_read(buffer_data
, buffer_size
, binary
);
249 LLVMDisposeMemoryBuffer(out_buffer
);
252 if (diag
.retval
!= 0)
253 pipe_debug_message(debug
, SHADER_INFO
, "LLVM compile failed");
257 LLVMTypeRef
tgsi2llvmtype(struct lp_build_tgsi_context
*bld_base
,
258 enum tgsi_opcode_type type
)
260 LLVMContextRef ctx
= bld_base
->base
.gallivm
->context
;
263 case TGSI_TYPE_UNSIGNED
:
264 case TGSI_TYPE_SIGNED
:
265 return LLVMInt32TypeInContext(ctx
);
266 case TGSI_TYPE_UNSIGNED64
:
267 case TGSI_TYPE_SIGNED64
:
268 return LLVMInt64TypeInContext(ctx
);
269 case TGSI_TYPE_DOUBLE
:
270 return LLVMDoubleTypeInContext(ctx
);
271 case TGSI_TYPE_UNTYPED
:
272 case TGSI_TYPE_FLOAT
:
273 return LLVMFloatTypeInContext(ctx
);
279 LLVMValueRef
bitcast(struct lp_build_tgsi_context
*bld_base
,
280 enum tgsi_opcode_type type
, LLVMValueRef value
)
282 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
283 LLVMTypeRef dst_type
= tgsi2llvmtype(bld_base
, type
);
286 return LLVMBuildBitCast(builder
, value
, dst_type
, "");
292 * Return a value that is equal to the given i32 \p index if it lies in [0,num)
293 * or an undefined value in the same interval otherwise.
295 LLVMValueRef
si_llvm_bound_index(struct si_shader_context
*ctx
,
299 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
300 LLVMBuilderRef builder
= gallivm
->builder
;
301 LLVMValueRef c_max
= LLVMConstInt(ctx
->i32
, num
- 1, 0);
304 if (util_is_power_of_two(num
)) {
305 index
= LLVMBuildAnd(builder
, index
, c_max
, "");
307 /* In theory, this MAX pattern should result in code that is
308 * as good as the bit-wise AND above.
310 * In practice, LLVM generates worse code (at the time of
311 * writing), because its value tracking is not strong enough.
313 cc
= LLVMBuildICmp(builder
, LLVMIntULE
, index
, c_max
, "");
314 index
= LLVMBuildSelect(builder
, cc
, index
, c_max
, "");
320 static struct si_llvm_flow
*
321 get_current_flow(struct si_shader_context
*ctx
)
323 if (ctx
->flow_depth
> 0)
324 return &ctx
->flow
[ctx
->flow_depth
- 1];
328 static struct si_llvm_flow
*
329 get_innermost_loop(struct si_shader_context
*ctx
)
331 for (unsigned i
= ctx
->flow_depth
; i
> 0; --i
) {
332 if (ctx
->flow
[i
- 1].loop_entry_block
)
333 return &ctx
->flow
[i
- 1];
338 static struct si_llvm_flow
*
339 push_flow(struct si_shader_context
*ctx
)
341 struct si_llvm_flow
*flow
;
343 if (ctx
->flow_depth
>= ctx
->flow_depth_max
) {
344 unsigned new_max
= MAX2(ctx
->flow_depth
<< 1, RADEON_LLVM_INITIAL_CF_DEPTH
);
345 ctx
->flow
= REALLOC(ctx
->flow
,
346 ctx
->flow_depth_max
* sizeof(*ctx
->flow
),
347 new_max
* sizeof(*ctx
->flow
));
348 ctx
->flow_depth_max
= new_max
;
351 flow
= &ctx
->flow
[ctx
->flow_depth
];
354 flow
->next_block
= NULL
;
355 flow
->loop_entry_block
= NULL
;
359 static LLVMValueRef
emit_swizzle(struct lp_build_tgsi_context
*bld_base
,
366 LLVMValueRef swizzles
[4];
368 LLVMInt32TypeInContext(bld_base
->base
.gallivm
->context
);
370 swizzles
[0] = LLVMConstInt(i32t
, swizzle_x
, 0);
371 swizzles
[1] = LLVMConstInt(i32t
, swizzle_y
, 0);
372 swizzles
[2] = LLVMConstInt(i32t
, swizzle_z
, 0);
373 swizzles
[3] = LLVMConstInt(i32t
, swizzle_w
, 0);
375 return LLVMBuildShuffleVector(bld_base
->base
.gallivm
->builder
,
377 LLVMGetUndef(LLVMTypeOf(value
)),
378 LLVMConstVector(swizzles
, 4), "");
382 * Return the description of the array covering the given temporary register
386 get_temp_array_id(struct lp_build_tgsi_context
*bld_base
,
388 const struct tgsi_ind_register
*reg
)
390 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
391 unsigned num_arrays
= ctx
->bld_base
.info
->array_max
[TGSI_FILE_TEMPORARY
];
394 if (reg
&& reg
->ArrayID
> 0 && reg
->ArrayID
<= num_arrays
)
397 for (i
= 0; i
< num_arrays
; i
++) {
398 const struct tgsi_array_info
*array
= &ctx
->temp_arrays
[i
];
400 if (reg_index
>= array
->range
.First
&& reg_index
<= array
->range
.Last
)
407 static struct tgsi_declaration_range
408 get_array_range(struct lp_build_tgsi_context
*bld_base
,
409 unsigned File
, unsigned reg_index
,
410 const struct tgsi_ind_register
*reg
)
412 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
413 struct tgsi_declaration_range range
;
415 if (File
== TGSI_FILE_TEMPORARY
) {
416 unsigned array_id
= get_temp_array_id(bld_base
, reg_index
, reg
);
418 return ctx
->temp_arrays
[array_id
- 1].range
;
422 range
.Last
= bld_base
->info
->file_max
[File
];
427 emit_array_index(struct si_shader_context
*ctx
,
428 const struct tgsi_ind_register
*reg
,
431 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
434 return LLVMConstInt(ctx
->i32
, offset
, 0);
436 LLVMValueRef addr
= LLVMBuildLoad(gallivm
->builder
, ctx
->addrs
[reg
->Index
][reg
->Swizzle
], "");
437 return LLVMBuildAdd(gallivm
->builder
, addr
, LLVMConstInt(ctx
->i32
, offset
, 0), "");
441 * For indirect registers, construct a pointer directly to the requested
442 * element using getelementptr if possible.
444 * Returns NULL if the insertelement/extractelement fallback for array access
448 get_pointer_into_array(struct si_shader_context
*ctx
,
452 const struct tgsi_ind_register
*reg_indirect
)
455 struct tgsi_array_info
*array
;
456 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
457 LLVMBuilderRef builder
= gallivm
->builder
;
458 LLVMValueRef idxs
[2];
462 if (file
!= TGSI_FILE_TEMPORARY
)
465 array_id
= get_temp_array_id(&ctx
->bld_base
, reg_index
, reg_indirect
);
469 alloca
= ctx
->temp_array_allocas
[array_id
- 1];
473 array
= &ctx
->temp_arrays
[array_id
- 1];
475 if (!(array
->writemask
& (1 << swizzle
)))
476 return ctx
->undef_alloca
;
478 index
= emit_array_index(ctx
, reg_indirect
,
479 reg_index
- ctx
->temp_arrays
[array_id
- 1].range
.First
);
481 /* Ensure that the index is within a valid range, to guard against
482 * VM faults and overwriting critical data (e.g. spilled resource
485 * TODO It should be possible to avoid the additional instructions
486 * if LLVM is changed so that it guarantuees:
487 * 1. the scratch space descriptor isolates the current wave (this
488 * could even save the scratch offset SGPR at the cost of an
489 * additional SALU instruction)
490 * 2. the memory for allocas must be allocated at the _end_ of the
491 * scratch space (after spilled registers)
493 index
= si_llvm_bound_index(ctx
, index
, array
->range
.Last
- array
->range
.First
+ 1);
495 index
= LLVMBuildMul(
497 LLVMConstInt(ctx
->i32
, util_bitcount(array
->writemask
), 0),
499 index
= LLVMBuildAdd(
501 LLVMConstInt(ctx
->i32
,
502 util_bitcount(array
->writemask
& ((1 << swizzle
) - 1)), 0),
504 idxs
[0] = ctx
->i32_0
;
506 return LLVMBuildGEP(builder
, alloca
, idxs
, 2, "");
510 si_llvm_emit_fetch_64bit(struct lp_build_tgsi_context
*bld_base
,
511 enum tgsi_opcode_type type
,
515 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
518 result
= LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base
->base
.gallivm
->context
, 32), bld_base
->base
.type
.length
* 2));
520 result
= LLVMBuildInsertElement(builder
,
522 bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, ptr
),
523 bld_base
->int_bld
.zero
, "");
524 result
= LLVMBuildInsertElement(builder
,
526 bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, ptr2
),
527 bld_base
->int_bld
.one
, "");
528 return bitcast(bld_base
, type
, result
);
532 emit_array_fetch(struct lp_build_tgsi_context
*bld_base
,
533 unsigned File
, enum tgsi_opcode_type type
,
534 struct tgsi_declaration_range range
,
537 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
539 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
541 unsigned i
, size
= range
.Last
- range
.First
+ 1;
542 LLVMTypeRef vec
= LLVMVectorType(tgsi2llvmtype(bld_base
, type
), size
);
543 LLVMValueRef result
= LLVMGetUndef(vec
);
545 struct tgsi_full_src_register tmp_reg
= {};
546 tmp_reg
.Register
.File
= File
;
548 for (i
= 0; i
< size
; ++i
) {
549 tmp_reg
.Register
.Index
= i
+ range
.First
;
550 LLVMValueRef temp
= si_llvm_emit_fetch(bld_base
, &tmp_reg
, type
, swizzle
);
551 result
= LLVMBuildInsertElement(builder
, result
, temp
,
552 LLVMConstInt(ctx
->i32
, i
, 0), "array_vector");
558 load_value_from_array(struct lp_build_tgsi_context
*bld_base
,
560 enum tgsi_opcode_type type
,
563 const struct tgsi_ind_register
*reg_indirect
)
565 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
566 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
567 LLVMBuilderRef builder
= gallivm
->builder
;
570 ptr
= get_pointer_into_array(ctx
, file
, swizzle
, reg_index
, reg_indirect
);
572 LLVMValueRef val
= LLVMBuildLoad(builder
, ptr
, "");
573 if (tgsi_type_is_64bit(type
)) {
574 LLVMValueRef ptr_hi
, val_hi
;
575 ptr_hi
= LLVMBuildGEP(builder
, ptr
, &ctx
->i32_1
, 1, "");
576 val_hi
= LLVMBuildLoad(builder
, ptr_hi
, "");
577 val
= si_llvm_emit_fetch_64bit(bld_base
, type
, val
, val_hi
);
582 struct tgsi_declaration_range range
=
583 get_array_range(bld_base
, file
, reg_index
, reg_indirect
);
585 emit_array_index(ctx
, reg_indirect
, reg_index
- range
.First
);
587 emit_array_fetch(bld_base
, file
, type
, range
, swizzle
);
588 return LLVMBuildExtractElement(builder
, array
, index
, "");
593 store_value_to_array(struct lp_build_tgsi_context
*bld_base
,
598 const struct tgsi_ind_register
*reg_indirect
)
600 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
601 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
602 LLVMBuilderRef builder
= gallivm
->builder
;
605 ptr
= get_pointer_into_array(ctx
, file
, chan_index
, reg_index
, reg_indirect
);
607 LLVMBuildStore(builder
, value
, ptr
);
610 struct tgsi_declaration_range range
= get_array_range(bld_base
, file
, reg_index
, reg_indirect
);
611 LLVMValueRef index
= emit_array_index(ctx
, reg_indirect
, reg_index
- range
.First
);
613 emit_array_fetch(bld_base
, file
, TGSI_TYPE_FLOAT
, range
, chan_index
);
614 LLVMValueRef temp_ptr
;
616 array
= LLVMBuildInsertElement(builder
, array
, value
, index
, "");
618 size
= range
.Last
- range
.First
+ 1;
619 for (i
= 0; i
< size
; ++i
) {
621 case TGSI_FILE_OUTPUT
:
622 temp_ptr
= ctx
->outputs
[i
+ range
.First
][chan_index
];
625 case TGSI_FILE_TEMPORARY
:
626 if (range
.First
+ i
>= ctx
->temps_count
)
628 temp_ptr
= ctx
->temps
[(i
+ range
.First
) * TGSI_NUM_CHANNELS
+ chan_index
];
634 value
= LLVMBuildExtractElement(builder
, array
,
635 LLVMConstInt(ctx
->i32
, i
, 0), "");
636 LLVMBuildStore(builder
, value
, temp_ptr
);
641 /* If this is true, preload FS inputs at the beginning of shaders. Otherwise,
642 * reload them at each use. This must be true if the shader is using
643 * derivatives and KILL, because KILL can leave the WQM and then a lazy
644 * input load isn't in the WQM anymore.
646 static bool si_preload_fs_inputs(struct si_shader_context
*ctx
)
648 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
650 return sel
->info
.uses_derivatives
&&
655 get_output_ptr(struct lp_build_tgsi_context
*bld_base
, unsigned index
,
658 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
660 assert(index
<= ctx
->bld_base
.info
->file_max
[TGSI_FILE_OUTPUT
]);
661 return ctx
->outputs
[index
][chan
];
664 LLVMValueRef
si_llvm_emit_fetch(struct lp_build_tgsi_context
*bld_base
,
665 const struct tgsi_full_src_register
*reg
,
666 enum tgsi_opcode_type type
,
669 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
670 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
671 LLVMValueRef result
= NULL
, ptr
, ptr2
;
674 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
676 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
677 values
[chan
] = si_llvm_emit_fetch(bld_base
, reg
, type
, chan
);
679 return lp_build_gather_values(&ctx
->gallivm
, values
,
683 if (reg
->Register
.Indirect
) {
684 LLVMValueRef load
= load_value_from_array(bld_base
, reg
->Register
.File
, type
,
685 swizzle
, reg
->Register
.Index
, ®
->Indirect
);
686 return bitcast(bld_base
, type
, load
);
689 switch(reg
->Register
.File
) {
690 case TGSI_FILE_IMMEDIATE
: {
691 LLVMTypeRef ctype
= tgsi2llvmtype(bld_base
, type
);
692 if (tgsi_type_is_64bit(type
)) {
693 result
= LLVMGetUndef(LLVMVectorType(ctx
->i32
, bld_base
->base
.type
.length
* 2));
694 result
= LLVMConstInsertElement(result
,
695 ctx
->imms
[reg
->Register
.Index
* TGSI_NUM_CHANNELS
+ swizzle
],
697 result
= LLVMConstInsertElement(result
,
698 ctx
->imms
[reg
->Register
.Index
* TGSI_NUM_CHANNELS
+ swizzle
+ 1],
700 return LLVMConstBitCast(result
, ctype
);
702 return LLVMConstBitCast(ctx
->imms
[reg
->Register
.Index
* TGSI_NUM_CHANNELS
+ swizzle
], ctype
);
706 case TGSI_FILE_INPUT
: {
707 unsigned index
= reg
->Register
.Index
;
708 LLVMValueRef input
[4];
710 /* I don't think doing this for vertex shaders is beneficial.
711 * For those, we want to make sure the VMEM loads are executed
712 * only once. Fragment shaders don't care much, because
713 * v_interp instructions are much cheaper than VMEM loads.
715 if (!si_preload_fs_inputs(ctx
) &&
716 ctx
->bld_base
.info
->processor
== PIPE_SHADER_FRAGMENT
)
717 ctx
->load_input(ctx
, index
, &ctx
->input_decls
[index
], input
);
719 memcpy(input
, &ctx
->inputs
[index
* 4], sizeof(input
));
721 result
= input
[swizzle
];
723 if (tgsi_type_is_64bit(type
)) {
725 ptr2
= input
[swizzle
+ 1];
726 return si_llvm_emit_fetch_64bit(bld_base
, type
, ptr
, ptr2
);
731 case TGSI_FILE_TEMPORARY
:
732 if (reg
->Register
.Index
>= ctx
->temps_count
)
733 return LLVMGetUndef(tgsi2llvmtype(bld_base
, type
));
734 ptr
= ctx
->temps
[reg
->Register
.Index
* TGSI_NUM_CHANNELS
+ swizzle
];
735 if (tgsi_type_is_64bit(type
)) {
736 ptr2
= ctx
->temps
[reg
->Register
.Index
* TGSI_NUM_CHANNELS
+ swizzle
+ 1];
737 return si_llvm_emit_fetch_64bit(bld_base
, type
,
738 LLVMBuildLoad(builder
, ptr
, ""),
739 LLVMBuildLoad(builder
, ptr2
, ""));
741 result
= LLVMBuildLoad(builder
, ptr
, "");
744 case TGSI_FILE_OUTPUT
:
745 ptr
= get_output_ptr(bld_base
, reg
->Register
.Index
, swizzle
);
746 if (tgsi_type_is_64bit(type
)) {
747 ptr2
= get_output_ptr(bld_base
, reg
->Register
.Index
, swizzle
+ 1);
748 return si_llvm_emit_fetch_64bit(bld_base
, type
,
749 LLVMBuildLoad(builder
, ptr
, ""),
750 LLVMBuildLoad(builder
, ptr2
, ""));
752 result
= LLVMBuildLoad(builder
, ptr
, "");
756 return LLVMGetUndef(tgsi2llvmtype(bld_base
, type
));
759 return bitcast(bld_base
, type
, result
);
762 static LLVMValueRef
fetch_system_value(struct lp_build_tgsi_context
*bld_base
,
763 const struct tgsi_full_src_register
*reg
,
764 enum tgsi_opcode_type type
,
767 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
768 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
769 LLVMValueRef cval
= ctx
->system_values
[reg
->Register
.Index
];
771 if (tgsi_type_is_64bit(type
)) {
774 assert(swizzle
== 0 || swizzle
== 2);
776 lo
= LLVMBuildExtractElement(
777 builder
, cval
, LLVMConstInt(ctx
->i32
, swizzle
, 0), "");
778 hi
= LLVMBuildExtractElement(
779 builder
, cval
, LLVMConstInt(ctx
->i32
, swizzle
+ 1, 0), "");
781 return si_llvm_emit_fetch_64bit(bld_base
, type
, lo
, hi
);
784 if (LLVMGetTypeKind(LLVMTypeOf(cval
)) == LLVMVectorTypeKind
) {
785 cval
= LLVMBuildExtractElement(
786 builder
, cval
, LLVMConstInt(ctx
->i32
, swizzle
, 0), "");
788 assert(swizzle
== 0);
791 return bitcast(bld_base
, type
, cval
);
794 static void emit_declaration(struct lp_build_tgsi_context
*bld_base
,
795 const struct tgsi_full_declaration
*decl
)
797 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
798 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
799 unsigned first
, last
, i
;
800 switch(decl
->Declaration
.File
) {
801 case TGSI_FILE_ADDRESS
:
804 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
806 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
807 ctx
->addrs
[idx
][chan
] = lp_build_alloca_undef(
815 case TGSI_FILE_TEMPORARY
:
818 LLVMValueRef array_alloca
= NULL
;
820 unsigned writemask
= decl
->Declaration
.UsageMask
;
821 first
= decl
->Range
.First
;
822 last
= decl
->Range
.Last
;
823 decl_size
= 4 * ((last
- first
) + 1);
825 if (decl
->Declaration
.Array
) {
826 unsigned id
= decl
->Array
.ArrayID
- 1;
829 writemask
&= ctx
->temp_arrays
[id
].writemask
;
830 ctx
->temp_arrays
[id
].writemask
= writemask
;
831 array_size
= ((last
- first
) + 1) * util_bitcount(writemask
);
833 /* If the array has more than 16 elements, store it
834 * in memory using an alloca that spans the entire
837 * Otherwise, store each array element individually.
838 * We will then generate vectors (per-channel, up to
839 * <16 x float> if the usagemask is a single bit) for
840 * indirect addressing.
842 * Note that 16 is the number of vector elements that
843 * LLVM will store in a register, so theoretically an
844 * array with up to 4 * 16 = 64 elements could be
845 * handled this way, but whether that's a good idea
846 * depends on VGPR register pressure elsewhere.
848 * FIXME: We shouldn't need to have the non-alloca
849 * code path for arrays. LLVM should be smart enough to
850 * promote allocas into registers when profitable.
852 * LLVM 3.8 crashes with this.
854 if ((HAVE_LLVM
>= 0x0309 && array_size
> 16) ||
855 /* TODO: VGPR indexing is buggy on GFX9. */
856 ctx
->screen
->b
.chip_class
== GFX9
) {
857 array_alloca
= LLVMBuildAlloca(builder
,
858 LLVMArrayType(ctx
->f32
,
859 array_size
), "array");
860 ctx
->temp_array_allocas
[id
] = array_alloca
;
864 if (!ctx
->temps_count
) {
865 ctx
->temps_count
= bld_base
->info
->file_max
[TGSI_FILE_TEMPORARY
] + 1;
866 ctx
->temps
= MALLOC(TGSI_NUM_CHANNELS
* ctx
->temps_count
* sizeof(LLVMValueRef
));
869 for (i
= 0; i
< decl_size
; ++i
) {
871 snprintf(name
, sizeof(name
), "TEMP%d.%c",
872 first
+ i
/ 4, "xyzw"[i
% 4]);
874 ctx
->temps
[first
* TGSI_NUM_CHANNELS
+ i
] =
875 lp_build_alloca_undef(&ctx
->gallivm
,
880 LLVMValueRef idxs
[2] = {
886 if (writemask
!= TGSI_WRITEMASK_XYZW
&&
887 !ctx
->undef_alloca
) {
888 /* Create a dummy alloca. We use it so that we
889 * have a pointer that is safe to load from if
890 * a shader ever reads from a channel that
891 * it never writes to.
893 ctx
->undef_alloca
= lp_build_alloca_undef(
898 for (i
= 0; i
< decl_size
; ++i
) {
900 if (writemask
& (1 << (i
% 4))) {
902 snprintf(name
, sizeof(name
), "TEMP%d.%c",
903 first
+ i
/ 4, "xyzw"[i
% 4]);
905 idxs
[1] = LLVMConstInt(ctx
->i32
, j
, 0);
906 ptr
= LLVMBuildGEP(builder
, array_alloca
, idxs
, 2, name
);
909 ptr
= ctx
->undef_alloca
;
911 ctx
->temps
[first
* TGSI_NUM_CHANNELS
+ i
] = ptr
;
916 case TGSI_FILE_INPUT
:
919 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
920 if (ctx
->load_input
&&
921 ctx
->input_decls
[idx
].Declaration
.File
!= TGSI_FILE_INPUT
) {
922 ctx
->input_decls
[idx
] = *decl
;
923 ctx
->input_decls
[idx
].Range
.First
= idx
;
924 ctx
->input_decls
[idx
].Range
.Last
= idx
;
925 ctx
->input_decls
[idx
].Semantic
.Index
+= idx
- decl
->Range
.First
;
927 if (si_preload_fs_inputs(ctx
) ||
928 bld_base
->info
->processor
!= PIPE_SHADER_FRAGMENT
)
929 ctx
->load_input(ctx
, idx
, &ctx
->input_decls
[idx
],
930 &ctx
->inputs
[idx
* 4]);
936 case TGSI_FILE_SYSTEM_VALUE
:
939 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
940 ctx
->load_system_value(ctx
, idx
, decl
);
945 case TGSI_FILE_OUTPUT
:
949 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
951 assert(idx
< RADEON_LLVM_MAX_OUTPUTS
);
952 if (ctx
->outputs
[idx
][0])
954 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
956 snprintf(name
, sizeof(name
), "OUT%d.%c",
957 idx
, "xyzw"[chan
% 4]);
959 ctx
->outputs
[idx
][chan
] = lp_build_alloca_undef(
967 case TGSI_FILE_MEMORY
:
968 ctx
->declare_memory_region(ctx
, decl
);
976 void si_llvm_emit_store(struct lp_build_tgsi_context
*bld_base
,
977 const struct tgsi_full_instruction
*inst
,
978 const struct tgsi_opcode_info
*info
,
981 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
982 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
983 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[0];
984 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
985 LLVMValueRef temp_ptr
, temp_ptr2
= NULL
;
986 unsigned chan
, chan_index
;
987 bool is_vec_store
= false;
988 enum tgsi_opcode_type dtype
= tgsi_opcode_infer_dst_type(inst
->Instruction
.Opcode
);
991 LLVMTypeKind k
= LLVMGetTypeKind(LLVMTypeOf(dst
[0]));
992 is_vec_store
= (k
== LLVMVectorTypeKind
);
996 LLVMValueRef values
[4] = {};
997 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst
, chan
) {
998 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, chan
, 0);
999 values
[chan
] = LLVMBuildExtractElement(gallivm
->builder
,
1002 bld_base
->emit_store(bld_base
, inst
, info
, values
);
1006 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1007 LLVMValueRef value
= dst
[chan_index
];
1009 if (tgsi_type_is_64bit(dtype
) && (chan_index
== 1 || chan_index
== 3))
1011 if (inst
->Instruction
.Saturate
)
1012 value
= ac_build_clamp(&ctx
->ac
, value
);
1014 if (reg
->Register
.File
== TGSI_FILE_ADDRESS
) {
1015 temp_ptr
= ctx
->addrs
[reg
->Register
.Index
][chan_index
];
1016 LLVMBuildStore(builder
, value
, temp_ptr
);
1020 if (!tgsi_type_is_64bit(dtype
))
1021 value
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, value
);
1023 if (reg
->Register
.Indirect
) {
1024 unsigned file
= reg
->Register
.File
;
1025 unsigned reg_index
= reg
->Register
.Index
;
1026 store_value_to_array(bld_base
, value
, file
, chan_index
,
1027 reg_index
, ®
->Indirect
);
1029 switch(reg
->Register
.File
) {
1030 case TGSI_FILE_OUTPUT
:
1031 temp_ptr
= ctx
->outputs
[reg
->Register
.Index
][chan_index
];
1032 if (tgsi_type_is_64bit(dtype
))
1033 temp_ptr2
= ctx
->outputs
[reg
->Register
.Index
][chan_index
+ 1];
1036 case TGSI_FILE_TEMPORARY
:
1038 if (reg
->Register
.Index
>= ctx
->temps_count
)
1041 temp_ptr
= ctx
->temps
[ TGSI_NUM_CHANNELS
* reg
->Register
.Index
+ chan_index
];
1042 if (tgsi_type_is_64bit(dtype
))
1043 temp_ptr2
= ctx
->temps
[ TGSI_NUM_CHANNELS
* reg
->Register
.Index
+ chan_index
+ 1];
1050 if (!tgsi_type_is_64bit(dtype
))
1051 LLVMBuildStore(builder
, value
, temp_ptr
);
1053 LLVMValueRef ptr
= LLVMBuildBitCast(builder
, value
,
1054 LLVMVectorType(ctx
->i32
, 2), "");
1056 value
= LLVMBuildExtractElement(builder
, ptr
,
1058 val2
= LLVMBuildExtractElement(builder
, ptr
,
1061 LLVMBuildStore(builder
, bitcast(bld_base
, TGSI_TYPE_FLOAT
, value
), temp_ptr
);
1062 LLVMBuildStore(builder
, bitcast(bld_base
, TGSI_TYPE_FLOAT
, val2
), temp_ptr2
);
1068 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
, int pc
)
1071 /* Subtract 1 so that the number shown is that of the corresponding
1072 * opcode in the TGSI dump, e.g. an if block has the same suffix as
1073 * the instruction number of the corresponding TGSI IF.
1075 snprintf(buf
, sizeof(buf
), "%s%d", base
, pc
- 1);
1076 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
1079 /* Append a basic block at the level of the parent flow.
1081 static LLVMBasicBlockRef
append_basic_block(struct si_shader_context
*ctx
,
1084 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
1086 assert(ctx
->flow_depth
>= 1);
1088 if (ctx
->flow_depth
>= 2) {
1089 struct si_llvm_flow
*flow
= &ctx
->flow
[ctx
->flow_depth
- 2];
1091 return LLVMInsertBasicBlockInContext(gallivm
->context
,
1092 flow
->next_block
, name
);
1095 return LLVMAppendBasicBlockInContext(gallivm
->context
, ctx
->main_fn
, name
);
1098 /* Emit a branch to the given default target for the current block if
1099 * applicable -- that is, if the current block does not already contain a
1100 * branch from a break or continue.
1102 static void emit_default_branch(LLVMBuilderRef builder
, LLVMBasicBlockRef target
)
1104 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
1105 LLVMBuildBr(builder
, target
);
1108 static void bgnloop_emit(const struct lp_build_tgsi_action
*action
,
1109 struct lp_build_tgsi_context
*bld_base
,
1110 struct lp_build_emit_data
*emit_data
)
1112 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1113 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
1114 struct si_llvm_flow
*flow
= push_flow(ctx
);
1115 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
1116 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
1117 set_basicblock_name(flow
->loop_entry_block
, "loop", bld_base
->pc
);
1118 LLVMBuildBr(gallivm
->builder
, flow
->loop_entry_block
);
1119 LLVMPositionBuilderAtEnd(gallivm
->builder
, flow
->loop_entry_block
);
1122 static void brk_emit(const struct lp_build_tgsi_action
*action
,
1123 struct lp_build_tgsi_context
*bld_base
,
1124 struct lp_build_emit_data
*emit_data
)
1126 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1127 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
1128 struct si_llvm_flow
*flow
= get_innermost_loop(ctx
);
1130 LLVMBuildBr(gallivm
->builder
, flow
->next_block
);
1133 static void cont_emit(const struct lp_build_tgsi_action
*action
,
1134 struct lp_build_tgsi_context
*bld_base
,
1135 struct lp_build_emit_data
*emit_data
)
1137 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1138 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
1139 struct si_llvm_flow
*flow
= get_innermost_loop(ctx
);
1141 LLVMBuildBr(gallivm
->builder
, flow
->loop_entry_block
);
1144 static void else_emit(const struct lp_build_tgsi_action
*action
,
1145 struct lp_build_tgsi_context
*bld_base
,
1146 struct lp_build_emit_data
*emit_data
)
1148 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1149 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
1150 struct si_llvm_flow
*current_branch
= get_current_flow(ctx
);
1151 LLVMBasicBlockRef endif_block
;
1153 assert(!current_branch
->loop_entry_block
);
1155 endif_block
= append_basic_block(ctx
, "ENDIF");
1156 emit_default_branch(gallivm
->builder
, endif_block
);
1158 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->next_block
);
1159 set_basicblock_name(current_branch
->next_block
, "else", bld_base
->pc
);
1161 current_branch
->next_block
= endif_block
;
1164 static void endif_emit(const struct lp_build_tgsi_action
*action
,
1165 struct lp_build_tgsi_context
*bld_base
,
1166 struct lp_build_emit_data
*emit_data
)
1168 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1169 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
1170 struct si_llvm_flow
*current_branch
= get_current_flow(ctx
);
1172 assert(!current_branch
->loop_entry_block
);
1174 emit_default_branch(gallivm
->builder
, current_branch
->next_block
);
1175 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->next_block
);
1176 set_basicblock_name(current_branch
->next_block
, "endif", bld_base
->pc
);
1181 static void endloop_emit(const struct lp_build_tgsi_action
*action
,
1182 struct lp_build_tgsi_context
*bld_base
,
1183 struct lp_build_emit_data
*emit_data
)
1185 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1186 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
1187 struct si_llvm_flow
*current_loop
= get_current_flow(ctx
);
1189 assert(current_loop
->loop_entry_block
);
1191 emit_default_branch(gallivm
->builder
, current_loop
->loop_entry_block
);
1193 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_loop
->next_block
);
1194 set_basicblock_name(current_loop
->next_block
, "endloop", bld_base
->pc
);
1198 static void if_cond_emit(const struct lp_build_tgsi_action
*action
,
1199 struct lp_build_tgsi_context
*bld_base
,
1200 struct lp_build_emit_data
*emit_data
,
1203 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1204 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
1205 struct si_llvm_flow
*flow
= push_flow(ctx
);
1206 LLVMBasicBlockRef if_block
;
1208 if_block
= append_basic_block(ctx
, "IF");
1209 flow
->next_block
= append_basic_block(ctx
, "ELSE");
1210 set_basicblock_name(if_block
, "if", bld_base
->pc
);
1211 LLVMBuildCondBr(gallivm
->builder
, cond
, if_block
, flow
->next_block
);
1212 LLVMPositionBuilderAtEnd(gallivm
->builder
, if_block
);
1215 static void if_emit(const struct lp_build_tgsi_action
*action
,
1216 struct lp_build_tgsi_context
*bld_base
,
1217 struct lp_build_emit_data
*emit_data
)
1219 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1222 cond
= LLVMBuildFCmp(gallivm
->builder
, LLVMRealUNE
,
1224 bld_base
->base
.zero
, "");
1226 if_cond_emit(action
, bld_base
, emit_data
, cond
);
1229 static void uif_emit(const struct lp_build_tgsi_action
*action
,
1230 struct lp_build_tgsi_context
*bld_base
,
1231 struct lp_build_emit_data
*emit_data
)
1233 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1236 cond
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
1237 bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, emit_data
->args
[0]),
1238 bld_base
->int_bld
.zero
, "");
1240 if_cond_emit(action
, bld_base
, emit_data
, cond
);
1243 static void emit_immediate(struct lp_build_tgsi_context
*bld_base
,
1244 const struct tgsi_full_immediate
*imm
)
1247 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1249 for (i
= 0; i
< 4; ++i
) {
1250 ctx
->imms
[ctx
->imms_num
* TGSI_NUM_CHANNELS
+ i
] =
1251 LLVMConstInt(ctx
->i32
, imm
->u
[i
].Uint
, false );
1257 void si_llvm_context_init(struct si_shader_context
*ctx
,
1258 struct si_screen
*sscreen
,
1259 LLVMTargetMachineRef tm
)
1261 struct lp_type type
;
1263 /* Initialize the gallivm object:
1264 * We are only using the module, context, and builder fields of this struct.
1265 * This should be enough for us to be able to pass our gallivm struct to the
1266 * helper functions in the gallivm module.
1268 memset(ctx
, 0, sizeof(*ctx
));
1269 ctx
->screen
= sscreen
;
1272 ctx
->gallivm
.context
= LLVMContextCreate();
1273 ctx
->gallivm
.module
= LLVMModuleCreateWithNameInContext("tgsi",
1274 ctx
->gallivm
.context
);
1275 LLVMSetTarget(ctx
->gallivm
.module
, "amdgcn--");
1277 #if HAVE_LLVM >= 0x0309
1278 LLVMTargetDataRef data_layout
= LLVMCreateTargetDataLayout(tm
);
1279 char *data_layout_str
= LLVMCopyStringRepOfTargetData(data_layout
);
1280 LLVMSetDataLayout(ctx
->gallivm
.module
, data_layout_str
);
1281 LLVMDisposeTargetData(data_layout
);
1282 LLVMDisposeMessage(data_layout_str
);
1285 bool unsafe_fpmath
= (sscreen
->b
.debug_flags
& DBG_UNSAFE_MATH
) != 0;
1286 enum lp_float_mode float_mode
=
1287 unsafe_fpmath
? LP_FLOAT_MODE_UNSAFE_FP_MATH
:
1288 LP_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH
;
1290 ctx
->gallivm
.builder
= lp_create_builder(ctx
->gallivm
.context
,
1293 ac_llvm_context_init(&ctx
->ac
, ctx
->gallivm
.context
);
1294 ctx
->ac
.module
= ctx
->gallivm
.module
;
1295 ctx
->ac
.builder
= ctx
->gallivm
.builder
;
1297 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
1299 type
.floating
= true;
1306 lp_build_context_init(&bld_base
->base
, &ctx
->gallivm
, type
);
1307 lp_build_context_init(&ctx
->bld_base
.uint_bld
, &ctx
->gallivm
, lp_uint_type(type
));
1308 lp_build_context_init(&ctx
->bld_base
.int_bld
, &ctx
->gallivm
, lp_int_type(type
));
1310 lp_build_context_init(&ctx
->bld_base
.dbl_bld
, &ctx
->gallivm
, type
);
1311 lp_build_context_init(&ctx
->bld_base
.uint64_bld
, &ctx
->gallivm
, lp_uint_type(type
));
1312 lp_build_context_init(&ctx
->bld_base
.int64_bld
, &ctx
->gallivm
, lp_int_type(type
));
1315 bld_base
->emit_swizzle
= emit_swizzle
;
1316 bld_base
->emit_declaration
= emit_declaration
;
1317 bld_base
->emit_immediate
= emit_immediate
;
1319 /* metadata allowing 2.5 ULP */
1320 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->gallivm
.context
,
1322 LLVMValueRef arg
= lp_build_const_float(&ctx
->gallivm
, 2.5);
1323 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->gallivm
.context
,
1326 bld_base
->op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
1327 bld_base
->op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
1328 bld_base
->op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
1329 bld_base
->op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
1330 bld_base
->op_actions
[TGSI_OPCODE_UIF
].emit
= uif_emit
;
1331 bld_base
->op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
1332 bld_base
->op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
1333 bld_base
->op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
1335 si_shader_context_init_alu(&ctx
->bld_base
);
1337 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->gallivm
.context
);
1338 ctx
->i1
= LLVMInt1TypeInContext(ctx
->gallivm
.context
);
1339 ctx
->i8
= LLVMInt8TypeInContext(ctx
->gallivm
.context
);
1340 ctx
->i32
= LLVMInt32TypeInContext(ctx
->gallivm
.context
);
1341 ctx
->i64
= LLVMInt64TypeInContext(ctx
->gallivm
.context
);
1342 ctx
->i128
= LLVMIntTypeInContext(ctx
->gallivm
.context
, 128);
1343 ctx
->f32
= LLVMFloatTypeInContext(ctx
->gallivm
.context
);
1344 ctx
->v16i8
= LLVMVectorType(ctx
->i8
, 16);
1345 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
1346 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
1347 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
1348 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
1350 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, 0);
1351 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, 0);
1354 /* Set the context to a certain TGSI shader. Can be called repeatedly
1355 * to change the shader. */
1356 void si_llvm_context_set_tgsi(struct si_shader_context
*ctx
,
1357 struct si_shader
*shader
)
1359 const struct tgsi_shader_info
*info
= NULL
;
1360 const struct tgsi_token
*tokens
= NULL
;
1362 if (shader
&& shader
->selector
) {
1363 info
= &shader
->selector
->info
;
1364 tokens
= shader
->selector
->tokens
;
1367 ctx
->shader
= shader
;
1368 ctx
->type
= info
? info
->processor
: -1;
1369 ctx
->bld_base
.info
= info
;
1371 /* Clean up the old contents. */
1372 FREE(ctx
->temp_arrays
);
1373 ctx
->temp_arrays
= NULL
;
1374 FREE(ctx
->temp_array_allocas
);
1375 ctx
->temp_array_allocas
= NULL
;
1383 ctx
->temps_count
= 0;
1385 if (!info
|| !tokens
)
1388 if (info
->array_max
[TGSI_FILE_TEMPORARY
] > 0) {
1389 int size
= info
->array_max
[TGSI_FILE_TEMPORARY
];
1391 ctx
->temp_arrays
= CALLOC(size
, sizeof(ctx
->temp_arrays
[0]));
1392 ctx
->temp_array_allocas
= CALLOC(size
, sizeof(ctx
->temp_array_allocas
[0]));
1394 tgsi_scan_arrays(tokens
, TGSI_FILE_TEMPORARY
, size
,
1397 if (info
->file_max
[TGSI_FILE_IMMEDIATE
] >= 0) {
1398 int size
= info
->file_max
[TGSI_FILE_IMMEDIATE
] + 1;
1399 ctx
->imms
= MALLOC(size
* TGSI_NUM_CHANNELS
* sizeof(LLVMValueRef
));
1402 /* Re-set these to start with a clean slate. */
1403 ctx
->bld_base
.num_instructions
= 0;
1404 ctx
->bld_base
.pc
= 0;
1405 memset(ctx
->outputs
, 0, sizeof(ctx
->outputs
));
1407 ctx
->bld_base
.emit_store
= si_llvm_emit_store
;
1408 ctx
->bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = si_llvm_emit_fetch
;
1409 ctx
->bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = si_llvm_emit_fetch
;
1410 ctx
->bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = si_llvm_emit_fetch
;
1411 ctx
->bld_base
.emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = si_llvm_emit_fetch
;
1412 ctx
->bld_base
.emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = fetch_system_value
;
1415 void si_llvm_create_func(struct si_shader_context
*ctx
,
1417 LLVMTypeRef
*return_types
, unsigned num_return_elems
,
1418 LLVMTypeRef
*ParamTypes
, unsigned ParamCount
)
1420 LLVMTypeRef main_fn_type
, ret_type
;
1421 LLVMBasicBlockRef main_fn_body
;
1423 if (num_return_elems
)
1424 ret_type
= LLVMStructTypeInContext(ctx
->gallivm
.context
,
1426 num_return_elems
, true);
1428 ret_type
= LLVMVoidTypeInContext(ctx
->gallivm
.context
);
1430 /* Setup the function */
1431 ctx
->return_type
= ret_type
;
1432 main_fn_type
= LLVMFunctionType(ret_type
, ParamTypes
, ParamCount
, 0);
1433 ctx
->main_fn
= LLVMAddFunction(ctx
->gallivm
.module
, name
, main_fn_type
);
1434 main_fn_body
= LLVMAppendBasicBlockInContext(ctx
->gallivm
.context
,
1435 ctx
->main_fn
, "main_body");
1436 LLVMPositionBuilderAtEnd(ctx
->gallivm
.builder
, main_fn_body
);
1439 void si_llvm_finalize_module(struct si_shader_context
*ctx
,
1442 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
1443 const char *triple
= LLVMGetTarget(gallivm
->module
);
1444 LLVMTargetLibraryInfoRef target_library_info
;
1446 /* Create the pass manager */
1447 gallivm
->passmgr
= LLVMCreatePassManager();
1449 target_library_info
= gallivm_create_target_library_info(triple
);
1450 LLVMAddTargetLibraryInfo(target_library_info
, gallivm
->passmgr
);
1453 LLVMAddVerifierPass(gallivm
->passmgr
);
1455 LLVMAddAlwaysInlinerPass(gallivm
->passmgr
);
1457 /* This pass should eliminate all the load and store instructions */
1458 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
1460 /* Add some optimization passes */
1461 LLVMAddScalarReplAggregatesPass(gallivm
->passmgr
);
1462 LLVMAddLICMPass(gallivm
->passmgr
);
1463 LLVMAddAggressiveDCEPass(gallivm
->passmgr
);
1464 LLVMAddCFGSimplificationPass(gallivm
->passmgr
);
1465 LLVMAddInstructionCombiningPass(gallivm
->passmgr
);
1468 LLVMRunPassManager(gallivm
->passmgr
, ctx
->gallivm
.module
);
1470 LLVMDisposeBuilder(gallivm
->builder
);
1471 LLVMDisposePassManager(gallivm
->passmgr
);
1472 gallivm_dispose_target_library_info(target_library_info
);
1475 void si_llvm_dispose(struct si_shader_context
*ctx
)
1477 LLVMDisposeModule(ctx
->gallivm
.module
);
1478 LLVMContextDispose(ctx
->gallivm
.context
);
1479 FREE(ctx
->temp_arrays
);
1480 ctx
->temp_arrays
= NULL
;
1481 FREE(ctx
->temp_array_allocas
);
1482 ctx
->temp_array_allocas
= NULL
;
1485 ctx
->temps_count
= 0;
1491 ctx
->flow_depth_max
= 0;