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 : $dst0 dst1",
591 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
592 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
596 def DOT4 : SDNode<"AMDGPUISD::DOT4",
597 SDTypeProfile<1, 8, [SDTCisFP<0>, SDTCisVT<1, f32>, SDTCisVT<2, f32>,
598 SDTCisVT<3, f32>, SDTCisVT<4, f32>, SDTCisVT<5, f32>,
599 SDTCisVT<6, f32>, SDTCisVT<7, f32>, SDTCisVT<8, f32>]>,
603 def TEXTURE_FETCH_Type : SDTypeProfile<1, 19, [SDTCisFP<0>]>;
605 def TEXTURE_FETCH: SDNode<"AMDGPUISD::TEXTURE_FETCH", TEXTURE_FETCH_Type, []>;
607 multiclass TexPattern<bits<32> TextureOp, Instruction inst, ValueType vt = v4f32> {
608 def : Pat<(TEXTURE_FETCH (i32 TextureOp), vt:$SRC_GPR,
609 (i32 imm:$srcx), (i32 imm:$srcy), (i32 imm:$srcz), (i32 imm:$srcw),
610 (i32 imm:$offsetx), (i32 imm:$offsety), (i32 imm:$offsetz),
611 (i32 imm:$DST_SEL_X), (i32 imm:$DST_SEL_Y), (i32 imm:$DST_SEL_Z),
612 (i32 imm:$DST_SEL_W),
613 (i32 imm:$RESOURCE_ID), (i32 imm:$SAMPLER_ID),
614 (i32 imm:$COORD_TYPE_X), (i32 imm:$COORD_TYPE_Y), (i32 imm:$COORD_TYPE_Z),
615 (i32 imm:$COORD_TYPE_W)),
616 (inst R600_Reg128:$SRC_GPR,
617 imm:$srcx, imm:$srcy, imm:$srcz, imm:$srcw,
618 imm:$offsetx, imm:$offsety, imm:$offsetz,
619 imm:$DST_SEL_X, imm:$DST_SEL_Y, imm:$DST_SEL_Z,
621 imm:$RESOURCE_ID, imm:$SAMPLER_ID,
622 imm:$COORD_TYPE_X, imm:$COORD_TYPE_Y, imm:$COORD_TYPE_Z,
626 //===----------------------------------------------------------------------===//
627 // Interpolation Instructions
628 //===----------------------------------------------------------------------===//
630 def INTERP_VEC_LOAD : AMDGPUShaderInst <
631 (outs R600_Reg128:$dst),
633 "INTERP_LOAD $src0 : $dst",
636 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
637 let bank_swizzle = 5;
640 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
641 let bank_swizzle = 5;
644 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
646 //===----------------------------------------------------------------------===//
647 // Export Instructions
648 //===----------------------------------------------------------------------===//
650 def ExportType : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
652 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
653 [SDNPHasChain, SDNPSideEffect]>;
656 field bits<32> Word0;
663 let Word0{12-0} = arraybase;
664 let Word0{14-13} = type;
665 let Word0{21-15} = gpr;
666 let Word0{22} = 0; // RW_REL
667 let Word0{29-23} = 0; // INDEX_GPR
668 let Word0{31-30} = elem_size;
671 class ExportSwzWord1 {
672 field bits<32> Word1;
681 let Word1{2-0} = sw_x;
682 let Word1{5-3} = sw_y;
683 let Word1{8-6} = sw_z;
684 let Word1{11-9} = sw_w;
687 class ExportBufWord1 {
688 field bits<32> Word1;
695 let Word1{11-0} = arraySize;
696 let Word1{15-12} = compMask;
699 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
700 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
702 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
703 0, 61, 0, 7, 7, 7, cf_inst, 0)
706 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
708 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
709 0, 61, 7, 0, 7, 7, cf_inst, 0)
712 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
714 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
717 def : Pat<(int_R600_store_dummy 1),
719 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
722 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
723 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
724 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
725 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
730 multiclass SteamOutputExportPattern<Instruction ExportInst,
731 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
733 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
734 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
735 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
736 4095, imm:$mask, buf0inst, 0)>;
738 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
739 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
740 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
741 4095, imm:$mask, buf1inst, 0)>;
743 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
744 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
745 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
746 4095, imm:$mask, buf2inst, 0)>;
748 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
749 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
750 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
751 4095, imm:$mask, buf3inst, 0)>;
754 // Export Instructions should not be duplicated by TailDuplication pass
755 // (which assumes that duplicable instruction are affected by exec mask)
756 let usesCustomInserter = 1, isNotDuplicable = 1 in {
758 class ExportSwzInst : InstR600ISA<(
760 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
761 i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
763 !strconcat("EXPORT", " $gpr"),
764 []>, ExportWord0, ExportSwzWord1 {
766 let Inst{31-0} = Word0;
767 let Inst{63-32} = Word1;
770 } // End usesCustomInserter = 1
772 class ExportBufInst : InstR600ISA<(
774 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
775 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
776 !strconcat("EXPORT", " $gpr"),
777 []>, ExportWord0, ExportBufWord1 {
779 let Inst{31-0} = Word0;
780 let Inst{63-32} = Word1;
783 //===----------------------------------------------------------------------===//
784 // Control Flow Instructions
785 //===----------------------------------------------------------------------===//
788 field bits<32> Word0;
791 bits<4> KCACHE_BANK0;
792 bits<4> KCACHE_BANK1;
793 bits<2> KCACHE_MODE0;
795 let Word0{21-0} = ADDR;
796 let Word0{25-22} = KCACHE_BANK0;
797 let Word0{29-26} = KCACHE_BANK1;
798 let Word0{31-30} = KCACHE_MODE0;
802 field bits<32> Word1;
804 bits<2> KCACHE_MODE1;
805 bits<8> KCACHE_ADDR0;
806 bits<8> KCACHE_ADDR1;
810 bits<1> WHOLE_QUAD_MODE;
813 let Word1{1-0} = KCACHE_MODE1;
814 let Word1{9-2} = KCACHE_ADDR0;
815 let Word1{17-10} = KCACHE_ADDR1;
816 let Word1{24-18} = COUNT;
817 let Word1{25} = ALT_CONST;
818 let Word1{29-26} = CF_INST;
819 let Word1{30} = WHOLE_QUAD_MODE;
820 let Word1{31} = BARRIER;
823 def KCACHE : InstFlag<"printKCache">;
825 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
826 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1,
827 KCACHE:$KCACHE_MODE0, KCACHE:$KCACHE_MODE1,
828 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1,
830 !strconcat(OpName, " $COUNT, @$ADDR, "
831 "KC0[$KCACHE_MODE0], KC1[$KCACHE_MODE1]"),
832 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
837 let WHOLE_QUAD_MODE = 0;
840 let Inst{31-0} = Word0;
841 let Inst{63-32} = Word1;
844 class CF_WORD0_R600 {
845 field bits<32> Word0;
852 class CF_WORD1_R600 {
853 field bits<32> Word1;
861 bits<1> END_OF_PROGRAM;
862 bits<1> VALID_PIXEL_MODE;
864 bits<1> WHOLE_QUAD_MODE;
867 let Word1{2-0} = POP_COUNT;
868 let Word1{7-3} = CF_CONST;
869 let Word1{9-8} = COND;
870 let Word1{12-10} = COUNT;
871 let Word1{18-13} = CALL_COUNT;
872 let Word1{19} = COUNT_3;
873 let Word1{21} = END_OF_PROGRAM;
874 let Word1{22} = VALID_PIXEL_MODE;
875 let Word1{29-23} = CF_INST;
876 let Word1{30} = WHOLE_QUAD_MODE;
877 let Word1{31} = BARRIER;
880 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
881 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
887 let VALID_PIXEL_MODE = 0;
891 let END_OF_PROGRAM = 0;
892 let WHOLE_QUAD_MODE = 0;
894 let Inst{31-0} = Word0;
895 let Inst{63-32} = Word1;
899 field bits<32> Word0;
902 bits<3> JUMPTABLE_SEL;
904 let Word0{23-0} = ADDR;
905 let Word0{26-24} = JUMPTABLE_SEL;
909 field bits<32> Word1;
915 bits<1> VALID_PIXEL_MODE;
916 bits<1> END_OF_PROGRAM;
920 let Word1{2-0} = POP_COUNT;
921 let Word1{7-3} = CF_CONST;
922 let Word1{9-8} = COND;
923 let Word1{15-10} = COUNT;
924 let Word1{20} = VALID_PIXEL_MODE;
925 let Word1{21} = END_OF_PROGRAM;
926 let Word1{29-22} = CF_INST;
927 let Word1{31} = BARRIER;
930 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
931 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
936 let JUMPTABLE_SEL = 0;
938 let VALID_PIXEL_MODE = 0;
940 let END_OF_PROGRAM = 0;
942 let Inst{31-0} = Word0;
943 let Inst{63-32} = Word1;
946 def CF_ALU : ALU_CLAUSE<8, "ALU">;
947 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
949 def FETCH_CLAUSE : AMDGPUInst <(outs),
950 (ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
956 def ALU_CLAUSE : AMDGPUInst <(outs),
957 (ins i32imm:$addr), "ALU clause starting at $addr:", [] > {
963 def LITERALS : AMDGPUInst <(outs),
964 (ins LITERAL:$literal1, LITERAL:$literal2), "$literal1, $literal2", [] > {
969 let Inst{31-0} = literal1;
970 let Inst{63-32} = literal2;
973 def PAD : AMDGPUInst <(outs), (ins), "PAD", [] > {
977 let Predicates = [isR600toCayman] in {
979 //===----------------------------------------------------------------------===//
980 // Common Instructions R600, R700, Evergreen, Cayman
981 //===----------------------------------------------------------------------===//
983 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
984 // Non-IEEE MUL: 0 * anything = 0
985 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
986 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
987 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
988 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
990 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
991 // so some of the instruction names don't match the asm string.
992 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
993 def SETE : R600_2OP <
995 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_EQ))]
1000 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GT))]
1003 def SGE : R600_2OP <
1005 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GE))]
1008 def SNE : R600_2OP <
1010 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_NE))]
1013 def SETE_DX10 : R600_2OP <
1015 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_EQ))]
1018 def SETGT_DX10 : R600_2OP <
1020 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GT))]
1023 def SETGE_DX10 : R600_2OP <
1025 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GE))]
1028 def SETNE_DX10 : R600_2OP <
1030 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_NE))]
1033 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
1034 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
1035 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
1036 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
1037 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
1039 def MOV : R600_1OP <0x19, "MOV", []>;
1041 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
1043 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
1044 (outs R600_Reg32:$dst),
1050 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
1052 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
1055 (MOV_IMM_I32 imm:$val)
1058 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
1061 (MOV_IMM_F32 fpimm:$val)
1064 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
1065 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
1066 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
1067 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
1069 let hasSideEffects = 1 in {
1071 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
1073 } // end hasSideEffects
1075 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
1076 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
1077 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
1078 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
1079 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
1080 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
1081 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
1082 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
1083 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
1084 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
1086 def SETE_INT : R600_2OP <
1088 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETEQ))]
1091 def SETGT_INT : R600_2OP <
1093 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGT))]
1096 def SETGE_INT : R600_2OP <
1098 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGE))]
1101 def SETNE_INT : R600_2OP <
1103 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETNE))]
1106 def SETGT_UINT : R600_2OP <
1108 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGT))]
1111 def SETGE_UINT : R600_2OP <
1113 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGE))]
1116 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
1117 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
1118 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
1119 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
1121 def CNDE_INT : R600_3OP <
1123 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_EQ))]
1126 def CNDGE_INT : R600_3OP <
1128 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GE))]
1131 def CNDGT_INT : R600_3OP <
1133 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GT))]
1136 //===----------------------------------------------------------------------===//
1137 // Texture instructions
1138 //===----------------------------------------------------------------------===//
1140 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1142 class R600_TEX <bits<11> inst, string opName> :
1143 InstR600 <(outs R600_Reg128:$DST_GPR),
1144 (ins R600_Reg128:$SRC_GPR,
1145 RSel:$srcx, RSel:$srcy, RSel:$srcz, RSel:$srcw,
1146 i32imm:$offsetx, i32imm:$offsety, i32imm:$offsetz,
1147 RSel:$DST_SEL_X, RSel:$DST_SEL_Y, RSel:$DST_SEL_Z, RSel:$DST_SEL_W,
1148 i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
1149 CT:$COORD_TYPE_X, CT:$COORD_TYPE_Y, CT:$COORD_TYPE_Z,
1152 " $DST_GPR.$DST_SEL_X$DST_SEL_Y$DST_SEL_Z$DST_SEL_W, "
1153 "$SRC_GPR.$srcx$srcy$srcz$srcw "
1154 "RID:$RESOURCE_ID SID:$SAMPLER_ID "
1155 "CT:$COORD_TYPE_X$COORD_TYPE_Y$COORD_TYPE_Z$COORD_TYPE_W"),
1157 NullALU>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
1158 let Inst{31-0} = Word0;
1159 let Inst{63-32} = Word1;
1161 let TEX_INST = inst{4-0};
1167 let FETCH_WHOLE_QUAD = 0;
1169 let SAMPLER_INDEX_MODE = 0;
1170 let RESOURCE_INDEX_MODE = 0;
1175 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1179 def TEX_SAMPLE : R600_TEX <0x10, "TEX_SAMPLE">;
1180 def TEX_SAMPLE_C : R600_TEX <0x18, "TEX_SAMPLE_C">;
1181 def TEX_SAMPLE_L : R600_TEX <0x11, "TEX_SAMPLE_L">;
1182 def TEX_SAMPLE_C_L : R600_TEX <0x19, "TEX_SAMPLE_C_L">;
1183 def TEX_SAMPLE_LB : R600_TEX <0x12, "TEX_SAMPLE_LB">;
1184 def TEX_SAMPLE_C_LB : R600_TEX <0x1A, "TEX_SAMPLE_C_LB">;
1185 def TEX_LD : R600_TEX <0x03, "TEX_LD">;
1186 def TEX_GET_TEXTURE_RESINFO : R600_TEX <0x04, "TEX_GET_TEXTURE_RESINFO">;
1187 def TEX_GET_GRADIENTS_H : R600_TEX <0x07, "TEX_GET_GRADIENTS_H">;
1188 def TEX_GET_GRADIENTS_V : R600_TEX <0x08, "TEX_GET_GRADIENTS_V">;
1189 def TEX_SET_GRADIENTS_H : R600_TEX <0x0B, "TEX_SET_GRADIENTS_H">;
1190 def TEX_SET_GRADIENTS_V : R600_TEX <0x0C, "TEX_SET_GRADIENTS_V">;
1191 def TEX_SAMPLE_G : R600_TEX <0x14, "TEX_SAMPLE_G">;
1192 def TEX_SAMPLE_C_G : R600_TEX <0x1C, "TEX_SAMPLE_C_G">;
1194 defm : TexPattern<0, TEX_SAMPLE>;
1195 defm : TexPattern<1, TEX_SAMPLE_C>;
1196 defm : TexPattern<2, TEX_SAMPLE_L>;
1197 defm : TexPattern<3, TEX_SAMPLE_C_L>;
1198 defm : TexPattern<4, TEX_SAMPLE_LB>;
1199 defm : TexPattern<5, TEX_SAMPLE_C_LB>;
1200 defm : TexPattern<6, TEX_LD, v4i32>;
1201 defm : TexPattern<7, TEX_GET_TEXTURE_RESINFO, v4i32>;
1202 defm : TexPattern<8, TEX_GET_GRADIENTS_H>;
1203 defm : TexPattern<9, TEX_GET_GRADIENTS_V>;
1205 //===----------------------------------------------------------------------===//
1206 // Helper classes for common instructions
1207 //===----------------------------------------------------------------------===//
1209 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
1214 class MULADD_Common <bits<5> inst> : R600_3OP <
1219 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
1220 inst, "MULADD_IEEE",
1221 [(set f32:$dst, (fadd (fmul f32:$src0, f32:$src1), f32:$src2))]
1224 class CNDE_Common <bits<5> inst> : R600_3OP <
1226 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_EQ))]
1229 class CNDGT_Common <bits<5> inst> : R600_3OP <
1231 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GT))]
1234 class CNDGE_Common <bits<5> inst> : R600_3OP <
1236 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GE))]
1240 let isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
1241 class R600_VEC2OP<list<dag> pattern> : InstR600 <(outs R600_Reg32:$dst), (ins
1243 UEM:$update_exec_mask_X, UP:$update_pred_X, WRITE:$write_X,
1244 OMOD:$omod_X, REL:$dst_rel_X, CLAMP:$clamp_X,
1245 R600_TReg32_X:$src0_X, NEG:$src0_neg_X, REL:$src0_rel_X, ABS:$src0_abs_X, SEL:$src0_sel_X,
1246 R600_TReg32_X:$src1_X, NEG:$src1_neg_X, REL:$src1_rel_X, ABS:$src1_abs_X, SEL:$src1_sel_X,
1247 R600_Pred:$pred_sel_X,
1249 UEM:$update_exec_mask_Y, UP:$update_pred_Y, WRITE:$write_Y,
1250 OMOD:$omod_Y, REL:$dst_rel_Y, CLAMP:$clamp_Y,
1251 R600_TReg32_Y:$src0_Y, NEG:$src0_neg_Y, REL:$src0_rel_Y, ABS:$src0_abs_Y, SEL:$src0_sel_Y,
1252 R600_TReg32_Y:$src1_Y, NEG:$src1_neg_Y, REL:$src1_rel_Y, ABS:$src1_abs_Y, SEL:$src1_sel_Y,
1253 R600_Pred:$pred_sel_Y,
1255 UEM:$update_exec_mask_Z, UP:$update_pred_Z, WRITE:$write_Z,
1256 OMOD:$omod_Z, REL:$dst_rel_Z, CLAMP:$clamp_Z,
1257 R600_TReg32_Z:$src0_Z, NEG:$src0_neg_Z, REL:$src0_rel_Z, ABS:$src0_abs_Z, SEL:$src0_sel_Z,
1258 R600_TReg32_Z:$src1_Z, NEG:$src1_neg_Z, REL:$src1_rel_Z, ABS:$src1_abs_Z, SEL:$src1_sel_Z,
1259 R600_Pred:$pred_sel_Z,
1261 UEM:$update_exec_mask_W, UP:$update_pred_W, WRITE:$write_W,
1262 OMOD:$omod_W, REL:$dst_rel_W, CLAMP:$clamp_W,
1263 R600_TReg32_W:$src0_W, NEG:$src0_neg_W, REL:$src0_rel_W, ABS:$src0_abs_W, SEL:$src0_sel_W,
1264 R600_TReg32_W:$src1_W, NEG:$src1_neg_W, REL:$src1_rel_W, ABS:$src1_abs_W, SEL:$src1_sel_W,
1265 R600_Pred:$pred_sel_W,
1266 LITERAL:$literal0, LITERAL:$literal1),
1272 def DOT_4 : R600_VEC2OP<[(set R600_Reg32:$dst, (DOT4
1273 R600_TReg32_X:$src0_X, R600_TReg32_X:$src1_X,
1274 R600_TReg32_Y:$src0_Y, R600_TReg32_Y:$src1_Y,
1275 R600_TReg32_Z:$src0_Z, R600_TReg32_Z:$src1_Z,
1276 R600_TReg32_W:$src0_W, R600_TReg32_W:$src1_W))]>;
1279 class DOT4_Common <bits<11> inst> : R600_2OP <inst, "DOT4", []>;
1282 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1283 multiclass CUBE_Common <bits<11> inst> {
1285 def _pseudo : InstR600 <
1286 (outs R600_Reg128:$dst),
1287 (ins R600_Reg128:$src),
1289 [(set v4f32:$dst, (int_AMDGPU_cube v4f32:$src))],
1295 def _real : R600_2OP <inst, "CUBE", []>;
1297 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1299 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1300 inst, "EXP_IEEE", fexp2
1303 let Itinerary = TransALU;
1306 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
1307 inst, "FLT_TO_INT", fp_to_sint
1310 let Itinerary = TransALU;
1313 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1314 inst, "INT_TO_FLT", sint_to_fp
1317 let Itinerary = TransALU;
1320 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1321 inst, "FLT_TO_UINT", fp_to_uint
1324 let Itinerary = TransALU;
1327 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1328 inst, "UINT_TO_FLT", uint_to_fp
1331 let Itinerary = TransALU;
1334 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1335 inst, "LOG_CLAMPED", []
1338 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1339 inst, "LOG_IEEE", flog2
1342 let Itinerary = TransALU;
1345 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1346 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1347 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1348 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1349 inst, "MULHI_INT", mulhs
1352 let Itinerary = TransALU;
1354 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1355 inst, "MULHI", mulhu
1358 let Itinerary = TransALU;
1360 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1361 inst, "MULLO_INT", mul
1364 let Itinerary = TransALU;
1366 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []> {
1368 let Itinerary = TransALU;
1371 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1372 inst, "RECIP_CLAMPED", []
1375 let Itinerary = TransALU;
1378 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1379 inst, "RECIP_IEEE", [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
1382 let Itinerary = TransALU;
1385 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1386 inst, "RECIP_UINT", AMDGPUurecip
1389 let Itinerary = TransALU;
1392 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1393 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1396 let Itinerary = TransALU;
1399 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1400 inst, "RECIPSQRT_IEEE", []
1403 let Itinerary = TransALU;
1406 class SIN_Common <bits<11> inst> : R600_1OP <
1410 let Itinerary = TransALU;
1413 class COS_Common <bits<11> inst> : R600_1OP <
1417 let Itinerary = TransALU;
1420 //===----------------------------------------------------------------------===//
1421 // Helper patterns for complex intrinsics
1422 //===----------------------------------------------------------------------===//
1424 multiclass DIV_Common <InstR600 recip_ieee> {
1426 (int_AMDGPU_div f32:$src0, f32:$src1),
1427 (MUL_IEEE $src0, (recip_ieee $src1))
1431 (fdiv f32:$src0, f32:$src1),
1432 (MUL_IEEE $src0, (recip_ieee $src1))
1436 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee>
1438 (int_TGSI_lit_z f32:$src_x, f32:$src_y, f32:$src_w),
1439 (exp_ieee (mul_lit (log_clamped (MAX $src_y, (f32 ZERO))), $src_w, $src_x))
1442 //===----------------------------------------------------------------------===//
1443 // R600 / R700 Instructions
1444 //===----------------------------------------------------------------------===//
1446 let Predicates = [isR600] in {
1448 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1449 def MULADD_r600 : MULADD_Common<0x10>;
1450 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1451 def CNDE_r600 : CNDE_Common<0x18>;
1452 def CNDGT_r600 : CNDGT_Common<0x19>;
1453 def CNDGE_r600 : CNDGE_Common<0x1A>;
1454 def DOT4_r600 : DOT4_Common<0x50>;
1455 defm CUBE_r600 : CUBE_Common<0x52>;
1456 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1457 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1458 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1459 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1460 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1461 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1462 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1463 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1464 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1465 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1466 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1467 def SIN_r600 : SIN_Common<0x6E>;
1468 def COS_r600 : COS_Common<0x6F>;
1469 def ASHR_r600 : ASHR_Common<0x70>;
1470 def LSHR_r600 : LSHR_Common<0x71>;
1471 def LSHL_r600 : LSHL_Common<0x72>;
1472 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1473 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1474 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1475 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1476 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1478 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1479 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL>;
1480 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1482 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
1484 def R600_ExportSwz : ExportSwzInst {
1485 let Word1{20-17} = 0; // BURST_COUNT
1486 let Word1{21} = eop;
1487 let Word1{22} = 1; // VALID_PIXEL_MODE
1488 let Word1{30-23} = inst;
1489 let Word1{31} = 1; // BARRIER
1491 defm : ExportPattern<R600_ExportSwz, 39>;
1493 def R600_ExportBuf : ExportBufInst {
1494 let Word1{20-17} = 0; // BURST_COUNT
1495 let Word1{21} = eop;
1496 let Word1{22} = 1; // VALID_PIXEL_MODE
1497 let Word1{30-23} = inst;
1498 let Word1{31} = 1; // BARRIER
1500 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1502 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1503 "TEX $COUNT @$ADDR"> {
1506 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1507 "VTX $COUNT @$ADDR"> {
1510 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1511 "LOOP_START_DX10 @$ADDR"> {
1515 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1519 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1520 "LOOP_BREAK @$ADDR"> {
1524 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1525 "CONTINUE @$ADDR"> {
1529 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1530 "JUMP @$ADDR POP:$POP_COUNT"> {
1533 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1534 "ELSE @$ADDR POP:$POP_COUNT"> {
1537 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1542 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1543 "POP @$ADDR POP:$POP_COUNT"> {
1546 def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
1550 let END_OF_PROGRAM = 1;
1555 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1557 class COS_PAT <InstR600 trig> : Pat<
1559 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1562 class SIN_PAT <InstR600 trig> : Pat<
1564 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1567 //===----------------------------------------------------------------------===//
1568 // R700 Only instructions
1569 //===----------------------------------------------------------------------===//
1571 let Predicates = [isR700] in {
1572 def SIN_r700 : SIN_Common<0x6E>;
1573 def COS_r700 : COS_Common<0x6F>;
1575 // R700 normalizes inputs to SIN/COS the same as EG
1576 def : SIN_PAT <SIN_r700>;
1577 def : COS_PAT <COS_r700>;
1580 //===----------------------------------------------------------------------===//
1581 // Evergreen Only instructions
1582 //===----------------------------------------------------------------------===//
1584 let Predicates = [isEG] in {
1586 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1587 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1589 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1590 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1591 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1592 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1593 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1594 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1595 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1596 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1597 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1598 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1599 def SIN_eg : SIN_Common<0x8D>;
1600 def COS_eg : COS_Common<0x8E>;
1602 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
1603 def : SIN_PAT <SIN_eg>;
1604 def : COS_PAT <COS_eg>;
1605 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
1606 } // End Predicates = [isEG]
1608 //===----------------------------------------------------------------------===//
1609 // Evergreen / Cayman Instructions
1610 //===----------------------------------------------------------------------===//
1612 let Predicates = [isEGorCayman] in {
1614 // BFE_UINT - bit_extract, an optimization for mask and shift
1619 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1624 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1625 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1626 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1627 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1628 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1629 [(set i32:$dst, (int_AMDIL_bit_extract_u32 i32:$src0, i32:$src1,
1633 def : BFEPattern <BFE_UINT_eg>;
1635 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>;
1636 defm : BFIPatterns <BFI_INT_eg>;
1638 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
1639 [(set i32:$dst, (AMDGPUbitalign i32:$src0, i32:$src1, i32:$src2))],
1643 def MULADD_eg : MULADD_Common<0x14>;
1644 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
1645 def ASHR_eg : ASHR_Common<0x15>;
1646 def LSHR_eg : LSHR_Common<0x16>;
1647 def LSHL_eg : LSHL_Common<0x17>;
1648 def CNDE_eg : CNDE_Common<0x19>;
1649 def CNDGT_eg : CNDGT_Common<0x1A>;
1650 def CNDGE_eg : CNDGE_Common<0x1B>;
1651 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1652 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1653 def DOT4_eg : DOT4_Common<0xBE>;
1654 defm CUBE_eg : CUBE_Common<0xC0>;
1656 let hasSideEffects = 1 in {
1657 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1660 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1662 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1666 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1668 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1672 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1674 // TRUNC is used for the FLT_TO_INT instructions to work around a
1675 // perceived problem where the rounding modes are applied differently
1676 // depending on the instruction and the slot they are in.
1678 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1679 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1681 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1682 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1683 // We should look into handling these cases separately.
1684 def : Pat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
1686 def : Pat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
1689 def : SHA256MaPattern <BFI_INT_eg, XOR_INT>;
1691 def EG_ExportSwz : ExportSwzInst {
1692 let Word1{19-16} = 0; // BURST_COUNT
1693 let Word1{20} = 1; // VALID_PIXEL_MODE
1694 let Word1{21} = eop;
1695 let Word1{29-22} = inst;
1696 let Word1{30} = 0; // MARK
1697 let Word1{31} = 1; // BARRIER
1699 defm : ExportPattern<EG_ExportSwz, 83>;
1701 def EG_ExportBuf : ExportBufInst {
1702 let Word1{19-16} = 0; // BURST_COUNT
1703 let Word1{20} = 1; // VALID_PIXEL_MODE
1704 let Word1{21} = eop;
1705 let Word1{29-22} = inst;
1706 let Word1{30} = 0; // MARK
1707 let Word1{31} = 1; // BARRIER
1709 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1711 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1712 "TEX $COUNT @$ADDR"> {
1715 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1716 "VTX $COUNT @$ADDR"> {
1719 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
1720 "LOOP_START_DX10 @$ADDR"> {
1724 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1728 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
1729 "LOOP_BREAK @$ADDR"> {
1733 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
1734 "CONTINUE @$ADDR"> {
1738 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1739 "JUMP @$ADDR POP:$POP_COUNT"> {
1742 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1743 "ELSE @$ADDR POP:$POP_COUNT"> {
1746 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
1751 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1752 "POP @$ADDR POP:$POP_COUNT"> {
1755 def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
1759 let END_OF_PROGRAM = 1;
1762 //===----------------------------------------------------------------------===//
1763 // Memory read/write instructions
1764 //===----------------------------------------------------------------------===//
1765 let usesCustomInserter = 1 in {
1767 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
1769 : EG_CF_RAT <0x57, 0x2, 0, (outs), ins, name, pattern> {
1771 // XXX: Have a separate instruction for non-indexed writes.
1777 let COMP_MASK = comp_mask;
1778 let BURST_COUNT = 0;
1784 } // End usesCustomInserter = 1
1787 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1788 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1789 0x1, "RAT_WRITE_CACHELESS_32_eg $rw_gpr, $index_gpr, $eop",
1790 [(global_store i32:$rw_gpr, i32:$index_gpr)]
1794 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1795 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1796 0xf, "RAT_WRITE_CACHELESS_128 $rw_gpr.XYZW, $index_gpr, $eop",
1797 [(global_store v4i32:$rw_gpr, i32:$index_gpr)]
1800 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1801 : InstR600ISA <outs, (ins MEMxi:$ptr), name, pattern>,
1802 VTX_WORD1_GPR, VTX_WORD0 {
1807 let FETCH_WHOLE_QUAD = 0;
1808 let BUFFER_ID = buffer_id;
1810 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1811 // to store vertex addresses in any channel, not just X.
1814 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
1815 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
1816 // however, based on my testing if USE_CONST_FIELDS is set, then all
1817 // these fields need to be set to 0.
1818 let USE_CONST_FIELDS = 0;
1819 let NUM_FORMAT_ALL = 1;
1820 let FORMAT_COMP_ALL = 0;
1821 let SRF_MODE_ALL = 0;
1823 let Inst{31-0} = Word0;
1824 let Inst{63-32} = Word1;
1825 // LLVM can only encode 64-bit instructions, so these fields are manually
1826 // encoded in R600CodeEmitter
1829 // bits<2> ENDIAN_SWAP = 0;
1830 // bits<1> CONST_BUF_NO_STRIDE = 0;
1831 // bits<1> MEGA_FETCH = 0;
1832 // bits<1> ALT_CONST = 0;
1833 // bits<2> BUFFER_INDEX_MODE = 0;
1837 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1838 // is done in R600CodeEmitter
1840 // Inst{79-64} = OFFSET;
1841 // Inst{81-80} = ENDIAN_SWAP;
1842 // Inst{82} = CONST_BUF_NO_STRIDE;
1843 // Inst{83} = MEGA_FETCH;
1844 // Inst{84} = ALT_CONST;
1845 // Inst{86-85} = BUFFER_INDEX_MODE;
1846 // Inst{95-86} = 0; Reserved
1848 // VTX_WORD3 (Padding)
1850 // Inst{127-96} = 0;
1855 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1856 : VTX_READ_eg <"VTX_READ_8 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1859 let MEGA_FETCH_COUNT = 1;
1861 let DST_SEL_Y = 7; // Masked
1862 let DST_SEL_Z = 7; // Masked
1863 let DST_SEL_W = 7; // Masked
1864 let DATA_FORMAT = 1; // FMT_8
1867 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1868 : VTX_READ_eg <"VTX_READ_16 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1870 let MEGA_FETCH_COUNT = 2;
1872 let DST_SEL_Y = 7; // Masked
1873 let DST_SEL_Z = 7; // Masked
1874 let DST_SEL_W = 7; // Masked
1875 let DATA_FORMAT = 5; // FMT_16
1879 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1880 : VTX_READ_eg <"VTX_READ_32 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1883 let MEGA_FETCH_COUNT = 4;
1885 let DST_SEL_Y = 7; // Masked
1886 let DST_SEL_Z = 7; // Masked
1887 let DST_SEL_W = 7; // Masked
1888 let DATA_FORMAT = 0xD; // COLOR_32
1890 // This is not really necessary, but there were some GPU hangs that appeared
1891 // to be caused by ALU instructions in the next instruction group that wrote
1892 // to the $ptr registers of the VTX_READ.
1894 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1895 // %T2_X<def> = MOV %ZERO
1896 //Adding this constraint prevents this from happening.
1897 let Constraints = "$ptr.ptr = $dst";
1900 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1901 : VTX_READ_eg <"VTX_READ_128 $dst.XYZW, $ptr", buffer_id, (outs R600_Reg128:$dst),
1904 let MEGA_FETCH_COUNT = 16;
1909 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1911 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1912 // that holds its buffer address to avoid potential hangs. We can't use
1913 // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
1914 // registers are different sizes.
1917 //===----------------------------------------------------------------------===//
1918 // VTX Read from parameter memory space
1919 //===----------------------------------------------------------------------===//
1921 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1922 [(set i32:$dst, (load_param_zexti8 ADDRVTX_READ:$ptr))]
1925 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1926 [(set i32:$dst, (load_param_zexti16 ADDRVTX_READ:$ptr))]
1929 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1930 [(set i32:$dst, (load_param ADDRVTX_READ:$ptr))]
1933 def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
1934 [(set v4i32:$dst, (load_param ADDRVTX_READ:$ptr))]
1937 //===----------------------------------------------------------------------===//
1938 // VTX Read from global memory space
1939 //===----------------------------------------------------------------------===//
1942 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1943 [(set i32:$dst, (zextloadi8_global ADDRVTX_READ:$ptr))]
1947 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1948 [(set i32:$dst, (global_load ADDRVTX_READ:$ptr))]
1952 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1953 [(set v4i32:$dst, (global_load ADDRVTX_READ:$ptr))]
1956 //===----------------------------------------------------------------------===//
1958 // XXX: We are currently storing all constants in the global address space.
1959 //===----------------------------------------------------------------------===//
1961 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1962 [(set i32:$dst, (constant_load ADDRVTX_READ:$ptr))]
1967 //===----------------------------------------------------------------------===//
1968 // Regist loads and stores - for indirect addressing
1969 //===----------------------------------------------------------------------===//
1971 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1973 let Predicates = [isCayman] in {
1975 let isVector = 1 in {
1977 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1979 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1980 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1981 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1982 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1983 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1984 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1985 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
1986 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1987 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1988 def SIN_cm : SIN_Common<0x8D>;
1989 def COS_cm : COS_Common<0x8E>;
1990 } // End isVector = 1
1992 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
1993 def : SIN_PAT <SIN_cm>;
1994 def : COS_PAT <COS_cm>;
1996 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1998 // RECIP_UINT emulation for Cayman
1999 // The multiplication scales from [0,1] to the unsigned integer range
2001 (AMDGPUurecip i32:$src0),
2002 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
2003 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
2006 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
2012 def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
2016 //===----------------------------------------------------------------------===//
2017 // Branch Instructions
2018 //===----------------------------------------------------------------------===//
2021 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
2022 "IF_PREDICATE_SET $src", []>;
2024 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
2025 "PREDICATED_BREAK $src", []>;
2027 //===----------------------------------------------------------------------===//
2028 // Pseudo instructions
2029 //===----------------------------------------------------------------------===//
2031 let isPseudo = 1 in {
2033 def PRED_X : InstR600 <
2034 (outs R600_Predicate_Bit:$dst),
2035 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
2037 let FlagOperandIdx = 3;
2040 let isTerminator = 1, isBranch = 1 in {
2041 def JUMP_COND : InstR600 <
2043 (ins brtarget:$target, R600_Predicate_Bit:$p),
2044 "JUMP $target ($p)",
2048 def JUMP : InstR600 <
2050 (ins brtarget:$target),
2055 let isPredicable = 1;
2059 } // End isTerminator = 1, isBranch = 1
2061 let usesCustomInserter = 1 in {
2063 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
2065 def MASK_WRITE : AMDGPUShaderInst <
2067 (ins R600_Reg32:$src),
2072 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
2076 (outs R600_Reg128:$dst),
2077 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
2078 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2079 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2080 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
2081 imm:$resourceId, imm:$samplerId, imm:$textureTarget))],
2086 def TXD_SHADOW: InstR600 <
2087 (outs R600_Reg128:$dst),
2088 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
2089 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2090 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2091 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
2092 imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))],
2097 } // End isPseudo = 1
2098 } // End usesCustomInserter = 1
2100 def CLAMP_R600 : CLAMP <R600_Reg32>;
2101 def FABS_R600 : FABS<R600_Reg32>;
2102 def FNEG_R600 : FNEG<R600_Reg32>;
2104 //===---------------------------------------------------------------------===//
2105 // Return instruction
2106 //===---------------------------------------------------------------------===//
2107 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
2108 usesCustomInserter = 1 in {
2109 def RETURN : ILFormat<(outs), (ins variable_ops),
2110 "RETURN", [(IL_retflag)]>;
2114 //===----------------------------------------------------------------------===//
2115 // Constant Buffer Addressing Support
2116 //===----------------------------------------------------------------------===//
2118 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
2119 def CONST_COPY : Instruction {
2120 let OutOperandList = (outs R600_Reg32:$dst);
2121 let InOperandList = (ins i32imm:$src);
2123 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
2124 let AsmString = "CONST_COPY";
2125 let neverHasSideEffects = 1;
2126 let isAsCheapAsAMove = 1;
2127 let Itinerary = NullALU;
2129 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
2131 def TEX_VTX_CONSTBUF :
2132 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
2133 [(set v4i32:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
2134 VTX_WORD1_GPR, VTX_WORD0 {
2138 let FETCH_WHOLE_QUAD = 0;
2142 let USE_CONST_FIELDS = 0;
2143 let NUM_FORMAT_ALL = 2;
2144 let FORMAT_COMP_ALL = 1;
2145 let SRF_MODE_ALL = 1;
2146 let MEGA_FETCH_COUNT = 16;
2151 let DATA_FORMAT = 35;
2153 let Inst{31-0} = Word0;
2154 let Inst{63-32} = Word1;
2156 // LLVM can only encode 64-bit instructions, so these fields are manually
2157 // encoded in R600CodeEmitter
2160 // bits<2> ENDIAN_SWAP = 0;
2161 // bits<1> CONST_BUF_NO_STRIDE = 0;
2162 // bits<1> MEGA_FETCH = 0;
2163 // bits<1> ALT_CONST = 0;
2164 // bits<2> BUFFER_INDEX_MODE = 0;
2168 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2169 // is done in R600CodeEmitter
2171 // Inst{79-64} = OFFSET;
2172 // Inst{81-80} = ENDIAN_SWAP;
2173 // Inst{82} = CONST_BUF_NO_STRIDE;
2174 // Inst{83} = MEGA_FETCH;
2175 // Inst{84} = ALT_CONST;
2176 // Inst{86-85} = BUFFER_INDEX_MODE;
2177 // Inst{95-86} = 0; Reserved
2179 // VTX_WORD3 (Padding)
2181 // Inst{127-96} = 0;
2186 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
2187 [(set v4f32:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
2188 VTX_WORD1_GPR, VTX_WORD0 {
2192 let FETCH_WHOLE_QUAD = 0;
2196 let USE_CONST_FIELDS = 1;
2197 let NUM_FORMAT_ALL = 0;
2198 let FORMAT_COMP_ALL = 0;
2199 let SRF_MODE_ALL = 1;
2200 let MEGA_FETCH_COUNT = 16;
2205 let DATA_FORMAT = 0;
2207 let Inst{31-0} = Word0;
2208 let Inst{63-32} = Word1;
2210 // LLVM can only encode 64-bit instructions, so these fields are manually
2211 // encoded in R600CodeEmitter
2214 // bits<2> ENDIAN_SWAP = 0;
2215 // bits<1> CONST_BUF_NO_STRIDE = 0;
2216 // bits<1> MEGA_FETCH = 0;
2217 // bits<1> ALT_CONST = 0;
2218 // bits<2> BUFFER_INDEX_MODE = 0;
2222 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2223 // is done in R600CodeEmitter
2225 // Inst{79-64} = OFFSET;
2226 // Inst{81-80} = ENDIAN_SWAP;
2227 // Inst{82} = CONST_BUF_NO_STRIDE;
2228 // Inst{83} = MEGA_FETCH;
2229 // Inst{84} = ALT_CONST;
2230 // Inst{86-85} = BUFFER_INDEX_MODE;
2231 // Inst{95-86} = 0; Reserved
2233 // VTX_WORD3 (Padding)
2235 // Inst{127-96} = 0;
2241 //===--------------------------------------------------------------------===//
2242 // Instructions support
2243 //===--------------------------------------------------------------------===//
2244 //===---------------------------------------------------------------------===//
2245 // Custom Inserter for Branches and returns, this eventually will be a
2247 //===---------------------------------------------------------------------===//
2248 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
2249 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
2250 "; Pseudo unconditional branch instruction",
2252 defm BRANCH_COND : BranchConditional<IL_brcond>;
2255 //===---------------------------------------------------------------------===//
2256 // Flow and Program control Instructions
2257 //===---------------------------------------------------------------------===//
2258 let isTerminator=1 in {
2259 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
2260 !strconcat("SWITCH", " $src"), []>;
2261 def CASE : ILFormat< (outs), (ins GPRI32:$src),
2262 !strconcat("CASE", " $src"), []>;
2263 def BREAK : ILFormat< (outs), (ins),
2265 def CONTINUE : ILFormat< (outs), (ins),
2267 def DEFAULT : ILFormat< (outs), (ins),
2269 def ELSE : ILFormat< (outs), (ins),
2271 def ENDSWITCH : ILFormat< (outs), (ins),
2273 def ENDMAIN : ILFormat< (outs), (ins),
2275 def END : ILFormat< (outs), (ins),
2277 def ENDFUNC : ILFormat< (outs), (ins),
2279 def ENDIF : ILFormat< (outs), (ins),
2281 def WHILELOOP : ILFormat< (outs), (ins),
2283 def ENDLOOP : ILFormat< (outs), (ins),
2285 def FUNC : ILFormat< (outs), (ins),
2287 def RETDYN : ILFormat< (outs), (ins),
2289 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2290 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
2291 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2292 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
2293 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2294 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
2295 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2296 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
2297 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2298 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
2299 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2300 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
2301 defm IFC : BranchInstr2<"IFC">;
2302 defm BREAKC : BranchInstr2<"BREAKC">;
2303 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
2306 //===----------------------------------------------------------------------===//
2308 //===----------------------------------------------------------------------===//
2310 // CND*_INT Pattterns for f32 True / False values
2312 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
2313 (selectcc i32:$src0, 0, f32:$src1, f32:$src2, cc),
2314 (cnd $src0, $src1, $src2)
2317 def : CND_INT_f32 <CNDE_INT, SETEQ>;
2318 def : CND_INT_f32 <CNDGT_INT, SETGT>;
2319 def : CND_INT_f32 <CNDGE_INT, SETGE>;
2321 //CNDGE_INT extra pattern
2323 (selectcc i32:$src0, -1, i32:$src1, i32:$src2, COND_GT),
2324 (CNDGE_INT $src0, $src1, $src2)
2330 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
2334 (int_AMDGPU_kill f32:$src0),
2335 (MASK_WRITE (KILLGT (f32 ZERO), $src0))
2340 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LT),
2346 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LE),
2350 // SETGT_DX10 reverse args
2352 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LT),
2353 (SETGT_DX10 $src1, $src0)
2356 // SETGE_DX10 reverse args
2358 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LE),
2359 (SETGE_DX10 $src1, $src0)
2362 // SETGT_INT reverse args
2364 (selectcc i32:$src0, i32:$src1, -1, 0, SETLT),
2365 (SETGT_INT $src1, $src0)
2368 // SETGE_INT reverse args
2370 (selectcc i32:$src0, i32:$src1, -1, 0, SETLE),
2371 (SETGE_INT $src1, $src0)
2374 // SETGT_UINT reverse args
2376 (selectcc i32:$src0, i32:$src1, -1, 0, SETULT),
2377 (SETGT_UINT $src1, $src0)
2380 // SETGE_UINT reverse args
2382 (selectcc i32:$src0, i32:$src1, -1, 0, SETULE),
2383 (SETGE_UINT $src1, $src0)
2386 // The next two patterns are special cases for handling 'true if ordered' and
2387 // 'true if unordered' conditionals. The assumption here is that the behavior of
2388 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
2390 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
2391 // We assume that SETE returns false when one of the operands is NAN and
2392 // SNE returns true when on of the operands is NAN
2394 //SETE - 'true if ordered'
2396 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETO),
2400 //SETE_DX10 - 'true if ordered'
2402 (selectcc f32:$src0, f32:$src1, -1, 0, SETO),
2403 (SETE_DX10 $src0, $src1)
2406 //SNE - 'true if unordered'
2408 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETUO),
2412 //SETNE_DX10 - 'true if ordered'
2414 (selectcc f32:$src0, f32:$src1, -1, 0, SETUO),
2415 (SETNE_DX10 $src0, $src1)
2418 def : Extract_Element <f32, v4f32, 0, sub0>;
2419 def : Extract_Element <f32, v4f32, 1, sub1>;
2420 def : Extract_Element <f32, v4f32, 2, sub2>;
2421 def : Extract_Element <f32, v4f32, 3, sub3>;
2423 def : Insert_Element <f32, v4f32, 0, sub0>;
2424 def : Insert_Element <f32, v4f32, 1, sub1>;
2425 def : Insert_Element <f32, v4f32, 2, sub2>;
2426 def : Insert_Element <f32, v4f32, 3, sub3>;
2428 def : Extract_Element <i32, v4i32, 0, sub0>;
2429 def : Extract_Element <i32, v4i32, 1, sub1>;
2430 def : Extract_Element <i32, v4i32, 2, sub2>;
2431 def : Extract_Element <i32, v4i32, 3, sub3>;
2433 def : Insert_Element <i32, v4i32, 0, sub0>;
2434 def : Insert_Element <i32, v4i32, 1, sub1>;
2435 def : Insert_Element <i32, v4i32, 2, sub2>;
2436 def : Insert_Element <i32, v4i32, 3, sub3>;
2438 def : Vector4_Build <v4f32, f32>;
2439 def : Vector4_Build <v4i32, i32>;
2441 // bitconvert patterns
2443 def : BitConvert <i32, f32, R600_Reg32>;
2444 def : BitConvert <f32, i32, R600_Reg32>;
2445 def : BitConvert <v4f32, v4i32, R600_Reg128>;
2446 def : BitConvert <v4i32, v4f32, R600_Reg128>;
2448 // DWORDADDR pattern
2449 def : DwordAddrPat <i32, R600_Reg32>;
2451 } // End isR600toCayman Predicate