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 // TableGen definitions for instructions which are available on R600 family
13 //===----------------------------------------------------------------------===//
15 include "R600Intrinsics.td"
16 include "R600InstrFormats.td"
18 class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern = []> :
19 InstR600 <outs, ins, asm, pattern, NullALU> {
21 let Namespace = "AMDGPU";
24 def MEMxi : Operand<iPTR> {
25 let MIOperandInfo = (ops R600_TReg32_X:$ptr, i32imm:$index);
26 let PrintMethod = "printMemOperand";
29 def MEMrr : Operand<iPTR> {
30 let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
33 // Operands for non-registers
35 class InstFlag<string PM = "printOperand", int Default = 0>
36 : OperandWithDefaultOps <i32, (ops (i32 Default))> {
40 // src_sel for ALU src operands, see also ALU_CONST, ALU_PARAM registers
41 def SEL : OperandWithDefaultOps <i32, (ops (i32 -1))> {
42 let PrintMethod = "printSel";
44 def BANK_SWIZZLE : OperandWithDefaultOps <i32, (ops (i32 0))> {
45 let PrintMethod = "printBankSwizzle";
48 def LITERAL : InstFlag<"printLiteral">;
50 def WRITE : InstFlag <"printWrite", 1>;
51 def OMOD : InstFlag <"printOMOD">;
52 def REL : InstFlag <"printRel">;
53 def CLAMP : InstFlag <"printClamp">;
54 def NEG : InstFlag <"printNeg">;
55 def ABS : InstFlag <"printAbs">;
56 def UEM : InstFlag <"printUpdateExecMask">;
57 def UP : InstFlag <"printUpdatePred">;
59 // XXX: The r600g finalizer in Mesa expects last to be one in most cases.
60 // Once we start using the packetizer in this backend we should have this
62 def LAST : InstFlag<"printLast", 1>;
63 def RSel : Operand<i32> {
64 let PrintMethod = "printRSel";
66 def CT: Operand<i32> {
67 let PrintMethod = "printCT";
70 def FRAMEri : Operand<iPTR> {
71 let MIOperandInfo = (ops R600_Reg32:$ptr, i32imm:$index);
74 def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
75 def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
76 def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
77 def ADDRGA_CONST_OFFSET : ComplexPattern<i32, 1, "SelectGlobalValueConstantOffset", [], []>;
78 def ADDRGA_VAR_OFFSET : ComplexPattern<i32, 2, "SelectGlobalValueVariableOffset", [], []>;
81 def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
85 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
87 // Class for instructions with only one source register.
88 // If you add new ins to this instruction, make sure they are listed before
89 // $literal, because the backend currently assumes that the last operand is
90 // a literal. Also be sure to update the enum R600Op1OperandIndex::ROI in
91 // R600Defines.h, R600InstrInfo::buildDefaultInstruction(),
92 // and R600InstrInfo::getOperandIdx().
93 class R600_1OP <bits<11> inst, string opName, list<dag> pattern,
94 InstrItinClass itin = AnyALU> :
95 InstR600 <(outs R600_Reg32:$dst),
96 (ins WRITE:$write, OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
97 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
98 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
99 BANK_SWIZZLE:$bank_swizzle),
100 !strconcat(" ", opName,
101 "$clamp $last $dst$write$dst_rel$omod, "
102 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
103 "$pred_sel $bank_swizzle"),
107 R600ALU_Word1_OP2 <inst> {
113 let update_exec_mask = 0;
115 let HasNativeOperands = 1;
118 let DisableEncoding = "$literal";
119 let UseNamedOperandTable = 1;
121 let Inst{31-0} = Word0;
122 let Inst{63-32} = Word1;
125 class R600_1OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
126 InstrItinClass itin = AnyALU> :
127 R600_1OP <inst, opName,
128 [(set R600_Reg32:$dst, (node R600_Reg32:$src0))], itin
131 // If you add or change the operands for R600_2OP instructions, you must
132 // also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
133 // R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
134 class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
135 InstrItinClass itin = AnyALU> :
136 InstR600 <(outs R600_Reg32:$dst),
137 (ins UEM:$update_exec_mask, UP:$update_pred, WRITE:$write,
138 OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
139 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
140 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs, SEL:$src1_sel,
141 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
142 BANK_SWIZZLE:$bank_swizzle),
143 !strconcat(" ", opName,
144 "$clamp $last $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
145 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
146 "$src1_neg$src1_abs$src1$src1_abs$src1_rel, "
147 "$pred_sel $bank_swizzle"),
151 R600ALU_Word1_OP2 <inst> {
153 let HasNativeOperands = 1;
156 let DisableEncoding = "$literal";
157 let UseNamedOperandTable = 1;
159 let Inst{31-0} = Word0;
160 let Inst{63-32} = Word1;
163 class R600_2OP_Helper <bits<11> inst, string opName,
164 SDPatternOperator node = null_frag,
165 InstrItinClass itin = AnyALU> :
166 R600_2OP <inst, opName,
167 [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
168 R600_Reg32:$src1))], itin
171 // If you add our change the operands for R600_3OP instructions, you must
172 // also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
173 // R600InstrInfo::buildDefaultInstruction(), and
174 // R600InstrInfo::getOperandIdx().
175 class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
176 InstrItinClass itin = AnyALU> :
177 InstR600 <(outs R600_Reg32:$dst),
178 (ins REL:$dst_rel, CLAMP:$clamp,
179 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, SEL:$src0_sel,
180 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
181 R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
182 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
183 BANK_SWIZZLE:$bank_swizzle),
184 !strconcat(" ", opName, "$clamp $last $dst$dst_rel, "
185 "$src0_neg$src0$src0_rel, "
186 "$src1_neg$src1$src1_rel, "
187 "$src2_neg$src2$src2_rel, "
193 R600ALU_Word1_OP3<inst>{
195 let HasNativeOperands = 1;
196 let DisableEncoding = "$literal";
198 let UseNamedOperandTable = 1;
201 let Inst{31-0} = Word0;
202 let Inst{63-32} = Word1;
205 class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
206 InstrItinClass itin = VecALU> :
207 InstR600 <(outs R600_Reg32:$dst),
215 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
217 def TEX_SHADOW : PatLeaf<
219 [{uint32_t TType = (uint32_t)N->getZExtValue();
220 return (TType >= 6 && TType <= 8) || TType == 13;
224 def TEX_RECT : PatLeaf<
226 [{uint32_t TType = (uint32_t)N->getZExtValue();
231 def TEX_ARRAY : PatLeaf<
233 [{uint32_t TType = (uint32_t)N->getZExtValue();
234 return TType == 9 || TType == 10 || TType == 16;
238 def TEX_SHADOW_ARRAY : PatLeaf<
240 [{uint32_t TType = (uint32_t)N->getZExtValue();
241 return TType == 11 || TType == 12 || TType == 17;
245 class EG_CF_RAT <bits <8> cfinst, bits <6> ratinst, bits<4> ratid, bits<4> mask,
246 dag outs, dag ins, string asm, list<dag> pattern> :
247 InstR600ISA <outs, ins, asm, pattern>,
248 CF_ALLOC_EXPORT_WORD0_RAT, CF_ALLOC_EXPORT_WORD1_BUF {
251 let rat_inst = ratinst;
253 // XXX: Have a separate instruction for non-indexed writes.
259 let comp_mask = mask;
262 let cf_inst = cfinst;
266 let Inst{31-0} = Word0;
267 let Inst{63-32} = Word1;
272 class VTX_READ <string name, dag outs, list<dag> pattern>
273 : InstR600ISA <outs, (ins MEMxi:$src_gpr, i8imm:$buffer_id), !strconcat(" ", name, ", #$buffer_id"), pattern>,
278 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
279 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
280 // however, based on my testing if USE_CONST_FIELDS is set, then all
281 // these fields need to be set to 0.
282 let USE_CONST_FIELDS = 0;
283 let NUM_FORMAT_ALL = 1;
284 let FORMAT_COMP_ALL = 0;
285 let SRF_MODE_ALL = 0;
287 let Inst{63-32} = Word1;
288 // LLVM can only encode 64-bit instructions, so these fields are manually
289 // encoded in R600CodeEmitter
292 // bits<2> ENDIAN_SWAP = 0;
293 // bits<1> CONST_BUF_NO_STRIDE = 0;
294 // bits<1> MEGA_FETCH = 0;
295 // bits<1> ALT_CONST = 0;
296 // bits<2> BUFFER_INDEX_MODE = 0;
298 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
299 // is done in R600CodeEmitter
301 // Inst{79-64} = OFFSET;
302 // Inst{81-80} = ENDIAN_SWAP;
303 // Inst{82} = CONST_BUF_NO_STRIDE;
304 // Inst{83} = MEGA_FETCH;
305 // Inst{84} = ALT_CONST;
306 // Inst{86-85} = BUFFER_INDEX_MODE;
307 // Inst{95-86} = 0; Reserved
309 // VTX_WORD3 (Padding)
316 class LoadParamFrag <PatFrag load_type> : PatFrag <
317 (ops node:$ptr), (load_type node:$ptr),
318 [{ return isConstantLoad(cast<LoadSDNode>(N), 0) ||
319 (cast<LoadSDNode>(N)->getAddressSpace() == AMDGPUASI.PARAM_I_ADDRESS); }]
322 def vtx_id3_az_extloadi8 : LoadParamFrag<az_extloadi8>;
323 def vtx_id3_az_extloadi16 : LoadParamFrag<az_extloadi16>;
324 def vtx_id3_load : LoadParamFrag<load>;
326 class LoadVtxId1 <PatFrag load> : PatFrag <
327 (ops node:$ptr), (load node:$ptr), [{
328 const MemSDNode *LD = cast<MemSDNode>(N);
329 return LD->getAddressSpace() == AMDGPUASI.GLOBAL_ADDRESS ||
330 (LD->getAddressSpace() == AMDGPUASI.CONSTANT_ADDRESS &&
331 !isa<GlobalValue>(GetUnderlyingObject(
332 LD->getMemOperand()->getValue(), CurDAG->getDataLayout())));
335 def vtx_id1_az_extloadi8 : LoadVtxId1 <az_extloadi8>;
336 def vtx_id1_az_extloadi16 : LoadVtxId1 <az_extloadi16>;
337 def vtx_id1_load : LoadVtxId1 <load>;
339 class LoadVtxId2 <PatFrag load> : PatFrag <
340 (ops node:$ptr), (load node:$ptr), [{
341 const MemSDNode *LD = cast<MemSDNode>(N);
342 return LD->getAddressSpace() == AMDGPUASI.CONSTANT_ADDRESS &&
343 isa<GlobalValue>(GetUnderlyingObject(
344 LD->getMemOperand()->getValue(), CurDAG->getDataLayout()));
347 def vtx_id2_az_extloadi8 : LoadVtxId2 <az_extloadi8>;
348 def vtx_id2_az_extloadi16 : LoadVtxId2 <az_extloadi16>;
349 def vtx_id2_load : LoadVtxId2 <load>;
351 def isR600 : Predicate<"Subtarget->getGeneration() <= R600Subtarget::R700">;
355 "Subtarget->getGeneration() <= R600Subtarget::NORTHERN_ISLANDS">;
357 //===----------------------------------------------------------------------===//
359 //===----------------------------------------------------------------------===//
361 def INTERP_PAIR_XY : AMDGPUShaderInst <
362 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
363 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
364 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
367 def INTERP_PAIR_ZW : AMDGPUShaderInst <
368 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
369 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
370 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
373 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
374 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
378 def DOT4 : SDNode<"AMDGPUISD::DOT4",
379 SDTypeProfile<1, 8, [SDTCisFP<0>, SDTCisVT<1, f32>, SDTCisVT<2, f32>,
380 SDTCisVT<3, f32>, SDTCisVT<4, f32>, SDTCisVT<5, f32>,
381 SDTCisVT<6, f32>, SDTCisVT<7, f32>, SDTCisVT<8, f32>]>,
385 def COS_HW : SDNode<"AMDGPUISD::COS_HW",
386 SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisFP<1>]>
389 def SIN_HW : SDNode<"AMDGPUISD::SIN_HW",
390 SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisFP<1>]>
393 def TEXTURE_FETCH_Type : SDTypeProfile<1, 19, [SDTCisFP<0>]>;
395 def TEXTURE_FETCH: SDNode<"AMDGPUISD::TEXTURE_FETCH", TEXTURE_FETCH_Type, []>;
397 multiclass TexPattern<bits<32> TextureOp, Instruction inst, ValueType vt = v4f32> {
398 def : Pat<(TEXTURE_FETCH (i32 TextureOp), vt:$SRC_GPR,
399 (i32 imm:$srcx), (i32 imm:$srcy), (i32 imm:$srcz), (i32 imm:$srcw),
400 (i32 imm:$offsetx), (i32 imm:$offsety), (i32 imm:$offsetz),
401 (i32 imm:$DST_SEL_X), (i32 imm:$DST_SEL_Y), (i32 imm:$DST_SEL_Z),
402 (i32 imm:$DST_SEL_W),
403 (i32 imm:$RESOURCE_ID), (i32 imm:$SAMPLER_ID),
404 (i32 imm:$COORD_TYPE_X), (i32 imm:$COORD_TYPE_Y), (i32 imm:$COORD_TYPE_Z),
405 (i32 imm:$COORD_TYPE_W)),
406 (inst R600_Reg128:$SRC_GPR,
407 imm:$srcx, imm:$srcy, imm:$srcz, imm:$srcw,
408 imm:$offsetx, imm:$offsety, imm:$offsetz,
409 imm:$DST_SEL_X, imm:$DST_SEL_Y, imm:$DST_SEL_Z,
411 imm:$RESOURCE_ID, imm:$SAMPLER_ID,
412 imm:$COORD_TYPE_X, imm:$COORD_TYPE_Y, imm:$COORD_TYPE_Z,
416 //===----------------------------------------------------------------------===//
417 // Interpolation Instructions
418 //===----------------------------------------------------------------------===//
420 def INTERP_VEC_LOAD : AMDGPUShaderInst <
421 (outs R600_Reg128:$dst),
423 "INTERP_LOAD $src0 : $dst">;
425 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
426 let bank_swizzle = 5;
429 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
430 let bank_swizzle = 5;
433 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
435 //===----------------------------------------------------------------------===//
436 // Export Instructions
437 //===----------------------------------------------------------------------===//
440 field bits<32> Word0;
447 let Word0{12-0} = arraybase;
448 let Word0{14-13} = type;
449 let Word0{21-15} = gpr;
450 let Word0{22} = 0; // RW_REL
451 let Word0{29-23} = 0; // INDEX_GPR
452 let Word0{31-30} = elem_size;
455 class ExportSwzWord1 {
456 field bits<32> Word1;
465 let Word1{2-0} = sw_x;
466 let Word1{5-3} = sw_y;
467 let Word1{8-6} = sw_z;
468 let Word1{11-9} = sw_w;
471 class ExportBufWord1 {
472 field bits<32> Word1;
479 let Word1{11-0} = arraySize;
480 let Word1{15-12} = compMask;
483 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
484 def : Pat<(R600_EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
485 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
486 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
487 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
492 multiclass SteamOutputExportPattern<Instruction ExportInst,
493 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
495 def : Pat<(int_r600_store_stream_output (v4f32 R600_Reg128:$src),
496 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
497 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
498 4095, imm:$mask, buf0inst, 0)>;
500 def : Pat<(int_r600_store_stream_output (v4f32 R600_Reg128:$src),
501 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
502 (ExportInst $src, 0, imm:$arraybase,
503 4095, imm:$mask, buf1inst, 0)>;
505 def : Pat<(int_r600_store_stream_output (v4f32 R600_Reg128:$src),
506 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
507 (ExportInst $src, 0, imm:$arraybase,
508 4095, imm:$mask, buf2inst, 0)>;
510 def : Pat<(int_r600_store_stream_output (v4f32 R600_Reg128:$src),
511 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
512 (ExportInst $src, 0, imm:$arraybase,
513 4095, imm:$mask, buf3inst, 0)>;
516 // Export Instructions should not be duplicated by TailDuplication pass
517 // (which assumes that duplicable instruction are affected by exec mask)
518 let usesCustomInserter = 1, isNotDuplicable = 1 in {
520 class ExportSwzInst : InstR600ISA<(
522 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
523 RSel:$sw_x, RSel:$sw_y, RSel:$sw_z, RSel:$sw_w, i32imm:$inst,
525 !strconcat("EXPORT", " $gpr.$sw_x$sw_y$sw_z$sw_w"),
526 []>, ExportWord0, ExportSwzWord1 {
528 let Inst{31-0} = Word0;
529 let Inst{63-32} = Word1;
533 } // End usesCustomInserter = 1
535 class ExportBufInst : InstR600ISA<(
537 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
538 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
539 !strconcat("EXPORT", " $gpr"),
540 []>, ExportWord0, ExportBufWord1 {
542 let Inst{31-0} = Word0;
543 let Inst{63-32} = Word1;
547 //===----------------------------------------------------------------------===//
548 // Control Flow Instructions
549 //===----------------------------------------------------------------------===//
552 def KCACHE : InstFlag<"printKCache">;
554 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
555 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1,
556 KCACHE:$KCACHE_MODE0, KCACHE:$KCACHE_MODE1,
557 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1,
558 i32imm:$COUNT, i32imm:$Enabled),
559 !strconcat(OpName, " $COUNT, @$ADDR, "
560 "KC0[$KCACHE_MODE0], KC1[$KCACHE_MODE1]"),
561 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
566 let WHOLE_QUAD_MODE = 0;
568 let isCodeGenOnly = 1;
569 let UseNamedOperandTable = 1;
571 let Inst{31-0} = Word0;
572 let Inst{63-32} = Word1;
575 class CF_WORD0_R600 {
576 field bits<32> Word0;
583 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
584 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
591 let VALID_PIXEL_MODE = 0;
593 let COUNT = CNT{2-0};
595 let COUNT_3 = CNT{3};
596 let END_OF_PROGRAM = 0;
597 let WHOLE_QUAD_MODE = 0;
599 let Inst{31-0} = Word0;
600 let Inst{63-32} = Word1;
603 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
604 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
609 let JUMPTABLE_SEL = 0;
611 let VALID_PIXEL_MODE = 0;
613 let END_OF_PROGRAM = 0;
615 let Inst{31-0} = Word0;
616 let Inst{63-32} = Word1;
619 def CF_ALU : ALU_CLAUSE<8, "ALU">;
620 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
621 def CF_ALU_POP_AFTER : ALU_CLAUSE<10, "ALU_POP_AFTER">;
622 def CF_ALU_CONTINUE : ALU_CLAUSE<13, "ALU_CONTINUE">;
623 def CF_ALU_BREAK : ALU_CLAUSE<14, "ALU_BREAK">;
624 def CF_ALU_ELSE_AFTER : ALU_CLAUSE<15, "ALU_ELSE_AFTER">;
626 def FETCH_CLAUSE : AMDGPUInst <(outs),
627 (ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
631 let isCodeGenOnly = 1;
634 def ALU_CLAUSE : AMDGPUInst <(outs),
635 (ins i32imm:$addr), "ALU clause starting at $addr:", [] > {
639 let isCodeGenOnly = 1;
642 def LITERALS : AMDGPUInst <(outs),
643 (ins LITERAL:$literal1, LITERAL:$literal2), "$literal1, $literal2", [] > {
644 let isCodeGenOnly = 1;
650 let Inst{31-0} = literal1;
651 let Inst{63-32} = literal2;
654 def PAD : AMDGPUInst <(outs), (ins), "PAD", [] > {
658 let Predicates = [isR600toCayman] in {
660 //===----------------------------------------------------------------------===//
661 // Common Instructions R600, R700, Evergreen, Cayman
662 //===----------------------------------------------------------------------===//
664 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
665 // Non-IEEE MUL: 0 * anything = 0
666 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE">;
667 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
668 // TODO: Do these actually match the regular fmin/fmax behavior?
669 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax_legacy>;
670 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin_legacy>;
671 // According to https://msdn.microsoft.com/en-us/library/windows/desktop/cc308050%28v=vs.85%29.aspx
672 // DX10 min/max returns the other operand if one is NaN,
673 // this matches http://llvm.org/docs/LangRef.html#llvm-minnum-intrinsic
674 def MAX_DX10 : R600_2OP_Helper <0x5, "MAX_DX10", fmaxnum>;
675 def MIN_DX10 : R600_2OP_Helper <0x6, "MIN_DX10", fminnum>;
677 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
678 // so some of the instruction names don't match the asm string.
679 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
680 def SETE : R600_2OP <
682 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_OEQ))]
687 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_OGT))]
692 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_OGE))]
697 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_UNE_NE))]
700 def SETE_DX10 : R600_2OP <
702 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_OEQ))]
705 def SETGT_DX10 : R600_2OP <
707 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_OGT))]
710 def SETGE_DX10 : R600_2OP <
712 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_OGE))]
715 // FIXME: This should probably be COND_ONE
716 def SETNE_DX10 : R600_2OP <
718 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_UNE_NE))]
721 // FIXME: Need combine for AMDGPUfract
722 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
723 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", ftrunc>;
724 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
725 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
726 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
728 def MOV : R600_1OP <0x19, "MOV", []>;
731 // This is a hack to get rid of DUMMY_CHAIN nodes.
732 // Most DUMMY_CHAINs should be eliminated during legalization, but undef
733 // values can sneak in some to selection.
734 let isPseudo = 1, isCodeGenOnly = 1 in {
735 def DUMMY_CHAIN : AMDGPUInst <
741 } // end let isPseudo = 1, isCodeGenOnly = 1
744 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
746 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
747 (outs R600_Reg32:$dst),
753 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
755 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
758 (MOV_IMM_I32 imm:$val)
761 def MOV_IMM_GLOBAL_ADDR : MOV_IMM<iPTR, i32imm>;
763 (AMDGPUconstdata_ptr tglobaladdr:$addr),
764 (MOV_IMM_GLOBAL_ADDR tglobaladdr:$addr)
768 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
771 (MOV_IMM_F32 fpimm:$val)
774 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
775 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
776 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
777 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
779 let hasSideEffects = 1 in {
781 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
783 } // end hasSideEffects
785 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
786 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
787 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
788 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
789 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
790 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
791 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", smax>;
792 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", smin>;
793 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", umax>;
794 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", umin>;
796 def SETE_INT : R600_2OP <
798 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETEQ))]
801 def SETGT_INT : R600_2OP <
803 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGT))]
806 def SETGE_INT : R600_2OP <
808 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGE))]
811 def SETNE_INT : R600_2OP <
813 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETNE))]
816 def SETGT_UINT : R600_2OP <
818 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGT))]
821 def SETGE_UINT : R600_2OP <
823 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGE))]
826 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
827 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
828 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
829 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
831 def CNDE_INT : R600_3OP <
833 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_EQ))]
836 def CNDGE_INT : R600_3OP <
838 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_SGE))]
841 def CNDGT_INT : R600_3OP <
843 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_SGT))]
846 //===----------------------------------------------------------------------===//
847 // Texture instructions
848 //===----------------------------------------------------------------------===//
850 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
852 class R600_TEX <bits<11> inst, string opName> :
853 InstR600 <(outs R600_Reg128:$DST_GPR),
854 (ins R600_Reg128:$SRC_GPR,
855 RSel:$srcx, RSel:$srcy, RSel:$srcz, RSel:$srcw,
856 i32imm:$offsetx, i32imm:$offsety, i32imm:$offsetz,
857 RSel:$DST_SEL_X, RSel:$DST_SEL_Y, RSel:$DST_SEL_Z, RSel:$DST_SEL_W,
858 i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
859 CT:$COORD_TYPE_X, CT:$COORD_TYPE_Y, CT:$COORD_TYPE_Z,
861 !strconcat(" ", opName,
862 " $DST_GPR.$DST_SEL_X$DST_SEL_Y$DST_SEL_Z$DST_SEL_W, "
863 "$SRC_GPR.$srcx$srcy$srcz$srcw "
864 "RID:$RESOURCE_ID SID:$SAMPLER_ID "
865 "CT:$COORD_TYPE_X$COORD_TYPE_Y$COORD_TYPE_Z$COORD_TYPE_W"),
867 NullALU>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
868 let Inst{31-0} = Word0;
869 let Inst{63-32} = Word1;
871 let TEX_INST = inst{4-0};
877 let FETCH_WHOLE_QUAD = 0;
879 let SAMPLER_INDEX_MODE = 0;
880 let RESOURCE_INDEX_MODE = 0;
885 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
889 def TEX_SAMPLE : R600_TEX <0x10, "TEX_SAMPLE">;
890 def TEX_SAMPLE_C : R600_TEX <0x18, "TEX_SAMPLE_C">;
891 def TEX_SAMPLE_L : R600_TEX <0x11, "TEX_SAMPLE_L">;
892 def TEX_SAMPLE_C_L : R600_TEX <0x19, "TEX_SAMPLE_C_L">;
893 def TEX_SAMPLE_LB : R600_TEX <0x12, "TEX_SAMPLE_LB">;
894 def TEX_SAMPLE_C_LB : R600_TEX <0x1A, "TEX_SAMPLE_C_LB">;
895 def TEX_LD : R600_TEX <0x03, "TEX_LD">;
896 def TEX_LDPTR : R600_TEX <0x03, "TEX_LDPTR"> {
899 def TEX_GET_TEXTURE_RESINFO : R600_TEX <0x04, "TEX_GET_TEXTURE_RESINFO">;
900 def TEX_GET_GRADIENTS_H : R600_TEX <0x07, "TEX_GET_GRADIENTS_H">;
901 def TEX_GET_GRADIENTS_V : R600_TEX <0x08, "TEX_GET_GRADIENTS_V">;
902 def TEX_SET_GRADIENTS_H : R600_TEX <0x0B, "TEX_SET_GRADIENTS_H">;
903 def TEX_SET_GRADIENTS_V : R600_TEX <0x0C, "TEX_SET_GRADIENTS_V">;
904 def TEX_SAMPLE_G : R600_TEX <0x14, "TEX_SAMPLE_G">;
905 def TEX_SAMPLE_C_G : R600_TEX <0x1C, "TEX_SAMPLE_C_G">;
907 defm : TexPattern<0, TEX_SAMPLE>;
908 defm : TexPattern<1, TEX_SAMPLE_C>;
909 defm : TexPattern<2, TEX_SAMPLE_L>;
910 defm : TexPattern<3, TEX_SAMPLE_C_L>;
911 defm : TexPattern<4, TEX_SAMPLE_LB>;
912 defm : TexPattern<5, TEX_SAMPLE_C_LB>;
913 defm : TexPattern<6, TEX_LD, v4i32>;
914 defm : TexPattern<7, TEX_GET_TEXTURE_RESINFO, v4i32>;
915 defm : TexPattern<8, TEX_GET_GRADIENTS_H>;
916 defm : TexPattern<9, TEX_GET_GRADIENTS_V>;
917 defm : TexPattern<10, TEX_LDPTR, v4i32>;
919 //===----------------------------------------------------------------------===//
920 // Helper classes for common instructions
921 //===----------------------------------------------------------------------===//
923 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
928 class MULADD_Common <bits<5> inst> : R600_3OP <
933 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
935 [(set f32:$dst, (fmad f32:$src0, f32:$src1, f32:$src2))]
938 class FMA_Common <bits<5> inst> : R600_3OP <
940 [(set f32:$dst, (fma f32:$src0, f32:$src1, f32:$src2))], VecALU
943 class CNDE_Common <bits<5> inst> : R600_3OP <
945 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_OEQ))]
948 class CNDGT_Common <bits<5> inst> : R600_3OP <
950 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_OGT))]
952 let Itinerary = VecALU;
955 class CNDGE_Common <bits<5> inst> : R600_3OP <
957 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_OGE))]
959 let Itinerary = VecALU;
963 let isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
964 class R600_VEC2OP<list<dag> pattern> : InstR600 <(outs R600_Reg32:$dst), (ins
966 UEM:$update_exec_mask_X, UP:$update_pred_X, WRITE:$write_X,
967 OMOD:$omod_X, REL:$dst_rel_X, CLAMP:$clamp_X,
968 R600_TReg32_X:$src0_X, NEG:$src0_neg_X, REL:$src0_rel_X, ABS:$src0_abs_X, SEL:$src0_sel_X,
969 R600_TReg32_X:$src1_X, NEG:$src1_neg_X, REL:$src1_rel_X, ABS:$src1_abs_X, SEL:$src1_sel_X,
970 R600_Pred:$pred_sel_X,
972 UEM:$update_exec_mask_Y, UP:$update_pred_Y, WRITE:$write_Y,
973 OMOD:$omod_Y, REL:$dst_rel_Y, CLAMP:$clamp_Y,
974 R600_TReg32_Y:$src0_Y, NEG:$src0_neg_Y, REL:$src0_rel_Y, ABS:$src0_abs_Y, SEL:$src0_sel_Y,
975 R600_TReg32_Y:$src1_Y, NEG:$src1_neg_Y, REL:$src1_rel_Y, ABS:$src1_abs_Y, SEL:$src1_sel_Y,
976 R600_Pred:$pred_sel_Y,
978 UEM:$update_exec_mask_Z, UP:$update_pred_Z, WRITE:$write_Z,
979 OMOD:$omod_Z, REL:$dst_rel_Z, CLAMP:$clamp_Z,
980 R600_TReg32_Z:$src0_Z, NEG:$src0_neg_Z, REL:$src0_rel_Z, ABS:$src0_abs_Z, SEL:$src0_sel_Z,
981 R600_TReg32_Z:$src1_Z, NEG:$src1_neg_Z, REL:$src1_rel_Z, ABS:$src1_abs_Z, SEL:$src1_sel_Z,
982 R600_Pred:$pred_sel_Z,
984 UEM:$update_exec_mask_W, UP:$update_pred_W, WRITE:$write_W,
985 OMOD:$omod_W, REL:$dst_rel_W, CLAMP:$clamp_W,
986 R600_TReg32_W:$src0_W, NEG:$src0_neg_W, REL:$src0_rel_W, ABS:$src0_abs_W, SEL:$src0_sel_W,
987 R600_TReg32_W:$src1_W, NEG:$src1_neg_W, REL:$src1_rel_W, ABS:$src1_abs_W, SEL:$src1_sel_W,
988 R600_Pred:$pred_sel_W,
989 LITERAL:$literal0, LITERAL:$literal1),
994 let UseNamedOperandTable = 1;
999 def DOT_4 : R600_VEC2OP<[(set R600_Reg32:$dst, (DOT4
1000 R600_TReg32_X:$src0_X, R600_TReg32_X:$src1_X,
1001 R600_TReg32_Y:$src0_Y, R600_TReg32_Y:$src1_Y,
1002 R600_TReg32_Z:$src0_Z, R600_TReg32_Z:$src1_Z,
1003 R600_TReg32_W:$src0_W, R600_TReg32_W:$src1_W))]>;
1006 class DOT4_Common <bits<11> inst> : R600_2OP <inst, "DOT4", []>;
1009 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
1010 multiclass CUBE_Common <bits<11> inst> {
1012 def _pseudo : InstR600 <
1013 (outs R600_Reg128:$dst),
1014 (ins R600_Reg128:$src0),
1016 [(set v4f32:$dst, (int_r600_cube v4f32:$src0))],
1020 let UseNamedOperandTable = 1;
1023 def _real : R600_2OP <inst, "CUBE", []>;
1025 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
1027 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1028 inst, "EXP_IEEE", fexp2
1030 let Itinerary = TransALU;
1033 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
1034 inst, "FLT_TO_INT", fp_to_sint
1036 let Itinerary = TransALU;
1039 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1040 inst, "INT_TO_FLT", sint_to_fp
1042 let Itinerary = TransALU;
1045 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1046 inst, "FLT_TO_UINT", fp_to_uint
1048 let Itinerary = TransALU;
1051 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1052 inst, "UINT_TO_FLT", uint_to_fp
1054 let Itinerary = TransALU;
1057 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1058 inst, "LOG_CLAMPED", []
1061 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1062 inst, "LOG_IEEE", flog2
1064 let Itinerary = TransALU;
1067 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1068 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1069 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1070 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1071 inst, "MULHI_INT", mulhs> {
1072 let Itinerary = TransALU;
1075 class MULHI_INT24_Common <bits<11> inst> : R600_2OP_Helper <
1076 inst, "MULHI_INT24", AMDGPUmulhi_i24> {
1077 let Itinerary = VecALU;
1080 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1081 inst, "MULHI", mulhu> {
1082 let Itinerary = TransALU;
1085 class MULHI_UINT24_Common <bits<11> inst> : R600_2OP_Helper <
1086 inst, "MULHI_UINT24", AMDGPUmulhi_u24> {
1087 let Itinerary = VecALU;
1090 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1091 inst, "MULLO_INT", mul> {
1092 let Itinerary = TransALU;
1094 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []> {
1095 let Itinerary = TransALU;
1098 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1099 inst, "RECIP_CLAMPED", []
1101 let Itinerary = TransALU;
1104 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1105 inst, "RECIP_IEEE", [(set f32:$dst, (AMDGPUrcp f32:$src0))]
1107 let Itinerary = TransALU;
1110 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1111 inst, "RECIP_UINT", AMDGPUurecip
1113 let Itinerary = TransALU;
1116 // Clamped to maximum.
1117 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1118 inst, "RECIPSQRT_CLAMPED", AMDGPUrsq_clamp
1120 let Itinerary = TransALU;
1123 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1124 inst, "RECIPSQRT_IEEE", AMDGPUrsq> {
1125 let Itinerary = TransALU;
1128 // TODO: There is also RECIPSQRT_FF which clamps to zero.
1130 class SIN_Common <bits<11> inst> : R600_1OP <
1131 inst, "SIN", [(set f32:$dst, (SIN_HW f32:$src0))]>{
1133 let Itinerary = TransALU;
1136 class COS_Common <bits<11> inst> : R600_1OP <
1137 inst, "COS", [(set f32:$dst, (COS_HW f32:$src0))]> {
1139 let Itinerary = TransALU;
1142 def CLAMP_R600 : CLAMP <R600_Reg32>;
1143 def FABS_R600 : FABS<R600_Reg32>;
1144 def FNEG_R600 : FNEG<R600_Reg32>;
1146 //===----------------------------------------------------------------------===//
1147 // Helper patterns for complex intrinsics
1148 //===----------------------------------------------------------------------===//
1150 // FIXME: Should be predicated on unsafe fp math.
1151 multiclass DIV_Common <InstR600 recip_ieee> {
1153 (fdiv f32:$src0, f32:$src1),
1154 (MUL_IEEE $src0, (recip_ieee $src1))
1157 def : RcpPat<recip_ieee, f32>;
1160 //===----------------------------------------------------------------------===//
1161 // R600 / R700 Instructions
1162 //===----------------------------------------------------------------------===//
1164 let Predicates = [isR600] in {
1166 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1167 def MULADD_r600 : MULADD_Common<0x10>;
1168 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1169 def CNDE_r600 : CNDE_Common<0x18>;
1170 def CNDGT_r600 : CNDGT_Common<0x19>;
1171 def CNDGE_r600 : CNDGE_Common<0x1A>;
1172 def DOT4_r600 : DOT4_Common<0x50>;
1173 defm CUBE_r600 : CUBE_Common<0x52>;
1174 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1175 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1176 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1177 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1178 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1179 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1180 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1181 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1182 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1183 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1184 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1185 def SIN_r600 : SIN_Common<0x6E>;
1186 def COS_r600 : COS_Common<0x6F>;
1187 def ASHR_r600 : ASHR_Common<0x70>;
1188 def LSHR_r600 : LSHR_Common<0x71>;
1189 def LSHL_r600 : LSHL_Common<0x72>;
1190 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1191 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1192 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1193 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1194 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1196 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1197 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL>;
1199 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
1200 def : RsqPat<RECIPSQRT_IEEE_r600, f32>;
1202 def R600_ExportSwz : ExportSwzInst {
1203 let Word1{20-17} = 0; // BURST_COUNT
1204 let Word1{21} = eop;
1205 let Word1{22} = 0; // VALID_PIXEL_MODE
1206 let Word1{30-23} = inst;
1207 let Word1{31} = 1; // BARRIER
1209 defm : ExportPattern<R600_ExportSwz, 39>;
1211 def R600_ExportBuf : ExportBufInst {
1212 let Word1{20-17} = 0; // BURST_COUNT
1213 let Word1{21} = eop;
1214 let Word1{22} = 0; // VALID_PIXEL_MODE
1215 let Word1{30-23} = inst;
1216 let Word1{31} = 1; // BARRIER
1218 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1220 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$CNT),
1221 "TEX $CNT @$ADDR"> {
1224 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$CNT),
1225 "VTX $CNT @$ADDR"> {
1228 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1229 "LOOP_START_DX10 @$ADDR"> {
1233 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1237 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1238 "LOOP_BREAK @$ADDR"> {
1242 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1243 "CONTINUE @$ADDR"> {
1247 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1248 "JUMP @$ADDR POP:$POP_COUNT"> {
1251 def CF_PUSH_ELSE_R600 : CF_CLAUSE_R600<12, (ins i32imm:$ADDR),
1252 "PUSH_ELSE @$ADDR"> {
1254 let POP_COUNT = 0; // FIXME?
1256 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1257 "ELSE @$ADDR POP:$POP_COUNT"> {
1260 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1265 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1266 "POP @$ADDR POP:$POP_COUNT"> {
1269 def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
1273 let END_OF_PROGRAM = 1;
1279 //===----------------------------------------------------------------------===//
1280 // Regist loads and stores - for indirect addressing
1281 //===----------------------------------------------------------------------===//
1283 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1285 // Hardcode channel to 0
1286 // NOTE: LSHR is not available here. LSHR is per family instruction
1288 (i32 (load_private ADDRIndirect:$addr) ),
1289 (R600_RegisterLoad FRAMEri:$addr, (i32 0))
1292 (store_private i32:$val, ADDRIndirect:$addr),
1293 (R600_RegisterStore i32:$val, FRAMEri:$addr, (i32 0))
1297 //===----------------------------------------------------------------------===//
1298 // Pseudo instructions
1299 //===----------------------------------------------------------------------===//
1301 let isPseudo = 1 in {
1303 def PRED_X : InstR600 <
1304 (outs R600_Predicate_Bit:$dst),
1305 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
1307 let FlagOperandIdx = 3;
1310 let isTerminator = 1, isBranch = 1 in {
1311 def JUMP_COND : InstR600 <
1313 (ins brtarget:$target, R600_Predicate_Bit:$p),
1314 "JUMP $target ($p)",
1318 def JUMP : InstR600 <
1320 (ins brtarget:$target),
1325 let isPredicable = 1;
1329 } // End isTerminator = 1, isBranch = 1
1331 let usesCustomInserter = 1 in {
1333 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
1335 def MASK_WRITE : AMDGPUShaderInst <
1337 (ins R600_Reg32:$src),
1342 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
1346 (outs R600_Reg128:$dst),
1347 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
1348 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1349 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget", [],
1354 def TXD_SHADOW: InstR600 <
1355 (outs R600_Reg128:$dst),
1356 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
1357 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1358 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1362 } // End isPseudo = 1
1363 } // End usesCustomInserter = 1
1366 //===----------------------------------------------------------------------===//
1367 // Constant Buffer Addressing Support
1368 //===----------------------------------------------------------------------===//
1370 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
1371 def CONST_COPY : Instruction {
1372 let OutOperandList = (outs R600_Reg32:$dst);
1373 let InOperandList = (ins i32imm:$src);
1375 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
1376 let AsmString = "CONST_COPY";
1377 let hasSideEffects = 0;
1378 let isAsCheapAsAMove = 1;
1379 let Itinerary = NullALU;
1381 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
1383 def TEX_VTX_CONSTBUF :
1384 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$buffer_id), "VTX_READ_eg $dst, $ptr",
1385 [(set v4i32:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$buffer_id)))]>,
1386 VTX_WORD1_GPR, VTX_WORD0_eg {
1390 let FETCH_WHOLE_QUAD = 0;
1394 let USE_CONST_FIELDS = 0;
1395 let NUM_FORMAT_ALL = 2;
1396 let FORMAT_COMP_ALL = 1;
1397 let SRF_MODE_ALL = 1;
1398 let MEGA_FETCH_COUNT = 16;
1403 let DATA_FORMAT = 35;
1405 let Inst{31-0} = Word0;
1406 let Inst{63-32} = Word1;
1408 // LLVM can only encode 64-bit instructions, so these fields are manually
1409 // encoded in R600CodeEmitter
1412 // bits<2> ENDIAN_SWAP = 0;
1413 // bits<1> CONST_BUF_NO_STRIDE = 0;
1414 // bits<1> MEGA_FETCH = 0;
1415 // bits<1> ALT_CONST = 0;
1416 // bits<2> BUFFER_INDEX_MODE = 0;
1420 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1421 // is done in R600CodeEmitter
1423 // Inst{79-64} = OFFSET;
1424 // Inst{81-80} = ENDIAN_SWAP;
1425 // Inst{82} = CONST_BUF_NO_STRIDE;
1426 // Inst{83} = MEGA_FETCH;
1427 // Inst{84} = ALT_CONST;
1428 // Inst{86-85} = BUFFER_INDEX_MODE;
1429 // Inst{95-86} = 0; Reserved
1431 // VTX_WORD3 (Padding)
1433 // Inst{127-96} = 0;
1438 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$buffer_id), "TEX_VTX_EXPLICIT_READ $dst, $ptr">,
1439 VTX_WORD1_GPR, VTX_WORD0_eg {
1443 let FETCH_WHOLE_QUAD = 0;
1447 let USE_CONST_FIELDS = 1;
1448 let NUM_FORMAT_ALL = 0;
1449 let FORMAT_COMP_ALL = 0;
1450 let SRF_MODE_ALL = 1;
1451 let MEGA_FETCH_COUNT = 16;
1456 let DATA_FORMAT = 0;
1458 let Inst{31-0} = Word0;
1459 let Inst{63-32} = Word1;
1461 // LLVM can only encode 64-bit instructions, so these fields are manually
1462 // encoded in R600CodeEmitter
1465 // bits<2> ENDIAN_SWAP = 0;
1466 // bits<1> CONST_BUF_NO_STRIDE = 0;
1467 // bits<1> MEGA_FETCH = 0;
1468 // bits<1> ALT_CONST = 0;
1469 // bits<2> BUFFER_INDEX_MODE = 0;
1473 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1474 // is done in R600CodeEmitter
1476 // Inst{79-64} = OFFSET;
1477 // Inst{81-80} = ENDIAN_SWAP;
1478 // Inst{82} = CONST_BUF_NO_STRIDE;
1479 // Inst{83} = MEGA_FETCH;
1480 // Inst{84} = ALT_CONST;
1481 // Inst{86-85} = BUFFER_INDEX_MODE;
1482 // Inst{95-86} = 0; Reserved
1484 // VTX_WORD3 (Padding)
1486 // Inst{127-96} = 0;
1490 //===---------------------------------------------------------------------===//
1491 // Flow and Program control Instructions
1492 //===---------------------------------------------------------------------===//
1493 class ILFormat<dag outs, dag ins, string asmstr, list<dag> pattern>
1496 let Namespace = "AMDGPU";
1497 dag OutOperandList = outs;
1498 dag InOperandList = ins;
1499 let Pattern = pattern;
1500 let AsmString = !strconcat(asmstr, "\n");
1502 let Itinerary = NullALU;
1503 bit hasIEEEFlag = 0;
1504 bit hasZeroOpFlag = 0;
1507 let hasSideEffects = 0;
1508 let isCodeGenOnly = 1;
1511 multiclass BranchConditional<SDNode Op, RegisterClass rci, RegisterClass rcf> {
1512 def _i32 : ILFormat<(outs),
1513 (ins brtarget:$target, rci:$src0),
1514 "; i32 Pseudo branch instruction",
1515 [(Op bb:$target, (i32 rci:$src0))]>;
1516 def _f32 : ILFormat<(outs),
1517 (ins brtarget:$target, rcf:$src0),
1518 "; f32 Pseudo branch instruction",
1519 [(Op bb:$target, (f32 rcf:$src0))]>;
1522 // Only scalar types should generate flow control
1523 multiclass BranchInstr<string name> {
1524 def _i32 : ILFormat<(outs), (ins R600_Reg32:$src),
1525 !strconcat(name, " $src"), []>;
1526 def _f32 : ILFormat<(outs), (ins R600_Reg32:$src),
1527 !strconcat(name, " $src"), []>;
1529 // Only scalar types should generate flow control
1530 multiclass BranchInstr2<string name> {
1531 def _i32 : ILFormat<(outs), (ins R600_Reg32:$src0, R600_Reg32:$src1),
1532 !strconcat(name, " $src0, $src1"), []>;
1533 def _f32 : ILFormat<(outs), (ins R600_Reg32:$src0, R600_Reg32:$src1),
1534 !strconcat(name, " $src0, $src1"), []>;
1537 //===---------------------------------------------------------------------===//
1538 // Custom Inserter for Branches and returns, this eventually will be a
1540 //===---------------------------------------------------------------------===//
1541 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
1542 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
1543 "; Pseudo unconditional branch instruction",
1545 defm BRANCH_COND : BranchConditional<IL_brcond, R600_Reg32, R600_Reg32>;
1548 //===---------------------------------------------------------------------===//
1549 // Return instruction
1550 //===---------------------------------------------------------------------===//
1551 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
1552 usesCustomInserter = 1 in {
1553 def RETURN : ILFormat<(outs), (ins variable_ops),
1554 "RETURN", [(AMDGPUendpgm)]
1558 //===----------------------------------------------------------------------===//
1559 // Branch Instructions
1560 //===----------------------------------------------------------------------===//
1562 def IF_PREDICATE_SET : ILFormat<(outs), (ins R600_Reg32:$src),
1563 "IF_PREDICATE_SET $src", []>;
1565 let isTerminator=1 in {
1566 def BREAK : ILFormat< (outs), (ins),
1568 def CONTINUE : ILFormat< (outs), (ins),
1570 def DEFAULT : ILFormat< (outs), (ins),
1572 def ELSE : ILFormat< (outs), (ins),
1574 def ENDSWITCH : ILFormat< (outs), (ins),
1576 def ENDMAIN : ILFormat< (outs), (ins),
1578 def END : ILFormat< (outs), (ins),
1580 def ENDFUNC : ILFormat< (outs), (ins),
1582 def ENDIF : ILFormat< (outs), (ins),
1584 def WHILELOOP : ILFormat< (outs), (ins),
1586 def ENDLOOP : ILFormat< (outs), (ins),
1588 def FUNC : ILFormat< (outs), (ins),
1590 def RETDYN : ILFormat< (outs), (ins),
1592 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1593 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
1594 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1595 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
1596 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1597 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
1598 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1599 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
1600 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1601 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
1602 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1603 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
1604 defm IFC : BranchInstr2<"IFC">;
1605 defm BREAKC : BranchInstr2<"BREAKC">;
1606 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
1609 //===----------------------------------------------------------------------===//
1610 // Indirect addressing pseudo instructions
1611 //===----------------------------------------------------------------------===//
1613 let isPseudo = 1 in {
1615 class ExtractVertical <RegisterClass vec_rc> : InstR600 <
1616 (outs R600_Reg32:$dst),
1617 (ins vec_rc:$vec, R600_Reg32:$index), "",
1622 let Constraints = "$dst = $vec" in {
1624 class InsertVertical <RegisterClass vec_rc> : InstR600 <
1626 (ins vec_rc:$vec, R600_Reg32:$value, R600_Reg32:$index), "",
1631 } // End Constraints = "$dst = $vec"
1633 } // End isPseudo = 1
1635 def R600_EXTRACT_ELT_V2 : ExtractVertical <R600_Reg64Vertical>;
1636 def R600_EXTRACT_ELT_V4 : ExtractVertical <R600_Reg128Vertical>;
1638 def R600_INSERT_ELT_V2 : InsertVertical <R600_Reg64Vertical>;
1639 def R600_INSERT_ELT_V4 : InsertVertical <R600_Reg128Vertical>;
1641 class ExtractVerticalPat <Instruction inst, ValueType vec_ty,
1642 ValueType scalar_ty> : Pat <
1643 (scalar_ty (extractelt vec_ty:$vec, i32:$index)),
1647 def : ExtractVerticalPat <R600_EXTRACT_ELT_V2, v2i32, i32>;
1648 def : ExtractVerticalPat <R600_EXTRACT_ELT_V2, v2f32, f32>;
1649 def : ExtractVerticalPat <R600_EXTRACT_ELT_V4, v4i32, i32>;
1650 def : ExtractVerticalPat <R600_EXTRACT_ELT_V4, v4f32, f32>;
1652 class InsertVerticalPat <Instruction inst, ValueType vec_ty,
1653 ValueType scalar_ty> : Pat <
1654 (vec_ty (insertelt vec_ty:$vec, scalar_ty:$value, i32:$index)),
1655 (inst $vec, $value, $index)
1658 def : InsertVerticalPat <R600_INSERT_ELT_V2, v2i32, i32>;
1659 def : InsertVerticalPat <R600_INSERT_ELT_V2, v2f32, f32>;
1660 def : InsertVerticalPat <R600_INSERT_ELT_V4, v4i32, i32>;
1661 def : InsertVerticalPat <R600_INSERT_ELT_V4, v4f32, f32>;
1663 //===----------------------------------------------------------------------===//
1665 //===----------------------------------------------------------------------===//
1667 // CND*_INT Patterns for f32 True / False values
1669 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
1670 (selectcc i32:$src0, 0, f32:$src1, f32:$src2, cc),
1671 (cnd $src0, $src1, $src2)
1674 def : CND_INT_f32 <CNDE_INT, SETEQ>;
1675 def : CND_INT_f32 <CNDGT_INT, SETGT>;
1676 def : CND_INT_f32 <CNDGE_INT, SETGE>;
1678 //CNDGE_INT extra pattern
1680 (selectcc i32:$src0, -1, i32:$src1, i32:$src2, COND_SGT),
1681 (CNDGE_INT $src0, $src1, $src2)
1687 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
1691 (int_AMDGPU_kill f32:$src0),
1692 (MASK_WRITE (KILLGT (f32 ZERO), $src0))
1695 def : Extract_Element <f32, v4f32, 0, sub0>;
1696 def : Extract_Element <f32, v4f32, 1, sub1>;
1697 def : Extract_Element <f32, v4f32, 2, sub2>;
1698 def : Extract_Element <f32, v4f32, 3, sub3>;
1700 def : Insert_Element <f32, v4f32, 0, sub0>;
1701 def : Insert_Element <f32, v4f32, 1, sub1>;
1702 def : Insert_Element <f32, v4f32, 2, sub2>;
1703 def : Insert_Element <f32, v4f32, 3, sub3>;
1705 def : Extract_Element <i32, v4i32, 0, sub0>;
1706 def : Extract_Element <i32, v4i32, 1, sub1>;
1707 def : Extract_Element <i32, v4i32, 2, sub2>;
1708 def : Extract_Element <i32, v4i32, 3, sub3>;
1710 def : Insert_Element <i32, v4i32, 0, sub0>;
1711 def : Insert_Element <i32, v4i32, 1, sub1>;
1712 def : Insert_Element <i32, v4i32, 2, sub2>;
1713 def : Insert_Element <i32, v4i32, 3, sub3>;
1715 def : Extract_Element <f32, v2f32, 0, sub0>;
1716 def : Extract_Element <f32, v2f32, 1, sub1>;
1718 def : Insert_Element <f32, v2f32, 0, sub0>;
1719 def : Insert_Element <f32, v2f32, 1, sub1>;
1721 def : Extract_Element <i32, v2i32, 0, sub0>;
1722 def : Extract_Element <i32, v2i32, 1, sub1>;
1724 def : Insert_Element <i32, v2i32, 0, sub0>;
1725 def : Insert_Element <i32, v2i32, 1, sub1>;
1727 // bitconvert patterns
1729 def : BitConvert <i32, f32, R600_Reg32>;
1730 def : BitConvert <f32, i32, R600_Reg32>;
1731 def : BitConvert <v2f32, v2i32, R600_Reg64>;
1732 def : BitConvert <v2i32, v2f32, R600_Reg64>;
1733 def : BitConvert <v4f32, v4i32, R600_Reg128>;
1734 def : BitConvert <v4i32, v4f32, R600_Reg128>;
1736 // DWORDADDR pattern
1737 def : DwordAddrPat <i32, R600_Reg32>;
1739 } // End isR600toCayman Predicate
1741 def getLDSNoRetOp : InstrMapping {
1742 let FilterClass = "R600_LDS_1A1D";
1743 let RowFields = ["BaseOp"];
1744 let ColFields = ["DisableEncoding"];
1745 let KeyCol = ["$dst"];
1746 let ValueCols = [[""""]];