1 //===-- R600Instructions.td - TODO: Add brief description -------===//
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 // TODO: Add full description
12 //===----------------------------------------------------------------------===//
14 include "R600Intrinsics.td"
16 class InstR600 <bits<32> inst, dag outs, dag ins, string asm, list<dag> pattern,
18 : AMDGPUInst <outs, ins, asm, pattern> {
25 let Namespace = "AMDIL";
26 let OutOperandList = outs;
27 let InOperandList = ins;
29 let Pattern = pattern;
32 let TSFlags{4} = Trig;
36 class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
37 AMDGPUInst <outs, ins, asm, pattern>
41 let Namespace = "AMDIL";
44 def MEMri : Operand<iPTRAny> {
45 let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
48 def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
62 class R600_1OP <bits<32> inst, string opName, list<dag> pattern,
63 InstrItinClass itin = AnyALU> :
65 (outs R600_Reg32:$dst),
66 (ins R600_Reg32:$src, variable_ops),
67 !strconcat(opName, " $dst, $src"),
72 class R600_2OP <bits<32> inst, string opName, list<dag> pattern,
73 InstrItinClass itin = AnyALU> :
75 (outs R600_Reg32:$dst),
76 (ins R600_Reg32:$src0, R600_Reg32:$src1, variable_ops),
77 !strconcat(opName, " $dst, $src0, $src1"),
82 class R600_3OP <bits<32> inst, string opName, list<dag> pattern,
83 InstrItinClass itin = AnyALU> :
85 (outs R600_Reg32:$dst),
86 (ins R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2, variable_ops),
87 !strconcat(opName, "$dst $src0, $src1, $src2"),
94 class R600_REDUCTION <bits<32> inst, dag ins, string asm, list<dag> pattern,
95 InstrItinClass itin = AnyALU> :
97 (outs R600_Reg32:$dst),
105 class R600_TEX <bits<32> inst, string opName, list<dag> pattern,
106 InstrItinClass itin = AnyALU> :
108 (outs R600_Reg128:$dst),
109 (ins R600_Reg128:$src0, i32imm:$src1, i32imm:$src2),
110 !strconcat(opName, "$dst, $src0, $src1, $src2"),
115 def TEX_SHADOW : PatLeaf<
117 [{uint32_t TType = (uint32_t)N->getZExtValue();
118 return (TType >= 6 && TType <= 8) || TType == 11 || TType == 12;
122 class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, dag outs, dag ins,
124 InstR600ISA <outs, ins, asm, []>
143 /* CF_ALLOC_EXPORT_WORD0_RAT */
144 let Inst{3-0} = RAT_ID;
145 let Inst{9-4} = rat_inst;
146 let Inst{10} = 0; /* Reserved */
147 let Inst{12-11} = RIM;
148 let Inst{14-13} = TYPE;
149 let Inst{21-15} = RW_GPR;
150 let Inst{22} = RW_REL;
151 let Inst{29-23} = INDEX_GPR;
152 let Inst{31-30} = ELEM_SIZE;
154 /* CF_ALLOC_EXPORT_WORD1_BUF */
155 /* XXX: We can't have auto encoding of 64-bit instructions until LLVM 3.1 :( */
157 let Inst{43-32} = ARRAY_SIZE;
158 let Inst{47-44} = COMP_MASK;
159 let Inst{51-48} = BURST_COUNT;
162 let Inst{61-54} = cf_inst;
164 let Inst{63} = BARRIER;
169 def store_global : PatFrag<(ops node:$value, node:$ptr),
170 (store node:$value, node:$ptr),
173 const PointerType *Type;
174 if ((src = cast<StoreSDNode>(N)->getSrcValue() &&
175 PT = dyn_cast<PointerType>(Src->getType()))) {
176 return PT->getAddressSpace() == 1;
183 def load_param : PatFrag<(ops node:$ptr),
187 const Value *Src = cast<LoadSDNode>(N)->getSrcValue();
189 PointerType * PT = dyn_cast<PointerType>(Src->getType());
190 return PT && PT->getAddressSpace() == AMDILAS::PARAM_I_ADDRESS;
195 //class EG_CF <bits<32> inst, string asm> :
196 // InstR600 <inst, (outs), (ins), asm, []>;
198 /* XXX: We will use this when we emit the real ISA.
212 let Inst{23-0} = ADDR;
213 let Inst{26-24} = JTS;
214 let Inst{34-32} = PC;
215 let Inst{39-35} = CF_CONST;
216 let Inst{41-40} = COND;
217 let Inst{47-42} = COUNT;
220 let Inst{61-54} = CF_INST;
225 def isR600 : Predicate<"Subtarget.device()"
226 "->getGeneration() == AMDILDeviceInfo::HD4XXX">;
227 def isEG : Predicate<"Subtarget.device()"
228 "->getGeneration() >= AMDILDeviceInfo::HD5XXX && "
229 "Subtarget.device()->getDeviceFlag() != OCL_DEVICE_CAYMAN">;
230 def isCayman : Predicate<"Subtarget.device()"
231 "->getDeviceFlag() == OCL_DEVICE_CAYMAN">;
232 def isEGorCayman : Predicate<"Subtarget.device()"
233 "->getGeneration() >= AMDILDeviceInfo::HD5XXX">;
235 def isR600toCayman : Predicate<
236 "Subtarget.device()->getGeneration() <= AMDILDeviceInfo::HD6XXX">;
239 let Predicates = [isR600toCayman] in {
241 /* ------------------------------------------- */
242 /* Common Instructions R600, R700, Evergreen, Cayman */
243 /* ------------------------------------------- */
244 let Gen = AMDGPUGen.R600_CAYMAN in {
248 [(set R600_Reg32:$dst, (fadd R600_Reg32:$src0, R600_Reg32:$src1))] > {
249 let AMDILOp = AMDILInst.ADD_f32;
251 // Non-IEEE MUL: 0 * anything = 0
254 [(set R600_Reg32:$dst, (int_AMDGPU_mul R600_Reg32:$src0, R600_Reg32:$src1))]
257 def MUL_IEEE : R600_2OP <
259 [(set R600_Reg32:$dst, (fmul R600_Reg32:$src0, R600_Reg32:$src1))]> {
260 let AMDILOp = AMDILInst.MUL_IEEE_f32;
265 [(set R600_Reg32:$dst, (int_AMDIL_max R600_Reg32:$src0, R600_Reg32:$src1))]> {
266 let AMDILOp = AMDILInst.MAX_f32;
271 [(set R600_Reg32:$dst, (int_AMDIL_min R600_Reg32:$src0, R600_Reg32:$src1))]> {
272 let AMDILOp = AMDILInst.MIN_f32;
275 /* For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
276 * so some of the instruction names don't match the asm string.
277 * XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
280 def SETE : R600_2OP <
282 [(set R600_Reg32:$dst, (int_AMDGPU_seq R600_Reg32:$src0, R600_Reg32:$src1))]> {
283 let AMDILOp = AMDILInst.FEQ;
288 [(set R600_Reg32:$dst, (int_AMDGPU_sgt R600_Reg32:$src0, R600_Reg32:$src1))]
293 [(set R600_Reg32:$dst, (int_AMDGPU_sge R600_Reg32:$src0, R600_Reg32:$src1))]> {
294 let AMDILOp = AMDILInst.FGE;
299 [(set R600_Reg32:$dst, (int_AMDGPU_sne R600_Reg32:$src0, R600_Reg32:$src1))]> {
300 let AMDILOp = AMDILInst.FNE;
303 def FRACT : R600_1OP <
306 let AMDILOp = AMDILInst.FRAC_f32;
309 def TRUNC : R600_1OP <
311 [(set R600_Reg32:$dst, (int_AMDGPU_trunc R600_Reg32:$src))]
314 def FLOOR : R600_1OP <
316 [(set R600_Reg32:$dst, (int_AMDGPU_floor R600_Reg32:$src))]
319 def MOV : R600_1OP <0x19, "MOV", []>;
321 def KILLGT : R600_2OP <
326 def AND_INT : R600_2OP <
329 let AMDILOp = AMDILInst.AND_i32;
332 def XOR_INT : R600_2OP <
337 def ADD_INT : R600_2OP <
338 0x34, "ADD_INT $dst, $src0, $src1",
340 let AMDILOp = AMDILInst.ADD_i32;
343 def SUB_INT : R600_2OP <
344 0x35, "SUB_INT $dst, $src0, $src1",
348 def SETE_INT : R600_2OP <
349 0x3A, "SETE_INT $dst, $src0, $src1",
351 let AMDILOp = AMDILInst.IEQ;
354 def SETGT_INT : R600_2OP <
355 0x3B, "SGT_INT $dst, $src0, $src1",
359 def SETGE_INT : R600_2OP <
360 0x3C, "SETGE_INT $dst, $src0, $src1",
362 let AMDILOp = AMDILInst.IGE;
365 def SETNE_INT : R600_2OP <
366 0x3D, "SETNE_INT $dst, $src0, $src1",
368 let AMDILOp = AMDILInst.INE;
371 def SETGT_UINT : R600_2OP <
372 0x3E, "SETGT_UINT $dst, $src0, $src1",
374 let AMDILOp = AMDILInst.UGT;
377 def SETGE_UINT : R600_2OP <
378 0x3F, "SETGE_UINT $dst, $src0, $src1",
380 let AMDILOp = AMDILInst.UGE;
383 def CNDE_INT : R600_3OP <
384 0x1C, "CNDE_INT $dst, $src0, $src1, $src2",
388 /* Texture instructions */
390 def TEX_SAMPLE : R600_TEX <
392 [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$src1, imm:$src2))]
395 def TEX_SAMPLE_C : R600_TEX <
396 0x18, "TEX_SAMPLE_C",
397 [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$src1, TEX_SHADOW:$src2))]
400 def TEX_SAMPLE_L : R600_TEX <
401 0x11, "TEX_SAMPLE_L",
402 [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$src1, imm:$src2))]
405 def TEX_SAMPLE_C_L : R600_TEX <
406 0x19, "TEX_SAMPLE_C_L",
407 [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$src1, TEX_SHADOW:$src2))]
410 def TEX_SAMPLE_LB : R600_TEX <
411 0x12, "TEX_SAMPLE_LB",
412 [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0, imm:$src1, imm:$src2))]
415 def TEX_SAMPLE_C_LB : R600_TEX <
416 0x1A, "TEX_SAMPLE_C_LB",
417 [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0, imm:$src1, TEX_SHADOW:$src2))]
420 def TEX_SAMPLE_G : R600_TEX <
421 0x14, "TEX_SAMPLE_G",
422 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, imm:$src1, imm:$src2))]
425 def TEX_SAMPLE_C_G : R600_TEX <
426 0x1C, "TEX_SAMPLE_C_G",
427 [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, imm:$src1, TEX_SHADOW:$src2))]
430 } // End Gen R600_CAYMAN
434 (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
437 /* Helper classes for common instructions */
439 class MUL_LIT_Common <bits<32> inst> : R600_3OP <
444 class MULADD_Common <bits<32> inst> : R600_3OP <
447 let AMDILOp = AMDILInst.MAD_f32;
450 class CNDE_Common <bits<32> inst> : R600_3OP <
453 let AMDILOp = AMDILInst.CMOVLOG_f32;
456 class CNDGT_Common <bits<32> inst> : R600_3OP <
461 class CNDGE_Common <bits<32> inst> : R600_3OP <
463 [(set R600_Reg32:$dst, (int_AMDGPU_cndlt R600_Reg32:$src0, R600_Reg32:$src2, R600_Reg32:$src1))]
466 class DOT4_Common <bits<32> inst> : R600_REDUCTION <
468 (ins R600_Reg128:$src0, R600_Reg128:$src1),
469 "DOT4 $dst $src0, $src1",
470 [(set R600_Reg32:$dst, (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1))]
473 class EXP_IEEE_Common <bits<32> inst> : R600_1OP <
476 let AMDILOp = AMDILInst.EXP_f32;
479 class FLT_TO_INT_Common <bits<32> inst> : R600_1OP <
480 inst, "FLT_TO_INT", []> {
481 let AMDILOp = AMDILInst.FTOI;
484 class INT_TO_FLT_Common <bits<32> inst> : R600_1OP <
485 inst, "INT_TO_FLT", []> {
486 let AMDILOp = AMDILInst.ITOF;
489 class LOG_CLAMPED_Common <bits<32> inst> : R600_1OP <
494 class LOG_IEEE_Common <bits<32> inst> : R600_1OP <
497 let AMDILOp = AMDILInst.LOG_f32;
500 class LSHL_Common <bits<32> inst> : R600_2OP <
501 inst, "LSHL $dst, $src0, $src1",
503 let AMDILOp = AMDILInst.SHL_i32;
506 class LSHR_Common <bits<32> inst> : R600_2OP <
507 inst, "LSHR $dst, $src0, $src1",
509 let AMDILOp = AMDILInst.USHR_i32;
512 class MULHI_INT_Common <bits<32> inst> : R600_2OP <
513 inst, "MULHI_INT $dst, $src0, $src1",
515 let AMDILOp = AMDILInst.SMULHI_i32;
518 class MULHI_UINT_Common <bits<32> inst> : R600_2OP <
519 inst, "MULHI $dst, $src0, $src1",
523 class MULLO_INT_Common <bits<32> inst> : R600_2OP <
524 inst, "MULLO_INT $dst, $src0, $src1",
526 let AMDILOp = AMDILInst.SMUL_i32;
529 class MULLO_UINT_Common <bits<32> inst> : R600_2OP <
530 inst, "MULLO_UINT $dst, $src0, $src1",
534 class RECIP_CLAMPED_Common <bits<32> inst> : R600_1OP <
535 inst, "RECIP_CLAMPED",
539 class RECIP_IEEE_Common <bits<32> inst> : R600_1OP <
541 [(set R600_Reg32:$dst, (int_AMDGPU_rcp R600_Reg32:$src))]> {
542 let AMDILOp = AMDILInst.RSQ_f32;
545 class RECIP_UINT_Common <bits<32> inst> : R600_1OP <
546 inst, "RECIP_INT $dst, $src",
550 class RECIPSQRT_CLAMPED_Common <bits<32> inst> : R600_1OP <
551 inst, "RECIPSQRT_CLAMPED",
552 [(set R600_Reg32:$dst, (int_AMDGPU_rsq R600_Reg32:$src))]
555 class RECIPSQRT_IEEE_Common <bits<32> inst> : R600_1OP <
556 inst, "RECIPSQRT_IEEE",
560 class SIN_Common <bits<32> inst> : R600_1OP <
563 let AMDILOp = AMDILInst.SIN_f32;
567 class COS_Common <bits<32> inst> : R600_1OP <
570 let AMDILOp = AMDILInst.COS_f32;
574 /* Helper patterns for complex intrinsics */
575 /* -------------------------------------- */
577 class DIV_Common <InstR600 recip_ieee> : Pat<
578 (int_AMDGPU_div R600_Reg32:$src0, R600_Reg32:$src1),
579 (MUL R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
582 class LRP_Common <InstR600 muladd> : Pat <
583 (int_AMDGPU_lrp R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
584 (muladd R600_Reg32:$src0, R600_Reg32:$src1, (MUL (SUB_f32 ONE, R600_Reg32:$src0), R600_Reg32:$src2))
587 class SSG_Common <InstR600 cndgt, InstR600 cndge> : Pat <
588 (int_AMDGPU_ssg R600_Reg32:$src),
589 (cndgt R600_Reg32:$src, (f32 ONE), (cndge R600_Reg32:$src, (f32 ZERO), (f32 NEG_ONE)))
592 class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee> : Pat <
593 (int_TGSI_lit_z R600_Reg32:$src_x, R600_Reg32:$src_y, R600_Reg32:$src_w),
594 (exp_ieee (mul_lit (log_clamped (MAX R600_Reg32:$src_y, (f32 ZERO))), R600_Reg32:$src_w, R600_Reg32:$src_x))
597 /* ---------------------- */
598 /* R600 / R700 Only Instructions */
599 /* ---------------------- */
601 let Predicates = [isR600] in {
603 let Gen = AMDGPUGen.R600 in {
605 def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
606 def MULADD_r600 : MULADD_Common<0x10>;
607 def CNDE_r600 : CNDE_Common<0x18>;
608 def CNDGT_r600 : CNDGT_Common<0x19>;
609 def CNDGE_r600 : CNDGE_Common<0x1A>;
610 def DOT4_r600 : DOT4_Common<0x50>;
611 def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
612 def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
613 def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
614 def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
615 def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
616 def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
617 def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
618 def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
619 def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
620 def SIN_r600 : SIN_Common<0x6E>;
621 def COS_r600 : COS_Common<0x6F>;
622 def LSHR_r600 : LSHR_Common<0x71>;
623 def LSHL_r600 : LSHL_Common<0x72>;
624 def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
625 def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
626 def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
627 def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
628 def RECIP_UINT_r600 : RECIP_UINT_Common <0x77>;
630 } // End AMDGPUGen.R600
632 def DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
633 def LRP_r600 : LRP_Common<MULADD_r600>;
634 def POW_r600 : POW_Common<LOG_IEEE_r600, EXP_IEEE_r600, MUL, GPRF32>;
635 def SSG_r600 : SSG_Common<CNDGT_r600, CNDGE_r600>;
636 def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
640 /* ----------------- */
641 /* R700+ Trig helper */
642 /* ----------------- */
645 class TRIG_HELPER_r700 <InstR600 trig_inst>: Pat <
646 (trig_inst R600_Reg32:$src),
647 (trig_inst (fmul R600_Reg32:$src, (PI))))
651 /* ---------------------- */
652 /* Evergreen Instructions */
653 /* ---------------------- */
656 let Predicates = [isEG] in {
658 let Gen = AMDGPUGen.EG in {
660 def RAT_WRITE_CACHELESS_eg :
661 EG_CF_RAT <0x57, 0x2, (outs), (ins R600_TReg32_X:$rw_gpr,
662 R600_TReg32_X:$index_gpr, i32imm:$rat_id), "">
665 let Inst{3-0} = RAT_ID;
666 let Inst{21-15} = RW_GPR;
667 let Inst{29-23} = INDEX_GPR;
668 /* Propery of the UAV */
669 let Inst{31-30} = ELEM_SIZE;
672 /* XXX: Have a separate instruction for non-indexed writes. */
688 def VTX_READ_eg : InstR600ISA < (outs R600_TReg32_X:$dst),
689 (ins R600_TReg32_X:$src, i32imm:$buffer_id),
690 "VTX_READ_eg $dst, $src", []>
697 /* If any of these field below need to be calculated at compile time, and
698 * a ins operand for them and move them to the list of operands above. */
700 /* XXX: This instruction is manual encoded, so none of these values are used.
703 bits<5> VC_INST = 0; //VC_INST_FETCH
704 bits<2> FETCH_TYPE = 2;
705 bits<1> FETCH_WHOLE_QUAD = 1;
707 bits<2> SRC_SEL_X = 0;
708 bits<6> MEGA_FETCH_COUNT = 4;
713 bits<3> DST_SEL_X = 0;
714 bits<3> DST_SEL_Y = 7; //Masked
715 bits<3> DST_SEL_Z = 7; //Masked
716 bits<3> DST_SEL_W = 7; //Masked
717 bits<1> USE_CONST_FIELDS = 1; //Masked
718 bits<6> DATA_FORMAT = 0;
719 bits<2> NUM_FORMAT_ALL = 0;
720 bits<1> FORMAT_COMP_ALL = 0;
721 bits<1> SRF_MODE_ALL = 0;
725 let Inst{4-0} = VC_INST;
726 let Inst{6-5} = FETCH_TYPE;
727 let Inst{7} = FETCH_WHOLE_QUAD;
728 let Inst{15-8} = BUFFER_ID;
729 let Inst{22-16} = SRC_GPR;
730 let Inst{23} = SRC_REL;
731 let Inst{25-24} = SRC_SEL_X;
732 let Inst{31-26} = MEGA_FETCH_COUNT;
734 /* DST_GPR is OK to leave uncommented, because LLVM 3.0 only prevents you
735 * from statically setting bits > 31. This field will be set by
736 * getMachineValueOp which can set bits > 31.
738 // let Inst{32-38} = DST_GPR;
740 /* XXX: Uncomment for LLVM 3.1 which supports 64-bit instructions */
743 let Inst{39} = DST_REL;
744 let Inst{40} = 0; //Reserved
745 let Inst{43-41} = DST_SEL_X;
746 let Inst{46-44} = DST_SEL_Y;
747 let Inst{49-47} = DST_SEL_Z;
748 let Inst{52-50} = DST_SEL_W;
749 let Inst{53} = USE_CONST_FIELDS;
750 let Inst{59-54} = DATA_FORMAT;
751 let Inst{61-60} = NUM_FORMAT_ALL;
752 let Inst{62} = FORMAT_COMP_ALL;
753 let Inst{63} = SRF_MODE_ALL;
759 } // End AMDGPUGen.EG
760 /* XXX: Need to convert PTR to rat_id */
762 def : Pat <(store_global (f32 R600_Reg32:$value), node:$ptr),
763 (RAT_WRITE_CACHELESS_eg (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)),
764 (f32 R600_Reg32:$value),
766 (f32 ZERO), 0, R600_Reg32:$ptr)>;
769 class VTX_Param_Read_Pattern <ValueType vt> : Pat <
770 (vt (load_param ADDRParam:$mem)),
771 (VTX_READ_eg (i32 R600_Reg32:$mem), 0)>;
773 def : VTX_Param_Read_Pattern <f32>;
774 def : VTX_Param_Read_Pattern <i32>;
776 } // End isEG Predicate
778 /* ------------------------------- */
779 /* Evergreen / Cayman Instructions */
780 /* ------------------------------- */
782 let Predicates = [isEGorCayman] in {
784 class TRIG_eg <InstR600 trig, Intrinsic intr> : Pat<
785 (intr R600_Reg32:$src),
786 (trig (MUL (MOV (LOADCONST_i32 CONST.TWO_PI_INV)), R600_Reg32:$src))
789 let Gen = AMDGPUGen.EG_CAYMAN in {
791 def MULADD_eg : MULADD_Common<0x14>;
792 def LSHR_eg : LSHR_Common<0x16>;
793 def LSHL_eg : LSHL_Common<0x17>;
794 def CNDE_eg : CNDE_Common<0x19>;
795 def CNDGT_eg : CNDGT_Common<0x1A>;
796 def CNDGE_eg : CNDGE_Common<0x1B>;
797 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
798 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50>;
799 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
800 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
801 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
802 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
803 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
804 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
805 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
806 def SIN_eg : SIN_Common<0x8D>;
807 def COS_eg : COS_Common<0x8E>;
808 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
809 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
810 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
811 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
812 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
813 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
814 def DOT4_eg : DOT4_Common<0xBE>;
816 } // End AMDGPUGen.EG_CAYMAN
818 def DIV_eg : DIV_Common<RECIP_IEEE_eg>;
819 def LRP_eg : LRP_Common<MULADD_eg>;
820 def POW_eg : POW_Common<LOG_IEEE_eg, EXP_IEEE_eg, MUL, GPRF32>;
821 def SSG_eg : SSG_Common<CNDGT_eg, CNDGE_eg>;
822 def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
824 def : TRIG_eg <SIN_eg, int_AMDGPU_sin>;
825 def : TRIG_eg <COS_eg, int_AMDGPU_cos>;
829 let Predicates = [isCayman] in {
831 let Gen = AMDGPUGen.CAYMAN in {
833 /* XXX: I'm not sure if this opcode is correct. */
834 def RECIP_UINT_cm : RECIP_UINT_Common<0x77>;
836 } // End AMDGPUGen.CAYMAN
840 /* Other Instructions */
842 let isCodeGenOnly = 1 in {
844 def SWIZZLE : AMDGPUShaderInst <
845 (outs GPRV4F32:$dst),
846 (ins GPRV4F32:$src0, i32imm:$src1),
847 "SWIZZLE $dst, $src0, $src1",
848 [(set GPRV4F32:$dst, (int_AMDGPU_swizzle GPRV4F32:$src0, imm:$src1))]
852 def LAST : AMDGPUShaderInst <
859 def GET_CHAN : AMDGPUShaderInst <
860 (outs R600_Reg32:$dst),
861 (ins R600_Reg128:$src0, i32imm:$src1),
862 "GET_CHAN $dst, $src0, $src1",
866 def SET_CHAN : AMDGPUShaderInst <
867 (outs R600_Reg128:$dst),
868 (ins R600_Reg32:$src0, i32imm:$src1),
869 "SET_CHAN $dst, $src0, $src1",
873 def MULLIT : AMDGPUShaderInst <
874 (outs R600_Reg128:$dst),
875 (ins R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2),
876 "MULLIT $dst, $src0, $src1",
877 [(set R600_Reg128:$dst, (int_AMDGPU_mullit R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2))]
880 let usesCustomInserter = 1, isPseudo = 1 in {
882 class R600PreloadInst <string asm, Intrinsic intr> : AMDGPUInst <
883 (outs R600_TReg32:$dst),
886 [(set R600_TReg32:$dst, (intr))]
889 def TGID_X : R600PreloadInst <"TGID_X", int_r600_read_tgid_x>;
890 def TGID_Y : R600PreloadInst <"TGID_Y", int_r600_read_tgid_y>;
891 def TGID_Z : R600PreloadInst <"TGID_Z", int_r600_read_tgid_z>;
893 def TIDIG_X : R600PreloadInst <"TIDIG_X", int_r600_read_tidig_x>;
894 def TIDIG_Y : R600PreloadInst <"TIDIG_Y", int_r600_read_tidig_y>;
895 def TIDIG_Z : R600PreloadInst <"TIDIG_Z", int_r600_read_tidig_z>;
897 def NGROUPS_X : R600PreloadInst <"NGROUPS_X", int_r600_read_ngroups_x>;
898 def NGROUPS_Y : R600PreloadInst <"NGROUPS_Y", int_r600_read_ngroups_y>;
899 def NGROUPS_Z : R600PreloadInst <"NGROUPS_Z", int_r600_read_ngroups_z>;
901 def GLOBAL_SIZE_X : R600PreloadInst <"GLOBAL_SIZE_X",
902 int_r600_read_global_size_x>;
903 def GLOBAL_SIZE_Y : R600PreloadInst <"GLOBAL_SIZE_Y",
904 int_r600_read_global_size_y>;
905 def GLOBAL_SIZE_Z : R600PreloadInst <"GLOBAL_SIZE_Z",
906 int_r600_read_global_size_z>;
908 def LOCAL_SIZE_X : R600PreloadInst <"LOCAL_SIZE_X",
909 int_r600_read_local_size_x>;
910 def LOCAL_SIZE_Y : R600PreloadInst <"LOCAL_SIZE_Y",
911 int_r600_read_local_size_y>;
912 def LOCAL_SIZE_Z : R600PreloadInst <"LOCAL_SIZE_Z",
913 int_r600_read_local_size_z>;
915 } // End usesCustomInserter = 1, isPseudo = 1
917 } // End isCodeGenOnly = 1
921 include "R600ShaderPatterns.td"
923 // We need this pattern to avoid having real registers in PHI nodes.
924 // For some reason this pattern only works when it comes after the other
927 (int_R600_load_input imm:$src),
928 (LOAD_INPUT imm:$src)
931 } // End isR600toCayman Predicate