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"
15 include "R600InstrFormats.td"
17 class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
18 InstR600 <outs, ins, asm, pattern, NullALU> {
20 let Namespace = "AMDGPU";
23 def MEMxi : Operand<iPTR> {
24 let MIOperandInfo = (ops R600_TReg32_X:$ptr, i32imm:$index);
25 let PrintMethod = "printMemOperand";
28 def MEMrr : Operand<iPTR> {
29 let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
32 // Operands for non-registers
34 class InstFlag<string PM = "printOperand", int Default = 0>
35 : OperandWithDefaultOps <i32, (ops (i32 Default))> {
39 // src_sel for ALU src operands, see also ALU_CONST, ALU_PARAM registers
40 def SEL : OperandWithDefaultOps <i32, (ops (i32 -1))> {
41 let PrintMethod = "printSel";
43 def BANK_SWIZZLE : OperandWithDefaultOps <i32, (ops (i32 0))> {
44 let PrintMethod = "printBankSwizzle";
47 def LITERAL : InstFlag<"printLiteral">;
49 def WRITE : InstFlag <"printWrite", 1>;
50 def OMOD : InstFlag <"printOMOD">;
51 def REL : InstFlag <"printRel">;
52 def CLAMP : InstFlag <"printClamp">;
53 def NEG : InstFlag <"printNeg">;
54 def ABS : InstFlag <"printAbs">;
55 def UEM : InstFlag <"printUpdateExecMask">;
56 def UP : InstFlag <"printUpdatePred">;
58 // XXX: The r600g finalizer in Mesa expects last to be one in most cases.
59 // Once we start using the packetizer in this backend we should have this
61 def LAST : InstFlag<"printLast", 1>;
62 def RSel : Operand<i32> {
63 let PrintMethod = "printRSel";
65 def CT: Operand<i32> {
66 let PrintMethod = "printCT";
69 def FRAMEri : Operand<iPTR> {
70 let MIOperandInfo = (ops R600_Reg32:$ptr, i32imm:$index);
73 def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
74 def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
75 def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
76 def ADDRGA_CONST_OFFSET : ComplexPattern<i32, 1, "SelectGlobalValueConstantOffset", [], []>;
77 def ADDRGA_VAR_OFFSET : ComplexPattern<i32, 2, "SelectGlobalValueVariableOffset", [], []>;
78 def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>;
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))]
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, SDPatternOperator node,
164 InstrItinClass itim = AnyALU> :
165 R600_2OP <inst, opName,
166 [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
170 // If you add our change the operands for R600_3OP instructions, you must
171 // also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
172 // R600InstrInfo::buildDefaultInstruction(), and
173 // R600InstrInfo::getOperandIdx().
174 class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
175 InstrItinClass itin = AnyALU> :
176 InstR600 <(outs R600_Reg32:$dst),
177 (ins REL:$dst_rel, CLAMP:$clamp,
178 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, SEL:$src0_sel,
179 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
180 R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
181 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
182 BANK_SWIZZLE:$bank_swizzle),
183 !strconcat(" ", opName, "$clamp $last $dst$dst_rel, "
184 "$src0_neg$src0$src0_rel, "
185 "$src1_neg$src1$src1_rel, "
186 "$src2_neg$src2$src2_rel, "
192 R600ALU_Word1_OP3<inst>{
194 let HasNativeOperands = 1;
195 let DisableEncoding = "$literal";
197 let UseNamedOperandTable = 1;
200 let Inst{31-0} = Word0;
201 let Inst{63-32} = Word1;
204 class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
205 InstrItinClass itin = VecALU> :
206 InstR600 <(outs R600_Reg32:$dst),
214 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
216 def TEX_SHADOW : PatLeaf<
218 [{uint32_t TType = (uint32_t)N->getZExtValue();
219 return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
223 def TEX_RECT : PatLeaf<
225 [{uint32_t TType = (uint32_t)N->getZExtValue();
230 def TEX_ARRAY : PatLeaf<
232 [{uint32_t TType = (uint32_t)N->getZExtValue();
233 return TType == 9 || TType == 10 || TType == 15 || TType == 16;
237 def TEX_SHADOW_ARRAY : PatLeaf<
239 [{uint32_t TType = (uint32_t)N->getZExtValue();
240 return TType == 11 || TType == 12 || TType == 17;
244 class EG_CF_RAT <bits <8> cfinst, bits <6> ratinst, bits<4> mask, dag outs,
245 dag ins, string asm, list<dag> pattern> :
246 InstR600ISA <outs, ins, asm, pattern>,
247 CF_ALLOC_EXPORT_WORD0_RAT, CF_ALLOC_EXPORT_WORD1_BUF {
250 let rat_inst = ratinst;
252 // XXX: Have a separate instruction for non-indexed writes.
258 let comp_mask = mask;
261 let cf_inst = cfinst;
265 let Inst{31-0} = Word0;
266 let Inst{63-32} = Word1;
270 class VTX_READ <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
271 : InstR600ISA <outs, (ins MEMxi:$src_gpr), name, pattern>,
276 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
277 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
278 // however, based on my testing if USE_CONST_FIELDS is set, then all
279 // these fields need to be set to 0.
280 let USE_CONST_FIELDS = 0;
281 let NUM_FORMAT_ALL = 1;
282 let FORMAT_COMP_ALL = 0;
283 let SRF_MODE_ALL = 0;
285 let Inst{63-32} = Word1;
286 // LLVM can only encode 64-bit instructions, so these fields are manually
287 // encoded in R600CodeEmitter
290 // bits<2> ENDIAN_SWAP = 0;
291 // bits<1> CONST_BUF_NO_STRIDE = 0;
292 // bits<1> MEGA_FETCH = 0;
293 // bits<1> ALT_CONST = 0;
294 // bits<2> BUFFER_INDEX_MODE = 0;
296 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
297 // is done in R600CodeEmitter
299 // Inst{79-64} = OFFSET;
300 // Inst{81-80} = ENDIAN_SWAP;
301 // Inst{82} = CONST_BUF_NO_STRIDE;
302 // Inst{83} = MEGA_FETCH;
303 // Inst{84} = ALT_CONST;
304 // Inst{86-85} = BUFFER_INDEX_MODE;
305 // Inst{95-86} = 0; Reserved
307 // VTX_WORD3 (Padding)
314 class LoadParamFrag <PatFrag load_type> : PatFrag <
315 (ops node:$ptr), (load_type node:$ptr),
316 [{ return isConstantLoad(dyn_cast<LoadSDNode>(N), 0); }]
319 def load_param : LoadParamFrag<load>;
320 def load_param_exti8 : LoadParamFrag<az_extloadi8>;
321 def load_param_exti16 : LoadParamFrag<az_extloadi16>;
323 def isR600 : Predicate<"Subtarget.getGeneration() <= AMDGPUSubtarget::R700">;
324 def isR700 : Predicate<"Subtarget.getGeneration() == AMDGPUSubtarget::R700">;
325 def isEG : Predicate<
326 "Subtarget.getGeneration() >= AMDGPUSubtarget::EVERGREEN && "
327 "Subtarget.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS && "
328 "!Subtarget.hasCaymanISA()">;
330 def isCayman : Predicate<"Subtarget.hasCaymanISA()">;
331 def isEGorCayman : Predicate<"Subtarget.getGeneration() == "
332 "AMDGPUSubtarget::EVERGREEN"
333 "|| Subtarget.getGeneration() =="
334 "AMDGPUSubtarget::NORTHERN_ISLANDS">;
336 def isR600toCayman : Predicate<
337 "Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS">;
339 //===----------------------------------------------------------------------===//
341 //===----------------------------------------------------------------------===//
343 def INTERP_PAIR_XY : AMDGPUShaderInst <
344 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
345 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
346 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
349 def INTERP_PAIR_ZW : AMDGPUShaderInst <
350 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
351 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
352 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
355 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
356 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
360 def DOT4 : SDNode<"AMDGPUISD::DOT4",
361 SDTypeProfile<1, 8, [SDTCisFP<0>, SDTCisVT<1, f32>, SDTCisVT<2, f32>,
362 SDTCisVT<3, f32>, SDTCisVT<4, f32>, SDTCisVT<5, f32>,
363 SDTCisVT<6, f32>, SDTCisVT<7, f32>, SDTCisVT<8, f32>]>,
367 def COS_HW : SDNode<"AMDGPUISD::COS_HW",
368 SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisFP<1>]>
371 def SIN_HW : SDNode<"AMDGPUISD::SIN_HW",
372 SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisFP<1>]>
375 def TEXTURE_FETCH_Type : SDTypeProfile<1, 19, [SDTCisFP<0>]>;
377 def TEXTURE_FETCH: SDNode<"AMDGPUISD::TEXTURE_FETCH", TEXTURE_FETCH_Type, []>;
379 multiclass TexPattern<bits<32> TextureOp, Instruction inst, ValueType vt = v4f32> {
380 def : Pat<(TEXTURE_FETCH (i32 TextureOp), vt:$SRC_GPR,
381 (i32 imm:$srcx), (i32 imm:$srcy), (i32 imm:$srcz), (i32 imm:$srcw),
382 (i32 imm:$offsetx), (i32 imm:$offsety), (i32 imm:$offsetz),
383 (i32 imm:$DST_SEL_X), (i32 imm:$DST_SEL_Y), (i32 imm:$DST_SEL_Z),
384 (i32 imm:$DST_SEL_W),
385 (i32 imm:$RESOURCE_ID), (i32 imm:$SAMPLER_ID),
386 (i32 imm:$COORD_TYPE_X), (i32 imm:$COORD_TYPE_Y), (i32 imm:$COORD_TYPE_Z),
387 (i32 imm:$COORD_TYPE_W)),
388 (inst R600_Reg128:$SRC_GPR,
389 imm:$srcx, imm:$srcy, imm:$srcz, imm:$srcw,
390 imm:$offsetx, imm:$offsety, imm:$offsetz,
391 imm:$DST_SEL_X, imm:$DST_SEL_Y, imm:$DST_SEL_Z,
393 imm:$RESOURCE_ID, imm:$SAMPLER_ID,
394 imm:$COORD_TYPE_X, imm:$COORD_TYPE_Y, imm:$COORD_TYPE_Z,
398 //===----------------------------------------------------------------------===//
399 // Interpolation Instructions
400 //===----------------------------------------------------------------------===//
402 def INTERP_VEC_LOAD : AMDGPUShaderInst <
403 (outs R600_Reg128:$dst),
405 "INTERP_LOAD $src0 : $dst",
408 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
409 let bank_swizzle = 5;
412 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
413 let bank_swizzle = 5;
416 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
418 //===----------------------------------------------------------------------===//
419 // Export Instructions
420 //===----------------------------------------------------------------------===//
422 def ExportType : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
424 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
425 [SDNPHasChain, SDNPSideEffect]>;
428 field bits<32> Word0;
435 let Word0{12-0} = arraybase;
436 let Word0{14-13} = type;
437 let Word0{21-15} = gpr;
438 let Word0{22} = 0; // RW_REL
439 let Word0{29-23} = 0; // INDEX_GPR
440 let Word0{31-30} = elem_size;
443 class ExportSwzWord1 {
444 field bits<32> Word1;
453 let Word1{2-0} = sw_x;
454 let Word1{5-3} = sw_y;
455 let Word1{8-6} = sw_z;
456 let Word1{11-9} = sw_w;
459 class ExportBufWord1 {
460 field bits<32> Word1;
467 let Word1{11-0} = arraySize;
468 let Word1{15-12} = compMask;
471 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
472 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
474 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
475 0, 61, 0, 7, 7, 7, cf_inst, 0)
478 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
480 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
481 0, 61, 7, 0, 7, 7, cf_inst, 0)
484 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
486 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
489 def : Pat<(int_R600_store_dummy 1),
491 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
494 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
495 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
496 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
497 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
502 multiclass SteamOutputExportPattern<Instruction ExportInst,
503 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
505 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
506 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
507 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
508 4095, imm:$mask, buf0inst, 0)>;
510 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
511 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
512 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
513 4095, imm:$mask, buf1inst, 0)>;
515 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
516 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
517 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
518 4095, imm:$mask, buf2inst, 0)>;
520 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
521 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
522 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
523 4095, imm:$mask, buf3inst, 0)>;
526 // Export Instructions should not be duplicated by TailDuplication pass
527 // (which assumes that duplicable instruction are affected by exec mask)
528 let usesCustomInserter = 1, isNotDuplicable = 1 in {
530 class ExportSwzInst : InstR600ISA<(
532 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
533 RSel:$sw_x, RSel:$sw_y, RSel:$sw_z, RSel:$sw_w, i32imm:$inst,
535 !strconcat("EXPORT", " $gpr.$sw_x$sw_y$sw_z$sw_w"),
536 []>, ExportWord0, ExportSwzWord1 {
538 let Inst{31-0} = Word0;
539 let Inst{63-32} = Word1;
542 } // End usesCustomInserter = 1
544 class ExportBufInst : InstR600ISA<(
546 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
547 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
548 !strconcat("EXPORT", " $gpr"),
549 []>, ExportWord0, ExportBufWord1 {
551 let Inst{31-0} = Word0;
552 let Inst{63-32} = Word1;
555 //===----------------------------------------------------------------------===//
556 // Control Flow Instructions
557 //===----------------------------------------------------------------------===//
560 def KCACHE : InstFlag<"printKCache">;
562 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
563 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1,
564 KCACHE:$KCACHE_MODE0, KCACHE:$KCACHE_MODE1,
565 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1,
566 i32imm:$COUNT, i32imm:$Enabled),
567 !strconcat(OpName, " $COUNT, @$ADDR, "
568 "KC0[$KCACHE_MODE0], KC1[$KCACHE_MODE1]"),
569 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
574 let WHOLE_QUAD_MODE = 0;
577 let Inst{31-0} = Word0;
578 let Inst{63-32} = Word1;
581 class CF_WORD0_R600 {
582 field bits<32> Word0;
589 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
590 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
597 let VALID_PIXEL_MODE = 0;
599 let COUNT = CNT{2-0};
601 let COUNT_3 = CNT{3};
602 let END_OF_PROGRAM = 0;
603 let WHOLE_QUAD_MODE = 0;
605 let Inst{31-0} = Word0;
606 let Inst{63-32} = Word1;
609 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
610 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
615 let JUMPTABLE_SEL = 0;
617 let VALID_PIXEL_MODE = 0;
619 let END_OF_PROGRAM = 0;
621 let Inst{31-0} = Word0;
622 let Inst{63-32} = Word1;
625 def CF_ALU : ALU_CLAUSE<8, "ALU">;
626 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
627 def CF_ALU_POP_AFTER : ALU_CLAUSE<10, "ALU_POP_AFTER">;
629 def FETCH_CLAUSE : AMDGPUInst <(outs),
630 (ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
636 def ALU_CLAUSE : AMDGPUInst <(outs),
637 (ins i32imm:$addr), "ALU clause starting at $addr:", [] > {
643 def LITERALS : AMDGPUInst <(outs),
644 (ins LITERAL:$literal1, LITERAL:$literal2), "$literal1, $literal2", [] > {
649 let Inst{31-0} = literal1;
650 let Inst{63-32} = literal2;
653 def PAD : AMDGPUInst <(outs), (ins), "PAD", [] > {
657 let Predicates = [isR600toCayman] in {
659 //===----------------------------------------------------------------------===//
660 // Common Instructions R600, R700, Evergreen, Cayman
661 //===----------------------------------------------------------------------===//
663 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
664 // Non-IEEE MUL: 0 * anything = 0
665 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
666 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
667 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
668 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
670 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
671 // so some of the instruction names don't match the asm string.
672 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
673 def SETE : R600_2OP <
675 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_EQ))]
680 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GT))]
685 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GE))]
690 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_NE))]
693 def SETE_DX10 : R600_2OP <
695 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_EQ))]
698 def SETGT_DX10 : R600_2OP <
700 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GT))]
703 def SETGE_DX10 : R600_2OP <
705 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GE))]
708 def SETNE_DX10 : R600_2OP <
710 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_NE))]
713 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
714 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
715 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
716 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
717 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
719 def MOV : R600_1OP <0x19, "MOV", []>;
721 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
723 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
724 (outs R600_Reg32:$dst),
730 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
732 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
735 (MOV_IMM_I32 imm:$val)
738 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
741 (MOV_IMM_F32 fpimm:$val)
744 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
745 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
746 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
747 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
749 let hasSideEffects = 1 in {
751 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
753 } // end hasSideEffects
755 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
756 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
757 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
758 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
759 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
760 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
761 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
762 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
763 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
764 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
766 def SETE_INT : R600_2OP <
768 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETEQ))]
771 def SETGT_INT : R600_2OP <
773 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGT))]
776 def SETGE_INT : R600_2OP <
778 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGE))]
781 def SETNE_INT : R600_2OP <
783 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETNE))]
786 def SETGT_UINT : R600_2OP <
788 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGT))]
791 def SETGE_UINT : R600_2OP <
793 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGE))]
796 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
797 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
798 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
799 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
801 def CNDE_INT : R600_3OP <
803 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_EQ))]
806 def CNDGE_INT : R600_3OP <
808 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GE))]
811 def CNDGT_INT : R600_3OP <
813 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GT))]
816 //===----------------------------------------------------------------------===//
817 // Texture instructions
818 //===----------------------------------------------------------------------===//
820 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
822 class R600_TEX <bits<11> inst, string opName> :
823 InstR600 <(outs R600_Reg128:$DST_GPR),
824 (ins R600_Reg128:$SRC_GPR,
825 RSel:$srcx, RSel:$srcy, RSel:$srcz, RSel:$srcw,
826 i32imm:$offsetx, i32imm:$offsety, i32imm:$offsetz,
827 RSel:$DST_SEL_X, RSel:$DST_SEL_Y, RSel:$DST_SEL_Z, RSel:$DST_SEL_W,
828 i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
829 CT:$COORD_TYPE_X, CT:$COORD_TYPE_Y, CT:$COORD_TYPE_Z,
832 " $DST_GPR.$DST_SEL_X$DST_SEL_Y$DST_SEL_Z$DST_SEL_W, "
833 "$SRC_GPR.$srcx$srcy$srcz$srcw "
834 "RID:$RESOURCE_ID SID:$SAMPLER_ID "
835 "CT:$COORD_TYPE_X$COORD_TYPE_Y$COORD_TYPE_Z$COORD_TYPE_W"),
837 NullALU>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
838 let Inst{31-0} = Word0;
839 let Inst{63-32} = Word1;
841 let TEX_INST = inst{4-0};
847 let FETCH_WHOLE_QUAD = 0;
849 let SAMPLER_INDEX_MODE = 0;
850 let RESOURCE_INDEX_MODE = 0;
855 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
859 def TEX_SAMPLE : R600_TEX <0x10, "TEX_SAMPLE">;
860 def TEX_SAMPLE_C : R600_TEX <0x18, "TEX_SAMPLE_C">;
861 def TEX_SAMPLE_L : R600_TEX <0x11, "TEX_SAMPLE_L">;
862 def TEX_SAMPLE_C_L : R600_TEX <0x19, "TEX_SAMPLE_C_L">;
863 def TEX_SAMPLE_LB : R600_TEX <0x12, "TEX_SAMPLE_LB">;
864 def TEX_SAMPLE_C_LB : R600_TEX <0x1A, "TEX_SAMPLE_C_LB">;
865 def TEX_LD : R600_TEX <0x03, "TEX_LD">;
866 def TEX_GET_TEXTURE_RESINFO : R600_TEX <0x04, "TEX_GET_TEXTURE_RESINFO">;
867 def TEX_GET_GRADIENTS_H : R600_TEX <0x07, "TEX_GET_GRADIENTS_H">;
868 def TEX_GET_GRADIENTS_V : R600_TEX <0x08, "TEX_GET_GRADIENTS_V">;
869 def TEX_SET_GRADIENTS_H : R600_TEX <0x0B, "TEX_SET_GRADIENTS_H">;
870 def TEX_SET_GRADIENTS_V : R600_TEX <0x0C, "TEX_SET_GRADIENTS_V">;
871 def TEX_SAMPLE_G : R600_TEX <0x14, "TEX_SAMPLE_G">;
872 def TEX_SAMPLE_C_G : R600_TEX <0x1C, "TEX_SAMPLE_C_G">;
874 defm : TexPattern<0, TEX_SAMPLE>;
875 defm : TexPattern<1, TEX_SAMPLE_C>;
876 defm : TexPattern<2, TEX_SAMPLE_L>;
877 defm : TexPattern<3, TEX_SAMPLE_C_L>;
878 defm : TexPattern<4, TEX_SAMPLE_LB>;
879 defm : TexPattern<5, TEX_SAMPLE_C_LB>;
880 defm : TexPattern<6, TEX_LD, v4i32>;
881 defm : TexPattern<7, TEX_GET_TEXTURE_RESINFO, v4i32>;
882 defm : TexPattern<8, TEX_GET_GRADIENTS_H>;
883 defm : TexPattern<9, TEX_GET_GRADIENTS_V>;
885 //===----------------------------------------------------------------------===//
886 // Helper classes for common instructions
887 //===----------------------------------------------------------------------===//
889 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
894 class MULADD_Common <bits<5> inst> : R600_3OP <
899 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
901 [(set f32:$dst, (fadd (fmul f32:$src0, f32:$src1), f32:$src2))]
904 class CNDE_Common <bits<5> inst> : R600_3OP <
906 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_EQ))]
909 class CNDGT_Common <bits<5> inst> : R600_3OP <
911 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GT))]
914 class CNDGE_Common <bits<5> inst> : R600_3OP <
916 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GE))]
920 let isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
921 class R600_VEC2OP<list<dag> pattern> : InstR600 <(outs R600_Reg32:$dst), (ins
923 UEM:$update_exec_mask_X, UP:$update_pred_X, WRITE:$write_X,
924 OMOD:$omod_X, REL:$dst_rel_X, CLAMP:$clamp_X,
925 R600_TReg32_X:$src0_X, NEG:$src0_neg_X, REL:$src0_rel_X, ABS:$src0_abs_X, SEL:$src0_sel_X,
926 R600_TReg32_X:$src1_X, NEG:$src1_neg_X, REL:$src1_rel_X, ABS:$src1_abs_X, SEL:$src1_sel_X,
927 R600_Pred:$pred_sel_X,
929 UEM:$update_exec_mask_Y, UP:$update_pred_Y, WRITE:$write_Y,
930 OMOD:$omod_Y, REL:$dst_rel_Y, CLAMP:$clamp_Y,
931 R600_TReg32_Y:$src0_Y, NEG:$src0_neg_Y, REL:$src0_rel_Y, ABS:$src0_abs_Y, SEL:$src0_sel_Y,
932 R600_TReg32_Y:$src1_Y, NEG:$src1_neg_Y, REL:$src1_rel_Y, ABS:$src1_abs_Y, SEL:$src1_sel_Y,
933 R600_Pred:$pred_sel_Y,
935 UEM:$update_exec_mask_Z, UP:$update_pred_Z, WRITE:$write_Z,
936 OMOD:$omod_Z, REL:$dst_rel_Z, CLAMP:$clamp_Z,
937 R600_TReg32_Z:$src0_Z, NEG:$src0_neg_Z, REL:$src0_rel_Z, ABS:$src0_abs_Z, SEL:$src0_sel_Z,
938 R600_TReg32_Z:$src1_Z, NEG:$src1_neg_Z, REL:$src1_rel_Z, ABS:$src1_abs_Z, SEL:$src1_sel_Z,
939 R600_Pred:$pred_sel_Z,
941 UEM:$update_exec_mask_W, UP:$update_pred_W, WRITE:$write_W,
942 OMOD:$omod_W, REL:$dst_rel_W, CLAMP:$clamp_W,
943 R600_TReg32_W:$src0_W, NEG:$src0_neg_W, REL:$src0_rel_W, ABS:$src0_abs_W, SEL:$src0_sel_W,
944 R600_TReg32_W:$src1_W, NEG:$src1_neg_W, REL:$src1_rel_W, ABS:$src1_abs_W, SEL:$src1_sel_W,
945 R600_Pred:$pred_sel_W,
946 LITERAL:$literal0, LITERAL:$literal1),
951 let UseNamedOperandTable = 1;
956 def DOT_4 : R600_VEC2OP<[(set R600_Reg32:$dst, (DOT4
957 R600_TReg32_X:$src0_X, R600_TReg32_X:$src1_X,
958 R600_TReg32_Y:$src0_Y, R600_TReg32_Y:$src1_Y,
959 R600_TReg32_Z:$src0_Z, R600_TReg32_Z:$src1_Z,
960 R600_TReg32_W:$src0_W, R600_TReg32_W:$src1_W))]>;
963 class DOT4_Common <bits<11> inst> : R600_2OP <inst, "DOT4", []>;
966 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
967 multiclass CUBE_Common <bits<11> inst> {
969 def _pseudo : InstR600 <
970 (outs R600_Reg128:$dst),
971 (ins R600_Reg128:$src0),
973 [(set v4f32:$dst, (int_AMDGPU_cube v4f32:$src0))],
977 let UseNamedOperandTable = 1;
980 def _real : R600_2OP <inst, "CUBE", []>;
982 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
984 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
985 inst, "EXP_IEEE", fexp2
988 let Itinerary = TransALU;
991 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
992 inst, "FLT_TO_INT", fp_to_sint
995 let Itinerary = TransALU;
998 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
999 inst, "INT_TO_FLT", sint_to_fp
1002 let Itinerary = TransALU;
1005 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1006 inst, "FLT_TO_UINT", fp_to_uint
1009 let Itinerary = TransALU;
1012 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1013 inst, "UINT_TO_FLT", uint_to_fp
1016 let Itinerary = TransALU;
1019 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1020 inst, "LOG_CLAMPED", []
1023 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1024 inst, "LOG_IEEE", flog2
1027 let Itinerary = TransALU;
1030 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1031 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1032 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1033 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1034 inst, "MULHI_INT", mulhs
1037 let Itinerary = TransALU;
1039 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1040 inst, "MULHI", mulhu
1043 let Itinerary = TransALU;
1045 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1046 inst, "MULLO_INT", mul
1049 let Itinerary = TransALU;
1051 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []> {
1053 let Itinerary = TransALU;
1056 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1057 inst, "RECIP_CLAMPED", []
1060 let Itinerary = TransALU;
1063 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1064 inst, "RECIP_IEEE", [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
1067 let Itinerary = TransALU;
1070 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1071 inst, "RECIP_UINT", AMDGPUurecip
1074 let Itinerary = TransALU;
1077 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1078 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1081 let Itinerary = TransALU;
1084 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1085 inst, "RECIPSQRT_IEEE", []
1088 let Itinerary = TransALU;
1091 class SIN_Common <bits<11> inst> : R600_1OP <
1092 inst, "SIN", [(set f32:$dst, (SIN_HW f32:$src0))]>{
1095 let Itinerary = TransALU;
1098 class COS_Common <bits<11> inst> : R600_1OP <
1099 inst, "COS", [(set f32:$dst, (COS_HW f32:$src0))]> {
1102 let Itinerary = TransALU;
1105 //===----------------------------------------------------------------------===//
1106 // Helper patterns for complex intrinsics
1107 //===----------------------------------------------------------------------===//
1109 multiclass DIV_Common <InstR600 recip_ieee> {
1111 (int_AMDGPU_div f32:$src0, f32:$src1),
1112 (MUL_IEEE $src0, (recip_ieee $src1))
1116 (fdiv f32:$src0, f32:$src1),
1117 (MUL_IEEE $src0, (recip_ieee $src1))
1121 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee>
1123 (int_TGSI_lit_z f32:$src_x, f32:$src_y, f32:$src_w),
1124 (exp_ieee (mul_lit (log_clamped (MAX $src_y, (f32 ZERO))), $src_w, $src_x))
1127 //===----------------------------------------------------------------------===//
1128 // R600 / R700 Instructions
1129 //===----------------------------------------------------------------------===//
1131 let Predicates = [isR600] in {
1133 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1134 def MULADD_r600 : MULADD_Common<0x10>;
1135 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1136 def CNDE_r600 : CNDE_Common<0x18>;
1137 def CNDGT_r600 : CNDGT_Common<0x19>;
1138 def CNDGE_r600 : CNDGE_Common<0x1A>;
1139 def DOT4_r600 : DOT4_Common<0x50>;
1140 defm CUBE_r600 : CUBE_Common<0x52>;
1141 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1142 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1143 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1144 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1145 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1146 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1147 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1148 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1149 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1150 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1151 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1152 def SIN_r600 : SIN_Common<0x6E>;
1153 def COS_r600 : COS_Common<0x6F>;
1154 def ASHR_r600 : ASHR_Common<0x70>;
1155 def LSHR_r600 : LSHR_Common<0x71>;
1156 def LSHL_r600 : LSHL_Common<0x72>;
1157 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1158 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1159 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1160 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1161 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1163 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1164 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL>;
1165 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1167 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
1169 def R600_ExportSwz : ExportSwzInst {
1170 let Word1{20-17} = 0; // BURST_COUNT
1171 let Word1{21} = eop;
1172 let Word1{22} = 1; // VALID_PIXEL_MODE
1173 let Word1{30-23} = inst;
1174 let Word1{31} = 1; // BARRIER
1176 defm : ExportPattern<R600_ExportSwz, 39>;
1178 def R600_ExportBuf : ExportBufInst {
1179 let Word1{20-17} = 0; // BURST_COUNT
1180 let Word1{21} = eop;
1181 let Word1{22} = 1; // VALID_PIXEL_MODE
1182 let Word1{30-23} = inst;
1183 let Word1{31} = 1; // BARRIER
1185 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1187 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$CNT),
1188 "TEX $CNT @$ADDR"> {
1191 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$CNT),
1192 "VTX $CNT @$ADDR"> {
1195 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1196 "LOOP_START_DX10 @$ADDR"> {
1200 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1204 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1205 "LOOP_BREAK @$ADDR"> {
1209 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1210 "CONTINUE @$ADDR"> {
1214 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1215 "JUMP @$ADDR POP:$POP_COUNT"> {
1218 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1219 "ELSE @$ADDR POP:$POP_COUNT"> {
1222 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1227 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1228 "POP @$ADDR POP:$POP_COUNT"> {
1231 def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
1235 let END_OF_PROGRAM = 1;
1240 //===----------------------------------------------------------------------===//
1241 // R700 Only instructions
1242 //===----------------------------------------------------------------------===//
1244 let Predicates = [isR700] in {
1245 def SIN_r700 : SIN_Common<0x6E>;
1246 def COS_r700 : COS_Common<0x6F>;
1249 //===----------------------------------------------------------------------===//
1250 // Evergreen Only instructions
1251 //===----------------------------------------------------------------------===//
1253 let Predicates = [isEG] in {
1255 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1256 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1258 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1259 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1260 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1261 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1262 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1263 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1264 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1265 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1266 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1267 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1268 def SIN_eg : SIN_Common<0x8D>;
1269 def COS_eg : COS_Common<0x8E>;
1271 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
1272 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
1274 //===----------------------------------------------------------------------===//
1275 // Memory read/write instructions
1276 //===----------------------------------------------------------------------===//
1277 let usesCustomInserter = 1 in {
1279 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> mask, string name,
1281 : EG_CF_RAT <0x57, 0x2, mask, (outs), ins, name, pattern> {
1284 } // End usesCustomInserter = 1
1287 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1288 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1289 0x1, "RAT_WRITE_CACHELESS_32_eg $rw_gpr, $index_gpr, $eop",
1290 [(global_store i32:$rw_gpr, i32:$index_gpr)]
1294 def RAT_WRITE_CACHELESS_64_eg : RAT_WRITE_CACHELESS_eg <
1295 (ins R600_Reg64:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1296 0x3, "RAT_WRITE_CACHELESS_64_eg $rw_gpr.XY, $index_gpr, $eop",
1297 [(global_store v2i32:$rw_gpr, i32:$index_gpr)]
1301 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1302 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1303 0xf, "RAT_WRITE_CACHELESS_128 $rw_gpr.XYZW, $index_gpr, $eop",
1304 [(global_store v4i32:$rw_gpr, i32:$index_gpr)]
1307 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1308 : VTX_WORD0_eg, VTX_READ<name, buffer_id, outs, pattern> {
1313 let FETCH_WHOLE_QUAD = 0;
1314 let BUFFER_ID = buffer_id;
1316 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1317 // to store vertex addresses in any channel, not just X.
1320 let Inst{31-0} = Word0;
1323 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1324 : VTX_READ_eg <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id,
1325 (outs R600_TReg32_X:$dst_gpr), pattern> {
1327 let MEGA_FETCH_COUNT = 1;
1329 let DST_SEL_Y = 7; // Masked
1330 let DST_SEL_Z = 7; // Masked
1331 let DST_SEL_W = 7; // Masked
1332 let DATA_FORMAT = 1; // FMT_8
1335 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1336 : VTX_READ_eg <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id,
1337 (outs R600_TReg32_X:$dst_gpr), pattern> {
1338 let MEGA_FETCH_COUNT = 2;
1340 let DST_SEL_Y = 7; // Masked
1341 let DST_SEL_Z = 7; // Masked
1342 let DST_SEL_W = 7; // Masked
1343 let DATA_FORMAT = 5; // FMT_16
1347 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1348 : VTX_READ_eg <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id,
1349 (outs R600_TReg32_X:$dst_gpr), pattern> {
1351 let MEGA_FETCH_COUNT = 4;
1353 let DST_SEL_Y = 7; // Masked
1354 let DST_SEL_Z = 7; // Masked
1355 let DST_SEL_W = 7; // Masked
1356 let DATA_FORMAT = 0xD; // COLOR_32
1358 // This is not really necessary, but there were some GPU hangs that appeared
1359 // to be caused by ALU instructions in the next instruction group that wrote
1360 // to the $src_gpr registers of the VTX_READ.
1362 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1363 // %T2_X<def> = MOV %ZERO
1364 //Adding this constraint prevents this from happening.
1365 let Constraints = "$src_gpr.ptr = $dst_gpr";
1368 class VTX_READ_64_eg <bits<8> buffer_id, list<dag> pattern>
1369 : VTX_READ_eg <"VTX_READ_64 $dst_gpr.XY, $src_gpr", buffer_id,
1370 (outs R600_Reg64:$dst_gpr), pattern> {
1372 let MEGA_FETCH_COUNT = 8;
1377 let DATA_FORMAT = 0x1D; // COLOR_32_32
1380 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1381 : VTX_READ_eg <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id,
1382 (outs R600_Reg128:$dst_gpr), pattern> {
1384 let MEGA_FETCH_COUNT = 16;
1389 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1391 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1392 // that holds its buffer address to avoid potential hangs. We can't use
1393 // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst
1394 // registers are different sizes.
1397 //===----------------------------------------------------------------------===//
1398 // VTX Read from parameter memory space
1399 //===----------------------------------------------------------------------===//
1401 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1402 [(set i32:$dst_gpr, (load_param_exti8 ADDRVTX_READ:$src_gpr))]
1405 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1406 [(set i32:$dst_gpr, (load_param_exti16 ADDRVTX_READ:$src_gpr))]
1409 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1410 [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1413 def VTX_READ_PARAM_64_eg : VTX_READ_64_eg <0,
1414 [(set v2i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1417 def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
1418 [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1421 //===----------------------------------------------------------------------===//
1422 // VTX Read from global memory space
1423 //===----------------------------------------------------------------------===//
1426 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1427 [(set i32:$dst_gpr, (az_extloadi8_global ADDRVTX_READ:$src_gpr))]
1430 def VTX_READ_GLOBAL_16_eg : VTX_READ_16_eg <1,
1431 [(set i32:$dst_gpr, (az_extloadi16_global ADDRVTX_READ:$src_gpr))]
1435 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1436 [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1440 def VTX_READ_GLOBAL_64_eg : VTX_READ_64_eg <1,
1441 [(set v2i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1445 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1446 [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1449 } // End Predicates = [isEG]
1451 //===----------------------------------------------------------------------===//
1452 // Evergreen / Cayman Instructions
1453 //===----------------------------------------------------------------------===//
1455 let Predicates = [isEGorCayman] in {
1457 // BFE_UINT - bit_extract, an optimization for mask and shift
1462 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1467 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1468 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1469 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1470 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1471 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1472 [(set i32:$dst, (int_AMDIL_bit_extract_u32 i32:$src0, i32:$src1,
1476 def : BFEPattern <BFE_UINT_eg>;
1478 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>;
1479 defm : BFIPatterns <BFI_INT_eg>;
1481 def MULADD_UINT24_eg : R600_3OP <0x10, "MULADD_UINT24",
1482 [(set i32:$dst, (add (mul U24:$src0, U24:$src1), i32:$src2))], VecALU
1484 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
1485 def : ROTRPattern <BIT_ALIGN_INT_eg>;
1487 def MULADD_eg : MULADD_Common<0x14>;
1488 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
1489 def ASHR_eg : ASHR_Common<0x15>;
1490 def LSHR_eg : LSHR_Common<0x16>;
1491 def LSHL_eg : LSHL_Common<0x17>;
1492 def CNDE_eg : CNDE_Common<0x19>;
1493 def CNDGT_eg : CNDGT_Common<0x1A>;
1494 def CNDGE_eg : CNDGE_Common<0x1B>;
1495 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1496 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1497 def MUL_UINT24_eg : R600_2OP <0xB5, "MUL_UINT24",
1498 [(set i32:$dst, (mul U24:$src0, U24:$src1))], VecALU
1500 def DOT4_eg : DOT4_Common<0xBE>;
1501 defm CUBE_eg : CUBE_Common<0xC0>;
1503 let hasSideEffects = 1 in {
1504 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1507 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1509 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1512 let Itinerary = AnyALU;
1515 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1517 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1521 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1523 def GROUP_BARRIER : InstR600 <
1524 (outs), (ins), " GROUP_BARRIER", [(int_AMDGPU_barrier_local)], AnyALU>,
1526 R600ALU_Word1_OP2 <0x54> {
1542 let bank_swizzle = 0;
1544 let update_exec_mask = 0;
1545 let update_pred = 0;
1547 let Inst{31-0} = Word0;
1548 let Inst{63-32} = Word1;
1553 //===----------------------------------------------------------------------===//
1555 //===----------------------------------------------------------------------===//
1556 class R600_LDS <bits<6> op, dag outs, dag ins, string asm,
1557 list<dag> pattern = []> :
1559 InstR600 <outs, ins, asm, pattern, XALU>,
1566 let Word1{27} = offset{0};
1567 let Word1{12} = offset{1};
1568 let Word1{28} = offset{2};
1569 let Word1{31} = offset{3};
1570 let Word0{12} = offset{4};
1571 let Word0{25} = offset{5};
1574 let Inst{31-0} = Word0;
1575 let Inst{63-32} = Word1;
1578 let HasNativeOperands = 1;
1579 let UseNamedOperandTable = 1;
1582 class R600_LDS_1A <bits<6> lds_op, string name, list<dag> pattern> : R600_LDS <
1584 (outs R600_Reg32:$dst),
1585 (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
1586 LAST:$last, R600_Pred:$pred_sel,
1587 BANK_SWIZZLE:$bank_swizzle),
1588 " "#name#" $last OQAP, $src0$src0_rel $pred_sel",
1598 let usesCustomInserter = 1;
1600 let DisableEncoding = "$dst";
1603 class R600_LDS_1A1D <bits<6> lds_op, string name, list<dag> pattern> :
1607 (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
1608 R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel,
1609 LAST:$last, R600_Pred:$pred_sel,
1610 BANK_SWIZZLE:$bank_swizzle),
1611 " "#name#" $last $src0$src0_rel, $src1$src1_rel, $pred_sel",
1620 def LDS_READ_RET : R600_LDS_1A <0x32, "LDS_READ_RET",
1621 [(set (i32 R600_Reg32:$dst), (local_load R600_Reg32:$src0))]
1624 def LDS_WRITE : R600_LDS_1A1D <0xD, "LDS_WRITE",
1625 [(local_store (i32 R600_Reg32:$src1), R600_Reg32:$src0)]
1628 // TRUNC is used for the FLT_TO_INT instructions to work around a
1629 // perceived problem where the rounding modes are applied differently
1630 // depending on the instruction and the slot they are in.
1632 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1633 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1635 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1636 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1637 // We should look into handling these cases separately.
1638 def : Pat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
1640 def : Pat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
1643 def : SHA256MaPattern <BFI_INT_eg, XOR_INT>;
1645 def EG_ExportSwz : ExportSwzInst {
1646 let Word1{19-16} = 0; // BURST_COUNT
1647 let Word1{20} = 1; // VALID_PIXEL_MODE
1648 let Word1{21} = eop;
1649 let Word1{29-22} = inst;
1650 let Word1{30} = 0; // MARK
1651 let Word1{31} = 1; // BARRIER
1653 defm : ExportPattern<EG_ExportSwz, 83>;
1655 def EG_ExportBuf : ExportBufInst {
1656 let Word1{19-16} = 0; // BURST_COUNT
1657 let Word1{20} = 1; // VALID_PIXEL_MODE
1658 let Word1{21} = eop;
1659 let Word1{29-22} = inst;
1660 let Word1{30} = 0; // MARK
1661 let Word1{31} = 1; // BARRIER
1663 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1665 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1666 "TEX $COUNT @$ADDR"> {
1669 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1670 "VTX $COUNT @$ADDR"> {
1673 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
1674 "LOOP_START_DX10 @$ADDR"> {
1678 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1682 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
1683 "LOOP_BREAK @$ADDR"> {
1687 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
1688 "CONTINUE @$ADDR"> {
1692 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1693 "JUMP @$ADDR POP:$POP_COUNT"> {
1696 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1697 "ELSE @$ADDR POP:$POP_COUNT"> {
1700 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
1705 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1706 "POP @$ADDR POP:$POP_COUNT"> {
1709 def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
1713 let END_OF_PROGRAM = 1;
1716 } // End Predicates = [isEGorCayman]
1718 //===----------------------------------------------------------------------===//
1719 // Regist loads and stores - for indirect addressing
1720 //===----------------------------------------------------------------------===//
1722 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1724 //===----------------------------------------------------------------------===//
1725 // Cayman Instructions
1726 //===----------------------------------------------------------------------===//
1728 let Predicates = [isCayman] in {
1730 def MULADD_INT24_cm : R600_3OP <0x08, "MULADD_INT24",
1731 [(set i32:$dst, (add (mul I24:$src0, I24:$src1), i32:$src2))], VecALU
1733 def MUL_INT24_cm : R600_2OP <0x5B, "MUL_INT24",
1734 [(set i32:$dst, (mul I24:$src0, I24:$src1))], VecALU
1737 let isVector = 1 in {
1739 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1741 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1742 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1743 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1744 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1745 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1746 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1747 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
1748 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1749 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1750 def SIN_cm : SIN_Common<0x8D>;
1751 def COS_cm : COS_Common<0x8E>;
1752 } // End isVector = 1
1754 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
1756 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1758 // RECIP_UINT emulation for Cayman
1759 // The multiplication scales from [0,1] to the unsigned integer range
1761 (AMDGPUurecip i32:$src0),
1762 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
1763 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
1766 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
1772 def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
1775 class RAT_STORE_DWORD_cm <bits<4> mask, dag ins, list<dag> pat> : EG_CF_RAT <
1776 0x57, 0x14, mask, (outs), ins,
1777 "EXPORT_RAT_INST_STORE_DWORD $rw_gpr, $index_gpr", pat
1779 let eop = 0; // This bit is not used on Cayman.
1782 def RAT_STORE_DWORD32_cm : RAT_STORE_DWORD_cm <0x1,
1783 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr),
1784 [(global_store i32:$rw_gpr, i32:$index_gpr)]
1787 def RAT_STORE_DWORD64_cm : RAT_STORE_DWORD_cm <0x3,
1788 (ins R600_Reg64:$rw_gpr, R600_TReg32_X:$index_gpr),
1789 [(global_store v2i32:$rw_gpr, i32:$index_gpr)]
1792 class VTX_READ_cm <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1793 : VTX_WORD0_cm, VTX_READ<name, buffer_id, outs, pattern> {
1798 let FETCH_WHOLE_QUAD = 0;
1799 let BUFFER_ID = buffer_id;
1801 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1802 // to store vertex addresses in any channel, not just X.
1805 let STRUCTURED_READ = 0;
1807 let COALESCED_READ = 0;
1809 let Inst{31-0} = Word0;
1812 class VTX_READ_8_cm <bits<8> buffer_id, list<dag> pattern>
1813 : VTX_READ_cm <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id,
1814 (outs R600_TReg32_X:$dst_gpr), pattern> {
1817 let DST_SEL_Y = 7; // Masked
1818 let DST_SEL_Z = 7; // Masked
1819 let DST_SEL_W = 7; // Masked
1820 let DATA_FORMAT = 1; // FMT_8
1823 class VTX_READ_16_cm <bits<8> buffer_id, list<dag> pattern>
1824 : VTX_READ_cm <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id,
1825 (outs R600_TReg32_X:$dst_gpr), pattern> {
1827 let DST_SEL_Y = 7; // Masked
1828 let DST_SEL_Z = 7; // Masked
1829 let DST_SEL_W = 7; // Masked
1830 let DATA_FORMAT = 5; // FMT_16
1834 class VTX_READ_32_cm <bits<8> buffer_id, list<dag> pattern>
1835 : VTX_READ_cm <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id,
1836 (outs R600_TReg32_X:$dst_gpr), pattern> {
1839 let DST_SEL_Y = 7; // Masked
1840 let DST_SEL_Z = 7; // Masked
1841 let DST_SEL_W = 7; // Masked
1842 let DATA_FORMAT = 0xD; // COLOR_32
1844 // This is not really necessary, but there were some GPU hangs that appeared
1845 // to be caused by ALU instructions in the next instruction group that wrote
1846 // to the $src_gpr registers of the VTX_READ.
1848 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1849 // %T2_X<def> = MOV %ZERO
1850 //Adding this constraint prevents this from happening.
1851 let Constraints = "$src_gpr.ptr = $dst_gpr";
1854 class VTX_READ_64_cm <bits<8> buffer_id, list<dag> pattern>
1855 : VTX_READ_cm <"VTX_READ_64 $dst_gpr, $src_gpr", buffer_id,
1856 (outs R600_Reg64:$dst_gpr), pattern> {
1862 let DATA_FORMAT = 0x1D; // COLOR_32_32
1865 class VTX_READ_128_cm <bits<8> buffer_id, list<dag> pattern>
1866 : VTX_READ_cm <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id,
1867 (outs R600_Reg128:$dst_gpr), pattern> {
1873 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1875 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1876 // that holds its buffer address to avoid potential hangs. We can't use
1877 // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst
1878 // registers are different sizes.
1881 //===----------------------------------------------------------------------===//
1882 // VTX Read from parameter memory space
1883 //===----------------------------------------------------------------------===//
1884 def VTX_READ_PARAM_8_cm : VTX_READ_8_cm <0,
1885 [(set i32:$dst_gpr, (load_param_exti8 ADDRVTX_READ:$src_gpr))]
1888 def VTX_READ_PARAM_16_cm : VTX_READ_16_cm <0,
1889 [(set i32:$dst_gpr, (load_param_exti16 ADDRVTX_READ:$src_gpr))]
1892 def VTX_READ_PARAM_32_cm : VTX_READ_32_cm <0,
1893 [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1896 def VTX_READ_PARAM_64_cm : VTX_READ_64_cm <0,
1897 [(set v2i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1900 def VTX_READ_PARAM_128_cm : VTX_READ_128_cm <0,
1901 [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1904 //===----------------------------------------------------------------------===//
1905 // VTX Read from global memory space
1906 //===----------------------------------------------------------------------===//
1909 def VTX_READ_GLOBAL_8_cm : VTX_READ_8_cm <1,
1910 [(set i32:$dst_gpr, (az_extloadi8_global ADDRVTX_READ:$src_gpr))]
1913 def VTX_READ_GLOBAL_16_cm : VTX_READ_16_cm <1,
1914 [(set i32:$dst_gpr, (az_extloadi16_global ADDRVTX_READ:$src_gpr))]
1918 def VTX_READ_GLOBAL_32_cm : VTX_READ_32_cm <1,
1919 [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1923 def VTX_READ_GLOBAL_64_cm : VTX_READ_64_cm <1,
1924 [(set v2i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1928 def VTX_READ_GLOBAL_128_cm : VTX_READ_128_cm <1,
1929 [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1934 //===----------------------------------------------------------------------===//
1935 // Branch Instructions
1936 //===----------------------------------------------------------------------===//
1939 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
1940 "IF_PREDICATE_SET $src", []>;
1942 //===----------------------------------------------------------------------===//
1943 // Pseudo instructions
1944 //===----------------------------------------------------------------------===//
1946 let isPseudo = 1 in {
1948 def PRED_X : InstR600 <
1949 (outs R600_Predicate_Bit:$dst),
1950 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
1952 let FlagOperandIdx = 3;
1955 let isTerminator = 1, isBranch = 1 in {
1956 def JUMP_COND : InstR600 <
1958 (ins brtarget:$target, R600_Predicate_Bit:$p),
1959 "JUMP $target ($p)",
1963 def JUMP : InstR600 <
1965 (ins brtarget:$target),
1970 let isPredicable = 1;
1974 } // End isTerminator = 1, isBranch = 1
1976 let usesCustomInserter = 1 in {
1978 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
1980 def MASK_WRITE : AMDGPUShaderInst <
1982 (ins R600_Reg32:$src),
1987 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
1991 (outs R600_Reg128:$dst),
1992 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
1993 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1994 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1995 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
1996 imm:$resourceId, imm:$samplerId, imm:$textureTarget))],
2001 def TXD_SHADOW: InstR600 <
2002 (outs R600_Reg128:$dst),
2003 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
2004 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
2005 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
2006 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
2007 imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))],
2012 } // End isPseudo = 1
2013 } // End usesCustomInserter = 1
2015 def CLAMP_R600 : CLAMP <R600_Reg32>;
2016 def FABS_R600 : FABS<R600_Reg32>;
2017 def FNEG_R600 : FNEG<R600_Reg32>;
2019 //===---------------------------------------------------------------------===//
2020 // Return instruction
2021 //===---------------------------------------------------------------------===//
2022 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
2023 usesCustomInserter = 1 in {
2024 def RETURN : ILFormat<(outs), (ins variable_ops),
2025 "RETURN", [(IL_retflag)]>;
2029 //===----------------------------------------------------------------------===//
2030 // Constant Buffer Addressing Support
2031 //===----------------------------------------------------------------------===//
2033 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
2034 def CONST_COPY : Instruction {
2035 let OutOperandList = (outs R600_Reg32:$dst);
2036 let InOperandList = (ins i32imm:$src);
2038 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
2039 let AsmString = "CONST_COPY";
2040 let neverHasSideEffects = 1;
2041 let isAsCheapAsAMove = 1;
2042 let Itinerary = NullALU;
2044 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
2046 def TEX_VTX_CONSTBUF :
2047 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
2048 [(set v4i32:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
2049 VTX_WORD1_GPR, VTX_WORD0_eg {
2053 let FETCH_WHOLE_QUAD = 0;
2057 let USE_CONST_FIELDS = 0;
2058 let NUM_FORMAT_ALL = 2;
2059 let FORMAT_COMP_ALL = 1;
2060 let SRF_MODE_ALL = 1;
2061 let MEGA_FETCH_COUNT = 16;
2066 let DATA_FORMAT = 35;
2068 let Inst{31-0} = Word0;
2069 let Inst{63-32} = Word1;
2071 // LLVM can only encode 64-bit instructions, so these fields are manually
2072 // encoded in R600CodeEmitter
2075 // bits<2> ENDIAN_SWAP = 0;
2076 // bits<1> CONST_BUF_NO_STRIDE = 0;
2077 // bits<1> MEGA_FETCH = 0;
2078 // bits<1> ALT_CONST = 0;
2079 // bits<2> BUFFER_INDEX_MODE = 0;
2083 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2084 // is done in R600CodeEmitter
2086 // Inst{79-64} = OFFSET;
2087 // Inst{81-80} = ENDIAN_SWAP;
2088 // Inst{82} = CONST_BUF_NO_STRIDE;
2089 // Inst{83} = MEGA_FETCH;
2090 // Inst{84} = ALT_CONST;
2091 // Inst{86-85} = BUFFER_INDEX_MODE;
2092 // Inst{95-86} = 0; Reserved
2094 // VTX_WORD3 (Padding)
2096 // Inst{127-96} = 0;
2101 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
2102 [(set v4f32:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
2103 VTX_WORD1_GPR, VTX_WORD0_eg {
2107 let FETCH_WHOLE_QUAD = 0;
2111 let USE_CONST_FIELDS = 1;
2112 let NUM_FORMAT_ALL = 0;
2113 let FORMAT_COMP_ALL = 0;
2114 let SRF_MODE_ALL = 1;
2115 let MEGA_FETCH_COUNT = 16;
2120 let DATA_FORMAT = 0;
2122 let Inst{31-0} = Word0;
2123 let Inst{63-32} = Word1;
2125 // LLVM can only encode 64-bit instructions, so these fields are manually
2126 // encoded in R600CodeEmitter
2129 // bits<2> ENDIAN_SWAP = 0;
2130 // bits<1> CONST_BUF_NO_STRIDE = 0;
2131 // bits<1> MEGA_FETCH = 0;
2132 // bits<1> ALT_CONST = 0;
2133 // bits<2> BUFFER_INDEX_MODE = 0;
2137 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
2138 // is done in R600CodeEmitter
2140 // Inst{79-64} = OFFSET;
2141 // Inst{81-80} = ENDIAN_SWAP;
2142 // Inst{82} = CONST_BUF_NO_STRIDE;
2143 // Inst{83} = MEGA_FETCH;
2144 // Inst{84} = ALT_CONST;
2145 // Inst{86-85} = BUFFER_INDEX_MODE;
2146 // Inst{95-86} = 0; Reserved
2148 // VTX_WORD3 (Padding)
2150 // Inst{127-96} = 0;
2156 //===--------------------------------------------------------------------===//
2157 // Instructions support
2158 //===--------------------------------------------------------------------===//
2159 //===---------------------------------------------------------------------===//
2160 // Custom Inserter for Branches and returns, this eventually will be a
2162 //===---------------------------------------------------------------------===//
2163 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
2164 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
2165 "; Pseudo unconditional branch instruction",
2167 defm BRANCH_COND : BranchConditional<IL_brcond>;
2170 //===---------------------------------------------------------------------===//
2171 // Flow and Program control Instructions
2172 //===---------------------------------------------------------------------===//
2173 let isTerminator=1 in {
2174 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
2175 !strconcat("SWITCH", " $src"), []>;
2176 def CASE : ILFormat< (outs), (ins GPRI32:$src),
2177 !strconcat("CASE", " $src"), []>;
2178 def BREAK : ILFormat< (outs), (ins),
2180 def CONTINUE : ILFormat< (outs), (ins),
2182 def DEFAULT : ILFormat< (outs), (ins),
2184 def ELSE : ILFormat< (outs), (ins),
2186 def ENDSWITCH : ILFormat< (outs), (ins),
2188 def ENDMAIN : ILFormat< (outs), (ins),
2190 def END : ILFormat< (outs), (ins),
2192 def ENDFUNC : ILFormat< (outs), (ins),
2194 def ENDIF : ILFormat< (outs), (ins),
2196 def WHILELOOP : ILFormat< (outs), (ins),
2198 def ENDLOOP : ILFormat< (outs), (ins),
2200 def FUNC : ILFormat< (outs), (ins),
2202 def RETDYN : ILFormat< (outs), (ins),
2204 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2205 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
2206 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2207 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
2208 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2209 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
2210 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2211 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
2212 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2213 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
2214 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2215 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
2216 defm IFC : BranchInstr2<"IFC">;
2217 defm BREAKC : BranchInstr2<"BREAKC">;
2218 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
2221 //===----------------------------------------------------------------------===//
2223 //===----------------------------------------------------------------------===//
2225 // CND*_INT Pattterns for f32 True / False values
2227 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
2228 (selectcc i32:$src0, 0, f32:$src1, f32:$src2, cc),
2229 (cnd $src0, $src1, $src2)
2232 def : CND_INT_f32 <CNDE_INT, SETEQ>;
2233 def : CND_INT_f32 <CNDGT_INT, SETGT>;
2234 def : CND_INT_f32 <CNDGE_INT, SETGE>;
2236 //CNDGE_INT extra pattern
2238 (selectcc i32:$src0, -1, i32:$src1, i32:$src2, COND_GT),
2239 (CNDGE_INT $src0, $src1, $src2)
2245 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
2249 (int_AMDGPU_kill f32:$src0),
2250 (MASK_WRITE (KILLGT (f32 ZERO), $src0))
2255 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LT),
2261 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LE),
2265 // SETGT_DX10 reverse args
2267 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LT),
2268 (SETGT_DX10 $src1, $src0)
2271 // SETGE_DX10 reverse args
2273 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LE),
2274 (SETGE_DX10 $src1, $src0)
2277 // SETGT_INT reverse args
2279 (selectcc i32:$src0, i32:$src1, -1, 0, SETLT),
2280 (SETGT_INT $src1, $src0)
2283 // SETGE_INT reverse args
2285 (selectcc i32:$src0, i32:$src1, -1, 0, SETLE),
2286 (SETGE_INT $src1, $src0)
2289 // SETGT_UINT reverse args
2291 (selectcc i32:$src0, i32:$src1, -1, 0, SETULT),
2292 (SETGT_UINT $src1, $src0)
2295 // SETGE_UINT reverse args
2297 (selectcc i32:$src0, i32:$src1, -1, 0, SETULE),
2298 (SETGE_UINT $src1, $src0)
2301 // The next two patterns are special cases for handling 'true if ordered' and
2302 // 'true if unordered' conditionals. The assumption here is that the behavior of
2303 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
2305 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
2306 // We assume that SETE returns false when one of the operands is NAN and
2307 // SNE returns true when on of the operands is NAN
2309 //SETE - 'true if ordered'
2311 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETO),
2315 //SETE_DX10 - 'true if ordered'
2317 (selectcc f32:$src0, f32:$src1, -1, 0, SETO),
2318 (SETE_DX10 $src0, $src1)
2321 //SNE - 'true if unordered'
2323 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETUO),
2327 //SETNE_DX10 - 'true if ordered'
2329 (selectcc f32:$src0, f32:$src1, -1, 0, SETUO),
2330 (SETNE_DX10 $src0, $src1)
2333 def : Extract_Element <f32, v4f32, 0, sub0>;
2334 def : Extract_Element <f32, v4f32, 1, sub1>;
2335 def : Extract_Element <f32, v4f32, 2, sub2>;
2336 def : Extract_Element <f32, v4f32, 3, sub3>;
2338 def : Insert_Element <f32, v4f32, 0, sub0>;
2339 def : Insert_Element <f32, v4f32, 1, sub1>;
2340 def : Insert_Element <f32, v4f32, 2, sub2>;
2341 def : Insert_Element <f32, v4f32, 3, sub3>;
2343 def : Extract_Element <i32, v4i32, 0, sub0>;
2344 def : Extract_Element <i32, v4i32, 1, sub1>;
2345 def : Extract_Element <i32, v4i32, 2, sub2>;
2346 def : Extract_Element <i32, v4i32, 3, sub3>;
2348 def : Insert_Element <i32, v4i32, 0, sub0>;
2349 def : Insert_Element <i32, v4i32, 1, sub1>;
2350 def : Insert_Element <i32, v4i32, 2, sub2>;
2351 def : Insert_Element <i32, v4i32, 3, sub3>;
2353 def : Vector4_Build <v4f32, f32>;
2354 def : Vector4_Build <v4i32, i32>;
2356 def : Extract_Element <f32, v2f32, 0, sub0>;
2357 def : Extract_Element <f32, v2f32, 1, sub1>;
2359 def : Insert_Element <f32, v2f32, 0, sub0>;
2360 def : Insert_Element <f32, v2f32, 1, sub1>;
2362 def : Extract_Element <i32, v2i32, 0, sub0>;
2363 def : Extract_Element <i32, v2i32, 1, sub1>;
2365 def : Insert_Element <i32, v2i32, 0, sub0>;
2366 def : Insert_Element <i32, v2i32, 1, sub1>;
2368 // bitconvert patterns
2370 def : BitConvert <i32, f32, R600_Reg32>;
2371 def : BitConvert <f32, i32, R600_Reg32>;
2372 def : BitConvert <v2f32, v2i32, R600_Reg64>;
2373 def : BitConvert <v2i32, v2f32, R600_Reg64>;
2374 def : BitConvert <v4f32, v4i32, R600_Reg128>;
2375 def : BitConvert <v4i32, v4f32, R600_Reg128>;
2377 // DWORDADDR pattern
2378 def : DwordAddrPat <i32, R600_Reg32>;
2380 } // End isR600toCayman Predicate