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";
119 let Inst{31-0} = Word0;
120 let Inst{63-32} = Word1;
123 class R600_1OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
124 InstrItinClass itin = AnyALU> :
125 R600_1OP <inst, opName,
126 [(set R600_Reg32:$dst, (node R600_Reg32:$src0))]
129 // If you add our change the operands for R600_2OP instructions, you must
130 // also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
131 // R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
132 class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
133 InstrItinClass itin = AnyALU> :
134 InstR600 <(outs R600_Reg32:$dst),
135 (ins UEM:$update_exec_mask, UP:$update_pred, WRITE:$write,
136 OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
137 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
138 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs, SEL:$src1_sel,
139 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
140 BANK_SWIZZLE:$bank_swizzle),
141 !strconcat(" ", opName,
142 "$clamp $last $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
143 "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
144 "$src1_neg$src1_abs$src1$src1_abs$src1_rel, "
145 "$pred_sel $bank_swizzle"),
149 R600ALU_Word1_OP2 <inst> {
151 let HasNativeOperands = 1;
153 let DisableEncoding = "$literal";
155 let Inst{31-0} = Word0;
156 let Inst{63-32} = Word1;
159 class R600_2OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
160 InstrItinClass itim = AnyALU> :
161 R600_2OP <inst, opName,
162 [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
166 // If you add our change the operands for R600_3OP instructions, you must
167 // also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
168 // R600InstrInfo::buildDefaultInstruction(), and
169 // R600InstrInfo::getOperandIdx().
170 class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
171 InstrItinClass itin = AnyALU> :
172 InstR600 <(outs R600_Reg32:$dst),
173 (ins REL:$dst_rel, CLAMP:$clamp,
174 R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, SEL:$src0_sel,
175 R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
176 R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
177 LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal,
178 BANK_SWIZZLE:$bank_swizzle),
179 !strconcat(" ", opName, "$clamp $last $dst$dst_rel, "
180 "$src0_neg$src0$src0_rel, "
181 "$src1_neg$src1$src1_rel, "
182 "$src2_neg$src2$src2_rel, "
188 R600ALU_Word1_OP3<inst>{
190 let HasNativeOperands = 1;
191 let DisableEncoding = "$literal";
194 let Inst{31-0} = Word0;
195 let Inst{63-32} = Word1;
198 class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
199 InstrItinClass itin = VecALU> :
200 InstR600 <(outs R600_Reg32:$dst),
208 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
210 def TEX_SHADOW : PatLeaf<
212 [{uint32_t TType = (uint32_t)N->getZExtValue();
213 return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
217 def TEX_RECT : PatLeaf<
219 [{uint32_t TType = (uint32_t)N->getZExtValue();
224 def TEX_ARRAY : PatLeaf<
226 [{uint32_t TType = (uint32_t)N->getZExtValue();
227 return TType == 9 || TType == 10 || TType == 15 || TType == 16;
231 def TEX_SHADOW_ARRAY : PatLeaf<
233 [{uint32_t TType = (uint32_t)N->getZExtValue();
234 return TType == 11 || TType == 12 || TType == 17;
238 class EG_CF_RAT <bits <8> cfinst, bits <6> ratinst, bits<4> ratid, dag outs,
239 dag ins, string asm, list<dag> pattern> :
240 InstR600ISA <outs, ins, asm, pattern>,
241 CF_ALLOC_EXPORT_WORD0_RAT, CF_ALLOC_EXPORT_WORD1_BUF {
243 let cf_inst = cfinst;
244 let rat_inst = ratinst;
247 let Inst{31-0} = Word0;
248 let Inst{63-32} = Word1;
252 class LoadParamFrag <PatFrag load_type> : PatFrag <
253 (ops node:$ptr), (load_type node:$ptr),
254 [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
257 def load_param : LoadParamFrag<load>;
258 def load_param_zexti8 : LoadParamFrag<zextloadi8>;
259 def load_param_zexti16 : LoadParamFrag<zextloadi16>;
261 def isR600 : Predicate<"Subtarget.getGeneration() <= AMDGPUSubtarget::R700">;
262 def isR700 : Predicate<"Subtarget.getGeneration() == AMDGPUSubtarget::R700">;
263 def isEG : Predicate<
264 "Subtarget.getGeneration() >= AMDGPUSubtarget::EVERGREEN && "
265 "Subtarget.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS && "
266 "!Subtarget.hasCaymanISA()">;
268 def isCayman : Predicate<"Subtarget.hasCaymanISA()">;
269 def isEGorCayman : Predicate<"Subtarget.getGeneration() == "
270 "AMDGPUSubtarget::EVERGREEN"
271 "|| Subtarget.getGeneration() =="
272 "AMDGPUSubtarget::NORTHERN_ISLANDS">;
274 def isR600toCayman : Predicate<
275 "Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS">;
277 //===----------------------------------------------------------------------===//
279 //===----------------------------------------------------------------------===//
281 def INTERP_PAIR_XY : AMDGPUShaderInst <
282 (outs R600_TReg32_X:$dst0, R600_TReg32_Y:$dst1),
283 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
284 "INTERP_PAIR_XY $src0 $src1 $src2 : $dst0 dst1",
287 def INTERP_PAIR_ZW : AMDGPUShaderInst <
288 (outs R600_TReg32_Z:$dst0, R600_TReg32_W:$dst1),
289 (ins i32imm:$src0, R600_TReg32_Y:$src1, R600_TReg32_X:$src2),
290 "INTERP_PAIR_ZW $src0 $src1 $src2 : $dst0 dst1",
293 def CONST_ADDRESS: SDNode<"AMDGPUISD::CONST_ADDRESS",
294 SDTypeProfile<1, -1, [SDTCisInt<0>, SDTCisPtrTy<1>]>,
298 def DOT4 : SDNode<"AMDGPUISD::DOT4",
299 SDTypeProfile<1, 8, [SDTCisFP<0>, SDTCisVT<1, f32>, SDTCisVT<2, f32>,
300 SDTCisVT<3, f32>, SDTCisVT<4, f32>, SDTCisVT<5, f32>,
301 SDTCisVT<6, f32>, SDTCisVT<7, f32>, SDTCisVT<8, f32>]>,
305 def TEXTURE_FETCH_Type : SDTypeProfile<1, 19, [SDTCisFP<0>]>;
307 def TEXTURE_FETCH: SDNode<"AMDGPUISD::TEXTURE_FETCH", TEXTURE_FETCH_Type, []>;
309 multiclass TexPattern<bits<32> TextureOp, Instruction inst, ValueType vt = v4f32> {
310 def : Pat<(TEXTURE_FETCH (i32 TextureOp), vt:$SRC_GPR,
311 (i32 imm:$srcx), (i32 imm:$srcy), (i32 imm:$srcz), (i32 imm:$srcw),
312 (i32 imm:$offsetx), (i32 imm:$offsety), (i32 imm:$offsetz),
313 (i32 imm:$DST_SEL_X), (i32 imm:$DST_SEL_Y), (i32 imm:$DST_SEL_Z),
314 (i32 imm:$DST_SEL_W),
315 (i32 imm:$RESOURCE_ID), (i32 imm:$SAMPLER_ID),
316 (i32 imm:$COORD_TYPE_X), (i32 imm:$COORD_TYPE_Y), (i32 imm:$COORD_TYPE_Z),
317 (i32 imm:$COORD_TYPE_W)),
318 (inst R600_Reg128:$SRC_GPR,
319 imm:$srcx, imm:$srcy, imm:$srcz, imm:$srcw,
320 imm:$offsetx, imm:$offsety, imm:$offsetz,
321 imm:$DST_SEL_X, imm:$DST_SEL_Y, imm:$DST_SEL_Z,
323 imm:$RESOURCE_ID, imm:$SAMPLER_ID,
324 imm:$COORD_TYPE_X, imm:$COORD_TYPE_Y, imm:$COORD_TYPE_Z,
328 //===----------------------------------------------------------------------===//
329 // Interpolation Instructions
330 //===----------------------------------------------------------------------===//
332 def INTERP_VEC_LOAD : AMDGPUShaderInst <
333 (outs R600_Reg128:$dst),
335 "INTERP_LOAD $src0 : $dst",
338 def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
339 let bank_swizzle = 5;
342 def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
343 let bank_swizzle = 5;
346 def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
348 //===----------------------------------------------------------------------===//
349 // Export Instructions
350 //===----------------------------------------------------------------------===//
352 def ExportType : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
354 def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
355 [SDNPHasChain, SDNPSideEffect]>;
358 field bits<32> Word0;
365 let Word0{12-0} = arraybase;
366 let Word0{14-13} = type;
367 let Word0{21-15} = gpr;
368 let Word0{22} = 0; // RW_REL
369 let Word0{29-23} = 0; // INDEX_GPR
370 let Word0{31-30} = elem_size;
373 class ExportSwzWord1 {
374 field bits<32> Word1;
383 let Word1{2-0} = sw_x;
384 let Word1{5-3} = sw_y;
385 let Word1{8-6} = sw_z;
386 let Word1{11-9} = sw_w;
389 class ExportBufWord1 {
390 field bits<32> Word1;
397 let Word1{11-0} = arraySize;
398 let Word1{15-12} = compMask;
401 multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
402 def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
404 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
405 0, 61, 0, 7, 7, 7, cf_inst, 0)
408 def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
410 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sub0),
411 0, 61, 7, 0, 7, 7, cf_inst, 0)
414 def : Pat<(int_R600_store_dummy (i32 imm:$type)),
416 (v4f32 (IMPLICIT_DEF)), imm:$type, 0, 7, 7, 7, 7, cf_inst, 0)
419 def : Pat<(int_R600_store_dummy 1),
421 (v4f32 (IMPLICIT_DEF)), 1, 60, 7, 7, 7, 7, cf_inst, 0)
424 def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 imm:$base), (i32 imm:$type),
425 (i32 imm:$swz_x), (i32 imm:$swz_y), (i32 imm:$swz_z), (i32 imm:$swz_w)),
426 (ExportInst R600_Reg128:$src, imm:$type, imm:$base,
427 imm:$swz_x, imm:$swz_y, imm:$swz_z, imm:$swz_w, cf_inst, 0)
432 multiclass SteamOutputExportPattern<Instruction ExportInst,
433 bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
435 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
436 (i32 imm:$arraybase), (i32 0), (i32 imm:$mask)),
437 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
438 4095, imm:$mask, buf0inst, 0)>;
440 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
441 (i32 imm:$arraybase), (i32 1), (i32 imm:$mask)),
442 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
443 4095, imm:$mask, buf1inst, 0)>;
445 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
446 (i32 imm:$arraybase), (i32 2), (i32 imm:$mask)),
447 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
448 4095, imm:$mask, buf2inst, 0)>;
450 def : Pat<(int_R600_store_stream_output (v4f32 R600_Reg128:$src),
451 (i32 imm:$arraybase), (i32 3), (i32 imm:$mask)),
452 (ExportInst R600_Reg128:$src, 0, imm:$arraybase,
453 4095, imm:$mask, buf3inst, 0)>;
456 // Export Instructions should not be duplicated by TailDuplication pass
457 // (which assumes that duplicable instruction are affected by exec mask)
458 let usesCustomInserter = 1, isNotDuplicable = 1 in {
460 class ExportSwzInst : InstR600ISA<(
462 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
463 i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
465 !strconcat("EXPORT", " $gpr"),
466 []>, ExportWord0, ExportSwzWord1 {
468 let Inst{31-0} = Word0;
469 let Inst{63-32} = Word1;
472 } // End usesCustomInserter = 1
474 class ExportBufInst : InstR600ISA<(
476 (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
477 i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
478 !strconcat("EXPORT", " $gpr"),
479 []>, ExportWord0, ExportBufWord1 {
481 let Inst{31-0} = Word0;
482 let Inst{63-32} = Word1;
485 //===----------------------------------------------------------------------===//
486 // Control Flow Instructions
487 //===----------------------------------------------------------------------===//
490 def KCACHE : InstFlag<"printKCache">;
492 class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
493 (ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1,
494 KCACHE:$KCACHE_MODE0, KCACHE:$KCACHE_MODE1,
495 i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1,
497 !strconcat(OpName, " $COUNT, @$ADDR, "
498 "KC0[$KCACHE_MODE0], KC1[$KCACHE_MODE1]"),
499 [] >, CF_ALU_WORD0, CF_ALU_WORD1 {
504 let WHOLE_QUAD_MODE = 0;
507 let Inst{31-0} = Word0;
508 let Inst{63-32} = Word1;
511 class CF_WORD0_R600 {
512 field bits<32> Word0;
519 class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
520 ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
526 let VALID_PIXEL_MODE = 0;
530 let END_OF_PROGRAM = 0;
531 let WHOLE_QUAD_MODE = 0;
533 let Inst{31-0} = Word0;
534 let Inst{63-32} = Word1;
537 class CF_CLAUSE_EG <bits<8> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
538 ins, AsmPrint, [] >, CF_WORD0_EG, CF_WORD1_EG {
543 let JUMPTABLE_SEL = 0;
545 let VALID_PIXEL_MODE = 0;
547 let END_OF_PROGRAM = 0;
549 let Inst{31-0} = Word0;
550 let Inst{63-32} = Word1;
553 def CF_ALU : ALU_CLAUSE<8, "ALU">;
554 def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
556 def FETCH_CLAUSE : AMDGPUInst <(outs),
557 (ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
563 def ALU_CLAUSE : AMDGPUInst <(outs),
564 (ins i32imm:$addr), "ALU clause starting at $addr:", [] > {
570 def LITERALS : AMDGPUInst <(outs),
571 (ins LITERAL:$literal1, LITERAL:$literal2), "$literal1, $literal2", [] > {
576 let Inst{31-0} = literal1;
577 let Inst{63-32} = literal2;
580 def PAD : AMDGPUInst <(outs), (ins), "PAD", [] > {
584 let Predicates = [isR600toCayman] in {
586 //===----------------------------------------------------------------------===//
587 // Common Instructions R600, R700, Evergreen, Cayman
588 //===----------------------------------------------------------------------===//
590 def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
591 // Non-IEEE MUL: 0 * anything = 0
592 def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
593 def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
594 def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
595 def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
597 // For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
598 // so some of the instruction names don't match the asm string.
599 // XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
600 def SETE : R600_2OP <
602 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_EQ))]
607 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GT))]
612 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GE))]
617 [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_NE))]
620 def SETE_DX10 : R600_2OP <
622 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_EQ))]
625 def SETGT_DX10 : R600_2OP <
627 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GT))]
630 def SETGE_DX10 : R600_2OP <
632 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GE))]
635 def SETNE_DX10 : R600_2OP <
637 [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_NE))]
640 def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
641 def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
642 def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
643 def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
644 def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
646 def MOV : R600_1OP <0x19, "MOV", []>;
648 let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
650 class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
651 (outs R600_Reg32:$dst),
657 } // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
659 def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
662 (MOV_IMM_I32 imm:$val)
665 def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
668 (MOV_IMM_F32 fpimm:$val)
671 def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
672 def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
673 def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
674 def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
676 let hasSideEffects = 1 in {
678 def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
680 } // end hasSideEffects
682 def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
683 def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
684 def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
685 def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
686 def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
687 def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
688 def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
689 def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
690 def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUumax>;
691 def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
693 def SETE_INT : R600_2OP <
695 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETEQ))]
698 def SETGT_INT : R600_2OP <
700 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGT))]
703 def SETGE_INT : R600_2OP <
705 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETGE))]
708 def SETNE_INT : R600_2OP <
710 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETNE))]
713 def SETGT_UINT : R600_2OP <
715 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGT))]
718 def SETGE_UINT : R600_2OP <
720 [(set i32:$dst, (selectcc i32:$src0, i32:$src1, -1, 0, SETUGE))]
723 def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
724 def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
725 def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
726 def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
728 def CNDE_INT : R600_3OP <
730 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_EQ))]
733 def CNDGE_INT : R600_3OP <
735 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GE))]
738 def CNDGT_INT : R600_3OP <
740 [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GT))]
743 //===----------------------------------------------------------------------===//
744 // Texture instructions
745 //===----------------------------------------------------------------------===//
747 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
749 class R600_TEX <bits<11> inst, string opName> :
750 InstR600 <(outs R600_Reg128:$DST_GPR),
751 (ins R600_Reg128:$SRC_GPR,
752 RSel:$srcx, RSel:$srcy, RSel:$srcz, RSel:$srcw,
753 i32imm:$offsetx, i32imm:$offsety, i32imm:$offsetz,
754 RSel:$DST_SEL_X, RSel:$DST_SEL_Y, RSel:$DST_SEL_Z, RSel:$DST_SEL_W,
755 i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
756 CT:$COORD_TYPE_X, CT:$COORD_TYPE_Y, CT:$COORD_TYPE_Z,
759 " $DST_GPR.$DST_SEL_X$DST_SEL_Y$DST_SEL_Z$DST_SEL_W, "
760 "$SRC_GPR.$srcx$srcy$srcz$srcw "
761 "RID:$RESOURCE_ID SID:$SAMPLER_ID "
762 "CT:$COORD_TYPE_X$COORD_TYPE_Y$COORD_TYPE_Z$COORD_TYPE_W"),
764 NullALU>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
765 let Inst{31-0} = Word0;
766 let Inst{63-32} = Word1;
768 let TEX_INST = inst{4-0};
774 let FETCH_WHOLE_QUAD = 0;
776 let SAMPLER_INDEX_MODE = 0;
777 let RESOURCE_INDEX_MODE = 0;
782 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
786 def TEX_SAMPLE : R600_TEX <0x10, "TEX_SAMPLE">;
787 def TEX_SAMPLE_C : R600_TEX <0x18, "TEX_SAMPLE_C">;
788 def TEX_SAMPLE_L : R600_TEX <0x11, "TEX_SAMPLE_L">;
789 def TEX_SAMPLE_C_L : R600_TEX <0x19, "TEX_SAMPLE_C_L">;
790 def TEX_SAMPLE_LB : R600_TEX <0x12, "TEX_SAMPLE_LB">;
791 def TEX_SAMPLE_C_LB : R600_TEX <0x1A, "TEX_SAMPLE_C_LB">;
792 def TEX_LD : R600_TEX <0x03, "TEX_LD">;
793 def TEX_GET_TEXTURE_RESINFO : R600_TEX <0x04, "TEX_GET_TEXTURE_RESINFO">;
794 def TEX_GET_GRADIENTS_H : R600_TEX <0x07, "TEX_GET_GRADIENTS_H">;
795 def TEX_GET_GRADIENTS_V : R600_TEX <0x08, "TEX_GET_GRADIENTS_V">;
796 def TEX_SET_GRADIENTS_H : R600_TEX <0x0B, "TEX_SET_GRADIENTS_H">;
797 def TEX_SET_GRADIENTS_V : R600_TEX <0x0C, "TEX_SET_GRADIENTS_V">;
798 def TEX_SAMPLE_G : R600_TEX <0x14, "TEX_SAMPLE_G">;
799 def TEX_SAMPLE_C_G : R600_TEX <0x1C, "TEX_SAMPLE_C_G">;
801 defm : TexPattern<0, TEX_SAMPLE>;
802 defm : TexPattern<1, TEX_SAMPLE_C>;
803 defm : TexPattern<2, TEX_SAMPLE_L>;
804 defm : TexPattern<3, TEX_SAMPLE_C_L>;
805 defm : TexPattern<4, TEX_SAMPLE_LB>;
806 defm : TexPattern<5, TEX_SAMPLE_C_LB>;
807 defm : TexPattern<6, TEX_LD, v4i32>;
808 defm : TexPattern<7, TEX_GET_TEXTURE_RESINFO, v4i32>;
809 defm : TexPattern<8, TEX_GET_GRADIENTS_H>;
810 defm : TexPattern<9, TEX_GET_GRADIENTS_V>;
812 //===----------------------------------------------------------------------===//
813 // Helper classes for common instructions
814 //===----------------------------------------------------------------------===//
816 class MUL_LIT_Common <bits<5> inst> : R600_3OP <
821 class MULADD_Common <bits<5> inst> : R600_3OP <
826 class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
828 [(set f32:$dst, (fadd (fmul f32:$src0, f32:$src1), f32:$src2))]
831 class CNDE_Common <bits<5> inst> : R600_3OP <
833 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_EQ))]
836 class CNDGT_Common <bits<5> inst> : R600_3OP <
838 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GT))]
841 class CNDGE_Common <bits<5> inst> : R600_3OP <
843 [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GE))]
847 let isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
848 class R600_VEC2OP<list<dag> pattern> : InstR600 <(outs R600_Reg32:$dst), (ins
850 UEM:$update_exec_mask_X, UP:$update_pred_X, WRITE:$write_X,
851 OMOD:$omod_X, REL:$dst_rel_X, CLAMP:$clamp_X,
852 R600_TReg32_X:$src0_X, NEG:$src0_neg_X, REL:$src0_rel_X, ABS:$src0_abs_X, SEL:$src0_sel_X,
853 R600_TReg32_X:$src1_X, NEG:$src1_neg_X, REL:$src1_rel_X, ABS:$src1_abs_X, SEL:$src1_sel_X,
854 R600_Pred:$pred_sel_X,
856 UEM:$update_exec_mask_Y, UP:$update_pred_Y, WRITE:$write_Y,
857 OMOD:$omod_Y, REL:$dst_rel_Y, CLAMP:$clamp_Y,
858 R600_TReg32_Y:$src0_Y, NEG:$src0_neg_Y, REL:$src0_rel_Y, ABS:$src0_abs_Y, SEL:$src0_sel_Y,
859 R600_TReg32_Y:$src1_Y, NEG:$src1_neg_Y, REL:$src1_rel_Y, ABS:$src1_abs_Y, SEL:$src1_sel_Y,
860 R600_Pred:$pred_sel_Y,
862 UEM:$update_exec_mask_Z, UP:$update_pred_Z, WRITE:$write_Z,
863 OMOD:$omod_Z, REL:$dst_rel_Z, CLAMP:$clamp_Z,
864 R600_TReg32_Z:$src0_Z, NEG:$src0_neg_Z, REL:$src0_rel_Z, ABS:$src0_abs_Z, SEL:$src0_sel_Z,
865 R600_TReg32_Z:$src1_Z, NEG:$src1_neg_Z, REL:$src1_rel_Z, ABS:$src1_abs_Z, SEL:$src1_sel_Z,
866 R600_Pred:$pred_sel_Z,
868 UEM:$update_exec_mask_W, UP:$update_pred_W, WRITE:$write_W,
869 OMOD:$omod_W, REL:$dst_rel_W, CLAMP:$clamp_W,
870 R600_TReg32_W:$src0_W, NEG:$src0_neg_W, REL:$src0_rel_W, ABS:$src0_abs_W, SEL:$src0_sel_W,
871 R600_TReg32_W:$src1_W, NEG:$src1_neg_W, REL:$src1_rel_W, ABS:$src1_abs_W, SEL:$src1_sel_W,
872 R600_Pred:$pred_sel_W,
873 LITERAL:$literal0, LITERAL:$literal1),
879 def DOT_4 : R600_VEC2OP<[(set R600_Reg32:$dst, (DOT4
880 R600_TReg32_X:$src0_X, R600_TReg32_X:$src1_X,
881 R600_TReg32_Y:$src0_Y, R600_TReg32_Y:$src1_Y,
882 R600_TReg32_Z:$src0_Z, R600_TReg32_Z:$src1_Z,
883 R600_TReg32_W:$src0_W, R600_TReg32_W:$src1_W))]>;
886 class DOT4_Common <bits<11> inst> : R600_2OP <inst, "DOT4", []>;
889 let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
890 multiclass CUBE_Common <bits<11> inst> {
892 def _pseudo : InstR600 <
893 (outs R600_Reg128:$dst),
894 (ins R600_Reg128:$src),
896 [(set v4f32:$dst, (int_AMDGPU_cube v4f32:$src))],
902 def _real : R600_2OP <inst, "CUBE", []>;
904 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
906 class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
907 inst, "EXP_IEEE", fexp2
910 let Itinerary = TransALU;
913 class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
914 inst, "FLT_TO_INT", fp_to_sint
917 let Itinerary = TransALU;
920 class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
921 inst, "INT_TO_FLT", sint_to_fp
924 let Itinerary = TransALU;
927 class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
928 inst, "FLT_TO_UINT", fp_to_uint
931 let Itinerary = TransALU;
934 class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
935 inst, "UINT_TO_FLT", uint_to_fp
938 let Itinerary = TransALU;
941 class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
942 inst, "LOG_CLAMPED", []
945 class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
946 inst, "LOG_IEEE", flog2
949 let Itinerary = TransALU;
952 class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
953 class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
954 class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
955 class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
956 inst, "MULHI_INT", mulhs
959 let Itinerary = TransALU;
961 class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
965 let Itinerary = TransALU;
967 class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
968 inst, "MULLO_INT", mul
971 let Itinerary = TransALU;
973 class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []> {
975 let Itinerary = TransALU;
978 class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
979 inst, "RECIP_CLAMPED", []
982 let Itinerary = TransALU;
985 class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
986 inst, "RECIP_IEEE", [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
989 let Itinerary = TransALU;
992 class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
993 inst, "RECIP_UINT", AMDGPUurecip
996 let Itinerary = TransALU;
999 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
1000 inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
1003 let Itinerary = TransALU;
1006 class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
1007 inst, "RECIPSQRT_IEEE", []
1010 let Itinerary = TransALU;
1013 class SIN_Common <bits<11> inst> : R600_1OP <
1017 let Itinerary = TransALU;
1020 class COS_Common <bits<11> inst> : R600_1OP <
1024 let Itinerary = TransALU;
1027 //===----------------------------------------------------------------------===//
1028 // Helper patterns for complex intrinsics
1029 //===----------------------------------------------------------------------===//
1031 multiclass DIV_Common <InstR600 recip_ieee> {
1033 (int_AMDGPU_div f32:$src0, f32:$src1),
1034 (MUL_IEEE $src0, (recip_ieee $src1))
1038 (fdiv f32:$src0, f32:$src1),
1039 (MUL_IEEE $src0, (recip_ieee $src1))
1043 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee>
1045 (int_TGSI_lit_z f32:$src_x, f32:$src_y, f32:$src_w),
1046 (exp_ieee (mul_lit (log_clamped (MAX $src_y, (f32 ZERO))), $src_w, $src_x))
1049 //===----------------------------------------------------------------------===//
1050 // R600 / R700 Instructions
1051 //===----------------------------------------------------------------------===//
1053 let Predicates = [isR600] in {
1055 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
1056 def MULADD_r600 : MULADD_Common<0x10>;
1057 def MULADD_IEEE_r600 : MULADD_IEEE_Common<0x14>;
1058 def CNDE_r600 : CNDE_Common<0x18>;
1059 def CNDGT_r600 : CNDGT_Common<0x19>;
1060 def CNDGE_r600 : CNDGE_Common<0x1A>;
1061 def DOT4_r600 : DOT4_Common<0x50>;
1062 defm CUBE_r600 : CUBE_Common<0x52>;
1063 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
1064 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
1065 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
1066 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
1067 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
1068 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
1069 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
1070 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
1071 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
1072 def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
1073 def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
1074 def SIN_r600 : SIN_Common<0x6E>;
1075 def COS_r600 : COS_Common<0x6F>;
1076 def ASHR_r600 : ASHR_Common<0x70>;
1077 def LSHR_r600 : LSHR_Common<0x71>;
1078 def LSHL_r600 : LSHL_Common<0x72>;
1079 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
1080 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
1081 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
1082 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
1083 def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
1085 defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
1086 def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL>;
1087 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
1089 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
1091 def R600_ExportSwz : ExportSwzInst {
1092 let Word1{20-17} = 0; // BURST_COUNT
1093 let Word1{21} = eop;
1094 let Word1{22} = 1; // VALID_PIXEL_MODE
1095 let Word1{30-23} = inst;
1096 let Word1{31} = 1; // BARRIER
1098 defm : ExportPattern<R600_ExportSwz, 39>;
1100 def R600_ExportBuf : ExportBufInst {
1101 let Word1{20-17} = 0; // BURST_COUNT
1102 let Word1{21} = eop;
1103 let Word1{22} = 1; // VALID_PIXEL_MODE
1104 let Word1{30-23} = inst;
1105 let Word1{31} = 1; // BARRIER
1107 defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
1109 def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1110 "TEX $COUNT @$ADDR"> {
1113 def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1114 "VTX $COUNT @$ADDR"> {
1117 def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
1118 "LOOP_START_DX10 @$ADDR"> {
1122 def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1126 def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
1127 "LOOP_BREAK @$ADDR"> {
1131 def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
1132 "CONTINUE @$ADDR"> {
1136 def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1137 "JUMP @$ADDR POP:$POP_COUNT"> {
1140 def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1141 "ELSE @$ADDR POP:$POP_COUNT"> {
1144 def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
1149 def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1150 "POP @$ADDR POP:$POP_COUNT"> {
1153 def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
1157 let END_OF_PROGRAM = 1;
1162 // Helper pattern for normalizing inputs to triginomic instructions for R700+
1164 class COS_PAT <InstR600 trig> : Pat<
1166 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1169 class SIN_PAT <InstR600 trig> : Pat<
1171 (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
1174 //===----------------------------------------------------------------------===//
1175 // R700 Only instructions
1176 //===----------------------------------------------------------------------===//
1178 let Predicates = [isR700] in {
1179 def SIN_r700 : SIN_Common<0x6E>;
1180 def COS_r700 : COS_Common<0x6F>;
1182 // R700 normalizes inputs to SIN/COS the same as EG
1183 def : SIN_PAT <SIN_r700>;
1184 def : COS_PAT <COS_r700>;
1187 //===----------------------------------------------------------------------===//
1188 // Evergreen Only instructions
1189 //===----------------------------------------------------------------------===//
1191 let Predicates = [isEG] in {
1193 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
1194 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
1196 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
1197 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
1198 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
1199 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
1200 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
1201 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
1202 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
1203 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
1204 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
1205 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
1206 def SIN_eg : SIN_Common<0x8D>;
1207 def COS_eg : COS_Common<0x8E>;
1209 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
1210 def : SIN_PAT <SIN_eg>;
1211 def : COS_PAT <COS_eg>;
1212 def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
1213 } // End Predicates = [isEG]
1215 //===----------------------------------------------------------------------===//
1216 // Evergreen / Cayman Instructions
1217 //===----------------------------------------------------------------------===//
1219 let Predicates = [isEGorCayman] in {
1221 // BFE_UINT - bit_extract, an optimization for mask and shift
1226 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
1231 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
1232 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
1233 // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
1234 // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
1235 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
1236 [(set i32:$dst, (int_AMDIL_bit_extract_u32 i32:$src0, i32:$src1,
1240 def : BFEPattern <BFE_UINT_eg>;
1242 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>;
1243 defm : BFIPatterns <BFI_INT_eg>;
1245 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
1246 def : ROTRPattern <BIT_ALIGN_INT_eg>;
1248 def MULADD_eg : MULADD_Common<0x14>;
1249 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
1250 def ASHR_eg : ASHR_Common<0x15>;
1251 def LSHR_eg : LSHR_Common<0x16>;
1252 def LSHL_eg : LSHL_Common<0x17>;
1253 def CNDE_eg : CNDE_Common<0x19>;
1254 def CNDGT_eg : CNDGT_Common<0x1A>;
1255 def CNDGE_eg : CNDGE_Common<0x1B>;
1256 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
1257 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
1258 def DOT4_eg : DOT4_Common<0xBE>;
1259 defm CUBE_eg : CUBE_Common<0xC0>;
1261 let hasSideEffects = 1 in {
1262 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
1265 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
1267 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
1271 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
1273 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
1277 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
1279 // TRUNC is used for the FLT_TO_INT instructions to work around a
1280 // perceived problem where the rounding modes are applied differently
1281 // depending on the instruction and the slot they are in.
1283 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
1284 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
1286 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
1287 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
1288 // We should look into handling these cases separately.
1289 def : Pat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
1291 def : Pat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
1294 def : SHA256MaPattern <BFI_INT_eg, XOR_INT>;
1296 def EG_ExportSwz : ExportSwzInst {
1297 let Word1{19-16} = 0; // BURST_COUNT
1298 let Word1{20} = 1; // VALID_PIXEL_MODE
1299 let Word1{21} = eop;
1300 let Word1{29-22} = inst;
1301 let Word1{30} = 0; // MARK
1302 let Word1{31} = 1; // BARRIER
1304 defm : ExportPattern<EG_ExportSwz, 83>;
1306 def EG_ExportBuf : ExportBufInst {
1307 let Word1{19-16} = 0; // BURST_COUNT
1308 let Word1{20} = 1; // VALID_PIXEL_MODE
1309 let Word1{21} = eop;
1310 let Word1{29-22} = inst;
1311 let Word1{30} = 0; // MARK
1312 let Word1{31} = 1; // BARRIER
1314 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
1316 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
1317 "TEX $COUNT @$ADDR"> {
1320 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
1321 "VTX $COUNT @$ADDR"> {
1324 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
1325 "LOOP_START_DX10 @$ADDR"> {
1329 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
1333 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
1334 "LOOP_BREAK @$ADDR"> {
1338 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
1339 "CONTINUE @$ADDR"> {
1343 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1344 "JUMP @$ADDR POP:$POP_COUNT"> {
1347 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1348 "ELSE @$ADDR POP:$POP_COUNT"> {
1351 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
1356 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
1357 "POP @$ADDR POP:$POP_COUNT"> {
1360 def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
1364 let END_OF_PROGRAM = 1;
1367 //===----------------------------------------------------------------------===//
1368 // Memory read/write instructions
1369 //===----------------------------------------------------------------------===//
1370 let usesCustomInserter = 1 in {
1372 class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> mask, string name,
1374 : EG_CF_RAT <0x57, 0x2, 0, (outs), ins, name, pattern> {
1376 // XXX: Have a separate instruction for non-indexed writes.
1382 let comp_mask = mask;
1383 let burst_count = 0;
1389 } // End usesCustomInserter = 1
1392 def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
1393 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1394 0x1, "RAT_WRITE_CACHELESS_32_eg $rw_gpr, $index_gpr, $eop",
1395 [(global_store i32:$rw_gpr, i32:$index_gpr)]
1399 def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
1400 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
1401 0xf, "RAT_WRITE_CACHELESS_128 $rw_gpr.XYZW, $index_gpr, $eop",
1402 [(global_store v4i32:$rw_gpr, i32:$index_gpr)]
1405 class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
1406 : InstR600ISA <outs, (ins MEMxi:$ptr), name, pattern>,
1407 VTX_WORD1_GPR, VTX_WORD0 {
1412 let FETCH_WHOLE_QUAD = 0;
1413 let BUFFER_ID = buffer_id;
1415 // XXX: We can infer this field based on the SRC_GPR. This would allow us
1416 // to store vertex addresses in any channel, not just X.
1419 // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
1420 // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
1421 // however, based on my testing if USE_CONST_FIELDS is set, then all
1422 // these fields need to be set to 0.
1423 let USE_CONST_FIELDS = 0;
1424 let NUM_FORMAT_ALL = 1;
1425 let FORMAT_COMP_ALL = 0;
1426 let SRF_MODE_ALL = 0;
1428 let Inst{31-0} = Word0;
1429 let Inst{63-32} = Word1;
1430 // LLVM can only encode 64-bit instructions, so these fields are manually
1431 // encoded in R600CodeEmitter
1434 // bits<2> ENDIAN_SWAP = 0;
1435 // bits<1> CONST_BUF_NO_STRIDE = 0;
1436 // bits<1> MEGA_FETCH = 0;
1437 // bits<1> ALT_CONST = 0;
1438 // bits<2> BUFFER_INDEX_MODE = 0;
1442 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1443 // is done in R600CodeEmitter
1445 // Inst{79-64} = OFFSET;
1446 // Inst{81-80} = ENDIAN_SWAP;
1447 // Inst{82} = CONST_BUF_NO_STRIDE;
1448 // Inst{83} = MEGA_FETCH;
1449 // Inst{84} = ALT_CONST;
1450 // Inst{86-85} = BUFFER_INDEX_MODE;
1451 // Inst{95-86} = 0; Reserved
1453 // VTX_WORD3 (Padding)
1455 // Inst{127-96} = 0;
1460 class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
1461 : VTX_READ_eg <"VTX_READ_8 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1464 let MEGA_FETCH_COUNT = 1;
1466 let DST_SEL_Y = 7; // Masked
1467 let DST_SEL_Z = 7; // Masked
1468 let DST_SEL_W = 7; // Masked
1469 let DATA_FORMAT = 1; // FMT_8
1472 class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
1473 : VTX_READ_eg <"VTX_READ_16 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1475 let MEGA_FETCH_COUNT = 2;
1477 let DST_SEL_Y = 7; // Masked
1478 let DST_SEL_Z = 7; // Masked
1479 let DST_SEL_W = 7; // Masked
1480 let DATA_FORMAT = 5; // FMT_16
1484 class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
1485 : VTX_READ_eg <"VTX_READ_32 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
1488 let MEGA_FETCH_COUNT = 4;
1490 let DST_SEL_Y = 7; // Masked
1491 let DST_SEL_Z = 7; // Masked
1492 let DST_SEL_W = 7; // Masked
1493 let DATA_FORMAT = 0xD; // COLOR_32
1495 // This is not really necessary, but there were some GPU hangs that appeared
1496 // to be caused by ALU instructions in the next instruction group that wrote
1497 // to the $ptr registers of the VTX_READ.
1499 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
1500 // %T2_X<def> = MOV %ZERO
1501 //Adding this constraint prevents this from happening.
1502 let Constraints = "$ptr.ptr = $dst";
1505 class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
1506 : VTX_READ_eg <"VTX_READ_128 $dst.XYZW, $ptr", buffer_id, (outs R600_Reg128:$dst),
1509 let MEGA_FETCH_COUNT = 16;
1514 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
1516 // XXX: Need to force VTX_READ_128 instructions to write to the same register
1517 // that holds its buffer address to avoid potential hangs. We can't use
1518 // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
1519 // registers are different sizes.
1522 //===----------------------------------------------------------------------===//
1523 // VTX Read from parameter memory space
1524 //===----------------------------------------------------------------------===//
1526 def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
1527 [(set i32:$dst, (load_param_zexti8 ADDRVTX_READ:$ptr))]
1530 def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
1531 [(set i32:$dst, (load_param_zexti16 ADDRVTX_READ:$ptr))]
1534 def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
1535 [(set i32:$dst, (load_param ADDRVTX_READ:$ptr))]
1538 def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
1539 [(set v4i32:$dst, (load_param ADDRVTX_READ:$ptr))]
1542 //===----------------------------------------------------------------------===//
1543 // VTX Read from global memory space
1544 //===----------------------------------------------------------------------===//
1547 def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
1548 [(set i32:$dst, (zextloadi8_global ADDRVTX_READ:$ptr))]
1552 def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
1553 [(set i32:$dst, (global_load ADDRVTX_READ:$ptr))]
1557 def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
1558 [(set v4i32:$dst, (global_load ADDRVTX_READ:$ptr))]
1561 //===----------------------------------------------------------------------===//
1563 // XXX: We are currently storing all constants in the global address space.
1564 //===----------------------------------------------------------------------===//
1566 def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
1567 [(set i32:$dst, (constant_load ADDRVTX_READ:$ptr))]
1572 //===----------------------------------------------------------------------===//
1573 // Regist loads and stores - for indirect addressing
1574 //===----------------------------------------------------------------------===//
1576 defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
1578 let Predicates = [isCayman] in {
1580 let isVector = 1 in {
1582 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
1584 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
1585 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
1586 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
1587 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
1588 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
1589 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
1590 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
1591 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
1592 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
1593 def SIN_cm : SIN_Common<0x8D>;
1594 def COS_cm : COS_Common<0x8E>;
1595 } // End isVector = 1
1597 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
1598 def : SIN_PAT <SIN_cm>;
1599 def : COS_PAT <COS_cm>;
1601 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
1603 // RECIP_UINT emulation for Cayman
1604 // The multiplication scales from [0,1] to the unsigned integer range
1606 (AMDGPUurecip i32:$src0),
1607 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
1608 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
1611 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
1617 def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
1621 //===----------------------------------------------------------------------===//
1622 // Branch Instructions
1623 //===----------------------------------------------------------------------===//
1626 def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
1627 "IF_PREDICATE_SET $src", []>;
1629 def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
1630 "PREDICATED_BREAK $src", []>;
1632 //===----------------------------------------------------------------------===//
1633 // Pseudo instructions
1634 //===----------------------------------------------------------------------===//
1636 let isPseudo = 1 in {
1638 def PRED_X : InstR600 <
1639 (outs R600_Predicate_Bit:$dst),
1640 (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
1642 let FlagOperandIdx = 3;
1645 let isTerminator = 1, isBranch = 1 in {
1646 def JUMP_COND : InstR600 <
1648 (ins brtarget:$target, R600_Predicate_Bit:$p),
1649 "JUMP $target ($p)",
1653 def JUMP : InstR600 <
1655 (ins brtarget:$target),
1660 let isPredicable = 1;
1664 } // End isTerminator = 1, isBranch = 1
1666 let usesCustomInserter = 1 in {
1668 let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
1670 def MASK_WRITE : AMDGPUShaderInst <
1672 (ins R600_Reg32:$src),
1677 } // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
1681 (outs R600_Reg128:$dst),
1682 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
1683 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1684 "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1685 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
1686 imm:$resourceId, imm:$samplerId, imm:$textureTarget))],
1691 def TXD_SHADOW: InstR600 <
1692 (outs R600_Reg128:$dst),
1693 (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2,
1694 i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
1695 "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
1696 [(set v4f32:$dst, (int_AMDGPU_txd v4f32:$src0, v4f32:$src1, v4f32:$src2,
1697 imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))],
1702 } // End isPseudo = 1
1703 } // End usesCustomInserter = 1
1705 def CLAMP_R600 : CLAMP <R600_Reg32>;
1706 def FABS_R600 : FABS<R600_Reg32>;
1707 def FNEG_R600 : FNEG<R600_Reg32>;
1709 //===---------------------------------------------------------------------===//
1710 // Return instruction
1711 //===---------------------------------------------------------------------===//
1712 let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
1713 usesCustomInserter = 1 in {
1714 def RETURN : ILFormat<(outs), (ins variable_ops),
1715 "RETURN", [(IL_retflag)]>;
1719 //===----------------------------------------------------------------------===//
1720 // Constant Buffer Addressing Support
1721 //===----------------------------------------------------------------------===//
1723 let usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
1724 def CONST_COPY : Instruction {
1725 let OutOperandList = (outs R600_Reg32:$dst);
1726 let InOperandList = (ins i32imm:$src);
1728 [(set R600_Reg32:$dst, (CONST_ADDRESS ADDRGA_CONST_OFFSET:$src))];
1729 let AsmString = "CONST_COPY";
1730 let neverHasSideEffects = 1;
1731 let isAsCheapAsAMove = 1;
1732 let Itinerary = NullALU;
1734 } // end usesCustomInserter = 1, isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU"
1736 def TEX_VTX_CONSTBUF :
1737 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
1738 [(set v4i32:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
1739 VTX_WORD1_GPR, VTX_WORD0 {
1743 let FETCH_WHOLE_QUAD = 0;
1747 let USE_CONST_FIELDS = 0;
1748 let NUM_FORMAT_ALL = 2;
1749 let FORMAT_COMP_ALL = 1;
1750 let SRF_MODE_ALL = 1;
1751 let MEGA_FETCH_COUNT = 16;
1756 let DATA_FORMAT = 35;
1758 let Inst{31-0} = Word0;
1759 let Inst{63-32} = Word1;
1761 // LLVM can only encode 64-bit instructions, so these fields are manually
1762 // encoded in R600CodeEmitter
1765 // bits<2> ENDIAN_SWAP = 0;
1766 // bits<1> CONST_BUF_NO_STRIDE = 0;
1767 // bits<1> MEGA_FETCH = 0;
1768 // bits<1> ALT_CONST = 0;
1769 // bits<2> BUFFER_INDEX_MODE = 0;
1773 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1774 // is done in R600CodeEmitter
1776 // Inst{79-64} = OFFSET;
1777 // Inst{81-80} = ENDIAN_SWAP;
1778 // Inst{82} = CONST_BUF_NO_STRIDE;
1779 // Inst{83} = MEGA_FETCH;
1780 // Inst{84} = ALT_CONST;
1781 // Inst{86-85} = BUFFER_INDEX_MODE;
1782 // Inst{95-86} = 0; Reserved
1784 // VTX_WORD3 (Padding)
1786 // Inst{127-96} = 0;
1791 InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
1792 [(set v4f32:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
1793 VTX_WORD1_GPR, VTX_WORD0 {
1797 let FETCH_WHOLE_QUAD = 0;
1801 let USE_CONST_FIELDS = 1;
1802 let NUM_FORMAT_ALL = 0;
1803 let FORMAT_COMP_ALL = 0;
1804 let SRF_MODE_ALL = 1;
1805 let MEGA_FETCH_COUNT = 16;
1810 let DATA_FORMAT = 0;
1812 let Inst{31-0} = Word0;
1813 let Inst{63-32} = Word1;
1815 // LLVM can only encode 64-bit instructions, so these fields are manually
1816 // encoded in R600CodeEmitter
1819 // bits<2> ENDIAN_SWAP = 0;
1820 // bits<1> CONST_BUF_NO_STRIDE = 0;
1821 // bits<1> MEGA_FETCH = 0;
1822 // bits<1> ALT_CONST = 0;
1823 // bits<2> BUFFER_INDEX_MODE = 0;
1827 // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
1828 // is done in R600CodeEmitter
1830 // Inst{79-64} = OFFSET;
1831 // Inst{81-80} = ENDIAN_SWAP;
1832 // Inst{82} = CONST_BUF_NO_STRIDE;
1833 // Inst{83} = MEGA_FETCH;
1834 // Inst{84} = ALT_CONST;
1835 // Inst{86-85} = BUFFER_INDEX_MODE;
1836 // Inst{95-86} = 0; Reserved
1838 // VTX_WORD3 (Padding)
1840 // Inst{127-96} = 0;
1846 //===--------------------------------------------------------------------===//
1847 // Instructions support
1848 //===--------------------------------------------------------------------===//
1849 //===---------------------------------------------------------------------===//
1850 // Custom Inserter for Branches and returns, this eventually will be a
1852 //===---------------------------------------------------------------------===//
1853 let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
1854 def BRANCH : ILFormat<(outs), (ins brtarget:$target),
1855 "; Pseudo unconditional branch instruction",
1857 defm BRANCH_COND : BranchConditional<IL_brcond>;
1860 //===---------------------------------------------------------------------===//
1861 // Flow and Program control Instructions
1862 //===---------------------------------------------------------------------===//
1863 let isTerminator=1 in {
1864 def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
1865 !strconcat("SWITCH", " $src"), []>;
1866 def CASE : ILFormat< (outs), (ins GPRI32:$src),
1867 !strconcat("CASE", " $src"), []>;
1868 def BREAK : ILFormat< (outs), (ins),
1870 def CONTINUE : ILFormat< (outs), (ins),
1872 def DEFAULT : ILFormat< (outs), (ins),
1874 def ELSE : ILFormat< (outs), (ins),
1876 def ENDSWITCH : ILFormat< (outs), (ins),
1878 def ENDMAIN : ILFormat< (outs), (ins),
1880 def END : ILFormat< (outs), (ins),
1882 def ENDFUNC : ILFormat< (outs), (ins),
1884 def ENDIF : ILFormat< (outs), (ins),
1886 def WHILELOOP : ILFormat< (outs), (ins),
1888 def ENDLOOP : ILFormat< (outs), (ins),
1890 def FUNC : ILFormat< (outs), (ins),
1892 def RETDYN : ILFormat< (outs), (ins),
1894 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1895 defm IF_LOGICALNZ : BranchInstr<"IF_LOGICALNZ">;
1896 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1897 defm IF_LOGICALZ : BranchInstr<"IF_LOGICALZ">;
1898 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1899 defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
1900 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1901 defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
1902 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1903 defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
1904 // This opcode has custom swizzle pattern encoded in Swizzle Encoder
1905 defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
1906 defm IFC : BranchInstr2<"IFC">;
1907 defm BREAKC : BranchInstr2<"BREAKC">;
1908 defm CONTINUEC : BranchInstr2<"CONTINUEC">;
1911 //===----------------------------------------------------------------------===//
1913 //===----------------------------------------------------------------------===//
1915 // CND*_INT Pattterns for f32 True / False values
1917 class CND_INT_f32 <InstR600 cnd, CondCode cc> : Pat <
1918 (selectcc i32:$src0, 0, f32:$src1, f32:$src2, cc),
1919 (cnd $src0, $src1, $src2)
1922 def : CND_INT_f32 <CNDE_INT, SETEQ>;
1923 def : CND_INT_f32 <CNDGT_INT, SETGT>;
1924 def : CND_INT_f32 <CNDGE_INT, SETGE>;
1926 //CNDGE_INT extra pattern
1928 (selectcc i32:$src0, -1, i32:$src1, i32:$src2, COND_GT),
1929 (CNDGE_INT $src0, $src1, $src2)
1935 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
1939 (int_AMDGPU_kill f32:$src0),
1940 (MASK_WRITE (KILLGT (f32 ZERO), $src0))
1945 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LT),
1951 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LE),
1955 // SETGT_DX10 reverse args
1957 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LT),
1958 (SETGT_DX10 $src1, $src0)
1961 // SETGE_DX10 reverse args
1963 (selectcc f32:$src0, f32:$src1, -1, 0, COND_LE),
1964 (SETGE_DX10 $src1, $src0)
1967 // SETGT_INT reverse args
1969 (selectcc i32:$src0, i32:$src1, -1, 0, SETLT),
1970 (SETGT_INT $src1, $src0)
1973 // SETGE_INT reverse args
1975 (selectcc i32:$src0, i32:$src1, -1, 0, SETLE),
1976 (SETGE_INT $src1, $src0)
1979 // SETGT_UINT reverse args
1981 (selectcc i32:$src0, i32:$src1, -1, 0, SETULT),
1982 (SETGT_UINT $src1, $src0)
1985 // SETGE_UINT reverse args
1987 (selectcc i32:$src0, i32:$src1, -1, 0, SETULE),
1988 (SETGE_UINT $src1, $src0)
1991 // The next two patterns are special cases for handling 'true if ordered' and
1992 // 'true if unordered' conditionals. The assumption here is that the behavior of
1993 // SETE and SNE conforms to the Direct3D 10 rules for floating point values
1995 // http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
1996 // We assume that SETE returns false when one of the operands is NAN and
1997 // SNE returns true when on of the operands is NAN
1999 //SETE - 'true if ordered'
2001 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETO),
2005 //SETE_DX10 - 'true if ordered'
2007 (selectcc f32:$src0, f32:$src1, -1, 0, SETO),
2008 (SETE_DX10 $src0, $src1)
2011 //SNE - 'true if unordered'
2013 (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETUO),
2017 //SETNE_DX10 - 'true if ordered'
2019 (selectcc f32:$src0, f32:$src1, -1, 0, SETUO),
2020 (SETNE_DX10 $src0, $src1)
2023 def : Extract_Element <f32, v4f32, 0, sub0>;
2024 def : Extract_Element <f32, v4f32, 1, sub1>;
2025 def : Extract_Element <f32, v4f32, 2, sub2>;
2026 def : Extract_Element <f32, v4f32, 3, sub3>;
2028 def : Insert_Element <f32, v4f32, 0, sub0>;
2029 def : Insert_Element <f32, v4f32, 1, sub1>;
2030 def : Insert_Element <f32, v4f32, 2, sub2>;
2031 def : Insert_Element <f32, v4f32, 3, sub3>;
2033 def : Extract_Element <i32, v4i32, 0, sub0>;
2034 def : Extract_Element <i32, v4i32, 1, sub1>;
2035 def : Extract_Element <i32, v4i32, 2, sub2>;
2036 def : Extract_Element <i32, v4i32, 3, sub3>;
2038 def : Insert_Element <i32, v4i32, 0, sub0>;
2039 def : Insert_Element <i32, v4i32, 1, sub1>;
2040 def : Insert_Element <i32, v4i32, 2, sub2>;
2041 def : Insert_Element <i32, v4i32, 3, sub3>;
2043 def : Vector4_Build <v4f32, f32>;
2044 def : Vector4_Build <v4i32, i32>;
2046 // bitconvert patterns
2048 def : BitConvert <i32, f32, R600_Reg32>;
2049 def : BitConvert <f32, i32, R600_Reg32>;
2050 def : BitConvert <v4f32, v4i32, R600_Reg128>;
2051 def : BitConvert <v4i32, v4f32, R600_Reg128>;
2053 // DWORDADDR pattern
2054 def : DwordAddrPat <i32, R600_Reg32>;
2056 } // End isR600toCayman Predicate