1 //===-- R600Instructions.td - R600 Instruction defs -------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // R600 Tablegen instruction definitions
12 //===----------------------------------------------------------------------===//
14 include "R600Intrinsics.td"
16 class InstR600 <bits<11> inst, dag outs, dag ins, string asm, list<dag> pattern,
18 : AMDGPUInst <outs, ins, asm, pattern> {
24 bits<2> FlagOperandIdx = 0;
27 bit HasNativeOperands = 0;
29 bits<11> op_code = inst;
31 let Namespace = "AMDGPU";
32 let OutOperandList = outs;
33 let InOperandList = ins;
35 let Pattern = pattern;
38 let TSFlags{4} = Trig;
41 // Vector instructions are instructions that must fill all slots in an
43 let TSFlags{6} = isVector;
44 let TSFlags{8-7} = FlagOperandIdx;
45 let TSFlags{9} = HasNativeOperands;
46 let TSFlags{10} = Op1;
47 let TSFlags{11} = Op2;
50 class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
51 AMDGPUInst <outs, ins, asm, pattern> {
54 let Namespace = "AMDGPU";
57 def MEMxi : Operand<iPTR> {
58 let MIOperandInfo = (ops R600_TReg32_X:$ptr, i32imm:$index);
59 let PrintMethod = "printMemOperand";
62 def MEMrr : Operand<iPTR> {
63 let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
66 // Operands for non-registers
68 class InstFlag<string PM = "printOperand", int Default = 0>
69 : OperandWithDefaultOps <i32, (ops (i32 Default))> {
73 // src_sel for ALU src operands, see also ALU_CONST, ALU_PARAM registers
74 def SEL : OperandWithDefaultOps <i32, (ops (i32 -1))> {
75 let PrintMethod = "printSel";
78 def LITERAL : InstFlag<"printLiteral">;
80 def WRITE : InstFlag <"printWrite", 1>;
81 def OMOD : InstFlag <"printOMOD">;
82 def REL : InstFlag <"printRel">;
83 def CLAMP : InstFlag <"printClamp">;
84 def NEG : InstFlag <"printNeg">;
85 def ABS : InstFlag <"printAbs">;
86 def UEM : InstFlag <"printUpdateExecMask">;
87 def UP : InstFlag <"printUpdatePred">;
89 // XXX: The r600g finalizer in Mesa expects last to be one in most cases.
90 // Once we start using the packetizer in this backend we should have this
92 def LAST : InstFlag<"printLast", 1>;
94 def FRAMEri : Operand<iPTR> {
95 let MIOperandInfo = (ops R600_Reg32:$ptr, i32imm:$index);
98 def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
99 def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
100 def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
101 def ADDRGA_CONST_OFFSET : ComplexPattern<i32, 1, "SelectGlobalValueConstantOffset", [], []>;
102 def ADDRGA_VAR_OFFSET : ComplexPattern<i32, 2, "SelectGlobalValueVariableOffset", [], []>;
103 def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>;
105 class R600ALU_Word0 {
106 field bits<32> Word0;
114 bits<3> index_mode = 0;
118 bits<9> src0_sel = src0{8-0};
119 bits<2> src0_chan = src0{10-9};
120 bits<9> src1_sel = src1{8-0};
121 bits<2> src1_chan = src1{10-9};
123 let Word0{8-0} = src0_sel;
124 let Word0{9} = src0_rel;
125 let Word0{11-10} = src0_chan;
126 let Word0{12} = src0_neg;
127 let Word0{21-13} = src1_sel;
128 let Word0{22} = src1_rel;
129 let Word0{24-23} = src1_chan;
130 let Word0{25} = src1_neg;
131 let Word0{28-26} = index_mode;
132 let Word0{30-29} = pred_sel;
133 let Word0{31} = last;
136 class R600ALU_Word1 {
137 field bits<32> Word1;
140 bits<3> bank_swizzle = 0;
144 bits<7> dst_sel = dst{6-0};
145 bits<2> dst_chan = dst{10-9};
147 let Word1{20-18} = bank_swizzle;
148 let Word1{27-21} = dst_sel;
149 let Word1{28} = dst_rel;
150 let Word1{30-29} = dst_chan;
151 let Word1{31} = clamp;
154 class R600ALU_Word1_OP2 <bits<11> alu_inst> : R600ALU_Word1{
158 bits<1> update_exec_mask;
163 let Word1{0} = src0_abs;
164 let Word1{1} = src1_abs;
165 let Word1{2} = update_exec_mask;
166 let Word1{3} = update_pred;
167 let Word1{4} = write;
168 let Word1{6-5} = omod;
169 let Word1{17-7} = alu_inst;
172 class R600ALU_Word1_OP3 <bits<5> alu_inst> : R600ALU_Word1{
178 bits<9> src2_sel = src2{8-0};
179 bits<2> src2_chan = src2{10-9};
181 let Word1{8-0} = src2_sel;
182 let Word1{9} = src2_rel;
183 let Word1{11-10} = src2_chan;
184 let Word1{12} = src2_neg;
185 let Word1{17-13} = alu_inst;
189 field bits<32> Word0;
193 bits<1> FETCH_WHOLE_QUAD;
197 bits<6> MEGA_FETCH_COUNT;
199 let Word0{4-0} = VC_INST;
200 let Word0{6-5} = FETCH_TYPE;
201 let Word0{7} = FETCH_WHOLE_QUAD;
202 let Word0{15-8} = BUFFER_ID;
203 let Word0{22-16} = SRC_GPR;
204 let Word0{23} = SRC_REL;
205 let Word0{25-24} = SRC_SEL_X;
206 let Word0{31-26} = MEGA_FETCH_COUNT;
209 class VTX_WORD1_GPR {
210 field bits<32> Word1;
217 bits<1> USE_CONST_FIELDS;
219 bits<2> NUM_FORMAT_ALL;
220 bits<1> FORMAT_COMP_ALL;
221 bits<1> SRF_MODE_ALL;
223 let Word1{6-0} = DST_GPR;
224 let Word1{7} = DST_REL;
225 let Word1{8} = 0; // Reserved
226 let Word1{11-9} = DST_SEL_X;
227 let Word1{14-12} = DST_SEL_Y;
228 let Word1{17-15} = DST_SEL_Z;
229 let Word1{20-18} = DST_SEL_W;
230 let Word1{21} = USE_CONST_FIELDS;
231 let Word1{27-22} = DATA_FORMAT;
232 let Word1{29-28} = NUM_FORMAT_ALL;
233 let Word1{30} = FORMAT_COMP_ALL;
234 let Word1{31} = SRF_MODE_ALL;
238 field bits<32> Word0;
242 bits<1> FETCH_WHOLE_QUAD;
247 bits<2> RESOURCE_INDEX_MODE;
248 bits<2> SAMPLER_INDEX_MODE;
250 let Word0{4-0} = TEX_INST;
251 let Word0{6-5} = INST_MOD;
252 let Word0{7} = FETCH_WHOLE_QUAD;
253 let Word0{15-8} = RESOURCE_ID;
254 let Word0{22-16} = SRC_GPR;
255 let Word0{23} = SRC_REL;
256 let Word0{24} = ALT_CONST;
257 let Word0{26-25} = RESOURCE_INDEX_MODE;
258 let Word0{28-27} = SAMPLER_INDEX_MODE;
262 field bits<32> Word1;
271 bits<1> COORD_TYPE_X;
272 bits<1> COORD_TYPE_Y;
273 bits<1> COORD_TYPE_Z;
274 bits<1> COORD_TYPE_W;
276 let Word1{6-0} = DST_GPR;
277 let Word1{7} = DST_REL;
278 let Word1{11-9} = DST_SEL_X;
279 let Word1{14-12} = DST_SEL_Y;
280 let Word1{17-15} = DST_SEL_Z;
281 let Word1{20-18} = DST_SEL_W;
282 let Word1{27-21} = LOD_BIAS;
283 let Word1{28} = COORD_TYPE_X;
284 let Word1{29} = COORD_TYPE_Y;
285 let Word1{30} = COORD_TYPE_Z;
286 let Word1{31} = COORD_TYPE_W;
290 field bits<32> Word2;
301 let Word2{4-0} = OFFSET_X;
302 let Word2{9-5} = OFFSET_Y;
303 let Word2{14-10} = OFFSET_Z;
304 let Word2{19-15} = SAMPLER_ID;
305 let Word2{22-20} = SRC_SEL_X;
306 let Word2{25-23} = SRC_SEL_Y;
307 let Word2{28-26} = SRC_SEL_Z;
308 let Word2{31-29} = SRC_SEL_W;
312 XXX: R600 subtarget uses a slightly different encoding than the other
313 subtargets. We currently handle this in R600MCCodeEmitter, but we may
314 want to use these instruction classes in the future.
316 class R600ALU_Word1_OP2_r600 : R600ALU_Word1_OP2 {
321 let Inst{37} = fog_merge;
322 let Inst{39-38} = omod;
323 let Inst{49-40} = alu_inst;
326 class R600ALU_Word1_OP2_r700 : R600ALU_Word1_OP2 {
330 let Inst{38-37} = omod;
331 let Inst{49-39} = alu_inst;
335 def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
339 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
341 // Class for instructions with only one source register.
342 // If you add new ins to this instruction, make sure they are listed before
343 // $literal, because the backend currently assumes that the last operand is
344 // a literal. Also be sure to update the enum R600Op1OperandIndex::ROI in
345 // R600Defines.h, R600InstrInfo::buildDefaultInstruction(),
346 // and R600InstrInfo::getOperandIdx().
347 class R600_1OP <bits<11> inst, string opName, list<dag> pattern,
348 InstrItinClass itin = AnyALU> :
350 (outs R600_Reg32:$dst),
351 (ins WRITE:$write, OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
352 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
353 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
354 !strconcat(" ", opName,
355 "$clamp $dst$write$dst_rel$omod, "
356 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
357 "$literal $pred_sel$last"),
361 R600ALU_Word1_OP2 <inst> {
367 let update_exec_mask = 0;
369 let HasNativeOperands = 1;
371 let DisableEncoding = "$literal";
373 let Inst{31-0} = Word0;
374 let Inst{63-32} = Word1;
377 class R600_1OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
378 InstrItinClass itin = AnyALU> :
379 R600_1OP <inst, opName,
380 [(set R600_Reg32:$dst, (node R600_Reg32:$src0))]
383 // If you add our change the operands for R600_2OP instructions, you must
384 // also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
385 // R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
386 class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
387 InstrItinClass itin = AnyALU> :
389 (outs R600_Reg32:$dst),
390 (ins UEM:$update_exec_mask, UP:$update_pred, WRITE:$write,
391 OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
392 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
393 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs, SEL:$src1_sel,
394 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
395 !strconcat(" ", opName,
396 "$clamp $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
397 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
398 "$src1_neg$src1_abs$src1$src1_abs$src1_rel, "
399 "$literal $pred_sel$last"),
403 R600ALU_Word1_OP2 <inst> {
405 let HasNativeOperands = 1;
407 let DisableEncoding = "$literal";
409 let Inst{31-0} = Word0;
410 let Inst{63-32} = Word1;
413 class R600_2OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
414 InstrItinClass itim = AnyALU> :
415 R600_2OP <inst, opName,
416 [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
420 // If you add our change the operands for R600_3OP instructions, you must
421 // also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
422 // R600InstrInfo::buildDefaultInstruction(), and
423 // R600InstrInfo::getOperandIdx().
424 class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
425 InstrItinClass itin = AnyALU> :
427 (outs R600_Reg32:$dst),
428 (ins REL:$dst_rel, CLAMP:$clamp,
429 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, SEL:$src0_sel,
430 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
431 R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
432 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
433 !strconcat(" ", opName, "$clamp $dst$dst_rel, "
434 "$src0_neg$src0$src0_rel, "
435 "$src1_neg$src1$src1_rel, "
436 "$src2_neg$src2$src2_rel, "
437 "$literal $pred_sel$last"),
441 R600ALU_Word1_OP3<inst>{
443 let HasNativeOperands = 1;
444 let DisableEncoding = "$literal";
447 let Inst{31-0} = Word0;
448 let Inst{63-32} = Word1;
451 class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
452 InstrItinClass itin = VecALU> :
454 (outs R600_Reg32:$dst),
460 class R600_TEX <bits<11> inst, string opName, list<dag> pattern,
461 InstrItinClass itin = AnyALU> :
463 (outs R600_Reg128:$DST_GPR),
464 (ins R600_Reg128:$SRC_GPR, i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID, i32imm:$textureTarget),
465 !strconcat(opName, "$DST_GPR, $SRC_GPR, $RESOURCE_ID, $SAMPLER_ID, $textureTarget"),
467 itin>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
468 let Inst{31-0} = Word0;
469 let Inst{63-32} = Word1;
471 let TEX_INST = inst{4-0};
481 let FETCH_WHOLE_QUAD = 0;
483 let SAMPLER_INDEX_MODE = 0;
485 let COORD_TYPE_X = 0;
486 let COORD_TYPE_Y = 0;
487 let COORD_TYPE_Z = 0;
488 let COORD_TYPE_W = 0;
491 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
493 def TEX_SHADOW : PatLeaf<
495 [{uint32_t TType = (uint32_t)N->getZExtValue();
496 return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
500 def TEX_RECT : PatLeaf<
502 [{uint32_t TType = (uint32_t)N->getZExtValue();
507 def TEX_ARRAY : PatLeaf<
509 [{uint32_t TType = (uint32_t)N->getZExtValue();
510 return TType == 9 || TType == 10 || TType == 15 || TType == 16;
514 def TEX_SHADOW_ARRAY : PatLeaf<
516 [{uint32_t TType = (uint32_t)N->getZExtValue();
517 return TType == 11 || TType == 12 || TType == 17;
521 class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, bits<4> rat_id, dag outs,
522 dag ins, string asm, list<dag> pattern> :
523 InstR600ISA <outs, ins, asm, pattern> {
540 // CF_ALLOC_EXPORT_WORD0_RAT
541 let Inst{3-0} = rat_id;
542 let Inst{9-4} = rat_inst;
543 let Inst{10} = 0; // Reserved
544 let Inst{12-11} = RIM;
545 let Inst{14-13} = TYPE;
546 let Inst{21-15} = RW_GPR;
547 let Inst{22} = RW_REL;
548 let Inst{29-23} = INDEX_GPR;
549 let Inst{31-30} = ELEM_SIZE;
551 // CF_ALLOC_EXPORT_WORD1_BUF
552 let Inst{43-32} = ARRAY_SIZE;
553 let Inst{47-44} = COMP_MASK;
554 let Inst{51-48} = BURST_COUNT;
557 let Inst{61-54} = cf_inst;
559 let Inst{63} = BARRIER;
562 class LoadParamFrag <PatFrag load_type> : PatFrag <
563 (ops node:$ptr), (load_type node:$ptr),
564 [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
567 def load_param : LoadParamFrag<load>;
568 def load_param_zexti8 : LoadParamFrag<zextloadi8>;
569 def load_param_zexti16 : LoadParamFrag<zextloadi16>;
571 def isR600 : Predicate<"Subtarget.device()"
572 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX">;
573 def isR700 : Predicate<"Subtarget.device()"
574 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX &&"
575 "Subtarget.device()->getDeviceFlag()"
576 ">= OCL_DEVICE_RV710">;
577 def isEG : Predicate<
578 "Subtarget.device()->getGeneration() >= AMDGPUDeviceInfo::HD5XXX && "
579 "Subtarget.device()->getGeneration() < AMDGPUDeviceInfo::HD7XXX && "
580 "Subtarget.device()->getDeviceFlag() != OCL_DEVICE_CAYMAN">;
582 def isCayman : Predicate<"Subtarget.device()"
583 "->getDeviceFlag() == OCL_DEVICE_CAYMAN">;
584 def isEGorCayman : Predicate<"Subtarget.device()"
585 "->getGeneration() == AMDGPUDeviceInfo::HD5XXX"
586 "|| Subtarget.device()->getGeneration() =="
587 "AMDGPUDeviceInfo::HD6XXX">;
589 def isR600toCayman : Predicate<
590 "Subtarget.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX">;
592 //===----------------------------------------------------------------------===//
594 //===----------------------------------------------------------------------===//
596 def INTERP_PAIR_XY : AMDGPUShaderInst <
597 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
598 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
599 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
602 def INTERP_PAIR_ZW : AMDGPUShaderInst <
603 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
604 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
605 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
608 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
609 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
613 //===----------------------------------------------------------------------===//
614 // Interpolation Instructions
615 //===----------------------------------------------------------------------===//
617 def INTERP_VEC_LOAD : AMDGPUShaderInst <
618 (outs R600_Reg128:$dst),
620 "INTERP_LOAD $src0 : $dst",
623 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
624 let bank_swizzle = 5;
627 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
628 let bank_swizzle = 5;
631 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
633 //===----------------------------------------------------------------------===//
634 // Export Instructions
635 //===----------------------------------------------------------------------===//
637 def ExportType : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
639 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
640 [SDNPHasChain, SDNPSideEffect]>;
643 field bits<32> Word0;
650 let Word0{12-0} = arraybase;
651 let Word0{14-13} = type;
652 let Word0{21-15} = gpr;
653 let Word0{22} = 0; // RW_REL
654 let Word0{29-23} = 0; // INDEX_GPR
655 let Word0{31-30} = elem_size;
658 class ExportSwzWord1 {
659 field bits<32> Word1;
668 let Word1{2-0} = sw_x;
669 let Word1{5-3} = sw_y;
670 let Word1{8-6} = sw_z;
671 let Word1{11-9} = sw_w;
674 class ExportBufWord1 {
675 field bits<32> Word1;
682 let Word1{11-0} = arraySize;
683 let Word1{15-12} = compMask;
686 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
687 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
689 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
690 0, 61, 0, 7, 7, 7, cf_inst, 0)
693 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
695 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
696 0, 61, 7, 0, 7, 7, cf_inst, 0)
699 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
701 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
704 def : Pat<(int_R600_store_dummy 1),
706 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
709 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
710 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
711 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
712 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
717 multiclass SteamOutputExportPattern<Instruction ExportInst,
718 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
720 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
721 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
722 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
723 4095, imm:$mask, buf0inst, 0)>;
725 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
726 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
727 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
728 4095, imm:$mask, buf1inst, 0)>;
730 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
731 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
732 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
733 4095, imm:$mask, buf2inst, 0)>;
735 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
736 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
737 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
738 4095, imm:$mask, buf3inst, 0)>;
741 let usesCustomInserter = 1 in {
743 class ExportSwzInst : InstR600ISA<(
745 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
746 i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
748 !strconcat("EXPORT", " $gpr"),
749 []>, ExportWord0, ExportSwzWord1 {
751 let Inst{31-0} = Word0;
752 let Inst{63-32} = Word1;
755 } // End usesCustomInserter = 1
757 class ExportBufInst : InstR600ISA<(
759 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
760 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
761 !strconcat("EXPORT", " $gpr"),
762 []>, ExportWord0, ExportBufWord1 {
764 let Inst{31-0} = Word0;
765 let Inst{63-32} = Word1;
768 //===----------------------------------------------------------------------===//
769 // Control Flow Instructions
770 //===----------------------------------------------------------------------===//
773 field bits<32> Word0;
776 bits<4> KCACHE_BANK0;
777 bits<4> KCACHE_BANK1;
778 bits<2> KCACHE_MODE0;
780 let Word0{21-0} = ADDR;
781 let Word0{25-22} = KCACHE_BANK0;
782 let Word0{29-26} = KCACHE_BANK1;
783 let Word0{31-30} = KCACHE_MODE0;
787 field bits<32> Word1;
789 bits<2> KCACHE_MODE1;
790 bits<8> KCACHE_ADDR0;
791 bits<8> KCACHE_ADDR1;
795 bits<1> WHOLE_QUAD_MODE;
798 let Word1{1-0} = KCACHE_MODE1;
799 let Word1{9-2} = KCACHE_ADDR0;
800 let Word1{17-10} = KCACHE_ADDR1;
801 let Word1{24-18} = COUNT;
802 let Word1{25} = ALT_CONST;
803 let Word1{29-26} = CF_INST;
804 let Word1{30} = WHOLE_QUAD_MODE;
805 let Word1{31} = BARRIER;
808 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
809 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1, i32imm:$KCACHE_MODE0, i32imm:$KCACHE_MODE1,
810 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1, i32imm:$COUNT),
811 !strconcat(OpName, " $COUNT, @$ADDR, "
812 "KC0[CB$KCACHE_BANK0:$KCACHE_ADDR0-$KCACHE_ADDR0+32]"
813 ", KC1[CB$KCACHE_BANK1:$KCACHE_ADDR1-$KCACHE_ADDR1+32]"),
814 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
819 let WHOLE_QUAD_MODE = 0;
822 let Inst{31-0} = Word0;
823 let Inst{63-32} = Word1;
826 class CF_WORD0_R600 {
827 field bits<32> Word0;
834 class CF_WORD1_R600 {
835 field bits<32> Word1;
843 bits<1> END_OF_PROGRAM;
844 bits<1> VALID_PIXEL_MODE;
846 bits<1> WHOLE_QUAD_MODE;
849 let Word1{2-0} = POP_COUNT;
850 let Word1{7-3} = CF_CONST;
851 let Word1{9-8} = COND;
852 let Word1{12-10} = COUNT;
853 let Word1{18-13} = CALL_COUNT;
854 let Word1{19} = COUNT_3;
855 let Word1{21} = END_OF_PROGRAM;
856 let Word1{22} = VALID_PIXEL_MODE;
857 let Word1{29-23} = CF_INST;
858 let Word1{30} = WHOLE_QUAD_MODE;
859 let Word1{31} = BARRIER;
862 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
863 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
869 let VALID_PIXEL_MODE = 0;
873 let END_OF_PROGRAM = 0;
874 let WHOLE_QUAD_MODE = 0;
876 let Inst{31-0} = Word0;
877 let Inst{63-32} = Word1;
881 field bits<32> Word0;
884 bits<3> JUMPTABLE_SEL;
886 let Word0{23-0} = ADDR;
887 let Word0{26-24} = JUMPTABLE_SEL;
891 field bits<32> Word1;
897 bits<1> VALID_PIXEL_MODE;
901 let Word1{2-0} = POP_COUNT;
902 let Word1{7-3} = CF_CONST;
903 let Word1{9-8} = COND;
904 let Word1{15-10} = COUNT;
905 let Word1{20} = VALID_PIXEL_MODE;
906 let Word1{29-22} = CF_INST;
907 let Word1{31} = BARRIER;
910 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
911 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
916 let JUMPTABLE_SEL = 0;
918 let VALID_PIXEL_MODE = 0;
921 let Inst{31-0} = Word0;
922 let Inst{63-32} = Word1;
925 def CF_ALU : ALU_CLAUSE<8, "ALU">;
926 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
928 def STACK_SIZE : AMDGPUInst <(outs),
929 (ins i32imm:$num), "nstack $num", [] > {
935 let Predicates = [isR600toCayman] in {
937 //===----------------------------------------------------------------------===//
938 // Common Instructions R600, R700, Evergreen, Cayman
939 //===----------------------------------------------------------------------===//
941 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
942 // Non-IEEE MUL: 0 * anything = 0
943 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
944 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
945 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
946 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
948 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
949 // so some of the instruction names don't match the asm string.
950 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
951 def SETE : R600_2OP <
953 [(set R600_Reg32:$dst,
954 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
960 [(set R600_Reg32:$dst,
961 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
967 [(set R600_Reg32:$dst,
968 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
974 [(set R600_Reg32:$dst,
975 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
979 def SETE_DX10 : R600_2OP <
981 [(set R600_Reg32:$dst,
982 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
986 def SETGT_DX10 : R600_2OP <
988 [(set R600_Reg32:$dst,
989 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
993 def SETGE_DX10 : R600_2OP <
995 [(set R600_Reg32:$dst,
996 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
1000 def SETNE_DX10 : R600_2OP <
1002 [(set R600_Reg32:$dst,
1003 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
1007 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
1008 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
1009 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
1010 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
1011 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
1013 def MOV : R600_1OP <0x19, "MOV", []>;
1015 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
1017 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
1018 (outs R600_Reg32:$dst),
1024 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
1026 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
1029 (MOV_IMM_I32 imm:$val)
1032 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
1035 (MOV_IMM_F32 fpimm:$val)
1038 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
1039 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
1040 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
1041 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
1043 let hasSideEffects = 1 in {
1045 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
1047 } // end hasSideEffects
1049 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
1050 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
1051 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
1052 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
1053 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
1054 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
1055 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
1056 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
1057 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
1058 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
1060 def SETE_INT : R600_2OP <
1062 [(set (i32 R600_Reg32:$dst),
1063 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETEQ))]
1066 def SETGT_INT : R600_2OP <
1068 [(set (i32 R600_Reg32:$dst),
1069 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGT))]
1072 def SETGE_INT : R600_2OP <
1074 [(set (i32 R600_Reg32:$dst),
1075 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGE))]
1078 def SETNE_INT : R600_2OP <
1080 [(set (i32 R600_Reg32:$dst),
1081 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETNE))]
1084 def SETGT_UINT : R600_2OP <
1086 [(set (i32 R600_Reg32:$dst),
1087 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGT))]
1090 def SETGE_UINT : R600_2OP <
1092 [(set (i32 R600_Reg32:$dst),
1093 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGE))]
1096 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
1097 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
1098 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
1099 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
1101 def CNDE_INT : R600_3OP <
1103 [(set (i32 R600_Reg32:$dst),
1104 (selectcc (i32 R600_Reg32:$src0), 0,
1105 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
1109 def CNDGE_INT : R600_3OP <
1111 [(set (i32 R600_Reg32:$dst),
1112 (selectcc (i32 R600_Reg32:$src0), 0,
1113 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
1117 def CNDGT_INT : R600_3OP <
1119 [(set (i32 R600_Reg32:$dst),
1120 (selectcc (i32 R600_Reg32:$src0), 0,
1121 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
1125 //===----------------------------------------------------------------------===//
1126 // Texture instructions
1127 //===----------------------------------------------------------------------===//
1129 def TEX_LD : R600_TEX <
1131 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txf R600_Reg128:$SRC_GPR,
1132 imm:$OFFSET_X, imm:$OFFSET_Y, imm:$OFFSET_Z, imm:$RESOURCE_ID,
1133 imm:$SAMPLER_ID, imm:$textureTarget))]
1135 let AsmString = "TEX_LD $DST_GPR, $SRC_GPR, $OFFSET_X, $OFFSET_Y, $OFFSET_Z,"
1136 "$RESOURCE_ID, $SAMPLER_ID, $textureTarget";
1137 let InOperandList = (ins R600_Reg128:$SRC_GPR, i32imm:$OFFSET_X,
1138 i32imm:$OFFSET_Y, i32imm:$OFFSET_Z, i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
1139 i32imm:$textureTarget);
1142 def TEX_GET_TEXTURE_RESINFO : R600_TEX <
1143 0x04, "TEX_GET_TEXTURE_RESINFO",
1144 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txq R600_Reg128:$SRC_GPR,
1145 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1148 def TEX_GET_GRADIENTS_H : R600_TEX <
1149 0x07, "TEX_GET_GRADIENTS_H",
1150 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_ddx R600_Reg128:$SRC_GPR,
1151 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1154 def TEX_GET_GRADIENTS_V : R600_TEX <
1155 0x08, "TEX_GET_GRADIENTS_V",
1156 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_ddy R600_Reg128:$SRC_GPR,
1157 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1160 def TEX_SET_GRADIENTS_H : R600_TEX <
1161 0x0B, "TEX_SET_GRADIENTS_H",
1165 def TEX_SET_GRADIENTS_V : R600_TEX <
1166 0x0C, "TEX_SET_GRADIENTS_V",
1170 def TEX_SAMPLE : R600_TEX <
1172 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_tex R600_Reg128:$SRC_GPR,
1173 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1176 def TEX_SAMPLE_C : R600_TEX <
1177 0x18, "TEX_SAMPLE_C",
1178 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_tex R600_Reg128:$SRC_GPR,
1179 imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
1182 def TEX_SAMPLE_L : R600_TEX <
1183 0x11, "TEX_SAMPLE_L",
1184 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txl R600_Reg128:$SRC_GPR,
1185 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1188 def TEX_SAMPLE_C_L : R600_TEX <
1189 0x19, "TEX_SAMPLE_C_L",
1190 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txl R600_Reg128:$SRC_GPR,
1191 imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
1194 def TEX_SAMPLE_LB : R600_TEX <
1195 0x12, "TEX_SAMPLE_LB",
1196 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txb R600_Reg128:$SRC_GPR,
1197 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1200 def TEX_SAMPLE_C_LB : R600_TEX <
1201 0x1A, "TEX_SAMPLE_C_LB",
1202 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txb R600_Reg128:$SRC_GPR,
1203 imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
1206 def TEX_SAMPLE_G : R600_TEX <
1207 0x14, "TEX_SAMPLE_G",
1211 def TEX_SAMPLE_C_G : R600_TEX <
1212 0x1C, "TEX_SAMPLE_C_G",
1216 //===----------------------------------------------------------------------===//
1217 // Helper classes for common instructions
1218 //===----------------------------------------------------------------------===//
1220 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
1225 class MULADD_Common <bits<5> inst> : R600_3OP <
1230 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
1231 inst, "MULADD_IEEE",
1232 [(set (f32 R600_Reg32:$dst),
1233 (fadd (fmul R600_Reg32:$src0, R600_Reg32:$src1), R600_Reg32:$src2))]
1236 class CNDE_Common <bits<5> inst> : R600_3OP <
1238 [(set R600_Reg32:$dst,
1239 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
1240 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
1244 class CNDGT_Common <bits<5> inst> : R600_3OP <
1246 [(set R600_Reg32:$dst,
1247 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
1248 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
1252 class CNDGE_Common <bits<5> inst> : R600_3OP <
1254 [(set R600_Reg32:$dst,
1255 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
1256 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
1260 multiclass DOT4_Common <bits<11> inst> {
1262 def _pseudo : R600_REDUCTION <inst,
1263 (ins R600_Reg128:$src0, R600_Reg128:$src1),
1264 "DOT4 $dst $src0, $src1",
1265 [(set R600_Reg32:$dst, (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1))]
1268 def _real : R600_2OP <inst, "DOT4", []>;
1271 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1272 multiclass CUBE_Common <bits<11> inst> {
1274 def _pseudo : InstR600 <
1276 (outs R600_Reg128:$dst),
1277 (ins R600_Reg128:$src),
1279 [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
1285 def _real : R600_2OP <inst, "CUBE", []>;
1287 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1289 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1290 inst, "EXP_IEEE", fexp2
1293 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
1294 inst, "FLT_TO_INT", fp_to_sint
1297 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1298 inst, "INT_TO_FLT", sint_to_fp
1301 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1302 inst, "FLT_TO_UINT", fp_to_uint
1305 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1306 inst, "UINT_TO_FLT", uint_to_fp
1309 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1310 inst, "LOG_CLAMPED", []
1313 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1314 inst, "LOG_IEEE", flog2
1317 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1318 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1319 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1320 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1321 inst, "MULHI_INT", mulhs
1323 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1324 inst, "MULHI", mulhu
1326 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1327 inst, "MULLO_INT", mul
1329 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []>;
1331 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1332 inst, "RECIP_CLAMPED", []
1335 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1336 inst, "RECIP_IEEE", [(set R600_Reg32:$dst, (fdiv FP_ONE, R600_Reg32:$src0))]
1339 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1340 inst, "RECIP_UINT", AMDGPUurecip
1343 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1344 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1347 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1348 inst, "RECIPSQRT_IEEE", []
1351 class SIN_Common <bits<11> inst> : R600_1OP <
1356 class COS_Common <bits<11> inst> : R600_1OP <
1361 //===----------------------------------------------------------------------===//
1362 // Helper patterns for complex intrinsics
1363 //===----------------------------------------------------------------------===//
1365 multiclass DIV_Common <InstR600 recip_ieee> {
1367 (int_AMDGPU_div R600_Reg32:$src0, R600_Reg32:$src1),
1368 (MUL_IEEE R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1372 (fdiv R600_Reg32:$src0, R600_Reg32:$src1),
1373 (MUL_IEEE R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1377 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee> : Pat <
1378 (int_TGSI_lit_z R600_Reg32:$src_x, R600_Reg32:$src_y, R600_Reg32:$src_w),
1379 (exp_ieee (mul_lit (log_clamped (MAX R600_Reg32:$src_y, (f32 ZERO))), R600_Reg32:$src_w, R600_Reg32:$src_x))
1382 //===----------------------------------------------------------------------===//
1383 // R600 / R700 Instructions
1384 //===----------------------------------------------------------------------===//
1386 let Predicates = [isR600] in {
1388 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1389 def MULADD_r600 : MULADD_Common<0x10>;
1390 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1391 def CNDE_r600 : CNDE_Common<0x18>;
1392 def CNDGT_r600 : CNDGT_Common<0x19>;
1393 def CNDGE_r600 : CNDGE_Common<0x1A>;
1394 defm DOT4_r600 : DOT4_Common<0x50>;
1395 defm CUBE_r600 : CUBE_Common<0x52>;
1396 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1397 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1398 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1399 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1400 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1401 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1402 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1403 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1404 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1405 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1406 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1407 def SIN_r600 : SIN_Common<0x6E>;
1408 def COS_r600 : COS_Common<0x6F>;
1409 def ASHR_r600 : ASHR_Common<0x70>;
1410 def LSHR_r600 : LSHR_Common<0x71>;
1411 def LSHL_r600 : LSHL_Common<0x72>;
1412 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1413 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1414 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1415 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1416 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1418 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1419 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL, R600_Reg32>;
1420 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1422 def : Pat<(fsqrt R600_Reg32:$src),
1423 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_r600 R600_Reg32:$src))>;
1425 def R600_ExportSwz : ExportSwzInst {
1426 let Word1{20-17} = 1; // BURST_COUNT
1427 let Word1{21} = eop;
1428 let Word1{22} = 1; // VALID_PIXEL_MODE
1429 let Word1{30-23} = inst;
1430 let Word1{31} = 1; // BARRIER
1432 defm : ExportPattern<R600_ExportSwz, 39>;
1434 def R600_ExportBuf : ExportBufInst {
1435 let Word1{20-17} = 1; // BURST_COUNT
1436 let Word1{21} = eop;
1437 let Word1{22} = 1; // VALID_PIXEL_MODE
1438 let Word1{30-23} = inst;
1439 let Word1{31} = 1; // BARRIER
1441 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1443 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1444 "TEX $COUNT @$ADDR"> {
1447 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1448 "VTX $COUNT @$ADDR"> {
1451 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1452 "LOOP_START_DX10 @$ADDR"> {
1456 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1460 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1461 "LOOP_BREAK @$ADDR"> {
1465 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1466 "CONTINUE @$ADDR"> {
1470 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1471 "JUMP @$ADDR POP:$POP_COUNT"> {
1474 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1475 "ELSE @$ADDR POP:$POP_COUNT"> {
1478 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1483 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1484 "POP @$ADDR POP:$POP_COUNT"> {
1490 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1492 class COS_PAT <InstR600 trig> : Pat<
1493 (fcos R600_Reg32:$src),
1494 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1497 class SIN_PAT <InstR600 trig> : Pat<
1498 (fsin R600_Reg32:$src),
1499 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1502 //===----------------------------------------------------------------------===//
1503 // R700 Only instructions
1504 //===----------------------------------------------------------------------===//
1506 let Predicates = [isR700] in {
1507 def SIN_r700 : SIN_Common<0x6E>;
1508 def COS_r700 : COS_Common<0x6F>;
1510 // R700 normalizes inputs to SIN/COS the same as EG
1511 def : SIN_PAT <SIN_r700>;
1512 def : COS_PAT <COS_r700>;
1515 //===----------------------------------------------------------------------===//
1516 // Evergreen Only instructions
1517 //===----------------------------------------------------------------------===//
1519 let Predicates = [isEG] in {
1521 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1522 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1524 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1525 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1526 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1527 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1528 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1529 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1530 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1531 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1532 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1533 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1534 def SIN_eg : SIN_Common<0x8D>;
1535 def COS_eg : COS_Common<0x8E>;
1537 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL, R600_Reg32>;
1538 def : SIN_PAT <SIN_eg>;
1539 def : COS_PAT <COS_eg>;
1540 def : Pat<(fsqrt R600_Reg32:$src),
1541 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_eg R600_Reg32:$src))>;
1542 } // End Predicates = [isEG]
1544 //===----------------------------------------------------------------------===//
1545 // Evergreen / Cayman Instructions
1546 //===----------------------------------------------------------------------===//
1548 let Predicates = [isEGorCayman] in {
1550 // BFE_UINT - bit_extract, an optimization for mask and shift
1555 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1560 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1561 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1562 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1563 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1564 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1565 [(set R600_Reg32:$dst, (int_AMDIL_bit_extract_u32 R600_Reg32:$src0,
1567 R600_Reg32:$src2))],
1571 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
1572 [(set R600_Reg32:$dst, (AMDGPUbitalign R600_Reg32:$src0, R600_Reg32:$src1,
1573 R600_Reg32:$src2))],
1577 def MULADD_eg : MULADD_Common<0x14>;
1578 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
1579 def ASHR_eg : ASHR_Common<0x15>;
1580 def LSHR_eg : LSHR_Common<0x16>;
1581 def LSHL_eg : LSHL_Common<0x17>;
1582 def CNDE_eg : CNDE_Common<0x19>;
1583 def CNDGT_eg : CNDGT_Common<0x1A>;
1584 def CNDGE_eg : CNDGE_Common<0x1B>;
1585 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1586 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1587 defm DOT4_eg : DOT4_Common<0xBE>;
1588 defm CUBE_eg : CUBE_Common<0xC0>;
1590 let hasSideEffects = 1 in {
1591 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1594 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1596 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1600 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1602 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1606 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1608 // TRUNC is used for the FLT_TO_INT instructions to work around a
1609 // perceived problem where the rounding modes are applied differently
1610 // depending on the instruction and the slot they are in.
1612 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1613 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1615 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1616 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1617 // We should look into handling these cases separately.
1618 def : Pat<(fp_to_sint R600_Reg32:$src0),
1619 (FLT_TO_INT_eg (TRUNC R600_Reg32:$src0))>;
1621 def : Pat<(fp_to_uint R600_Reg32:$src0),
1622 (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src0))>;
1624 def EG_ExportSwz : ExportSwzInst {
1625 let Word1{19-16} = 1; // BURST_COUNT
1626 let Word1{20} = 1; // VALID_PIXEL_MODE
1627 let Word1{21} = eop;
1628 let Word1{29-22} = inst;
1629 let Word1{30} = 0; // MARK
1630 let Word1{31} = 1; // BARRIER
1632 defm : ExportPattern<EG_ExportSwz, 83>;
1634 def EG_ExportBuf : ExportBufInst {
1635 let Word1{19-16} = 1; // BURST_COUNT
1636 let Word1{20} = 1; // VALID_PIXEL_MODE
1637 let Word1{21} = eop;
1638 let Word1{29-22} = inst;
1639 let Word1{30} = 0; // MARK
1640 let Word1{31} = 1; // BARRIER
1642 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1644 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1645 "TEX $COUNT @$ADDR"> {
1648 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1649 "VTX $COUNT @$ADDR"> {
1652 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
1653 "LOOP_START_DX10 @$ADDR"> {
1657 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1661 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
1662 "LOOP_BREAK @$ADDR"> {
1666 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
1667 "CONTINUE @$ADDR"> {
1671 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1672 "JUMP @$ADDR POP:$POP_COUNT"> {
1675 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1676 "ELSE @$ADDR POP:$POP_COUNT"> {
1679 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
1684 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1685 "POP @$ADDR POP:$POP_COUNT"> {
1690 //===----------------------------------------------------------------------===//
1691 // Memory read/write instructions
1692 //===----------------------------------------------------------------------===//
1693 let usesCustomInserter = 1 in {
1695 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
1697 : EG_CF_RAT <0x57, 0x2, 0, (outs), ins,
1698 !strconcat(name, " $rw_gpr, $index_gpr, $eop"), pattern> {
1700 // XXX: Have a separate instruction for non-indexed writes.
1706 let COMP_MASK = comp_mask;
1707 let BURST_COUNT = 0;
1713 } // End usesCustomInserter = 1
1716 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1717 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1718 0x1, "RAT_WRITE_CACHELESS_32_eg",
1719 [(global_store (i32 R600_TReg32_X:$rw_gpr), R600_TReg32_X:$index_gpr)]
1723 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1724 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1725 0xf, "RAT_WRITE_CACHELESS_128",
1726 [(global_store (v4i32 R600_Reg128:$rw_gpr), R600_TReg32_X:$index_gpr)]
1729 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1730 : InstR600ISA <outs, (ins MEMxi:$ptr), name#" $dst, $ptr", pattern>,
1731 VTX_WORD1_GPR, VTX_WORD0 {
1736 let FETCH_WHOLE_QUAD = 0;
1737 let BUFFER_ID = buffer_id;
1739 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1740 // to store vertex addresses in any channel, not just X.
1743 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
1744 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
1745 // however, based on my testing if USE_CONST_FIELDS is set, then all
1746 // these fields need to be set to 0.
1747 let USE_CONST_FIELDS = 0;
1748 let NUM_FORMAT_ALL = 1;
1749 let FORMAT_COMP_ALL = 0;
1750 let SRF_MODE_ALL = 0;
1752 let Inst{31-0} = Word0;
1753 let Inst{63-32} = Word1;
1754 // LLVM can only encode 64-bit instructions, so these fields are manually
1755 // encoded in R600CodeEmitter
1758 // bits<2> ENDIAN_SWAP = 0;
1759 // bits<1> CONST_BUF_NO_STRIDE = 0;
1760 // bits<1> MEGA_FETCH = 0;
1761 // bits<1> ALT_CONST = 0;
1762 // bits<2> BUFFER_INDEX_MODE = 0;
1766 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1767 // is done in R600CodeEmitter
1769 // Inst{79-64} = OFFSET;
1770 // Inst{81-80} = ENDIAN_SWAP;
1771 // Inst{82} = CONST_BUF_NO_STRIDE;
1772 // Inst{83} = MEGA_FETCH;
1773 // Inst{84} = ALT_CONST;
1774 // Inst{86-85} = BUFFER_INDEX_MODE;
1775 // Inst{95-86} = 0; Reserved
1777 // VTX_WORD3 (Padding)
1779 // Inst{127-96} = 0;
1782 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1783 : VTX_READ_eg <"VTX_READ_8", buffer_id, (outs R600_TReg32_X:$dst),
1786 let MEGA_FETCH_COUNT = 1;
1788 let DST_SEL_Y = 7; // Masked
1789 let DST_SEL_Z = 7; // Masked
1790 let DST_SEL_W = 7; // Masked
1791 let DATA_FORMAT = 1; // FMT_8
1794 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1795 : VTX_READ_eg <"VTX_READ_16", buffer_id, (outs R600_TReg32_X:$dst),
1797 let MEGA_FETCH_COUNT = 2;
1799 let DST_SEL_Y = 7; // Masked
1800 let DST_SEL_Z = 7; // Masked
1801 let DST_SEL_W = 7; // Masked
1802 let DATA_FORMAT = 5; // FMT_16
1806 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1807 : VTX_READ_eg <"VTX_READ_32", buffer_id, (outs R600_TReg32_X:$dst),
1810 let MEGA_FETCH_COUNT = 4;
1812 let DST_SEL_Y = 7; // Masked
1813 let DST_SEL_Z = 7; // Masked
1814 let DST_SEL_W = 7; // Masked
1815 let DATA_FORMAT = 0xD; // COLOR_32
1817 // This is not really necessary, but there were some GPU hangs that appeared
1818 // to be caused by ALU instructions in the next instruction group that wrote
1819 // to the $ptr registers of the VTX_READ.
1821 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1822 // %T2_X<def> = MOV %ZERO
1823 //Adding this constraint prevents this from happening.
1824 let Constraints = "$ptr.ptr = $dst";
1827 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1828 : VTX_READ_eg <"VTX_READ_128", buffer_id, (outs R600_Reg128:$dst),
1831 let MEGA_FETCH_COUNT = 16;
1836 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1838 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1839 // that holds its buffer address to avoid potential hangs. We can't use
1840 // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
1841 // registers are different sizes.
1844 //===----------------------------------------------------------------------===//
1845 // VTX Read from parameter memory space
1846 //===----------------------------------------------------------------------===//
1848 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1849 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti8 ADDRVTX_READ:$ptr))]
1852 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1853 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti16 ADDRVTX_READ:$ptr))]
1856 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1857 [(set (i32 R600_TReg32_X:$dst), (load_param ADDRVTX_READ:$ptr))]
1860 def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
1861 [(set (v4i32 R600_Reg128:$dst), (load_param ADDRVTX_READ:$ptr))]
1864 //===----------------------------------------------------------------------===//
1865 // VTX Read from global memory space
1866 //===----------------------------------------------------------------------===//
1869 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1870 [(set (i32 R600_TReg32_X:$dst), (zextloadi8_global ADDRVTX_READ:$ptr))]
1874 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1875 [(set (i32 R600_TReg32_X:$dst), (global_load ADDRVTX_READ:$ptr))]
1879 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1880 [(set (v4i32 R600_Reg128:$dst), (global_load ADDRVTX_READ:$ptr))]
1883 //===----------------------------------------------------------------------===//
1885 // XXX: We are currently storing all constants in the global address space.
1886 //===----------------------------------------------------------------------===//
1888 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1889 [(set (i32 R600_TReg32_X:$dst), (constant_load ADDRVTX_READ:$ptr))]
1894 //===----------------------------------------------------------------------===//
1895 // Regist loads and stores - for indirect addressing
1896 //===----------------------------------------------------------------------===//
1898 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1900 let Predicates = [isCayman] in {
1902 let isVector = 1 in {
1904 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1906 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1907 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1908 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1909 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1910 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1911 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1912 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
1913 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1914 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1915 def SIN_cm : SIN_Common<0x8D>;
1916 def COS_cm : COS_Common<0x8E>;
1917 } // End isVector = 1
1919 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL, R600_Reg32>;
1920 def : SIN_PAT <SIN_cm>;
1921 def : COS_PAT <COS_cm>;
1923 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1925 // RECIP_UINT emulation for Cayman
1927 (AMDGPUurecip R600_Reg32:$src0),
1928 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg R600_Reg32:$src0)),
1929 (MOV_IMM_I32 0x4f800000)))
1933 def : Pat<(fsqrt R600_Reg32:$src),
1934 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm R600_Reg32:$src))>;
1938 //===----------------------------------------------------------------------===//
1939 // Branch Instructions
1940 //===----------------------------------------------------------------------===//
1943 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
1944 "IF_PREDICATE_SET $src", []>;
1946 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
1947 "PREDICATED_BREAK $src", []>;
1949 //===----------------------------------------------------------------------===//
1950 // Pseudo instructions
1951 //===----------------------------------------------------------------------===//
1953 let isPseudo = 1 in {
1955 def PRED_X : InstR600 <
1956 0, (outs R600_Predicate_Bit:$dst),
1957 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
1959 let FlagOperandIdx = 3;
1962 let isTerminator = 1, isBranch = 1 in {
1963 def JUMP_COND : InstR600 <0x10,
1965 (ins brtarget:$target, R600_Predicate_Bit:$p),
1966 "JUMP $target ($p)",
1970 def JUMP : InstR600 <0x10,
1972 (ins brtarget:$target),
1977 let isPredicable = 1;
1981 } // End isTerminator = 1, isBranch = 1
1983 let usesCustomInserter = 1 in {
1985 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
1987 def MASK_WRITE : AMDGPUShaderInst <
1989 (ins R600_Reg32:$src),
1994 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
1997 def TXD: AMDGPUShaderInst <
1998 (outs R600_Reg128:$dst),
1999 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2000 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2001 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
2004 def TXD_SHADOW: AMDGPUShaderInst <
2005 (outs R600_Reg128:$dst),
2006 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2007 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2008 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
2011 } // End isPseudo = 1
2012 } // End usesCustomInserter = 1
2014 def CLAMP_R600 : CLAMP <R600_Reg32>;
2015 def FABS_R600 : FABS<R600_Reg32>;
2016 def FNEG_R600 : FNEG<R600_Reg32>;
2018 //===---------------------------------------------------------------------===//
2019 // Return instruction
2020 //===---------------------------------------------------------------------===//
2021 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
2022 usesCustomInserter = 1 in {
2023 def RETURN : ILFormat<(outs), (ins variable_ops),
2024 "RETURN", [(IL_retflag)]>;
2028 //===----------------------------------------------------------------------===//
2029 // Constant Buffer Addressing Support
2030 //===----------------------------------------------------------------------===//
2032 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
2033 def CONST_COPY : Instruction {
2034 let OutOperandList = (outs R600_Reg32:$dst);
2035 let InOperandList = (ins i32imm:$src);
2037 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
2038 let AsmString = "CONST_COPY";
2039 let neverHasSideEffects = 1;
2040 let isAsCheapAsAMove = 1;
2041 let Itinerary = NullALU;
2043 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
2045 def TEX_VTX_CONSTBUF :
2046 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
2047 [(set R600_Reg128:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
2048 VTX_WORD1_GPR, VTX_WORD0 {
2052 let FETCH_WHOLE_QUAD = 0;
2056 let USE_CONST_FIELDS = 0;
2057 let NUM_FORMAT_ALL = 2;
2058 let FORMAT_COMP_ALL = 1;
2059 let SRF_MODE_ALL = 1;
2060 let MEGA_FETCH_COUNT = 16;
2065 let DATA_FORMAT = 35;
2067 let Inst{31-0} = Word0;
2068 let Inst{63-32} = Word1;
2070 // LLVM can only encode 64-bit instructions, so these fields are manually
2071 // encoded in R600CodeEmitter
2074 // bits<2> ENDIAN_SWAP = 0;
2075 // bits<1> CONST_BUF_NO_STRIDE = 0;
2076 // bits<1> MEGA_FETCH = 0;
2077 // bits<1> ALT_CONST = 0;
2078 // bits<2> BUFFER_INDEX_MODE = 0;
2082 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2083 // is done in R600CodeEmitter
2085 // Inst{79-64} = OFFSET;
2086 // Inst{81-80} = ENDIAN_SWAP;
2087 // Inst{82} = CONST_BUF_NO_STRIDE;
2088 // Inst{83} = MEGA_FETCH;
2089 // Inst{84} = ALT_CONST;
2090 // Inst{86-85} = BUFFER_INDEX_MODE;
2091 // Inst{95-86} = 0; Reserved
2093 // VTX_WORD3 (Padding)
2095 // Inst{127-96} = 0;
2099 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
2100 [(set R600_Reg128:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
2101 VTX_WORD1_GPR, VTX_WORD0 {
2105 let FETCH_WHOLE_QUAD = 0;
2109 let USE_CONST_FIELDS = 1;
2110 let NUM_FORMAT_ALL = 0;
2111 let FORMAT_COMP_ALL = 0;
2112 let SRF_MODE_ALL = 1;
2113 let MEGA_FETCH_COUNT = 16;
2118 let DATA_FORMAT = 0;
2120 let Inst{31-0} = Word0;
2121 let Inst{63-32} = Word1;
2123 // LLVM can only encode 64-bit instructions, so these fields are manually
2124 // encoded in R600CodeEmitter
2127 // bits<2> ENDIAN_SWAP = 0;
2128 // bits<1> CONST_BUF_NO_STRIDE = 0;
2129 // bits<1> MEGA_FETCH = 0;
2130 // bits<1> ALT_CONST = 0;
2131 // bits<2> BUFFER_INDEX_MODE = 0;
2135 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2136 // is done in R600CodeEmitter
2138 // Inst{79-64} = OFFSET;
2139 // Inst{81-80} = ENDIAN_SWAP;
2140 // Inst{82} = CONST_BUF_NO_STRIDE;
2141 // Inst{83} = MEGA_FETCH;
2142 // Inst{84} = ALT_CONST;
2143 // Inst{86-85} = BUFFER_INDEX_MODE;
2144 // Inst{95-86} = 0; Reserved
2146 // VTX_WORD3 (Padding)
2148 // Inst{127-96} = 0;
2153 //===--------------------------------------------------------------------===//
2154 // Instructions support
2155 //===--------------------------------------------------------------------===//
2156 //===---------------------------------------------------------------------===//
2157 // Custom Inserter for Branches and returns, this eventually will be a
2159 //===---------------------------------------------------------------------===//
2160 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
2161 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
2162 "; Pseudo unconditional branch instruction",
2164 defm BRANCH_COND : BranchConditional<IL_brcond>;
2167 //===---------------------------------------------------------------------===//
2168 // Flow and Program control Instructions
2169 //===---------------------------------------------------------------------===//
2170 let isTerminator=1 in {
2171 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
2172 !strconcat("SWITCH", " $src"), []>;
2173 def CASE : ILFormat< (outs), (ins GPRI32:$src),
2174 !strconcat("CASE", " $src"), []>;
2175 def BREAK : ILFormat< (outs), (ins),
2177 def CONTINUE : ILFormat< (outs), (ins),
2179 def DEFAULT : ILFormat< (outs), (ins),
2181 def ELSE : ILFormat< (outs), (ins),
2183 def ENDSWITCH : ILFormat< (outs), (ins),
2185 def ENDMAIN : ILFormat< (outs), (ins),
2187 def END : ILFormat< (outs), (ins),
2189 def ENDFUNC : ILFormat< (outs), (ins),
2191 def ENDIF : ILFormat< (outs), (ins),
2193 def WHILELOOP : ILFormat< (outs), (ins),
2195 def ENDLOOP : ILFormat< (outs), (ins),
2197 def FUNC : ILFormat< (outs), (ins),
2199 def RETDYN : ILFormat< (outs), (ins),
2201 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2202 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
2203 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2204 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
2205 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2206 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
2207 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2208 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
2209 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2210 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
2211 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2212 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
2213 defm IFC : BranchInstr2<"IFC">;
2214 defm BREAKC : BranchInstr2<"BREAKC">;
2215 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
2218 //===----------------------------------------------------------------------===//
2220 //===----------------------------------------------------------------------===//
2222 // CND*_INT Pattterns for f32 True / False values
2224 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
2225 (selectcc (i32 R600_Reg32:$src0), 0, (f32 R600_Reg32:$src1),
2226 R600_Reg32:$src2, cc),
2227 (cnd R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2)
2230 def : CND_INT_f32 <CNDE_INT, SETEQ>;
2231 def : CND_INT_f32 <CNDGT_INT, SETGT>;
2232 def : CND_INT_f32 <CNDGE_INT, SETGE>;
2234 //CNDGE_INT extra pattern
2236 (selectcc (i32 R600_Reg32:$src0), -1, (i32 R600_Reg32:$src1),
2237 (i32 R600_Reg32:$src2), COND_GT),
2238 (CNDGE_INT R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2)
2244 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
2248 (int_AMDGPU_kill R600_Reg32:$src0),
2249 (MASK_WRITE (KILLGT (f32 ZERO), (f32 R600_Reg32:$src0)))
2254 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LT),
2255 (SGT R600_Reg32:$src1, R600_Reg32:$src0)
2260 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LE),
2261 (SGE R600_Reg32:$src1, R600_Reg32:$src0)
2264 // SETGT_DX10 reverse args
2266 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, COND_LT),
2267 (SETGT_DX10 R600_Reg32:$src1, R600_Reg32:$src0)
2270 // SETGE_DX10 reverse args
2272 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, COND_LE),
2273 (SETGE_DX10 R600_Reg32:$src1, R600_Reg32:$src0)
2276 // SETGT_INT reverse args
2278 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLT),
2279 (SETGT_INT R600_Reg32:$src1, R600_Reg32:$src0)
2282 // SETGE_INT reverse args
2284 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLE),
2285 (SETGE_INT R600_Reg32:$src1, R600_Reg32:$src0)
2288 // SETGT_UINT reverse args
2290 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULT),
2291 (SETGT_UINT R600_Reg32:$src1, R600_Reg32:$src0)
2294 // SETGE_UINT reverse args
2296 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULE),
2297 (SETGE_UINT R600_Reg32:$src1, R600_Reg32:$src0)
2300 // The next two patterns are special cases for handling 'true if ordered' and
2301 // 'true if unordered' conditionals. The assumption here is that the behavior of
2302 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
2304 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
2305 // We assume that SETE returns false when one of the operands is NAN and
2306 // SNE returns true when on of the operands is NAN
2308 //SETE - 'true if ordered'
2310 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETO),
2311 (SETE R600_Reg32:$src0, R600_Reg32:$src1)
2314 //SETE_DX10 - 'true if ordered'
2316 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETO),
2317 (SETE_DX10 R600_Reg32:$src0, R600_Reg32:$src1)
2320 //SNE - 'true if unordered'
2322 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETUO),
2323 (SNE R600_Reg32:$src0, R600_Reg32:$src1)
2326 //SETNE_DX10 - 'true if ordered'
2328 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUO),
2329 (SETNE_DX10 R600_Reg32:$src0, R600_Reg32:$src1)
2332 def : Extract_Element <f32, v4f32, R600_Reg128, 0, sub0>;
2333 def : Extract_Element <f32, v4f32, R600_Reg128, 1, sub1>;
2334 def : Extract_Element <f32, v4f32, R600_Reg128, 2, sub2>;
2335 def : Extract_Element <f32, v4f32, R600_Reg128, 3, sub3>;
2337 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 0, sub0>;
2338 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 1, sub1>;
2339 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 2, sub2>;
2340 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 3, sub3>;
2342 def : Extract_Element <i32, v4i32, R600_Reg128, 0, sub0>;
2343 def : Extract_Element <i32, v4i32, R600_Reg128, 1, sub1>;
2344 def : Extract_Element <i32, v4i32, R600_Reg128, 2, sub2>;
2345 def : Extract_Element <i32, v4i32, R600_Reg128, 3, sub3>;
2347 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 0, sub0>;
2348 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 1, sub1>;
2349 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 2, sub2>;
2350 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 3, sub3>;
2352 def : Vector4_Build <v4f32, R600_Reg128, f32, R600_Reg32>;
2353 def : Vector4_Build <v4i32, R600_Reg128, i32, R600_Reg32>;
2355 // bitconvert patterns
2357 def : BitConvert <i32, f32, R600_Reg32>;
2358 def : BitConvert <f32, i32, R600_Reg32>;
2359 def : BitConvert <v4f32, v4i32, R600_Reg128>;
2360 def : BitConvert <v4i32, v4f32, R600_Reg128>;
2362 // DWORDADDR pattern
2363 def : DwordAddrPat <i32, R600_Reg32>;
2365 } // End isR600toCayman Predicate