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:$ptr),
547 [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
550 def load_param : LoadParamFrag<load>;
551 def load_param_zexti8 : LoadParamFrag<zextloadi8>;
552 def load_param_zexti16 : LoadParamFrag<zextloadi16>;
554 def isR600 : Predicate<"Subtarget.device()"
555 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX">;
556 def isR700 : Predicate<"Subtarget.device()"
557 "->getGeneration() == AMDGPUDeviceInfo::HD4XXX &&"
558 "Subtarget.device()->getDeviceFlag()"
559 ">= OCL_DEVICE_RV710">;
560 def isEG : Predicate<
561 "Subtarget.device()->getGeneration() >= AMDGPUDeviceInfo::HD5XXX && "
562 "Subtarget.device()->getGeneration() < AMDGPUDeviceInfo::HD7XXX && "
563 "Subtarget.device()->getDeviceFlag() != OCL_DEVICE_CAYMAN">;
565 def isCayman : Predicate<"Subtarget.device()"
566 "->getDeviceFlag() == OCL_DEVICE_CAYMAN">;
567 def isEGorCayman : Predicate<"Subtarget.device()"
568 "->getGeneration() == AMDGPUDeviceInfo::HD5XXX"
569 "|| Subtarget.device()->getGeneration() =="
570 "AMDGPUDeviceInfo::HD6XXX">;
572 def isR600toCayman : Predicate<
573 "Subtarget.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX">;
575 //===----------------------------------------------------------------------===//
577 //===----------------------------------------------------------------------===//
579 def INTERP_PAIR_XY : AMDGPUShaderInst <
580 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
581 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
582 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
585 def INTERP_PAIR_ZW : AMDGPUShaderInst <
586 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
587 (ins i32imm:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
588 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
591 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
592 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
596 def TEXTURE_FETCH_Type : SDTypeProfile<1, 19, [SDTCisFP<0>]>;
598 def TEXTURE_FETCH: SDNode<"AMDGPUISD::TEXTURE_FETCH", TEXTURE_FETCH_Type, []>;
600 multiclass TexPattern<bits<32> TextureOp, Instruction inst, ValueType vt = v4f32> {
601 def : Pat<(TEXTURE_FETCH (i32 TextureOp), vt:$SRC_GPR,
602 (i32 imm:$srcx), (i32 imm:$srcy), (i32 imm:$srcz), (i32 imm:$srcw),
603 (i32 imm:$offsetx), (i32 imm:$offsety), (i32 imm:$offsetz),
604 (i32 imm:$DST_SEL_X), (i32 imm:$DST_SEL_Y), (i32 imm:$DST_SEL_Z),
605 (i32 imm:$DST_SEL_W),
606 (i32 imm:$RESOURCE_ID), (i32 imm:$SAMPLER_ID),
607 (i32 imm:$COORD_TYPE_X), (i32 imm:$COORD_TYPE_Y), (i32 imm:$COORD_TYPE_Z),
608 (i32 imm:$COORD_TYPE_W)),
609 (inst R600_Reg128:$SRC_GPR,
610 imm:$srcx, imm:$srcy, imm:$srcz, imm:$srcw,
611 imm:$offsetx, imm:$offsety, imm:$offsetz,
612 imm:$DST_SEL_X, imm:$DST_SEL_Y, imm:$DST_SEL_Z,
614 imm:$RESOURCE_ID, imm:$SAMPLER_ID,
615 imm:$COORD_TYPE_X, imm:$COORD_TYPE_Y, imm:$COORD_TYPE_Z,
619 //===----------------------------------------------------------------------===//
620 // Interpolation Instructions
621 //===----------------------------------------------------------------------===//
623 def INTERP_VEC_LOAD : AMDGPUShaderInst <
624 (outs R600_Reg128:$dst),
626 "INTERP_LOAD $src0 : $dst",
629 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
630 let bank_swizzle = 5;
633 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
634 let bank_swizzle = 5;
637 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
639 //===----------------------------------------------------------------------===//
640 // Export Instructions
641 //===----------------------------------------------------------------------===//
643 def ExportType : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
645 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
646 [SDNPHasChain, SDNPSideEffect]>;
649 field bits<32> Word0;
656 let Word0{12-0} = arraybase;
657 let Word0{14-13} = type;
658 let Word0{21-15} = gpr;
659 let Word0{22} = 0; // RW_REL
660 let Word0{29-23} = 0; // INDEX_GPR
661 let Word0{31-30} = elem_size;
664 class ExportSwzWord1 {
665 field bits<32> Word1;
674 let Word1{2-0} = sw_x;
675 let Word1{5-3} = sw_y;
676 let Word1{8-6} = sw_z;
677 let Word1{11-9} = sw_w;
680 class ExportBufWord1 {
681 field bits<32> Word1;
688 let Word1{11-0} = arraySize;
689 let Word1{15-12} = compMask;
692 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
693 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
695 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
696 0, 61, 0, 7, 7, 7, cf_inst, 0)
699 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
701 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
702 0, 61, 7, 0, 7, 7, cf_inst, 0)
705 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
707 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
710 def : Pat<(int_R600_store_dummy 1),
712 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
715 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
716 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
717 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
718 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
723 multiclass SteamOutputExportPattern<Instruction ExportInst,
724 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
726 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
727 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
728 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
729 4095, imm:$mask, buf0inst, 0)>;
731 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
732 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
733 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
734 4095, imm:$mask, buf1inst, 0)>;
736 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
737 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
738 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
739 4095, imm:$mask, buf2inst, 0)>;
741 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
742 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
743 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
744 4095, imm:$mask, buf3inst, 0)>;
747 // Export Instructions should not be duplicated by TailDuplication pass
748 // (which assumes that duplicable instruction are affected by exec mask)
749 let usesCustomInserter = 1, isNotDuplicable = 1 in {
751 class ExportSwzInst : InstR600ISA<(
753 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
754 i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
756 !strconcat("EXPORT", " $gpr"),
757 []>, ExportWord0, ExportSwzWord1 {
759 let Inst{31-0} = Word0;
760 let Inst{63-32} = Word1;
763 } // End usesCustomInserter = 1
765 class ExportBufInst : InstR600ISA<(
767 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
768 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
769 !strconcat("EXPORT", " $gpr"),
770 []>, ExportWord0, ExportBufWord1 {
772 let Inst{31-0} = Word0;
773 let Inst{63-32} = Word1;
776 //===----------------------------------------------------------------------===//
777 // Control Flow Instructions
778 //===----------------------------------------------------------------------===//
781 field bits<32> Word0;
784 bits<4> KCACHE_BANK0;
785 bits<4> KCACHE_BANK1;
786 bits<2> KCACHE_MODE0;
788 let Word0{21-0} = ADDR;
789 let Word0{25-22} = KCACHE_BANK0;
790 let Word0{29-26} = KCACHE_BANK1;
791 let Word0{31-30} = KCACHE_MODE0;
795 field bits<32> Word1;
797 bits<2> KCACHE_MODE1;
798 bits<8> KCACHE_ADDR0;
799 bits<8> KCACHE_ADDR1;
803 bits<1> WHOLE_QUAD_MODE;
806 let Word1{1-0} = KCACHE_MODE1;
807 let Word1{9-2} = KCACHE_ADDR0;
808 let Word1{17-10} = KCACHE_ADDR1;
809 let Word1{24-18} = COUNT;
810 let Word1{25} = ALT_CONST;
811 let Word1{29-26} = CF_INST;
812 let Word1{30} = WHOLE_QUAD_MODE;
813 let Word1{31} = BARRIER;
816 def KCACHE : InstFlag<"printKCache">;
818 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
819 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1,
820 KCACHE:$KCACHE_MODE0, KCACHE:$KCACHE_MODE1,
821 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1,
823 !strconcat(OpName, " $COUNT, @$ADDR, "
824 "KC0[$KCACHE_MODE0], KC1[$KCACHE_MODE1]"),
825 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
830 let WHOLE_QUAD_MODE = 0;
833 let Inst{31-0} = Word0;
834 let Inst{63-32} = Word1;
837 class CF_WORD0_R600 {
838 field bits<32> Word0;
845 class CF_WORD1_R600 {
846 field bits<32> Word1;
854 bits<1> END_OF_PROGRAM;
855 bits<1> VALID_PIXEL_MODE;
857 bits<1> WHOLE_QUAD_MODE;
860 let Word1{2-0} = POP_COUNT;
861 let Word1{7-3} = CF_CONST;
862 let Word1{9-8} = COND;
863 let Word1{12-10} = COUNT;
864 let Word1{18-13} = CALL_COUNT;
865 let Word1{19} = COUNT_3;
866 let Word1{21} = END_OF_PROGRAM;
867 let Word1{22} = VALID_PIXEL_MODE;
868 let Word1{29-23} = CF_INST;
869 let Word1{30} = WHOLE_QUAD_MODE;
870 let Word1{31} = BARRIER;
873 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
874 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
880 let VALID_PIXEL_MODE = 0;
884 let END_OF_PROGRAM = 0;
885 let WHOLE_QUAD_MODE = 0;
887 let Inst{31-0} = Word0;
888 let Inst{63-32} = Word1;
892 field bits<32> Word0;
895 bits<3> JUMPTABLE_SEL;
897 let Word0{23-0} = ADDR;
898 let Word0{26-24} = JUMPTABLE_SEL;
902 field bits<32> Word1;
908 bits<1> VALID_PIXEL_MODE;
909 bits<1> END_OF_PROGRAM;
913 let Word1{2-0} = POP_COUNT;
914 let Word1{7-3} = CF_CONST;
915 let Word1{9-8} = COND;
916 let Word1{15-10} = COUNT;
917 let Word1{20} = VALID_PIXEL_MODE;
918 let Word1{21} = END_OF_PROGRAM;
919 let Word1{29-22} = CF_INST;
920 let Word1{31} = BARRIER;
923 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
924 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
929 let JUMPTABLE_SEL = 0;
931 let VALID_PIXEL_MODE = 0;
933 let END_OF_PROGRAM = 0;
935 let Inst{31-0} = Word0;
936 let Inst{63-32} = Word1;
939 def CF_ALU : ALU_CLAUSE<8, "ALU">;
940 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
942 def FETCH_CLAUSE : AMDGPUInst <(outs),
943 (ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
949 def ALU_CLAUSE : AMDGPUInst <(outs),
950 (ins i32imm:$addr), "ALU clause starting at $addr:", [] > {
956 def LITERALS : AMDGPUInst <(outs),
957 (ins LITERAL:$literal1, LITERAL:$literal2), "$literal1, $literal2", [] > {
962 let Inst{31-0} = literal1;
963 let Inst{63-32} = literal2;
966 def PAD : AMDGPUInst <(outs), (ins), "PAD", [] > {
970 let Predicates = [isR600toCayman] in {
972 //===----------------------------------------------------------------------===//
973 // Common Instructions R600, R700, Evergreen, Cayman
974 //===----------------------------------------------------------------------===//
976 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
977 // Non-IEEE MUL: 0 * anything = 0
978 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
979 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
980 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
981 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
983 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
984 // so some of the instruction names don't match the asm string.
985 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
986 def SETE : R600_2OP <
988 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_EQ))]
993 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GT))]
998 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GE))]
1001 def SNE : R600_2OP <
1003 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_NE))]
1006 def SETE_DX10 : R600_2OP <
1008 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_EQ))]
1011 def SETGT_DX10 : R600_2OP <
1013 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GT))]
1016 def SETGE_DX10 : R600_2OP <
1018 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GE))]
1021 def SETNE_DX10 : R600_2OP <
1023 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_NE))]
1026 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
1027 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
1028 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
1029 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
1030 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
1032 def MOV : R600_1OP <0x19, "MOV", []>;
1034 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
1036 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
1037 (outs R600_Reg32:$dst),
1043 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
1045 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
1048 (MOV_IMM_I32 imm:$val)
1051 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
1054 (MOV_IMM_F32 fpimm:$val)
1057 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
1058 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
1059 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
1060 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
1062 let hasSideEffects = 1 in {
1064 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
1066 } // end hasSideEffects
1068 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
1069 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
1070 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
1071 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
1072 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
1073 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
1074 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
1075 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
1076 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
1077 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
1079 def SETE_INT : R600_2OP <
1081 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETEQ))]
1084 def SETGT_INT : R600_2OP <
1086 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGT))]
1089 def SETGE_INT : R600_2OP <
1091 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGE))]
1094 def SETNE_INT : R600_2OP <
1096 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETNE))]
1099 def SETGT_UINT : R600_2OP <
1101 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGT))]
1104 def SETGE_UINT : R600_2OP <
1106 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGE))]
1109 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
1110 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
1111 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
1112 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
1114 def CNDE_INT : R600_3OP <
1116 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_EQ))]
1119 def CNDGE_INT : R600_3OP <
1121 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GE))]
1124 def CNDGT_INT : R600_3OP <
1126 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GT))]
1129 //===----------------------------------------------------------------------===//
1130 // Texture instructions
1131 //===----------------------------------------------------------------------===//
1133 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1135 class R600_TEX <bits<11> inst, string opName> :
1136 InstR600 <(outs R600_Reg128:$DST_GPR),
1137 (ins R600_Reg128:$SRC_GPR,
1138 RSel:$srcx, RSel:$srcy, RSel:$srcz, RSel:$srcw,
1139 i32imm:$offsetx, i32imm:$offsety, i32imm:$offsetz,
1140 RSel:$DST_SEL_X, RSel:$DST_SEL_Y, RSel:$DST_SEL_Z, RSel:$DST_SEL_W,
1141 i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
1142 CT:$COORD_TYPE_X, CT:$COORD_TYPE_Y, CT:$COORD_TYPE_Z,
1145 " $DST_GPR.$DST_SEL_X$DST_SEL_Y$DST_SEL_Z$DST_SEL_W, "
1146 "$SRC_GPR.$srcx$srcy$srcz$srcw "
1147 "RID:$RESOURCE_ID SID:$SAMPLER_ID "
1148 "CT:$COORD_TYPE_X$COORD_TYPE_Y$COORD_TYPE_Z$COORD_TYPE_W"),
1150 NullALU>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
1151 let Inst{31-0} = Word0;
1152 let Inst{63-32} = Word1;
1154 let TEX_INST = inst{4-0};
1160 let FETCH_WHOLE_QUAD = 0;
1162 let SAMPLER_INDEX_MODE = 0;
1163 let RESOURCE_INDEX_MODE = 0;
1168 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1172 def TEX_SAMPLE : R600_TEX <0x10, "TEX_SAMPLE">;
1173 def TEX_SAMPLE_C : R600_TEX <0x18, "TEX_SAMPLE_C">;
1174 def TEX_SAMPLE_L : R600_TEX <0x11, "TEX_SAMPLE_L">;
1175 def TEX_SAMPLE_C_L : R600_TEX <0x19, "TEX_SAMPLE_C_L">;
1176 def TEX_SAMPLE_LB : R600_TEX <0x12, "TEX_SAMPLE_LB">;
1177 def TEX_SAMPLE_C_LB : R600_TEX <0x1A, "TEX_SAMPLE_C_LB">;
1178 def TEX_LD : R600_TEX <0x03, "TEX_LD">;
1179 def TEX_GET_TEXTURE_RESINFO : R600_TEX <0x04, "TEX_GET_TEXTURE_RESINFO">;
1180 def TEX_GET_GRADIENTS_H : R600_TEX <0x07, "TEX_GET_GRADIENTS_H">;
1181 def TEX_GET_GRADIENTS_V : R600_TEX <0x08, "TEX_GET_GRADIENTS_V">;
1182 def TEX_SET_GRADIENTS_H : R600_TEX <0x0B, "TEX_SET_GRADIENTS_H">;
1183 def TEX_SET_GRADIENTS_V : R600_TEX <0x0C, "TEX_SET_GRADIENTS_V">;
1184 def TEX_SAMPLE_G : R600_TEX <0x14, "TEX_SAMPLE_G">;
1185 def TEX_SAMPLE_C_G : R600_TEX <0x1C, "TEX_SAMPLE_C_G">;
1187 defm : TexPattern<0, TEX_SAMPLE>;
1188 defm : TexPattern<1, TEX_SAMPLE_C>;
1189 defm : TexPattern<2, TEX_SAMPLE_L>;
1190 defm : TexPattern<3, TEX_SAMPLE_C_L>;
1191 defm : TexPattern<4, TEX_SAMPLE_LB>;
1192 defm : TexPattern<5, TEX_SAMPLE_C_LB>;
1193 defm : TexPattern<6, TEX_LD, v4i32>;
1194 defm : TexPattern<7, TEX_GET_TEXTURE_RESINFO, v4i32>;
1195 defm : TexPattern<8, TEX_GET_GRADIENTS_H>;
1196 defm : TexPattern<9, TEX_GET_GRADIENTS_V>;
1198 //===----------------------------------------------------------------------===//
1199 // Helper classes for common instructions
1200 //===----------------------------------------------------------------------===//
1202 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
1207 class MULADD_Common <bits<5> inst> : R600_3OP <
1212 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
1213 inst, "MULADD_IEEE",
1214 [(set f32:$dst, (fadd (fmul f32:$src0, f32:$src1), f32:$src2))]
1217 class CNDE_Common <bits<5> inst> : R600_3OP <
1219 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_EQ))]
1222 class CNDGT_Common <bits<5> inst> : R600_3OP <
1224 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GT))]
1227 class CNDGE_Common <bits<5> inst> : R600_3OP <
1229 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GE))]
1232 multiclass DOT4_Common <bits<11> inst> {
1234 def _pseudo : R600_REDUCTION <inst,
1235 (ins R600_Reg128:$src0, R600_Reg128:$src1),
1236 "DOT4 $dst $src0, $src1",
1237 [(set f32:$dst, (int_AMDGPU_dp4 v4f32:$src0, v4f32:$src1))]
1240 def _real : R600_2OP <inst, "DOT4", []>;
1243 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1244 multiclass CUBE_Common <bits<11> inst> {
1246 def _pseudo : InstR600 <
1247 (outs R600_Reg128:$dst),
1248 (ins R600_Reg128:$src),
1250 [(set v4f32:$dst, (int_AMDGPU_cube v4f32:$src))],
1256 def _real : R600_2OP <inst, "CUBE", []>;
1258 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1260 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1261 inst, "EXP_IEEE", fexp2
1264 let Itinerary = TransALU;
1267 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
1268 inst, "FLT_TO_INT", fp_to_sint
1271 let Itinerary = TransALU;
1274 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1275 inst, "INT_TO_FLT", sint_to_fp
1278 let Itinerary = TransALU;
1281 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1282 inst, "FLT_TO_UINT", fp_to_uint
1285 let Itinerary = TransALU;
1288 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1289 inst, "UINT_TO_FLT", uint_to_fp
1292 let Itinerary = TransALU;
1295 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1296 inst, "LOG_CLAMPED", []
1299 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1300 inst, "LOG_IEEE", flog2
1303 let Itinerary = TransALU;
1306 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1307 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1308 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1309 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1310 inst, "MULHI_INT", mulhs
1313 let Itinerary = TransALU;
1315 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1316 inst, "MULHI", mulhu
1319 let Itinerary = TransALU;
1321 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1322 inst, "MULLO_INT", mul
1325 let Itinerary = TransALU;
1327 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []> {
1329 let Itinerary = TransALU;
1332 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1333 inst, "RECIP_CLAMPED", []
1336 let Itinerary = TransALU;
1339 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1340 inst, "RECIP_IEEE", [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
1343 let Itinerary = TransALU;
1346 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1347 inst, "RECIP_UINT", AMDGPUurecip
1350 let Itinerary = TransALU;
1353 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1354 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1357 let Itinerary = TransALU;
1360 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1361 inst, "RECIPSQRT_IEEE", []
1364 let Itinerary = TransALU;
1367 class SIN_Common <bits<11> inst> : R600_1OP <
1371 let Itinerary = TransALU;
1374 class COS_Common <bits<11> inst> : R600_1OP <
1378 let Itinerary = TransALU;
1381 //===----------------------------------------------------------------------===//
1382 // Helper patterns for complex intrinsics
1383 //===----------------------------------------------------------------------===//
1385 multiclass DIV_Common <InstR600 recip_ieee> {
1387 (int_AMDGPU_div f32:$src0, f32:$src1),
1388 (MUL_IEEE $src0, (recip_ieee $src1))
1392 (fdiv f32:$src0, f32:$src1),
1393 (MUL_IEEE $src0, (recip_ieee $src1))
1397 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee>
1399 (int_TGSI_lit_z f32:$src_x, f32:$src_y, f32:$src_w),
1400 (exp_ieee (mul_lit (log_clamped (MAX $src_y, (f32 ZERO))), $src_w, $src_x))
1403 //===----------------------------------------------------------------------===//
1404 // R600 / R700 Instructions
1405 //===----------------------------------------------------------------------===//
1407 let Predicates = [isR600] in {
1409 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1410 def MULADD_r600 : MULADD_Common<0x10>;
1411 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1412 def CNDE_r600 : CNDE_Common<0x18>;
1413 def CNDGT_r600 : CNDGT_Common<0x19>;
1414 def CNDGE_r600 : CNDGE_Common<0x1A>;
1415 defm DOT4_r600 : DOT4_Common<0x50>;
1416 defm CUBE_r600 : CUBE_Common<0x52>;
1417 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1418 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1419 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1420 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1421 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1422 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1423 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1424 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1425 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1426 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1427 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1428 def SIN_r600 : SIN_Common<0x6E>;
1429 def COS_r600 : COS_Common<0x6F>;
1430 def ASHR_r600 : ASHR_Common<0x70>;
1431 def LSHR_r600 : LSHR_Common<0x71>;
1432 def LSHL_r600 : LSHL_Common<0x72>;
1433 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1434 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1435 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1436 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1437 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1439 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1440 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL>;
1441 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1443 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
1445 def R600_ExportSwz : ExportSwzInst {
1446 let Word1{20-17} = 0; // BURST_COUNT
1447 let Word1{21} = eop;
1448 let Word1{22} = 1; // VALID_PIXEL_MODE
1449 let Word1{30-23} = inst;
1450 let Word1{31} = 1; // BARRIER
1452 defm : ExportPattern<R600_ExportSwz, 39>;
1454 def R600_ExportBuf : ExportBufInst {
1455 let Word1{20-17} = 0; // BURST_COUNT
1456 let Word1{21} = eop;
1457 let Word1{22} = 1; // VALID_PIXEL_MODE
1458 let Word1{30-23} = inst;
1459 let Word1{31} = 1; // BARRIER
1461 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1463 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1464 "TEX $COUNT @$ADDR"> {
1467 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1468 "VTX $COUNT @$ADDR"> {
1471 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1472 "LOOP_START_DX10 @$ADDR"> {
1476 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1480 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1481 "LOOP_BREAK @$ADDR"> {
1485 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1486 "CONTINUE @$ADDR"> {
1490 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1491 "JUMP @$ADDR POP:$POP_COUNT"> {
1494 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1495 "ELSE @$ADDR POP:$POP_COUNT"> {
1498 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1503 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1504 "POP @$ADDR POP:$POP_COUNT"> {
1507 def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
1511 let END_OF_PROGRAM = 1;
1516 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1518 class COS_PAT <InstR600 trig> : Pat<
1520 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1523 class SIN_PAT <InstR600 trig> : Pat<
1525 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1528 //===----------------------------------------------------------------------===//
1529 // R700 Only instructions
1530 //===----------------------------------------------------------------------===//
1532 let Predicates = [isR700] in {
1533 def SIN_r700 : SIN_Common<0x6E>;
1534 def COS_r700 : COS_Common<0x6F>;
1536 // R700 normalizes inputs to SIN/COS the same as EG
1537 def : SIN_PAT <SIN_r700>;
1538 def : COS_PAT <COS_r700>;
1541 //===----------------------------------------------------------------------===//
1542 // Evergreen Only instructions
1543 //===----------------------------------------------------------------------===//
1545 let Predicates = [isEG] in {
1547 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1548 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1550 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1551 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1552 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1553 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1554 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1555 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1556 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1557 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1558 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1559 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1560 def SIN_eg : SIN_Common<0x8D>;
1561 def COS_eg : COS_Common<0x8E>;
1563 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
1564 def : SIN_PAT <SIN_eg>;
1565 def : COS_PAT <COS_eg>;
1566 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
1567 } // End Predicates = [isEG]
1569 //===----------------------------------------------------------------------===//
1570 // Evergreen / Cayman Instructions
1571 //===----------------------------------------------------------------------===//
1573 let Predicates = [isEGorCayman] in {
1575 // BFE_UINT - bit_extract, an optimization for mask and shift
1580 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1585 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1586 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1587 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1588 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1589 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1590 [(set i32:$dst, (int_AMDIL_bit_extract_u32 i32:$src0, i32:$src1,
1594 def : BFEPattern <BFE_UINT_eg>;
1596 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>;
1597 defm : BFIPatterns <BFI_INT_eg>;
1599 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
1600 [(set i32:$dst, (AMDGPUbitalign i32:$src0, i32:$src1, i32:$src2))],
1604 def MULADD_eg : MULADD_Common<0x14>;
1605 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
1606 def ASHR_eg : ASHR_Common<0x15>;
1607 def LSHR_eg : LSHR_Common<0x16>;
1608 def LSHL_eg : LSHL_Common<0x17>;
1609 def CNDE_eg : CNDE_Common<0x19>;
1610 def CNDGT_eg : CNDGT_Common<0x1A>;
1611 def CNDGE_eg : CNDGE_Common<0x1B>;
1612 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1613 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1614 defm DOT4_eg : DOT4_Common<0xBE>;
1615 defm CUBE_eg : CUBE_Common<0xC0>;
1617 let hasSideEffects = 1 in {
1618 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1621 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1623 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1627 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1629 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1633 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1635 // TRUNC is used for the FLT_TO_INT instructions to work around a
1636 // perceived problem where the rounding modes are applied differently
1637 // depending on the instruction and the slot they are in.
1639 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1640 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1642 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1643 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1644 // We should look into handling these cases separately.
1645 def : Pat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
1647 def : Pat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
1650 def : SHA256MaPattern <BFI_INT_eg, XOR_INT>;
1652 def EG_ExportSwz : ExportSwzInst {
1653 let Word1{19-16} = 0; // BURST_COUNT
1654 let Word1{20} = 1; // VALID_PIXEL_MODE
1655 let Word1{21} = eop;
1656 let Word1{29-22} = inst;
1657 let Word1{30} = 0; // MARK
1658 let Word1{31} = 1; // BARRIER
1660 defm : ExportPattern<EG_ExportSwz, 83>;
1662 def EG_ExportBuf : ExportBufInst {
1663 let Word1{19-16} = 0; // BURST_COUNT
1664 let Word1{20} = 1; // VALID_PIXEL_MODE
1665 let Word1{21} = eop;
1666 let Word1{29-22} = inst;
1667 let Word1{30} = 0; // MARK
1668 let Word1{31} = 1; // BARRIER
1670 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1672 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1673 "TEX $COUNT @$ADDR"> {
1676 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1677 "VTX $COUNT @$ADDR"> {
1680 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
1681 "LOOP_START_DX10 @$ADDR"> {
1685 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1689 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
1690 "LOOP_BREAK @$ADDR"> {
1694 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
1695 "CONTINUE @$ADDR"> {
1699 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1700 "JUMP @$ADDR POP:$POP_COUNT"> {
1703 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1704 "ELSE @$ADDR POP:$POP_COUNT"> {
1707 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
1712 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1713 "POP @$ADDR POP:$POP_COUNT"> {
1716 def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
1720 let END_OF_PROGRAM = 1;
1723 //===----------------------------------------------------------------------===//
1724 // Memory read/write instructions
1725 //===----------------------------------------------------------------------===//
1726 let usesCustomInserter = 1 in {
1728 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
1730 : EG_CF_RAT <0x57, 0x2, 0, (outs), ins, name, pattern> {
1732 // XXX: Have a separate instruction for non-indexed writes.
1738 let COMP_MASK = comp_mask;
1739 let BURST_COUNT = 0;
1745 } // End usesCustomInserter = 1
1748 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1749 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1750 0x1, "RAT_WRITE_CACHELESS_32_eg $rw_gpr, $index_gpr, $eop",
1751 [(global_store i32:$rw_gpr, i32:$index_gpr)]
1755 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1756 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1757 0xf, "RAT_WRITE_CACHELESS_128 $rw_gpr.XYZW, $index_gpr, $eop",
1758 [(global_store v4i32:$rw_gpr, i32:$index_gpr)]
1761 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1762 : InstR600ISA <outs, (ins MEMxi:$ptr), name, pattern>,
1763 VTX_WORD1_GPR, VTX_WORD0 {
1768 let FETCH_WHOLE_QUAD = 0;
1769 let BUFFER_ID = buffer_id;
1771 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1772 // to store vertex addresses in any channel, not just X.
1775 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
1776 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
1777 // however, based on my testing if USE_CONST_FIELDS is set, then all
1778 // these fields need to be set to 0.
1779 let USE_CONST_FIELDS = 0;
1780 let NUM_FORMAT_ALL = 1;
1781 let FORMAT_COMP_ALL = 0;
1782 let SRF_MODE_ALL = 0;
1784 let Inst{31-0} = Word0;
1785 let Inst{63-32} = Word1;
1786 // LLVM can only encode 64-bit instructions, so these fields are manually
1787 // encoded in R600CodeEmitter
1790 // bits<2> ENDIAN_SWAP = 0;
1791 // bits<1> CONST_BUF_NO_STRIDE = 0;
1792 // bits<1> MEGA_FETCH = 0;
1793 // bits<1> ALT_CONST = 0;
1794 // bits<2> BUFFER_INDEX_MODE = 0;
1798 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1799 // is done in R600CodeEmitter
1801 // Inst{79-64} = OFFSET;
1802 // Inst{81-80} = ENDIAN_SWAP;
1803 // Inst{82} = CONST_BUF_NO_STRIDE;
1804 // Inst{83} = MEGA_FETCH;
1805 // Inst{84} = ALT_CONST;
1806 // Inst{86-85} = BUFFER_INDEX_MODE;
1807 // Inst{95-86} = 0; Reserved
1809 // VTX_WORD3 (Padding)
1811 // Inst{127-96} = 0;
1816 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1817 : VTX_READ_eg <"VTX_READ_8 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1820 let MEGA_FETCH_COUNT = 1;
1822 let DST_SEL_Y = 7; // Masked
1823 let DST_SEL_Z = 7; // Masked
1824 let DST_SEL_W = 7; // Masked
1825 let DATA_FORMAT = 1; // FMT_8
1828 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1829 : VTX_READ_eg <"VTX_READ_16 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1831 let MEGA_FETCH_COUNT = 2;
1833 let DST_SEL_Y = 7; // Masked
1834 let DST_SEL_Z = 7; // Masked
1835 let DST_SEL_W = 7; // Masked
1836 let DATA_FORMAT = 5; // FMT_16
1840 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1841 : VTX_READ_eg <"VTX_READ_32 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1844 let MEGA_FETCH_COUNT = 4;
1846 let DST_SEL_Y = 7; // Masked
1847 let DST_SEL_Z = 7; // Masked
1848 let DST_SEL_W = 7; // Masked
1849 let DATA_FORMAT = 0xD; // COLOR_32
1851 // This is not really necessary, but there were some GPU hangs that appeared
1852 // to be caused by ALU instructions in the next instruction group that wrote
1853 // to the $ptr registers of the VTX_READ.
1855 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1856 // %T2_X<def> = MOV %ZERO
1857 //Adding this constraint prevents this from happening.
1858 let Constraints = "$ptr.ptr = $dst";
1861 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1862 : VTX_READ_eg <"VTX_READ_128 $dst.XYZW, $ptr", buffer_id, (outs R600_Reg128:$dst),
1865 let MEGA_FETCH_COUNT = 16;
1870 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1872 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1873 // that holds its buffer address to avoid potential hangs. We can't use
1874 // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
1875 // registers are different sizes.
1878 //===----------------------------------------------------------------------===//
1879 // VTX Read from parameter memory space
1880 //===----------------------------------------------------------------------===//
1882 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1883 [(set i32:$dst, (load_param_zexti8 ADDRVTX_READ:$ptr))]
1886 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1887 [(set i32:$dst, (load_param_zexti16 ADDRVTX_READ:$ptr))]
1890 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1891 [(set i32:$dst, (load_param ADDRVTX_READ:$ptr))]
1894 def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
1895 [(set v4i32:$dst, (load_param ADDRVTX_READ:$ptr))]
1898 //===----------------------------------------------------------------------===//
1899 // VTX Read from global memory space
1900 //===----------------------------------------------------------------------===//
1903 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1904 [(set i32:$dst, (zextloadi8_global ADDRVTX_READ:$ptr))]
1908 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1909 [(set i32:$dst, (global_load ADDRVTX_READ:$ptr))]
1913 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1914 [(set v4i32:$dst, (global_load ADDRVTX_READ:$ptr))]
1917 //===----------------------------------------------------------------------===//
1919 // XXX: We are currently storing all constants in the global address space.
1920 //===----------------------------------------------------------------------===//
1922 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1923 [(set i32:$dst, (constant_load ADDRVTX_READ:$ptr))]
1928 //===----------------------------------------------------------------------===//
1929 // Regist loads and stores - for indirect addressing
1930 //===----------------------------------------------------------------------===//
1932 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1934 let Predicates = [isCayman] in {
1936 let isVector = 1 in {
1938 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1940 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1941 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1942 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1943 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1944 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1945 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1946 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
1947 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1948 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1949 def SIN_cm : SIN_Common<0x8D>;
1950 def COS_cm : COS_Common<0x8E>;
1951 } // End isVector = 1
1953 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
1954 def : SIN_PAT <SIN_cm>;
1955 def : COS_PAT <COS_cm>;
1957 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1959 // RECIP_UINT emulation for Cayman
1960 // The multiplication scales from [0,1] to the unsigned integer range
1962 (AMDGPUurecip i32:$src0),
1963 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
1964 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
1967 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
1973 def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
1977 //===----------------------------------------------------------------------===//
1978 // Branch Instructions
1979 //===----------------------------------------------------------------------===//
1982 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
1983 "IF_PREDICATE_SET $src", []>;
1985 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
1986 "PREDICATED_BREAK $src", []>;
1988 //===----------------------------------------------------------------------===//
1989 // Pseudo instructions
1990 //===----------------------------------------------------------------------===//
1992 let isPseudo = 1 in {
1994 def PRED_X : InstR600 <
1995 (outs R600_Predicate_Bit:$dst),
1996 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
1998 let FlagOperandIdx = 3;
2001 let isTerminator = 1, isBranch = 1 in {
2002 def JUMP_COND : InstR600 <
2004 (ins brtarget:$target, R600_Predicate_Bit:$p),
2005 "JUMP $target ($p)",
2009 def JUMP : InstR600 <
2011 (ins brtarget:$target),
2016 let isPredicable = 1;
2020 } // End isTerminator = 1, isBranch = 1
2022 let usesCustomInserter = 1 in {
2024 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
2026 def MASK_WRITE : AMDGPUShaderInst <
2028 (ins R600_Reg32:$src),
2033 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
2037 (outs R600_Reg128:$dst),
2038 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
2039 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2040 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2041 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
2042 imm:$resourceId, imm:$samplerId, imm:$textureTarget))],
2047 def TXD_SHADOW: InstR600 <
2048 (outs R600_Reg128:$dst),
2049 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
2050 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2051 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2052 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
2053 imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))],
2058 } // End isPseudo = 1
2059 } // End usesCustomInserter = 1
2061 def CLAMP_R600 : CLAMP <R600_Reg32>;
2062 def FABS_R600 : FABS<R600_Reg32>;
2063 def FNEG_R600 : FNEG<R600_Reg32>;
2065 //===---------------------------------------------------------------------===//
2066 // Return instruction
2067 //===---------------------------------------------------------------------===//
2068 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
2069 usesCustomInserter = 1 in {
2070 def RETURN : ILFormat<(outs), (ins variable_ops),
2071 "RETURN", [(IL_retflag)]>;
2075 //===----------------------------------------------------------------------===//
2076 // Constant Buffer Addressing Support
2077 //===----------------------------------------------------------------------===//
2079 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
2080 def CONST_COPY : Instruction {
2081 let OutOperandList = (outs R600_Reg32:$dst);
2082 let InOperandList = (ins i32imm:$src);
2084 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
2085 let AsmString = "CONST_COPY";
2086 let neverHasSideEffects = 1;
2087 let isAsCheapAsAMove = 1;
2088 let Itinerary = NullALU;
2090 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
2092 def TEX_VTX_CONSTBUF :
2093 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
2094 [(set v4i32:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
2095 VTX_WORD1_GPR, VTX_WORD0 {
2099 let FETCH_WHOLE_QUAD = 0;
2103 let USE_CONST_FIELDS = 0;
2104 let NUM_FORMAT_ALL = 2;
2105 let FORMAT_COMP_ALL = 1;
2106 let SRF_MODE_ALL = 1;
2107 let MEGA_FETCH_COUNT = 16;
2112 let DATA_FORMAT = 35;
2114 let Inst{31-0} = Word0;
2115 let Inst{63-32} = Word1;
2117 // LLVM can only encode 64-bit instructions, so these fields are manually
2118 // encoded in R600CodeEmitter
2121 // bits<2> ENDIAN_SWAP = 0;
2122 // bits<1> CONST_BUF_NO_STRIDE = 0;
2123 // bits<1> MEGA_FETCH = 0;
2124 // bits<1> ALT_CONST = 0;
2125 // bits<2> BUFFER_INDEX_MODE = 0;
2129 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2130 // is done in R600CodeEmitter
2132 // Inst{79-64} = OFFSET;
2133 // Inst{81-80} = ENDIAN_SWAP;
2134 // Inst{82} = CONST_BUF_NO_STRIDE;
2135 // Inst{83} = MEGA_FETCH;
2136 // Inst{84} = ALT_CONST;
2137 // Inst{86-85} = BUFFER_INDEX_MODE;
2138 // Inst{95-86} = 0; Reserved
2140 // VTX_WORD3 (Padding)
2142 // Inst{127-96} = 0;
2147 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
2148 [(set v4f32:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
2149 VTX_WORD1_GPR, VTX_WORD0 {
2153 let FETCH_WHOLE_QUAD = 0;
2157 let USE_CONST_FIELDS = 1;
2158 let NUM_FORMAT_ALL = 0;
2159 let FORMAT_COMP_ALL = 0;
2160 let SRF_MODE_ALL = 1;
2161 let MEGA_FETCH_COUNT = 16;
2166 let DATA_FORMAT = 0;
2168 let Inst{31-0} = Word0;
2169 let Inst{63-32} = Word1;
2171 // LLVM can only encode 64-bit instructions, so these fields are manually
2172 // encoded in R600CodeEmitter
2175 // bits<2> ENDIAN_SWAP = 0;
2176 // bits<1> CONST_BUF_NO_STRIDE = 0;
2177 // bits<1> MEGA_FETCH = 0;
2178 // bits<1> ALT_CONST = 0;
2179 // bits<2> BUFFER_INDEX_MODE = 0;
2183 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2184 // is done in R600CodeEmitter
2186 // Inst{79-64} = OFFSET;
2187 // Inst{81-80} = ENDIAN_SWAP;
2188 // Inst{82} = CONST_BUF_NO_STRIDE;
2189 // Inst{83} = MEGA_FETCH;
2190 // Inst{84} = ALT_CONST;
2191 // Inst{86-85} = BUFFER_INDEX_MODE;
2192 // Inst{95-86} = 0; Reserved
2194 // VTX_WORD3 (Padding)
2196 // Inst{127-96} = 0;
2202 //===--------------------------------------------------------------------===//
2203 // Instructions support
2204 //===--------------------------------------------------------------------===//
2205 //===---------------------------------------------------------------------===//
2206 // Custom Inserter for Branches and returns, this eventually will be a
2208 //===---------------------------------------------------------------------===//
2209 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
2210 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
2211 "; Pseudo unconditional branch instruction",
2213 defm BRANCH_COND : BranchConditional<IL_brcond>;
2216 //===---------------------------------------------------------------------===//
2217 // Flow and Program control Instructions
2218 //===---------------------------------------------------------------------===//
2219 let isTerminator=1 in {
2220 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
2221 !strconcat("SWITCH", " $src"), []>;
2222 def CASE : ILFormat< (outs), (ins GPRI32:$src),
2223 !strconcat("CASE", " $src"), []>;
2224 def BREAK : ILFormat< (outs), (ins),
2226 def CONTINUE : ILFormat< (outs), (ins),
2228 def DEFAULT : ILFormat< (outs), (ins),
2230 def ELSE : ILFormat< (outs), (ins),
2232 def ENDSWITCH : ILFormat< (outs), (ins),
2234 def ENDMAIN : ILFormat< (outs), (ins),
2236 def END : ILFormat< (outs), (ins),
2238 def ENDFUNC : ILFormat< (outs), (ins),
2240 def ENDIF : ILFormat< (outs), (ins),
2242 def WHILELOOP : ILFormat< (outs), (ins),
2244 def ENDLOOP : ILFormat< (outs), (ins),
2246 def FUNC : ILFormat< (outs), (ins),
2248 def RETDYN : ILFormat< (outs), (ins),
2250 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2251 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
2252 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2253 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
2254 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2255 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
2256 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2257 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
2258 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2259 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
2260 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2261 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
2262 defm IFC : BranchInstr2<"IFC">;
2263 defm BREAKC : BranchInstr2<"BREAKC">;
2264 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
2267 //===----------------------------------------------------------------------===//
2269 //===----------------------------------------------------------------------===//
2271 // CND*_INT Pattterns for f32 True / False values
2273 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
2274 (selectcc i32:$src0, 0, f32:$src1, f32:$src2, cc),
2275 (cnd $src0, $src1, $src2)
2278 def : CND_INT_f32 <CNDE_INT, SETEQ>;
2279 def : CND_INT_f32 <CNDGT_INT, SETGT>;
2280 def : CND_INT_f32 <CNDGE_INT, SETGE>;
2282 //CNDGE_INT extra pattern
2284 (selectcc i32:$src0, -1, i32:$src1, i32:$src2, COND_GT),
2285 (CNDGE_INT $src0, $src1, $src2)
2291 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
2295 (int_AMDGPU_kill f32:$src0),
2296 (MASK_WRITE (KILLGT (f32 ZERO), $src0))
2301 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LT),
2307 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LE),
2311 // SETGT_DX10 reverse args
2313 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LT),
2314 (SETGT_DX10 $src1, $src0)
2317 // SETGE_DX10 reverse args
2319 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LE),
2320 (SETGE_DX10 $src1, $src0)
2323 // SETGT_INT reverse args
2325 (selectcc i32:$src0, i32:$src1, -1, 0, SETLT),
2326 (SETGT_INT $src1, $src0)
2329 // SETGE_INT reverse args
2331 (selectcc i32:$src0, i32:$src1, -1, 0, SETLE),
2332 (SETGE_INT $src1, $src0)
2335 // SETGT_UINT reverse args
2337 (selectcc i32:$src0, i32:$src1, -1, 0, SETULT),
2338 (SETGT_UINT $src1, $src0)
2341 // SETGE_UINT reverse args
2343 (selectcc i32:$src0, i32:$src1, -1, 0, SETULE),
2344 (SETGE_UINT $src1, $src0)
2347 // The next two patterns are special cases for handling 'true if ordered' and
2348 // 'true if unordered' conditionals. The assumption here is that the behavior of
2349 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
2351 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
2352 // We assume that SETE returns false when one of the operands is NAN and
2353 // SNE returns true when on of the operands is NAN
2355 //SETE - 'true if ordered'
2357 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETO),
2361 //SETE_DX10 - 'true if ordered'
2363 (selectcc f32:$src0, f32:$src1, -1, 0, SETO),
2364 (SETE_DX10 $src0, $src1)
2367 //SNE - 'true if unordered'
2369 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETUO),
2373 //SETNE_DX10 - 'true if ordered'
2375 (selectcc f32:$src0, f32:$src1, -1, 0, SETUO),
2376 (SETNE_DX10 $src0, $src1)
2379 def : Extract_Element <f32, v4f32, 0, sub0>;
2380 def : Extract_Element <f32, v4f32, 1, sub1>;
2381 def : Extract_Element <f32, v4f32, 2, sub2>;
2382 def : Extract_Element <f32, v4f32, 3, sub3>;
2384 def : Insert_Element <f32, v4f32, 0, sub0>;
2385 def : Insert_Element <f32, v4f32, 1, sub1>;
2386 def : Insert_Element <f32, v4f32, 2, sub2>;
2387 def : Insert_Element <f32, v4f32, 3, sub3>;
2389 def : Extract_Element <i32, v4i32, 0, sub0>;
2390 def : Extract_Element <i32, v4i32, 1, sub1>;
2391 def : Extract_Element <i32, v4i32, 2, sub2>;
2392 def : Extract_Element <i32, v4i32, 3, sub3>;
2394 def : Insert_Element <i32, v4i32, 0, sub0>;
2395 def : Insert_Element <i32, v4i32, 1, sub1>;
2396 def : Insert_Element <i32, v4i32, 2, sub2>;
2397 def : Insert_Element <i32, v4i32, 3, sub3>;
2399 def : Vector4_Build <v4f32, f32>;
2400 def : Vector4_Build <v4i32, i32>;
2402 // bitconvert patterns
2404 def : BitConvert <i32, f32, R600_Reg32>;
2405 def : BitConvert <f32, i32, R600_Reg32>;
2406 def : BitConvert <v4f32, v4i32, R600_Reg128>;
2407 def : BitConvert <v4i32, v4f32, R600_Reg128>;
2409 // DWORDADDR pattern
2410 def : DwordAddrPat <i32, R600_Reg32>;
2412 } // End isR600toCayman Predicate