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 <dag outs, dag ins, string asm, list<dag> pattern,
18 : AMDGPUInst <outs, ins, asm, pattern> {
25 bits<2> FlagOperandIdx = 0;
28 bit HasNativeOperands = 0;
32 let Namespace = "AMDGPU";
33 let OutOperandList = outs;
34 let InOperandList = ins;
36 let Pattern = pattern;
39 let TSFlags{0} = TransOnly;
40 let TSFlags{4} = Trig;
43 // Vector instructions are instructions that must fill all slots in an
45 let TSFlags{6} = isVector;
46 let TSFlags{8-7} = FlagOperandIdx;
47 let TSFlags{9} = HasNativeOperands;
48 let TSFlags{10} = Op1;
49 let TSFlags{11} = Op2;
50 let TSFlags{12} = VTXInst;
51 let TSFlags{13} = TEXInst;
54 class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
55 InstR600 <outs, ins, asm, pattern, NullALU> {
57 let Namespace = "AMDGPU";
60 def MEMxi : Operand<iPTR> {
61 let MIOperandInfo = (ops R600_TReg32_X:$ptr, i32imm:$index);
62 let PrintMethod = "printMemOperand";
65 def MEMrr : Operand<iPTR> {
66 let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
69 // Operands for non-registers
71 class InstFlag<string PM = "printOperand", int Default = 0>
72 : OperandWithDefaultOps <i32, (ops (i32 Default))> {
76 // src_sel for ALU src operands, see also ALU_CONST, ALU_PARAM registers
77 def SEL : OperandWithDefaultOps <i32, (ops (i32 -1))> {
78 let PrintMethod = "printSel";
80 def BANK_SWIZZLE : OperandWithDefaultOps <i32, (ops (i32 0))> {
81 let PrintMethod = "printSel";
84 def LITERAL : InstFlag<"printLiteral">;
86 def WRITE : InstFlag <"printWrite", 1>;
87 def OMOD : InstFlag <"printOMOD">;
88 def REL : InstFlag <"printRel">;
89 def CLAMP : InstFlag <"printClamp">;
90 def NEG : InstFlag <"printNeg">;
91 def ABS : InstFlag <"printAbs">;
92 def UEM : InstFlag <"printUpdateExecMask">;
93 def UP : InstFlag <"printUpdatePred">;
95 // XXX: The r600g finalizer in Mesa expects last to be one in most cases.
96 // Once we start using the packetizer in this backend we should have this
98 def LAST : InstFlag<"printLast", 1>;
100 def FRAMEri : Operand<iPTR> {
101 let MIOperandInfo = (ops R600_Reg32:$ptr, i32imm:$index);
104 def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
105 def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
106 def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
107 def ADDRGA_CONST_OFFSET : ComplexPattern<i32, 1, "SelectGlobalValueConstantOffset", [], []>;
108 def ADDRGA_VAR_OFFSET : ComplexPattern<i32, 2, "SelectGlobalValueVariableOffset", [], []>;
109 def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>;
111 class R600ALU_Word0 {
112 field bits<32> Word0;
120 bits<3> index_mode = 0;
124 bits<9> src0_sel = src0{8-0};
125 bits<2> src0_chan = src0{10-9};
126 bits<9> src1_sel = src1{8-0};
127 bits<2> src1_chan = src1{10-9};
129 let Word0{8-0} = src0_sel;
130 let Word0{9} = src0_rel;
131 let Word0{11-10} = src0_chan;
132 let Word0{12} = src0_neg;
133 let Word0{21-13} = src1_sel;
134 let Word0{22} = src1_rel;
135 let Word0{24-23} = src1_chan;
136 let Word0{25} = src1_neg;
137 let Word0{28-26} = index_mode;
138 let Word0{30-29} = pred_sel;
139 let Word0{31} = last;
142 class R600ALU_Word1 {
143 field bits<32> Word1;
146 bits<3> bank_swizzle;
150 bits<7> dst_sel = dst{6-0};
151 bits<2> dst_chan = dst{10-9};
153 let Word1{20-18} = bank_swizzle;
154 let Word1{27-21} = dst_sel;
155 let Word1{28} = dst_rel;
156 let Word1{30-29} = dst_chan;
157 let Word1{31} = clamp;
160 class R600ALU_Word1_OP2 <bits<11> alu_inst> : R600ALU_Word1{
164 bits<1> update_exec_mask;
169 let Word1{0} = src0_abs;
170 let Word1{1} = src1_abs;
171 let Word1{2} = update_exec_mask;
172 let Word1{3} = update_pred;
173 let Word1{4} = write;
174 let Word1{6-5} = omod;
175 let Word1{17-7} = alu_inst;
178 class R600ALU_Word1_OP3 <bits<5> alu_inst> : R600ALU_Word1{
184 bits<9> src2_sel = src2{8-0};
185 bits<2> src2_chan = src2{10-9};
187 let Word1{8-0} = src2_sel;
188 let Word1{9} = src2_rel;
189 let Word1{11-10} = src2_chan;
190 let Word1{12} = src2_neg;
191 let Word1{17-13} = alu_inst;
195 field bits<32> Word0;
199 bits<1> FETCH_WHOLE_QUAD;
203 bits<6> MEGA_FETCH_COUNT;
205 let Word0{4-0} = VC_INST;
206 let Word0{6-5} = FETCH_TYPE;
207 let Word0{7} = FETCH_WHOLE_QUAD;
208 let Word0{15-8} = BUFFER_ID;
209 let Word0{22-16} = SRC_GPR;
210 let Word0{23} = SRC_REL;
211 let Word0{25-24} = SRC_SEL_X;
212 let Word0{31-26} = MEGA_FETCH_COUNT;
215 class VTX_WORD1_GPR {
216 field bits<32> Word1;
223 bits<1> USE_CONST_FIELDS;
225 bits<2> NUM_FORMAT_ALL;
226 bits<1> FORMAT_COMP_ALL;
227 bits<1> SRF_MODE_ALL;
229 let Word1{6-0} = DST_GPR;
230 let Word1{7} = DST_REL;
231 let Word1{8} = 0; // Reserved
232 let Word1{11-9} = DST_SEL_X;
233 let Word1{14-12} = DST_SEL_Y;
234 let Word1{17-15} = DST_SEL_Z;
235 let Word1{20-18} = DST_SEL_W;
236 let Word1{21} = USE_CONST_FIELDS;
237 let Word1{27-22} = DATA_FORMAT;
238 let Word1{29-28} = NUM_FORMAT_ALL;
239 let Word1{30} = FORMAT_COMP_ALL;
240 let Word1{31} = SRF_MODE_ALL;
244 field bits<32> Word0;
248 bits<1> FETCH_WHOLE_QUAD;
253 bits<2> RESOURCE_INDEX_MODE;
254 bits<2> SAMPLER_INDEX_MODE;
256 let Word0{4-0} = TEX_INST;
257 let Word0{6-5} = INST_MOD;
258 let Word0{7} = FETCH_WHOLE_QUAD;
259 let Word0{15-8} = RESOURCE_ID;
260 let Word0{22-16} = SRC_GPR;
261 let Word0{23} = SRC_REL;
262 let Word0{24} = ALT_CONST;
263 let Word0{26-25} = RESOURCE_INDEX_MODE;
264 let Word0{28-27} = SAMPLER_INDEX_MODE;
268 field bits<32> Word1;
277 bits<1> COORD_TYPE_X;
278 bits<1> COORD_TYPE_Y;
279 bits<1> COORD_TYPE_Z;
280 bits<1> COORD_TYPE_W;
282 let Word1{6-0} = DST_GPR;
283 let Word1{7} = DST_REL;
284 let Word1{11-9} = DST_SEL_X;
285 let Word1{14-12} = DST_SEL_Y;
286 let Word1{17-15} = DST_SEL_Z;
287 let Word1{20-18} = DST_SEL_W;
288 let Word1{27-21} = LOD_BIAS;
289 let Word1{28} = COORD_TYPE_X;
290 let Word1{29} = COORD_TYPE_Y;
291 let Word1{30} = COORD_TYPE_Z;
292 let Word1{31} = COORD_TYPE_W;
296 field bits<32> Word2;
307 let Word2{4-0} = OFFSET_X;
308 let Word2{9-5} = OFFSET_Y;
309 let Word2{14-10} = OFFSET_Z;
310 let Word2{19-15} = SAMPLER_ID;
311 let Word2{22-20} = SRC_SEL_X;
312 let Word2{25-23} = SRC_SEL_Y;
313 let Word2{28-26} = SRC_SEL_Z;
314 let Word2{31-29} = SRC_SEL_W;
318 XXX: R600 subtarget uses a slightly different encoding than the other
319 subtargets. We currently handle this in R600MCCodeEmitter, but we may
320 want to use these instruction classes in the future.
322 class R600ALU_Word1_OP2_r600 : R600ALU_Word1_OP2 {
327 let Inst{37} = fog_merge;
328 let Inst{39-38} = omod;
329 let Inst{49-40} = alu_inst;
332 class R600ALU_Word1_OP2_r700 : R600ALU_Word1_OP2 {
336 let Inst{38-37} = omod;
337 let Inst{49-39} = alu_inst;
341 def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
345 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
347 // Class for instructions with only one source register.
348 // If you add new ins to this instruction, make sure they are listed before
349 // $literal, because the backend currently assumes that the last operand is
350 // a literal. Also be sure to update the enum R600Op1OperandIndex::ROI in
351 // R600Defines.h, R600InstrInfo::buildDefaultInstruction(),
352 // and R600InstrInfo::getOperandIdx().
353 class R600_1OP <bits<11> inst, string opName, list<dag> pattern,
354 InstrItinClass itin = AnyALU> :
355 InstR600 <(outs R600_Reg32:$dst),
356 (ins WRITE:$write, OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
357 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
358 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
359 BANK_SWIZZLE:$bank_swizzle),
360 !strconcat(" ", opName,
361 "$clamp $dst$write$dst_rel$omod, "
362 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
363 "$literal $pred_sel$last"),
367 R600ALU_Word1_OP2 <inst> {
373 let update_exec_mask = 0;
375 let HasNativeOperands = 1;
377 let DisableEncoding = "$literal";
379 let Inst{31-0} = Word0;
380 let Inst{63-32} = Word1;
383 class R600_1OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
384 InstrItinClass itin = AnyALU> :
385 R600_1OP <inst, opName,
386 [(set R600_Reg32:$dst, (node R600_Reg32:$src0))]
389 // If you add our change the operands for R600_2OP instructions, you must
390 // also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
391 // R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
392 class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
393 InstrItinClass itin = AnyALU> :
394 InstR600 <(outs R600_Reg32:$dst),
395 (ins UEM:$update_exec_mask, UP:$update_pred, WRITE:$write,
396 OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
397 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
398 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs, SEL:$src1_sel,
399 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
400 BANK_SWIZZLE:$bank_swizzle),
401 !strconcat(" ", opName,
402 "$clamp $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
403 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
404 "$src1_neg$src1_abs$src1$src1_abs$src1_rel, "
405 "$literal $pred_sel$last"),
409 R600ALU_Word1_OP2 <inst> {
411 let HasNativeOperands = 1;
413 let DisableEncoding = "$literal";
415 let Inst{31-0} = Word0;
416 let Inst{63-32} = Word1;
419 class R600_2OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
420 InstrItinClass itim = AnyALU> :
421 R600_2OP <inst, opName,
422 [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
426 // If you add our change the operands for R600_3OP instructions, you must
427 // also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
428 // R600InstrInfo::buildDefaultInstruction(), and
429 // R600InstrInfo::getOperandIdx().
430 class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
431 InstrItinClass itin = AnyALU> :
432 InstR600 <(outs R600_Reg32:$dst),
433 (ins REL:$dst_rel, CLAMP:$clamp,
434 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, SEL:$src0_sel,
435 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
436 R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
437 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
438 BANK_SWIZZLE:$bank_swizzle),
439 !strconcat(" ", opName, "$clamp $dst$dst_rel, "
440 "$src0_neg$src0$src0_rel, "
441 "$src1_neg$src1$src1_rel, "
442 "$src2_neg$src2$src2_rel, "
443 "$literal $pred_sel$last"),
447 R600ALU_Word1_OP3<inst>{
449 let HasNativeOperands = 1;
450 let DisableEncoding = "$literal";
453 let Inst{31-0} = Word0;
454 let Inst{63-32} = Word1;
457 class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
458 InstrItinClass itin = VecALU> :
459 InstR600 <(outs R600_Reg32:$dst),
465 class R600_TEX <bits<11> inst, string opName, list<dag> pattern,
466 InstrItinClass itin = AnyALU> :
467 InstR600 <(outs R600_Reg128:$DST_GPR),
468 (ins R600_Reg128:$SRC_GPR, i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID, i32imm:$textureTarget),
469 !strconcat(opName, "$DST_GPR, $SRC_GPR, $RESOURCE_ID, $SAMPLER_ID, $textureTarget"),
471 itin>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
472 let Inst{31-0} = Word0;
473 let Inst{63-32} = Word1;
475 let TEX_INST = inst{4-0};
485 let FETCH_WHOLE_QUAD = 0;
487 let SAMPLER_INDEX_MODE = 0;
488 let RESOURCE_INDEX_MODE = 0;
490 let COORD_TYPE_X = 0;
491 let COORD_TYPE_Y = 0;
492 let COORD_TYPE_Z = 0;
493 let COORD_TYPE_W = 0;
498 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
500 def TEX_SHADOW : PatLeaf<
502 [{uint32_t TType = (uint32_t)N->getZExtValue();
503 return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
507 def TEX_RECT : PatLeaf<
509 [{uint32_t TType = (uint32_t)N->getZExtValue();
514 def TEX_ARRAY : PatLeaf<
516 [{uint32_t TType = (uint32_t)N->getZExtValue();
517 return TType == 9 || TType == 10 || TType == 15 || TType == 16;
521 def TEX_SHADOW_ARRAY : PatLeaf<
523 [{uint32_t TType = (uint32_t)N->getZExtValue();
524 return TType == 11 || TType == 12 || TType == 17;
528 class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, bits<4> rat_id, dag outs,
529 dag ins, string asm, list<dag> pattern> :
530 InstR600ISA <outs, ins, asm, pattern> {
547 // CF_ALLOC_EXPORT_WORD0_RAT
548 let Inst{3-0} = rat_id;
549 let Inst{9-4} = rat_inst;
550 let Inst{10} = 0; // Reserved
551 let Inst{12-11} = RIM;
552 let Inst{14-13} = TYPE;
553 let Inst{21-15} = RW_GPR;
554 let Inst{22} = RW_REL;
555 let Inst{29-23} = INDEX_GPR;
556 let Inst{31-30} = ELEM_SIZE;
558 // CF_ALLOC_EXPORT_WORD1_BUF
559 let Inst{43-32} = ARRAY_SIZE;
560 let Inst{47-44} = COMP_MASK;
561 let Inst{51-48} = BURST_COUNT;
564 let Inst{61-54} = cf_inst;
566 let Inst{63} = BARRIER;
569 class LoadParamFrag <PatFrag load_type> : PatFrag <
570 (ops node:$ptr), (load_type node:$ptr),
571 [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
574 def load_param : LoadParamFrag<load>;
575 def load_param_zexti8 : LoadParamFrag<zextloadi8>;
576 def load_param_zexti16 : LoadParamFrag<zextloadi16>;
578 def isR600 : Predicate<"Subtarget.device()"
579 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX">;
580 def isR700 : Predicate<"Subtarget.device()"
581 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX &&"
582 "Subtarget.device()->getDeviceFlag()"
583 ">= OCL_DEVICE_RV710">;
584 def isEG : Predicate<
585 "Subtarget.device()->getGeneration() >= AMDGPUDeviceInfo::HD5XXX && "
586 "Subtarget.device()->getGeneration() < AMDGPUDeviceInfo::HD7XXX && "
587 "Subtarget.device()->getDeviceFlag() != OCL_DEVICE_CAYMAN">;
589 def isCayman : Predicate<"Subtarget.device()"
590 "->getDeviceFlag() == OCL_DEVICE_CAYMAN">;
591 def isEGorCayman : Predicate<"Subtarget.device()"
592 "->getGeneration() == AMDGPUDeviceInfo::HD5XXX"
593 "|| Subtarget.device()->getGeneration() =="
594 "AMDGPUDeviceInfo::HD6XXX">;
596 def isR600toCayman : Predicate<
597 "Subtarget.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX">;
599 //===----------------------------------------------------------------------===//
601 //===----------------------------------------------------------------------===//
603 def INTERP_PAIR_XY : AMDGPUShaderInst <
604 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
605 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
606 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
609 def INTERP_PAIR_ZW : AMDGPUShaderInst <
610 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
611 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
612 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
615 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
616 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
620 //===----------------------------------------------------------------------===//
621 // Interpolation Instructions
622 //===----------------------------------------------------------------------===//
624 def INTERP_VEC_LOAD : AMDGPUShaderInst <
625 (outs R600_Reg128:$dst),
627 "INTERP_LOAD $src0 : $dst",
630 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
631 let bank_swizzle = 5;
634 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
635 let bank_swizzle = 5;
638 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
640 //===----------------------------------------------------------------------===//
641 // Export Instructions
642 //===----------------------------------------------------------------------===//
644 def ExportType : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
646 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
647 [SDNPHasChain, SDNPSideEffect]>;
650 field bits<32> Word0;
657 let Word0{12-0} = arraybase;
658 let Word0{14-13} = type;
659 let Word0{21-15} = gpr;
660 let Word0{22} = 0; // RW_REL
661 let Word0{29-23} = 0; // INDEX_GPR
662 let Word0{31-30} = elem_size;
665 class ExportSwzWord1 {
666 field bits<32> Word1;
675 let Word1{2-0} = sw_x;
676 let Word1{5-3} = sw_y;
677 let Word1{8-6} = sw_z;
678 let Word1{11-9} = sw_w;
681 class ExportBufWord1 {
682 field bits<32> Word1;
689 let Word1{11-0} = arraySize;
690 let Word1{15-12} = compMask;
693 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
694 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
696 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
697 0, 61, 0, 7, 7, 7, cf_inst, 0)
700 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
702 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
703 0, 61, 7, 0, 7, 7, cf_inst, 0)
706 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
708 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
711 def : Pat<(int_R600_store_dummy 1),
713 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
716 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
717 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
718 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
719 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
724 multiclass SteamOutputExportPattern<Instruction ExportInst,
725 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
727 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
728 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
729 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
730 4095, imm:$mask, buf0inst, 0)>;
732 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
733 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
734 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
735 4095, imm:$mask, buf1inst, 0)>;
737 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
738 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
739 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
740 4095, imm:$mask, buf2inst, 0)>;
742 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
743 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
744 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
745 4095, imm:$mask, buf3inst, 0)>;
748 // Export Instructions should not be duplicated by TailDuplication pass
749 // (which assumes that duplicable instruction are affected by exec mask)
750 let usesCustomInserter = 1, isNotDuplicable = 1 in {
752 class ExportSwzInst : InstR600ISA<(
754 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
755 i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
757 !strconcat("EXPORT", " $gpr"),
758 []>, ExportWord0, ExportSwzWord1 {
760 let Inst{31-0} = Word0;
761 let Inst{63-32} = Word1;
764 } // End usesCustomInserter = 1
766 class ExportBufInst : InstR600ISA<(
768 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
769 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
770 !strconcat("EXPORT", " $gpr"),
771 []>, ExportWord0, ExportBufWord1 {
773 let Inst{31-0} = Word0;
774 let Inst{63-32} = Word1;
777 //===----------------------------------------------------------------------===//
778 // Control Flow Instructions
779 //===----------------------------------------------------------------------===//
782 field bits<32> Word0;
785 bits<4> KCACHE_BANK0;
786 bits<4> KCACHE_BANK1;
787 bits<2> KCACHE_MODE0;
789 let Word0{21-0} = ADDR;
790 let Word0{25-22} = KCACHE_BANK0;
791 let Word0{29-26} = KCACHE_BANK1;
792 let Word0{31-30} = KCACHE_MODE0;
796 field bits<32> Word1;
798 bits<2> KCACHE_MODE1;
799 bits<8> KCACHE_ADDR0;
800 bits<8> KCACHE_ADDR1;
804 bits<1> WHOLE_QUAD_MODE;
807 let Word1{1-0} = KCACHE_MODE1;
808 let Word1{9-2} = KCACHE_ADDR0;
809 let Word1{17-10} = KCACHE_ADDR1;
810 let Word1{24-18} = COUNT;
811 let Word1{25} = ALT_CONST;
812 let Word1{29-26} = CF_INST;
813 let Word1{30} = WHOLE_QUAD_MODE;
814 let Word1{31} = BARRIER;
817 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
818 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1, i32imm:$KCACHE_MODE0, i32imm:$KCACHE_MODE1,
819 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1, i32imm:$COUNT),
820 !strconcat(OpName, " $COUNT, @$ADDR, "
821 "KC0[CB$KCACHE_BANK0:$KCACHE_ADDR0-$KCACHE_ADDR0+32]"
822 ", KC1[CB$KCACHE_BANK1:$KCACHE_ADDR1-$KCACHE_ADDR1+32]"),
823 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
828 let WHOLE_QUAD_MODE = 0;
831 let Inst{31-0} = Word0;
832 let Inst{63-32} = Word1;
835 class CF_WORD0_R600 {
836 field bits<32> Word0;
843 class CF_WORD1_R600 {
844 field bits<32> Word1;
852 bits<1> END_OF_PROGRAM;
853 bits<1> VALID_PIXEL_MODE;
855 bits<1> WHOLE_QUAD_MODE;
858 let Word1{2-0} = POP_COUNT;
859 let Word1{7-3} = CF_CONST;
860 let Word1{9-8} = COND;
861 let Word1{12-10} = COUNT;
862 let Word1{18-13} = CALL_COUNT;
863 let Word1{19} = COUNT_3;
864 let Word1{21} = END_OF_PROGRAM;
865 let Word1{22} = VALID_PIXEL_MODE;
866 let Word1{29-23} = CF_INST;
867 let Word1{30} = WHOLE_QUAD_MODE;
868 let Word1{31} = BARRIER;
871 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
872 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
878 let VALID_PIXEL_MODE = 0;
882 let END_OF_PROGRAM = 0;
883 let WHOLE_QUAD_MODE = 0;
885 let Inst{31-0} = Word0;
886 let Inst{63-32} = Word1;
890 field bits<32> Word0;
893 bits<3> JUMPTABLE_SEL;
895 let Word0{23-0} = ADDR;
896 let Word0{26-24} = JUMPTABLE_SEL;
900 field bits<32> Word1;
906 bits<1> VALID_PIXEL_MODE;
907 bits<1> END_OF_PROGRAM;
911 let Word1{2-0} = POP_COUNT;
912 let Word1{7-3} = CF_CONST;
913 let Word1{9-8} = COND;
914 let Word1{15-10} = COUNT;
915 let Word1{20} = VALID_PIXEL_MODE;
916 let Word1{21} = END_OF_PROGRAM;
917 let Word1{29-22} = CF_INST;
918 let Word1{31} = BARRIER;
921 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
922 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
927 let JUMPTABLE_SEL = 0;
929 let VALID_PIXEL_MODE = 0;
931 let END_OF_PROGRAM = 0;
933 let Inst{31-0} = Word0;
934 let Inst{63-32} = Word1;
937 def CF_ALU : ALU_CLAUSE<8, "ALU">;
938 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
940 def FETCH_CLAUSE : AMDGPUInst <(outs),
941 (ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
947 def ALU_CLAUSE : AMDGPUInst <(outs),
948 (ins i32imm:$addr), "ALU clause starting at $addr:", [] > {
954 def LITERALS : AMDGPUInst <(outs),
955 (ins LITERAL:$literal1, LITERAL:$literal2), "$literal1, $literal2", [] > {
960 let Inst{31-0} = literal1;
961 let Inst{63-32} = literal2;
964 def PAD : AMDGPUInst <(outs), (ins), "PAD", [] > {
968 let Predicates = [isR600toCayman] in {
970 //===----------------------------------------------------------------------===//
971 // Common Instructions R600, R700, Evergreen, Cayman
972 //===----------------------------------------------------------------------===//
974 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
975 // Non-IEEE MUL: 0 * anything = 0
976 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
977 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
978 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
979 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
981 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
982 // so some of the instruction names don't match the asm string.
983 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
984 def SETE : R600_2OP <
986 [(set R600_Reg32:$dst,
987 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
993 [(set R600_Reg32:$dst,
994 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
1000 [(set R600_Reg32:$dst,
1001 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
1005 def SNE : R600_2OP <
1007 [(set R600_Reg32:$dst,
1008 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
1012 def SETE_DX10 : R600_2OP <
1014 [(set R600_Reg32:$dst,
1015 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
1019 def SETGT_DX10 : R600_2OP <
1021 [(set R600_Reg32:$dst,
1022 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
1026 def SETGE_DX10 : R600_2OP <
1028 [(set R600_Reg32:$dst,
1029 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
1033 def SETNE_DX10 : R600_2OP <
1035 [(set R600_Reg32:$dst,
1036 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
1040 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
1041 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
1042 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
1043 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
1044 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
1046 def MOV : R600_1OP <0x19, "MOV", []>;
1048 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
1050 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
1051 (outs R600_Reg32:$dst),
1057 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
1059 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
1062 (MOV_IMM_I32 imm:$val)
1065 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
1068 (MOV_IMM_F32 fpimm:$val)
1071 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
1072 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
1073 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
1074 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
1076 let hasSideEffects = 1 in {
1078 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
1080 } // end hasSideEffects
1082 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
1083 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
1084 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
1085 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
1086 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
1087 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
1088 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
1089 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
1090 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
1091 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
1093 def SETE_INT : R600_2OP <
1095 [(set (i32 R600_Reg32:$dst),
1096 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETEQ))]
1099 def SETGT_INT : R600_2OP <
1101 [(set (i32 R600_Reg32:$dst),
1102 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGT))]
1105 def SETGE_INT : R600_2OP <
1107 [(set (i32 R600_Reg32:$dst),
1108 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGE))]
1111 def SETNE_INT : R600_2OP <
1113 [(set (i32 R600_Reg32:$dst),
1114 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETNE))]
1117 def SETGT_UINT : R600_2OP <
1119 [(set (i32 R600_Reg32:$dst),
1120 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGT))]
1123 def SETGE_UINT : R600_2OP <
1125 [(set (i32 R600_Reg32:$dst),
1126 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGE))]
1129 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
1130 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
1131 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
1132 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
1134 def CNDE_INT : R600_3OP <
1136 [(set (i32 R600_Reg32:$dst),
1137 (selectcc (i32 R600_Reg32:$src0), 0,
1138 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
1142 def CNDGE_INT : R600_3OP <
1144 [(set (i32 R600_Reg32:$dst),
1145 (selectcc (i32 R600_Reg32:$src0), 0,
1146 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
1150 def CNDGT_INT : R600_3OP <
1152 [(set (i32 R600_Reg32:$dst),
1153 (selectcc (i32 R600_Reg32:$src0), 0,
1154 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
1158 //===----------------------------------------------------------------------===//
1159 // Texture instructions
1160 //===----------------------------------------------------------------------===//
1162 def TEX_LD : R600_TEX <
1164 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txf R600_Reg128:$SRC_GPR,
1165 imm:$OFFSET_X, imm:$OFFSET_Y, imm:$OFFSET_Z, imm:$RESOURCE_ID,
1166 imm:$SAMPLER_ID, imm:$textureTarget))]
1168 let AsmString = "TEX_LD $DST_GPR, $SRC_GPR, $OFFSET_X, $OFFSET_Y, $OFFSET_Z,"
1169 "$RESOURCE_ID, $SAMPLER_ID, $textureTarget";
1170 let InOperandList = (ins R600_Reg128:$SRC_GPR, i32imm:$OFFSET_X,
1171 i32imm:$OFFSET_Y, i32imm:$OFFSET_Z, i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
1172 i32imm:$textureTarget);
1175 def TEX_GET_TEXTURE_RESINFO : R600_TEX <
1176 0x04, "TEX_GET_TEXTURE_RESINFO",
1177 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txq R600_Reg128:$SRC_GPR,
1178 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1181 def TEX_GET_GRADIENTS_H : R600_TEX <
1182 0x07, "TEX_GET_GRADIENTS_H",
1183 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_ddx R600_Reg128:$SRC_GPR,
1184 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1187 def TEX_GET_GRADIENTS_V : R600_TEX <
1188 0x08, "TEX_GET_GRADIENTS_V",
1189 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_ddy R600_Reg128:$SRC_GPR,
1190 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1193 def TEX_SET_GRADIENTS_H : R600_TEX <
1194 0x0B, "TEX_SET_GRADIENTS_H",
1198 def TEX_SET_GRADIENTS_V : R600_TEX <
1199 0x0C, "TEX_SET_GRADIENTS_V",
1203 def TEX_SAMPLE : R600_TEX <
1205 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_tex R600_Reg128:$SRC_GPR,
1206 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1209 def TEX_SAMPLE_C : R600_TEX <
1210 0x18, "TEX_SAMPLE_C",
1211 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_tex R600_Reg128:$SRC_GPR,
1212 imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
1215 def TEX_SAMPLE_L : R600_TEX <
1216 0x11, "TEX_SAMPLE_L",
1217 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txl R600_Reg128:$SRC_GPR,
1218 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1221 def TEX_SAMPLE_C_L : R600_TEX <
1222 0x19, "TEX_SAMPLE_C_L",
1223 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txl R600_Reg128:$SRC_GPR,
1224 imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
1227 def TEX_SAMPLE_LB : R600_TEX <
1228 0x12, "TEX_SAMPLE_LB",
1229 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txb R600_Reg128:$SRC_GPR,
1230 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1233 def TEX_SAMPLE_C_LB : R600_TEX <
1234 0x1A, "TEX_SAMPLE_C_LB",
1235 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txb R600_Reg128:$SRC_GPR,
1236 imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
1239 def TEX_SAMPLE_G : R600_TEX <
1240 0x14, "TEX_SAMPLE_G",
1244 def TEX_SAMPLE_C_G : R600_TEX <
1245 0x1C, "TEX_SAMPLE_C_G",
1249 //===----------------------------------------------------------------------===//
1250 // Helper classes for common instructions
1251 //===----------------------------------------------------------------------===//
1253 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
1258 class MULADD_Common <bits<5> inst> : R600_3OP <
1263 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
1264 inst, "MULADD_IEEE",
1265 [(set (f32 R600_Reg32:$dst),
1266 (fadd (fmul R600_Reg32:$src0, R600_Reg32:$src1), R600_Reg32:$src2))]
1269 class CNDE_Common <bits<5> inst> : R600_3OP <
1271 [(set R600_Reg32:$dst,
1272 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
1273 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
1277 class CNDGT_Common <bits<5> inst> : R600_3OP <
1279 [(set R600_Reg32:$dst,
1280 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
1281 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
1285 class CNDGE_Common <bits<5> inst> : R600_3OP <
1287 [(set R600_Reg32:$dst,
1288 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
1289 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
1293 multiclass DOT4_Common <bits<11> inst> {
1295 def _pseudo : R600_REDUCTION <inst,
1296 (ins R600_Reg128:$src0, R600_Reg128:$src1),
1297 "DOT4 $dst $src0, $src1",
1298 [(set R600_Reg32:$dst, (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1))]
1301 def _real : R600_2OP <inst, "DOT4", []>;
1304 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1305 multiclass CUBE_Common <bits<11> inst> {
1307 def _pseudo : InstR600 <
1308 (outs R600_Reg128:$dst),
1309 (ins R600_Reg128:$src),
1311 [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
1317 def _real : R600_2OP <inst, "CUBE", []>;
1319 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1321 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1322 inst, "EXP_IEEE", fexp2
1325 let Itinerary = TransALU;
1328 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
1329 inst, "FLT_TO_INT", fp_to_sint
1332 let Itinerary = TransALU;
1335 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1336 inst, "INT_TO_FLT", sint_to_fp
1339 let Itinerary = TransALU;
1342 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1343 inst, "FLT_TO_UINT", fp_to_uint
1346 let Itinerary = TransALU;
1349 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1350 inst, "UINT_TO_FLT", uint_to_fp
1353 let Itinerary = TransALU;
1356 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1357 inst, "LOG_CLAMPED", []
1360 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1361 inst, "LOG_IEEE", flog2
1364 let Itinerary = TransALU;
1367 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1368 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1369 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1370 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1371 inst, "MULHI_INT", mulhs
1374 let Itinerary = TransALU;
1376 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1377 inst, "MULHI", mulhu
1380 let Itinerary = TransALU;
1382 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1383 inst, "MULLO_INT", mul
1386 let Itinerary = TransALU;
1388 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []> {
1390 let Itinerary = TransALU;
1393 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1394 inst, "RECIP_CLAMPED", []
1397 let Itinerary = TransALU;
1400 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1401 inst, "RECIP_IEEE", [(set R600_Reg32:$dst, (fdiv FP_ONE, R600_Reg32:$src0))]
1404 let Itinerary = TransALU;
1407 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1408 inst, "RECIP_UINT", AMDGPUurecip
1411 let Itinerary = TransALU;
1414 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1415 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1418 let Itinerary = TransALU;
1421 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1422 inst, "RECIPSQRT_IEEE", []
1425 let Itinerary = TransALU;
1428 class SIN_Common <bits<11> inst> : R600_1OP <
1432 let Itinerary = TransALU;
1435 class COS_Common <bits<11> inst> : R600_1OP <
1439 let Itinerary = TransALU;
1442 //===----------------------------------------------------------------------===//
1443 // Helper patterns for complex intrinsics
1444 //===----------------------------------------------------------------------===//
1446 multiclass DIV_Common <InstR600 recip_ieee> {
1448 (int_AMDGPU_div R600_Reg32:$src0, R600_Reg32:$src1),
1449 (MUL_IEEE R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1453 (fdiv R600_Reg32:$src0, R600_Reg32:$src1),
1454 (MUL_IEEE R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1458 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee> : Pat <
1459 (int_TGSI_lit_z R600_Reg32:$src_x, R600_Reg32:$src_y, R600_Reg32:$src_w),
1460 (exp_ieee (mul_lit (log_clamped (MAX R600_Reg32:$src_y, (f32 ZERO))), R600_Reg32:$src_w, R600_Reg32:$src_x))
1463 //===----------------------------------------------------------------------===//
1464 // R600 / R700 Instructions
1465 //===----------------------------------------------------------------------===//
1467 let Predicates = [isR600] in {
1469 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1470 def MULADD_r600 : MULADD_Common<0x10>;
1471 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1472 def CNDE_r600 : CNDE_Common<0x18>;
1473 def CNDGT_r600 : CNDGT_Common<0x19>;
1474 def CNDGE_r600 : CNDGE_Common<0x1A>;
1475 defm DOT4_r600 : DOT4_Common<0x50>;
1476 defm CUBE_r600 : CUBE_Common<0x52>;
1477 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1478 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1479 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1480 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1481 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1482 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1483 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1484 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1485 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1486 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1487 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1488 def SIN_r600 : SIN_Common<0x6E>;
1489 def COS_r600 : COS_Common<0x6F>;
1490 def ASHR_r600 : ASHR_Common<0x70>;
1491 def LSHR_r600 : LSHR_Common<0x71>;
1492 def LSHL_r600 : LSHL_Common<0x72>;
1493 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1494 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1495 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1496 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1497 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1499 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1500 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL, R600_Reg32>;
1501 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1503 def : Pat<(fsqrt R600_Reg32:$src),
1504 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_r600 R600_Reg32:$src))>;
1506 def R600_ExportSwz : ExportSwzInst {
1507 let Word1{20-17} = 0; // BURST_COUNT
1508 let Word1{21} = eop;
1509 let Word1{22} = 1; // VALID_PIXEL_MODE
1510 let Word1{30-23} = inst;
1511 let Word1{31} = 1; // BARRIER
1513 defm : ExportPattern<R600_ExportSwz, 39>;
1515 def R600_ExportBuf : ExportBufInst {
1516 let Word1{20-17} = 0; // BURST_COUNT
1517 let Word1{21} = eop;
1518 let Word1{22} = 1; // VALID_PIXEL_MODE
1519 let Word1{30-23} = inst;
1520 let Word1{31} = 1; // BARRIER
1522 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1524 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1525 "TEX $COUNT @$ADDR"> {
1528 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1529 "VTX $COUNT @$ADDR"> {
1532 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1533 "LOOP_START_DX10 @$ADDR"> {
1537 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1541 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1542 "LOOP_BREAK @$ADDR"> {
1546 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1547 "CONTINUE @$ADDR"> {
1551 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1552 "JUMP @$ADDR POP:$POP_COUNT"> {
1555 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1556 "ELSE @$ADDR POP:$POP_COUNT"> {
1559 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1564 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1565 "POP @$ADDR POP:$POP_COUNT"> {
1568 def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
1572 let END_OF_PROGRAM = 1;
1577 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1579 class COS_PAT <InstR600 trig> : Pat<
1580 (fcos R600_Reg32:$src),
1581 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1584 class SIN_PAT <InstR600 trig> : Pat<
1585 (fsin R600_Reg32:$src),
1586 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1589 //===----------------------------------------------------------------------===//
1590 // R700 Only instructions
1591 //===----------------------------------------------------------------------===//
1593 let Predicates = [isR700] in {
1594 def SIN_r700 : SIN_Common<0x6E>;
1595 def COS_r700 : COS_Common<0x6F>;
1597 // R700 normalizes inputs to SIN/COS the same as EG
1598 def : SIN_PAT <SIN_r700>;
1599 def : COS_PAT <COS_r700>;
1602 //===----------------------------------------------------------------------===//
1603 // Evergreen Only instructions
1604 //===----------------------------------------------------------------------===//
1606 let Predicates = [isEG] in {
1608 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1609 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1611 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1612 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1613 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1614 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1615 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1616 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1617 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1618 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1619 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1620 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1621 def SIN_eg : SIN_Common<0x8D>;
1622 def COS_eg : COS_Common<0x8E>;
1624 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL, R600_Reg32>;
1625 def : SIN_PAT <SIN_eg>;
1626 def : COS_PAT <COS_eg>;
1627 def : Pat<(fsqrt R600_Reg32:$src),
1628 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_eg R600_Reg32:$src))>;
1629 } // End Predicates = [isEG]
1631 //===----------------------------------------------------------------------===//
1632 // Evergreen / Cayman Instructions
1633 //===----------------------------------------------------------------------===//
1635 let Predicates = [isEGorCayman] in {
1637 // BFE_UINT - bit_extract, an optimization for mask and shift
1642 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1647 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1648 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1649 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1650 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1651 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1652 [(set R600_Reg32:$dst, (int_AMDIL_bit_extract_u32 R600_Reg32:$src0,
1654 R600_Reg32:$src2))],
1658 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", []>;
1659 defm : BFIPatterns <BFI_INT_eg>;
1661 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
1662 [(set R600_Reg32:$dst, (AMDGPUbitalign R600_Reg32:$src0, R600_Reg32:$src1,
1663 R600_Reg32:$src2))],
1667 def MULADD_eg : MULADD_Common<0x14>;
1668 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
1669 def ASHR_eg : ASHR_Common<0x15>;
1670 def LSHR_eg : LSHR_Common<0x16>;
1671 def LSHL_eg : LSHL_Common<0x17>;
1672 def CNDE_eg : CNDE_Common<0x19>;
1673 def CNDGT_eg : CNDGT_Common<0x1A>;
1674 def CNDGE_eg : CNDGE_Common<0x1B>;
1675 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1676 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1677 defm DOT4_eg : DOT4_Common<0xBE>;
1678 defm CUBE_eg : CUBE_Common<0xC0>;
1680 let hasSideEffects = 1 in {
1681 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1684 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1686 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1690 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1692 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1696 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1698 // TRUNC is used for the FLT_TO_INT instructions to work around a
1699 // perceived problem where the rounding modes are applied differently
1700 // depending on the instruction and the slot they are in.
1702 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1703 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1705 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1706 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1707 // We should look into handling these cases separately.
1708 def : Pat<(fp_to_sint R600_Reg32:$src0),
1709 (FLT_TO_INT_eg (TRUNC R600_Reg32:$src0))>;
1711 def : Pat<(fp_to_uint R600_Reg32:$src0),
1712 (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src0))>;
1714 def EG_ExportSwz : ExportSwzInst {
1715 let Word1{19-16} = 0; // BURST_COUNT
1716 let Word1{20} = 1; // VALID_PIXEL_MODE
1717 let Word1{21} = eop;
1718 let Word1{29-22} = inst;
1719 let Word1{30} = 0; // MARK
1720 let Word1{31} = 1; // BARRIER
1722 defm : ExportPattern<EG_ExportSwz, 83>;
1724 def EG_ExportBuf : ExportBufInst {
1725 let Word1{19-16} = 0; // BURST_COUNT
1726 let Word1{20} = 1; // VALID_PIXEL_MODE
1727 let Word1{21} = eop;
1728 let Word1{29-22} = inst;
1729 let Word1{30} = 0; // MARK
1730 let Word1{31} = 1; // BARRIER
1732 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1734 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1735 "TEX $COUNT @$ADDR"> {
1738 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1739 "VTX $COUNT @$ADDR"> {
1742 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
1743 "LOOP_START_DX10 @$ADDR"> {
1747 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1751 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
1752 "LOOP_BREAK @$ADDR"> {
1756 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
1757 "CONTINUE @$ADDR"> {
1761 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1762 "JUMP @$ADDR POP:$POP_COUNT"> {
1765 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1766 "ELSE @$ADDR POP:$POP_COUNT"> {
1769 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
1774 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1775 "POP @$ADDR POP:$POP_COUNT"> {
1778 def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
1782 let END_OF_PROGRAM = 1;
1785 //===----------------------------------------------------------------------===//
1786 // Memory read/write instructions
1787 //===----------------------------------------------------------------------===//
1788 let usesCustomInserter = 1 in {
1790 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
1792 : EG_CF_RAT <0x57, 0x2, 0, (outs), ins,
1793 !strconcat(name, " $rw_gpr, $index_gpr, $eop"), pattern> {
1795 // XXX: Have a separate instruction for non-indexed writes.
1801 let COMP_MASK = comp_mask;
1802 let BURST_COUNT = 0;
1808 } // End usesCustomInserter = 1
1811 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1812 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1813 0x1, "RAT_WRITE_CACHELESS_32_eg",
1814 [(global_store (i32 R600_TReg32_X:$rw_gpr), R600_TReg32_X:$index_gpr)]
1818 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1819 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1820 0xf, "RAT_WRITE_CACHELESS_128",
1821 [(global_store (v4i32 R600_Reg128:$rw_gpr), R600_TReg32_X:$index_gpr)]
1824 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1825 : InstR600ISA <outs, (ins MEMxi:$ptr), name#" $dst, $ptr", pattern>,
1826 VTX_WORD1_GPR, VTX_WORD0 {
1831 let FETCH_WHOLE_QUAD = 0;
1832 let BUFFER_ID = buffer_id;
1834 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1835 // to store vertex addresses in any channel, not just X.
1838 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
1839 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
1840 // however, based on my testing if USE_CONST_FIELDS is set, then all
1841 // these fields need to be set to 0.
1842 let USE_CONST_FIELDS = 0;
1843 let NUM_FORMAT_ALL = 1;
1844 let FORMAT_COMP_ALL = 0;
1845 let SRF_MODE_ALL = 0;
1847 let Inst{31-0} = Word0;
1848 let Inst{63-32} = Word1;
1849 // LLVM can only encode 64-bit instructions, so these fields are manually
1850 // encoded in R600CodeEmitter
1853 // bits<2> ENDIAN_SWAP = 0;
1854 // bits<1> CONST_BUF_NO_STRIDE = 0;
1855 // bits<1> MEGA_FETCH = 0;
1856 // bits<1> ALT_CONST = 0;
1857 // bits<2> BUFFER_INDEX_MODE = 0;
1861 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1862 // is done in R600CodeEmitter
1864 // Inst{79-64} = OFFSET;
1865 // Inst{81-80} = ENDIAN_SWAP;
1866 // Inst{82} = CONST_BUF_NO_STRIDE;
1867 // Inst{83} = MEGA_FETCH;
1868 // Inst{84} = ALT_CONST;
1869 // Inst{86-85} = BUFFER_INDEX_MODE;
1870 // Inst{95-86} = 0; Reserved
1872 // VTX_WORD3 (Padding)
1874 // Inst{127-96} = 0;
1879 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1880 : VTX_READ_eg <"VTX_READ_8", buffer_id, (outs R600_TReg32_X:$dst),
1883 let MEGA_FETCH_COUNT = 1;
1885 let DST_SEL_Y = 7; // Masked
1886 let DST_SEL_Z = 7; // Masked
1887 let DST_SEL_W = 7; // Masked
1888 let DATA_FORMAT = 1; // FMT_8
1891 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1892 : VTX_READ_eg <"VTX_READ_16", buffer_id, (outs R600_TReg32_X:$dst),
1894 let MEGA_FETCH_COUNT = 2;
1896 let DST_SEL_Y = 7; // Masked
1897 let DST_SEL_Z = 7; // Masked
1898 let DST_SEL_W = 7; // Masked
1899 let DATA_FORMAT = 5; // FMT_16
1903 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1904 : VTX_READ_eg <"VTX_READ_32", buffer_id, (outs R600_TReg32_X:$dst),
1907 let MEGA_FETCH_COUNT = 4;
1909 let DST_SEL_Y = 7; // Masked
1910 let DST_SEL_Z = 7; // Masked
1911 let DST_SEL_W = 7; // Masked
1912 let DATA_FORMAT = 0xD; // COLOR_32
1914 // This is not really necessary, but there were some GPU hangs that appeared
1915 // to be caused by ALU instructions in the next instruction group that wrote
1916 // to the $ptr registers of the VTX_READ.
1918 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1919 // %T2_X<def> = MOV %ZERO
1920 //Adding this constraint prevents this from happening.
1921 let Constraints = "$ptr.ptr = $dst";
1924 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1925 : VTX_READ_eg <"VTX_READ_128", buffer_id, (outs R600_Reg128:$dst),
1928 let MEGA_FETCH_COUNT = 16;
1933 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1935 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1936 // that holds its buffer address to avoid potential hangs. We can't use
1937 // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
1938 // registers are different sizes.
1941 //===----------------------------------------------------------------------===//
1942 // VTX Read from parameter memory space
1943 //===----------------------------------------------------------------------===//
1945 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1946 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti8 ADDRVTX_READ:$ptr))]
1949 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1950 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti16 ADDRVTX_READ:$ptr))]
1953 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1954 [(set (i32 R600_TReg32_X:$dst), (load_param ADDRVTX_READ:$ptr))]
1957 def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
1958 [(set (v4i32 R600_Reg128:$dst), (load_param ADDRVTX_READ:$ptr))]
1961 //===----------------------------------------------------------------------===//
1962 // VTX Read from global memory space
1963 //===----------------------------------------------------------------------===//
1966 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1967 [(set (i32 R600_TReg32_X:$dst), (zextloadi8_global ADDRVTX_READ:$ptr))]
1971 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1972 [(set (i32 R600_TReg32_X:$dst), (global_load ADDRVTX_READ:$ptr))]
1976 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1977 [(set (v4i32 R600_Reg128:$dst), (global_load ADDRVTX_READ:$ptr))]
1980 //===----------------------------------------------------------------------===//
1982 // XXX: We are currently storing all constants in the global address space.
1983 //===----------------------------------------------------------------------===//
1985 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1986 [(set (i32 R600_TReg32_X:$dst), (constant_load ADDRVTX_READ:$ptr))]
1991 //===----------------------------------------------------------------------===//
1992 // Regist loads and stores - for indirect addressing
1993 //===----------------------------------------------------------------------===//
1995 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1997 let Predicates = [isCayman] in {
1999 let isVector = 1 in {
2001 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
2003 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
2004 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
2005 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
2006 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
2007 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
2008 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
2009 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
2010 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
2011 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
2012 def SIN_cm : SIN_Common<0x8D>;
2013 def COS_cm : COS_Common<0x8E>;
2014 } // End isVector = 1
2016 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL, R600_Reg32>;
2017 def : SIN_PAT <SIN_cm>;
2018 def : COS_PAT <COS_cm>;
2020 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
2022 // RECIP_UINT emulation for Cayman
2023 // The multiplication scales from [0,1] to the unsigned integer range
2025 (AMDGPUurecip R600_Reg32:$src0),
2026 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg R600_Reg32:$src0)),
2027 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
2030 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
2036 def : Pat<(fsqrt R600_Reg32:$src),
2037 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm R600_Reg32:$src))>;
2041 //===----------------------------------------------------------------------===//
2042 // Branch Instructions
2043 //===----------------------------------------------------------------------===//
2046 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
2047 "IF_PREDICATE_SET $src", []>;
2049 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
2050 "PREDICATED_BREAK $src", []>;
2052 //===----------------------------------------------------------------------===//
2053 // Pseudo instructions
2054 //===----------------------------------------------------------------------===//
2056 let isPseudo = 1 in {
2058 def PRED_X : InstR600 <
2059 (outs R600_Predicate_Bit:$dst),
2060 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
2062 let FlagOperandIdx = 3;
2065 let isTerminator = 1, isBranch = 1 in {
2066 def JUMP_COND : InstR600 <
2068 (ins brtarget:$target, R600_Predicate_Bit:$p),
2069 "JUMP $target ($p)",
2073 def JUMP : InstR600 <
2075 (ins brtarget:$target),
2080 let isPredicable = 1;
2084 } // End isTerminator = 1, isBranch = 1
2086 let usesCustomInserter = 1 in {
2088 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
2090 def MASK_WRITE : AMDGPUShaderInst <
2092 (ins R600_Reg32:$src),
2097 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
2101 (outs R600_Reg128:$dst),
2102 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2103 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2104 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, imm:$textureTarget))], NullALU> {
2108 def TXD_SHADOW: InstR600 <
2109 (outs R600_Reg128:$dst),
2110 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2111 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2112 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))], NullALU
2116 } // End isPseudo = 1
2117 } // End usesCustomInserter = 1
2119 def CLAMP_R600 : CLAMP <R600_Reg32>;
2120 def FABS_R600 : FABS<R600_Reg32>;
2121 def FNEG_R600 : FNEG<R600_Reg32>;
2123 //===---------------------------------------------------------------------===//
2124 // Return instruction
2125 //===---------------------------------------------------------------------===//
2126 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
2127 usesCustomInserter = 1 in {
2128 def RETURN : ILFormat<(outs), (ins variable_ops),
2129 "RETURN", [(IL_retflag)]>;
2133 //===----------------------------------------------------------------------===//
2134 // Constant Buffer Addressing Support
2135 //===----------------------------------------------------------------------===//
2137 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
2138 def CONST_COPY : Instruction {
2139 let OutOperandList = (outs R600_Reg32:$dst);
2140 let InOperandList = (ins i32imm:$src);
2142 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
2143 let AsmString = "CONST_COPY";
2144 let neverHasSideEffects = 1;
2145 let isAsCheapAsAMove = 1;
2146 let Itinerary = NullALU;
2148 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
2150 def TEX_VTX_CONSTBUF :
2151 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
2152 [(set R600_Reg128:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
2153 VTX_WORD1_GPR, VTX_WORD0 {
2157 let FETCH_WHOLE_QUAD = 0;
2161 let USE_CONST_FIELDS = 0;
2162 let NUM_FORMAT_ALL = 2;
2163 let FORMAT_COMP_ALL = 1;
2164 let SRF_MODE_ALL = 1;
2165 let MEGA_FETCH_COUNT = 16;
2170 let DATA_FORMAT = 35;
2172 let Inst{31-0} = Word0;
2173 let Inst{63-32} = Word1;
2175 // LLVM can only encode 64-bit instructions, so these fields are manually
2176 // encoded in R600CodeEmitter
2179 // bits<2> ENDIAN_SWAP = 0;
2180 // bits<1> CONST_BUF_NO_STRIDE = 0;
2181 // bits<1> MEGA_FETCH = 0;
2182 // bits<1> ALT_CONST = 0;
2183 // bits<2> BUFFER_INDEX_MODE = 0;
2187 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2188 // is done in R600CodeEmitter
2190 // Inst{79-64} = OFFSET;
2191 // Inst{81-80} = ENDIAN_SWAP;
2192 // Inst{82} = CONST_BUF_NO_STRIDE;
2193 // Inst{83} = MEGA_FETCH;
2194 // Inst{84} = ALT_CONST;
2195 // Inst{86-85} = BUFFER_INDEX_MODE;
2196 // Inst{95-86} = 0; Reserved
2198 // VTX_WORD3 (Padding)
2200 // Inst{127-96} = 0;
2205 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
2206 [(set R600_Reg128:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
2207 VTX_WORD1_GPR, VTX_WORD0 {
2211 let FETCH_WHOLE_QUAD = 0;
2215 let USE_CONST_FIELDS = 1;
2216 let NUM_FORMAT_ALL = 0;
2217 let FORMAT_COMP_ALL = 0;
2218 let SRF_MODE_ALL = 1;
2219 let MEGA_FETCH_COUNT = 16;
2224 let DATA_FORMAT = 0;
2226 let Inst{31-0} = Word0;
2227 let Inst{63-32} = Word1;
2229 // LLVM can only encode 64-bit instructions, so these fields are manually
2230 // encoded in R600CodeEmitter
2233 // bits<2> ENDIAN_SWAP = 0;
2234 // bits<1> CONST_BUF_NO_STRIDE = 0;
2235 // bits<1> MEGA_FETCH = 0;
2236 // bits<1> ALT_CONST = 0;
2237 // bits<2> BUFFER_INDEX_MODE = 0;
2241 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2242 // is done in R600CodeEmitter
2244 // Inst{79-64} = OFFSET;
2245 // Inst{81-80} = ENDIAN_SWAP;
2246 // Inst{82} = CONST_BUF_NO_STRIDE;
2247 // Inst{83} = MEGA_FETCH;
2248 // Inst{84} = ALT_CONST;
2249 // Inst{86-85} = BUFFER_INDEX_MODE;
2250 // Inst{95-86} = 0; Reserved
2252 // VTX_WORD3 (Padding)
2254 // Inst{127-96} = 0;
2260 //===--------------------------------------------------------------------===//
2261 // Instructions support
2262 //===--------------------------------------------------------------------===//
2263 //===---------------------------------------------------------------------===//
2264 // Custom Inserter for Branches and returns, this eventually will be a
2266 //===---------------------------------------------------------------------===//
2267 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
2268 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
2269 "; Pseudo unconditional branch instruction",
2271 defm BRANCH_COND : BranchConditional<IL_brcond>;
2274 //===---------------------------------------------------------------------===//
2275 // Flow and Program control Instructions
2276 //===---------------------------------------------------------------------===//
2277 let isTerminator=1 in {
2278 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
2279 !strconcat("SWITCH", " $src"), []>;
2280 def CASE : ILFormat< (outs), (ins GPRI32:$src),
2281 !strconcat("CASE", " $src"), []>;
2282 def BREAK : ILFormat< (outs), (ins),
2284 def CONTINUE : ILFormat< (outs), (ins),
2286 def DEFAULT : ILFormat< (outs), (ins),
2288 def ELSE : ILFormat< (outs), (ins),
2290 def ENDSWITCH : ILFormat< (outs), (ins),
2292 def ENDMAIN : ILFormat< (outs), (ins),
2294 def END : ILFormat< (outs), (ins),
2296 def ENDFUNC : ILFormat< (outs), (ins),
2298 def ENDIF : ILFormat< (outs), (ins),
2300 def WHILELOOP : ILFormat< (outs), (ins),
2302 def ENDLOOP : ILFormat< (outs), (ins),
2304 def FUNC : ILFormat< (outs), (ins),
2306 def RETDYN : ILFormat< (outs), (ins),
2308 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2309 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
2310 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2311 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
2312 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2313 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
2314 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2315 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
2316 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2317 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
2318 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2319 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
2320 defm IFC : BranchInstr2<"IFC">;
2321 defm BREAKC : BranchInstr2<"BREAKC">;
2322 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
2325 //===----------------------------------------------------------------------===//
2327 //===----------------------------------------------------------------------===//
2329 // CND*_INT Pattterns for f32 True / False values
2331 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
2332 (selectcc (i32 R600_Reg32:$src0), 0, (f32 R600_Reg32:$src1),
2333 R600_Reg32:$src2, cc),
2334 (cnd R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2)
2337 def : CND_INT_f32 <CNDE_INT, SETEQ>;
2338 def : CND_INT_f32 <CNDGT_INT, SETGT>;
2339 def : CND_INT_f32 <CNDGE_INT, SETGE>;
2341 //CNDGE_INT extra pattern
2343 (selectcc (i32 R600_Reg32:$src0), -1, (i32 R600_Reg32:$src1),
2344 (i32 R600_Reg32:$src2), COND_GT),
2345 (CNDGE_INT R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2)
2351 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
2355 (int_AMDGPU_kill R600_Reg32:$src0),
2356 (MASK_WRITE (KILLGT (f32 ZERO), (f32 R600_Reg32:$src0)))
2361 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LT),
2362 (SGT R600_Reg32:$src1, R600_Reg32:$src0)
2367 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LE),
2368 (SGE R600_Reg32:$src1, R600_Reg32:$src0)
2371 // SETGT_DX10 reverse args
2373 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, COND_LT),
2374 (SETGT_DX10 R600_Reg32:$src1, R600_Reg32:$src0)
2377 // SETGE_DX10 reverse args
2379 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, COND_LE),
2380 (SETGE_DX10 R600_Reg32:$src1, R600_Reg32:$src0)
2383 // SETGT_INT reverse args
2385 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLT),
2386 (SETGT_INT R600_Reg32:$src1, R600_Reg32:$src0)
2389 // SETGE_INT reverse args
2391 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLE),
2392 (SETGE_INT R600_Reg32:$src1, R600_Reg32:$src0)
2395 // SETGT_UINT reverse args
2397 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULT),
2398 (SETGT_UINT R600_Reg32:$src1, R600_Reg32:$src0)
2401 // SETGE_UINT reverse args
2403 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULE),
2404 (SETGE_UINT R600_Reg32:$src1, R600_Reg32:$src0)
2407 // The next two patterns are special cases for handling 'true if ordered' and
2408 // 'true if unordered' conditionals. The assumption here is that the behavior of
2409 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
2411 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
2412 // We assume that SETE returns false when one of the operands is NAN and
2413 // SNE returns true when on of the operands is NAN
2415 //SETE - 'true if ordered'
2417 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETO),
2418 (SETE R600_Reg32:$src0, R600_Reg32:$src1)
2421 //SETE_DX10 - 'true if ordered'
2423 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETO),
2424 (SETE_DX10 R600_Reg32:$src0, R600_Reg32:$src1)
2427 //SNE - 'true if unordered'
2429 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETUO),
2430 (SNE R600_Reg32:$src0, R600_Reg32:$src1)
2433 //SETNE_DX10 - 'true if ordered'
2435 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUO),
2436 (SETNE_DX10 R600_Reg32:$src0, R600_Reg32:$src1)
2439 def : Extract_Element <f32, v4f32, R600_Reg128, 0, sub0>;
2440 def : Extract_Element <f32, v4f32, R600_Reg128, 1, sub1>;
2441 def : Extract_Element <f32, v4f32, R600_Reg128, 2, sub2>;
2442 def : Extract_Element <f32, v4f32, R600_Reg128, 3, sub3>;
2444 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 0, sub0>;
2445 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 1, sub1>;
2446 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 2, sub2>;
2447 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 3, sub3>;
2449 def : Extract_Element <i32, v4i32, R600_Reg128, 0, sub0>;
2450 def : Extract_Element <i32, v4i32, R600_Reg128, 1, sub1>;
2451 def : Extract_Element <i32, v4i32, R600_Reg128, 2, sub2>;
2452 def : Extract_Element <i32, v4i32, R600_Reg128, 3, sub3>;
2454 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 0, sub0>;
2455 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 1, sub1>;
2456 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 2, sub2>;
2457 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 3, sub3>;
2459 def : Vector4_Build <v4f32, R600_Reg128, f32, R600_Reg32>;
2460 def : Vector4_Build <v4i32, R600_Reg128, i32, R600_Reg32>;
2462 // bitconvert patterns
2464 def : BitConvert <i32, f32, R600_Reg32>;
2465 def : BitConvert <f32, i32, R600_Reg32>;
2466 def : BitConvert <v4f32, v4i32, R600_Reg128>;
2467 def : BitConvert <v4i32, v4f32, R600_Reg128>;
2469 // DWORDADDR pattern
2470 def : DwordAddrPat <i32, R600_Reg32>;
2472 } // End isR600toCayman Predicate