1 //===-- R600Instructions.td - R600 Instruction defs -------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // R600 Tablegen instruction definitions
12 //===----------------------------------------------------------------------===//
14 include "R600Intrinsics.td"
16 class InstR600 <bits<11> inst, dag outs, dag ins, string asm, list<dag> pattern,
18 : AMDGPUInst <outs, ins, asm, pattern> {
24 bits<2> FlagOperandIdx = 0;
27 bit HasNativeOperands = 0;
29 bits<11> op_code = inst;
31 let Namespace = "AMDGPU";
32 let OutOperandList = outs;
33 let InOperandList = ins;
35 let Pattern = pattern;
38 let TSFlags{4} = Trig;
41 // Vector instructions are instructions that must fill all slots in an
43 let TSFlags{6} = isVector;
44 let TSFlags{8-7} = FlagOperandIdx;
45 let TSFlags{9} = HasNativeOperands;
46 let TSFlags{10} = Op1;
47 let TSFlags{11} = Op2;
50 class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
51 AMDGPUInst <outs, ins, asm, pattern> {
54 let Namespace = "AMDGPU";
57 def MEMxi : Operand<iPTR> {
58 let MIOperandInfo = (ops R600_TReg32_X:$ptr, i32imm:$index);
59 let PrintMethod = "printMemOperand";
62 def MEMrr : Operand<iPTR> {
63 let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
66 // Operands for non-registers
68 class InstFlag<string PM = "printOperand", int Default = 0>
69 : OperandWithDefaultOps <i32, (ops (i32 Default))> {
73 // src_sel for ALU src operands, see also ALU_CONST, ALU_PARAM registers
74 def SEL : OperandWithDefaultOps <i32, (ops (i32 -1))> {
75 let PrintMethod = "printSel";
78 def LITERAL : InstFlag<"printLiteral">;
80 def WRITE : InstFlag <"printWrite", 1>;
81 def OMOD : InstFlag <"printOMOD">;
82 def REL : InstFlag <"printRel">;
83 def CLAMP : InstFlag <"printClamp">;
84 def NEG : InstFlag <"printNeg">;
85 def ABS : InstFlag <"printAbs">;
86 def UEM : InstFlag <"printUpdateExecMask">;
87 def UP : InstFlag <"printUpdatePred">;
89 // XXX: The r600g finalizer in Mesa expects last to be one in most cases.
90 // Once we start using the packetizer in this backend we should have this
92 def LAST : InstFlag<"printLast", 1>;
94 def FRAMEri : Operand<iPTR> {
95 let MIOperandInfo = (ops R600_Reg32:$ptr, i32imm:$index);
98 def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
99 def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
100 def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
101 def ADDRGA_CONST_OFFSET : ComplexPattern<i32, 1, "SelectGlobalValueConstantOffset", [], []>;
102 def ADDRGA_VAR_OFFSET : ComplexPattern<i32, 2, "SelectGlobalValueVariableOffset", [], []>;
103 def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>;
105 class R600ALU_Word0 {
106 field bits<32> Word0;
114 bits<3> index_mode = 0;
118 bits<9> src0_sel = src0{8-0};
119 bits<2> src0_chan = src0{10-9};
120 bits<9> src1_sel = src1{8-0};
121 bits<2> src1_chan = src1{10-9};
123 let Word0{8-0} = src0_sel;
124 let Word0{9} = src0_rel;
125 let Word0{11-10} = src0_chan;
126 let Word0{12} = src0_neg;
127 let Word0{21-13} = src1_sel;
128 let Word0{22} = src1_rel;
129 let Word0{24-23} = src1_chan;
130 let Word0{25} = src1_neg;
131 let Word0{28-26} = index_mode;
132 let Word0{30-29} = pred_sel;
133 let Word0{31} = last;
136 class R600ALU_Word1 {
137 field bits<32> Word1;
140 bits<3> bank_swizzle = 0;
144 bits<7> dst_sel = dst{6-0};
145 bits<2> dst_chan = dst{10-9};
147 let Word1{20-18} = bank_swizzle;
148 let Word1{27-21} = dst_sel;
149 let Word1{28} = dst_rel;
150 let Word1{30-29} = dst_chan;
151 let Word1{31} = clamp;
154 class R600ALU_Word1_OP2 <bits<11> alu_inst> : R600ALU_Word1{
158 bits<1> update_exec_mask;
163 let Word1{0} = src0_abs;
164 let Word1{1} = src1_abs;
165 let Word1{2} = update_exec_mask;
166 let Word1{3} = update_pred;
167 let Word1{4} = write;
168 let Word1{6-5} = omod;
169 let Word1{17-7} = alu_inst;
172 class R600ALU_Word1_OP3 <bits<5> alu_inst> : R600ALU_Word1{
178 bits<9> src2_sel = src2{8-0};
179 bits<2> src2_chan = src2{10-9};
181 let Word1{8-0} = src2_sel;
182 let Word1{9} = src2_rel;
183 let Word1{11-10} = src2_chan;
184 let Word1{12} = src2_neg;
185 let Word1{17-13} = alu_inst;
189 field bits<32> Word0;
193 bits<1> FETCH_WHOLE_QUAD;
197 bits<6> MEGA_FETCH_COUNT;
199 let Word0{4-0} = VC_INST;
200 let Word0{6-5} = FETCH_TYPE;
201 let Word0{7} = FETCH_WHOLE_QUAD;
202 let Word0{15-8} = BUFFER_ID;
203 let Word0{22-16} = SRC_GPR;
204 let Word0{23} = SRC_REL;
205 let Word0{25-24} = SRC_SEL_X;
206 let Word0{31-26} = MEGA_FETCH_COUNT;
209 class VTX_WORD1_GPR {
210 field bits<32> Word1;
217 bits<1> USE_CONST_FIELDS;
219 bits<2> NUM_FORMAT_ALL;
220 bits<1> FORMAT_COMP_ALL;
221 bits<1> SRF_MODE_ALL;
223 let Word1{6-0} = DST_GPR;
224 let Word1{7} = DST_REL;
225 let Word1{8} = 0; // Reserved
226 let Word1{11-9} = DST_SEL_X;
227 let Word1{14-12} = DST_SEL_Y;
228 let Word1{17-15} = DST_SEL_Z;
229 let Word1{20-18} = DST_SEL_W;
230 let Word1{21} = USE_CONST_FIELDS;
231 let Word1{27-22} = DATA_FORMAT;
232 let Word1{29-28} = NUM_FORMAT_ALL;
233 let Word1{30} = FORMAT_COMP_ALL;
234 let Word1{31} = SRF_MODE_ALL;
238 XXX: R600 subtarget uses a slightly different encoding than the other
239 subtargets. We currently handle this in R600MCCodeEmitter, but we may
240 want to use these instruction classes in the future.
242 class R600ALU_Word1_OP2_r600 : R600ALU_Word1_OP2 {
247 let Inst{37} = fog_merge;
248 let Inst{39-38} = omod;
249 let Inst{49-40} = alu_inst;
252 class R600ALU_Word1_OP2_r700 : R600ALU_Word1_OP2 {
256 let Inst{38-37} = omod;
257 let Inst{49-39} = alu_inst;
261 def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
265 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
267 // Class for instructions with only one source register.
268 // If you add new ins to this instruction, make sure they are listed before
269 // $literal, because the backend currently assumes that the last operand is
270 // a literal. Also be sure to update the enum R600Op1OperandIndex::ROI in
271 // R600Defines.h, R600InstrInfo::buildDefaultInstruction(),
272 // and R600InstrInfo::getOperandIdx().
273 class R600_1OP <bits<11> inst, string opName, list<dag> pattern,
274 InstrItinClass itin = AnyALU> :
276 (outs R600_Reg32:$dst),
277 (ins WRITE:$write, OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
278 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
279 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
281 "$clamp $dst$write$dst_rel$omod, "
282 "$src0_neg$src0_abs$src0$src0_sel$src0_abs$src0_rel, "
283 "$literal $pred_sel$last"),
287 R600ALU_Word1_OP2 <inst> {
293 let update_exec_mask = 0;
295 let HasNativeOperands = 1;
297 let DisableEncoding = "$literal";
299 let Inst{31-0} = Word0;
300 let Inst{63-32} = Word1;
303 class R600_1OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
304 InstrItinClass itin = AnyALU> :
305 R600_1OP <inst, opName,
306 [(set R600_Reg32:$dst, (node R600_Reg32:$src0))]
309 // If you add our change the operands for R600_2OP instructions, you must
310 // also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
311 // R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
312 class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
313 InstrItinClass itin = AnyALU> :
315 (outs R600_Reg32:$dst),
316 (ins UEM:$update_exec_mask, UP:$update_pred, WRITE:$write,
317 OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
318 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
319 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs, SEL:$src1_sel,
320 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
322 "$clamp $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
323 "$src0_neg$src0_abs$src0$src0_sel$src0_abs$src0_rel, "
324 "$src1_neg$src1_abs$src1$src1_sel$src1_abs$src1_rel, "
325 "$literal $pred_sel$last"),
329 R600ALU_Word1_OP2 <inst> {
331 let HasNativeOperands = 1;
333 let DisableEncoding = "$literal";
335 let Inst{31-0} = Word0;
336 let Inst{63-32} = Word1;
339 class R600_2OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
340 InstrItinClass itim = AnyALU> :
341 R600_2OP <inst, opName,
342 [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
346 // If you add our change the operands for R600_3OP instructions, you must
347 // also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
348 // R600InstrInfo::buildDefaultInstruction(), and
349 // R600InstrInfo::getOperandIdx().
350 class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
351 InstrItinClass itin = AnyALU> :
353 (outs R600_Reg32:$dst),
354 (ins REL:$dst_rel, CLAMP:$clamp,
355 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, SEL:$src0_sel,
356 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
357 R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
358 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
359 !strconcat(opName, "$clamp $dst$dst_rel, "
360 "$src0_neg$src0$src0_sel$src0_rel, "
361 "$src1_neg$src1$src1_sel$src1_rel, "
362 "$src2_neg$src2$src2_sel$src2_rel, "
363 "$literal $pred_sel$last"),
367 R600ALU_Word1_OP3<inst>{
369 let HasNativeOperands = 1;
370 let DisableEncoding = "$literal";
373 let Inst{31-0} = Word0;
374 let Inst{63-32} = Word1;
377 class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
378 InstrItinClass itin = VecALU> :
380 (outs R600_Reg32:$dst),
386 class R600_TEX <bits<11> inst, string opName, list<dag> pattern,
387 InstrItinClass itin = AnyALU> :
389 (outs R600_Reg128:$dst),
390 (ins R600_Reg128:$src0, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
391 !strconcat(opName, "$dst, $src0, $resourceId, $samplerId, $textureTarget"),
394 let Inst {10-0} = inst;
397 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
399 def TEX_SHADOW : PatLeaf<
401 [{uint32_t TType = (uint32_t)N->getZExtValue();
402 return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
406 def TEX_RECT : PatLeaf<
408 [{uint32_t TType = (uint32_t)N->getZExtValue();
413 class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, bits<4> rat_id, dag outs,
414 dag ins, string asm, list<dag> pattern> :
415 InstR600ISA <outs, ins, asm, pattern> {
432 // CF_ALLOC_EXPORT_WORD0_RAT
433 let Inst{3-0} = rat_id;
434 let Inst{9-4} = rat_inst;
435 let Inst{10} = 0; // Reserved
436 let Inst{12-11} = RIM;
437 let Inst{14-13} = TYPE;
438 let Inst{21-15} = RW_GPR;
439 let Inst{22} = RW_REL;
440 let Inst{29-23} = INDEX_GPR;
441 let Inst{31-30} = ELEM_SIZE;
443 // CF_ALLOC_EXPORT_WORD1_BUF
444 let Inst{43-32} = ARRAY_SIZE;
445 let Inst{47-44} = COMP_MASK;
446 let Inst{51-48} = BURST_COUNT;
449 let Inst{61-54} = cf_inst;
451 let Inst{63} = BARRIER;
454 class LoadParamFrag <PatFrag load_type> : PatFrag <
455 (ops node:$ptr), (load_type node:$ptr),
456 [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
459 def load_param : LoadParamFrag<load>;
460 def load_param_zexti8 : LoadParamFrag<zextloadi8>;
461 def load_param_zexti16 : LoadParamFrag<zextloadi16>;
463 def isR600 : Predicate<"Subtarget.device()"
464 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX">;
465 def isR700 : Predicate<"Subtarget.device()"
466 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX &&"
467 "Subtarget.device()->getDeviceFlag()"
468 ">= OCL_DEVICE_RV710">;
469 def isEG : Predicate<
470 "Subtarget.device()->getGeneration() >= AMDGPUDeviceInfo::HD5XXX && "
471 "Subtarget.device()->getGeneration() < AMDGPUDeviceInfo::HD7XXX && "
472 "Subtarget.device()->getDeviceFlag() != OCL_DEVICE_CAYMAN">;
474 def isCayman : Predicate<"Subtarget.device()"
475 "->getDeviceFlag() == OCL_DEVICE_CAYMAN">;
476 def isEGorCayman : Predicate<"Subtarget.device()"
477 "->getGeneration() == AMDGPUDeviceInfo::HD5XXX"
478 "|| Subtarget.device()->getGeneration() =="
479 "AMDGPUDeviceInfo::HD6XXX">;
481 def isR600toCayman : Predicate<
482 "Subtarget.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX">;
484 //===----------------------------------------------------------------------===//
486 //===----------------------------------------------------------------------===//
488 def INTERP_PAIR_XY : AMDGPUShaderInst <
489 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
490 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
491 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
494 def INTERP_PAIR_ZW : AMDGPUShaderInst <
495 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
496 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
497 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
500 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
501 SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
505 //===----------------------------------------------------------------------===//
506 // Interpolation Instructions
507 //===----------------------------------------------------------------------===//
509 def INTERP_VEC_LOAD : AMDGPUShaderInst <
510 (outs R600_Reg128:$dst),
512 "INTERP_LOAD $src0 : $dst",
515 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
516 let bank_swizzle = 5;
519 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
520 let bank_swizzle = 5;
523 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
525 //===----------------------------------------------------------------------===//
526 // Export Instructions
527 //===----------------------------------------------------------------------===//
529 def ExportType : SDTypeProfile<0, 5, [SDTCisFP<0>, SDTCisInt<1>]>;
531 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
532 [SDNPHasChain, SDNPSideEffect]>;
535 field bits<32> Word0;
542 let Word0{12-0} = arraybase;
543 let Word0{14-13} = type;
544 let Word0{21-15} = gpr;
545 let Word0{22} = 0; // RW_REL
546 let Word0{29-23} = 0; // INDEX_GPR
547 let Word0{31-30} = elem_size;
550 class ExportSwzWord1 {
551 field bits<32> Word1;
560 let Word1{2-0} = sw_x;
561 let Word1{5-3} = sw_y;
562 let Word1{8-6} = sw_z;
563 let Word1{11-9} = sw_w;
566 class ExportBufWord1 {
567 field bits<32> Word1;
574 let Word1{11-0} = arraySize;
575 let Word1{15-12} = compMask;
578 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
579 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
581 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sel_x),
582 0, 61, 0, 7, 7, 7, cf_inst, 0)
585 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
587 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sel_x),
588 0, 61, 7, 0, 7, 7, cf_inst, 0)
591 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
593 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
596 def : Pat<(int_R600_store_dummy 1),
598 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
601 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 0),
602 (i32 imm:$type), (i32 imm:$arraybase), (i32 imm)),
603 (ExportInst R600_Reg128:$src, imm:$type, imm:$arraybase,
604 0, 1, 2, 3, cf_inst, 0)
606 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 1),
607 (i32 imm:$type), (i32 imm:$arraybase), (i32 imm)),
608 (ExportInst R600_Reg128:$src, imm:$type, imm:$arraybase,
609 0, 1, 2, 3, cf_inst, 0)
612 def : Pat<(int_R600_store_swizzle (v4f32 R600_Reg128:$src), imm:$arraybase,
614 (ExportInst R600_Reg128:$src, imm:$type, imm:$arraybase,
615 0, 1, 2, 3, cf_inst, 0)
619 multiclass SteamOutputExportPattern<Instruction ExportInst,
620 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
622 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
623 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
624 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
625 4095, imm:$mask, buf0inst, 0)>;
627 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
628 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
629 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
630 4095, imm:$mask, buf1inst, 0)>;
632 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
633 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
634 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
635 4095, imm:$mask, buf2inst, 0)>;
637 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
638 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
639 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
640 4095, imm:$mask, buf3inst, 0)>;
643 let isTerminator = 1, usesCustomInserter = 1 in {
645 class ExportSwzInst : InstR600ISA<(
647 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
648 i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
650 !strconcat("EXPORT", " $gpr"),
651 []>, ExportWord0, ExportSwzWord1 {
653 let Inst{31-0} = Word0;
654 let Inst{63-32} = Word1;
657 } // End isTerminator = 1, usesCustomInserter = 1
659 class ExportBufInst : InstR600ISA<(
661 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
662 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
663 !strconcat("EXPORT", " $gpr"),
664 []>, ExportWord0, ExportBufWord1 {
666 let Inst{31-0} = Word0;
667 let Inst{63-32} = Word1;
670 let Predicates = [isR600toCayman] in {
672 //===----------------------------------------------------------------------===//
673 // Common Instructions R600, R700, Evergreen, Cayman
674 //===----------------------------------------------------------------------===//
676 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
677 // Non-IEEE MUL: 0 * anything = 0
678 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
679 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
680 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
681 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
683 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
684 // so some of the instruction names don't match the asm string.
685 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
686 def SETE : R600_2OP <
688 [(set R600_Reg32:$dst,
689 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
695 [(set R600_Reg32:$dst,
696 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
702 [(set R600_Reg32:$dst,
703 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
709 [(set R600_Reg32:$dst,
710 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
714 def SETE_DX10 : R600_2OP <
716 [(set R600_Reg32:$dst,
717 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
721 def SETGT_DX10 : R600_2OP <
723 [(set R600_Reg32:$dst,
724 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
728 def SETGE_DX10 : R600_2OP <
730 [(set R600_Reg32:$dst,
731 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
735 def SETNE_DX10 : R600_2OP <
737 [(set R600_Reg32:$dst,
738 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, (i32 -1), (i32 0),
742 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
743 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
744 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
745 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
746 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
748 def MOV : R600_1OP <0x19, "MOV", []>;
750 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
752 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
753 (outs R600_Reg32:$dst),
759 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
761 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
764 (MOV_IMM_I32 imm:$val)
767 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
770 (MOV_IMM_F32 fpimm:$val)
773 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
774 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
775 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
776 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
778 let hasSideEffects = 1 in {
780 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
782 } // end hasSideEffects
784 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
785 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
786 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
787 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
788 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
789 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
790 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
791 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
792 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
793 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
795 def SETE_INT : R600_2OP <
797 [(set (i32 R600_Reg32:$dst),
798 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETEQ))]
801 def SETGT_INT : R600_2OP <
803 [(set (i32 R600_Reg32:$dst),
804 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGT))]
807 def SETGE_INT : R600_2OP <
809 [(set (i32 R600_Reg32:$dst),
810 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGE))]
813 def SETNE_INT : R600_2OP <
815 [(set (i32 R600_Reg32:$dst),
816 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETNE))]
819 def SETGT_UINT : R600_2OP <
821 [(set (i32 R600_Reg32:$dst),
822 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGT))]
825 def SETGE_UINT : R600_2OP <
827 [(set (i32 R600_Reg32:$dst),
828 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGE))]
831 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
832 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
833 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
834 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
836 def CNDE_INT : R600_3OP <
838 [(set (i32 R600_Reg32:$dst),
839 (selectcc (i32 R600_Reg32:$src0), 0,
840 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
844 def CNDGE_INT : R600_3OP <
846 [(set (i32 R600_Reg32:$dst),
847 (selectcc (i32 R600_Reg32:$src0), 0,
848 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
852 def CNDGT_INT : R600_3OP <
854 [(set (i32 R600_Reg32:$dst),
855 (selectcc (i32 R600_Reg32:$src0), 0,
856 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
860 //===----------------------------------------------------------------------===//
861 // Texture instructions
862 //===----------------------------------------------------------------------===//
864 def TEX_LD : R600_TEX <
866 [(set R600_Reg128:$dst, (int_AMDGPU_txf R600_Reg128:$src0, imm:$src1, imm:$src2, imm:$src3, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
868 let AsmString = "TEX_LD $dst, $src0, $src1, $src2, $src3, $resourceId, $samplerId, $textureTarget";
869 let InOperandList = (ins R600_Reg128:$src0, i32imm:$src1, i32imm:$src2, i32imm:$src3, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget);
872 def TEX_GET_TEXTURE_RESINFO : R600_TEX <
873 0x04, "TEX_GET_TEXTURE_RESINFO",
874 [(set R600_Reg128:$dst, (int_AMDGPU_txq R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
877 def TEX_GET_GRADIENTS_H : R600_TEX <
878 0x07, "TEX_GET_GRADIENTS_H",
879 [(set R600_Reg128:$dst, (int_AMDGPU_ddx R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
882 def TEX_GET_GRADIENTS_V : R600_TEX <
883 0x08, "TEX_GET_GRADIENTS_V",
884 [(set R600_Reg128:$dst, (int_AMDGPU_ddy R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
887 def TEX_SET_GRADIENTS_H : R600_TEX <
888 0x0B, "TEX_SET_GRADIENTS_H",
892 def TEX_SET_GRADIENTS_V : R600_TEX <
893 0x0C, "TEX_SET_GRADIENTS_V",
897 def TEX_SAMPLE : R600_TEX <
899 [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
902 def TEX_SAMPLE_C : R600_TEX <
903 0x18, "TEX_SAMPLE_C",
904 [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
907 def TEX_SAMPLE_L : R600_TEX <
908 0x11, "TEX_SAMPLE_L",
909 [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
912 def TEX_SAMPLE_C_L : R600_TEX <
913 0x19, "TEX_SAMPLE_C_L",
914 [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
917 def TEX_SAMPLE_LB : R600_TEX <
918 0x12, "TEX_SAMPLE_LB",
919 [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0,imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
922 def TEX_SAMPLE_C_LB : R600_TEX <
923 0x1A, "TEX_SAMPLE_C_LB",
924 [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
927 def TEX_SAMPLE_G : R600_TEX <
928 0x14, "TEX_SAMPLE_G",
932 def TEX_SAMPLE_C_G : R600_TEX <
933 0x1C, "TEX_SAMPLE_C_G",
937 //===----------------------------------------------------------------------===//
938 // Helper classes for common instructions
939 //===----------------------------------------------------------------------===//
941 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
946 class MULADD_Common <bits<5> inst> : R600_3OP <
948 [(set (f32 R600_Reg32:$dst),
949 (IL_mad R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2))]
952 class CNDE_Common <bits<5> inst> : R600_3OP <
954 [(set R600_Reg32:$dst,
955 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
956 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
960 class CNDGT_Common <bits<5> inst> : R600_3OP <
962 [(set R600_Reg32:$dst,
963 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
964 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
968 class CNDGE_Common <bits<5> inst> : R600_3OP <
970 [(set R600_Reg32:$dst,
971 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
972 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
976 multiclass DOT4_Common <bits<11> inst> {
978 def _pseudo : R600_REDUCTION <inst,
979 (ins R600_Reg128:$src0, R600_Reg128:$src1),
980 "DOT4 $dst $src0, $src1",
981 [(set R600_Reg32:$dst, (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1))]
984 def _real : R600_2OP <inst, "DOT4", []>;
987 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
988 multiclass CUBE_Common <bits<11> inst> {
990 def _pseudo : InstR600 <
992 (outs R600_Reg128:$dst),
993 (ins R600_Reg128:$src),
995 [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
1001 def _real : R600_2OP <inst, "CUBE", []>;
1003 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1005 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1006 inst, "EXP_IEEE", fexp2
1009 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
1010 inst, "FLT_TO_INT", fp_to_sint
1013 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1014 inst, "INT_TO_FLT", sint_to_fp
1017 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1018 inst, "FLT_TO_UINT", fp_to_uint
1021 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1022 inst, "UINT_TO_FLT", uint_to_fp
1025 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1026 inst, "LOG_CLAMPED", []
1029 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1030 inst, "LOG_IEEE", flog2
1033 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1034 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1035 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1036 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1037 inst, "MULHI_INT", mulhs
1039 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1040 inst, "MULHI", mulhu
1042 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1043 inst, "MULLO_INT", mul
1045 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []>;
1047 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1048 inst, "RECIP_CLAMPED", []
1051 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1052 inst, "RECIP_IEEE", [(set R600_Reg32:$dst, (fdiv FP_ONE, R600_Reg32:$src0))]
1055 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1056 inst, "RECIP_UINT", AMDGPUurecip
1059 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1060 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1063 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1064 inst, "RECIPSQRT_IEEE", []
1067 class SIN_Common <bits<11> inst> : R600_1OP <
1072 class COS_Common <bits<11> inst> : R600_1OP <
1077 //===----------------------------------------------------------------------===//
1078 // Helper patterns for complex intrinsics
1079 //===----------------------------------------------------------------------===//
1081 multiclass DIV_Common <InstR600 recip_ieee> {
1083 (int_AMDGPU_div R600_Reg32:$src0, R600_Reg32:$src1),
1084 (MUL R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1088 (fdiv R600_Reg32:$src0, R600_Reg32:$src1),
1089 (MUL R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1093 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee> : Pat <
1094 (int_TGSI_lit_z R600_Reg32:$src_x, R600_Reg32:$src_y, R600_Reg32:$src_w),
1095 (exp_ieee (mul_lit (log_clamped (MAX R600_Reg32:$src_y, (f32 ZERO))), R600_Reg32:$src_w, R600_Reg32:$src_x))
1098 //===----------------------------------------------------------------------===//
1099 // R600 / R700 Instructions
1100 //===----------------------------------------------------------------------===//
1102 let Predicates = [isR600] in {
1104 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1105 def MULADD_r600 : MULADD_Common<0x10>;
1106 def CNDE_r600 : CNDE_Common<0x18>;
1107 def CNDGT_r600 : CNDGT_Common<0x19>;
1108 def CNDGE_r600 : CNDGE_Common<0x1A>;
1109 defm DOT4_r600 : DOT4_Common<0x50>;
1110 defm CUBE_r600 : CUBE_Common<0x52>;
1111 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1112 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1113 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1114 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1115 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1116 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1117 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1118 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1119 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1120 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1121 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1122 def SIN_r600 : SIN_Common<0x6E>;
1123 def COS_r600 : COS_Common<0x6F>;
1124 def ASHR_r600 : ASHR_Common<0x70>;
1125 def LSHR_r600 : LSHR_Common<0x71>;
1126 def LSHL_r600 : LSHL_Common<0x72>;
1127 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1128 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1129 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1130 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1131 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1133 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1134 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1136 def : Pat<(fsqrt R600_Reg32:$src),
1137 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_r600 R600_Reg32:$src))>;
1139 def R600_ExportSwz : ExportSwzInst {
1140 let Word1{20-17} = 1; // BURST_COUNT
1141 let Word1{21} = eop;
1142 let Word1{22} = 1; // VALID_PIXEL_MODE
1143 let Word1{30-23} = inst;
1144 let Word1{31} = 1; // BARRIER
1146 defm : ExportPattern<R600_ExportSwz, 39>;
1148 def R600_ExportBuf : ExportBufInst {
1149 let Word1{20-17} = 1; // BURST_COUNT
1150 let Word1{21} = eop;
1151 let Word1{22} = 1; // VALID_PIXEL_MODE
1152 let Word1{30-23} = inst;
1153 let Word1{31} = 1; // BARRIER
1155 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1158 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1160 class COS_PAT <InstR600 trig> : Pat<
1161 (fcos R600_Reg32:$src),
1162 (trig (MUL (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1165 class SIN_PAT <InstR600 trig> : Pat<
1166 (fsin R600_Reg32:$src),
1167 (trig (MUL (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1170 //===----------------------------------------------------------------------===//
1171 // R700 Only instructions
1172 //===----------------------------------------------------------------------===//
1174 let Predicates = [isR700] in {
1175 def SIN_r700 : SIN_Common<0x6E>;
1176 def COS_r700 : COS_Common<0x6F>;
1178 // R700 normalizes inputs to SIN/COS the same as EG
1179 def : SIN_PAT <SIN_r700>;
1180 def : COS_PAT <COS_r700>;
1183 //===----------------------------------------------------------------------===//
1184 // Evergreen Only instructions
1185 //===----------------------------------------------------------------------===//
1187 let Predicates = [isEG] in {
1189 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1190 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1192 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1193 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1194 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1195 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1196 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1197 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1198 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1199 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1200 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1201 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1202 def SIN_eg : SIN_Common<0x8D>;
1203 def COS_eg : COS_Common<0x8E>;
1205 def : SIN_PAT <SIN_eg>;
1206 def : COS_PAT <COS_eg>;
1207 def : Pat<(fsqrt R600_Reg32:$src),
1208 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_eg R600_Reg32:$src))>;
1209 } // End Predicates = [isEG]
1211 //===----------------------------------------------------------------------===//
1212 // Evergreen / Cayman Instructions
1213 //===----------------------------------------------------------------------===//
1215 let Predicates = [isEGorCayman] in {
1217 // BFE_UINT - bit_extract, an optimization for mask and shift
1222 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1227 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1228 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1229 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1230 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1231 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1232 [(set R600_Reg32:$dst, (int_AMDIL_bit_extract_u32 R600_Reg32:$src0,
1234 R600_Reg32:$src2))],
1238 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
1239 [(set R600_Reg32:$dst, (AMDGPUbitalign R600_Reg32:$src0, R600_Reg32:$src1,
1240 R600_Reg32:$src2))],
1244 def MULADD_eg : MULADD_Common<0x14>;
1245 def ASHR_eg : ASHR_Common<0x15>;
1246 def LSHR_eg : LSHR_Common<0x16>;
1247 def LSHL_eg : LSHL_Common<0x17>;
1248 def CNDE_eg : CNDE_Common<0x19>;
1249 def CNDGT_eg : CNDGT_Common<0x1A>;
1250 def CNDGE_eg : CNDGE_Common<0x1B>;
1251 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1252 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1253 defm DOT4_eg : DOT4_Common<0xBE>;
1254 defm CUBE_eg : CUBE_Common<0xC0>;
1256 let hasSideEffects = 1 in {
1257 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1260 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1262 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1266 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1268 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1272 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1274 // TRUNC is used for the FLT_TO_INT instructions to work around a
1275 // perceived problem where the rounding modes are applied differently
1276 // depending on the instruction and the slot they are in.
1278 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1279 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1281 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1282 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1283 // We should look into handling these cases separately.
1284 def : Pat<(fp_to_sint R600_Reg32:$src0),
1285 (FLT_TO_INT_eg (TRUNC R600_Reg32:$src0))>;
1287 def : Pat<(fp_to_uint R600_Reg32:$src0),
1288 (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src0))>;
1290 def EG_ExportSwz : ExportSwzInst {
1291 let Word1{19-16} = 1; // BURST_COUNT
1292 let Word1{20} = 1; // VALID_PIXEL_MODE
1293 let Word1{21} = eop;
1294 let Word1{29-22} = inst;
1295 let Word1{30} = 0; // MARK
1296 let Word1{31} = 1; // BARRIER
1298 defm : ExportPattern<EG_ExportSwz, 83>;
1300 def EG_ExportBuf : ExportBufInst {
1301 let Word1{19-16} = 1; // BURST_COUNT
1302 let Word1{20} = 1; // VALID_PIXEL_MODE
1303 let Word1{21} = eop;
1304 let Word1{29-22} = inst;
1305 let Word1{30} = 0; // MARK
1306 let Word1{31} = 1; // BARRIER
1308 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1310 //===----------------------------------------------------------------------===//
1311 // Memory read/write instructions
1312 //===----------------------------------------------------------------------===//
1313 let usesCustomInserter = 1 in {
1315 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
1317 : EG_CF_RAT <0x57, 0x2, 0, (outs), ins,
1318 !strconcat(name, " $rw_gpr, $index_gpr, $eop"), pattern> {
1320 // XXX: Have a separate instruction for non-indexed writes.
1326 let COMP_MASK = comp_mask;
1327 let BURST_COUNT = 0;
1333 } // End usesCustomInserter = 1
1336 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1337 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1338 0x1, "RAT_WRITE_CACHELESS_32_eg",
1339 [(global_store (i32 R600_TReg32_X:$rw_gpr), R600_TReg32_X:$index_gpr)]
1343 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1344 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1345 0xf, "RAT_WRITE_CACHELESS_128",
1346 [(global_store (v4i32 R600_Reg128:$rw_gpr), R600_TReg32_X:$index_gpr)]
1349 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1350 : InstR600ISA <outs, (ins MEMxi:$ptr), name#" $dst, $ptr", pattern>,
1351 VTX_WORD1_GPR, VTX_WORD0 {
1356 let FETCH_WHOLE_QUAD = 0;
1357 let BUFFER_ID = buffer_id;
1359 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1360 // to store vertex addresses in any channel, not just X.
1363 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
1364 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
1365 // however, based on my testing if USE_CONST_FIELDS is set, then all
1366 // these fields need to be set to 0.
1367 let USE_CONST_FIELDS = 0;
1368 let NUM_FORMAT_ALL = 1;
1369 let FORMAT_COMP_ALL = 0;
1370 let SRF_MODE_ALL = 0;
1372 let Inst{31-0} = Word0;
1373 let Inst{63-32} = Word1;
1374 // LLVM can only encode 64-bit instructions, so these fields are manually
1375 // encoded in R600CodeEmitter
1378 // bits<2> ENDIAN_SWAP = 0;
1379 // bits<1> CONST_BUF_NO_STRIDE = 0;
1380 // bits<1> MEGA_FETCH = 0;
1381 // bits<1> ALT_CONST = 0;
1382 // bits<2> BUFFER_INDEX_MODE = 0;
1386 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1387 // is done in R600CodeEmitter
1389 // Inst{79-64} = OFFSET;
1390 // Inst{81-80} = ENDIAN_SWAP;
1391 // Inst{82} = CONST_BUF_NO_STRIDE;
1392 // Inst{83} = MEGA_FETCH;
1393 // Inst{84} = ALT_CONST;
1394 // Inst{86-85} = BUFFER_INDEX_MODE;
1395 // Inst{95-86} = 0; Reserved
1397 // VTX_WORD3 (Padding)
1399 // Inst{127-96} = 0;
1402 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1403 : VTX_READ_eg <"VTX_READ_8", buffer_id, (outs R600_TReg32_X:$dst),
1406 let MEGA_FETCH_COUNT = 1;
1408 let DST_SEL_Y = 7; // Masked
1409 let DST_SEL_Z = 7; // Masked
1410 let DST_SEL_W = 7; // Masked
1411 let DATA_FORMAT = 1; // FMT_8
1414 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1415 : VTX_READ_eg <"VTX_READ_16", buffer_id, (outs R600_TReg32_X:$dst),
1417 let MEGA_FETCH_COUNT = 2;
1419 let DST_SEL_Y = 7; // Masked
1420 let DST_SEL_Z = 7; // Masked
1421 let DST_SEL_W = 7; // Masked
1422 let DATA_FORMAT = 5; // FMT_16
1426 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1427 : VTX_READ_eg <"VTX_READ_32", buffer_id, (outs R600_TReg32_X:$dst),
1430 let MEGA_FETCH_COUNT = 4;
1432 let DST_SEL_Y = 7; // Masked
1433 let DST_SEL_Z = 7; // Masked
1434 let DST_SEL_W = 7; // Masked
1435 let DATA_FORMAT = 0xD; // COLOR_32
1437 // This is not really necessary, but there were some GPU hangs that appeared
1438 // to be caused by ALU instructions in the next instruction group that wrote
1439 // to the $ptr registers of the VTX_READ.
1441 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1442 // %T2_X<def> = MOV %ZERO
1443 //Adding this constraint prevents this from happening.
1444 let Constraints = "$ptr.ptr = $dst";
1447 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1448 : VTX_READ_eg <"VTX_READ_128", buffer_id, (outs R600_Reg128:$dst),
1451 let MEGA_FETCH_COUNT = 16;
1456 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1458 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1459 // that holds its buffer address to avoid potential hangs. We can't use
1460 // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
1461 // registers are different sizes.
1464 //===----------------------------------------------------------------------===//
1465 // VTX Read from parameter memory space
1466 //===----------------------------------------------------------------------===//
1468 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1469 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti8 ADDRVTX_READ:$ptr))]
1472 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1473 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti16 ADDRVTX_READ:$ptr))]
1476 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1477 [(set (i32 R600_TReg32_X:$dst), (load_param ADDRVTX_READ:$ptr))]
1480 //===----------------------------------------------------------------------===//
1481 // VTX Read from global memory space
1482 //===----------------------------------------------------------------------===//
1485 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1486 [(set (i32 R600_TReg32_X:$dst), (zextloadi8_global ADDRVTX_READ:$ptr))]
1490 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1491 [(set (i32 R600_TReg32_X:$dst), (global_load ADDRVTX_READ:$ptr))]
1495 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1496 [(set (v4i32 R600_Reg128:$dst), (global_load ADDRVTX_READ:$ptr))]
1499 //===----------------------------------------------------------------------===//
1501 // XXX: We are currently storing all constants in the global address space.
1502 //===----------------------------------------------------------------------===//
1504 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1505 [(set (i32 R600_TReg32_X:$dst), (constant_load ADDRVTX_READ:$ptr))]
1510 //===----------------------------------------------------------------------===//
1511 // Regist loads and stores - for indirect addressing
1512 //===----------------------------------------------------------------------===//
1514 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1516 let Predicates = [isCayman] in {
1518 let isVector = 1 in {
1520 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1522 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1523 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1524 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1525 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1526 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1527 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1528 def LOG_IEEE_ : LOG_IEEE_Common<0x83>;
1529 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1530 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1531 def SIN_cm : SIN_Common<0x8D>;
1532 def COS_cm : COS_Common<0x8E>;
1533 } // End isVector = 1
1535 def : SIN_PAT <SIN_cm>;
1536 def : COS_PAT <COS_cm>;
1538 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1540 // RECIP_UINT emulation for Cayman
1542 (AMDGPUurecip R600_Reg32:$src0),
1543 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg R600_Reg32:$src0)),
1544 (MOV_IMM_I32 0x4f800000)))
1548 def : Pat<(fsqrt R600_Reg32:$src),
1549 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm R600_Reg32:$src))>;
1553 //===----------------------------------------------------------------------===//
1554 // Branch Instructions
1555 //===----------------------------------------------------------------------===//
1558 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
1559 "IF_PREDICATE_SET $src", []>;
1561 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
1562 "PREDICATED_BREAK $src", []>;
1564 //===----------------------------------------------------------------------===//
1565 // Pseudo instructions
1566 //===----------------------------------------------------------------------===//
1568 let isPseudo = 1 in {
1570 def PRED_X : InstR600 <
1571 0, (outs R600_Predicate_Bit:$dst),
1572 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
1574 let FlagOperandIdx = 3;
1577 let isTerminator = 1, isBranch = 1, isBarrier = 1 in {
1579 def JUMP : InstR600 <0x10,
1581 (ins brtarget:$target, R600_Pred:$p),
1582 "JUMP $target ($p)",
1586 } // End isTerminator = 1, isBranch = 1, isBarrier = 1
1588 let usesCustomInserter = 1 in {
1590 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
1592 def MASK_WRITE : AMDGPUShaderInst <
1594 (ins R600_Reg32:$src),
1599 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
1602 def TXD: AMDGPUShaderInst <
1603 (outs R600_Reg128:$dst),
1604 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1605 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1606 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
1609 def TXD_SHADOW: AMDGPUShaderInst <
1610 (outs R600_Reg128:$dst),
1611 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1612 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1613 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
1616 } // End isPseudo = 1
1617 } // End usesCustomInserter = 1
1619 def CLAMP_R600 : CLAMP <R600_Reg32>;
1620 def FABS_R600 : FABS<R600_Reg32>;
1621 def FNEG_R600 : FNEG<R600_Reg32>;
1623 //===---------------------------------------------------------------------===//
1624 // Return instruction
1625 //===---------------------------------------------------------------------===//
1626 let isTerminator = 1, isReturn = 1, isBarrier = 1, hasCtrlDep = 1,
1627 usesCustomInserter = 1 in {
1628 def RETURN : ILFormat<(outs), (ins variable_ops),
1629 "RETURN", [(IL_retflag)]>;
1633 //===----------------------------------------------------------------------===//
1634 // Constant Buffer Addressing Support
1635 //===----------------------------------------------------------------------===//
1637 let isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
1638 def CONST_COPY : Instruction {
1639 let OutOperandList = (outs R600_Reg32:$dst);
1640 let InOperandList = (ins i32imm:$src);
1641 let Pattern = [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
1642 let AsmString = "CONST_COPY";
1643 let neverHasSideEffects = 1;
1644 let isAsCheapAsAMove = 1;
1645 let Itinerary = NullALU;
1647 } // end isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
1649 def TEX_VTX_CONSTBUF :
1650 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr), "VTX_READ_eg $dst, $ptr",
1651 [(set R600_Reg128:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr))]>,
1652 VTX_WORD1_GPR, VTX_WORD0 {
1656 let FETCH_WHOLE_QUAD = 0;
1661 let USE_CONST_FIELDS = 0;
1662 let NUM_FORMAT_ALL = 2;
1663 let FORMAT_COMP_ALL = 1;
1664 let SRF_MODE_ALL = 1;
1665 let MEGA_FETCH_COUNT = 16;
1670 let DATA_FORMAT = 35;
1672 let Inst{31-0} = Word0;
1673 let Inst{63-32} = Word1;
1675 // LLVM can only encode 64-bit instructions, so these fields are manually
1676 // encoded in R600CodeEmitter
1679 // bits<2> ENDIAN_SWAP = 0;
1680 // bits<1> CONST_BUF_NO_STRIDE = 0;
1681 // bits<1> MEGA_FETCH = 0;
1682 // bits<1> ALT_CONST = 0;
1683 // bits<2> BUFFER_INDEX_MODE = 0;
1687 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1688 // is done in R600CodeEmitter
1690 // Inst{79-64} = OFFSET;
1691 // Inst{81-80} = ENDIAN_SWAP;
1692 // Inst{82} = CONST_BUF_NO_STRIDE;
1693 // Inst{83} = MEGA_FETCH;
1694 // Inst{84} = ALT_CONST;
1695 // Inst{86-85} = BUFFER_INDEX_MODE;
1696 // Inst{95-86} = 0; Reserved
1698 // VTX_WORD3 (Padding)
1700 // Inst{127-96} = 0;
1704 //===--------------------------------------------------------------------===//
1705 // Instructions support
1706 //===--------------------------------------------------------------------===//
1707 //===---------------------------------------------------------------------===//
1708 // Custom Inserter for Branches and returns, this eventually will be a
1710 //===---------------------------------------------------------------------===//
1711 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
1712 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
1713 "; Pseudo unconditional branch instruction",
1715 defm BRANCH_COND : BranchConditional<IL_brcond>;
1718 //===---------------------------------------------------------------------===//
1719 // Flow and Program control Instructions
1720 //===---------------------------------------------------------------------===//
1721 let isTerminator=1 in {
1722 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
1723 !strconcat("SWITCH", " $src"), []>;
1724 def CASE : ILFormat< (outs), (ins GPRI32:$src),
1725 !strconcat("CASE", " $src"), []>;
1726 def BREAK : ILFormat< (outs), (ins),
1728 def CONTINUE : ILFormat< (outs), (ins),
1730 def DEFAULT : ILFormat< (outs), (ins),
1732 def ELSE : ILFormat< (outs), (ins),
1734 def ENDSWITCH : ILFormat< (outs), (ins),
1736 def ENDMAIN : ILFormat< (outs), (ins),
1738 def END : ILFormat< (outs), (ins),
1740 def ENDFUNC : ILFormat< (outs), (ins),
1742 def ENDIF : ILFormat< (outs), (ins),
1744 def WHILELOOP : ILFormat< (outs), (ins),
1746 def ENDLOOP : ILFormat< (outs), (ins),
1748 def FUNC : ILFormat< (outs), (ins),
1750 def RETDYN : ILFormat< (outs), (ins),
1752 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1753 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
1754 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1755 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
1756 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1757 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
1758 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1759 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
1760 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1761 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
1762 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1763 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
1764 defm IFC : BranchInstr2<"IFC">;
1765 defm BREAKC : BranchInstr2<"BREAKC">;
1766 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
1769 //===----------------------------------------------------------------------===//
1771 //===----------------------------------------------------------------------===//
1773 //CNDGE_INT extra pattern
1775 (selectcc (i32 R600_Reg32:$src0), -1, (i32 R600_Reg32:$src1),
1776 (i32 R600_Reg32:$src2), COND_GT),
1777 (CNDGE_INT R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2)
1783 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
1787 (int_AMDGPU_kill R600_Reg32:$src0),
1788 (MASK_WRITE (KILLGT (f32 ZERO), (f32 R600_Reg32:$src0)))
1793 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LT),
1794 (SGT R600_Reg32:$src1, R600_Reg32:$src0)
1799 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LE),
1800 (SGE R600_Reg32:$src1, R600_Reg32:$src0)
1803 // SETGT_DX10 reverse args
1805 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, COND_LT),
1806 (SETGT_DX10 R600_Reg32:$src1, R600_Reg32:$src0)
1809 // SETGE_DX10 reverse args
1811 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, COND_LE),
1812 (SETGE_DX10 R600_Reg32:$src1, R600_Reg32:$src0)
1815 // SETGT_INT reverse args
1817 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLT),
1818 (SETGT_INT R600_Reg32:$src1, R600_Reg32:$src0)
1821 // SETGE_INT reverse args
1823 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLE),
1824 (SETGE_INT R600_Reg32:$src1, R600_Reg32:$src0)
1827 // SETGT_UINT reverse args
1829 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULT),
1830 (SETGT_UINT R600_Reg32:$src1, R600_Reg32:$src0)
1833 // SETGE_UINT reverse args
1835 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULE),
1836 (SETGE_UINT R600_Reg32:$src1, R600_Reg32:$src0)
1839 // The next two patterns are special cases for handling 'true if ordered' and
1840 // 'true if unordered' conditionals. The assumption here is that the behavior of
1841 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
1843 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
1844 // We assume that SETE returns false when one of the operands is NAN and
1845 // SNE returns true when on of the operands is NAN
1847 //SETE - 'true if ordered'
1849 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETO),
1850 (SETE R600_Reg32:$src0, R600_Reg32:$src1)
1853 //SETE_DX10 - 'true if ordered'
1855 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETO),
1856 (SETE_DX10 R600_Reg32:$src0, R600_Reg32:$src1)
1859 //SNE - 'true if unordered'
1861 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETUO),
1862 (SNE R600_Reg32:$src0, R600_Reg32:$src1)
1865 //SETNE_DX10 - 'true if ordered'
1867 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUO),
1868 (SETNE_DX10 R600_Reg32:$src0, R600_Reg32:$src1)
1871 def : Extract_Element <f32, v4f32, R600_Reg128, 0, sel_x>;
1872 def : Extract_Element <f32, v4f32, R600_Reg128, 1, sel_y>;
1873 def : Extract_Element <f32, v4f32, R600_Reg128, 2, sel_z>;
1874 def : Extract_Element <f32, v4f32, R600_Reg128, 3, sel_w>;
1876 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 0, sel_x>;
1877 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 1, sel_y>;
1878 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 2, sel_z>;
1879 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 3, sel_w>;
1881 def : Extract_Element <i32, v4i32, R600_Reg128, 0, sel_x>;
1882 def : Extract_Element <i32, v4i32, R600_Reg128, 1, sel_y>;
1883 def : Extract_Element <i32, v4i32, R600_Reg128, 2, sel_z>;
1884 def : Extract_Element <i32, v4i32, R600_Reg128, 3, sel_w>;
1886 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 0, sel_x>;
1887 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 1, sel_y>;
1888 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 2, sel_z>;
1889 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 3, sel_w>;
1891 def : Vector_Build <v4f32, R600_Reg128, f32, R600_Reg32>;
1892 def : Vector_Build <v4i32, R600_Reg128, i32, R600_Reg32>;
1894 // bitconvert patterns
1896 def : BitConvert <i32, f32, R600_Reg32>;
1897 def : BitConvert <f32, i32, R600_Reg32>;
1898 def : BitConvert <v4f32, v4i32, R600_Reg128>;
1899 def : BitConvert <v4i32, v4f32, R600_Reg128>;
1901 // DWORDADDR pattern
1902 def : DwordAddrPat <i32, R600_Reg32>;
1904 } // End isR600toCayman Predicate