1 //===-- R600Instructions.td - R600 Instruction defs -------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // R600 Tablegen instruction definitions
12 //===----------------------------------------------------------------------===//
14 include "R600Intrinsics.td"
16 class InstR600 <dag outs, dag ins, string asm, list<dag> pattern,
18 : AMDGPUInst <outs, ins, asm, pattern> {
25 bits<2> FlagOperandIdx = 0;
28 bit HasNativeOperands = 0;
32 let Namespace = "AMDGPU";
33 let OutOperandList = outs;
34 let InOperandList = ins;
36 let Pattern = pattern;
39 let TSFlags{0} = TransOnly;
40 let TSFlags{4} = Trig;
43 // Vector instructions are instructions that must fill all slots in an
45 let TSFlags{6} = isVector;
46 let TSFlags{8-7} = FlagOperandIdx;
47 let TSFlags{9} = HasNativeOperands;
48 let TSFlags{10} = Op1;
49 let TSFlags{11} = Op2;
50 let TSFlags{12} = VTXInst;
51 let TSFlags{13} = TEXInst;
54 class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
55 InstR600 <outs, ins, asm, pattern, NullALU> {
57 let Namespace = "AMDGPU";
60 def MEMxi : Operand<iPTR> {
61 let MIOperandInfo = (ops R600_TReg32_X:$ptr, i32imm:$index);
62 let PrintMethod = "printMemOperand";
65 def MEMrr : Operand<iPTR> {
66 let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
69 // Operands for non-registers
71 class InstFlag<string PM = "printOperand", int Default = 0>
72 : OperandWithDefaultOps <i32, (ops (i32 Default))> {
76 // src_sel for ALU src operands, see also ALU_CONST, ALU_PARAM registers
77 def SEL : OperandWithDefaultOps <i32, (ops (i32 -1))> {
78 let PrintMethod = "printSel";
80 def BANK_SWIZZLE : OperandWithDefaultOps <i32, (ops (i32 0))> {
81 let PrintMethod = "printBankSwizzle";
84 def LITERAL : InstFlag<"printLiteral">;
86 def WRITE : InstFlag <"printWrite", 1>;
87 def OMOD : InstFlag <"printOMOD">;
88 def REL : InstFlag <"printRel">;
89 def CLAMP : InstFlag <"printClamp">;
90 def NEG : InstFlag <"printNeg">;
91 def ABS : InstFlag <"printAbs">;
92 def UEM : InstFlag <"printUpdateExecMask">;
93 def UP : InstFlag <"printUpdatePred">;
95 // XXX: The r600g finalizer in Mesa expects last to be one in most cases.
96 // Once we start using the packetizer in this backend we should have this
98 def LAST : InstFlag<"printLast", 1>;
99 def RSel : Operand<i32> {
100 let PrintMethod = "printRSel";
102 def CT: Operand<i32> {
103 let PrintMethod = "printCT";
106 def FRAMEri : Operand<iPTR> {
107 let MIOperandInfo = (ops R600_Reg32:$ptr, i32imm:$index);
110 def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
111 def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
112 def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
113 def ADDRGA_CONST_OFFSET : ComplexPattern<i32, 1, "SelectGlobalValueConstantOffset", [], []>;
114 def ADDRGA_VAR_OFFSET : ComplexPattern<i32, 2, "SelectGlobalValueVariableOffset", [], []>;
115 def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>;
117 class R600ALU_Word0 {
118 field bits<32> Word0;
126 bits<3> index_mode = 0;
130 bits<9> src0_sel = src0{8-0};
131 bits<2> src0_chan = src0{10-9};
132 bits<9> src1_sel = src1{8-0};
133 bits<2> src1_chan = src1{10-9};
135 let Word0{8-0} = src0_sel;
136 let Word0{9} = src0_rel;
137 let Word0{11-10} = src0_chan;
138 let Word0{12} = src0_neg;
139 let Word0{21-13} = src1_sel;
140 let Word0{22} = src1_rel;
141 let Word0{24-23} = src1_chan;
142 let Word0{25} = src1_neg;
143 let Word0{28-26} = index_mode;
144 let Word0{30-29} = pred_sel;
145 let Word0{31} = last;
148 class R600ALU_Word1 {
149 field bits<32> Word1;
152 bits<3> bank_swizzle;
156 bits<7> dst_sel = dst{6-0};
157 bits<2> dst_chan = dst{10-9};
159 let Word1{20-18} = bank_swizzle;
160 let Word1{27-21} = dst_sel;
161 let Word1{28} = dst_rel;
162 let Word1{30-29} = dst_chan;
163 let Word1{31} = clamp;
166 class R600ALU_Word1_OP2 <bits<11> alu_inst> : R600ALU_Word1{
170 bits<1> update_exec_mask;
175 let Word1{0} = src0_abs;
176 let Word1{1} = src1_abs;
177 let Word1{2} = update_exec_mask;
178 let Word1{3} = update_pred;
179 let Word1{4} = write;
180 let Word1{6-5} = omod;
181 let Word1{17-7} = alu_inst;
184 class R600ALU_Word1_OP3 <bits<5> alu_inst> : R600ALU_Word1{
190 bits<9> src2_sel = src2{8-0};
191 bits<2> src2_chan = src2{10-9};
193 let Word1{8-0} = src2_sel;
194 let Word1{9} = src2_rel;
195 let Word1{11-10} = src2_chan;
196 let Word1{12} = src2_neg;
197 let Word1{17-13} = alu_inst;
201 field bits<32> Word0;
205 bits<1> FETCH_WHOLE_QUAD;
209 bits<6> MEGA_FETCH_COUNT;
211 let Word0{4-0} = VC_INST;
212 let Word0{6-5} = FETCH_TYPE;
213 let Word0{7} = FETCH_WHOLE_QUAD;
214 let Word0{15-8} = BUFFER_ID;
215 let Word0{22-16} = SRC_GPR;
216 let Word0{23} = SRC_REL;
217 let Word0{25-24} = SRC_SEL_X;
218 let Word0{31-26} = MEGA_FETCH_COUNT;
221 class VTX_WORD1_GPR {
222 field bits<32> Word1;
229 bits<1> USE_CONST_FIELDS;
231 bits<2> NUM_FORMAT_ALL;
232 bits<1> FORMAT_COMP_ALL;
233 bits<1> SRF_MODE_ALL;
235 let Word1{6-0} = DST_GPR;
236 let Word1{7} = DST_REL;
237 let Word1{8} = 0; // Reserved
238 let Word1{11-9} = DST_SEL_X;
239 let Word1{14-12} = DST_SEL_Y;
240 let Word1{17-15} = DST_SEL_Z;
241 let Word1{20-18} = DST_SEL_W;
242 let Word1{21} = USE_CONST_FIELDS;
243 let Word1{27-22} = DATA_FORMAT;
244 let Word1{29-28} = NUM_FORMAT_ALL;
245 let Word1{30} = FORMAT_COMP_ALL;
246 let Word1{31} = SRF_MODE_ALL;
250 field bits<32> Word0;
254 bits<1> FETCH_WHOLE_QUAD;
259 bits<2> RESOURCE_INDEX_MODE;
260 bits<2> SAMPLER_INDEX_MODE;
262 let Word0{4-0} = TEX_INST;
263 let Word0{6-5} = INST_MOD;
264 let Word0{7} = FETCH_WHOLE_QUAD;
265 let Word0{15-8} = RESOURCE_ID;
266 let Word0{22-16} = SRC_GPR;
267 let Word0{23} = SRC_REL;
268 let Word0{24} = ALT_CONST;
269 let Word0{26-25} = RESOURCE_INDEX_MODE;
270 let Word0{28-27} = SAMPLER_INDEX_MODE;
274 field bits<32> Word1;
283 bits<1> COORD_TYPE_X;
284 bits<1> COORD_TYPE_Y;
285 bits<1> COORD_TYPE_Z;
286 bits<1> COORD_TYPE_W;
288 let Word1{6-0} = DST_GPR;
289 let Word1{7} = DST_REL;
290 let Word1{11-9} = DST_SEL_X;
291 let Word1{14-12} = DST_SEL_Y;
292 let Word1{17-15} = DST_SEL_Z;
293 let Word1{20-18} = DST_SEL_W;
294 let Word1{27-21} = LOD_BIAS;
295 let Word1{28} = COORD_TYPE_X;
296 let Word1{29} = COORD_TYPE_Y;
297 let Word1{30} = COORD_TYPE_Z;
298 let Word1{31} = COORD_TYPE_W;
302 field bits<32> Word2;
313 let Word2{4-0} = OFFSET_X;
314 let Word2{9-5} = OFFSET_Y;
315 let Word2{14-10} = OFFSET_Z;
316 let Word2{19-15} = SAMPLER_ID;
317 let Word2{22-20} = SRC_SEL_X;
318 let Word2{25-23} = SRC_SEL_Y;
319 let Word2{28-26} = SRC_SEL_Z;
320 let Word2{31-29} = SRC_SEL_W;
324 XXX: R600 subtarget uses a slightly different encoding than the other
325 subtargets. We currently handle this in R600MCCodeEmitter, but we may
326 want to use these instruction classes in the future.
328 class R600ALU_Word1_OP2_r600 : R600ALU_Word1_OP2 {
333 let Inst{37} = fog_merge;
334 let Inst{39-38} = omod;
335 let Inst{49-40} = alu_inst;
338 class R600ALU_Word1_OP2_r700 : R600ALU_Word1_OP2 {
342 let Inst{38-37} = omod;
343 let Inst{49-39} = alu_inst;
347 def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
351 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
353 // Class for instructions with only one source register.
354 // If you add new ins to this instruction, make sure they are listed before
355 // $literal, because the backend currently assumes that the last operand is
356 // a literal. Also be sure to update the enum R600Op1OperandIndex::ROI in
357 // R600Defines.h, R600InstrInfo::buildDefaultInstruction(),
358 // and R600InstrInfo::getOperandIdx().
359 class R600_1OP <bits<11> inst, string opName, list<dag> pattern,
360 InstrItinClass itin = AnyALU> :
361 InstR600 <(outs R600_Reg32:$dst),
362 (ins WRITE:$write, OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
363 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
364 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
365 BANK_SWIZZLE:$bank_swizzle),
366 !strconcat(" ", opName,
367 "$clamp $last $dst$write$dst_rel$omod, "
368 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
369 "$pred_sel $bank_swizzle"),
373 R600ALU_Word1_OP2 <inst> {
379 let update_exec_mask = 0;
381 let HasNativeOperands = 1;
383 let DisableEncoding = "$literal";
385 let Inst{31-0} = Word0;
386 let Inst{63-32} = Word1;
389 class R600_1OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
390 InstrItinClass itin = AnyALU> :
391 R600_1OP <inst, opName,
392 [(set R600_Reg32:$dst, (node R600_Reg32:$src0))]
395 // If you add our change the operands for R600_2OP instructions, you must
396 // also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
397 // R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
398 class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
399 InstrItinClass itin = AnyALU> :
400 InstR600 <(outs R600_Reg32:$dst),
401 (ins UEM:$update_exec_mask, UP:$update_pred, WRITE:$write,
402 OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
403 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
404 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs, SEL:$src1_sel,
405 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
406 BANK_SWIZZLE:$bank_swizzle),
407 !strconcat(" ", opName,
408 "$clamp $last $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
409 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
410 "$src1_neg$src1_abs$src1$src1_abs$src1_rel, "
411 "$pred_sel $bank_swizzle"),
415 R600ALU_Word1_OP2 <inst> {
417 let HasNativeOperands = 1;
419 let DisableEncoding = "$literal";
421 let Inst{31-0} = Word0;
422 let Inst{63-32} = Word1;
425 class R600_2OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
426 InstrItinClass itim = AnyALU> :
427 R600_2OP <inst, opName,
428 [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
432 // If you add our change the operands for R600_3OP instructions, you must
433 // also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
434 // R600InstrInfo::buildDefaultInstruction(), and
435 // R600InstrInfo::getOperandIdx().
436 class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
437 InstrItinClass itin = AnyALU> :
438 InstR600 <(outs R600_Reg32:$dst),
439 (ins REL:$dst_rel, CLAMP:$clamp,
440 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, SEL:$src0_sel,
441 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
442 R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
443 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
444 BANK_SWIZZLE:$bank_swizzle),
445 !strconcat(" ", opName, "$clamp $last $dst$dst_rel, "
446 "$src0_neg$src0$src0_rel, "
447 "$src1_neg$src1$src1_rel, "
448 "$src2_neg$src2$src2_rel, "
454 R600ALU_Word1_OP3<inst>{
456 let HasNativeOperands = 1;
457 let DisableEncoding = "$literal";
460 let Inst{31-0} = Word0;
461 let Inst{63-32} = Word1;
464 class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
465 InstrItinClass itin = VecALU> :
466 InstR600 <(outs R600_Reg32:$dst),
474 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
476 def TEX_SHADOW : PatLeaf<
478 [{uint32_t TType = (uint32_t)N->getZExtValue();
479 return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
483 def TEX_RECT : PatLeaf<
485 [{uint32_t TType = (uint32_t)N->getZExtValue();
490 def TEX_ARRAY : PatLeaf<
492 [{uint32_t TType = (uint32_t)N->getZExtValue();
493 return TType == 9 || TType == 10 || TType == 15 || TType == 16;
497 def TEX_SHADOW_ARRAY : PatLeaf<
499 [{uint32_t TType = (uint32_t)N->getZExtValue();
500 return TType == 11 || TType == 12 || TType == 17;
504 class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, bits<4> rat_id, dag outs,
505 dag ins, string asm, list<dag> pattern> :
506 InstR600ISA <outs, ins, asm, pattern> {
523 // CF_ALLOC_EXPORT_WORD0_RAT
524 let Inst{3-0} = rat_id;
525 let Inst{9-4} = rat_inst;
526 let Inst{10} = 0; // Reserved
527 let Inst{12-11} = RIM;
528 let Inst{14-13} = TYPE;
529 let Inst{21-15} = RW_GPR;
530 let Inst{22} = RW_REL;
531 let Inst{29-23} = INDEX_GPR;
532 let Inst{31-30} = ELEM_SIZE;
534 // CF_ALLOC_EXPORT_WORD1_BUF
535 let Inst{43-32} = ARRAY_SIZE;
536 let Inst{47-44} = COMP_MASK;
537 let Inst{51-48} = BURST_COUNT;
540 let Inst{61-54} = cf_inst;
542 let Inst{63} = BARRIER;
545 class LoadParamFrag <PatFrag load_type> : PatFrag <
546 (ops node:$ptr), (load_type node:$ptr),
547 [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
550 def load_param : LoadParamFrag<load>;
551 def load_param_zexti8 : LoadParamFrag<zextloadi8>;
552 def load_param_zexti16 : LoadParamFrag<zextloadi16>;
554 def isR600 : Predicate<"Subtarget.getGeneration() <= AMDGPUSubtarget::R700">;
555 def isR700 : Predicate<"Subtarget.getGeneration() == AMDGPUSubtarget::R700">;
556 def isEG : Predicate<
557 "Subtarget.getGeneration() >= AMDGPUSubtarget::EVERGREEN && "
558 "Subtarget.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS && "
559 "!Subtarget.hasCaymanISA()">;
561 def isCayman : Predicate<"Subtarget.hasCaymanISA()">;
562 def isEGorCayman : Predicate<"Subtarget.getGeneration() == "
563 "AMDGPUSubtarget::EVERGREEN"
564 "|| Subtarget.getGeneration() =="
565 "AMDGPUSubtarget::NORTHERN_ISLANDS">;
567 def isR600toCayman : Predicate<
568 "Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS">;
570 //===----------------------------------------------------------------------===//
572 //===----------------------------------------------------------------------===//
574 def INTERP_PAIR_XY : AMDGPUShaderInst <
575 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
576 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
577 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
580 def INTERP_PAIR_ZW : AMDGPUShaderInst <
581 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
582 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
583 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
586 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
587 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
591 def DOT4 : SDNode<"AMDGPUISD::DOT4",
592 SDTypeProfile<1, 8, [SDTCisFP<0>, SDTCisVT<1, f32>, SDTCisVT<2, f32>,
593 SDTCisVT<3, f32>, SDTCisVT<4, f32>, SDTCisVT<5, f32>,
594 SDTCisVT<6, f32>, SDTCisVT<7, f32>, SDTCisVT<8, f32>]>,
598 def TEXTURE_FETCH_Type : SDTypeProfile<1, 19, [SDTCisFP<0>]>;
600 def TEXTURE_FETCH: SDNode<"AMDGPUISD::TEXTURE_FETCH", TEXTURE_FETCH_Type, []>;
602 multiclass TexPattern<bits<32> TextureOp, Instruction inst, ValueType vt = v4f32> {
603 def : Pat<(TEXTURE_FETCH (i32 TextureOp), vt:$SRC_GPR,
604 (i32 imm:$srcx), (i32 imm:$srcy), (i32 imm:$srcz), (i32 imm:$srcw),
605 (i32 imm:$offsetx), (i32 imm:$offsety), (i32 imm:$offsetz),
606 (i32 imm:$DST_SEL_X), (i32 imm:$DST_SEL_Y), (i32 imm:$DST_SEL_Z),
607 (i32 imm:$DST_SEL_W),
608 (i32 imm:$RESOURCE_ID), (i32 imm:$SAMPLER_ID),
609 (i32 imm:$COORD_TYPE_X), (i32 imm:$COORD_TYPE_Y), (i32 imm:$COORD_TYPE_Z),
610 (i32 imm:$COORD_TYPE_W)),
611 (inst R600_Reg128:$SRC_GPR,
612 imm:$srcx, imm:$srcy, imm:$srcz, imm:$srcw,
613 imm:$offsetx, imm:$offsety, imm:$offsetz,
614 imm:$DST_SEL_X, imm:$DST_SEL_Y, imm:$DST_SEL_Z,
616 imm:$RESOURCE_ID, imm:$SAMPLER_ID,
617 imm:$COORD_TYPE_X, imm:$COORD_TYPE_Y, imm:$COORD_TYPE_Z,
621 //===----------------------------------------------------------------------===//
622 // Interpolation Instructions
623 //===----------------------------------------------------------------------===//
625 def INTERP_VEC_LOAD : AMDGPUShaderInst <
626 (outs R600_Reg128:$dst),
628 "INTERP_LOAD $src0 : $dst",
631 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
632 let bank_swizzle = 5;
635 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
636 let bank_swizzle = 5;
639 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
641 //===----------------------------------------------------------------------===//
642 // Export Instructions
643 //===----------------------------------------------------------------------===//
645 def ExportType : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
647 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
648 [SDNPHasChain, SDNPSideEffect]>;
651 field bits<32> Word0;
658 let Word0{12-0} = arraybase;
659 let Word0{14-13} = type;
660 let Word0{21-15} = gpr;
661 let Word0{22} = 0; // RW_REL
662 let Word0{29-23} = 0; // INDEX_GPR
663 let Word0{31-30} = elem_size;
666 class ExportSwzWord1 {
667 field bits<32> Word1;
676 let Word1{2-0} = sw_x;
677 let Word1{5-3} = sw_y;
678 let Word1{8-6} = sw_z;
679 let Word1{11-9} = sw_w;
682 class ExportBufWord1 {
683 field bits<32> Word1;
690 let Word1{11-0} = arraySize;
691 let Word1{15-12} = compMask;
694 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
695 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
697 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
698 0, 61, 0, 7, 7, 7, cf_inst, 0)
701 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
703 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
704 0, 61, 7, 0, 7, 7, cf_inst, 0)
707 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
709 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
712 def : Pat<(int_R600_store_dummy 1),
714 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
717 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
718 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
719 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
720 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
725 multiclass SteamOutputExportPattern<Instruction ExportInst,
726 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
728 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
729 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
730 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
731 4095, imm:$mask, buf0inst, 0)>;
733 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
734 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
735 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
736 4095, imm:$mask, buf1inst, 0)>;
738 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
739 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
740 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
741 4095, imm:$mask, buf2inst, 0)>;
743 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
744 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
745 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
746 4095, imm:$mask, buf3inst, 0)>;
749 // Export Instructions should not be duplicated by TailDuplication pass
750 // (which assumes that duplicable instruction are affected by exec mask)
751 let usesCustomInserter = 1, isNotDuplicable = 1 in {
753 class ExportSwzInst : InstR600ISA<(
755 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
756 i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
758 !strconcat("EXPORT", " $gpr"),
759 []>, ExportWord0, ExportSwzWord1 {
761 let Inst{31-0} = Word0;
762 let Inst{63-32} = Word1;
765 } // End usesCustomInserter = 1
767 class ExportBufInst : InstR600ISA<(
769 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
770 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
771 !strconcat("EXPORT", " $gpr"),
772 []>, ExportWord0, ExportBufWord1 {
774 let Inst{31-0} = Word0;
775 let Inst{63-32} = Word1;
778 //===----------------------------------------------------------------------===//
779 // Control Flow Instructions
780 //===----------------------------------------------------------------------===//
783 field bits<32> Word0;
786 bits<4> KCACHE_BANK0;
787 bits<4> KCACHE_BANK1;
788 bits<2> KCACHE_MODE0;
790 let Word0{21-0} = ADDR;
791 let Word0{25-22} = KCACHE_BANK0;
792 let Word0{29-26} = KCACHE_BANK1;
793 let Word0{31-30} = KCACHE_MODE0;
797 field bits<32> Word1;
799 bits<2> KCACHE_MODE1;
800 bits<8> KCACHE_ADDR0;
801 bits<8> KCACHE_ADDR1;
805 bits<1> WHOLE_QUAD_MODE;
808 let Word1{1-0} = KCACHE_MODE1;
809 let Word1{9-2} = KCACHE_ADDR0;
810 let Word1{17-10} = KCACHE_ADDR1;
811 let Word1{24-18} = COUNT;
812 let Word1{25} = ALT_CONST;
813 let Word1{29-26} = CF_INST;
814 let Word1{30} = WHOLE_QUAD_MODE;
815 let Word1{31} = BARRIER;
818 def KCACHE : InstFlag<"printKCache">;
820 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
821 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1,
822 KCACHE:$KCACHE_MODE0, KCACHE:$KCACHE_MODE1,
823 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1,
825 !strconcat(OpName, " $COUNT, @$ADDR, "
826 "KC0[$KCACHE_MODE0], KC1[$KCACHE_MODE1]"),
827 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
832 let WHOLE_QUAD_MODE = 0;
835 let Inst{31-0} = Word0;
836 let Inst{63-32} = Word1;
839 class CF_WORD0_R600 {
840 field bits<32> Word0;
847 class CF_WORD1_R600 {
848 field bits<32> Word1;
856 bits<1> END_OF_PROGRAM;
857 bits<1> VALID_PIXEL_MODE;
859 bits<1> WHOLE_QUAD_MODE;
862 let Word1{2-0} = POP_COUNT;
863 let Word1{7-3} = CF_CONST;
864 let Word1{9-8} = COND;
865 let Word1{12-10} = COUNT;
866 let Word1{18-13} = CALL_COUNT;
867 let Word1{19} = COUNT_3;
868 let Word1{21} = END_OF_PROGRAM;
869 let Word1{22} = VALID_PIXEL_MODE;
870 let Word1{29-23} = CF_INST;
871 let Word1{30} = WHOLE_QUAD_MODE;
872 let Word1{31} = BARRIER;
875 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
876 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
882 let VALID_PIXEL_MODE = 0;
886 let END_OF_PROGRAM = 0;
887 let WHOLE_QUAD_MODE = 0;
889 let Inst{31-0} = Word0;
890 let Inst{63-32} = Word1;
894 field bits<32> Word0;
897 bits<3> JUMPTABLE_SEL;
899 let Word0{23-0} = ADDR;
900 let Word0{26-24} = JUMPTABLE_SEL;
904 field bits<32> Word1;
910 bits<1> VALID_PIXEL_MODE;
911 bits<1> END_OF_PROGRAM;
915 let Word1{2-0} = POP_COUNT;
916 let Word1{7-3} = CF_CONST;
917 let Word1{9-8} = COND;
918 let Word1{15-10} = COUNT;
919 let Word1{20} = VALID_PIXEL_MODE;
920 let Word1{21} = END_OF_PROGRAM;
921 let Word1{29-22} = CF_INST;
922 let Word1{31} = BARRIER;
925 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
926 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
931 let JUMPTABLE_SEL = 0;
933 let VALID_PIXEL_MODE = 0;
935 let END_OF_PROGRAM = 0;
937 let Inst{31-0} = Word0;
938 let Inst{63-32} = Word1;
941 def CF_ALU : ALU_CLAUSE<8, "ALU">;
942 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
944 def FETCH_CLAUSE : AMDGPUInst <(outs),
945 (ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
951 def ALU_CLAUSE : AMDGPUInst <(outs),
952 (ins i32imm:$addr), "ALU clause starting at $addr:", [] > {
958 def LITERALS : AMDGPUInst <(outs),
959 (ins LITERAL:$literal1, LITERAL:$literal2), "$literal1, $literal2", [] > {
964 let Inst{31-0} = literal1;
965 let Inst{63-32} = literal2;
968 def PAD : AMDGPUInst <(outs), (ins), "PAD", [] > {
972 let Predicates = [isR600toCayman] in {
974 //===----------------------------------------------------------------------===//
975 // Common Instructions R600, R700, Evergreen, Cayman
976 //===----------------------------------------------------------------------===//
978 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
979 // Non-IEEE MUL: 0 * anything = 0
980 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
981 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
982 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
983 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
985 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
986 // so some of the instruction names don't match the asm string.
987 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
988 def SETE : R600_2OP <
990 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_EQ))]
995 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GT))]
1000 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GE))]
1003 def SNE : R600_2OP <
1005 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_NE))]
1008 def SETE_DX10 : R600_2OP <
1010 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_EQ))]
1013 def SETGT_DX10 : R600_2OP <
1015 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GT))]
1018 def SETGE_DX10 : R600_2OP <
1020 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GE))]
1023 def SETNE_DX10 : R600_2OP <
1025 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_NE))]
1028 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
1029 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
1030 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
1031 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
1032 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
1034 def MOV : R600_1OP <0x19, "MOV", []>;
1036 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
1038 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
1039 (outs R600_Reg32:$dst),
1045 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
1047 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
1050 (MOV_IMM_I32 imm:$val)
1053 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
1056 (MOV_IMM_F32 fpimm:$val)
1059 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
1060 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
1061 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
1062 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
1064 let hasSideEffects = 1 in {
1066 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
1068 } // end hasSideEffects
1070 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
1071 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
1072 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
1073 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
1074 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
1075 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
1076 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
1077 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
1078 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
1079 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
1081 def SETE_INT : R600_2OP <
1083 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETEQ))]
1086 def SETGT_INT : R600_2OP <
1088 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGT))]
1091 def SETGE_INT : R600_2OP <
1093 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGE))]
1096 def SETNE_INT : R600_2OP <
1098 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETNE))]
1101 def SETGT_UINT : R600_2OP <
1103 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGT))]
1106 def SETGE_UINT : R600_2OP <
1108 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGE))]
1111 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
1112 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
1113 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
1114 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
1116 def CNDE_INT : R600_3OP <
1118 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_EQ))]
1121 def CNDGE_INT : R600_3OP <
1123 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GE))]
1126 def CNDGT_INT : R600_3OP <
1128 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GT))]
1131 //===----------------------------------------------------------------------===//
1132 // Texture instructions
1133 //===----------------------------------------------------------------------===//
1135 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1137 class R600_TEX <bits<11> inst, string opName> :
1138 InstR600 <(outs R600_Reg128:$DST_GPR),
1139 (ins R600_Reg128:$SRC_GPR,
1140 RSel:$srcx, RSel:$srcy, RSel:$srcz, RSel:$srcw,
1141 i32imm:$offsetx, i32imm:$offsety, i32imm:$offsetz,
1142 RSel:$DST_SEL_X, RSel:$DST_SEL_Y, RSel:$DST_SEL_Z, RSel:$DST_SEL_W,
1143 i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
1144 CT:$COORD_TYPE_X, CT:$COORD_TYPE_Y, CT:$COORD_TYPE_Z,
1147 " $DST_GPR.$DST_SEL_X$DST_SEL_Y$DST_SEL_Z$DST_SEL_W, "
1148 "$SRC_GPR.$srcx$srcy$srcz$srcw "
1149 "RID:$RESOURCE_ID SID:$SAMPLER_ID "
1150 "CT:$COORD_TYPE_X$COORD_TYPE_Y$COORD_TYPE_Z$COORD_TYPE_W"),
1152 NullALU>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
1153 let Inst{31-0} = Word0;
1154 let Inst{63-32} = Word1;
1156 let TEX_INST = inst{4-0};
1162 let FETCH_WHOLE_QUAD = 0;
1164 let SAMPLER_INDEX_MODE = 0;
1165 let RESOURCE_INDEX_MODE = 0;
1170 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1174 def TEX_SAMPLE : R600_TEX <0x10, "TEX_SAMPLE">;
1175 def TEX_SAMPLE_C : R600_TEX <0x18, "TEX_SAMPLE_C">;
1176 def TEX_SAMPLE_L : R600_TEX <0x11, "TEX_SAMPLE_L">;
1177 def TEX_SAMPLE_C_L : R600_TEX <0x19, "TEX_SAMPLE_C_L">;
1178 def TEX_SAMPLE_LB : R600_TEX <0x12, "TEX_SAMPLE_LB">;
1179 def TEX_SAMPLE_C_LB : R600_TEX <0x1A, "TEX_SAMPLE_C_LB">;
1180 def TEX_LD : R600_TEX <0x03, "TEX_LD">;
1181 def TEX_GET_TEXTURE_RESINFO : R600_TEX <0x04, "TEX_GET_TEXTURE_RESINFO">;
1182 def TEX_GET_GRADIENTS_H : R600_TEX <0x07, "TEX_GET_GRADIENTS_H">;
1183 def TEX_GET_GRADIENTS_V : R600_TEX <0x08, "TEX_GET_GRADIENTS_V">;
1184 def TEX_SET_GRADIENTS_H : R600_TEX <0x0B, "TEX_SET_GRADIENTS_H">;
1185 def TEX_SET_GRADIENTS_V : R600_TEX <0x0C, "TEX_SET_GRADIENTS_V">;
1186 def TEX_SAMPLE_G : R600_TEX <0x14, "TEX_SAMPLE_G">;
1187 def TEX_SAMPLE_C_G : R600_TEX <0x1C, "TEX_SAMPLE_C_G">;
1189 defm : TexPattern<0, TEX_SAMPLE>;
1190 defm : TexPattern<1, TEX_SAMPLE_C>;
1191 defm : TexPattern<2, TEX_SAMPLE_L>;
1192 defm : TexPattern<3, TEX_SAMPLE_C_L>;
1193 defm : TexPattern<4, TEX_SAMPLE_LB>;
1194 defm : TexPattern<5, TEX_SAMPLE_C_LB>;
1195 defm : TexPattern<6, TEX_LD, v4i32>;
1196 defm : TexPattern<7, TEX_GET_TEXTURE_RESINFO, v4i32>;
1197 defm : TexPattern<8, TEX_GET_GRADIENTS_H>;
1198 defm : TexPattern<9, TEX_GET_GRADIENTS_V>;
1200 //===----------------------------------------------------------------------===//
1201 // Helper classes for common instructions
1202 //===----------------------------------------------------------------------===//
1204 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
1209 class MULADD_Common <bits<5> inst> : R600_3OP <
1214 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
1215 inst, "MULADD_IEEE",
1216 [(set f32:$dst, (fadd (fmul f32:$src0, f32:$src1), f32:$src2))]
1219 class CNDE_Common <bits<5> inst> : R600_3OP <
1221 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_EQ))]
1224 class CNDGT_Common <bits<5> inst> : R600_3OP <
1226 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GT))]
1229 class CNDGE_Common <bits<5> inst> : R600_3OP <
1231 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GE))]
1235 let isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
1236 class R600_VEC2OP<list<dag> pattern> : InstR600 <(outs R600_Reg32:$dst), (ins
1238 UEM:$update_exec_mask_X, UP:$update_pred_X, WRITE:$write_X,
1239 OMOD:$omod_X, REL:$dst_rel_X, CLAMP:$clamp_X,
1240 R600_TReg32_X:$src0_X, NEG:$src0_neg_X, REL:$src0_rel_X, ABS:$src0_abs_X, SEL:$src0_sel_X,
1241 R600_TReg32_X:$src1_X, NEG:$src1_neg_X, REL:$src1_rel_X, ABS:$src1_abs_X, SEL:$src1_sel_X,
1242 R600_Pred:$pred_sel_X,
1244 UEM:$update_exec_mask_Y, UP:$update_pred_Y, WRITE:$write_Y,
1245 OMOD:$omod_Y, REL:$dst_rel_Y, CLAMP:$clamp_Y,
1246 R600_TReg32_Y:$src0_Y, NEG:$src0_neg_Y, REL:$src0_rel_Y, ABS:$src0_abs_Y, SEL:$src0_sel_Y,
1247 R600_TReg32_Y:$src1_Y, NEG:$src1_neg_Y, REL:$src1_rel_Y, ABS:$src1_abs_Y, SEL:$src1_sel_Y,
1248 R600_Pred:$pred_sel_Y,
1250 UEM:$update_exec_mask_Z, UP:$update_pred_Z, WRITE:$write_Z,
1251 OMOD:$omod_Z, REL:$dst_rel_Z, CLAMP:$clamp_Z,
1252 R600_TReg32_Z:$src0_Z, NEG:$src0_neg_Z, REL:$src0_rel_Z, ABS:$src0_abs_Z, SEL:$src0_sel_Z,
1253 R600_TReg32_Z:$src1_Z, NEG:$src1_neg_Z, REL:$src1_rel_Z, ABS:$src1_abs_Z, SEL:$src1_sel_Z,
1254 R600_Pred:$pred_sel_Z,
1256 UEM:$update_exec_mask_W, UP:$update_pred_W, WRITE:$write_W,
1257 OMOD:$omod_W, REL:$dst_rel_W, CLAMP:$clamp_W,
1258 R600_TReg32_W:$src0_W, NEG:$src0_neg_W, REL:$src0_rel_W, ABS:$src0_abs_W, SEL:$src0_sel_W,
1259 R600_TReg32_W:$src1_W, NEG:$src1_neg_W, REL:$src1_rel_W, ABS:$src1_abs_W, SEL:$src1_sel_W,
1260 R600_Pred:$pred_sel_W,
1261 LITERAL:$literal0, LITERAL:$literal1),
1267 def DOT_4 : R600_VEC2OP<[(set R600_Reg32:$dst, (DOT4
1268 R600_TReg32_X:$src0_X, R600_TReg32_X:$src1_X,
1269 R600_TReg32_Y:$src0_Y, R600_TReg32_Y:$src1_Y,
1270 R600_TReg32_Z:$src0_Z, R600_TReg32_Z:$src1_Z,
1271 R600_TReg32_W:$src0_W, R600_TReg32_W:$src1_W))]>;
1274 class DOT4_Common <bits<11> inst> : R600_2OP <inst, "DOT4", []>;
1277 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1278 multiclass CUBE_Common <bits<11> inst> {
1280 def _pseudo : InstR600 <
1281 (outs R600_Reg128:$dst),
1282 (ins R600_Reg128:$src),
1284 [(set v4f32:$dst, (int_AMDGPU_cube v4f32:$src))],
1290 def _real : R600_2OP <inst, "CUBE", []>;
1292 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1294 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1295 inst, "EXP_IEEE", fexp2
1298 let Itinerary = TransALU;
1301 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
1302 inst, "FLT_TO_INT", fp_to_sint
1305 let Itinerary = TransALU;
1308 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1309 inst, "INT_TO_FLT", sint_to_fp
1312 let Itinerary = TransALU;
1315 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1316 inst, "FLT_TO_UINT", fp_to_uint
1319 let Itinerary = TransALU;
1322 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1323 inst, "UINT_TO_FLT", uint_to_fp
1326 let Itinerary = TransALU;
1329 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1330 inst, "LOG_CLAMPED", []
1333 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1334 inst, "LOG_IEEE", flog2
1337 let Itinerary = TransALU;
1340 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1341 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1342 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1343 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1344 inst, "MULHI_INT", mulhs
1347 let Itinerary = TransALU;
1349 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1350 inst, "MULHI", mulhu
1353 let Itinerary = TransALU;
1355 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1356 inst, "MULLO_INT", mul
1359 let Itinerary = TransALU;
1361 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []> {
1363 let Itinerary = TransALU;
1366 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1367 inst, "RECIP_CLAMPED", []
1370 let Itinerary = TransALU;
1373 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1374 inst, "RECIP_IEEE", [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
1377 let Itinerary = TransALU;
1380 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1381 inst, "RECIP_UINT", AMDGPUurecip
1384 let Itinerary = TransALU;
1387 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1388 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1391 let Itinerary = TransALU;
1394 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1395 inst, "RECIPSQRT_IEEE", []
1398 let Itinerary = TransALU;
1401 class SIN_Common <bits<11> inst> : R600_1OP <
1405 let Itinerary = TransALU;
1408 class COS_Common <bits<11> inst> : R600_1OP <
1412 let Itinerary = TransALU;
1415 //===----------------------------------------------------------------------===//
1416 // Helper patterns for complex intrinsics
1417 //===----------------------------------------------------------------------===//
1419 multiclass DIV_Common <InstR600 recip_ieee> {
1421 (int_AMDGPU_div f32:$src0, f32:$src1),
1422 (MUL_IEEE $src0, (recip_ieee $src1))
1426 (fdiv f32:$src0, f32:$src1),
1427 (MUL_IEEE $src0, (recip_ieee $src1))
1431 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee>
1433 (int_TGSI_lit_z f32:$src_x, f32:$src_y, f32:$src_w),
1434 (exp_ieee (mul_lit (log_clamped (MAX $src_y, (f32 ZERO))), $src_w, $src_x))
1437 //===----------------------------------------------------------------------===//
1438 // R600 / R700 Instructions
1439 //===----------------------------------------------------------------------===//
1441 let Predicates = [isR600] in {
1443 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1444 def MULADD_r600 : MULADD_Common<0x10>;
1445 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1446 def CNDE_r600 : CNDE_Common<0x18>;
1447 def CNDGT_r600 : CNDGT_Common<0x19>;
1448 def CNDGE_r600 : CNDGE_Common<0x1A>;
1449 def DOT4_r600 : DOT4_Common<0x50>;
1450 defm CUBE_r600 : CUBE_Common<0x52>;
1451 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1452 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1453 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1454 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1455 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1456 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1457 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1458 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1459 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1460 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1461 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1462 def SIN_r600 : SIN_Common<0x6E>;
1463 def COS_r600 : COS_Common<0x6F>;
1464 def ASHR_r600 : ASHR_Common<0x70>;
1465 def LSHR_r600 : LSHR_Common<0x71>;
1466 def LSHL_r600 : LSHL_Common<0x72>;
1467 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1468 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1469 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1470 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1471 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1473 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1474 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL>;
1475 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1477 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
1479 def R600_ExportSwz : ExportSwzInst {
1480 let Word1{20-17} = 0; // BURST_COUNT
1481 let Word1{21} = eop;
1482 let Word1{22} = 1; // VALID_PIXEL_MODE
1483 let Word1{30-23} = inst;
1484 let Word1{31} = 1; // BARRIER
1486 defm : ExportPattern<R600_ExportSwz, 39>;
1488 def R600_ExportBuf : ExportBufInst {
1489 let Word1{20-17} = 0; // BURST_COUNT
1490 let Word1{21} = eop;
1491 let Word1{22} = 1; // VALID_PIXEL_MODE
1492 let Word1{30-23} = inst;
1493 let Word1{31} = 1; // BARRIER
1495 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1497 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1498 "TEX $COUNT @$ADDR"> {
1501 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1502 "VTX $COUNT @$ADDR"> {
1505 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1506 "LOOP_START_DX10 @$ADDR"> {
1510 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1514 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1515 "LOOP_BREAK @$ADDR"> {
1519 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1520 "CONTINUE @$ADDR"> {
1524 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1525 "JUMP @$ADDR POP:$POP_COUNT"> {
1528 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1529 "ELSE @$ADDR POP:$POP_COUNT"> {
1532 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1537 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1538 "POP @$ADDR POP:$POP_COUNT"> {
1541 def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
1545 let END_OF_PROGRAM = 1;
1550 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1552 class COS_PAT <InstR600 trig> : Pat<
1554 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1557 class SIN_PAT <InstR600 trig> : Pat<
1559 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1562 //===----------------------------------------------------------------------===//
1563 // R700 Only instructions
1564 //===----------------------------------------------------------------------===//
1566 let Predicates = [isR700] in {
1567 def SIN_r700 : SIN_Common<0x6E>;
1568 def COS_r700 : COS_Common<0x6F>;
1570 // R700 normalizes inputs to SIN/COS the same as EG
1571 def : SIN_PAT <SIN_r700>;
1572 def : COS_PAT <COS_r700>;
1575 //===----------------------------------------------------------------------===//
1576 // Evergreen Only instructions
1577 //===----------------------------------------------------------------------===//
1579 let Predicates = [isEG] in {
1581 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1582 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1584 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1585 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1586 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1587 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1588 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1589 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1590 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1591 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1592 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1593 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1594 def SIN_eg : SIN_Common<0x8D>;
1595 def COS_eg : COS_Common<0x8E>;
1597 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
1598 def : SIN_PAT <SIN_eg>;
1599 def : COS_PAT <COS_eg>;
1600 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
1601 } // End Predicates = [isEG]
1603 //===----------------------------------------------------------------------===//
1604 // Evergreen / Cayman Instructions
1605 //===----------------------------------------------------------------------===//
1607 let Predicates = [isEGorCayman] in {
1609 // BFE_UINT - bit_extract, an optimization for mask and shift
1614 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1619 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1620 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1621 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1622 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1623 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1624 [(set i32:$dst, (int_AMDIL_bit_extract_u32 i32:$src0, i32:$src1,
1628 def : BFEPattern <BFE_UINT_eg>;
1630 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>;
1631 defm : BFIPatterns <BFI_INT_eg>;
1633 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
1634 def : ROTRPattern <BIT_ALIGN_INT_eg>;
1636 def MULADD_eg : MULADD_Common<0x14>;
1637 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
1638 def ASHR_eg : ASHR_Common<0x15>;
1639 def LSHR_eg : LSHR_Common<0x16>;
1640 def LSHL_eg : LSHL_Common<0x17>;
1641 def CNDE_eg : CNDE_Common<0x19>;
1642 def CNDGT_eg : CNDGT_Common<0x1A>;
1643 def CNDGE_eg : CNDGE_Common<0x1B>;
1644 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1645 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1646 def DOT4_eg : DOT4_Common<0xBE>;
1647 defm CUBE_eg : CUBE_Common<0xC0>;
1649 let hasSideEffects = 1 in {
1650 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1653 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1655 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1659 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1661 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1665 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1667 // TRUNC is used for the FLT_TO_INT instructions to work around a
1668 // perceived problem where the rounding modes are applied differently
1669 // depending on the instruction and the slot they are in.
1671 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1672 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1674 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1675 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1676 // We should look into handling these cases separately.
1677 def : Pat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
1679 def : Pat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
1682 def : SHA256MaPattern <BFI_INT_eg, XOR_INT>;
1684 def EG_ExportSwz : ExportSwzInst {
1685 let Word1{19-16} = 0; // BURST_COUNT
1686 let Word1{20} = 1; // VALID_PIXEL_MODE
1687 let Word1{21} = eop;
1688 let Word1{29-22} = inst;
1689 let Word1{30} = 0; // MARK
1690 let Word1{31} = 1; // BARRIER
1692 defm : ExportPattern<EG_ExportSwz, 83>;
1694 def EG_ExportBuf : ExportBufInst {
1695 let Word1{19-16} = 0; // BURST_COUNT
1696 let Word1{20} = 1; // VALID_PIXEL_MODE
1697 let Word1{21} = eop;
1698 let Word1{29-22} = inst;
1699 let Word1{30} = 0; // MARK
1700 let Word1{31} = 1; // BARRIER
1702 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1704 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1705 "TEX $COUNT @$ADDR"> {
1708 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1709 "VTX $COUNT @$ADDR"> {
1712 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
1713 "LOOP_START_DX10 @$ADDR"> {
1717 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1721 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
1722 "LOOP_BREAK @$ADDR"> {
1726 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
1727 "CONTINUE @$ADDR"> {
1731 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1732 "JUMP @$ADDR POP:$POP_COUNT"> {
1735 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1736 "ELSE @$ADDR POP:$POP_COUNT"> {
1739 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
1744 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1745 "POP @$ADDR POP:$POP_COUNT"> {
1748 def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
1752 let END_OF_PROGRAM = 1;
1755 //===----------------------------------------------------------------------===//
1756 // Memory read/write instructions
1757 //===----------------------------------------------------------------------===//
1758 let usesCustomInserter = 1 in {
1760 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
1762 : EG_CF_RAT <0x57, 0x2, 0, (outs), ins, name, pattern> {
1764 // XXX: Have a separate instruction for non-indexed writes.
1770 let COMP_MASK = comp_mask;
1771 let BURST_COUNT = 0;
1777 } // End usesCustomInserter = 1
1780 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1781 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1782 0x1, "RAT_WRITE_CACHELESS_32_eg $rw_gpr, $index_gpr, $eop",
1783 [(global_store i32:$rw_gpr, i32:$index_gpr)]
1787 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1788 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1789 0xf, "RAT_WRITE_CACHELESS_128 $rw_gpr.XYZW, $index_gpr, $eop",
1790 [(global_store v4i32:$rw_gpr, i32:$index_gpr)]
1793 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1794 : InstR600ISA <outs, (ins MEMxi:$ptr), name, pattern>,
1795 VTX_WORD1_GPR, VTX_WORD0 {
1800 let FETCH_WHOLE_QUAD = 0;
1801 let BUFFER_ID = buffer_id;
1803 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1804 // to store vertex addresses in any channel, not just X.
1807 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
1808 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
1809 // however, based on my testing if USE_CONST_FIELDS is set, then all
1810 // these fields need to be set to 0.
1811 let USE_CONST_FIELDS = 0;
1812 let NUM_FORMAT_ALL = 1;
1813 let FORMAT_COMP_ALL = 0;
1814 let SRF_MODE_ALL = 0;
1816 let Inst{31-0} = Word0;
1817 let Inst{63-32} = Word1;
1818 // LLVM can only encode 64-bit instructions, so these fields are manually
1819 // encoded in R600CodeEmitter
1822 // bits<2> ENDIAN_SWAP = 0;
1823 // bits<1> CONST_BUF_NO_STRIDE = 0;
1824 // bits<1> MEGA_FETCH = 0;
1825 // bits<1> ALT_CONST = 0;
1826 // bits<2> BUFFER_INDEX_MODE = 0;
1830 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1831 // is done in R600CodeEmitter
1833 // Inst{79-64} = OFFSET;
1834 // Inst{81-80} = ENDIAN_SWAP;
1835 // Inst{82} = CONST_BUF_NO_STRIDE;
1836 // Inst{83} = MEGA_FETCH;
1837 // Inst{84} = ALT_CONST;
1838 // Inst{86-85} = BUFFER_INDEX_MODE;
1839 // Inst{95-86} = 0; Reserved
1841 // VTX_WORD3 (Padding)
1843 // Inst{127-96} = 0;
1848 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1849 : VTX_READ_eg <"VTX_READ_8 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1852 let MEGA_FETCH_COUNT = 1;
1854 let DST_SEL_Y = 7; // Masked
1855 let DST_SEL_Z = 7; // Masked
1856 let DST_SEL_W = 7; // Masked
1857 let DATA_FORMAT = 1; // FMT_8
1860 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1861 : VTX_READ_eg <"VTX_READ_16 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1863 let MEGA_FETCH_COUNT = 2;
1865 let DST_SEL_Y = 7; // Masked
1866 let DST_SEL_Z = 7; // Masked
1867 let DST_SEL_W = 7; // Masked
1868 let DATA_FORMAT = 5; // FMT_16
1872 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1873 : VTX_READ_eg <"VTX_READ_32 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1876 let MEGA_FETCH_COUNT = 4;
1878 let DST_SEL_Y = 7; // Masked
1879 let DST_SEL_Z = 7; // Masked
1880 let DST_SEL_W = 7; // Masked
1881 let DATA_FORMAT = 0xD; // COLOR_32
1883 // This is not really necessary, but there were some GPU hangs that appeared
1884 // to be caused by ALU instructions in the next instruction group that wrote
1885 // to the $ptr registers of the VTX_READ.
1887 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1888 // %T2_X<def> = MOV %ZERO
1889 //Adding this constraint prevents this from happening.
1890 let Constraints = "$ptr.ptr = $dst";
1893 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1894 : VTX_READ_eg <"VTX_READ_128 $dst.XYZW, $ptr", buffer_id, (outs R600_Reg128:$dst),
1897 let MEGA_FETCH_COUNT = 16;
1902 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1904 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1905 // that holds its buffer address to avoid potential hangs. We can't use
1906 // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
1907 // registers are different sizes.
1910 //===----------------------------------------------------------------------===//
1911 // VTX Read from parameter memory space
1912 //===----------------------------------------------------------------------===//
1914 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1915 [(set i32:$dst, (load_param_zexti8 ADDRVTX_READ:$ptr))]
1918 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1919 [(set i32:$dst, (load_param_zexti16 ADDRVTX_READ:$ptr))]
1922 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1923 [(set i32:$dst, (load_param ADDRVTX_READ:$ptr))]
1926 def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
1927 [(set v4i32:$dst, (load_param ADDRVTX_READ:$ptr))]
1930 //===----------------------------------------------------------------------===//
1931 // VTX Read from global memory space
1932 //===----------------------------------------------------------------------===//
1935 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1936 [(set i32:$dst, (zextloadi8_global ADDRVTX_READ:$ptr))]
1940 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1941 [(set i32:$dst, (global_load ADDRVTX_READ:$ptr))]
1945 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1946 [(set v4i32:$dst, (global_load ADDRVTX_READ:$ptr))]
1949 //===----------------------------------------------------------------------===//
1951 // XXX: We are currently storing all constants in the global address space.
1952 //===----------------------------------------------------------------------===//
1954 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1955 [(set i32:$dst, (constant_load ADDRVTX_READ:$ptr))]
1960 //===----------------------------------------------------------------------===//
1961 // Regist loads and stores - for indirect addressing
1962 //===----------------------------------------------------------------------===//
1964 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1966 let Predicates = [isCayman] in {
1968 let isVector = 1 in {
1970 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1972 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1973 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1974 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1975 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1976 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1977 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1978 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
1979 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1980 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1981 def SIN_cm : SIN_Common<0x8D>;
1982 def COS_cm : COS_Common<0x8E>;
1983 } // End isVector = 1
1985 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
1986 def : SIN_PAT <SIN_cm>;
1987 def : COS_PAT <COS_cm>;
1989 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1991 // RECIP_UINT emulation for Cayman
1992 // The multiplication scales from [0,1] to the unsigned integer range
1994 (AMDGPUurecip i32:$src0),
1995 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
1996 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
1999 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
2005 def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
2009 //===----------------------------------------------------------------------===//
2010 // Branch Instructions
2011 //===----------------------------------------------------------------------===//
2014 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
2015 "IF_PREDICATE_SET $src", []>;
2017 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
2018 "PREDICATED_BREAK $src", []>;
2020 //===----------------------------------------------------------------------===//
2021 // Pseudo instructions
2022 //===----------------------------------------------------------------------===//
2024 let isPseudo = 1 in {
2026 def PRED_X : InstR600 <
2027 (outs R600_Predicate_Bit:$dst),
2028 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
2030 let FlagOperandIdx = 3;
2033 let isTerminator = 1, isBranch = 1 in {
2034 def JUMP_COND : InstR600 <
2036 (ins brtarget:$target, R600_Predicate_Bit:$p),
2037 "JUMP $target ($p)",
2041 def JUMP : InstR600 <
2043 (ins brtarget:$target),
2048 let isPredicable = 1;
2052 } // End isTerminator = 1, isBranch = 1
2054 let usesCustomInserter = 1 in {
2056 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
2058 def MASK_WRITE : AMDGPUShaderInst <
2060 (ins R600_Reg32:$src),
2065 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
2069 (outs R600_Reg128:$dst),
2070 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
2071 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2072 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2073 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
2074 imm:$resourceId, imm:$samplerId, imm:$textureTarget))],
2079 def TXD_SHADOW: InstR600 <
2080 (outs R600_Reg128:$dst),
2081 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
2082 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2083 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2084 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
2085 imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))],
2090 } // End isPseudo = 1
2091 } // End usesCustomInserter = 1
2093 def CLAMP_R600 : CLAMP <R600_Reg32>;
2094 def FABS_R600 : FABS<R600_Reg32>;
2095 def FNEG_R600 : FNEG<R600_Reg32>;
2097 //===---------------------------------------------------------------------===//
2098 // Return instruction
2099 //===---------------------------------------------------------------------===//
2100 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
2101 usesCustomInserter = 1 in {
2102 def RETURN : ILFormat<(outs), (ins variable_ops),
2103 "RETURN", [(IL_retflag)]>;
2107 //===----------------------------------------------------------------------===//
2108 // Constant Buffer Addressing Support
2109 //===----------------------------------------------------------------------===//
2111 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
2112 def CONST_COPY : Instruction {
2113 let OutOperandList = (outs R600_Reg32:$dst);
2114 let InOperandList = (ins i32imm:$src);
2116 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
2117 let AsmString = "CONST_COPY";
2118 let neverHasSideEffects = 1;
2119 let isAsCheapAsAMove = 1;
2120 let Itinerary = NullALU;
2122 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
2124 def TEX_VTX_CONSTBUF :
2125 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
2126 [(set v4i32:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
2127 VTX_WORD1_GPR, VTX_WORD0 {
2131 let FETCH_WHOLE_QUAD = 0;
2135 let USE_CONST_FIELDS = 0;
2136 let NUM_FORMAT_ALL = 2;
2137 let FORMAT_COMP_ALL = 1;
2138 let SRF_MODE_ALL = 1;
2139 let MEGA_FETCH_COUNT = 16;
2144 let DATA_FORMAT = 35;
2146 let Inst{31-0} = Word0;
2147 let Inst{63-32} = Word1;
2149 // LLVM can only encode 64-bit instructions, so these fields are manually
2150 // encoded in R600CodeEmitter
2153 // bits<2> ENDIAN_SWAP = 0;
2154 // bits<1> CONST_BUF_NO_STRIDE = 0;
2155 // bits<1> MEGA_FETCH = 0;
2156 // bits<1> ALT_CONST = 0;
2157 // bits<2> BUFFER_INDEX_MODE = 0;
2161 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2162 // is done in R600CodeEmitter
2164 // Inst{79-64} = OFFSET;
2165 // Inst{81-80} = ENDIAN_SWAP;
2166 // Inst{82} = CONST_BUF_NO_STRIDE;
2167 // Inst{83} = MEGA_FETCH;
2168 // Inst{84} = ALT_CONST;
2169 // Inst{86-85} = BUFFER_INDEX_MODE;
2170 // Inst{95-86} = 0; Reserved
2172 // VTX_WORD3 (Padding)
2174 // Inst{127-96} = 0;
2179 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
2180 [(set v4f32:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
2181 VTX_WORD1_GPR, VTX_WORD0 {
2185 let FETCH_WHOLE_QUAD = 0;
2189 let USE_CONST_FIELDS = 1;
2190 let NUM_FORMAT_ALL = 0;
2191 let FORMAT_COMP_ALL = 0;
2192 let SRF_MODE_ALL = 1;
2193 let MEGA_FETCH_COUNT = 16;
2198 let DATA_FORMAT = 0;
2200 let Inst{31-0} = Word0;
2201 let Inst{63-32} = Word1;
2203 // LLVM can only encode 64-bit instructions, so these fields are manually
2204 // encoded in R600CodeEmitter
2207 // bits<2> ENDIAN_SWAP = 0;
2208 // bits<1> CONST_BUF_NO_STRIDE = 0;
2209 // bits<1> MEGA_FETCH = 0;
2210 // bits<1> ALT_CONST = 0;
2211 // bits<2> BUFFER_INDEX_MODE = 0;
2215 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2216 // is done in R600CodeEmitter
2218 // Inst{79-64} = OFFSET;
2219 // Inst{81-80} = ENDIAN_SWAP;
2220 // Inst{82} = CONST_BUF_NO_STRIDE;
2221 // Inst{83} = MEGA_FETCH;
2222 // Inst{84} = ALT_CONST;
2223 // Inst{86-85} = BUFFER_INDEX_MODE;
2224 // Inst{95-86} = 0; Reserved
2226 // VTX_WORD3 (Padding)
2228 // Inst{127-96} = 0;
2234 //===--------------------------------------------------------------------===//
2235 // Instructions support
2236 //===--------------------------------------------------------------------===//
2237 //===---------------------------------------------------------------------===//
2238 // Custom Inserter for Branches and returns, this eventually will be a
2240 //===---------------------------------------------------------------------===//
2241 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
2242 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
2243 "; Pseudo unconditional branch instruction",
2245 defm BRANCH_COND : BranchConditional<IL_brcond>;
2248 //===---------------------------------------------------------------------===//
2249 // Flow and Program control Instructions
2250 //===---------------------------------------------------------------------===//
2251 let isTerminator=1 in {
2252 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
2253 !strconcat("SWITCH", " $src"), []>;
2254 def CASE : ILFormat< (outs), (ins GPRI32:$src),
2255 !strconcat("CASE", " $src"), []>;
2256 def BREAK : ILFormat< (outs), (ins),
2258 def CONTINUE : ILFormat< (outs), (ins),
2260 def DEFAULT : ILFormat< (outs), (ins),
2262 def ELSE : ILFormat< (outs), (ins),
2264 def ENDSWITCH : ILFormat< (outs), (ins),
2266 def ENDMAIN : ILFormat< (outs), (ins),
2268 def END : ILFormat< (outs), (ins),
2270 def ENDFUNC : ILFormat< (outs), (ins),
2272 def ENDIF : ILFormat< (outs), (ins),
2274 def WHILELOOP : ILFormat< (outs), (ins),
2276 def ENDLOOP : ILFormat< (outs), (ins),
2278 def FUNC : ILFormat< (outs), (ins),
2280 def RETDYN : ILFormat< (outs), (ins),
2282 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2283 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
2284 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2285 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
2286 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2287 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
2288 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2289 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
2290 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2291 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
2292 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2293 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
2294 defm IFC : BranchInstr2<"IFC">;
2295 defm BREAKC : BranchInstr2<"BREAKC">;
2296 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
2299 //===----------------------------------------------------------------------===//
2301 //===----------------------------------------------------------------------===//
2303 // CND*_INT Pattterns for f32 True / False values
2305 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
2306 (selectcc i32:$src0, 0, f32:$src1, f32:$src2, cc),
2307 (cnd $src0, $src1, $src2)
2310 def : CND_INT_f32 <CNDE_INT, SETEQ>;
2311 def : CND_INT_f32 <CNDGT_INT, SETGT>;
2312 def : CND_INT_f32 <CNDGE_INT, SETGE>;
2314 //CNDGE_INT extra pattern
2316 (selectcc i32:$src0, -1, i32:$src1, i32:$src2, COND_GT),
2317 (CNDGE_INT $src0, $src1, $src2)
2323 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
2327 (int_AMDGPU_kill f32:$src0),
2328 (MASK_WRITE (KILLGT (f32 ZERO), $src0))
2333 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LT),
2339 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LE),
2343 // SETGT_DX10 reverse args
2345 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LT),
2346 (SETGT_DX10 $src1, $src0)
2349 // SETGE_DX10 reverse args
2351 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LE),
2352 (SETGE_DX10 $src1, $src0)
2355 // SETGT_INT reverse args
2357 (selectcc i32:$src0, i32:$src1, -1, 0, SETLT),
2358 (SETGT_INT $src1, $src0)
2361 // SETGE_INT reverse args
2363 (selectcc i32:$src0, i32:$src1, -1, 0, SETLE),
2364 (SETGE_INT $src1, $src0)
2367 // SETGT_UINT reverse args
2369 (selectcc i32:$src0, i32:$src1, -1, 0, SETULT),
2370 (SETGT_UINT $src1, $src0)
2373 // SETGE_UINT reverse args
2375 (selectcc i32:$src0, i32:$src1, -1, 0, SETULE),
2376 (SETGE_UINT $src1, $src0)
2379 // The next two patterns are special cases for handling 'true if ordered' and
2380 // 'true if unordered' conditionals. The assumption here is that the behavior of
2381 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
2383 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
2384 // We assume that SETE returns false when one of the operands is NAN and
2385 // SNE returns true when on of the operands is NAN
2387 //SETE - 'true if ordered'
2389 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETO),
2393 //SETE_DX10 - 'true if ordered'
2395 (selectcc f32:$src0, f32:$src1, -1, 0, SETO),
2396 (SETE_DX10 $src0, $src1)
2399 //SNE - 'true if unordered'
2401 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETUO),
2405 //SETNE_DX10 - 'true if ordered'
2407 (selectcc f32:$src0, f32:$src1, -1, 0, SETUO),
2408 (SETNE_DX10 $src0, $src1)
2411 def : Extract_Element <f32, v4f32, 0, sub0>;
2412 def : Extract_Element <f32, v4f32, 1, sub1>;
2413 def : Extract_Element <f32, v4f32, 2, sub2>;
2414 def : Extract_Element <f32, v4f32, 3, sub3>;
2416 def : Insert_Element <f32, v4f32, 0, sub0>;
2417 def : Insert_Element <f32, v4f32, 1, sub1>;
2418 def : Insert_Element <f32, v4f32, 2, sub2>;
2419 def : Insert_Element <f32, v4f32, 3, sub3>;
2421 def : Extract_Element <i32, v4i32, 0, sub0>;
2422 def : Extract_Element <i32, v4i32, 1, sub1>;
2423 def : Extract_Element <i32, v4i32, 2, sub2>;
2424 def : Extract_Element <i32, v4i32, 3, sub3>;
2426 def : Insert_Element <i32, v4i32, 0, sub0>;
2427 def : Insert_Element <i32, v4i32, 1, sub1>;
2428 def : Insert_Element <i32, v4i32, 2, sub2>;
2429 def : Insert_Element <i32, v4i32, 3, sub3>;
2431 def : Vector4_Build <v4f32, f32>;
2432 def : Vector4_Build <v4i32, i32>;
2434 // bitconvert patterns
2436 def : BitConvert <i32, f32, R600_Reg32>;
2437 def : BitConvert <f32, i32, R600_Reg32>;
2438 def : BitConvert <v4f32, v4i32, R600_Reg128>;
2439 def : BitConvert <v4i32, v4f32, R600_Reg128>;
2441 // DWORDADDR pattern
2442 def : DwordAddrPat <i32, R600_Reg32>;
2444 } // End isR600toCayman Predicate