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:$src1, FP_ONE, FP_ZERO,
714 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
715 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
716 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
717 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
718 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
720 def MOV : R600_1OP <0x19, "MOV", []>;
722 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
724 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
725 (outs R600_Reg32:$dst),
731 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
733 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
736 (MOV_IMM_I32 imm:$val)
739 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
742 (MOV_IMM_F32 fpimm:$val)
745 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
746 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
747 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
748 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
750 let hasSideEffects = 1 in {
752 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
754 } // end hasSideEffects
756 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
757 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
758 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
759 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
760 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
761 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
762 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
763 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
764 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
765 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
767 def SETE_INT : R600_2OP <
769 [(set (i32 R600_Reg32:$dst),
770 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETEQ))]
773 def SETGT_INT : R600_2OP <
775 [(set (i32 R600_Reg32:$dst),
776 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGT))]
779 def SETGE_INT : R600_2OP <
781 [(set (i32 R600_Reg32:$dst),
782 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGE))]
785 def SETNE_INT : R600_2OP <
787 [(set (i32 R600_Reg32:$dst),
788 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETNE))]
791 def SETGT_UINT : R600_2OP <
793 [(set (i32 R600_Reg32:$dst),
794 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGT))]
797 def SETGE_UINT : R600_2OP <
799 [(set (i32 R600_Reg32:$dst),
800 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGE))]
803 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
804 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
805 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
806 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
808 def CNDE_INT : R600_3OP <
810 [(set (i32 R600_Reg32:$dst),
811 (selectcc (i32 R600_Reg32:$src0), 0,
812 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
816 def CNDGE_INT : R600_3OP <
818 [(set (i32 R600_Reg32:$dst),
819 (selectcc (i32 R600_Reg32:$src0), 0,
820 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
824 def CNDGT_INT : R600_3OP <
826 [(set (i32 R600_Reg32:$dst),
827 (selectcc (i32 R600_Reg32:$src0), 0,
828 (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
832 //===----------------------------------------------------------------------===//
833 // Texture instructions
834 //===----------------------------------------------------------------------===//
836 def TEX_LD : R600_TEX <
838 [(set R600_Reg128:$dst, (int_AMDGPU_txf R600_Reg128:$src0, imm:$src1, imm:$src2, imm:$src3, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
840 let AsmString = "TEX_LD $dst, $src0, $src1, $src2, $src3, $resourceId, $samplerId, $textureTarget";
841 let InOperandList = (ins R600_Reg128:$src0, i32imm:$src1, i32imm:$src2, i32imm:$src3, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget);
844 def TEX_GET_TEXTURE_RESINFO : R600_TEX <
845 0x04, "TEX_GET_TEXTURE_RESINFO",
846 [(set R600_Reg128:$dst, (int_AMDGPU_txq R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
849 def TEX_GET_GRADIENTS_H : R600_TEX <
850 0x07, "TEX_GET_GRADIENTS_H",
851 [(set R600_Reg128:$dst, (int_AMDGPU_ddx R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
854 def TEX_GET_GRADIENTS_V : R600_TEX <
855 0x08, "TEX_GET_GRADIENTS_V",
856 [(set R600_Reg128:$dst, (int_AMDGPU_ddy R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
859 def TEX_SET_GRADIENTS_H : R600_TEX <
860 0x0B, "TEX_SET_GRADIENTS_H",
864 def TEX_SET_GRADIENTS_V : R600_TEX <
865 0x0C, "TEX_SET_GRADIENTS_V",
869 def TEX_SAMPLE : R600_TEX <
871 [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
874 def TEX_SAMPLE_C : R600_TEX <
875 0x18, "TEX_SAMPLE_C",
876 [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
879 def TEX_SAMPLE_L : R600_TEX <
880 0x11, "TEX_SAMPLE_L",
881 [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
884 def TEX_SAMPLE_C_L : R600_TEX <
885 0x19, "TEX_SAMPLE_C_L",
886 [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
889 def TEX_SAMPLE_LB : R600_TEX <
890 0x12, "TEX_SAMPLE_LB",
891 [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0,imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
894 def TEX_SAMPLE_C_LB : R600_TEX <
895 0x1A, "TEX_SAMPLE_C_LB",
896 [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
899 def TEX_SAMPLE_G : R600_TEX <
900 0x14, "TEX_SAMPLE_G",
904 def TEX_SAMPLE_C_G : R600_TEX <
905 0x1C, "TEX_SAMPLE_C_G",
909 //===----------------------------------------------------------------------===//
910 // Helper classes for common instructions
911 //===----------------------------------------------------------------------===//
913 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
918 class MULADD_Common <bits<5> inst> : R600_3OP <
920 [(set (f32 R600_Reg32:$dst),
921 (IL_mad R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2))]
924 class CNDE_Common <bits<5> inst> : R600_3OP <
926 [(set R600_Reg32:$dst,
927 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
928 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
932 class CNDGT_Common <bits<5> inst> : R600_3OP <
934 [(set R600_Reg32:$dst,
935 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
936 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
940 class CNDGE_Common <bits<5> inst> : R600_3OP <
942 [(set R600_Reg32:$dst,
943 (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
944 (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
948 multiclass DOT4_Common <bits<11> inst> {
950 def _pseudo : R600_REDUCTION <inst,
951 (ins R600_Reg128:$src0, R600_Reg128:$src1),
952 "DOT4 $dst $src0, $src1",
953 [(set R600_Reg32:$dst, (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1))]
956 def _real : R600_2OP <inst, "DOT4", []>;
959 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
960 multiclass CUBE_Common <bits<11> inst> {
962 def _pseudo : InstR600 <
964 (outs R600_Reg128:$dst),
965 (ins R600_Reg128:$src),
967 [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
973 def _real : R600_2OP <inst, "CUBE", []>;
975 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
977 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
978 inst, "EXP_IEEE", fexp2
981 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
982 inst, "FLT_TO_INT", fp_to_sint
985 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
986 inst, "INT_TO_FLT", sint_to_fp
989 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
990 inst, "FLT_TO_UINT", fp_to_uint
993 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
994 inst, "UINT_TO_FLT", uint_to_fp
997 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
998 inst, "LOG_CLAMPED", []
1001 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1002 inst, "LOG_IEEE", flog2
1005 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1006 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1007 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1008 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1009 inst, "MULHI_INT", mulhs
1011 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1012 inst, "MULHI", mulhu
1014 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1015 inst, "MULLO_INT", mul
1017 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []>;
1019 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1020 inst, "RECIP_CLAMPED", []
1023 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1024 inst, "RECIP_IEEE", [(set R600_Reg32:$dst, (fdiv FP_ONE, R600_Reg32:$src0))]
1027 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1028 inst, "RECIP_UINT", AMDGPUurecip
1031 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1032 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1035 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1036 inst, "RECIPSQRT_IEEE", []
1039 class SIN_Common <bits<11> inst> : R600_1OP <
1044 class COS_Common <bits<11> inst> : R600_1OP <
1049 //===----------------------------------------------------------------------===//
1050 // Helper patterns for complex intrinsics
1051 //===----------------------------------------------------------------------===//
1053 multiclass DIV_Common <InstR600 recip_ieee> {
1055 (int_AMDGPU_div R600_Reg32:$src0, R600_Reg32:$src1),
1056 (MUL R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1060 (fdiv R600_Reg32:$src0, R600_Reg32:$src1),
1061 (MUL R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
1065 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee> : Pat <
1066 (int_TGSI_lit_z R600_Reg32:$src_x, R600_Reg32:$src_y, R600_Reg32:$src_w),
1067 (exp_ieee (mul_lit (log_clamped (MAX R600_Reg32:$src_y, (f32 ZERO))), R600_Reg32:$src_w, R600_Reg32:$src_x))
1070 //===----------------------------------------------------------------------===//
1071 // R600 / R700 Instructions
1072 //===----------------------------------------------------------------------===//
1074 let Predicates = [isR600] in {
1076 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1077 def MULADD_r600 : MULADD_Common<0x10>;
1078 def CNDE_r600 : CNDE_Common<0x18>;
1079 def CNDGT_r600 : CNDGT_Common<0x19>;
1080 def CNDGE_r600 : CNDGE_Common<0x1A>;
1081 defm DOT4_r600 : DOT4_Common<0x50>;
1082 defm CUBE_r600 : CUBE_Common<0x52>;
1083 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1084 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1085 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1086 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1087 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1088 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1089 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1090 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1091 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1092 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1093 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1094 def SIN_r600 : SIN_Common<0x6E>;
1095 def COS_r600 : COS_Common<0x6F>;
1096 def ASHR_r600 : ASHR_Common<0x70>;
1097 def LSHR_r600 : LSHR_Common<0x71>;
1098 def LSHL_r600 : LSHL_Common<0x72>;
1099 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1100 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1101 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1102 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1103 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1105 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1106 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1108 def : Pat<(fsqrt R600_Reg32:$src),
1109 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_r600 R600_Reg32:$src))>;
1111 def R600_ExportSwz : ExportSwzInst {
1112 let Word1{20-17} = 1; // BURST_COUNT
1113 let Word1{21} = eop;
1114 let Word1{22} = 1; // VALID_PIXEL_MODE
1115 let Word1{30-23} = inst;
1116 let Word1{31} = 1; // BARRIER
1118 defm : ExportPattern<R600_ExportSwz, 39>;
1120 def R600_ExportBuf : ExportBufInst {
1121 let Word1{20-17} = 1; // BURST_COUNT
1122 let Word1{21} = eop;
1123 let Word1{22} = 1; // VALID_PIXEL_MODE
1124 let Word1{30-23} = inst;
1125 let Word1{31} = 1; // BARRIER
1127 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1130 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1132 class COS_PAT <InstR600 trig> : Pat<
1133 (fcos R600_Reg32:$src),
1134 (trig (MUL (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1137 class SIN_PAT <InstR600 trig> : Pat<
1138 (fsin R600_Reg32:$src),
1139 (trig (MUL (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
1142 //===----------------------------------------------------------------------===//
1143 // R700 Only instructions
1144 //===----------------------------------------------------------------------===//
1146 let Predicates = [isR700] in {
1147 def SIN_r700 : SIN_Common<0x6E>;
1148 def COS_r700 : COS_Common<0x6F>;
1150 // R700 normalizes inputs to SIN/COS the same as EG
1151 def : SIN_PAT <SIN_r700>;
1152 def : COS_PAT <COS_r700>;
1155 //===----------------------------------------------------------------------===//
1156 // Evergreen Only instructions
1157 //===----------------------------------------------------------------------===//
1159 let Predicates = [isEG] in {
1161 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1162 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1164 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1165 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1166 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1167 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1168 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1169 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1170 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1171 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1172 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1173 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1174 def SIN_eg : SIN_Common<0x8D>;
1175 def COS_eg : COS_Common<0x8E>;
1177 def : SIN_PAT <SIN_eg>;
1178 def : COS_PAT <COS_eg>;
1179 def : Pat<(fsqrt R600_Reg32:$src),
1180 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_eg R600_Reg32:$src))>;
1181 } // End Predicates = [isEG]
1183 //===----------------------------------------------------------------------===//
1184 // Evergreen / Cayman Instructions
1185 //===----------------------------------------------------------------------===//
1187 let Predicates = [isEGorCayman] in {
1189 // BFE_UINT - bit_extract, an optimization for mask and shift
1194 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1199 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1200 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1201 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1202 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1203 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1204 [(set R600_Reg32:$dst, (int_AMDIL_bit_extract_u32 R600_Reg32:$src0,
1206 R600_Reg32:$src2))],
1210 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
1211 [(set R600_Reg32:$dst, (AMDGPUbitalign R600_Reg32:$src0, R600_Reg32:$src1,
1212 R600_Reg32:$src2))],
1216 def MULADD_eg : MULADD_Common<0x14>;
1217 def ASHR_eg : ASHR_Common<0x15>;
1218 def LSHR_eg : LSHR_Common<0x16>;
1219 def LSHL_eg : LSHL_Common<0x17>;
1220 def CNDE_eg : CNDE_Common<0x19>;
1221 def CNDGT_eg : CNDGT_Common<0x1A>;
1222 def CNDGE_eg : CNDGE_Common<0x1B>;
1223 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1224 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1225 defm DOT4_eg : DOT4_Common<0xBE>;
1226 defm CUBE_eg : CUBE_Common<0xC0>;
1228 let hasSideEffects = 1 in {
1229 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1232 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1234 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1238 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1240 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1244 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1246 // TRUNC is used for the FLT_TO_INT instructions to work around a
1247 // perceived problem where the rounding modes are applied differently
1248 // depending on the instruction and the slot they are in.
1250 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1251 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1253 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1254 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1255 // We should look into handling these cases separately.
1256 def : Pat<(fp_to_sint R600_Reg32:$src0),
1257 (FLT_TO_INT_eg (TRUNC R600_Reg32:$src0))>;
1259 def : Pat<(fp_to_uint R600_Reg32:$src0),
1260 (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src0))>;
1262 def EG_ExportSwz : ExportSwzInst {
1263 let Word1{19-16} = 1; // BURST_COUNT
1264 let Word1{20} = 1; // VALID_PIXEL_MODE
1265 let Word1{21} = eop;
1266 let Word1{29-22} = inst;
1267 let Word1{30} = 0; // MARK
1268 let Word1{31} = 1; // BARRIER
1270 defm : ExportPattern<EG_ExportSwz, 83>;
1272 def EG_ExportBuf : ExportBufInst {
1273 let Word1{19-16} = 1; // BURST_COUNT
1274 let Word1{20} = 1; // VALID_PIXEL_MODE
1275 let Word1{21} = eop;
1276 let Word1{29-22} = inst;
1277 let Word1{30} = 0; // MARK
1278 let Word1{31} = 1; // BARRIER
1280 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1282 //===----------------------------------------------------------------------===//
1283 // Memory read/write instructions
1284 //===----------------------------------------------------------------------===//
1285 let usesCustomInserter = 1 in {
1287 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
1289 : EG_CF_RAT <0x57, 0x2, 0, (outs), ins,
1290 !strconcat(name, " $rw_gpr, $index_gpr, $eop"), pattern> {
1292 // XXX: Have a separate instruction for non-indexed writes.
1298 let COMP_MASK = comp_mask;
1299 let BURST_COUNT = 0;
1305 } // End usesCustomInserter = 1
1308 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1309 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1310 0x1, "RAT_WRITE_CACHELESS_32_eg",
1311 [(global_store (i32 R600_TReg32_X:$rw_gpr), R600_TReg32_X:$index_gpr)]
1315 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1316 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1317 0xf, "RAT_WRITE_CACHELESS_128",
1318 [(global_store (v4i32 R600_Reg128:$rw_gpr), R600_TReg32_X:$index_gpr)]
1321 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1322 : InstR600ISA <outs, (ins MEMxi:$ptr), name#" $dst, $ptr", pattern>,
1323 VTX_WORD1_GPR, VTX_WORD0 {
1328 let FETCH_WHOLE_QUAD = 0;
1329 let BUFFER_ID = buffer_id;
1331 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1332 // to store vertex addresses in any channel, not just X.
1335 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
1336 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
1337 // however, based on my testing if USE_CONST_FIELDS is set, then all
1338 // these fields need to be set to 0.
1339 let USE_CONST_FIELDS = 0;
1340 let NUM_FORMAT_ALL = 1;
1341 let FORMAT_COMP_ALL = 0;
1342 let SRF_MODE_ALL = 0;
1344 let Inst{31-0} = Word0;
1345 let Inst{63-32} = Word1;
1346 // LLVM can only encode 64-bit instructions, so these fields are manually
1347 // encoded in R600CodeEmitter
1350 // bits<2> ENDIAN_SWAP = 0;
1351 // bits<1> CONST_BUF_NO_STRIDE = 0;
1352 // bits<1> MEGA_FETCH = 0;
1353 // bits<1> ALT_CONST = 0;
1354 // bits<2> BUFFER_INDEX_MODE = 0;
1358 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1359 // is done in R600CodeEmitter
1361 // Inst{79-64} = OFFSET;
1362 // Inst{81-80} = ENDIAN_SWAP;
1363 // Inst{82} = CONST_BUF_NO_STRIDE;
1364 // Inst{83} = MEGA_FETCH;
1365 // Inst{84} = ALT_CONST;
1366 // Inst{86-85} = BUFFER_INDEX_MODE;
1367 // Inst{95-86} = 0; Reserved
1369 // VTX_WORD3 (Padding)
1371 // Inst{127-96} = 0;
1374 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1375 : VTX_READ_eg <"VTX_READ_8", buffer_id, (outs R600_TReg32_X:$dst),
1378 let MEGA_FETCH_COUNT = 1;
1380 let DST_SEL_Y = 7; // Masked
1381 let DST_SEL_Z = 7; // Masked
1382 let DST_SEL_W = 7; // Masked
1383 let DATA_FORMAT = 1; // FMT_8
1386 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1387 : VTX_READ_eg <"VTX_READ_16", buffer_id, (outs R600_TReg32_X:$dst),
1389 let MEGA_FETCH_COUNT = 2;
1391 let DST_SEL_Y = 7; // Masked
1392 let DST_SEL_Z = 7; // Masked
1393 let DST_SEL_W = 7; // Masked
1394 let DATA_FORMAT = 5; // FMT_16
1398 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1399 : VTX_READ_eg <"VTX_READ_32", buffer_id, (outs R600_TReg32_X:$dst),
1402 let MEGA_FETCH_COUNT = 4;
1404 let DST_SEL_Y = 7; // Masked
1405 let DST_SEL_Z = 7; // Masked
1406 let DST_SEL_W = 7; // Masked
1407 let DATA_FORMAT = 0xD; // COLOR_32
1409 // This is not really necessary, but there were some GPU hangs that appeared
1410 // to be caused by ALU instructions in the next instruction group that wrote
1411 // to the $ptr registers of the VTX_READ.
1413 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1414 // %T2_X<def> = MOV %ZERO
1415 //Adding this constraint prevents this from happening.
1416 let Constraints = "$ptr.ptr = $dst";
1419 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1420 : VTX_READ_eg <"VTX_READ_128", buffer_id, (outs R600_Reg128:$dst),
1423 let MEGA_FETCH_COUNT = 16;
1428 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1430 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1431 // that holds its buffer address to avoid potential hangs. We can't use
1432 // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
1433 // registers are different sizes.
1436 //===----------------------------------------------------------------------===//
1437 // VTX Read from parameter memory space
1438 //===----------------------------------------------------------------------===//
1440 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1441 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti8 ADDRVTX_READ:$ptr))]
1444 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1445 [(set (i32 R600_TReg32_X:$dst), (load_param_zexti16 ADDRVTX_READ:$ptr))]
1448 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1449 [(set (i32 R600_TReg32_X:$dst), (load_param ADDRVTX_READ:$ptr))]
1452 //===----------------------------------------------------------------------===//
1453 // VTX Read from global memory space
1454 //===----------------------------------------------------------------------===//
1457 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1458 [(set (i32 R600_TReg32_X:$dst), (zextloadi8_global ADDRVTX_READ:$ptr))]
1462 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1463 [(set (i32 R600_TReg32_X:$dst), (global_load ADDRVTX_READ:$ptr))]
1467 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1468 [(set (v4i32 R600_Reg128:$dst), (global_load ADDRVTX_READ:$ptr))]
1471 //===----------------------------------------------------------------------===//
1473 // XXX: We are currently storing all constants in the global address space.
1474 //===----------------------------------------------------------------------===//
1476 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1477 [(set (i32 R600_TReg32_X:$dst), (constant_load ADDRVTX_READ:$ptr))]
1482 //===----------------------------------------------------------------------===//
1483 // Regist loads and stores - for indirect addressing
1484 //===----------------------------------------------------------------------===//
1486 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1488 let Predicates = [isCayman] in {
1490 let isVector = 1 in {
1492 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1494 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1495 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1496 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1497 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1498 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1499 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1500 def LOG_IEEE_ : LOG_IEEE_Common<0x83>;
1501 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1502 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1503 def SIN_cm : SIN_Common<0x8D>;
1504 def COS_cm : COS_Common<0x8E>;
1505 } // End isVector = 1
1507 def : SIN_PAT <SIN_cm>;
1508 def : COS_PAT <COS_cm>;
1510 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1512 // RECIP_UINT emulation for Cayman
1514 (AMDGPUurecip R600_Reg32:$src0),
1515 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg R600_Reg32:$src0)),
1516 (MOV_IMM_I32 0x4f800000)))
1520 def : Pat<(fsqrt R600_Reg32:$src),
1521 (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm R600_Reg32:$src))>;
1525 //===----------------------------------------------------------------------===//
1526 // Branch Instructions
1527 //===----------------------------------------------------------------------===//
1530 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
1531 "IF_PREDICATE_SET $src", []>;
1533 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
1534 "PREDICATED_BREAK $src", []>;
1536 //===----------------------------------------------------------------------===//
1537 // Pseudo instructions
1538 //===----------------------------------------------------------------------===//
1540 let isPseudo = 1 in {
1542 def PRED_X : InstR600 <
1543 0, (outs R600_Predicate_Bit:$dst),
1544 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
1546 let FlagOperandIdx = 3;
1549 let isTerminator = 1, isBranch = 1, isBarrier = 1 in {
1551 def JUMP : InstR600 <0x10,
1553 (ins brtarget:$target, R600_Pred:$p),
1554 "JUMP $target ($p)",
1558 } // End isTerminator = 1, isBranch = 1, isBarrier = 1
1560 let usesCustomInserter = 1 in {
1562 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
1564 def MASK_WRITE : AMDGPUShaderInst <
1566 (ins R600_Reg32:$src),
1571 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
1574 def TXD: AMDGPUShaderInst <
1575 (outs R600_Reg128:$dst),
1576 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1577 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1578 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
1581 def TXD_SHADOW: AMDGPUShaderInst <
1582 (outs R600_Reg128:$dst),
1583 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1584 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1585 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
1588 } // End isPseudo = 1
1589 } // End usesCustomInserter = 1
1591 def CLAMP_R600 : CLAMP <R600_Reg32>;
1592 def FABS_R600 : FABS<R600_Reg32>;
1593 def FNEG_R600 : FNEG<R600_Reg32>;
1595 //===---------------------------------------------------------------------===//
1596 // Return instruction
1597 //===---------------------------------------------------------------------===//
1598 let isTerminator = 1, isReturn = 1, isBarrier = 1, hasCtrlDep = 1,
1599 usesCustomInserter = 1 in {
1600 def RETURN : ILFormat<(outs), (ins variable_ops),
1601 "RETURN", [(IL_retflag)]>;
1605 //===----------------------------------------------------------------------===//
1606 // Constant Buffer Addressing Support
1607 //===----------------------------------------------------------------------===//
1609 let isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
1610 def CONST_COPY : Instruction {
1611 let OutOperandList = (outs R600_Reg32:$dst);
1612 let InOperandList = (ins i32imm:$src);
1613 let Pattern = [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
1614 let AsmString = "CONST_COPY";
1615 let neverHasSideEffects = 1;
1616 let isAsCheapAsAMove = 1;
1617 let Itinerary = NullALU;
1619 } // end isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
1621 def TEX_VTX_CONSTBUF :
1622 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr), "VTX_READ_eg $dst, $ptr",
1623 [(set R600_Reg128:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr))]>,
1624 VTX_WORD1_GPR, VTX_WORD0 {
1628 let FETCH_WHOLE_QUAD = 0;
1633 let USE_CONST_FIELDS = 0;
1634 let NUM_FORMAT_ALL = 2;
1635 let FORMAT_COMP_ALL = 1;
1636 let SRF_MODE_ALL = 1;
1637 let MEGA_FETCH_COUNT = 16;
1642 let DATA_FORMAT = 35;
1644 let Inst{31-0} = Word0;
1645 let Inst{63-32} = Word1;
1647 // LLVM can only encode 64-bit instructions, so these fields are manually
1648 // encoded in R600CodeEmitter
1651 // bits<2> ENDIAN_SWAP = 0;
1652 // bits<1> CONST_BUF_NO_STRIDE = 0;
1653 // bits<1> MEGA_FETCH = 0;
1654 // bits<1> ALT_CONST = 0;
1655 // bits<2> BUFFER_INDEX_MODE = 0;
1659 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1660 // is done in R600CodeEmitter
1662 // Inst{79-64} = OFFSET;
1663 // Inst{81-80} = ENDIAN_SWAP;
1664 // Inst{82} = CONST_BUF_NO_STRIDE;
1665 // Inst{83} = MEGA_FETCH;
1666 // Inst{84} = ALT_CONST;
1667 // Inst{86-85} = BUFFER_INDEX_MODE;
1668 // Inst{95-86} = 0; Reserved
1670 // VTX_WORD3 (Padding)
1672 // Inst{127-96} = 0;
1676 //===--------------------------------------------------------------------===//
1677 // Instructions support
1678 //===--------------------------------------------------------------------===//
1679 //===---------------------------------------------------------------------===//
1680 // Custom Inserter for Branches and returns, this eventually will be a
1682 //===---------------------------------------------------------------------===//
1683 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
1684 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
1685 "; Pseudo unconditional branch instruction",
1687 defm BRANCH_COND : BranchConditional<IL_brcond>;
1690 //===---------------------------------------------------------------------===//
1691 // Flow and Program control Instructions
1692 //===---------------------------------------------------------------------===//
1693 let isTerminator=1 in {
1694 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
1695 !strconcat("SWITCH", " $src"), []>;
1696 def CASE : ILFormat< (outs), (ins GPRI32:$src),
1697 !strconcat("CASE", " $src"), []>;
1698 def BREAK : ILFormat< (outs), (ins),
1700 def CONTINUE : ILFormat< (outs), (ins),
1702 def DEFAULT : ILFormat< (outs), (ins),
1704 def ELSE : ILFormat< (outs), (ins),
1706 def ENDSWITCH : ILFormat< (outs), (ins),
1708 def ENDMAIN : ILFormat< (outs), (ins),
1710 def END : ILFormat< (outs), (ins),
1712 def ENDFUNC : ILFormat< (outs), (ins),
1714 def ENDIF : ILFormat< (outs), (ins),
1716 def WHILELOOP : ILFormat< (outs), (ins),
1718 def ENDLOOP : ILFormat< (outs), (ins),
1720 def FUNC : ILFormat< (outs), (ins),
1722 def RETDYN : ILFormat< (outs), (ins),
1724 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1725 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
1726 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1727 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
1728 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1729 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
1730 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1731 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
1732 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1733 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
1734 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1735 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
1736 defm IFC : BranchInstr2<"IFC">;
1737 defm BREAKC : BranchInstr2<"BREAKC">;
1738 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
1741 //===----------------------------------------------------------------------===//
1743 //===----------------------------------------------------------------------===//
1745 //CNDGE_INT extra pattern
1747 (selectcc (i32 R600_Reg32:$src0), -1, (i32 R600_Reg32:$src1),
1748 (i32 R600_Reg32:$src2), COND_GT),
1749 (CNDGE_INT R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2)
1755 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
1759 (int_AMDGPU_kill R600_Reg32:$src0),
1760 (MASK_WRITE (KILLGT (f32 ZERO), (f32 R600_Reg32:$src0)))
1765 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LT),
1766 (SGT R600_Reg32:$src1, R600_Reg32:$src0)
1771 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LE),
1772 (SGE R600_Reg32:$src1, R600_Reg32:$src0)
1775 // SETGT_INT reverse args
1777 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLT),
1778 (SETGT_INT R600_Reg32:$src1, R600_Reg32:$src0)
1781 // SETGE_INT reverse args
1783 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLE),
1784 (SETGE_INT R600_Reg32:$src1, R600_Reg32:$src0)
1787 // SETGT_UINT reverse args
1789 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULT),
1790 (SETGT_UINT R600_Reg32:$src1, R600_Reg32:$src0)
1793 // SETGE_UINT reverse args
1795 (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULE),
1796 (SETGE_UINT R600_Reg32:$src1, R600_Reg32:$src0)
1799 // The next two patterns are special cases for handling 'true if ordered' and
1800 // 'true if unordered' conditionals. The assumption here is that the behavior of
1801 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
1803 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
1804 // We assume that SETE returns false when one of the operands is NAN and
1805 // SNE returns true when on of the operands is NAN
1807 //SETE - 'true if ordered'
1809 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETO),
1810 (SETE R600_Reg32:$src0, R600_Reg32:$src1)
1813 //SNE - 'true if unordered'
1815 (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETUO),
1816 (SNE R600_Reg32:$src0, R600_Reg32:$src1)
1819 def : Extract_Element <f32, v4f32, R600_Reg128, 0, sel_x>;
1820 def : Extract_Element <f32, v4f32, R600_Reg128, 1, sel_y>;
1821 def : Extract_Element <f32, v4f32, R600_Reg128, 2, sel_z>;
1822 def : Extract_Element <f32, v4f32, R600_Reg128, 3, sel_w>;
1824 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 0, sel_x>;
1825 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 1, sel_y>;
1826 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 2, sel_z>;
1827 def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 3, sel_w>;
1829 def : Extract_Element <i32, v4i32, R600_Reg128, 0, sel_x>;
1830 def : Extract_Element <i32, v4i32, R600_Reg128, 1, sel_y>;
1831 def : Extract_Element <i32, v4i32, R600_Reg128, 2, sel_z>;
1832 def : Extract_Element <i32, v4i32, R600_Reg128, 3, sel_w>;
1834 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 0, sel_x>;
1835 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 1, sel_y>;
1836 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 2, sel_z>;
1837 def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 3, sel_w>;
1839 def : Vector_Build <v4f32, R600_Reg128, f32, R600_Reg32>;
1840 def : Vector_Build <v4i32, R600_Reg128, i32, R600_Reg32>;
1842 // bitconvert patterns
1844 def : BitConvert <i32, f32, R600_Reg32>;
1845 def : BitConvert <f32, i32, R600_Reg32>;
1846 def : BitConvert <v4f32, v4i32, R600_Reg128>;
1847 def : BitConvert <v4i32, v4f32, R600_Reg128>;
1849 // DWORDADDR pattern
1850 def : DwordAddrPat <i32, R600_Reg32>;
1852 } // End isR600toCayman Predicate