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> {
24 bits<2> FlagOperandIdx = 0;
27 bit HasNativeOperands = 0;
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;
48 let TSFlags{12} = VTXInst;
49 let TSFlags{13} = TEXInst;
52 class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
53 InstR600 <outs, ins, asm, pattern, NullALU> {
55 let Namespace = "AMDGPU";
58 def MEMxi : Operand<iPTR> {
59 let MIOperandInfo = (ops R600_TReg32_X:$ptr, i32imm:$index);
60 let PrintMethod = "printMemOperand";
63 def MEMrr : Operand<iPTR> {
64 let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
67 // Operands for non-registers
69 class InstFlag<string PM = "printOperand", int Default = 0>
70 : OperandWithDefaultOps <i32, (ops (i32 Default))> {
74 // src_sel for ALU src operands, see also ALU_CONST, ALU_PARAM registers
75 def SEL : OperandWithDefaultOps <i32, (ops (i32 -1))> {
76 let PrintMethod = "printSel";
79 def LITERAL : InstFlag<"printLiteral">;
81 def WRITE : InstFlag <"printWrite", 1>;
82 def OMOD : InstFlag <"printOMOD">;
83 def REL : InstFlag <"printRel">;
84 def CLAMP : InstFlag <"printClamp">;
85 def NEG : InstFlag <"printNeg">;
86 def ABS : InstFlag <"printAbs">;
87 def UEM : InstFlag <"printUpdateExecMask">;
88 def UP : InstFlag <"printUpdatePred">;
90 // XXX: The r600g finalizer in Mesa expects last to be one in most cases.
91 // Once we start using the packetizer in this backend we should have this
93 def LAST : InstFlag<"printLast", 1>;
95 def FRAMEri : Operand<iPTR> {
96 let MIOperandInfo = (ops R600_Reg32:$ptr, i32imm:$index);
99 def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
100 def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
101 def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
102 def ADDRGA_CONST_OFFSET : ComplexPattern<i32, 1, "SelectGlobalValueConstantOffset", [], []>;
103 def ADDRGA_VAR_OFFSET : ComplexPattern<i32, 2, "SelectGlobalValueVariableOffset", [], []>;
104 def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>;
106 class R600ALU_Word0 {
107 field bits<32> Word0;
115 bits<3> index_mode = 0;
119 bits<9> src0_sel = src0{8-0};
120 bits<2> src0_chan = src0{10-9};
121 bits<9> src1_sel = src1{8-0};
122 bits<2> src1_chan = src1{10-9};
124 let Word0{8-0} = src0_sel;
125 let Word0{9} = src0_rel;
126 let Word0{11-10} = src0_chan;
127 let Word0{12} = src0_neg;
128 let Word0{21-13} = src1_sel;
129 let Word0{22} = src1_rel;
130 let Word0{24-23} = src1_chan;
131 let Word0{25} = src1_neg;
132 let Word0{28-26} = index_mode;
133 let Word0{30-29} = pred_sel;
134 let Word0{31} = last;
137 class R600ALU_Word1 {
138 field bits<32> Word1;
141 bits<3> bank_swizzle = 0;
145 bits<7> dst_sel = dst{6-0};
146 bits<2> dst_chan = dst{10-9};
148 let Word1{20-18} = bank_swizzle;
149 let Word1{27-21} = dst_sel;
150 let Word1{28} = dst_rel;
151 let Word1{30-29} = dst_chan;
152 let Word1{31} = clamp;
155 class R600ALU_Word1_OP2 <bits<11> alu_inst> : R600ALU_Word1{
159 bits<1> update_exec_mask;
164 let Word1{0} = src0_abs;
165 let Word1{1} = src1_abs;
166 let Word1{2} = update_exec_mask;
167 let Word1{3} = update_pred;
168 let Word1{4} = write;
169 let Word1{6-5} = omod;
170 let Word1{17-7} = alu_inst;
173 class R600ALU_Word1_OP3 <bits<5> alu_inst> : R600ALU_Word1{
179 bits<9> src2_sel = src2{8-0};
180 bits<2> src2_chan = src2{10-9};
182 let Word1{8-0} = src2_sel;
183 let Word1{9} = src2_rel;
184 let Word1{11-10} = src2_chan;
185 let Word1{12} = src2_neg;
186 let Word1{17-13} = alu_inst;
190 field bits<32> Word0;
194 bits<1> FETCH_WHOLE_QUAD;
198 bits<6> MEGA_FETCH_COUNT;
200 let Word0{4-0} = VC_INST;
201 let Word0{6-5} = FETCH_TYPE;
202 let Word0{7} = FETCH_WHOLE_QUAD;
203 let Word0{15-8} = BUFFER_ID;
204 let Word0{22-16} = SRC_GPR;
205 let Word0{23} = SRC_REL;
206 let Word0{25-24} = SRC_SEL_X;
207 let Word0{31-26} = MEGA_FETCH_COUNT;
210 class VTX_WORD1_GPR {
211 field bits<32> Word1;
218 bits<1> USE_CONST_FIELDS;
220 bits<2> NUM_FORMAT_ALL;
221 bits<1> FORMAT_COMP_ALL;
222 bits<1> SRF_MODE_ALL;
224 let Word1{6-0} = DST_GPR;
225 let Word1{7} = DST_REL;
226 let Word1{8} = 0; // Reserved
227 let Word1{11-9} = DST_SEL_X;
228 let Word1{14-12} = DST_SEL_Y;
229 let Word1{17-15} = DST_SEL_Z;
230 let Word1{20-18} = DST_SEL_W;
231 let Word1{21} = USE_CONST_FIELDS;
232 let Word1{27-22} = DATA_FORMAT;
233 let Word1{29-28} = NUM_FORMAT_ALL;
234 let Word1{30} = FORMAT_COMP_ALL;
235 let Word1{31} = SRF_MODE_ALL;
239 field bits<32> Word0;
243 bits<1> FETCH_WHOLE_QUAD;
248 bits<2> RESOURCE_INDEX_MODE;
249 bits<2> SAMPLER_INDEX_MODE;
251 let Word0{4-0} = TEX_INST;
252 let Word0{6-5} = INST_MOD;
253 let Word0{7} = FETCH_WHOLE_QUAD;
254 let Word0{15-8} = RESOURCE_ID;
255 let Word0{22-16} = SRC_GPR;
256 let Word0{23} = SRC_REL;
257 let Word0{24} = ALT_CONST;
258 let Word0{26-25} = RESOURCE_INDEX_MODE;
259 let Word0{28-27} = SAMPLER_INDEX_MODE;
263 field bits<32> Word1;
272 bits<1> COORD_TYPE_X;
273 bits<1> COORD_TYPE_Y;
274 bits<1> COORD_TYPE_Z;
275 bits<1> COORD_TYPE_W;
277 let Word1{6-0} = DST_GPR;
278 let Word1{7} = DST_REL;
279 let Word1{11-9} = DST_SEL_X;
280 let Word1{14-12} = DST_SEL_Y;
281 let Word1{17-15} = DST_SEL_Z;
282 let Word1{20-18} = DST_SEL_W;
283 let Word1{27-21} = LOD_BIAS;
284 let Word1{28} = COORD_TYPE_X;
285 let Word1{29} = COORD_TYPE_Y;
286 let Word1{30} = COORD_TYPE_Z;
287 let Word1{31} = COORD_TYPE_W;
291 field bits<32> Word2;
302 let Word2{4-0} = OFFSET_X;
303 let Word2{9-5} = OFFSET_Y;
304 let Word2{14-10} = OFFSET_Z;
305 let Word2{19-15} = SAMPLER_ID;
306 let Word2{22-20} = SRC_SEL_X;
307 let Word2{25-23} = SRC_SEL_Y;
308 let Word2{28-26} = SRC_SEL_Z;
309 let Word2{31-29} = SRC_SEL_W;
313 XXX: R600 subtarget uses a slightly different encoding than the other
314 subtargets. We currently handle this in R600MCCodeEmitter, but we may
315 want to use these instruction classes in the future.
317 class R600ALU_Word1_OP2_r600 : R600ALU_Word1_OP2 {
322 let Inst{37} = fog_merge;
323 let Inst{39-38} = omod;
324 let Inst{49-40} = alu_inst;
327 class R600ALU_Word1_OP2_r700 : R600ALU_Word1_OP2 {
331 let Inst{38-37} = omod;
332 let Inst{49-39} = alu_inst;
336 def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
340 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
342 // Class for instructions with only one source register.
343 // If you add new ins to this instruction, make sure they are listed before
344 // $literal, because the backend currently assumes that the last operand is
345 // a literal. Also be sure to update the enum R600Op1OperandIndex::ROI in
346 // R600Defines.h, R600InstrInfo::buildDefaultInstruction(),
347 // and R600InstrInfo::getOperandIdx().
348 class R600_1OP <bits<11> inst, string opName, list<dag> pattern,
349 InstrItinClass itin = AnyALU> :
350 InstR600 <(outs R600_Reg32:$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:$src0))]
383 // If you add our change the operands for R600_2OP instructions, you must
384 // also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
385 // R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
386 class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
387 InstrItinClass itin = AnyALU> :
388 InstR600 <(outs R600_Reg32:$dst),
389 (ins UEM:$update_exec_mask, UP:$update_pred, WRITE:$write,
390 OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
391 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
392 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs, SEL:$src1_sel,
393 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
394 !strconcat(" ", opName,
395 "$clamp $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
396 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
397 "$src1_neg$src1_abs$src1$src1_abs$src1_rel, "
398 "$literal $pred_sel$last"),
402 R600ALU_Word1_OP2 <inst> {
404 let HasNativeOperands = 1;
406 let DisableEncoding = "$literal";
408 let Inst{31-0} = Word0;
409 let Inst{63-32} = Word1;
412 class R600_2OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
413 InstrItinClass itim = AnyALU> :
414 R600_2OP <inst, opName,
415 [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
419 // If you add our change the operands for R600_3OP instructions, you must
420 // also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
421 // R600InstrInfo::buildDefaultInstruction(), and
422 // R600InstrInfo::getOperandIdx().
423 class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
424 InstrItinClass itin = AnyALU> :
425 InstR600 <(outs R600_Reg32:$dst),
426 (ins REL:$dst_rel, CLAMP:$clamp,
427 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, SEL:$src0_sel,
428 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
429 R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
430 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
431 !strconcat(" ", opName, "$clamp $dst$dst_rel, "
432 "$src0_neg$src0$src0_rel, "
433 "$src1_neg$src1$src1_rel, "
434 "$src2_neg$src2$src2_rel, "
435 "$literal $pred_sel$last"),
439 R600ALU_Word1_OP3<inst>{
441 let HasNativeOperands = 1;
442 let DisableEncoding = "$literal";
445 let Inst{31-0} = Word0;
446 let Inst{63-32} = Word1;
449 class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
450 InstrItinClass itin = VecALU> :
451 InstR600 <(outs R600_Reg32:$dst),
457 class R600_TEX <bits<11> inst, string opName, list<dag> pattern,
458 InstrItinClass itin = AnyALU> :
459 InstR600 <(outs R600_Reg128:$DST_GPR),
460 (ins R600_Reg128:$SRC_GPR, i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID, i32imm:$textureTarget),
461 !strconcat(opName, "$DST_GPR, $SRC_GPR, $RESOURCE_ID, $SAMPLER_ID, $textureTarget"),
463 itin>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
464 let Inst{31-0} = Word0;
465 let Inst{63-32} = Word1;
467 let TEX_INST = inst{4-0};
477 let FETCH_WHOLE_QUAD = 0;
479 let SAMPLER_INDEX_MODE = 0;
480 let RESOURCE_INDEX_MODE = 0;
482 let COORD_TYPE_X = 0;
483 let COORD_TYPE_Y = 0;
484 let COORD_TYPE_Z = 0;
485 let COORD_TYPE_W = 0;
490 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
492 def TEX_SHADOW : PatLeaf<
494 [{uint32_t TType = (uint32_t)N->getZExtValue();
495 return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
499 def TEX_RECT : PatLeaf<
501 [{uint32_t TType = (uint32_t)N->getZExtValue();
506 def TEX_ARRAY : PatLeaf<
508 [{uint32_t TType = (uint32_t)N->getZExtValue();
509 return TType == 9 || TType == 10 || TType == 15 || TType == 16;
513 def TEX_SHADOW_ARRAY : PatLeaf<
515 [{uint32_t TType = (uint32_t)N->getZExtValue();
516 return TType == 11 || TType == 12 || TType == 17;
520 class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, bits<4> rat_id, dag outs,
521 dag ins, string asm, list<dag> pattern> :
522 InstR600ISA <outs, ins, asm, pattern> {
539 // CF_ALLOC_EXPORT_WORD0_RAT
540 let Inst{3-0} = rat_id;
541 let Inst{9-4} = rat_inst;
542 let Inst{10} = 0; // Reserved
543 let Inst{12-11} = RIM;
544 let Inst{14-13} = TYPE;
545 let Inst{21-15} = RW_GPR;
546 let Inst{22} = RW_REL;
547 let Inst{29-23} = INDEX_GPR;
548 let Inst{31-30} = ELEM_SIZE;
550 // CF_ALLOC_EXPORT_WORD1_BUF
551 let Inst{43-32} = ARRAY_SIZE;
552 let Inst{47-44} = COMP_MASK;
553 let Inst{51-48} = BURST_COUNT;
556 let Inst{61-54} = cf_inst;
558 let Inst{63} = BARRIER;
561 class LoadParamFrag <PatFrag load_type> : PatFrag <
562 (ops node:$ptr), (load_type node:$ptr),
563 [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
566 def load_param : LoadParamFrag<load>;
567 def load_param_zexti8 : LoadParamFrag<zextloadi8>;
568 def load_param_zexti16 : LoadParamFrag<zextloadi16>;
570 def isR600 : Predicate<"Subtarget.device()"
571 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX">;
572 def isR700 : Predicate<"Subtarget.device()"
573 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX &&"
574 "Subtarget.device()->getDeviceFlag()"
575 ">= OCL_DEVICE_RV710">;
576 def isEG : Predicate<
577 "Subtarget.device()->getGeneration() >= AMDGPUDeviceInfo::HD5XXX && "
578 "Subtarget.device()->getGeneration() < AMDGPUDeviceInfo::HD7XXX && "
579 "Subtarget.device()->getDeviceFlag() != OCL_DEVICE_CAYMAN">;
581 def isCayman : Predicate<"Subtarget.device()"
582 "->getDeviceFlag() == OCL_DEVICE_CAYMAN">;
583 def isEGorCayman : Predicate<"Subtarget.device()"
584 "->getGeneration() == AMDGPUDeviceInfo::HD5XXX"
585 "|| Subtarget.device()->getGeneration() =="
586 "AMDGPUDeviceInfo::HD6XXX">;
588 def isR600toCayman : Predicate<
589 "Subtarget.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX">;
591 //===----------------------------------------------------------------------===//
593 //===----------------------------------------------------------------------===//
595 def INTERP_PAIR_XY : AMDGPUShaderInst <
596 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
597 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
598 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
601 def INTERP_PAIR_ZW : AMDGPUShaderInst <
602 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
603 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
604 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
607 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
608 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
612 //===----------------------------------------------------------------------===//
613 // Interpolation Instructions
614 //===----------------------------------------------------------------------===//
616 def INTERP_VEC_LOAD : AMDGPUShaderInst <
617 (outs R600_Reg128:$dst),
619 "INTERP_LOAD $src0 : $dst",
622 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
623 let bank_swizzle = 5;
626 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
627 let bank_swizzle = 5;
630 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
632 //===----------------------------------------------------------------------===//
633 // Export Instructions
634 //===----------------------------------------------------------------------===//
636 def ExportType : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
638 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
639 [SDNPHasChain, SDNPSideEffect]>;
642 field bits<32> Word0;
649 let Word0{12-0} = arraybase;
650 let Word0{14-13} = type;
651 let Word0{21-15} = gpr;
652 let Word0{22} = 0; // RW_REL
653 let Word0{29-23} = 0; // INDEX_GPR
654 let Word0{31-30} = elem_size;
657 class ExportSwzWord1 {
658 field bits<32> Word1;
667 let Word1{2-0} = sw_x;
668 let Word1{5-3} = sw_y;
669 let Word1{8-6} = sw_z;
670 let Word1{11-9} = sw_w;
673 class ExportBufWord1 {
674 field bits<32> Word1;
681 let Word1{11-0} = arraySize;
682 let Word1{15-12} = compMask;
685 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
686 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
688 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
689 0, 61, 0, 7, 7, 7, cf_inst, 0)
692 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
694 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
695 0, 61, 7, 0, 7, 7, cf_inst, 0)
698 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
700 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
703 def : Pat<(int_R600_store_dummy 1),
705 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
708 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
709 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
710 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
711 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
716 multiclass SteamOutputExportPattern<Instruction ExportInst,
717 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
719 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
720 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
721 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
722 4095, imm:$mask, buf0inst, 0)>;
724 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
725 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
726 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
727 4095, imm:$mask, buf1inst, 0)>;
729 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
730 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
731 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
732 4095, imm:$mask, buf2inst, 0)>;
734 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
735 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
736 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
737 4095, imm:$mask, buf3inst, 0)>;
740 // Export Instructions should not be duplicated by TailDuplication pass
741 // (which assumes that duplicable instruction are affected by exec mask)
742 let usesCustomInserter = 1, isNotDuplicable = 1 in {
744 class ExportSwzInst : InstR600ISA<(
746 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
747 i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
749 !strconcat("EXPORT", " $gpr"),
750 []>, ExportWord0, ExportSwzWord1 {
752 let Inst{31-0} = Word0;
753 let Inst{63-32} = Word1;
756 } // End usesCustomInserter = 1
758 class ExportBufInst : InstR600ISA<(
760 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
761 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
762 !strconcat("EXPORT", " $gpr"),
763 []>, ExportWord0, ExportBufWord1 {
765 let Inst{31-0} = Word0;
766 let Inst{63-32} = Word1;
769 //===----------------------------------------------------------------------===//
770 // Control Flow Instructions
771 //===----------------------------------------------------------------------===//
774 field bits<32> Word0;
777 bits<4> KCACHE_BANK0;
778 bits<4> KCACHE_BANK1;
779 bits<2> KCACHE_MODE0;
781 let Word0{21-0} = ADDR;
782 let Word0{25-22} = KCACHE_BANK0;
783 let Word0{29-26} = KCACHE_BANK1;
784 let Word0{31-30} = KCACHE_MODE0;
788 field bits<32> Word1;
790 bits<2> KCACHE_MODE1;
791 bits<8> KCACHE_ADDR0;
792 bits<8> KCACHE_ADDR1;
796 bits<1> WHOLE_QUAD_MODE;
799 let Word1{1-0} = KCACHE_MODE1;
800 let Word1{9-2} = KCACHE_ADDR0;
801 let Word1{17-10} = KCACHE_ADDR1;
802 let Word1{24-18} = COUNT;
803 let Word1{25} = ALT_CONST;
804 let Word1{29-26} = CF_INST;
805 let Word1{30} = WHOLE_QUAD_MODE;
806 let Word1{31} = BARRIER;
809 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
810 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1, i32imm:$KCACHE_MODE0, i32imm:$KCACHE_MODE1,
811 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1, i32imm:$COUNT),
812 !strconcat(OpName, " $COUNT, @$ADDR, "
813 "KC0[CB$KCACHE_BANK0:$KCACHE_ADDR0-$KCACHE_ADDR0+32]"
814 ", KC1[CB$KCACHE_BANK1:$KCACHE_ADDR1-$KCACHE_ADDR1+32]"),
815 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
820 let WHOLE_QUAD_MODE = 0;
823 let Inst{31-0} = Word0;
824 let Inst{63-32} = Word1;
827 class CF_WORD0_R600 {
828 field bits<32> Word0;
835 class CF_WORD1_R600 {
836 field bits<32> Word1;
844 bits<1> END_OF_PROGRAM;
845 bits<1> VALID_PIXEL_MODE;
847 bits<1> WHOLE_QUAD_MODE;
850 let Word1{2-0} = POP_COUNT;
851 let Word1{7-3} = CF_CONST;
852 let Word1{9-8} = COND;
853 let Word1{12-10} = COUNT;
854 let Word1{18-13} = CALL_COUNT;
855 let Word1{19} = COUNT_3;
856 let Word1{21} = END_OF_PROGRAM;
857 let Word1{22} = VALID_PIXEL_MODE;
858 let Word1{29-23} = CF_INST;
859 let Word1{30} = WHOLE_QUAD_MODE;
860 let Word1{31} = BARRIER;
863 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
864 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
870 let VALID_PIXEL_MODE = 0;
874 let END_OF_PROGRAM = 0;
875 let WHOLE_QUAD_MODE = 0;
877 let Inst{31-0} = Word0;
878 let Inst{63-32} = Word1;
882 field bits<32> Word0;
885 bits<3> JUMPTABLE_SEL;
887 let Word0{23-0} = ADDR;
888 let Word0{26-24} = JUMPTABLE_SEL;
892 field bits<32> Word1;
898 bits<1> VALID_PIXEL_MODE;
899 bits<1> END_OF_PROGRAM;
903 let Word1{2-0} = POP_COUNT;
904 let Word1{7-3} = CF_CONST;
905 let Word1{9-8} = COND;
906 let Word1{15-10} = COUNT;
907 let Word1{20} = VALID_PIXEL_MODE;
908 let Word1{21} = END_OF_PROGRAM;
909 let Word1{29-22} = CF_INST;
910 let Word1{31} = BARRIER;
913 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
914 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
919 let JUMPTABLE_SEL = 0;
921 let VALID_PIXEL_MODE = 0;
923 let END_OF_PROGRAM = 0;
925 let Inst{31-0} = Word0;
926 let Inst{63-32} = Word1;
929 def CF_ALU : ALU_CLAUSE<8, "ALU">;
930 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
932 def FETCH_CLAUSE : AMDGPUInst <(outs),
933 (ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
939 def PAD : AMDGPUInst <(outs), (ins), "PAD", [] > {
943 let Predicates = [isR600toCayman] in {
945 //===----------------------------------------------------------------------===//
946 // Common Instructions R600, R700, Evergreen, Cayman
947 //===----------------------------------------------------------------------===//
949 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
950 // Non-IEEE MUL: 0 * anything = 0
951 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
952 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
953 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
954 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
956 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
957 // so some of the instruction names don't match the asm string.
958 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
959 def SETE : R600_2OP <
961 [(set R600_Reg32:$dst,
962 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
968 [(set R600_Reg32:$dst,
969 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
975 [(set R600_Reg32:$dst,
976 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
982 [(set R600_Reg32:$dst,
983 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
987 def SETE_DX10 : R600_2OP <
989 [(set R600_Reg32:$dst,
990 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
994 def SETGT_DX10 : R600_2OP <
996 [(set R600_Reg32:$dst,
997 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
1001 def SETGE_DX10 : R600_2OP <
1003 [(set R600_Reg32:$dst,
1004 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
1008 def SETNE_DX10 : R600_2OP <
1010 [(set R600_Reg32:$dst,
1011 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
1015 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
1016 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
1017 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
1018 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
1019 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
1021 def MOV : R600_1OP <0x19, "MOV", []>;
1023 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
1025 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
1026 (outs R600_Reg32:$dst),
1032 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
1034 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
1037 (MOV_IMM_I32 imm:$val)
1040 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
1043 (MOV_IMM_F32 fpimm:$val)
1046 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
1047 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
1048 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
1049 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
1051 let hasSideEffects = 1 in {
1053 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
1055 } // end hasSideEffects
1057 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
1058 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
1059 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
1060 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
1061 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
1062 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
1063 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
1064 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
1065 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
1066 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
1068 def SETE_INT : R600_2OP <
1070 [(set (i32 R600_Reg32:$dst),
1071 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETEQ))]
1074 def SETGT_INT : R600_2OP <
1076 [(set (i32 R600_Reg32:$dst),
1077 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGT))]
1080 def SETGE_INT : R600_2OP <
1082 [(set (i32 R600_Reg32:$dst),
1083 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGE))]
1086 def SETNE_INT : R600_2OP <
1088 [(set (i32 R600_Reg32:$dst),
1089 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETNE))]
1092 def SETGT_UINT : R600_2OP <
1094 [(set (i32 R600_Reg32:$dst),
1095 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGT))]
1098 def SETGE_UINT : R600_2OP <
1100 [(set (i32 R600_Reg32:$dst),
1101 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGE))]
1104 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
1105 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
1106 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
1107 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
1109 def CNDE_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 CNDGE_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 def CNDGT_INT : R600_3OP <
1127 [(set (i32 R600_Reg32:$dst),
1128 (selectcc (i32 R600_Reg32:$src0), 0,
1129 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
1133 //===----------------------------------------------------------------------===//
1134 // Texture instructions
1135 //===----------------------------------------------------------------------===//
1137 def TEX_LD : R600_TEX <
1139 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txf R600_Reg128:$SRC_GPR,
1140 imm:$OFFSET_X, imm:$OFFSET_Y, imm:$OFFSET_Z, imm:$RESOURCE_ID,
1141 imm:$SAMPLER_ID, imm:$textureTarget))]
1143 let AsmString = "TEX_LD $DST_GPR, $SRC_GPR, $OFFSET_X, $OFFSET_Y, $OFFSET_Z,"
1144 "$RESOURCE_ID, $SAMPLER_ID, $textureTarget";
1145 let InOperandList = (ins R600_Reg128:$SRC_GPR, i32imm:$OFFSET_X,
1146 i32imm:$OFFSET_Y, i32imm:$OFFSET_Z, i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
1147 i32imm:$textureTarget);
1150 def TEX_GET_TEXTURE_RESINFO : R600_TEX <
1151 0x04, "TEX_GET_TEXTURE_RESINFO",
1152 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txq R600_Reg128:$SRC_GPR,
1153 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1156 def TEX_GET_GRADIENTS_H : R600_TEX <
1157 0x07, "TEX_GET_GRADIENTS_H",
1158 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_ddx R600_Reg128:$SRC_GPR,
1159 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1162 def TEX_GET_GRADIENTS_V : R600_TEX <
1163 0x08, "TEX_GET_GRADIENTS_V",
1164 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_ddy R600_Reg128:$SRC_GPR,
1165 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1168 def TEX_SET_GRADIENTS_H : R600_TEX <
1169 0x0B, "TEX_SET_GRADIENTS_H",
1173 def TEX_SET_GRADIENTS_V : R600_TEX <
1174 0x0C, "TEX_SET_GRADIENTS_V",
1178 def TEX_SAMPLE : R600_TEX <
1180 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_tex R600_Reg128:$SRC_GPR,
1181 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1184 def TEX_SAMPLE_C : R600_TEX <
1185 0x18, "TEX_SAMPLE_C",
1186 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_tex R600_Reg128:$SRC_GPR,
1187 imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
1190 def TEX_SAMPLE_L : R600_TEX <
1191 0x11, "TEX_SAMPLE_L",
1192 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txl R600_Reg128:$SRC_GPR,
1193 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1196 def TEX_SAMPLE_C_L : R600_TEX <
1197 0x19, "TEX_SAMPLE_C_L",
1198 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txl R600_Reg128:$SRC_GPR,
1199 imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
1202 def TEX_SAMPLE_LB : R600_TEX <
1203 0x12, "TEX_SAMPLE_LB",
1204 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txb R600_Reg128:$SRC_GPR,
1205 imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
1208 def TEX_SAMPLE_C_LB : R600_TEX <
1209 0x1A, "TEX_SAMPLE_C_LB",
1210 [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txb R600_Reg128:$SRC_GPR,
1211 imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
1214 def TEX_SAMPLE_G : R600_TEX <
1215 0x14, "TEX_SAMPLE_G",
1219 def TEX_SAMPLE_C_G : R600_TEX <
1220 0x1C, "TEX_SAMPLE_C_G",
1224 //===----------------------------------------------------------------------===//
1225 // Helper classes for common instructions
1226 //===----------------------------------------------------------------------===//
1228 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
1233 class MULADD_Common <bits<5> inst> : R600_3OP <
1238 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
1239 inst, "MULADD_IEEE",
1240 [(set (f32 R600_Reg32:$dst),
1241 (fadd (fmul R600_Reg32:$src0, R600_Reg32:$src1), R600_Reg32:$src2))]
1244 class CNDE_Common <bits<5> inst> : R600_3OP <
1246 [(set R600_Reg32:$dst,
1247 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
1248 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
1252 class CNDGT_Common <bits<5> inst> : R600_3OP <
1254 [(set R600_Reg32:$dst,
1255 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
1256 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
1260 class CNDGE_Common <bits<5> inst> : R600_3OP <
1262 [(set R600_Reg32:$dst,
1263 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
1264 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
1268 multiclass DOT4_Common <bits<11> inst> {
1270 def _pseudo : R600_REDUCTION <inst,
1271 (ins R600_Reg128:$src0, R600_Reg128:$src1),
1272 "DOT4 $dst $src0, $src1",
1273 [(set R600_Reg32:$dst, (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1))]
1276 def _real : R600_2OP <inst, "DOT4", []>;
1279 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1280 multiclass CUBE_Common <bits<11> inst> {
1282 def _pseudo : InstR600 <
1283 (outs R600_Reg128:$dst),
1284 (ins R600_Reg128:$src),
1286 [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
1292 def _real : R600_2OP <inst, "CUBE", []>;
1294 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1296 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1297 inst, "EXP_IEEE", fexp2
1300 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
1301 inst, "FLT_TO_INT", fp_to_sint
1304 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1305 inst, "INT_TO_FLT", sint_to_fp
1308 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1309 inst, "FLT_TO_UINT", fp_to_uint
1312 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1313 inst, "UINT_TO_FLT", uint_to_fp
1316 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1317 inst, "LOG_CLAMPED", []
1320 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1321 inst, "LOG_IEEE", flog2
1324 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1325 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1326 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1327 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1328 inst, "MULHI_INT", mulhs
1330 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1331 inst, "MULHI", mulhu
1333 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1334 inst, "MULLO_INT", mul
1336 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []>;
1338 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1339 inst, "RECIP_CLAMPED", []
1342 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1343 inst, "RECIP_IEEE", [(set R600_Reg32:$dst, (fdiv FP_ONE, R600_Reg32:$src0))]
1346 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1347 inst, "RECIP_UINT", AMDGPUurecip
1350 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1351 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1354 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1355 inst, "RECIPSQRT_IEEE", []
1358 class SIN_Common <bits<11> inst> : R600_1OP <
1363 class COS_Common <bits<11> inst> : R600_1OP <
1368 //===----------------------------------------------------------------------===//
1369 // Helper patterns for complex intrinsics
1370 //===----------------------------------------------------------------------===//
1372 multiclass DIV_Common <InstR600 recip_ieee> {
1374 (int_AMDGPU_div R600_Reg32:$src0, R600_Reg32:$src1),
1375 (MUL_IEEE R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1379 (fdiv R600_Reg32:$src0, R600_Reg32:$src1),
1380 (MUL_IEEE R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1384 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee> : Pat <
1385 (int_TGSI_lit_z R600_Reg32:$src_x, R600_Reg32:$src_y, R600_Reg32:$src_w),
1386 (exp_ieee (mul_lit (log_clamped (MAX R600_Reg32:$src_y, (f32 ZERO))), R600_Reg32:$src_w, R600_Reg32:$src_x))
1389 //===----------------------------------------------------------------------===//
1390 // R600 / R700 Instructions
1391 //===----------------------------------------------------------------------===//
1393 let Predicates = [isR600] in {
1395 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1396 def MULADD_r600 : MULADD_Common<0x10>;
1397 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1398 def CNDE_r600 : CNDE_Common<0x18>;
1399 def CNDGT_r600 : CNDGT_Common<0x19>;
1400 def CNDGE_r600 : CNDGE_Common<0x1A>;
1401 defm DOT4_r600 : DOT4_Common<0x50>;
1402 defm CUBE_r600 : CUBE_Common<0x52>;
1403 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1404 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1405 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1406 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1407 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1408 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1409 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1410 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1411 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1412 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1413 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1414 def SIN_r600 : SIN_Common<0x6E>;
1415 def COS_r600 : COS_Common<0x6F>;
1416 def ASHR_r600 : ASHR_Common<0x70>;
1417 def LSHR_r600 : LSHR_Common<0x71>;
1418 def LSHL_r600 : LSHL_Common<0x72>;
1419 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1420 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1421 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1422 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1423 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1425 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1426 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL, R600_Reg32>;
1427 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1429 def : Pat<(fsqrt R600_Reg32:$src),
1430 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_r600 R600_Reg32:$src))>;
1432 def R600_ExportSwz : ExportSwzInst {
1433 let Word1{20-17} = 0; // BURST_COUNT
1434 let Word1{21} = eop;
1435 let Word1{22} = 1; // VALID_PIXEL_MODE
1436 let Word1{30-23} = inst;
1437 let Word1{31} = 1; // BARRIER
1439 defm : ExportPattern<R600_ExportSwz, 39>;
1441 def R600_ExportBuf : ExportBufInst {
1442 let Word1{20-17} = 0; // BURST_COUNT
1443 let Word1{21} = eop;
1444 let Word1{22} = 1; // VALID_PIXEL_MODE
1445 let Word1{30-23} = inst;
1446 let Word1{31} = 1; // BARRIER
1448 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1450 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1451 "TEX $COUNT @$ADDR"> {
1454 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1455 "VTX $COUNT @$ADDR"> {
1458 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1459 "LOOP_START_DX10 @$ADDR"> {
1463 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1467 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1468 "LOOP_BREAK @$ADDR"> {
1472 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1473 "CONTINUE @$ADDR"> {
1477 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1478 "JUMP @$ADDR POP:$POP_COUNT"> {
1481 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1482 "ELSE @$ADDR POP:$POP_COUNT"> {
1485 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1490 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1491 "POP @$ADDR POP:$POP_COUNT"> {
1494 def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
1498 let END_OF_PROGRAM = 1;
1503 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1505 class COS_PAT <InstR600 trig> : Pat<
1506 (fcos R600_Reg32:$src),
1507 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1510 class SIN_PAT <InstR600 trig> : Pat<
1511 (fsin R600_Reg32:$src),
1512 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1515 //===----------------------------------------------------------------------===//
1516 // R700 Only instructions
1517 //===----------------------------------------------------------------------===//
1519 let Predicates = [isR700] in {
1520 def SIN_r700 : SIN_Common<0x6E>;
1521 def COS_r700 : COS_Common<0x6F>;
1523 // R700 normalizes inputs to SIN/COS the same as EG
1524 def : SIN_PAT <SIN_r700>;
1525 def : COS_PAT <COS_r700>;
1528 //===----------------------------------------------------------------------===//
1529 // Evergreen Only instructions
1530 //===----------------------------------------------------------------------===//
1532 let Predicates = [isEG] in {
1534 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1535 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1537 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1538 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1539 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1540 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1541 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1542 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1543 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1544 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1545 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1546 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1547 def SIN_eg : SIN_Common<0x8D>;
1548 def COS_eg : COS_Common<0x8E>;
1550 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL, R600_Reg32>;
1551 def : SIN_PAT <SIN_eg>;
1552 def : COS_PAT <COS_eg>;
1553 def : Pat<(fsqrt R600_Reg32:$src),
1554 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_eg R600_Reg32:$src))>;
1555 } // End Predicates = [isEG]
1557 //===----------------------------------------------------------------------===//
1558 // Evergreen / Cayman Instructions
1559 //===----------------------------------------------------------------------===//
1561 let Predicates = [isEGorCayman] in {
1563 // BFE_UINT - bit_extract, an optimization for mask and shift
1568 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1573 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1574 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1575 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1576 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1577 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1578 [(set R600_Reg32:$dst, (int_AMDIL_bit_extract_u32 R600_Reg32:$src0,
1580 R600_Reg32:$src2))],
1584 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", []>;
1585 defm : BFIPatterns <BFI_INT_eg>;
1587 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
1588 [(set R600_Reg32:$dst, (AMDGPUbitalign R600_Reg32:$src0, R600_Reg32:$src1,
1589 R600_Reg32:$src2))],
1593 def MULADD_eg : MULADD_Common<0x14>;
1594 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
1595 def ASHR_eg : ASHR_Common<0x15>;
1596 def LSHR_eg : LSHR_Common<0x16>;
1597 def LSHL_eg : LSHL_Common<0x17>;
1598 def CNDE_eg : CNDE_Common<0x19>;
1599 def CNDGT_eg : CNDGT_Common<0x1A>;
1600 def CNDGE_eg : CNDGE_Common<0x1B>;
1601 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1602 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1603 defm DOT4_eg : DOT4_Common<0xBE>;
1604 defm CUBE_eg : CUBE_Common<0xC0>;
1606 let hasSideEffects = 1 in {
1607 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1610 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1612 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1616 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1618 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1622 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1624 // TRUNC is used for the FLT_TO_INT instructions to work around a
1625 // perceived problem where the rounding modes are applied differently
1626 // depending on the instruction and the slot they are in.
1628 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1629 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1631 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1632 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1633 // We should look into handling these cases separately.
1634 def : Pat<(fp_to_sint R600_Reg32:$src0),
1635 (FLT_TO_INT_eg (TRUNC R600_Reg32:$src0))>;
1637 def : Pat<(fp_to_uint R600_Reg32:$src0),
1638 (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src0))>;
1640 def EG_ExportSwz : ExportSwzInst {
1641 let Word1{19-16} = 0; // BURST_COUNT
1642 let Word1{20} = 1; // VALID_PIXEL_MODE
1643 let Word1{21} = eop;
1644 let Word1{29-22} = inst;
1645 let Word1{30} = 0; // MARK
1646 let Word1{31} = 1; // BARRIER
1648 defm : ExportPattern<EG_ExportSwz, 83>;
1650 def EG_ExportBuf : ExportBufInst {
1651 let Word1{19-16} = 0; // BURST_COUNT
1652 let Word1{20} = 1; // VALID_PIXEL_MODE
1653 let Word1{21} = eop;
1654 let Word1{29-22} = inst;
1655 let Word1{30} = 0; // MARK
1656 let Word1{31} = 1; // BARRIER
1658 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1660 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1661 "TEX $COUNT @$ADDR"> {
1664 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1665 "VTX $COUNT @$ADDR"> {
1668 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
1669 "LOOP_START_DX10 @$ADDR"> {
1673 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1677 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
1678 "LOOP_BREAK @$ADDR"> {
1682 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
1683 "CONTINUE @$ADDR"> {
1687 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1688 "JUMP @$ADDR POP:$POP_COUNT"> {
1691 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1692 "ELSE @$ADDR POP:$POP_COUNT"> {
1695 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
1700 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1701 "POP @$ADDR POP:$POP_COUNT"> {
1704 def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
1708 let END_OF_PROGRAM = 1;
1711 //===----------------------------------------------------------------------===//
1712 // Memory read/write instructions
1713 //===----------------------------------------------------------------------===//
1714 let usesCustomInserter = 1 in {
1716 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
1718 : EG_CF_RAT <0x57, 0x2, 0, (outs), ins,
1719 !strconcat(name, " $rw_gpr, $index_gpr, $eop"), pattern> {
1721 // XXX: Have a separate instruction for non-indexed writes.
1727 let COMP_MASK = comp_mask;
1728 let BURST_COUNT = 0;
1734 } // End usesCustomInserter = 1
1737 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1738 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1739 0x1, "RAT_WRITE_CACHELESS_32_eg",
1740 [(global_store (i32 R600_TReg32_X:$rw_gpr), R600_TReg32_X:$index_gpr)]
1744 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1745 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1746 0xf, "RAT_WRITE_CACHELESS_128",
1747 [(global_store (v4i32 R600_Reg128:$rw_gpr), R600_TReg32_X:$index_gpr)]
1750 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1751 : InstR600ISA <outs, (ins MEMxi:$ptr), name#" $dst, $ptr", pattern>,
1752 VTX_WORD1_GPR, VTX_WORD0 {
1757 let FETCH_WHOLE_QUAD = 0;
1758 let BUFFER_ID = buffer_id;
1760 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1761 // to store vertex addresses in any channel, not just X.
1764 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
1765 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
1766 // however, based on my testing if USE_CONST_FIELDS is set, then all
1767 // these fields need to be set to 0.
1768 let USE_CONST_FIELDS = 0;
1769 let NUM_FORMAT_ALL = 1;
1770 let FORMAT_COMP_ALL = 0;
1771 let SRF_MODE_ALL = 0;
1773 let Inst{31-0} = Word0;
1774 let Inst{63-32} = Word1;
1775 // LLVM can only encode 64-bit instructions, so these fields are manually
1776 // encoded in R600CodeEmitter
1779 // bits<2> ENDIAN_SWAP = 0;
1780 // bits<1> CONST_BUF_NO_STRIDE = 0;
1781 // bits<1> MEGA_FETCH = 0;
1782 // bits<1> ALT_CONST = 0;
1783 // bits<2> BUFFER_INDEX_MODE = 0;
1787 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1788 // is done in R600CodeEmitter
1790 // Inst{79-64} = OFFSET;
1791 // Inst{81-80} = ENDIAN_SWAP;
1792 // Inst{82} = CONST_BUF_NO_STRIDE;
1793 // Inst{83} = MEGA_FETCH;
1794 // Inst{84} = ALT_CONST;
1795 // Inst{86-85} = BUFFER_INDEX_MODE;
1796 // Inst{95-86} = 0; Reserved
1798 // VTX_WORD3 (Padding)
1800 // Inst{127-96} = 0;
1805 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1806 : VTX_READ_eg <"VTX_READ_8", buffer_id, (outs R600_TReg32_X:$dst),
1809 let MEGA_FETCH_COUNT = 1;
1811 let DST_SEL_Y = 7; // Masked
1812 let DST_SEL_Z = 7; // Masked
1813 let DST_SEL_W = 7; // Masked
1814 let DATA_FORMAT = 1; // FMT_8
1817 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1818 : VTX_READ_eg <"VTX_READ_16", buffer_id, (outs R600_TReg32_X:$dst),
1820 let MEGA_FETCH_COUNT = 2;
1822 let DST_SEL_Y = 7; // Masked
1823 let DST_SEL_Z = 7; // Masked
1824 let DST_SEL_W = 7; // Masked
1825 let DATA_FORMAT = 5; // FMT_16
1829 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1830 : VTX_READ_eg <"VTX_READ_32", buffer_id, (outs R600_TReg32_X:$dst),
1833 let MEGA_FETCH_COUNT = 4;
1835 let DST_SEL_Y = 7; // Masked
1836 let DST_SEL_Z = 7; // Masked
1837 let DST_SEL_W = 7; // Masked
1838 let DATA_FORMAT = 0xD; // COLOR_32
1840 // This is not really necessary, but there were some GPU hangs that appeared
1841 // to be caused by ALU instructions in the next instruction group that wrote
1842 // to the $ptr registers of the VTX_READ.
1844 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1845 // %T2_X<def> = MOV %ZERO
1846 //Adding this constraint prevents this from happening.
1847 let Constraints = "$ptr.ptr = $dst";
1850 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1851 : VTX_READ_eg <"VTX_READ_128", buffer_id, (outs R600_Reg128:$dst),
1854 let MEGA_FETCH_COUNT = 16;
1859 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1861 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1862 // that holds its buffer address to avoid potential hangs. We can't use
1863 // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
1864 // registers are different sizes.
1867 //===----------------------------------------------------------------------===//
1868 // VTX Read from parameter memory space
1869 //===----------------------------------------------------------------------===//
1871 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1872 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti8 ADDRVTX_READ:$ptr))]
1875 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1876 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti16 ADDRVTX_READ:$ptr))]
1879 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1880 [(set (i32 R600_TReg32_X:$dst), (load_param ADDRVTX_READ:$ptr))]
1883 def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
1884 [(set (v4i32 R600_Reg128:$dst), (load_param ADDRVTX_READ:$ptr))]
1887 //===----------------------------------------------------------------------===//
1888 // VTX Read from global memory space
1889 //===----------------------------------------------------------------------===//
1892 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1893 [(set (i32 R600_TReg32_X:$dst), (zextloadi8_global ADDRVTX_READ:$ptr))]
1897 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1898 [(set (i32 R600_TReg32_X:$dst), (global_load ADDRVTX_READ:$ptr))]
1902 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1903 [(set (v4i32 R600_Reg128:$dst), (global_load ADDRVTX_READ:$ptr))]
1906 //===----------------------------------------------------------------------===//
1908 // XXX: We are currently storing all constants in the global address space.
1909 //===----------------------------------------------------------------------===//
1911 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1912 [(set (i32 R600_TReg32_X:$dst), (constant_load ADDRVTX_READ:$ptr))]
1917 //===----------------------------------------------------------------------===//
1918 // Regist loads and stores - for indirect addressing
1919 //===----------------------------------------------------------------------===//
1921 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1923 let Predicates = [isCayman] in {
1925 let isVector = 1 in {
1927 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1929 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1930 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1931 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1932 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1933 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1934 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1935 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
1936 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1937 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1938 def SIN_cm : SIN_Common<0x8D>;
1939 def COS_cm : COS_Common<0x8E>;
1940 } // End isVector = 1
1942 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL, R600_Reg32>;
1943 def : SIN_PAT <SIN_cm>;
1944 def : COS_PAT <COS_cm>;
1946 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1948 // RECIP_UINT emulation for Cayman
1949 // The multiplication scales from [0,1] to the unsigned integer range
1951 (AMDGPUurecip R600_Reg32:$src0),
1952 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg R600_Reg32:$src0)),
1953 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
1956 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
1962 def : Pat<(fsqrt R600_Reg32:$src),
1963 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm R600_Reg32:$src))>;
1967 //===----------------------------------------------------------------------===//
1968 // Branch Instructions
1969 //===----------------------------------------------------------------------===//
1972 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
1973 "IF_PREDICATE_SET $src", []>;
1975 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
1976 "PREDICATED_BREAK $src", []>;
1978 //===----------------------------------------------------------------------===//
1979 // Pseudo instructions
1980 //===----------------------------------------------------------------------===//
1982 let isPseudo = 1 in {
1984 def PRED_X : InstR600 <
1985 (outs R600_Predicate_Bit:$dst),
1986 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
1988 let FlagOperandIdx = 3;
1991 let isTerminator = 1, isBranch = 1 in {
1992 def JUMP_COND : InstR600 <
1994 (ins brtarget:$target, R600_Predicate_Bit:$p),
1995 "JUMP $target ($p)",
1999 def JUMP : InstR600 <
2001 (ins brtarget:$target),
2006 let isPredicable = 1;
2010 } // End isTerminator = 1, isBranch = 1
2012 let usesCustomInserter = 1 in {
2014 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
2016 def MASK_WRITE : AMDGPUShaderInst <
2018 (ins R600_Reg32:$src),
2023 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
2027 (outs R600_Reg128:$dst),
2028 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2029 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2030 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, imm:$textureTarget))], NullALU> {
2034 def TXD_SHADOW: InstR600 <
2035 (outs R600_Reg128:$dst),
2036 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2037 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2038 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))], NullALU
2042 } // End isPseudo = 1
2043 } // End usesCustomInserter = 1
2045 def CLAMP_R600 : CLAMP <R600_Reg32>;
2046 def FABS_R600 : FABS<R600_Reg32>;
2047 def FNEG_R600 : FNEG<R600_Reg32>;
2049 //===---------------------------------------------------------------------===//
2050 // Return instruction
2051 //===---------------------------------------------------------------------===//
2052 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
2053 usesCustomInserter = 1 in {
2054 def RETURN : ILFormat<(outs), (ins variable_ops),
2055 "RETURN", [(IL_retflag)]>;
2059 //===----------------------------------------------------------------------===//
2060 // Constant Buffer Addressing Support
2061 //===----------------------------------------------------------------------===//
2063 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
2064 def CONST_COPY : Instruction {
2065 let OutOperandList = (outs R600_Reg32:$dst);
2066 let InOperandList = (ins i32imm:$src);
2068 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
2069 let AsmString = "CONST_COPY";
2070 let neverHasSideEffects = 1;
2071 let isAsCheapAsAMove = 1;
2072 let Itinerary = NullALU;
2074 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
2076 def TEX_VTX_CONSTBUF :
2077 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
2078 [(set R600_Reg128:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
2079 VTX_WORD1_GPR, VTX_WORD0 {
2083 let FETCH_WHOLE_QUAD = 0;
2087 let USE_CONST_FIELDS = 0;
2088 let NUM_FORMAT_ALL = 2;
2089 let FORMAT_COMP_ALL = 1;
2090 let SRF_MODE_ALL = 1;
2091 let MEGA_FETCH_COUNT = 16;
2096 let DATA_FORMAT = 35;
2098 let Inst{31-0} = Word0;
2099 let Inst{63-32} = Word1;
2101 // LLVM can only encode 64-bit instructions, so these fields are manually
2102 // encoded in R600CodeEmitter
2105 // bits<2> ENDIAN_SWAP = 0;
2106 // bits<1> CONST_BUF_NO_STRIDE = 0;
2107 // bits<1> MEGA_FETCH = 0;
2108 // bits<1> ALT_CONST = 0;
2109 // bits<2> BUFFER_INDEX_MODE = 0;
2113 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2114 // is done in R600CodeEmitter
2116 // Inst{79-64} = OFFSET;
2117 // Inst{81-80} = ENDIAN_SWAP;
2118 // Inst{82} = CONST_BUF_NO_STRIDE;
2119 // Inst{83} = MEGA_FETCH;
2120 // Inst{84} = ALT_CONST;
2121 // Inst{86-85} = BUFFER_INDEX_MODE;
2122 // Inst{95-86} = 0; Reserved
2124 // VTX_WORD3 (Padding)
2126 // Inst{127-96} = 0;
2131 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
2132 [(set R600_Reg128:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
2133 VTX_WORD1_GPR, VTX_WORD0 {
2137 let FETCH_WHOLE_QUAD = 0;
2141 let USE_CONST_FIELDS = 1;
2142 let NUM_FORMAT_ALL = 0;
2143 let FORMAT_COMP_ALL = 0;
2144 let SRF_MODE_ALL = 1;
2145 let MEGA_FETCH_COUNT = 16;
2150 let DATA_FORMAT = 0;
2152 let Inst{31-0} = Word0;
2153 let Inst{63-32} = Word1;
2155 // LLVM can only encode 64-bit instructions, so these fields are manually
2156 // encoded in R600CodeEmitter
2159 // bits<2> ENDIAN_SWAP = 0;
2160 // bits<1> CONST_BUF_NO_STRIDE = 0;
2161 // bits<1> MEGA_FETCH = 0;
2162 // bits<1> ALT_CONST = 0;
2163 // bits<2> BUFFER_INDEX_MODE = 0;
2167 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2168 // is done in R600CodeEmitter
2170 // Inst{79-64} = OFFSET;
2171 // Inst{81-80} = ENDIAN_SWAP;
2172 // Inst{82} = CONST_BUF_NO_STRIDE;
2173 // Inst{83} = MEGA_FETCH;
2174 // Inst{84} = ALT_CONST;
2175 // Inst{86-85} = BUFFER_INDEX_MODE;
2176 // Inst{95-86} = 0; Reserved
2178 // VTX_WORD3 (Padding)
2180 // Inst{127-96} = 0;
2186 //===--------------------------------------------------------------------===//
2187 // Instructions support
2188 //===--------------------------------------------------------------------===//
2189 //===---------------------------------------------------------------------===//
2190 // Custom Inserter for Branches and returns, this eventually will be a
2192 //===---------------------------------------------------------------------===//
2193 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
2194 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
2195 "; Pseudo unconditional branch instruction",
2197 defm BRANCH_COND : BranchConditional<IL_brcond>;
2200 //===---------------------------------------------------------------------===//
2201 // Flow and Program control Instructions
2202 //===---------------------------------------------------------------------===//
2203 let isTerminator=1 in {
2204 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
2205 !strconcat("SWITCH", " $src"), []>;
2206 def CASE : ILFormat< (outs), (ins GPRI32:$src),
2207 !strconcat("CASE", " $src"), []>;
2208 def BREAK : ILFormat< (outs), (ins),
2210 def CONTINUE : ILFormat< (outs), (ins),
2212 def DEFAULT : ILFormat< (outs), (ins),
2214 def ELSE : ILFormat< (outs), (ins),
2216 def ENDSWITCH : ILFormat< (outs), (ins),
2218 def ENDMAIN : ILFormat< (outs), (ins),
2220 def END : ILFormat< (outs), (ins),
2222 def ENDFUNC : ILFormat< (outs), (ins),
2224 def ENDIF : ILFormat< (outs), (ins),
2226 def WHILELOOP : ILFormat< (outs), (ins),
2228 def ENDLOOP : ILFormat< (outs), (ins),
2230 def FUNC : ILFormat< (outs), (ins),
2232 def RETDYN : ILFormat< (outs), (ins),
2234 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2235 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
2236 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2237 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
2238 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2239 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
2240 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2241 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
2242 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2243 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
2244 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2245 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
2246 defm IFC : BranchInstr2<"IFC">;
2247 defm BREAKC : BranchInstr2<"BREAKC">;
2248 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
2251 //===----------------------------------------------------------------------===//
2253 //===----------------------------------------------------------------------===//
2255 // CND*_INT Pattterns for f32 True / False values
2257 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
2258 (selectcc (i32 R600_Reg32:$src0), 0, (f32 R600_Reg32:$src1),
2259 R600_Reg32:$src2, cc),
2260 (cnd R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2)
2263 def : CND_INT_f32 <CNDE_INT, SETEQ>;
2264 def : CND_INT_f32 <CNDGT_INT, SETGT>;
2265 def : CND_INT_f32 <CNDGE_INT, SETGE>;
2267 //CNDGE_INT extra pattern
2269 (selectcc (i32 R600_Reg32:$src0), -1, (i32 R600_Reg32:$src1),
2270 (i32 R600_Reg32:$src2), COND_GT),
2271 (CNDGE_INT R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2)
2277 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
2281 (int_AMDGPU_kill R600_Reg32:$src0),
2282 (MASK_WRITE (KILLGT (f32 ZERO), (f32 R600_Reg32:$src0)))
2287 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LT),
2288 (SGT R600_Reg32:$src1, R600_Reg32:$src0)
2293 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LE),
2294 (SGE R600_Reg32:$src1, R600_Reg32:$src0)
2297 // SETGT_DX10 reverse args
2299 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, COND_LT),
2300 (SETGT_DX10 R600_Reg32:$src1, R600_Reg32:$src0)
2303 // SETGE_DX10 reverse args
2305 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, COND_LE),
2306 (SETGE_DX10 R600_Reg32:$src1, R600_Reg32:$src0)
2309 // SETGT_INT reverse args
2311 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLT),
2312 (SETGT_INT R600_Reg32:$src1, R600_Reg32:$src0)
2315 // SETGE_INT reverse args
2317 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLE),
2318 (SETGE_INT R600_Reg32:$src1, R600_Reg32:$src0)
2321 // SETGT_UINT reverse args
2323 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULT),
2324 (SETGT_UINT R600_Reg32:$src1, R600_Reg32:$src0)
2327 // SETGE_UINT reverse args
2329 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULE),
2330 (SETGE_UINT R600_Reg32:$src1, R600_Reg32:$src0)
2333 // The next two patterns are special cases for handling 'true if ordered' and
2334 // 'true if unordered' conditionals. The assumption here is that the behavior of
2335 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
2337 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
2338 // We assume that SETE returns false when one of the operands is NAN and
2339 // SNE returns true when on of the operands is NAN
2341 //SETE - 'true if ordered'
2343 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETO),
2344 (SETE R600_Reg32:$src0, R600_Reg32:$src1)
2347 //SETE_DX10 - 'true if ordered'
2349 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETO),
2350 (SETE_DX10 R600_Reg32:$src0, R600_Reg32:$src1)
2353 //SNE - 'true if unordered'
2355 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETUO),
2356 (SNE R600_Reg32:$src0, R600_Reg32:$src1)
2359 //SETNE_DX10 - 'true if ordered'
2361 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUO),
2362 (SETNE_DX10 R600_Reg32:$src0, R600_Reg32:$src1)
2365 def : Extract_Element <f32, v4f32, R600_Reg128, 0, sub0>;
2366 def : Extract_Element <f32, v4f32, R600_Reg128, 1, sub1>;
2367 def : Extract_Element <f32, v4f32, R600_Reg128, 2, sub2>;
2368 def : Extract_Element <f32, v4f32, R600_Reg128, 3, sub3>;
2370 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 0, sub0>;
2371 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 1, sub1>;
2372 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 2, sub2>;
2373 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 3, sub3>;
2375 def : Extract_Element <i32, v4i32, R600_Reg128, 0, sub0>;
2376 def : Extract_Element <i32, v4i32, R600_Reg128, 1, sub1>;
2377 def : Extract_Element <i32, v4i32, R600_Reg128, 2, sub2>;
2378 def : Extract_Element <i32, v4i32, R600_Reg128, 3, sub3>;
2380 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 0, sub0>;
2381 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 1, sub1>;
2382 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 2, sub2>;
2383 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 3, sub3>;
2385 def : Vector4_Build <v4f32, R600_Reg128, f32, R600_Reg32>;
2386 def : Vector4_Build <v4i32, R600_Reg128, i32, R600_Reg32>;
2388 // bitconvert patterns
2390 def : BitConvert <i32, f32, R600_Reg32>;
2391 def : BitConvert <f32, i32, R600_Reg32>;
2392 def : BitConvert <v4f32, v4i32, R600_Reg128>;
2393 def : BitConvert <v4i32, v4f32, R600_Reg128>;
2395 // DWORDADDR pattern
2396 def : DwordAddrPat <i32, R600_Reg32>;
2398 } // End isR600toCayman Predicate