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 $bank_swizzle"),
107 R600ALU_Word1_OP2 <inst> {
113 let update_exec_mask = 0;
115 let HasNativeOperands = 1;
117 let DisableEncoding = "$literal";
118 let UseNamedOperandTable = 1;
120 let Inst{31-0} = Word0;
121 let Inst{63-32} = Word1;
124 class R600_1OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
125 InstrItinClass itin = AnyALU> :
126 R600_1OP <inst, opName,
127 [(set R600_Reg32:$dst, (node R600_Reg32:$src0))]
130 // If you add or change the operands for R600_2OP instructions, you must
131 // also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
132 // R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
133 class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
134 InstrItinClass itin = AnyALU> :
135 InstR600 <(outs R600_Reg32:$dst),
136 (ins UEM:$update_exec_mask, UP:$update_pred, WRITE:$write,
137 OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
138 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
139 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs, SEL:$src1_sel,
140 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
141 BANK_SWIZZLE:$bank_swizzle),
142 !strconcat(" ", opName,
143 "$clamp $last $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
144 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
145 "$src1_neg$src1_abs$src1$src1_abs$src1_rel, "
146 "$pred_sel $bank_swizzle"),
150 R600ALU_Word1_OP2 <inst> {
152 let HasNativeOperands = 1;
154 let DisableEncoding = "$literal";
155 let UseNamedOperandTable = 1;
157 let Inst{31-0} = Word0;
158 let Inst{63-32} = Word1;
161 class R600_2OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
162 InstrItinClass itim = AnyALU> :
163 R600_2OP <inst, opName,
164 [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
168 // If you add our change the operands for R600_3OP instructions, you must
169 // also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
170 // R600InstrInfo::buildDefaultInstruction(), and
171 // R600InstrInfo::getOperandIdx().
172 class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
173 InstrItinClass itin = AnyALU> :
174 InstR600 <(outs R600_Reg32:$dst),
175 (ins REL:$dst_rel, CLAMP:$clamp,
176 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, SEL:$src0_sel,
177 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
178 R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
179 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
180 BANK_SWIZZLE:$bank_swizzle),
181 !strconcat(" ", opName, "$clamp $last $dst$dst_rel, "
182 "$src0_neg$src0$src0_rel, "
183 "$src1_neg$src1$src1_rel, "
184 "$src2_neg$src2$src2_rel, "
190 R600ALU_Word1_OP3<inst>{
192 let HasNativeOperands = 1;
193 let DisableEncoding = "$literal";
195 let UseNamedOperandTable = 1;
197 let Inst{31-0} = Word0;
198 let Inst{63-32} = Word1;
201 class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
202 InstrItinClass itin = VecALU> :
203 InstR600 <(outs R600_Reg32:$dst),
211 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
213 def TEX_SHADOW : PatLeaf<
215 [{uint32_t TType = (uint32_t)N->getZExtValue();
216 return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
220 def TEX_RECT : PatLeaf<
222 [{uint32_t TType = (uint32_t)N->getZExtValue();
227 def TEX_ARRAY : PatLeaf<
229 [{uint32_t TType = (uint32_t)N->getZExtValue();
230 return TType == 9 || TType == 10 || TType == 15 || TType == 16;
234 def TEX_SHADOW_ARRAY : PatLeaf<
236 [{uint32_t TType = (uint32_t)N->getZExtValue();
237 return TType == 11 || TType == 12 || TType == 17;
241 class EG_CF_RAT <bits <8> cfinst, bits <6> ratinst, bits<4> mask, dag outs,
242 dag ins, string asm, list<dag> pattern> :
243 InstR600ISA <outs, ins, asm, pattern>,
244 CF_ALLOC_EXPORT_WORD0_RAT, CF_ALLOC_EXPORT_WORD1_BUF {
247 let rat_inst = ratinst;
249 // XXX: Have a separate instruction for non-indexed writes.
255 let comp_mask = mask;
258 let cf_inst = cfinst;
262 let Inst{31-0} = Word0;
263 let Inst{63-32} = Word1;
267 class VTX_READ <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
268 : InstR600ISA <outs, (ins MEMxi:$src_gpr), name, pattern>,
273 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
274 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
275 // however, based on my testing if USE_CONST_FIELDS is set, then all
276 // these fields need to be set to 0.
277 let USE_CONST_FIELDS = 0;
278 let NUM_FORMAT_ALL = 1;
279 let FORMAT_COMP_ALL = 0;
280 let SRF_MODE_ALL = 0;
282 let Inst{63-32} = Word1;
283 // LLVM can only encode 64-bit instructions, so these fields are manually
284 // encoded in R600CodeEmitter
287 // bits<2> ENDIAN_SWAP = 0;
288 // bits<1> CONST_BUF_NO_STRIDE = 0;
289 // bits<1> MEGA_FETCH = 0;
290 // bits<1> ALT_CONST = 0;
291 // bits<2> BUFFER_INDEX_MODE = 0;
293 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
294 // is done in R600CodeEmitter
296 // Inst{79-64} = OFFSET;
297 // Inst{81-80} = ENDIAN_SWAP;
298 // Inst{82} = CONST_BUF_NO_STRIDE;
299 // Inst{83} = MEGA_FETCH;
300 // Inst{84} = ALT_CONST;
301 // Inst{86-85} = BUFFER_INDEX_MODE;
302 // Inst{95-86} = 0; Reserved
304 // VTX_WORD3 (Padding)
311 class LoadParamFrag <PatFrag load_type> : PatFrag <
312 (ops node:$ptr), (load_type node:$ptr),
313 [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
316 def load_param : LoadParamFrag<load>;
317 def load_param_zexti8 : LoadParamFrag<zextloadi8>;
318 def load_param_zexti16 : LoadParamFrag<zextloadi16>;
320 def isR600 : Predicate<"Subtarget.getGeneration() <= AMDGPUSubtarget::R700">;
321 def isR700 : Predicate<"Subtarget.getGeneration() == AMDGPUSubtarget::R700">;
322 def isEG : Predicate<
323 "Subtarget.getGeneration() >= AMDGPUSubtarget::EVERGREEN && "
324 "Subtarget.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS && "
325 "!Subtarget.hasCaymanISA()">;
327 def isCayman : Predicate<"Subtarget.hasCaymanISA()">;
328 def isEGorCayman : Predicate<"Subtarget.getGeneration() == "
329 "AMDGPUSubtarget::EVERGREEN"
330 "|| Subtarget.getGeneration() =="
331 "AMDGPUSubtarget::NORTHERN_ISLANDS">;
333 def isR600toCayman : Predicate<
334 "Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS">;
336 //===----------------------------------------------------------------------===//
338 //===----------------------------------------------------------------------===//
340 def INTERP_PAIR_XY : AMDGPUShaderInst <
341 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
342 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
343 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
346 def INTERP_PAIR_ZW : AMDGPUShaderInst <
347 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
348 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
349 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
352 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
353 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
357 def DOT4 : SDNode<"AMDGPUISD::DOT4",
358 SDTypeProfile<1, 8, [SDTCisFP<0>, SDTCisVT<1, f32>, SDTCisVT<2, f32>,
359 SDTCisVT<3, f32>, SDTCisVT<4, f32>, SDTCisVT<5, f32>,
360 SDTCisVT<6, f32>, SDTCisVT<7, f32>, SDTCisVT<8, f32>]>,
364 def TEXTURE_FETCH_Type : SDTypeProfile<1, 19, [SDTCisFP<0>]>;
366 def TEXTURE_FETCH: SDNode<"AMDGPUISD::TEXTURE_FETCH", TEXTURE_FETCH_Type, []>;
368 multiclass TexPattern<bits<32> TextureOp, Instruction inst, ValueType vt = v4f32> {
369 def : Pat<(TEXTURE_FETCH (i32 TextureOp), vt:$SRC_GPR,
370 (i32 imm:$srcx), (i32 imm:$srcy), (i32 imm:$srcz), (i32 imm:$srcw),
371 (i32 imm:$offsetx), (i32 imm:$offsety), (i32 imm:$offsetz),
372 (i32 imm:$DST_SEL_X), (i32 imm:$DST_SEL_Y), (i32 imm:$DST_SEL_Z),
373 (i32 imm:$DST_SEL_W),
374 (i32 imm:$RESOURCE_ID), (i32 imm:$SAMPLER_ID),
375 (i32 imm:$COORD_TYPE_X), (i32 imm:$COORD_TYPE_Y), (i32 imm:$COORD_TYPE_Z),
376 (i32 imm:$COORD_TYPE_W)),
377 (inst R600_Reg128:$SRC_GPR,
378 imm:$srcx, imm:$srcy, imm:$srcz, imm:$srcw,
379 imm:$offsetx, imm:$offsety, imm:$offsetz,
380 imm:$DST_SEL_X, imm:$DST_SEL_Y, imm:$DST_SEL_Z,
382 imm:$RESOURCE_ID, imm:$SAMPLER_ID,
383 imm:$COORD_TYPE_X, imm:$COORD_TYPE_Y, imm:$COORD_TYPE_Z,
387 //===----------------------------------------------------------------------===//
388 // Interpolation Instructions
389 //===----------------------------------------------------------------------===//
391 def INTERP_VEC_LOAD : AMDGPUShaderInst <
392 (outs R600_Reg128:$dst),
394 "INTERP_LOAD $src0 : $dst",
397 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
398 let bank_swizzle = 5;
401 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
402 let bank_swizzle = 5;
405 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
407 //===----------------------------------------------------------------------===//
408 // Export Instructions
409 //===----------------------------------------------------------------------===//
411 def ExportType : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
413 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
414 [SDNPHasChain, SDNPSideEffect]>;
417 field bits<32> Word0;
424 let Word0{12-0} = arraybase;
425 let Word0{14-13} = type;
426 let Word0{21-15} = gpr;
427 let Word0{22} = 0; // RW_REL
428 let Word0{29-23} = 0; // INDEX_GPR
429 let Word0{31-30} = elem_size;
432 class ExportSwzWord1 {
433 field bits<32> Word1;
442 let Word1{2-0} = sw_x;
443 let Word1{5-3} = sw_y;
444 let Word1{8-6} = sw_z;
445 let Word1{11-9} = sw_w;
448 class ExportBufWord1 {
449 field bits<32> Word1;
456 let Word1{11-0} = arraySize;
457 let Word1{15-12} = compMask;
460 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
461 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
463 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
464 0, 61, 0, 7, 7, 7, cf_inst, 0)
467 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
469 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
470 0, 61, 7, 0, 7, 7, cf_inst, 0)
473 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
475 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
478 def : Pat<(int_R600_store_dummy 1),
480 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
483 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
484 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
485 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
486 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
491 multiclass SteamOutputExportPattern<Instruction ExportInst,
492 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
494 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
495 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
496 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
497 4095, imm:$mask, buf0inst, 0)>;
499 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
500 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
501 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
502 4095, imm:$mask, buf1inst, 0)>;
504 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
505 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
506 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
507 4095, imm:$mask, buf2inst, 0)>;
509 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
510 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
511 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
512 4095, imm:$mask, buf3inst, 0)>;
515 // Export Instructions should not be duplicated by TailDuplication pass
516 // (which assumes that duplicable instruction are affected by exec mask)
517 let usesCustomInserter = 1, isNotDuplicable = 1 in {
519 class ExportSwzInst : InstR600ISA<(
521 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
522 i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
524 !strconcat("EXPORT", " $gpr"),
525 []>, ExportWord0, ExportSwzWord1 {
527 let Inst{31-0} = Word0;
528 let Inst{63-32} = Word1;
531 } // End usesCustomInserter = 1
533 class ExportBufInst : InstR600ISA<(
535 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
536 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
537 !strconcat("EXPORT", " $gpr"),
538 []>, ExportWord0, ExportBufWord1 {
540 let Inst{31-0} = Word0;
541 let Inst{63-32} = Word1;
544 //===----------------------------------------------------------------------===//
545 // Control Flow Instructions
546 //===----------------------------------------------------------------------===//
549 def KCACHE : InstFlag<"printKCache">;
551 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
552 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1,
553 KCACHE:$KCACHE_MODE0, KCACHE:$KCACHE_MODE1,
554 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1,
556 !strconcat(OpName, " $COUNT, @$ADDR, "
557 "KC0[$KCACHE_MODE0], KC1[$KCACHE_MODE1]"),
558 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
563 let WHOLE_QUAD_MODE = 0;
566 let Inst{31-0} = Word0;
567 let Inst{63-32} = Word1;
570 class CF_WORD0_R600 {
571 field bits<32> Word0;
578 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
579 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
586 let VALID_PIXEL_MODE = 0;
588 let COUNT = CNT{2-0};
590 let COUNT_3 = CNT{3};
591 let END_OF_PROGRAM = 0;
592 let WHOLE_QUAD_MODE = 0;
594 let Inst{31-0} = Word0;
595 let Inst{63-32} = Word1;
598 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
599 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
604 let JUMPTABLE_SEL = 0;
606 let VALID_PIXEL_MODE = 0;
608 let END_OF_PROGRAM = 0;
610 let Inst{31-0} = Word0;
611 let Inst{63-32} = Word1;
614 def CF_ALU : ALU_CLAUSE<8, "ALU">;
615 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
617 def FETCH_CLAUSE : AMDGPUInst <(outs),
618 (ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
624 def ALU_CLAUSE : AMDGPUInst <(outs),
625 (ins i32imm:$addr), "ALU clause starting at $addr:", [] > {
631 def LITERALS : AMDGPUInst <(outs),
632 (ins LITERAL:$literal1, LITERAL:$literal2), "$literal1, $literal2", [] > {
637 let Inst{31-0} = literal1;
638 let Inst{63-32} = literal2;
641 def PAD : AMDGPUInst <(outs), (ins), "PAD", [] > {
645 let Predicates = [isR600toCayman] in {
647 //===----------------------------------------------------------------------===//
648 // Common Instructions R600, R700, Evergreen, Cayman
649 //===----------------------------------------------------------------------===//
651 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
652 // Non-IEEE MUL: 0 * anything = 0
653 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
654 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
655 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
656 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
658 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
659 // so some of the instruction names don't match the asm string.
660 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
661 def SETE : R600_2OP <
663 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_EQ))]
668 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GT))]
673 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GE))]
678 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_NE))]
681 def SETE_DX10 : R600_2OP <
683 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_EQ))]
686 def SETGT_DX10 : R600_2OP <
688 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GT))]
691 def SETGE_DX10 : R600_2OP <
693 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GE))]
696 def SETNE_DX10 : R600_2OP <
698 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_NE))]
701 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
702 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
703 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
704 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
705 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
707 def MOV : R600_1OP <0x19, "MOV", []>;
709 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
711 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
712 (outs R600_Reg32:$dst),
718 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
720 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
723 (MOV_IMM_I32 imm:$val)
726 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
729 (MOV_IMM_F32 fpimm:$val)
732 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
733 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
734 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
735 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
737 let hasSideEffects = 1 in {
739 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
741 } // end hasSideEffects
743 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
744 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
745 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
746 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
747 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
748 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
749 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
750 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
751 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
752 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
754 def SETE_INT : R600_2OP <
756 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETEQ))]
759 def SETGT_INT : R600_2OP <
761 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGT))]
764 def SETGE_INT : R600_2OP <
766 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGE))]
769 def SETNE_INT : R600_2OP <
771 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETNE))]
774 def SETGT_UINT : R600_2OP <
776 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGT))]
779 def SETGE_UINT : R600_2OP <
781 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGE))]
784 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
785 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
786 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
787 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
789 def CNDE_INT : R600_3OP <
791 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_EQ))]
794 def CNDGE_INT : R600_3OP <
796 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GE))]
799 def CNDGT_INT : R600_3OP <
801 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GT))]
804 //===----------------------------------------------------------------------===//
805 // Texture instructions
806 //===----------------------------------------------------------------------===//
808 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
810 class R600_TEX <bits<11> inst, string opName> :
811 InstR600 <(outs R600_Reg128:$DST_GPR),
812 (ins R600_Reg128:$SRC_GPR,
813 RSel:$srcx, RSel:$srcy, RSel:$srcz, RSel:$srcw,
814 i32imm:$offsetx, i32imm:$offsety, i32imm:$offsetz,
815 RSel:$DST_SEL_X, RSel:$DST_SEL_Y, RSel:$DST_SEL_Z, RSel:$DST_SEL_W,
816 i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
817 CT:$COORD_TYPE_X, CT:$COORD_TYPE_Y, CT:$COORD_TYPE_Z,
820 " $DST_GPR.$DST_SEL_X$DST_SEL_Y$DST_SEL_Z$DST_SEL_W, "
821 "$SRC_GPR.$srcx$srcy$srcz$srcw "
822 "RID:$RESOURCE_ID SID:$SAMPLER_ID "
823 "CT:$COORD_TYPE_X$COORD_TYPE_Y$COORD_TYPE_Z$COORD_TYPE_W"),
825 NullALU>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
826 let Inst{31-0} = Word0;
827 let Inst{63-32} = Word1;
829 let TEX_INST = inst{4-0};
835 let FETCH_WHOLE_QUAD = 0;
837 let SAMPLER_INDEX_MODE = 0;
838 let RESOURCE_INDEX_MODE = 0;
843 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
847 def TEX_SAMPLE : R600_TEX <0x10, "TEX_SAMPLE">;
848 def TEX_SAMPLE_C : R600_TEX <0x18, "TEX_SAMPLE_C">;
849 def TEX_SAMPLE_L : R600_TEX <0x11, "TEX_SAMPLE_L">;
850 def TEX_SAMPLE_C_L : R600_TEX <0x19, "TEX_SAMPLE_C_L">;
851 def TEX_SAMPLE_LB : R600_TEX <0x12, "TEX_SAMPLE_LB">;
852 def TEX_SAMPLE_C_LB : R600_TEX <0x1A, "TEX_SAMPLE_C_LB">;
853 def TEX_LD : R600_TEX <0x03, "TEX_LD">;
854 def TEX_GET_TEXTURE_RESINFO : R600_TEX <0x04, "TEX_GET_TEXTURE_RESINFO">;
855 def TEX_GET_GRADIENTS_H : R600_TEX <0x07, "TEX_GET_GRADIENTS_H">;
856 def TEX_GET_GRADIENTS_V : R600_TEX <0x08, "TEX_GET_GRADIENTS_V">;
857 def TEX_SET_GRADIENTS_H : R600_TEX <0x0B, "TEX_SET_GRADIENTS_H">;
858 def TEX_SET_GRADIENTS_V : R600_TEX <0x0C, "TEX_SET_GRADIENTS_V">;
859 def TEX_SAMPLE_G : R600_TEX <0x14, "TEX_SAMPLE_G">;
860 def TEX_SAMPLE_C_G : R600_TEX <0x1C, "TEX_SAMPLE_C_G">;
862 defm : TexPattern<0, TEX_SAMPLE>;
863 defm : TexPattern<1, TEX_SAMPLE_C>;
864 defm : TexPattern<2, TEX_SAMPLE_L>;
865 defm : TexPattern<3, TEX_SAMPLE_C_L>;
866 defm : TexPattern<4, TEX_SAMPLE_LB>;
867 defm : TexPattern<5, TEX_SAMPLE_C_LB>;
868 defm : TexPattern<6, TEX_LD, v4i32>;
869 defm : TexPattern<7, TEX_GET_TEXTURE_RESINFO, v4i32>;
870 defm : TexPattern<8, TEX_GET_GRADIENTS_H>;
871 defm : TexPattern<9, TEX_GET_GRADIENTS_V>;
873 //===----------------------------------------------------------------------===//
874 // Helper classes for common instructions
875 //===----------------------------------------------------------------------===//
877 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
882 class MULADD_Common <bits<5> inst> : R600_3OP <
887 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
889 [(set f32:$dst, (fadd (fmul f32:$src0, f32:$src1), f32:$src2))]
892 class CNDE_Common <bits<5> inst> : R600_3OP <
894 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_EQ))]
897 class CNDGT_Common <bits<5> inst> : R600_3OP <
899 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GT))]
902 class CNDGE_Common <bits<5> inst> : R600_3OP <
904 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GE))]
908 let isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
909 class R600_VEC2OP<list<dag> pattern> : InstR600 <(outs R600_Reg32:$dst), (ins
911 UEM:$update_exec_mask_X, UP:$update_pred_X, WRITE:$write_X,
912 OMOD:$omod_X, REL:$dst_rel_X, CLAMP:$clamp_X,
913 R600_TReg32_X:$src0_X, NEG:$src0_neg_X, REL:$src0_rel_X, ABS:$src0_abs_X, SEL:$src0_sel_X,
914 R600_TReg32_X:$src1_X, NEG:$src1_neg_X, REL:$src1_rel_X, ABS:$src1_abs_X, SEL:$src1_sel_X,
915 R600_Pred:$pred_sel_X,
917 UEM:$update_exec_mask_Y, UP:$update_pred_Y, WRITE:$write_Y,
918 OMOD:$omod_Y, REL:$dst_rel_Y, CLAMP:$clamp_Y,
919 R600_TReg32_Y:$src0_Y, NEG:$src0_neg_Y, REL:$src0_rel_Y, ABS:$src0_abs_Y, SEL:$src0_sel_Y,
920 R600_TReg32_Y:$src1_Y, NEG:$src1_neg_Y, REL:$src1_rel_Y, ABS:$src1_abs_Y, SEL:$src1_sel_Y,
921 R600_Pred:$pred_sel_Y,
923 UEM:$update_exec_mask_Z, UP:$update_pred_Z, WRITE:$write_Z,
924 OMOD:$omod_Z, REL:$dst_rel_Z, CLAMP:$clamp_Z,
925 R600_TReg32_Z:$src0_Z, NEG:$src0_neg_Z, REL:$src0_rel_Z, ABS:$src0_abs_Z, SEL:$src0_sel_Z,
926 R600_TReg32_Z:$src1_Z, NEG:$src1_neg_Z, REL:$src1_rel_Z, ABS:$src1_abs_Z, SEL:$src1_sel_Z,
927 R600_Pred:$pred_sel_Z,
929 UEM:$update_exec_mask_W, UP:$update_pred_W, WRITE:$write_W,
930 OMOD:$omod_W, REL:$dst_rel_W, CLAMP:$clamp_W,
931 R600_TReg32_W:$src0_W, NEG:$src0_neg_W, REL:$src0_rel_W, ABS:$src0_abs_W, SEL:$src0_sel_W,
932 R600_TReg32_W:$src1_W, NEG:$src1_neg_W, REL:$src1_rel_W, ABS:$src1_abs_W, SEL:$src1_sel_W,
933 R600_Pred:$pred_sel_W,
934 LITERAL:$literal0, LITERAL:$literal1),
939 let UseNamedOperandTable = 1;
944 def DOT_4 : R600_VEC2OP<[(set R600_Reg32:$dst, (DOT4
945 R600_TReg32_X:$src0_X, R600_TReg32_X:$src1_X,
946 R600_TReg32_Y:$src0_Y, R600_TReg32_Y:$src1_Y,
947 R600_TReg32_Z:$src0_Z, R600_TReg32_Z:$src1_Z,
948 R600_TReg32_W:$src0_W, R600_TReg32_W:$src1_W))]>;
951 class DOT4_Common <bits<11> inst> : R600_2OP <inst, "DOT4", []>;
954 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
955 multiclass CUBE_Common <bits<11> inst> {
957 def _pseudo : InstR600 <
958 (outs R600_Reg128:$dst),
959 (ins R600_Reg128:$src0),
961 [(set v4f32:$dst, (int_AMDGPU_cube v4f32:$src0))],
965 let UseNamedOperandTable = 1;
968 def _real : R600_2OP <inst, "CUBE", []>;
970 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
972 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
973 inst, "EXP_IEEE", fexp2
976 let Itinerary = TransALU;
979 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
980 inst, "FLT_TO_INT", fp_to_sint
983 let Itinerary = TransALU;
986 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
987 inst, "INT_TO_FLT", sint_to_fp
990 let Itinerary = TransALU;
993 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
994 inst, "FLT_TO_UINT", fp_to_uint
997 let Itinerary = TransALU;
1000 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
1001 inst, "UINT_TO_FLT", uint_to_fp
1004 let Itinerary = TransALU;
1007 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
1008 inst, "LOG_CLAMPED", []
1011 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
1012 inst, "LOG_IEEE", flog2
1015 let Itinerary = TransALU;
1018 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
1019 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
1020 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
1021 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
1022 inst, "MULHI_INT", mulhs
1025 let Itinerary = TransALU;
1027 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
1028 inst, "MULHI", mulhu
1031 let Itinerary = TransALU;
1033 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
1034 inst, "MULLO_INT", mul
1037 let Itinerary = TransALU;
1039 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []> {
1041 let Itinerary = TransALU;
1044 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
1045 inst, "RECIP_CLAMPED", []
1048 let Itinerary = TransALU;
1051 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
1052 inst, "RECIP_IEEE", [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
1055 let Itinerary = TransALU;
1058 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
1059 inst, "RECIP_UINT", AMDGPUurecip
1062 let Itinerary = TransALU;
1065 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1066 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1069 let Itinerary = TransALU;
1072 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1073 inst, "RECIPSQRT_IEEE", []
1076 let Itinerary = TransALU;
1079 class SIN_Common <bits<11> inst> : R600_1OP <
1083 let Itinerary = TransALU;
1086 class COS_Common <bits<11> inst> : R600_1OP <
1090 let Itinerary = TransALU;
1093 //===----------------------------------------------------------------------===//
1094 // Helper patterns for complex intrinsics
1095 //===----------------------------------------------------------------------===//
1097 multiclass DIV_Common <InstR600 recip_ieee> {
1099 (int_AMDGPU_div f32:$src0, f32:$src1),
1100 (MUL_IEEE $src0, (recip_ieee $src1))
1104 (fdiv f32:$src0, f32:$src1),
1105 (MUL_IEEE $src0, (recip_ieee $src1))
1109 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee>
1111 (int_TGSI_lit_z f32:$src_x, f32:$src_y, f32:$src_w),
1112 (exp_ieee (mul_lit (log_clamped (MAX $src_y, (f32 ZERO))), $src_w, $src_x))
1115 //===----------------------------------------------------------------------===//
1116 // R600 / R700 Instructions
1117 //===----------------------------------------------------------------------===//
1119 let Predicates = [isR600] in {
1121 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1122 def MULADD_r600 : MULADD_Common<0x10>;
1123 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1124 def CNDE_r600 : CNDE_Common<0x18>;
1125 def CNDGT_r600 : CNDGT_Common<0x19>;
1126 def CNDGE_r600 : CNDGE_Common<0x1A>;
1127 def DOT4_r600 : DOT4_Common<0x50>;
1128 defm CUBE_r600 : CUBE_Common<0x52>;
1129 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1130 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1131 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1132 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1133 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1134 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1135 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1136 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1137 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1138 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1139 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1140 def SIN_r600 : SIN_Common<0x6E>;
1141 def COS_r600 : COS_Common<0x6F>;
1142 def ASHR_r600 : ASHR_Common<0x70>;
1143 def LSHR_r600 : LSHR_Common<0x71>;
1144 def LSHL_r600 : LSHL_Common<0x72>;
1145 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1146 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1147 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1148 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1149 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1151 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1152 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL>;
1153 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1155 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
1157 def R600_ExportSwz : ExportSwzInst {
1158 let Word1{20-17} = 0; // BURST_COUNT
1159 let Word1{21} = eop;
1160 let Word1{22} = 1; // VALID_PIXEL_MODE
1161 let Word1{30-23} = inst;
1162 let Word1{31} = 1; // BARRIER
1164 defm : ExportPattern<R600_ExportSwz, 39>;
1166 def R600_ExportBuf : ExportBufInst {
1167 let Word1{20-17} = 0; // BURST_COUNT
1168 let Word1{21} = eop;
1169 let Word1{22} = 1; // VALID_PIXEL_MODE
1170 let Word1{30-23} = inst;
1171 let Word1{31} = 1; // BARRIER
1173 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1175 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$CNT),
1176 "TEX $CNT @$ADDR"> {
1179 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$CNT),
1180 "VTX $CNT @$ADDR"> {
1183 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1184 "LOOP_START_DX10 @$ADDR"> {
1188 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1192 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1193 "LOOP_BREAK @$ADDR"> {
1197 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1198 "CONTINUE @$ADDR"> {
1202 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1203 "JUMP @$ADDR POP:$POP_COUNT"> {
1206 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1207 "ELSE @$ADDR POP:$POP_COUNT"> {
1210 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1215 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1216 "POP @$ADDR POP:$POP_COUNT"> {
1219 def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
1223 let END_OF_PROGRAM = 1;
1228 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1230 class COS_PAT <InstR600 trig> : Pat<
1232 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1235 class SIN_PAT <InstR600 trig> : Pat<
1237 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1240 //===----------------------------------------------------------------------===//
1241 // R700 Only instructions
1242 //===----------------------------------------------------------------------===//
1244 let Predicates = [isR700] in {
1245 def SIN_r700 : SIN_Common<0x6E>;
1246 def COS_r700 : COS_Common<0x6F>;
1248 // R700 normalizes inputs to SIN/COS the same as EG
1249 def : SIN_PAT <SIN_r700>;
1250 def : COS_PAT <COS_r700>;
1253 //===----------------------------------------------------------------------===//
1254 // Evergreen Only instructions
1255 //===----------------------------------------------------------------------===//
1257 let Predicates = [isEG] in {
1259 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1260 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1262 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1263 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1264 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1265 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1266 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1267 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1268 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1269 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1270 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1271 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1272 def SIN_eg : SIN_Common<0x8D>;
1273 def COS_eg : COS_Common<0x8E>;
1275 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
1276 def : SIN_PAT <SIN_eg>;
1277 def : COS_PAT <COS_eg>;
1278 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
1280 //===----------------------------------------------------------------------===//
1281 // Memory read/write instructions
1282 //===----------------------------------------------------------------------===//
1283 let usesCustomInserter = 1 in {
1285 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> mask, string name,
1287 : EG_CF_RAT <0x57, 0x2, mask, (outs), ins, name, pattern> {
1290 } // End usesCustomInserter = 1
1293 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1294 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1295 0x1, "RAT_WRITE_CACHELESS_32_eg $rw_gpr, $index_gpr, $eop",
1296 [(global_store i32:$rw_gpr, i32:$index_gpr)]
1300 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1301 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1302 0xf, "RAT_WRITE_CACHELESS_128 $rw_gpr.XYZW, $index_gpr, $eop",
1303 [(global_store v4i32:$rw_gpr, i32:$index_gpr)]
1306 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1307 : VTX_WORD0_eg, VTX_READ<name, buffer_id, outs, pattern> {
1312 let FETCH_WHOLE_QUAD = 0;
1313 let BUFFER_ID = buffer_id;
1315 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1316 // to store vertex addresses in any channel, not just X.
1319 let Inst{31-0} = Word0;
1322 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1323 : VTX_READ_eg <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id,
1324 (outs R600_TReg32_X:$dst_gpr), pattern> {
1326 let MEGA_FETCH_COUNT = 1;
1328 let DST_SEL_Y = 7; // Masked
1329 let DST_SEL_Z = 7; // Masked
1330 let DST_SEL_W = 7; // Masked
1331 let DATA_FORMAT = 1; // FMT_8
1334 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1335 : VTX_READ_eg <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id,
1336 (outs R600_TReg32_X:$dst_gpr), pattern> {
1337 let MEGA_FETCH_COUNT = 2;
1339 let DST_SEL_Y = 7; // Masked
1340 let DST_SEL_Z = 7; // Masked
1341 let DST_SEL_W = 7; // Masked
1342 let DATA_FORMAT = 5; // FMT_16
1346 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1347 : VTX_READ_eg <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id,
1348 (outs R600_TReg32_X:$dst_gpr), pattern> {
1350 let MEGA_FETCH_COUNT = 4;
1352 let DST_SEL_Y = 7; // Masked
1353 let DST_SEL_Z = 7; // Masked
1354 let DST_SEL_W = 7; // Masked
1355 let DATA_FORMAT = 0xD; // COLOR_32
1357 // This is not really necessary, but there were some GPU hangs that appeared
1358 // to be caused by ALU instructions in the next instruction group that wrote
1359 // to the $src_gpr registers of the VTX_READ.
1361 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1362 // %T2_X<def> = MOV %ZERO
1363 //Adding this constraint prevents this from happening.
1364 let Constraints = "$src_gpr.ptr = $dst_gpr";
1367 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1368 : VTX_READ_eg <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id,
1369 (outs R600_Reg128:$dst_gpr), pattern> {
1371 let MEGA_FETCH_COUNT = 16;
1376 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1378 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1379 // that holds its buffer address to avoid potential hangs. We can't use
1380 // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst
1381 // registers are different sizes.
1384 //===----------------------------------------------------------------------===//
1385 // VTX Read from parameter memory space
1386 //===----------------------------------------------------------------------===//
1388 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1389 [(set i32:$dst_gpr, (load_param_zexti8 ADDRVTX_READ:$src_gpr))]
1392 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1393 [(set i32:$dst_gpr, (load_param_zexti16 ADDRVTX_READ:$src_gpr))]
1396 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1397 [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1400 def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
1401 [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1404 //===----------------------------------------------------------------------===//
1405 // VTX Read from global memory space
1406 //===----------------------------------------------------------------------===//
1409 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1410 [(set i32:$dst_gpr, (zextloadi8_global ADDRVTX_READ:$src_gpr))]
1414 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1415 [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1419 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1420 [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1423 //===----------------------------------------------------------------------===//
1425 // XXX: We are currently storing all constants in the global address space.
1426 //===----------------------------------------------------------------------===//
1428 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1429 [(set i32:$dst_gpr, (constant_load ADDRVTX_READ:$src_gpr))]
1433 } // End Predicates = [isEG]
1435 //===----------------------------------------------------------------------===//
1436 // Evergreen / Cayman Instructions
1437 //===----------------------------------------------------------------------===//
1439 let Predicates = [isEGorCayman] in {
1441 // BFE_UINT - bit_extract, an optimization for mask and shift
1446 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1451 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1452 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1453 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1454 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1455 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1456 [(set i32:$dst, (int_AMDIL_bit_extract_u32 i32:$src0, i32:$src1,
1460 def : BFEPattern <BFE_UINT_eg>;
1462 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>;
1463 defm : BFIPatterns <BFI_INT_eg>;
1465 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
1466 def : ROTRPattern <BIT_ALIGN_INT_eg>;
1468 def MULADD_eg : MULADD_Common<0x14>;
1469 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
1470 def ASHR_eg : ASHR_Common<0x15>;
1471 def LSHR_eg : LSHR_Common<0x16>;
1472 def LSHL_eg : LSHL_Common<0x17>;
1473 def CNDE_eg : CNDE_Common<0x19>;
1474 def CNDGT_eg : CNDGT_Common<0x1A>;
1475 def CNDGE_eg : CNDGE_Common<0x1B>;
1476 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1477 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1478 def DOT4_eg : DOT4_Common<0xBE>;
1479 defm CUBE_eg : CUBE_Common<0xC0>;
1481 let hasSideEffects = 1 in {
1482 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1485 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1487 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1491 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1493 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1497 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1499 // TRUNC is used for the FLT_TO_INT instructions to work around a
1500 // perceived problem where the rounding modes are applied differently
1501 // depending on the instruction and the slot they are in.
1503 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1504 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1506 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1507 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1508 // We should look into handling these cases separately.
1509 def : Pat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
1511 def : Pat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
1514 def : SHA256MaPattern <BFI_INT_eg, XOR_INT>;
1516 def EG_ExportSwz : ExportSwzInst {
1517 let Word1{19-16} = 0; // BURST_COUNT
1518 let Word1{20} = 1; // VALID_PIXEL_MODE
1519 let Word1{21} = eop;
1520 let Word1{29-22} = inst;
1521 let Word1{30} = 0; // MARK
1522 let Word1{31} = 1; // BARRIER
1524 defm : ExportPattern<EG_ExportSwz, 83>;
1526 def EG_ExportBuf : ExportBufInst {
1527 let Word1{19-16} = 0; // BURST_COUNT
1528 let Word1{20} = 1; // VALID_PIXEL_MODE
1529 let Word1{21} = eop;
1530 let Word1{29-22} = inst;
1531 let Word1{30} = 0; // MARK
1532 let Word1{31} = 1; // BARRIER
1534 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1536 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1537 "TEX $COUNT @$ADDR"> {
1540 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1541 "VTX $COUNT @$ADDR"> {
1544 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
1545 "LOOP_START_DX10 @$ADDR"> {
1549 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1553 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
1554 "LOOP_BREAK @$ADDR"> {
1558 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
1559 "CONTINUE @$ADDR"> {
1563 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1564 "JUMP @$ADDR POP:$POP_COUNT"> {
1567 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1568 "ELSE @$ADDR POP:$POP_COUNT"> {
1571 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
1576 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1577 "POP @$ADDR POP:$POP_COUNT"> {
1580 def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
1584 let END_OF_PROGRAM = 1;
1587 } // End Predicates = [isEGorCayman]
1589 //===----------------------------------------------------------------------===//
1590 // Regist loads and stores - for indirect addressing
1591 //===----------------------------------------------------------------------===//
1593 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1595 //===----------------------------------------------------------------------===//
1596 // Cayman Instructions
1597 //===----------------------------------------------------------------------===//
1599 let Predicates = [isCayman] in {
1601 let isVector = 1 in {
1603 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1605 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1606 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1607 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1608 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1609 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1610 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1611 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
1612 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1613 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1614 def SIN_cm : SIN_Common<0x8D>;
1615 def COS_cm : COS_Common<0x8E>;
1616 } // End isVector = 1
1618 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
1619 def : SIN_PAT <SIN_cm>;
1620 def : COS_PAT <COS_cm>;
1622 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1624 // RECIP_UINT emulation for Cayman
1625 // The multiplication scales from [0,1] to the unsigned integer range
1627 (AMDGPUurecip i32:$src0),
1628 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
1629 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
1632 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
1638 def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
1641 def RAT_STORE_DWORD_cm : EG_CF_RAT <
1642 0x57, 0x14, 0x1, (outs),
1643 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr),
1644 "EXPORT_RAT_INST_STORE_DWORD $rw_gpr, $index_gpr",
1645 [(global_store i32:$rw_gpr, i32:$index_gpr)]
1647 let eop = 0; // This bit is not used on Cayman.
1650 class VTX_READ_cm <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1651 : VTX_WORD0_cm, VTX_READ<name, buffer_id, outs, pattern> {
1656 let FETCH_WHOLE_QUAD = 0;
1657 let BUFFER_ID = buffer_id;
1659 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1660 // to store vertex addresses in any channel, not just X.
1663 let STRUCTURED_READ = 0;
1665 let COALESCED_READ = 0;
1667 let Inst{31-0} = Word0;
1670 class VTX_READ_8_cm <bits<8> buffer_id, list<dag> pattern>
1671 : VTX_READ_cm <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id,
1672 (outs R600_TReg32_X:$dst_gpr), pattern> {
1675 let DST_SEL_Y = 7; // Masked
1676 let DST_SEL_Z = 7; // Masked
1677 let DST_SEL_W = 7; // Masked
1678 let DATA_FORMAT = 1; // FMT_8
1681 class VTX_READ_16_cm <bits<8> buffer_id, list<dag> pattern>
1682 : VTX_READ_cm <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id,
1683 (outs R600_TReg32_X:$dst_gpr), pattern> {
1685 let DST_SEL_Y = 7; // Masked
1686 let DST_SEL_Z = 7; // Masked
1687 let DST_SEL_W = 7; // Masked
1688 let DATA_FORMAT = 5; // FMT_16
1692 class VTX_READ_32_cm <bits<8> buffer_id, list<dag> pattern>
1693 : VTX_READ_cm <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id,
1694 (outs R600_TReg32_X:$dst_gpr), pattern> {
1697 let DST_SEL_Y = 7; // Masked
1698 let DST_SEL_Z = 7; // Masked
1699 let DST_SEL_W = 7; // Masked
1700 let DATA_FORMAT = 0xD; // COLOR_32
1702 // This is not really necessary, but there were some GPU hangs that appeared
1703 // to be caused by ALU instructions in the next instruction group that wrote
1704 // to the $src_gpr registers of the VTX_READ.
1706 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1707 // %T2_X<def> = MOV %ZERO
1708 //Adding this constraint prevents this from happening.
1709 let Constraints = "$src_gpr.ptr = $dst_gpr";
1712 class VTX_READ_128_cm <bits<8> buffer_id, list<dag> pattern>
1713 : VTX_READ_cm <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id,
1714 (outs R600_Reg128:$dst_gpr), pattern> {
1720 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1722 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1723 // that holds its buffer address to avoid potential hangs. We can't use
1724 // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst
1725 // registers are different sizes.
1728 //===----------------------------------------------------------------------===//
1729 // VTX Read from parameter memory space
1730 //===----------------------------------------------------------------------===//
1731 def VTX_READ_PARAM_8_cm : VTX_READ_8_cm <0,
1732 [(set i32:$dst_gpr, (load_param_zexti8 ADDRVTX_READ:$src_gpr))]
1735 def VTX_READ_PARAM_16_cm : VTX_READ_16_cm <0,
1736 [(set i32:$dst_gpr, (load_param_zexti16 ADDRVTX_READ:$src_gpr))]
1739 def VTX_READ_PARAM_32_cm : VTX_READ_32_cm <0,
1740 [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1743 def VTX_READ_PARAM_128_cm : VTX_READ_128_cm <0,
1744 [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
1747 //===----------------------------------------------------------------------===//
1748 // VTX Read from global memory space
1749 //===----------------------------------------------------------------------===//
1752 def VTX_READ_GLOBAL_8_cm : VTX_READ_8_cm <1,
1753 [(set i32:$dst_gpr, (zextloadi8_global ADDRVTX_READ:$src_gpr))]
1757 def VTX_READ_GLOBAL_32_cm : VTX_READ_32_cm <1,
1758 [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1762 def VTX_READ_GLOBAL_128_cm : VTX_READ_128_cm <1,
1763 [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
1766 //===----------------------------------------------------------------------===//
1768 // XXX: We are currently storing all constants in the global address space.
1769 //===----------------------------------------------------------------------===//
1771 def CONSTANT_LOAD_cm : VTX_READ_32_cm <1,
1772 [(set i32:$dst_gpr, (constant_load ADDRVTX_READ:$src_gpr))]
1777 //===----------------------------------------------------------------------===//
1778 // Branch Instructions
1779 //===----------------------------------------------------------------------===//
1782 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
1783 "IF_PREDICATE_SET $src", []>;
1785 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
1786 "PREDICATED_BREAK $src", []>;
1788 //===----------------------------------------------------------------------===//
1789 // Pseudo instructions
1790 //===----------------------------------------------------------------------===//
1792 let isPseudo = 1 in {
1794 def PRED_X : InstR600 <
1795 (outs R600_Predicate_Bit:$dst),
1796 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
1798 let FlagOperandIdx = 3;
1801 let isTerminator = 1, isBranch = 1 in {
1802 def JUMP_COND : InstR600 <
1804 (ins brtarget:$target, R600_Predicate_Bit:$p),
1805 "JUMP $target ($p)",
1809 def JUMP : InstR600 <
1811 (ins brtarget:$target),
1816 let isPredicable = 1;
1820 } // End isTerminator = 1, isBranch = 1
1822 let usesCustomInserter = 1 in {
1824 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
1826 def MASK_WRITE : AMDGPUShaderInst <
1828 (ins R600_Reg32:$src),
1833 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
1837 (outs R600_Reg128:$dst),
1838 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
1839 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1840 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1841 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
1842 imm:$resourceId, imm:$samplerId, imm:$textureTarget))],
1847 def TXD_SHADOW: InstR600 <
1848 (outs R600_Reg128:$dst),
1849 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
1850 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1851 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1852 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
1853 imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))],
1858 } // End isPseudo = 1
1859 } // End usesCustomInserter = 1
1861 def CLAMP_R600 : CLAMP <R600_Reg32>;
1862 def FABS_R600 : FABS<R600_Reg32>;
1863 def FNEG_R600 : FNEG<R600_Reg32>;
1865 //===---------------------------------------------------------------------===//
1866 // Return instruction
1867 //===---------------------------------------------------------------------===//
1868 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
1869 usesCustomInserter = 1 in {
1870 def RETURN : ILFormat<(outs), (ins variable_ops),
1871 "RETURN", [(IL_retflag)]>;
1875 //===----------------------------------------------------------------------===//
1876 // Constant Buffer Addressing Support
1877 //===----------------------------------------------------------------------===//
1879 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
1880 def CONST_COPY : Instruction {
1881 let OutOperandList = (outs R600_Reg32:$dst);
1882 let InOperandList = (ins i32imm:$src);
1884 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
1885 let AsmString = "CONST_COPY";
1886 let neverHasSideEffects = 1;
1887 let isAsCheapAsAMove = 1;
1888 let Itinerary = NullALU;
1890 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
1892 def TEX_VTX_CONSTBUF :
1893 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
1894 [(set v4i32:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
1895 VTX_WORD1_GPR, VTX_WORD0_eg {
1899 let FETCH_WHOLE_QUAD = 0;
1903 let USE_CONST_FIELDS = 0;
1904 let NUM_FORMAT_ALL = 2;
1905 let FORMAT_COMP_ALL = 1;
1906 let SRF_MODE_ALL = 1;
1907 let MEGA_FETCH_COUNT = 16;
1912 let DATA_FORMAT = 35;
1914 let Inst{31-0} = Word0;
1915 let Inst{63-32} = Word1;
1917 // LLVM can only encode 64-bit instructions, so these fields are manually
1918 // encoded in R600CodeEmitter
1921 // bits<2> ENDIAN_SWAP = 0;
1922 // bits<1> CONST_BUF_NO_STRIDE = 0;
1923 // bits<1> MEGA_FETCH = 0;
1924 // bits<1> ALT_CONST = 0;
1925 // bits<2> BUFFER_INDEX_MODE = 0;
1929 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1930 // is done in R600CodeEmitter
1932 // Inst{79-64} = OFFSET;
1933 // Inst{81-80} = ENDIAN_SWAP;
1934 // Inst{82} = CONST_BUF_NO_STRIDE;
1935 // Inst{83} = MEGA_FETCH;
1936 // Inst{84} = ALT_CONST;
1937 // Inst{86-85} = BUFFER_INDEX_MODE;
1938 // Inst{95-86} = 0; Reserved
1940 // VTX_WORD3 (Padding)
1942 // Inst{127-96} = 0;
1947 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
1948 [(set v4f32:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
1949 VTX_WORD1_GPR, VTX_WORD0_eg {
1953 let FETCH_WHOLE_QUAD = 0;
1957 let USE_CONST_FIELDS = 1;
1958 let NUM_FORMAT_ALL = 0;
1959 let FORMAT_COMP_ALL = 0;
1960 let SRF_MODE_ALL = 1;
1961 let MEGA_FETCH_COUNT = 16;
1966 let DATA_FORMAT = 0;
1968 let Inst{31-0} = Word0;
1969 let Inst{63-32} = Word1;
1971 // LLVM can only encode 64-bit instructions, so these fields are manually
1972 // encoded in R600CodeEmitter
1975 // bits<2> ENDIAN_SWAP = 0;
1976 // bits<1> CONST_BUF_NO_STRIDE = 0;
1977 // bits<1> MEGA_FETCH = 0;
1978 // bits<1> ALT_CONST = 0;
1979 // bits<2> BUFFER_INDEX_MODE = 0;
1983 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1984 // is done in R600CodeEmitter
1986 // Inst{79-64} = OFFSET;
1987 // Inst{81-80} = ENDIAN_SWAP;
1988 // Inst{82} = CONST_BUF_NO_STRIDE;
1989 // Inst{83} = MEGA_FETCH;
1990 // Inst{84} = ALT_CONST;
1991 // Inst{86-85} = BUFFER_INDEX_MODE;
1992 // Inst{95-86} = 0; Reserved
1994 // VTX_WORD3 (Padding)
1996 // Inst{127-96} = 0;
2002 //===--------------------------------------------------------------------===//
2003 // Instructions support
2004 //===--------------------------------------------------------------------===//
2005 //===---------------------------------------------------------------------===//
2006 // Custom Inserter for Branches and returns, this eventually will be a
2008 //===---------------------------------------------------------------------===//
2009 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
2010 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
2011 "; Pseudo unconditional branch instruction",
2013 defm BRANCH_COND : BranchConditional<IL_brcond>;
2016 //===---------------------------------------------------------------------===//
2017 // Flow and Program control Instructions
2018 //===---------------------------------------------------------------------===//
2019 let isTerminator=1 in {
2020 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
2021 !strconcat("SWITCH", " $src"), []>;
2022 def CASE : ILFormat< (outs), (ins GPRI32:$src),
2023 !strconcat("CASE", " $src"), []>;
2024 def BREAK : ILFormat< (outs), (ins),
2026 def CONTINUE : ILFormat< (outs), (ins),
2028 def DEFAULT : ILFormat< (outs), (ins),
2030 def ELSE : ILFormat< (outs), (ins),
2032 def ENDSWITCH : ILFormat< (outs), (ins),
2034 def ENDMAIN : ILFormat< (outs), (ins),
2036 def END : ILFormat< (outs), (ins),
2038 def ENDFUNC : ILFormat< (outs), (ins),
2040 def ENDIF : ILFormat< (outs), (ins),
2042 def WHILELOOP : ILFormat< (outs), (ins),
2044 def ENDLOOP : ILFormat< (outs), (ins),
2046 def FUNC : ILFormat< (outs), (ins),
2048 def RETDYN : ILFormat< (outs), (ins),
2050 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2051 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
2052 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2053 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
2054 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2055 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
2056 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2057 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
2058 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2059 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
2060 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
2061 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
2062 defm IFC : BranchInstr2<"IFC">;
2063 defm BREAKC : BranchInstr2<"BREAKC">;
2064 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
2067 //===----------------------------------------------------------------------===//
2069 //===----------------------------------------------------------------------===//
2071 // CND*_INT Pattterns for f32 True / False values
2073 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
2074 (selectcc i32:$src0, 0, f32:$src1, f32:$src2, cc),
2075 (cnd $src0, $src1, $src2)
2078 def : CND_INT_f32 <CNDE_INT, SETEQ>;
2079 def : CND_INT_f32 <CNDGT_INT, SETGT>;
2080 def : CND_INT_f32 <CNDGE_INT, SETGE>;
2082 //CNDGE_INT extra pattern
2084 (selectcc i32:$src0, -1, i32:$src1, i32:$src2, COND_GT),
2085 (CNDGE_INT $src0, $src1, $src2)
2091 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
2095 (int_AMDGPU_kill f32:$src0),
2096 (MASK_WRITE (KILLGT (f32 ZERO), $src0))
2101 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LT),
2107 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LE),
2111 // SETGT_DX10 reverse args
2113 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LT),
2114 (SETGT_DX10 $src1, $src0)
2117 // SETGE_DX10 reverse args
2119 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LE),
2120 (SETGE_DX10 $src1, $src0)
2123 // SETGT_INT reverse args
2125 (selectcc i32:$src0, i32:$src1, -1, 0, SETLT),
2126 (SETGT_INT $src1, $src0)
2129 // SETGE_INT reverse args
2131 (selectcc i32:$src0, i32:$src1, -1, 0, SETLE),
2132 (SETGE_INT $src1, $src0)
2135 // SETGT_UINT reverse args
2137 (selectcc i32:$src0, i32:$src1, -1, 0, SETULT),
2138 (SETGT_UINT $src1, $src0)
2141 // SETGE_UINT reverse args
2143 (selectcc i32:$src0, i32:$src1, -1, 0, SETULE),
2144 (SETGE_UINT $src1, $src0)
2147 // The next two patterns are special cases for handling 'true if ordered' and
2148 // 'true if unordered' conditionals. The assumption here is that the behavior of
2149 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
2151 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
2152 // We assume that SETE returns false when one of the operands is NAN and
2153 // SNE returns true when on of the operands is NAN
2155 //SETE - 'true if ordered'
2157 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETO),
2161 //SETE_DX10 - 'true if ordered'
2163 (selectcc f32:$src0, f32:$src1, -1, 0, SETO),
2164 (SETE_DX10 $src0, $src1)
2167 //SNE - 'true if unordered'
2169 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETUO),
2173 //SETNE_DX10 - 'true if ordered'
2175 (selectcc f32:$src0, f32:$src1, -1, 0, SETUO),
2176 (SETNE_DX10 $src0, $src1)
2179 def : Extract_Element <f32, v4f32, 0, sub0>;
2180 def : Extract_Element <f32, v4f32, 1, sub1>;
2181 def : Extract_Element <f32, v4f32, 2, sub2>;
2182 def : Extract_Element <f32, v4f32, 3, sub3>;
2184 def : Insert_Element <f32, v4f32, 0, sub0>;
2185 def : Insert_Element <f32, v4f32, 1, sub1>;
2186 def : Insert_Element <f32, v4f32, 2, sub2>;
2187 def : Insert_Element <f32, v4f32, 3, sub3>;
2189 def : Extract_Element <i32, v4i32, 0, sub0>;
2190 def : Extract_Element <i32, v4i32, 1, sub1>;
2191 def : Extract_Element <i32, v4i32, 2, sub2>;
2192 def : Extract_Element <i32, v4i32, 3, sub3>;
2194 def : Insert_Element <i32, v4i32, 0, sub0>;
2195 def : Insert_Element <i32, v4i32, 1, sub1>;
2196 def : Insert_Element <i32, v4i32, 2, sub2>;
2197 def : Insert_Element <i32, v4i32, 3, sub3>;
2199 def : Vector4_Build <v4f32, f32>;
2200 def : Vector4_Build <v4i32, i32>;
2202 // bitconvert patterns
2204 def : BitConvert <i32, f32, R600_Reg32>;
2205 def : BitConvert <f32, i32, R600_Reg32>;
2206 def : BitConvert <v4f32, v4i32, R600_Reg128>;
2207 def : BitConvert <v4i32, v4f32, R600_Reg128>;
2209 // DWORDADDR pattern
2210 def : DwordAddrPat <i32, R600_Reg32>;
2212 } // End isR600toCayman Predicate