2 * Copyright 2016 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 #include "si_shader_internal.h"
26 #include "ac_llvm_util.h"
28 void si_llvm_emit_kill(struct ac_shader_abi
*abi
, LLVMValueRef visible
)
30 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
31 LLVMBuilderRef builder
= ctx
->ac
.builder
;
33 if (ctx
->shader
->selector
->force_correct_derivs_after_kill
) {
34 /* Kill immediately while maintaining WQM. */
35 ac_build_kill_if_false(&ctx
->ac
,
36 ac_build_wqm_vote(&ctx
->ac
, visible
));
38 LLVMValueRef mask
= LLVMBuildLoad(builder
, ctx
->postponed_kill
, "");
39 mask
= LLVMBuildAnd(builder
, mask
, visible
, "");
40 LLVMBuildStore(builder
, mask
, ctx
->postponed_kill
);
44 ac_build_kill_if_false(&ctx
->ac
, visible
);
47 static void kil_emit(const struct lp_build_tgsi_action
*action
,
48 struct lp_build_tgsi_context
*bld_base
,
49 struct lp_build_emit_data
*emit_data
)
51 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
54 if (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_KILL_IF
) {
55 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
56 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
57 LLVMBuilderRef builder
= ctx
->ac
.builder
;
59 LLVMValueRef conds
[TGSI_NUM_CHANNELS
];
61 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
62 LLVMValueRef value
= lp_build_emit_fetch(bld_base
, inst
, 0, i
);
63 /* UGE because NaN shouldn't get killed */
64 conds
[i
] = LLVMBuildFCmp(builder
, LLVMRealUGE
, value
,
68 /* And the conditions together */
69 for (i
= TGSI_NUM_CHANNELS
- 1; i
> 0; i
--) {
70 conds
[i
- 1] = LLVMBuildAnd(builder
, conds
[i
], conds
[i
- 1], "");
74 assert(emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_KILL
);
75 visible
= ctx
->i1false
;
78 si_llvm_emit_kill(&ctx
->abi
, visible
);
81 static void emit_icmp(const struct lp_build_tgsi_action
*action
,
82 struct lp_build_tgsi_context
*bld_base
,
83 struct lp_build_emit_data
*emit_data
)
86 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
88 switch (emit_data
->inst
->Instruction
.Opcode
) {
89 case TGSI_OPCODE_USEQ
:
90 case TGSI_OPCODE_U64SEQ
: pred
= LLVMIntEQ
; break;
91 case TGSI_OPCODE_USNE
:
92 case TGSI_OPCODE_U64SNE
: pred
= LLVMIntNE
; break;
93 case TGSI_OPCODE_USGE
:
94 case TGSI_OPCODE_U64SGE
: pred
= LLVMIntUGE
; break;
95 case TGSI_OPCODE_USLT
:
96 case TGSI_OPCODE_U64SLT
: pred
= LLVMIntULT
; break;
97 case TGSI_OPCODE_ISGE
:
98 case TGSI_OPCODE_I64SGE
: pred
= LLVMIntSGE
; break;
99 case TGSI_OPCODE_ISLT
:
100 case TGSI_OPCODE_I64SLT
: pred
= LLVMIntSLT
; break;
102 assert(!"unknown instruction");
107 LLVMValueRef v
= LLVMBuildICmp(ctx
->ac
.builder
, pred
,
108 emit_data
->args
[0], emit_data
->args
[1],"");
110 v
= LLVMBuildSExtOrBitCast(ctx
->ac
.builder
, v
, ctx
->i32
, "");
112 emit_data
->output
[emit_data
->chan
] = v
;
115 static void emit_ucmp(const struct lp_build_tgsi_action
*action
,
116 struct lp_build_tgsi_context
*bld_base
,
117 struct lp_build_emit_data
*emit_data
)
119 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
120 LLVMValueRef arg0
= ac_to_integer(&ctx
->ac
, emit_data
->args
[0]);
122 LLVMValueRef v
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
, arg0
,
125 emit_data
->output
[emit_data
->chan
] =
126 LLVMBuildSelect(ctx
->ac
.builder
, v
, emit_data
->args
[1], emit_data
->args
[2], "");
129 static void emit_cmp(const struct lp_build_tgsi_action
*action
,
130 struct lp_build_tgsi_context
*bld_base
,
131 struct lp_build_emit_data
*emit_data
)
133 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
134 LLVMValueRef cond
, *args
= emit_data
->args
;
136 cond
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealOLT
, args
[0],
139 emit_data
->output
[emit_data
->chan
] =
140 LLVMBuildSelect(ctx
->ac
.builder
, cond
, args
[1], args
[2], "");
143 static void emit_set_cond(const struct lp_build_tgsi_action
*action
,
144 struct lp_build_tgsi_context
*bld_base
,
145 struct lp_build_emit_data
*emit_data
)
147 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
148 LLVMRealPredicate pred
;
151 /* Use ordered for everything but NE (which is usual for
154 switch (emit_data
->inst
->Instruction
.Opcode
) {
155 case TGSI_OPCODE_SGE
: pred
= LLVMRealOGE
; break;
156 case TGSI_OPCODE_SEQ
: pred
= LLVMRealOEQ
; break;
157 case TGSI_OPCODE_SLE
: pred
= LLVMRealOLE
; break;
158 case TGSI_OPCODE_SLT
: pred
= LLVMRealOLT
; break;
159 case TGSI_OPCODE_SNE
: pred
= LLVMRealUNE
; break;
160 case TGSI_OPCODE_SGT
: pred
= LLVMRealOGT
; break;
161 default: assert(!"unknown instruction"); pred
= 0; break;
164 cond
= LLVMBuildFCmp(ctx
->ac
.builder
,
165 pred
, emit_data
->args
[0], emit_data
->args
[1], "");
167 emit_data
->output
[emit_data
->chan
] = LLVMBuildSelect(ctx
->ac
.builder
,
168 cond
, ctx
->ac
.f32_1
, ctx
->ac
.f32_0
, "");
171 static void emit_fcmp(const struct lp_build_tgsi_action
*action
,
172 struct lp_build_tgsi_context
*bld_base
,
173 struct lp_build_emit_data
*emit_data
)
175 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
176 LLVMRealPredicate pred
;
178 /* Use ordered for everything but NE (which is usual for
181 switch (emit_data
->inst
->Instruction
.Opcode
) {
182 case TGSI_OPCODE_FSEQ
: pred
= LLVMRealOEQ
; break;
183 case TGSI_OPCODE_FSGE
: pred
= LLVMRealOGE
; break;
184 case TGSI_OPCODE_FSLT
: pred
= LLVMRealOLT
; break;
185 case TGSI_OPCODE_FSNE
: pred
= LLVMRealUNE
; break;
186 default: assert(!"unknown instruction"); pred
= 0; break;
189 LLVMValueRef v
= LLVMBuildFCmp(ctx
->ac
.builder
, pred
,
190 emit_data
->args
[0], emit_data
->args
[1],"");
192 v
= LLVMBuildSExtOrBitCast(ctx
->ac
.builder
, v
, ctx
->i32
, "");
194 emit_data
->output
[emit_data
->chan
] = v
;
197 static void emit_dcmp(const struct lp_build_tgsi_action
*action
,
198 struct lp_build_tgsi_context
*bld_base
,
199 struct lp_build_emit_data
*emit_data
)
201 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
202 LLVMRealPredicate pred
;
204 /* Use ordered for everything but NE (which is usual for
207 switch (emit_data
->inst
->Instruction
.Opcode
) {
208 case TGSI_OPCODE_DSEQ
: pred
= LLVMRealOEQ
; break;
209 case TGSI_OPCODE_DSGE
: pred
= LLVMRealOGE
; break;
210 case TGSI_OPCODE_DSLT
: pred
= LLVMRealOLT
; break;
211 case TGSI_OPCODE_DSNE
: pred
= LLVMRealUNE
; break;
212 default: assert(!"unknown instruction"); pred
= 0; break;
215 LLVMValueRef v
= LLVMBuildFCmp(ctx
->ac
.builder
, pred
,
216 emit_data
->args
[0], emit_data
->args
[1],"");
218 v
= LLVMBuildSExtOrBitCast(ctx
->ac
.builder
, v
, ctx
->i32
, "");
220 emit_data
->output
[emit_data
->chan
] = v
;
223 static void emit_not(const struct lp_build_tgsi_action
*action
,
224 struct lp_build_tgsi_context
*bld_base
,
225 struct lp_build_emit_data
*emit_data
)
227 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
228 LLVMValueRef v
= ac_to_integer(&ctx
->ac
, emit_data
->args
[0]);
229 emit_data
->output
[emit_data
->chan
] = LLVMBuildNot(ctx
->ac
.builder
, v
, "");
232 static void emit_arl(const struct lp_build_tgsi_action
*action
,
233 struct lp_build_tgsi_context
*bld_base
,
234 struct lp_build_emit_data
*emit_data
)
236 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
237 LLVMValueRef floor_index
=
238 ac_build_intrinsic(&ctx
->ac
, "llvm.floor.f32", ctx
->f32
,
239 &emit_data
->args
[0], 1, AC_FUNC_ATTR_READNONE
);
240 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToSI(ctx
->ac
.builder
,
241 floor_index
, ctx
->i32
, "");
244 static void emit_and(const struct lp_build_tgsi_action
*action
,
245 struct lp_build_tgsi_context
*bld_base
,
246 struct lp_build_emit_data
*emit_data
)
248 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
249 emit_data
->output
[emit_data
->chan
] = LLVMBuildAnd(ctx
->ac
.builder
,
250 emit_data
->args
[0], emit_data
->args
[1], "");
253 static void emit_or(const struct lp_build_tgsi_action
*action
,
254 struct lp_build_tgsi_context
*bld_base
,
255 struct lp_build_emit_data
*emit_data
)
257 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
258 emit_data
->output
[emit_data
->chan
] = LLVMBuildOr(ctx
->ac
.builder
,
259 emit_data
->args
[0], emit_data
->args
[1], "");
262 static void emit_uadd(const struct lp_build_tgsi_action
*action
,
263 struct lp_build_tgsi_context
*bld_base
,
264 struct lp_build_emit_data
*emit_data
)
266 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
267 emit_data
->output
[emit_data
->chan
] = LLVMBuildAdd(ctx
->ac
.builder
,
268 emit_data
->args
[0], emit_data
->args
[1], "");
271 static void emit_udiv(const struct lp_build_tgsi_action
*action
,
272 struct lp_build_tgsi_context
*bld_base
,
273 struct lp_build_emit_data
*emit_data
)
275 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
276 emit_data
->output
[emit_data
->chan
] = LLVMBuildUDiv(ctx
->ac
.builder
,
277 emit_data
->args
[0], emit_data
->args
[1], "");
280 static void emit_idiv(const struct lp_build_tgsi_action
*action
,
281 struct lp_build_tgsi_context
*bld_base
,
282 struct lp_build_emit_data
*emit_data
)
284 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
285 emit_data
->output
[emit_data
->chan
] = LLVMBuildSDiv(ctx
->ac
.builder
,
286 emit_data
->args
[0], emit_data
->args
[1], "");
289 static void emit_mod(const struct lp_build_tgsi_action
*action
,
290 struct lp_build_tgsi_context
*bld_base
,
291 struct lp_build_emit_data
*emit_data
)
293 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
294 emit_data
->output
[emit_data
->chan
] = LLVMBuildSRem(ctx
->ac
.builder
,
295 emit_data
->args
[0], emit_data
->args
[1], "");
298 static void emit_umod(const struct lp_build_tgsi_action
*action
,
299 struct lp_build_tgsi_context
*bld_base
,
300 struct lp_build_emit_data
*emit_data
)
302 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
303 emit_data
->output
[emit_data
->chan
] = LLVMBuildURem(ctx
->ac
.builder
,
304 emit_data
->args
[0], emit_data
->args
[1], "");
307 static void emit_shl(const struct lp_build_tgsi_action
*action
,
308 struct lp_build_tgsi_context
*bld_base
,
309 struct lp_build_emit_data
*emit_data
)
311 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
312 emit_data
->output
[emit_data
->chan
] = LLVMBuildShl(ctx
->ac
.builder
,
313 emit_data
->args
[0], emit_data
->args
[1], "");
316 static void emit_ushr(const struct lp_build_tgsi_action
*action
,
317 struct lp_build_tgsi_context
*bld_base
,
318 struct lp_build_emit_data
*emit_data
)
320 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
321 emit_data
->output
[emit_data
->chan
] = LLVMBuildLShr(ctx
->ac
.builder
,
322 emit_data
->args
[0], emit_data
->args
[1], "");
324 static void emit_ishr(const struct lp_build_tgsi_action
*action
,
325 struct lp_build_tgsi_context
*bld_base
,
326 struct lp_build_emit_data
*emit_data
)
328 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
329 emit_data
->output
[emit_data
->chan
] = LLVMBuildAShr(ctx
->ac
.builder
,
330 emit_data
->args
[0], emit_data
->args
[1], "");
333 static void emit_xor(const struct lp_build_tgsi_action
*action
,
334 struct lp_build_tgsi_context
*bld_base
,
335 struct lp_build_emit_data
*emit_data
)
337 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
338 emit_data
->output
[emit_data
->chan
] = LLVMBuildXor(ctx
->ac
.builder
,
339 emit_data
->args
[0], emit_data
->args
[1], "");
342 static void emit_ssg(const struct lp_build_tgsi_action
*action
,
343 struct lp_build_tgsi_context
*bld_base
,
344 struct lp_build_emit_data
*emit_data
)
346 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
350 if (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_I64SSG
) {
351 val
= ac_build_isign(&ctx
->ac
, emit_data
->args
[0], 64);
352 } else if (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_ISSG
) {
353 val
= ac_build_isign(&ctx
->ac
, emit_data
->args
[0], 32);
354 } else if (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_DSSG
) {
355 val
= ac_build_fsign(&ctx
->ac
, emit_data
->args
[0], 64);
357 val
= ac_build_fsign(&ctx
->ac
, emit_data
->args
[0], 32);
360 emit_data
->output
[emit_data
->chan
] = val
;
363 static void emit_ineg(const struct lp_build_tgsi_action
*action
,
364 struct lp_build_tgsi_context
*bld_base
,
365 struct lp_build_emit_data
*emit_data
)
367 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
368 emit_data
->output
[emit_data
->chan
] = LLVMBuildNeg(ctx
->ac
.builder
,
369 emit_data
->args
[0], "");
372 static void emit_dneg(const struct lp_build_tgsi_action
*action
,
373 struct lp_build_tgsi_context
*bld_base
,
374 struct lp_build_emit_data
*emit_data
)
376 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
377 emit_data
->output
[emit_data
->chan
] = LLVMBuildFNeg(ctx
->ac
.builder
,
378 emit_data
->args
[0], "");
381 static void emit_frac(const struct lp_build_tgsi_action
*action
,
382 struct lp_build_tgsi_context
*bld_base
,
383 struct lp_build_emit_data
*emit_data
)
385 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
388 if (emit_data
->info
->opcode
== TGSI_OPCODE_FRC
)
390 else if (emit_data
->info
->opcode
== TGSI_OPCODE_DFRAC
)
397 emit_data
->output
[emit_data
->chan
] =
398 ac_build_fract(&ctx
->ac
, emit_data
->args
[0], bitsize
);
401 static void emit_f2i(const struct lp_build_tgsi_action
*action
,
402 struct lp_build_tgsi_context
*bld_base
,
403 struct lp_build_emit_data
*emit_data
)
405 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
406 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToSI(ctx
->ac
.builder
,
407 emit_data
->args
[0], ctx
->i32
, "");
410 static void emit_f2u(const struct lp_build_tgsi_action
*action
,
411 struct lp_build_tgsi_context
*bld_base
,
412 struct lp_build_emit_data
*emit_data
)
414 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
415 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToUI(ctx
->ac
.builder
,
416 emit_data
->args
[0], ctx
->i32
, "");
419 static void emit_i2f(const struct lp_build_tgsi_action
*action
,
420 struct lp_build_tgsi_context
*bld_base
,
421 struct lp_build_emit_data
*emit_data
)
423 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
424 emit_data
->output
[emit_data
->chan
] = LLVMBuildSIToFP(ctx
->ac
.builder
,
425 emit_data
->args
[0], ctx
->f32
, "");
428 static void emit_u2f(const struct lp_build_tgsi_action
*action
,
429 struct lp_build_tgsi_context
*bld_base
,
430 struct lp_build_emit_data
*emit_data
)
432 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
433 emit_data
->output
[emit_data
->chan
] = LLVMBuildUIToFP(ctx
->ac
.builder
,
434 emit_data
->args
[0], ctx
->f32
, "");
438 build_tgsi_intrinsic_nomem(const struct lp_build_tgsi_action
*action
,
439 struct lp_build_tgsi_context
*bld_base
,
440 struct lp_build_emit_data
*emit_data
)
442 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
443 emit_data
->output
[emit_data
->chan
] =
444 ac_build_intrinsic(&ctx
->ac
, action
->intr_name
,
445 emit_data
->dst_type
, emit_data
->args
,
446 emit_data
->arg_count
, AC_FUNC_ATTR_READNONE
);
449 static void emit_bfi(const struct lp_build_tgsi_action
*action
,
450 struct lp_build_tgsi_context
*bld_base
,
451 struct lp_build_emit_data
*emit_data
)
453 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
454 LLVMBuilderRef builder
= ctx
->ac
.builder
;
455 LLVMValueRef bfi_args
[3];
456 LLVMValueRef bfi_sm5
;
459 // Calculate the bitmask: (((1 << src3) - 1) << src2
460 bfi_args
[0] = LLVMBuildShl(builder
,
461 LLVMBuildSub(builder
,
462 LLVMBuildShl(builder
,
464 emit_data
->args
[3], ""),
466 emit_data
->args
[2], "");
468 bfi_args
[1] = LLVMBuildShl(builder
, emit_data
->args
[1],
469 emit_data
->args
[2], "");
471 bfi_args
[2] = emit_data
->args
[0];
474 * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
475 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
478 LLVMBuildXor(builder
, bfi_args
[2],
479 LLVMBuildAnd(builder
, bfi_args
[0],
480 LLVMBuildXor(builder
, bfi_args
[1], bfi_args
[2],
483 /* Since shifts of >= 32 bits are undefined in LLVM IR, the backend
484 * uses the convenient V_BFI lowering for the above, which follows SM5
485 * and disagrees with GLSL semantics when bits (src3) is 32.
487 cond
= LLVMBuildICmp(builder
, LLVMIntUGE
, emit_data
->args
[3],
488 LLVMConstInt(ctx
->i32
, 32, 0), "");
489 emit_data
->output
[emit_data
->chan
] =
490 LLVMBuildSelect(builder
, cond
, emit_data
->args
[1], bfi_sm5
, "");
493 static void emit_bfe(const struct lp_build_tgsi_action
*action
,
494 struct lp_build_tgsi_context
*bld_base
,
495 struct lp_build_emit_data
*emit_data
)
497 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
499 /* FIXME: LLVM 7 returns incorrect result when count is 0.
500 * https://bugs.freedesktop.org/show_bug.cgi?id=107276
502 LLVMValueRef zero
= ctx
->i32_0
;
503 LLVMValueRef bfe_sm5
=
504 ac_build_bfe(&ctx
->ac
, emit_data
->args
[0],
505 emit_data
->args
[1], emit_data
->args
[2],
506 emit_data
->info
->opcode
== TGSI_OPCODE_IBFE
);
508 /* Correct for GLSL semantics. */
509 LLVMValueRef cond
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntUGE
, emit_data
->args
[2],
510 LLVMConstInt(ctx
->i32
, 32, 0), "");
511 LLVMValueRef cond2
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, emit_data
->args
[2],
513 bfe_sm5
= LLVMBuildSelect(ctx
->ac
.builder
, cond
, emit_data
->args
[0], bfe_sm5
, "");
514 emit_data
->output
[emit_data
->chan
] =
515 LLVMBuildSelect(ctx
->ac
.builder
, cond2
, zero
, bfe_sm5
, "");
518 /* this is ffs in C */
519 static void emit_lsb(const struct lp_build_tgsi_action
*action
,
520 struct lp_build_tgsi_context
*bld_base
,
521 struct lp_build_emit_data
*emit_data
)
523 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
525 emit_data
->output
[emit_data
->chan
] = ac_find_lsb(&ctx
->ac
, emit_data
->dst_type
, emit_data
->args
[0]);
528 /* Find the last bit set. */
529 static void emit_umsb(const struct lp_build_tgsi_action
*action
,
530 struct lp_build_tgsi_context
*bld_base
,
531 struct lp_build_emit_data
*emit_data
)
533 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
535 emit_data
->output
[emit_data
->chan
] =
536 ac_build_umsb(&ctx
->ac
, emit_data
->args
[0], emit_data
->dst_type
);
539 /* Find the last bit opposite of the sign bit. */
540 static void emit_imsb(const struct lp_build_tgsi_action
*action
,
541 struct lp_build_tgsi_context
*bld_base
,
542 struct lp_build_emit_data
*emit_data
)
544 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
545 emit_data
->output
[emit_data
->chan
] =
546 ac_build_imsb(&ctx
->ac
, emit_data
->args
[0],
547 emit_data
->dst_type
);
550 static void emit_iabs(const struct lp_build_tgsi_action
*action
,
551 struct lp_build_tgsi_context
*bld_base
,
552 struct lp_build_emit_data
*emit_data
)
554 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
556 emit_data
->output
[emit_data
->chan
] =
557 ac_build_imax(&ctx
->ac
, emit_data
->args
[0],
558 LLVMBuildNeg(ctx
->ac
.builder
, emit_data
->args
[0], ""));
561 static void emit_minmax_int(const struct lp_build_tgsi_action
*action
,
562 struct lp_build_tgsi_context
*bld_base
,
563 struct lp_build_emit_data
*emit_data
)
565 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
568 switch (emit_data
->info
->opcode
) {
571 case TGSI_OPCODE_IMAX
:
572 case TGSI_OPCODE_I64MAX
:
575 case TGSI_OPCODE_IMIN
:
576 case TGSI_OPCODE_I64MIN
:
579 case TGSI_OPCODE_UMAX
:
580 case TGSI_OPCODE_U64MAX
:
583 case TGSI_OPCODE_UMIN
:
584 case TGSI_OPCODE_U64MIN
:
589 emit_data
->output
[emit_data
->chan
] =
590 LLVMBuildSelect(ctx
->ac
.builder
,
591 LLVMBuildICmp(ctx
->ac
.builder
, op
, emit_data
->args
[0],
592 emit_data
->args
[1], ""),
594 emit_data
->args
[1], "");
597 static void emit_pk2h(const struct lp_build_tgsi_action
*action
,
598 struct lp_build_tgsi_context
*bld_base
,
599 struct lp_build_emit_data
*emit_data
)
601 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
603 lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
),
604 lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Y
),
608 /* From the GLSL 4.50 spec:
609 * "The rounding mode cannot be set and is undefined."
611 * v_cvt_pkrtz_f16 rounds to zero, but it's fastest.
613 emit_data
->output
[emit_data
->chan
] =
614 LLVMBuildBitCast(ctx
->ac
.builder
, ac_build_cvt_pkrtz_f16(&ctx
->ac
, v
),
618 static void emit_up2h(const struct lp_build_tgsi_action
*action
,
619 struct lp_build_tgsi_context
*bld_base
,
620 struct lp_build_emit_data
*emit_data
)
622 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
624 LLVMValueRef const16
, input
, val
;
627 i16
= LLVMInt16TypeInContext(ctx
->ac
.context
);
628 const16
= LLVMConstInt(ctx
->i32
, 16, 0);
629 input
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
631 for (i
= 0; i
< 2; i
++) {
632 val
= i
== 1 ? LLVMBuildLShr(ctx
->ac
.builder
, input
, const16
, "") : input
;
633 val
= LLVMBuildTrunc(ctx
->ac
.builder
, val
, i16
, "");
634 val
= ac_to_float(&ctx
->ac
, val
);
635 emit_data
->output
[i
] = LLVMBuildFPExt(ctx
->ac
.builder
, val
, ctx
->f32
, "");
639 static void emit_fdiv(const struct lp_build_tgsi_action
*action
,
640 struct lp_build_tgsi_context
*bld_base
,
641 struct lp_build_emit_data
*emit_data
)
643 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
645 emit_data
->output
[emit_data
->chan
] =
646 ac_build_fdiv(&ctx
->ac
, emit_data
->args
[0], emit_data
->args
[1]);
649 /* 1/sqrt is translated to rsq for f32 if fp32 denormals are not enabled in
650 * the target machine. f64 needs global unsafe math flags to get rsq. */
651 static void emit_rsq(const struct lp_build_tgsi_action
*action
,
652 struct lp_build_tgsi_context
*bld_base
,
653 struct lp_build_emit_data
*emit_data
)
655 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
658 ac_build_intrinsic(&ctx
->ac
, "llvm.sqrt.f32", ctx
->f32
,
659 &emit_data
->args
[0], 1, AC_FUNC_ATTR_READNONE
);
661 emit_data
->output
[emit_data
->chan
] =
662 ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
, sqrt
);
665 static void dfracexp_emit(const struct lp_build_tgsi_action
*action
,
666 struct lp_build_tgsi_context
*bld_base
,
667 struct lp_build_emit_data
*emit_data
)
669 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
670 LLVMValueRef in
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
672 emit_data
->output
[emit_data
->chan
] =
673 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.frexp.mant.f64",
674 ctx
->ac
.f64
, &in
, 1, 0);
675 emit_data
->output1
[emit_data
->chan
] =
676 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.frexp.exp.i32.f64",
677 ctx
->ac
.i32
, &in
, 1, 0);
680 void si_shader_context_init_alu(struct lp_build_tgsi_context
*bld_base
)
682 lp_set_default_actions(bld_base
);
684 bld_base
->op_actions
[TGSI_OPCODE_AND
].emit
= emit_and
;
685 bld_base
->op_actions
[TGSI_OPCODE_ARL
].emit
= emit_arl
;
686 bld_base
->op_actions
[TGSI_OPCODE_BFI
].emit
= emit_bfi
;
687 bld_base
->op_actions
[TGSI_OPCODE_BREV
].emit
= build_tgsi_intrinsic_nomem
;
688 bld_base
->op_actions
[TGSI_OPCODE_BREV
].intr_name
= "llvm.bitreverse.i32";
689 bld_base
->op_actions
[TGSI_OPCODE_CEIL
].emit
= build_tgsi_intrinsic_nomem
;
690 bld_base
->op_actions
[TGSI_OPCODE_CEIL
].intr_name
= "llvm.ceil.f32";
691 bld_base
->op_actions
[TGSI_OPCODE_CMP
].emit
= emit_cmp
;
692 bld_base
->op_actions
[TGSI_OPCODE_COS
].emit
= build_tgsi_intrinsic_nomem
;
693 bld_base
->op_actions
[TGSI_OPCODE_COS
].intr_name
= "llvm.cos.f32";
694 bld_base
->op_actions
[TGSI_OPCODE_DABS
].emit
= build_tgsi_intrinsic_nomem
;
695 bld_base
->op_actions
[TGSI_OPCODE_DABS
].intr_name
= "llvm.fabs.f64";
696 bld_base
->op_actions
[TGSI_OPCODE_DCEIL
].emit
= build_tgsi_intrinsic_nomem
;
697 bld_base
->op_actions
[TGSI_OPCODE_DCEIL
].intr_name
= "llvm.ceil.f64";
698 bld_base
->op_actions
[TGSI_OPCODE_DFLR
].emit
= build_tgsi_intrinsic_nomem
;
699 bld_base
->op_actions
[TGSI_OPCODE_DFLR
].intr_name
= "llvm.floor.f64";
700 bld_base
->op_actions
[TGSI_OPCODE_DFMA
].emit
= build_tgsi_intrinsic_nomem
;
701 bld_base
->op_actions
[TGSI_OPCODE_DFMA
].intr_name
= "llvm.fma.f64";
702 bld_base
->op_actions
[TGSI_OPCODE_DFRAC
].emit
= emit_frac
;
703 bld_base
->op_actions
[TGSI_OPCODE_DIV
].emit
= emit_fdiv
;
704 bld_base
->op_actions
[TGSI_OPCODE_DNEG
].emit
= emit_dneg
;
705 bld_base
->op_actions
[TGSI_OPCODE_DROUND
].emit
= build_tgsi_intrinsic_nomem
;
706 bld_base
->op_actions
[TGSI_OPCODE_DROUND
].intr_name
= "llvm.rint.f64";
707 bld_base
->op_actions
[TGSI_OPCODE_DSEQ
].emit
= emit_dcmp
;
708 bld_base
->op_actions
[TGSI_OPCODE_DSGE
].emit
= emit_dcmp
;
709 bld_base
->op_actions
[TGSI_OPCODE_DSLT
].emit
= emit_dcmp
;
710 bld_base
->op_actions
[TGSI_OPCODE_DSNE
].emit
= emit_dcmp
;
711 bld_base
->op_actions
[TGSI_OPCODE_DSSG
].emit
= emit_ssg
;
712 bld_base
->op_actions
[TGSI_OPCODE_DRSQ
].emit
= build_tgsi_intrinsic_nomem
;
713 bld_base
->op_actions
[TGSI_OPCODE_DRSQ
].intr_name
= "llvm.amdgcn.rsq.f64";
714 bld_base
->op_actions
[TGSI_OPCODE_DSQRT
].emit
= build_tgsi_intrinsic_nomem
;
715 bld_base
->op_actions
[TGSI_OPCODE_DSQRT
].intr_name
= "llvm.sqrt.f64";
716 bld_base
->op_actions
[TGSI_OPCODE_DTRUNC
].emit
= build_tgsi_intrinsic_nomem
;
717 bld_base
->op_actions
[TGSI_OPCODE_DTRUNC
].intr_name
= "llvm.trunc.f64";
718 bld_base
->op_actions
[TGSI_OPCODE_DFRACEXP
].emit
= dfracexp_emit
;
719 bld_base
->op_actions
[TGSI_OPCODE_DLDEXP
].emit
= build_tgsi_intrinsic_nomem
;
720 bld_base
->op_actions
[TGSI_OPCODE_DLDEXP
].intr_name
= "llvm.amdgcn.ldexp.f64";
721 bld_base
->op_actions
[TGSI_OPCODE_EX2
].emit
= build_tgsi_intrinsic_nomem
;
722 bld_base
->op_actions
[TGSI_OPCODE_EX2
].intr_name
= "llvm.exp2.f32";
723 bld_base
->op_actions
[TGSI_OPCODE_FLR
].emit
= build_tgsi_intrinsic_nomem
;
724 bld_base
->op_actions
[TGSI_OPCODE_FLR
].intr_name
= "llvm.floor.f32";
725 bld_base
->op_actions
[TGSI_OPCODE_FMA
].emit
=
726 bld_base
->op_actions
[TGSI_OPCODE_MAD
].emit
;
727 bld_base
->op_actions
[TGSI_OPCODE_FRC
].emit
= emit_frac
;
728 bld_base
->op_actions
[TGSI_OPCODE_F2I
].emit
= emit_f2i
;
729 bld_base
->op_actions
[TGSI_OPCODE_F2U
].emit
= emit_f2u
;
730 bld_base
->op_actions
[TGSI_OPCODE_FSEQ
].emit
= emit_fcmp
;
731 bld_base
->op_actions
[TGSI_OPCODE_FSGE
].emit
= emit_fcmp
;
732 bld_base
->op_actions
[TGSI_OPCODE_FSLT
].emit
= emit_fcmp
;
733 bld_base
->op_actions
[TGSI_OPCODE_FSNE
].emit
= emit_fcmp
;
734 bld_base
->op_actions
[TGSI_OPCODE_IABS
].emit
= emit_iabs
;
735 bld_base
->op_actions
[TGSI_OPCODE_IBFE
].emit
= emit_bfe
;
736 bld_base
->op_actions
[TGSI_OPCODE_IDIV
].emit
= emit_idiv
;
737 bld_base
->op_actions
[TGSI_OPCODE_IMAX
].emit
= emit_minmax_int
;
738 bld_base
->op_actions
[TGSI_OPCODE_IMIN
].emit
= emit_minmax_int
;
739 bld_base
->op_actions
[TGSI_OPCODE_IMSB
].emit
= emit_imsb
;
740 bld_base
->op_actions
[TGSI_OPCODE_INEG
].emit
= emit_ineg
;
741 bld_base
->op_actions
[TGSI_OPCODE_ISHR
].emit
= emit_ishr
;
742 bld_base
->op_actions
[TGSI_OPCODE_ISGE
].emit
= emit_icmp
;
743 bld_base
->op_actions
[TGSI_OPCODE_ISLT
].emit
= emit_icmp
;
744 bld_base
->op_actions
[TGSI_OPCODE_ISSG
].emit
= emit_ssg
;
745 bld_base
->op_actions
[TGSI_OPCODE_I2F
].emit
= emit_i2f
;
746 bld_base
->op_actions
[TGSI_OPCODE_KILL_IF
].emit
= kil_emit
;
747 bld_base
->op_actions
[TGSI_OPCODE_KILL
].emit
= kil_emit
;
748 bld_base
->op_actions
[TGSI_OPCODE_LDEXP
].emit
= build_tgsi_intrinsic_nomem
;
749 bld_base
->op_actions
[TGSI_OPCODE_LDEXP
].intr_name
= "llvm.amdgcn.ldexp.f32";
750 bld_base
->op_actions
[TGSI_OPCODE_LSB
].emit
= emit_lsb
;
751 bld_base
->op_actions
[TGSI_OPCODE_LG2
].emit
= build_tgsi_intrinsic_nomem
;
752 bld_base
->op_actions
[TGSI_OPCODE_LG2
].intr_name
= "llvm.log2.f32";
753 bld_base
->op_actions
[TGSI_OPCODE_MAX
].emit
= build_tgsi_intrinsic_nomem
;
754 bld_base
->op_actions
[TGSI_OPCODE_MAX
].intr_name
= "llvm.maxnum.f32";
755 bld_base
->op_actions
[TGSI_OPCODE_MIN
].emit
= build_tgsi_intrinsic_nomem
;
756 bld_base
->op_actions
[TGSI_OPCODE_MIN
].intr_name
= "llvm.minnum.f32";
757 bld_base
->op_actions
[TGSI_OPCODE_MOD
].emit
= emit_mod
;
758 bld_base
->op_actions
[TGSI_OPCODE_UMSB
].emit
= emit_umsb
;
759 bld_base
->op_actions
[TGSI_OPCODE_NOT
].emit
= emit_not
;
760 bld_base
->op_actions
[TGSI_OPCODE_OR
].emit
= emit_or
;
761 bld_base
->op_actions
[TGSI_OPCODE_PK2H
].emit
= emit_pk2h
;
762 bld_base
->op_actions
[TGSI_OPCODE_POPC
].emit
= build_tgsi_intrinsic_nomem
;
763 bld_base
->op_actions
[TGSI_OPCODE_POPC
].intr_name
= "llvm.ctpop.i32";
764 bld_base
->op_actions
[TGSI_OPCODE_POW
].emit
= build_tgsi_intrinsic_nomem
;
765 bld_base
->op_actions
[TGSI_OPCODE_POW
].intr_name
= "llvm.pow.f32";
766 bld_base
->op_actions
[TGSI_OPCODE_ROUND
].emit
= build_tgsi_intrinsic_nomem
;
767 bld_base
->op_actions
[TGSI_OPCODE_ROUND
].intr_name
= "llvm.rint.f32";
768 bld_base
->op_actions
[TGSI_OPCODE_RSQ
].emit
= emit_rsq
;
769 bld_base
->op_actions
[TGSI_OPCODE_SGE
].emit
= emit_set_cond
;
770 bld_base
->op_actions
[TGSI_OPCODE_SEQ
].emit
= emit_set_cond
;
771 bld_base
->op_actions
[TGSI_OPCODE_SHL
].emit
= emit_shl
;
772 bld_base
->op_actions
[TGSI_OPCODE_SLE
].emit
= emit_set_cond
;
773 bld_base
->op_actions
[TGSI_OPCODE_SLT
].emit
= emit_set_cond
;
774 bld_base
->op_actions
[TGSI_OPCODE_SNE
].emit
= emit_set_cond
;
775 bld_base
->op_actions
[TGSI_OPCODE_SGT
].emit
= emit_set_cond
;
776 bld_base
->op_actions
[TGSI_OPCODE_SIN
].emit
= build_tgsi_intrinsic_nomem
;
777 bld_base
->op_actions
[TGSI_OPCODE_SIN
].intr_name
= "llvm.sin.f32";
778 bld_base
->op_actions
[TGSI_OPCODE_SQRT
].emit
= build_tgsi_intrinsic_nomem
;
779 bld_base
->op_actions
[TGSI_OPCODE_SQRT
].intr_name
= "llvm.sqrt.f32";
780 bld_base
->op_actions
[TGSI_OPCODE_SSG
].emit
= emit_ssg
;
781 bld_base
->op_actions
[TGSI_OPCODE_TRUNC
].emit
= build_tgsi_intrinsic_nomem
;
782 bld_base
->op_actions
[TGSI_OPCODE_TRUNC
].intr_name
= "llvm.trunc.f32";
783 bld_base
->op_actions
[TGSI_OPCODE_UADD
].emit
= emit_uadd
;
784 bld_base
->op_actions
[TGSI_OPCODE_UBFE
].emit
= emit_bfe
;
785 bld_base
->op_actions
[TGSI_OPCODE_UDIV
].emit
= emit_udiv
;
786 bld_base
->op_actions
[TGSI_OPCODE_UMAX
].emit
= emit_minmax_int
;
787 bld_base
->op_actions
[TGSI_OPCODE_UMIN
].emit
= emit_minmax_int
;
788 bld_base
->op_actions
[TGSI_OPCODE_UMOD
].emit
= emit_umod
;
789 bld_base
->op_actions
[TGSI_OPCODE_USEQ
].emit
= emit_icmp
;
790 bld_base
->op_actions
[TGSI_OPCODE_USGE
].emit
= emit_icmp
;
791 bld_base
->op_actions
[TGSI_OPCODE_USHR
].emit
= emit_ushr
;
792 bld_base
->op_actions
[TGSI_OPCODE_USLT
].emit
= emit_icmp
;
793 bld_base
->op_actions
[TGSI_OPCODE_USNE
].emit
= emit_icmp
;
794 bld_base
->op_actions
[TGSI_OPCODE_U2F
].emit
= emit_u2f
;
795 bld_base
->op_actions
[TGSI_OPCODE_XOR
].emit
= emit_xor
;
796 bld_base
->op_actions
[TGSI_OPCODE_UCMP
].emit
= emit_ucmp
;
797 bld_base
->op_actions
[TGSI_OPCODE_UP2H
].emit
= emit_up2h
;
799 bld_base
->op_actions
[TGSI_OPCODE_I64MAX
].emit
= emit_minmax_int
;
800 bld_base
->op_actions
[TGSI_OPCODE_I64MIN
].emit
= emit_minmax_int
;
801 bld_base
->op_actions
[TGSI_OPCODE_U64MAX
].emit
= emit_minmax_int
;
802 bld_base
->op_actions
[TGSI_OPCODE_U64MIN
].emit
= emit_minmax_int
;
803 bld_base
->op_actions
[TGSI_OPCODE_I64ABS
].emit
= emit_iabs
;
804 bld_base
->op_actions
[TGSI_OPCODE_I64SSG
].emit
= emit_ssg
;
805 bld_base
->op_actions
[TGSI_OPCODE_I64NEG
].emit
= emit_ineg
;
807 bld_base
->op_actions
[TGSI_OPCODE_U64SEQ
].emit
= emit_icmp
;
808 bld_base
->op_actions
[TGSI_OPCODE_U64SNE
].emit
= emit_icmp
;
809 bld_base
->op_actions
[TGSI_OPCODE_U64SGE
].emit
= emit_icmp
;
810 bld_base
->op_actions
[TGSI_OPCODE_U64SLT
].emit
= emit_icmp
;
811 bld_base
->op_actions
[TGSI_OPCODE_I64SGE
].emit
= emit_icmp
;
812 bld_base
->op_actions
[TGSI_OPCODE_I64SLT
].emit
= emit_icmp
;
814 bld_base
->op_actions
[TGSI_OPCODE_U64ADD
].emit
= emit_uadd
;
815 bld_base
->op_actions
[TGSI_OPCODE_U64SHL
].emit
= emit_shl
;
816 bld_base
->op_actions
[TGSI_OPCODE_U64SHR
].emit
= emit_ushr
;
817 bld_base
->op_actions
[TGSI_OPCODE_I64SHR
].emit
= emit_ishr
;
819 bld_base
->op_actions
[TGSI_OPCODE_U64MOD
].emit
= emit_umod
;
820 bld_base
->op_actions
[TGSI_OPCODE_I64MOD
].emit
= emit_mod
;
821 bld_base
->op_actions
[TGSI_OPCODE_U64DIV
].emit
= emit_udiv
;
822 bld_base
->op_actions
[TGSI_OPCODE_I64DIV
].emit
= emit_idiv
;