2 * Copyright © 2018 Valve Corporation
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
34 #include "ac_binary.h"
35 #include "amd_family.h"
36 #include "aco_opcodes.h"
39 struct radv_nir_compiler_options
;
40 struct radv_shader_args
;
41 struct radv_shader_info
;
45 extern uint64_t debug_flags
;
49 DEBUG_VALIDATE_RA
= 0x2,
54 * Representation of the instruction's microcode encoding format
55 * Note: Some Vector ALU Formats can be combined, such that:
56 * - VOP2* | VOP3A represents a VOP2 instruction in VOP3A encoding
57 * - VOP2* | DPP represents a VOP2 instruction with data parallel primitive.
58 * - VOP2* | SDWA represents a VOP2 instruction with sub-dword addressing.
60 * (*) The same is applicable for VOP1 and VOPC instructions.
62 enum class Format
: std::uint16_t {
63 /* Pseudo Instruction Format */
65 /* Scalar ALU & Control Formats */
71 /* Scalar Memory Format */
75 /* Vector Memory Buffer Formats */
78 /* Vector Memory Image Format */
89 PSEUDO_REDUCTION
= 18,
91 /* Vector ALU Formats */
99 /* Vector Parameter Interpolation Format */
105 enum barrier_interaction
: uint8_t {
107 barrier_buffer
= 0x1,
109 barrier_atomic
= 0x4,
110 barrier_shared
= 0x8,
111 /* used for geometry shaders to ensure vertex data writes are before the
112 * GS_DONE s_sendmsg. */
113 barrier_gs_data
= 0x10,
114 /* used for geometry shaders to ensure s_sendmsg instructions are in-order. */
115 barrier_gs_sendmsg
= 0x20,
116 /* used by barriers. created by s_barrier */
117 barrier_barrier
= 0x40,
129 /* Note that v_rcp_f32, v_exp_f32, v_log_f32, v_sqrt_f32, v_rsq_f32 and
130 * v_mad_f32/v_madak_f32/v_madmk_f32/v_mac_f32 always flush denormals. */
131 fp_denorm_flush
= 0x0,
132 fp_denorm_keep
= 0x3,
136 /* matches encoding of the MODE register */
140 fp_round round16_64
:2;
142 unsigned denorm16_64
:2;
146 /* if false, optimizations which may remove infs/nan/-0.0 can be done */
147 bool preserve_signed_zero_inf_nan32
:1;
148 bool preserve_signed_zero_inf_nan16_64
:1;
149 /* if false, optimizations which may remove denormal flushing can be done */
150 bool must_flush_denorms32
:1;
151 bool must_flush_denorms16_64
:1;
152 bool care_about_round32
:1;
153 bool care_about_round16_64
:1;
155 /* Returns true if instructions using the mode "other" can safely use the
156 * current one instead. */
157 bool canReplace(float_mode other
) const noexcept
{
158 return val
== other
.val
&&
159 (preserve_signed_zero_inf_nan32
|| !other
.preserve_signed_zero_inf_nan32
) &&
160 (preserve_signed_zero_inf_nan16_64
|| !other
.preserve_signed_zero_inf_nan16_64
) &&
161 (must_flush_denorms32
|| !other
.must_flush_denorms32
) &&
162 (must_flush_denorms16_64
|| !other
.must_flush_denorms16_64
) &&
163 (care_about_round32
|| !other
.care_about_round32
) &&
164 (care_about_round16_64
|| !other
.care_about_round16_64
);
168 constexpr Format
asVOP3(Format format
) {
169 return (Format
) ((uint32_t) Format::VOP3
| (uint32_t) format
);
197 /* these are used for WWM and spills to vgpr */
198 v1_linear
= v1
| (1 << 6),
199 v2_linear
= v2
| (1 << 6),
202 RegClass() = default;
203 constexpr RegClass(RC rc
)
205 constexpr RegClass(RegType type
, unsigned size
)
206 : rc((RC
) ((type
== RegType::vgpr
? 1 << 5 : 0) | size
)) {}
208 constexpr operator RC() const { return rc
; }
209 explicit operator bool() = delete;
211 constexpr RegType
type() const { return rc
<= RC::s16
? RegType::sgpr
: RegType::vgpr
; }
212 constexpr unsigned size() const { return (unsigned) rc
& 0x1F; }
213 constexpr bool is_linear() const { return rc
<= RC::s16
|| rc
& (1 << 6); }
214 constexpr RegClass
as_linear() const { return RegClass((RC
) (rc
| (1 << 6))); }
220 /* transitional helper expressions */
221 static constexpr RegClass s1
{RegClass::s1
};
222 static constexpr RegClass s2
{RegClass::s2
};
223 static constexpr RegClass s3
{RegClass::s3
};
224 static constexpr RegClass s4
{RegClass::s4
};
225 static constexpr RegClass s8
{RegClass::s8
};
226 static constexpr RegClass s16
{RegClass::s16
};
227 static constexpr RegClass v1
{RegClass::v1
};
228 static constexpr RegClass v2
{RegClass::v2
};
229 static constexpr RegClass v3
{RegClass::v3
};
230 static constexpr RegClass v4
{RegClass::v4
};
231 static constexpr RegClass v5
{RegClass::v5
};
232 static constexpr RegClass v6
{RegClass::v6
};
233 static constexpr RegClass v7
{RegClass::v7
};
234 static constexpr RegClass v8
{RegClass::v8
};
238 * Each temporary virtual register has a
239 * register class (i.e. size and type)
244 constexpr Temp(uint32_t id
, RegClass cls
) noexcept
245 : id_(id
), reg_class(cls
) {}
247 constexpr uint32_t id() const noexcept
{ return id_
; }
248 constexpr RegClass
regClass() const noexcept
{ return reg_class
; }
250 constexpr unsigned size() const noexcept
{ return reg_class
.size(); }
251 constexpr RegType
type() const noexcept
{ return reg_class
.type(); }
252 constexpr bool is_linear() const noexcept
{ return reg_class
.is_linear(); }
254 constexpr bool operator <(Temp other
) const noexcept
{ return id() < other
.id(); }
255 constexpr bool operator==(Temp other
) const noexcept
{ return id() == other
.id(); }
256 constexpr bool operator!=(Temp other
) const noexcept
{ return id() != other
.id(); }
265 * Represents the physical register for each
266 * Operand and Definition.
269 constexpr PhysReg() = default;
270 explicit constexpr PhysReg(unsigned r
) : reg_b(r
<< 2) {}
271 constexpr unsigned reg() const { return reg_b
>> 2; }
272 constexpr unsigned byte() const { return reg_b
& 0x3; }
273 constexpr operator unsigned() const { return reg(); }
278 /* helper expressions for special registers */
279 static constexpr PhysReg m0
{124};
280 static constexpr PhysReg vcc
{106};
281 static constexpr PhysReg vcc_hi
{107};
282 static constexpr PhysReg sgpr_null
{125}; /* GFX10+ */
283 static constexpr PhysReg exec
{126};
284 static constexpr PhysReg exec_lo
{126};
285 static constexpr PhysReg exec_hi
{127};
286 static constexpr PhysReg vccz
{251};
287 static constexpr PhysReg execz
{252};
288 static constexpr PhysReg scc
{253};
292 * Initially, each Operand refers to either
293 * a temporary virtual register
294 * or to a constant value
295 * Temporary registers get mapped to physical register during RA
296 * Constant values are inlined into the instruction sequence.
302 : reg_(PhysReg
{128}), isTemp_(false), isFixed_(true), isConstant_(false),
303 isKill_(false), isUndef_(true), isFirstKill_(false), is64BitConst_(false),
304 isLateKill_(false) {}
306 explicit Operand(Temp r
) noexcept
313 setFixed(PhysReg
{128});
316 explicit Operand(uint32_t v
, bool is64bit
= false) noexcept
320 is64BitConst_
= is64bit
;
322 setFixed(PhysReg
{128 + v
});
323 else if (v
>= 0xFFFFFFF0) /* [-16 .. -1] */
324 setFixed(PhysReg
{192 - v
});
325 else if (v
== 0x3f000000) /* 0.5 */
326 setFixed(PhysReg
{240});
327 else if (v
== 0xbf000000) /* -0.5 */
328 setFixed(PhysReg
{241});
329 else if (v
== 0x3f800000) /* 1.0 */
330 setFixed(PhysReg
{242});
331 else if (v
== 0xbf800000) /* -1.0 */
332 setFixed(PhysReg
{243});
333 else if (v
== 0x40000000) /* 2.0 */
334 setFixed(PhysReg
{244});
335 else if (v
== 0xc0000000) /* -2.0 */
336 setFixed(PhysReg
{245});
337 else if (v
== 0x40800000) /* 4.0 */
338 setFixed(PhysReg
{246});
339 else if (v
== 0xc0800000) /* -4.0 */
340 setFixed(PhysReg
{247});
341 else { /* Literal Constant */
342 assert(!is64bit
&& "attempt to create a 64-bit literal constant");
343 setFixed(PhysReg
{255});
346 explicit Operand(uint64_t v
) noexcept
349 is64BitConst_
= true;
351 data_
.i
= (uint32_t) v
;
352 setFixed(PhysReg
{128 + (uint32_t) v
});
353 } else if (v
>= 0xFFFFFFFFFFFFFFF0) { /* [-16 .. -1] */
354 data_
.i
= (uint32_t) v
;
355 setFixed(PhysReg
{192 - (uint32_t) v
});
356 } else if (v
== 0x3FE0000000000000) { /* 0.5 */
357 data_
.i
= 0x3f000000;
358 setFixed(PhysReg
{240});
359 } else if (v
== 0xBFE0000000000000) { /* -0.5 */
360 data_
.i
= 0xbf000000;
361 setFixed(PhysReg
{241});
362 } else if (v
== 0x3FF0000000000000) { /* 1.0 */
363 data_
.i
= 0x3f800000;
364 setFixed(PhysReg
{242});
365 } else if (v
== 0xBFF0000000000000) { /* -1.0 */
366 data_
.i
= 0xbf800000;
367 setFixed(PhysReg
{243});
368 } else if (v
== 0x4000000000000000) { /* 2.0 */
369 data_
.i
= 0x40000000;
370 setFixed(PhysReg
{244});
371 } else if (v
== 0xC000000000000000) { /* -2.0 */
372 data_
.i
= 0xc0000000;
373 setFixed(PhysReg
{245});
374 } else if (v
== 0x4010000000000000) { /* 4.0 */
375 data_
.i
= 0x40800000;
376 setFixed(PhysReg
{246});
377 } else if (v
== 0xC010000000000000) { /* -4.0 */
378 data_
.i
= 0xc0800000;
379 setFixed(PhysReg
{247});
380 } else { /* Literal Constant: we don't know if it is a long or double.*/
382 assert(false && "attempt to create a 64-bit literal constant");
385 explicit Operand(RegClass type
) noexcept
388 data_
.temp
= Temp(0, type
);
389 setFixed(PhysReg
{128});
391 explicit Operand(PhysReg reg
, RegClass type
) noexcept
393 data_
.temp
= Temp(0, type
);
397 constexpr bool isTemp() const noexcept
402 constexpr void setTemp(Temp t
) noexcept
{
403 assert(!isConstant_
);
408 constexpr Temp
getTemp() const noexcept
413 constexpr uint32_t tempId() const noexcept
415 return data_
.temp
.id();
418 constexpr bool hasRegClass() const noexcept
420 return isTemp() || isUndefined();
423 constexpr RegClass
regClass() const noexcept
425 return data_
.temp
.regClass();
428 constexpr unsigned size() const noexcept
431 return is64BitConst_
? 2 : 1;
433 return data_
.temp
.size();
436 constexpr bool isFixed() const noexcept
441 constexpr PhysReg
physReg() const noexcept
446 constexpr void setFixed(PhysReg reg
) noexcept
448 isFixed_
= reg
!= unsigned(-1);
452 constexpr bool isConstant() const noexcept
457 constexpr bool isLiteral() const noexcept
459 return isConstant() && reg_
== 255;
462 constexpr bool isUndefined() const noexcept
467 constexpr uint32_t constantValue() const noexcept
472 constexpr bool constantEquals(uint32_t cmp
) const noexcept
474 return isConstant() && constantValue() == cmp
;
477 constexpr uint64_t constantValue64(bool signext
=false) const noexcept
482 else if (reg_
<= 208)
483 return 0xFFFFFFFFFFFFFFFF - (reg_
- 193);
487 return 0x3FE0000000000000;
489 return 0xBFE0000000000000;
491 return 0x3FF0000000000000;
493 return 0xBFF0000000000000;
495 return 0x4000000000000000;
497 return 0xC000000000000000;
499 return 0x4010000000000000;
501 return 0xC010000000000000;
504 return (signext
&& (data_
.i
& 0x80000000u
) ? 0xffffffff00000000ull
: 0ull) | data_
.i
;
507 /* Indicates that the killed operand's live range intersects with the
508 * instruction's definitions. Unlike isKill() and isFirstKill(), this is
509 * not set by liveness analysis. */
510 constexpr void setLateKill(bool flag
) noexcept
515 constexpr bool isLateKill() const noexcept
520 constexpr void setKill(bool flag
) noexcept
527 constexpr bool isKill() const noexcept
529 return isKill_
|| isFirstKill();
532 constexpr void setFirstKill(bool flag
) noexcept
539 /* When there are multiple operands killing the same temporary,
540 * isFirstKill() is only returns true for the first one. */
541 constexpr bool isFirstKill() const noexcept
546 constexpr bool isKillBeforeDef() const noexcept
548 return isKill() && !isLateKill();
551 constexpr bool isFirstKillBeforeDef() const noexcept
553 return isFirstKill() && !isLateKill();
556 constexpr bool operator == (Operand other
) const noexcept
558 if (other
.size() != size())
560 if (isFixed() != other
.isFixed() || isKillBeforeDef() != other
.isKillBeforeDef())
562 if (isFixed() && other
.isFixed() && physReg() != other
.physReg())
565 return other
.isLiteral() && other
.constantValue() == constantValue();
566 else if (isConstant())
567 return other
.isConstant() && other
.physReg() == physReg();
568 else if (isUndefined())
569 return other
.isUndefined() && other
.regClass() == regClass();
571 return other
.isTemp() && other
.getTemp() == getTemp();
577 Temp temp
= Temp(0, s1
);
584 uint8_t isConstant_
:1;
587 uint8_t isFirstKill_
:1;
588 uint8_t is64BitConst_
:1;
589 uint8_t isLateKill_
:1;
591 /* can't initialize bit-fields in c++11, so work around using a union */
592 uint8_t control_
= 0;
598 * Definitions are the results of Instructions
599 * and refer to temporary virtual registers
600 * which are later mapped to physical registers
602 class Definition final
605 constexpr Definition() : temp(Temp(0, s1
)), reg_(0), isFixed_(0), hasHint_(0), isKill_(0) {}
606 Definition(uint32_t index
, RegClass type
) noexcept
607 : temp(index
, type
) {}
608 explicit Definition(Temp tmp
) noexcept
610 Definition(PhysReg reg
, RegClass type
) noexcept
611 : temp(Temp(0, type
))
615 Definition(uint32_t tmpId
, PhysReg reg
, RegClass type
) noexcept
616 : temp(Temp(tmpId
, type
))
621 constexpr bool isTemp() const noexcept
626 constexpr Temp
getTemp() const noexcept
631 constexpr uint32_t tempId() const noexcept
636 constexpr void setTemp(Temp t
) noexcept
{
640 constexpr RegClass
regClass() const noexcept
642 return temp
.regClass();
645 constexpr unsigned size() const noexcept
650 constexpr bool isFixed() const noexcept
655 constexpr PhysReg
physReg() const noexcept
660 constexpr void setFixed(PhysReg reg
) noexcept
666 constexpr void setHint(PhysReg reg
) noexcept
672 constexpr bool hasHint() const noexcept
677 constexpr void setKill(bool flag
) noexcept
682 constexpr bool isKill() const noexcept
688 Temp temp
= Temp(0, s1
);
696 /* can't initialize bit-fields in c++11, so work around using a union */
697 uint8_t control_
= 0;
708 aco::span
<Operand
> operands
;
709 aco::span
<Definition
> definitions
;
711 constexpr bool isVALU() const noexcept
713 return ((uint16_t) format
& (uint16_t) Format::VOP1
) == (uint16_t) Format::VOP1
714 || ((uint16_t) format
& (uint16_t) Format::VOP2
) == (uint16_t) Format::VOP2
715 || ((uint16_t) format
& (uint16_t) Format::VOPC
) == (uint16_t) Format::VOPC
716 || ((uint16_t) format
& (uint16_t) Format::VOP3A
) == (uint16_t) Format::VOP3A
717 || ((uint16_t) format
& (uint16_t) Format::VOP3B
) == (uint16_t) Format::VOP3B
718 || ((uint16_t) format
& (uint16_t) Format::VOP3P
) == (uint16_t) Format::VOP3P
;
721 constexpr bool isSALU() const noexcept
723 return format
== Format::SOP1
||
724 format
== Format::SOP2
||
725 format
== Format::SOPC
||
726 format
== Format::SOPK
||
727 format
== Format::SOPP
;
730 constexpr bool isVMEM() const noexcept
732 return format
== Format::MTBUF
||
733 format
== Format::MUBUF
||
734 format
== Format::MIMG
;
737 constexpr bool isDPP() const noexcept
739 return (uint16_t) format
& (uint16_t) Format::DPP
;
742 constexpr bool isVOP3() const noexcept
744 return ((uint16_t) format
& (uint16_t) Format::VOP3A
) ||
745 ((uint16_t) format
& (uint16_t) Format::VOP3B
) ||
746 format
== Format::VOP3P
;
749 constexpr bool isSDWA() const noexcept
751 return (uint16_t) format
& (uint16_t) Format::SDWA
;
754 constexpr bool isFlatOrGlobal() const noexcept
756 return format
== Format::FLAT
|| format
== Format::GLOBAL
;
759 constexpr bool usesModifiers() const noexcept
;
761 constexpr bool reads_exec() const noexcept
763 for (const Operand
& op
: operands
) {
764 if (op
.isFixed() && op
.physReg() == exec
)
771 struct SOPK_instruction
: public Instruction
{
775 struct SOPP_instruction
: public Instruction
{
780 struct SOPC_instruction
: public Instruction
{
783 struct SOP1_instruction
: public Instruction
{
786 struct SOP2_instruction
: public Instruction
{
790 * Scalar Memory Format:
791 * For s_(buffer_)load_dword*:
792 * Operand(0): SBASE - SGPR-pair which provides base address
793 * Operand(1): Offset - immediate (un)signed offset or SGPR
794 * Operand(2) / Definition(0): SDATA - SGPR for read / write result
795 * Operand(n-1): SOffset - SGPR offset (Vega only)
797 * Having no operands is also valid for instructions such as s_dcache_inv.
800 struct SMEM_instruction
: public Instruction
{
801 bool glc
: 1; /* VI+: globally coherent */
802 bool dlc
: 1; /* NAVI: device level coherent */
803 bool nv
: 1; /* VEGA only: Non-volatile */
804 bool can_reorder
: 1;
805 bool disable_wqm
: 1;
806 barrier_interaction barrier
;
809 struct VOP1_instruction
: public Instruction
{
812 struct VOP2_instruction
: public Instruction
{
815 struct VOPC_instruction
: public Instruction
{
818 struct VOP3A_instruction
: public Instruction
{
827 * Data Parallel Primitives Format:
828 * This format can be used for VOP1, VOP2 or VOPC instructions.
829 * The swizzle applies to the src0 operand.
832 struct DPP_instruction
: public Instruction
{
836 uint8_t row_mask
: 4;
837 uint8_t bank_mask
: 4;
841 struct Interp_instruction
: public Instruction
{
847 * Local and Global Data Sharing instructions
848 * Operand(0): ADDR - VGPR which supplies the address.
849 * Operand(1): DATA0 - First data VGPR.
850 * Operand(2): DATA1 - Second data VGPR.
851 * Operand(n-1): M0 - LDS size.
852 * Definition(0): VDST - Destination VGPR when results returned to VGPRs.
855 struct DS_instruction
: public Instruction
{
862 * Vector Memory Untyped-buffer Instructions
863 * Operand(0): SRSRC - Specifies which SGPR supplies T# (resource constant)
864 * Operand(1): VADDR - Address source. Can carry an index and/or offset
865 * Operand(2): SOFFSET - SGPR to supply unsigned byte offset. (SGPR, M0, or inline constant)
866 * Operand(3) / Definition(0): VDATA - Vector GPR for write result / read data
869 struct MUBUF_instruction
: public Instruction
{
870 uint16_t offset
: 12; /* Unsigned byte offset - 12 bit */
871 bool offen
: 1; /* Supply an offset from VGPR (VADDR) */
872 bool idxen
: 1; /* Supply an index from VGPR (VADDR) */
873 bool addr64
: 1; /* SI, CIK: Address size is 64-bit */
874 bool glc
: 1; /* globally coherent */
875 bool dlc
: 1; /* NAVI: device level coherent */
876 bool slc
: 1; /* system level coherent */
877 bool tfe
: 1; /* texture fail enable */
878 bool lds
: 1; /* Return read-data to LDS instead of VGPRs */
879 bool disable_wqm
: 1; /* Require an exec mask without helper invocations */
880 bool can_reorder
: 1;
881 barrier_interaction barrier
;
885 * Vector Memory Typed-buffer Instructions
886 * Operand(0): SRSRC - Specifies which SGPR supplies T# (resource constant)
887 * Operand(1): VADDR - Address source. Can carry an index and/or offset
888 * Operand(2): SOFFSET - SGPR to supply unsigned byte offset. (SGPR, M0, or inline constant)
889 * Operand(3) / Definition(0): VDATA - Vector GPR for write result / read data
892 struct MTBUF_instruction
: public Instruction
{
893 uint16_t offset
; /* Unsigned byte offset - 12 bit */
894 uint8_t dfmt
: 4; /* Data Format of data in memory buffer */
895 uint8_t nfmt
: 3; /* Numeric format of data in memory */
896 bool offen
: 1; /* Supply an offset from VGPR (VADDR) */
897 bool idxen
: 1; /* Supply an index from VGPR (VADDR) */
898 bool glc
: 1; /* globally coherent */
899 bool dlc
: 1; /* NAVI: device level coherent */
900 bool slc
: 1; /* system level coherent */
901 bool tfe
: 1; /* texture fail enable */
902 bool disable_wqm
: 1; /* Require an exec mask without helper invocations */
903 bool can_reorder
: 1;
904 barrier_interaction barrier
;
908 * Vector Memory Image Instructions
909 * Operand(0) SRSRC - Scalar GPR that specifies the resource constant.
910 * Operand(1): SSAMP - Scalar GPR that specifies sampler constant.
911 * or VDATA - Vector GPR for write data.
912 * Operand(2): VADDR - Address source. Can carry an offset or an index.
913 * Definition(0): VDATA - Vector GPR for read result.
916 struct MIMG_instruction
: public Instruction
{
917 uint8_t dmask
; /* Data VGPR enable mask */
918 uint8_t dim
: 3; /* NAVI: dimensionality */
919 bool unrm
: 1; /* Force address to be un-normalized */
920 bool dlc
: 1; /* NAVI: device level coherent */
921 bool glc
: 1; /* globally coherent */
922 bool slc
: 1; /* system level coherent */
923 bool tfe
: 1; /* texture fail enable */
924 bool da
: 1; /* declare an array */
925 bool lwe
: 1; /* Force data to be un-normalized */
926 bool r128
: 1; /* NAVI: Texture resource size */
927 bool a16
: 1; /* VEGA, NAVI: Address components are 16-bits */
928 bool d16
: 1; /* Convert 32-bit data to 16-bit data */
929 bool disable_wqm
: 1; /* Require an exec mask without helper invocations */
930 bool can_reorder
: 1;
931 barrier_interaction barrier
;
935 * Flat/Scratch/Global Instructions
938 * Operand(2) / Definition(0): DATA/VDST
941 struct FLAT_instruction
: public Instruction
{
942 uint16_t offset
; /* Vega/Navi only */
943 bool slc
: 1; /* system level coherent */
944 bool glc
: 1; /* globally coherent */
945 bool dlc
: 1; /* NAVI: device level coherent */
948 bool disable_wqm
: 1; /* Require an exec mask without helper invocations */
949 bool can_reorder
: 1;
950 barrier_interaction barrier
;
953 struct Export_instruction
: public Instruction
{
954 uint8_t enabled_mask
;
961 struct Pseudo_instruction
: public Instruction
{
963 PhysReg scratch_sgpr
; /* might not be valid if it's not needed */
966 struct Pseudo_branch_instruction
: public Instruction
{
967 /* target[0] is the block index of the branch target.
968 * For conditional branches, target[1] contains the fall-through alternative.
969 * A value of 0 means the target has not been initialized (BB0 cannot be a branch target).
974 struct Pseudo_barrier_instruction
: public Instruction
{
991 gfx10_wave64_bpermute
995 * Subgroup Reduction Instructions, everything except for the data to be
996 * reduced and the result as inserted by setup_reduce_temp().
997 * Operand(0): data to be reduced
998 * Operand(1): reduce temporary
999 * Operand(2): vector temporary
1000 * Definition(0): result
1001 * Definition(1): scalar temporary
1002 * Definition(2): scalar identity temporary (not used to store identity on GFX10)
1003 * Definition(3): scc clobber
1004 * Definition(4): vcc clobber
1007 struct Pseudo_reduction_instruction
: public Instruction
{
1009 unsigned cluster_size
; // must be 0 for scans
1012 struct instr_deleter_functor
{
1013 void operator()(void* p
) {
1018 template<typename T
>
1019 using aco_ptr
= std::unique_ptr
<T
, instr_deleter_functor
>;
1021 template<typename T
>
1022 T
* create_instruction(aco_opcode opcode
, Format format
, uint32_t num_operands
, uint32_t num_definitions
)
1024 std::size_t size
= sizeof(T
) + num_operands
* sizeof(Operand
) + num_definitions
* sizeof(Definition
);
1025 char *data
= (char*) calloc(1, size
);
1026 T
* inst
= (T
*) data
;
1028 inst
->opcode
= opcode
;
1029 inst
->format
= format
;
1031 uint16_t operands_offset
= data
+ sizeof(T
) - (char*)&inst
->operands
;
1032 inst
->operands
= aco::span
<Operand
>(operands_offset
, num_operands
);
1033 uint16_t definitions_offset
= (char*)inst
->operands
.end() - (char*)&inst
->definitions
;
1034 inst
->definitions
= aco::span
<Definition
>(definitions_offset
, num_definitions
);
1039 constexpr bool Instruction::usesModifiers() const noexcept
1041 if (isDPP() || isSDWA())
1045 const VOP3A_instruction
*vop3
= static_cast<const VOP3A_instruction
*>(this);
1046 for (unsigned i
= 0; i
< operands
.size(); i
++) {
1047 if (vop3
->abs
[i
] || vop3
->neg
[i
])
1050 return vop3
->opsel
|| vop3
->clamp
|| vop3
->omod
;
1053 constexpr bool is_phi(Instruction
* instr
)
1055 return instr
->opcode
== aco_opcode::p_phi
|| instr
->opcode
== aco_opcode::p_linear_phi
;
1058 static inline bool is_phi(aco_ptr
<Instruction
>& instr
)
1060 return is_phi(instr
.get());
1063 barrier_interaction
get_barrier_interaction(Instruction
* instr
);
1065 bool is_dead(const std::vector
<uint16_t>& uses
, Instruction
*instr
);
1068 /* uniform indicates that leaving this block,
1069 * all actives lanes stay active */
1070 block_kind_uniform
= 1 << 0,
1071 block_kind_top_level
= 1 << 1,
1072 block_kind_loop_preheader
= 1 << 2,
1073 block_kind_loop_header
= 1 << 3,
1074 block_kind_loop_exit
= 1 << 4,
1075 block_kind_continue
= 1 << 5,
1076 block_kind_break
= 1 << 6,
1077 block_kind_continue_or_break
= 1 << 7,
1078 block_kind_discard
= 1 << 8,
1079 block_kind_branch
= 1 << 9,
1080 block_kind_merge
= 1 << 10,
1081 block_kind_invert
= 1 << 11,
1082 block_kind_uses_discard_if
= 1 << 12,
1083 block_kind_needs_lowering
= 1 << 13,
1084 block_kind_uses_demote
= 1 << 14,
1085 block_kind_export_end
= 1 << 15,
1089 struct RegisterDemand
{
1090 constexpr RegisterDemand() = default;
1091 constexpr RegisterDemand(const int16_t v
, const int16_t s
) noexcept
1092 : vgpr
{v
}, sgpr
{s
} {}
1096 constexpr friend bool operator==(const RegisterDemand a
, const RegisterDemand b
) noexcept
{
1097 return a
.vgpr
== b
.vgpr
&& a
.sgpr
== b
.sgpr
;
1100 constexpr bool exceeds(const RegisterDemand other
) const noexcept
{
1101 return vgpr
> other
.vgpr
|| sgpr
> other
.sgpr
;
1104 constexpr RegisterDemand
operator+(const Temp t
) const noexcept
{
1105 if (t
.type() == RegType::sgpr
)
1106 return RegisterDemand( vgpr
, sgpr
+ t
.size() );
1108 return RegisterDemand( vgpr
+ t
.size(), sgpr
);
1111 constexpr RegisterDemand
operator+(const RegisterDemand other
) const noexcept
{
1112 return RegisterDemand(vgpr
+ other
.vgpr
, sgpr
+ other
.sgpr
);
1115 constexpr RegisterDemand
operator-(const RegisterDemand other
) const noexcept
{
1116 return RegisterDemand(vgpr
- other
.vgpr
, sgpr
- other
.sgpr
);
1119 constexpr RegisterDemand
& operator+=(const RegisterDemand other
) noexcept
{
1125 constexpr RegisterDemand
& operator-=(const RegisterDemand other
) noexcept
{
1131 constexpr RegisterDemand
& operator+=(const Temp t
) noexcept
{
1132 if (t
.type() == RegType::sgpr
)
1139 constexpr RegisterDemand
& operator-=(const Temp t
) noexcept
{
1140 if (t
.type() == RegType::sgpr
)
1147 constexpr void update(const RegisterDemand other
) noexcept
{
1148 vgpr
= std::max(vgpr
, other
.vgpr
);
1149 sgpr
= std::max(sgpr
, other
.sgpr
);
1158 unsigned offset
= 0;
1159 std::vector
<aco_ptr
<Instruction
>> instructions
;
1160 std::vector
<unsigned> logical_preds
;
1161 std::vector
<unsigned> linear_preds
;
1162 std::vector
<unsigned> logical_succs
;
1163 std::vector
<unsigned> linear_succs
;
1164 RegisterDemand register_demand
= RegisterDemand();
1165 uint16_t loop_nest_depth
= 0;
1167 int logical_idom
= -1;
1168 int linear_idom
= -1;
1169 Temp live_out_exec
= Temp();
1171 /* this information is needed for predecessors to blocks with phis when
1172 * moving out of ssa */
1173 bool scc_live_out
= false;
1174 PhysReg scratch_sgpr
= PhysReg(); /* only needs to be valid if scc_live_out != false */
1176 Block(unsigned idx
) : index(idx
) {}
1177 Block() : index(0) {}
1180 using Stage
= uint16_t;
1182 /* software stages */
1183 static constexpr Stage sw_vs
= 1 << 0;
1184 static constexpr Stage sw_gs
= 1 << 1;
1185 static constexpr Stage sw_tcs
= 1 << 2;
1186 static constexpr Stage sw_tes
= 1 << 3;
1187 static constexpr Stage sw_fs
= 1 << 4;
1188 static constexpr Stage sw_cs
= 1 << 5;
1189 static constexpr Stage sw_gs_copy
= 1 << 6;
1190 static constexpr Stage sw_mask
= 0x7f;
1192 /* hardware stages (can't be OR'd, just a mask for convenience when testing multiple) */
1193 static constexpr Stage hw_vs
= 1 << 7;
1194 static constexpr Stage hw_es
= 1 << 8; /* not on GFX9. combined into GS on GFX9 (and GFX10/legacy). */
1195 static constexpr Stage hw_gs
= 1 << 9;
1196 static constexpr Stage hw_ls
= 1 << 10; /* not on GFX9. combined into HS on GFX9 (and GFX10/legacy). */
1197 static constexpr Stage hw_hs
= 1 << 11;
1198 static constexpr Stage hw_fs
= 1 << 12;
1199 static constexpr Stage hw_cs
= 1 << 13;
1200 static constexpr Stage hw_mask
= 0x7f << 7;
1202 /* possible settings of Program::stage */
1203 static constexpr Stage vertex_vs
= sw_vs
| hw_vs
;
1204 static constexpr Stage fragment_fs
= sw_fs
| hw_fs
;
1205 static constexpr Stage compute_cs
= sw_cs
| hw_cs
;
1206 static constexpr Stage tess_eval_vs
= sw_tes
| hw_vs
;
1207 static constexpr Stage gs_copy_vs
= sw_gs_copy
| hw_vs
;
1209 static constexpr Stage ngg_vertex_gs
= sw_vs
| hw_gs
;
1210 static constexpr Stage ngg_vertex_geometry_gs
= sw_vs
| sw_gs
| hw_gs
;
1211 static constexpr Stage ngg_tess_eval_geometry_gs
= sw_tes
| sw_gs
| hw_gs
;
1212 static constexpr Stage ngg_vertex_tess_control_hs
= sw_vs
| sw_tcs
| hw_hs
;
1213 /* GFX9 (and GFX10 if NGG isn't used) */
1214 static constexpr Stage vertex_geometry_gs
= sw_vs
| sw_gs
| hw_gs
;
1215 static constexpr Stage vertex_tess_control_hs
= sw_vs
| sw_tcs
| hw_hs
;
1216 static constexpr Stage tess_eval_geometry_gs
= sw_tes
| sw_gs
| hw_gs
;
1218 static constexpr Stage vertex_ls
= sw_vs
| hw_ls
; /* vertex before tesselation control */
1219 static constexpr Stage vertex_es
= sw_vs
| hw_es
; /* vertex before geometry */
1220 static constexpr Stage tess_control_hs
= sw_tcs
| hw_hs
;
1221 static constexpr Stage tess_eval_es
= sw_tes
| hw_es
; /* tesselation evaluation before geometry */
1222 static constexpr Stage geometry_gs
= sw_gs
| hw_gs
;
1226 statistic_instructions
,
1230 statistic_vmem_clauses
,
1231 statistic_smem_clauses
,
1232 statistic_vmem_score
,
1233 statistic_smem_score
,
1234 statistic_sgpr_presched
,
1235 statistic_vgpr_presched
,
1239 class Program final
{
1241 float_mode next_fp_mode
;
1242 std::vector
<Block
> blocks
;
1243 RegisterDemand max_reg_demand
= RegisterDemand();
1244 uint16_t num_waves
= 0;
1245 uint16_t max_waves
= 0; /* maximum number of waves, regardless of register usage */
1246 ac_shader_config
* config
;
1247 struct radv_shader_info
*info
;
1248 enum chip_class chip_class
;
1249 enum radeon_family family
;
1252 Stage stage
; /* Stage */
1253 bool needs_exact
= false; /* there exists an instruction with disable_wqm = true */
1254 bool needs_wqm
= false; /* there exists a p_wqm instruction */
1255 bool wb_smem_l1_on_end
= false;
1257 std::vector
<uint8_t> constant_data
;
1258 Temp private_segment_buffer
;
1259 Temp scratch_offset
;
1261 uint16_t min_waves
= 0;
1262 uint16_t lds_alloc_granule
;
1263 uint32_t lds_limit
; /* in bytes */
1264 bool has_16bank_lds
;
1265 uint16_t vgpr_limit
;
1266 uint16_t sgpr_limit
;
1267 uint16_t physical_sgprs
;
1268 uint16_t sgpr_alloc_granule
; /* minus one. must be power of two */
1269 uint16_t vgpr_alloc_granule
; /* minus one. must be power of two */
1270 unsigned workgroup_size
; /* if known; otherwise UINT_MAX */
1272 bool xnack_enabled
= false;
1274 bool needs_vcc
= false;
1275 bool needs_flat_scr
= false;
1277 bool collect_statistics
= false;
1278 uint32_t statistics
[num_statistics
];
1280 uint32_t allocateId()
1282 assert(allocationID
<= 16777215);
1283 return allocationID
++;
1286 uint32_t peekAllocationId()
1288 return allocationID
;
1291 void setAllocationId(uint32_t id
)
1296 Block
* create_and_insert_block() {
1297 blocks
.emplace_back(blocks
.size());
1298 blocks
.back().fp_mode
= next_fp_mode
;
1299 return &blocks
.back();
1302 Block
* insert_block(Block
&& block
) {
1303 block
.index
= blocks
.size();
1304 block
.fp_mode
= next_fp_mode
;
1305 blocks
.emplace_back(std::move(block
));
1306 return &blocks
.back();
1310 uint32_t allocationID
= 1;
1314 /* live temps out per block */
1315 std::vector
<std::set
<Temp
>> live_out
;
1316 /* register demand (sgpr/vgpr) per instruction per block */
1317 std::vector
<std::vector
<RegisterDemand
>> register_demand
;
1320 void select_program(Program
*program
,
1321 unsigned shader_count
,
1322 struct nir_shader
*const *shaders
,
1323 ac_shader_config
* config
,
1324 struct radv_shader_args
*args
);
1325 void select_gs_copy_shader(Program
*program
, struct nir_shader
*gs_shader
,
1326 ac_shader_config
* config
,
1327 struct radv_shader_args
*args
);
1329 void lower_wqm(Program
* program
, live
& live_vars
,
1330 const struct radv_nir_compiler_options
*options
);
1331 void lower_bool_phis(Program
* program
);
1332 void calc_min_waves(Program
* program
);
1333 void update_vgpr_sgpr_demand(Program
* program
, const RegisterDemand new_demand
);
1334 live
live_var_analysis(Program
* program
, const struct radv_nir_compiler_options
*options
);
1335 std::vector
<uint16_t> dead_code_analysis(Program
*program
);
1336 void dominator_tree(Program
* program
);
1337 void insert_exec_mask(Program
*program
);
1338 void value_numbering(Program
* program
);
1339 void optimize(Program
* program
);
1340 void setup_reduce_temp(Program
* program
);
1341 void lower_to_cssa(Program
* program
, live
& live_vars
, const struct radv_nir_compiler_options
*options
);
1342 void register_allocation(Program
*program
, std::vector
<std::set
<Temp
>> live_out_per_block
);
1343 void ssa_elimination(Program
* program
);
1344 void lower_to_hw_instr(Program
* program
);
1345 void schedule_program(Program
* program
, live
& live_vars
);
1346 void spill(Program
* program
, live
& live_vars
, const struct radv_nir_compiler_options
*options
);
1347 void insert_wait_states(Program
* program
);
1348 void insert_NOPs(Program
* program
);
1349 unsigned emit_program(Program
* program
, std::vector
<uint32_t>& code
);
1350 void print_asm(Program
*program
, std::vector
<uint32_t>& binary
,
1351 unsigned exec_size
, std::ostream
& out
);
1352 void validate(Program
* program
, FILE *output
);
1353 bool validate_ra(Program
* program
, const struct radv_nir_compiler_options
*options
, FILE *output
);
1355 void perfwarn(bool cond
, const char *msg
, Instruction
*instr
=NULL
);
1357 #define perfwarn(program, cond, msg, ...) do {} while(0)
1360 void collect_presched_stats(Program
*program
);
1361 void collect_preasm_stats(Program
*program
);
1362 void collect_postasm_stats(Program
*program
, const std::vector
<uint32_t>& code
);
1364 void aco_print_instr(Instruction
*instr
, FILE *output
);
1365 void aco_print_program(Program
*program
, FILE *output
);
1367 /* utilities for dealing with register demand */
1368 RegisterDemand
get_live_changes(aco_ptr
<Instruction
>& instr
);
1369 RegisterDemand
get_temp_registers(aco_ptr
<Instruction
>& instr
);
1370 RegisterDemand
get_demand_before(RegisterDemand demand
, aco_ptr
<Instruction
>& instr
, aco_ptr
<Instruction
>& instr_before
);
1372 /* number of sgprs that need to be allocated but might notbe addressable as s0-s105 */
1373 uint16_t get_extra_sgprs(Program
*program
);
1375 /* get number of sgprs/vgprs allocated required to address a number of sgprs/vgprs */
1376 uint16_t get_sgpr_alloc(Program
*program
, uint16_t addressable_sgprs
);
1377 uint16_t get_vgpr_alloc(Program
*program
, uint16_t addressable_vgprs
);
1379 /* return number of addressable sgprs/vgprs for max_waves */
1380 uint16_t get_addr_sgpr_from_waves(Program
*program
, uint16_t max_waves
);
1381 uint16_t get_addr_vgpr_from_waves(Program
*program
, uint16_t max_waves
);
1384 const int16_t opcode_gfx7
[static_cast<int>(aco_opcode::num_opcodes
)];
1385 const int16_t opcode_gfx9
[static_cast<int>(aco_opcode::num_opcodes
)];
1386 const int16_t opcode_gfx10
[static_cast<int>(aco_opcode::num_opcodes
)];
1387 const std::bitset
<static_cast<int>(aco_opcode::num_opcodes
)> can_use_input_modifiers
;
1388 const std::bitset
<static_cast<int>(aco_opcode::num_opcodes
)> can_use_output_modifiers
;
1389 const std::bitset
<static_cast<int>(aco_opcode::num_opcodes
)> is_atomic
;
1390 const char *name
[static_cast<int>(aco_opcode::num_opcodes
)];
1391 const aco::Format format
[static_cast<int>(aco_opcode::num_opcodes
)];
1394 extern const Info instr_info
;
1398 #endif /* ACO_IR_H */