1 //===-- EvergreenInstructions.td - EG Instruction defs ----*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // TableGen definitions for instructions which are:
11 // - Available to Evergreen and newer VLIW4/VLIW5 GPUs
12 // - Available only on Evergreen family GPUs.
14 //===----------------------------------------------------------------------===//
17 "Subtarget->getGeneration() >= AMDGPUSubtarget::EVERGREEN && "
18 "!Subtarget->hasCaymanISA()"
21 def isEGorCayman : Predicate<
22 "Subtarget->getGeneration() == AMDGPUSubtarget::EVERGREEN ||"
23 "Subtarget->getGeneration() == AMDGPUSubtarget::NORTHERN_ISLANDS"
26 class EGPat<dag pattern, dag result> : AMDGPUPat<pattern, result> {
27 let SubtargetPredicate = isEG;
30 class EGOrCaymanPat<dag pattern, dag result> : AMDGPUPat<pattern, result> {
31 let SubtargetPredicate = isEGorCayman;
34 //===----------------------------------------------------------------------===//
35 // Evergreen / Cayman store instructions
36 //===----------------------------------------------------------------------===//
38 let SubtargetPredicate = isEGorCayman in {
40 class CF_MEM_RAT_CACHELESS <bits<6> rat_inst, bits<4> rat_id, bits<4> mask, dag ins,
41 string name, list<dag> pattern>
42 : EG_CF_RAT <0x57, rat_inst, rat_id, mask, (outs), ins,
43 "MEM_RAT_CACHELESS "#name, pattern>;
45 class CF_MEM_RAT <bits<6> rat_inst, bits<4> rat_id, bits<4> mask, dag ins,
46 dag outs, string name, list<dag> pattern>
47 : EG_CF_RAT <0x56, rat_inst, rat_id, mask, outs, ins,
48 "MEM_RAT "#name, pattern>;
50 class CF_MEM_RAT_STORE_TYPED<bits<1> has_eop>
51 : CF_MEM_RAT <0x1, ?, 0xf, (ins R600_Reg128:$rw_gpr, R600_Reg128:$index_gpr,
52 i32imm:$rat_id, InstFlag:$eop), (outs),
53 "STORE_TYPED RAT($rat_id) $rw_gpr, $index_gpr"
54 #!if(has_eop, ", $eop", ""),
55 [(int_r600_rat_store_typed R600_Reg128:$rw_gpr,
56 R600_Reg128:$index_gpr,
59 def RAT_MSKOR : CF_MEM_RAT <0x11, 0, 0xf,
60 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr), (outs),
61 "MSKOR $rw_gpr.XW, $index_gpr",
62 [(mskor_global v4i32:$rw_gpr, i32:$index_gpr)]
68 multiclass RAT_ATOMIC<bits<6> op_ret, bits<6> op_noret, string name> {
69 let Constraints = "$rw_gpr = $out_gpr", eop = 0, mayStore = 1 in {
70 def _RTN: CF_MEM_RAT <op_ret, 0, 0xf,
71 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr),
72 (outs R600_Reg128:$out_gpr),
73 name ## "_RTN" ## " $rw_gpr, $index_gpr", [] >;
74 def _NORET: CF_MEM_RAT <op_noret, 0, 0xf,
75 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr),
76 (outs R600_Reg128:$out_gpr),
77 name ## " $rw_gpr, $index_gpr", [] >;
81 // Swap no-ret is just store. Raw store to cached target
82 // can only store on dword, which exactly matches swap_no_ret.
83 defm RAT_ATOMIC_XCHG_INT : RAT_ATOMIC<1, 34, "ATOMIC_XCHG_INT">;
84 defm RAT_ATOMIC_CMPXCHG_INT : RAT_ATOMIC<4, 36, "ATOMIC_CMPXCHG_INT">;
85 defm RAT_ATOMIC_ADD : RAT_ATOMIC<7, 39, "ATOMIC_ADD">;
86 defm RAT_ATOMIC_SUB : RAT_ATOMIC<8, 40, "ATOMIC_SUB">;
87 defm RAT_ATOMIC_RSUB : RAT_ATOMIC<9, 41, "ATOMIC_RSUB">;
88 defm RAT_ATOMIC_MIN_INT : RAT_ATOMIC<10, 42, "ATOMIC_MIN_INT">;
89 defm RAT_ATOMIC_MIN_UINT : RAT_ATOMIC<11, 43, "ATOMIC_MIN_UINT">;
90 defm RAT_ATOMIC_MAX_INT : RAT_ATOMIC<12, 44, "ATOMIC_MAX_INT">;
91 defm RAT_ATOMIC_MAX_UINT : RAT_ATOMIC<13, 45, "ATOMIC_MAX_UINT">;
92 defm RAT_ATOMIC_AND : RAT_ATOMIC<14, 46, "ATOMIC_AND">;
93 defm RAT_ATOMIC_OR : RAT_ATOMIC<15, 47, "ATOMIC_OR">;
94 defm RAT_ATOMIC_XOR : RAT_ATOMIC<16, 48, "ATOMIC_XOR">;
95 defm RAT_ATOMIC_INC_UINT : RAT_ATOMIC<18, 50, "ATOMIC_INC_UINT">;
96 defm RAT_ATOMIC_DEC_UINT : RAT_ATOMIC<19, 51, "ATOMIC_DEC_UINT">;
98 } // End SubtargetPredicate = isEGorCayman
100 //===----------------------------------------------------------------------===//
101 // Evergreen Only instructions
102 //===----------------------------------------------------------------------===//
104 let SubtargetPredicate = isEG in {
106 def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
107 defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
109 def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
110 def MULHI_INT_eg : MULHI_INT_Common<0x90>;
111 def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
112 def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
113 def MULHI_UINT24_eg : MULHI_UINT24_Common<0xb2>;
115 def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
116 def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
117 def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
118 def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
119 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
120 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
121 def : RsqPat<RECIPSQRT_IEEE_eg, f32>;
122 def SIN_eg : SIN_Common<0x8D>;
123 def COS_eg : COS_Common<0x8E>;
125 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
126 def : EGPat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
127 } // End SubtargetPredicate = isEG
129 //===----------------------------------------------------------------------===//
130 // Memory read/write instructions
131 //===----------------------------------------------------------------------===//
133 let usesCustomInserter = 1 in {
136 def RAT_WRITE_CACHELESS_32_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x1,
137 (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
138 "STORE_RAW $rw_gpr, $index_gpr, $eop",
139 [(store_global i32:$rw_gpr, i32:$index_gpr)]
143 def RAT_WRITE_CACHELESS_64_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x3,
144 (ins R600_Reg64:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
145 "STORE_RAW $rw_gpr.XY, $index_gpr, $eop",
146 [(store_global v2i32:$rw_gpr, i32:$index_gpr)]
150 def RAT_WRITE_CACHELESS_128_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0xf,
151 (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
152 "STORE_RAW $rw_gpr.XYZW, $index_gpr, $eop",
153 [(store_global v4i32:$rw_gpr, i32:$index_gpr)]
156 def RAT_STORE_TYPED_eg: CF_MEM_RAT_STORE_TYPED<1>;
158 } // End usesCustomInserter = 1
160 class VTX_READ_eg <string name, dag outs>
161 : VTX_WORD0_eg, VTX_READ<name, outs, []> {
166 let FETCH_WHOLE_QUAD = 0;
168 // XXX: We can infer this field based on the SRC_GPR. This would allow us
169 // to store vertex addresses in any channel, not just X.
172 let Inst{31-0} = Word0;
176 : VTX_READ_eg <"VTX_READ_8 $dst_gpr, $src_gpr",
177 (outs R600_TReg32_X:$dst_gpr)> {
179 let MEGA_FETCH_COUNT = 1;
181 let DST_SEL_Y = 7; // Masked
182 let DST_SEL_Z = 7; // Masked
183 let DST_SEL_W = 7; // Masked
184 let DATA_FORMAT = 1; // FMT_8
188 : VTX_READ_eg <"VTX_READ_16 $dst_gpr, $src_gpr",
189 (outs R600_TReg32_X:$dst_gpr)> {
190 let MEGA_FETCH_COUNT = 2;
192 let DST_SEL_Y = 7; // Masked
193 let DST_SEL_Z = 7; // Masked
194 let DST_SEL_W = 7; // Masked
195 let DATA_FORMAT = 5; // FMT_16
200 : VTX_READ_eg <"VTX_READ_32 $dst_gpr, $src_gpr",
201 (outs R600_TReg32_X:$dst_gpr)> {
203 let MEGA_FETCH_COUNT = 4;
205 let DST_SEL_Y = 7; // Masked
206 let DST_SEL_Z = 7; // Masked
207 let DST_SEL_W = 7; // Masked
208 let DATA_FORMAT = 0xD; // COLOR_32
210 // This is not really necessary, but there were some GPU hangs that appeared
211 // to be caused by ALU instructions in the next instruction group that wrote
212 // to the $src_gpr registers of the VTX_READ.
214 // %t3_x = VTX_READ_PARAM_32_eg killed %t2_x, 24
216 //Adding this constraint prevents this from happening.
217 let Constraints = "$src_gpr.ptr = $dst_gpr";
221 : VTX_READ_eg <"VTX_READ_64 $dst_gpr.XY, $src_gpr",
222 (outs R600_Reg64:$dst_gpr)> {
224 let MEGA_FETCH_COUNT = 8;
229 let DATA_FORMAT = 0x1D; // COLOR_32_32
233 : VTX_READ_eg <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr",
234 (outs R600_Reg128:$dst_gpr)> {
236 let MEGA_FETCH_COUNT = 16;
241 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
243 // XXX: Need to force VTX_READ_128 instructions to write to the same register
244 // that holds its buffer address to avoid potential hangs. We can't use
245 // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst
246 // registers are different sizes.
249 //===----------------------------------------------------------------------===//
250 // VTX Read from parameter memory space
251 //===----------------------------------------------------------------------===//
252 def : EGPat<(i32:$dst_gpr (vtx_id3_az_extloadi8 ADDRVTX_READ:$src_gpr)),
253 (VTX_READ_8_eg MEMxi:$src_gpr, 3)>;
254 def : EGPat<(i32:$dst_gpr (vtx_id3_az_extloadi16 ADDRVTX_READ:$src_gpr)),
255 (VTX_READ_16_eg MEMxi:$src_gpr, 3)>;
256 def : EGPat<(i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)),
257 (VTX_READ_32_eg MEMxi:$src_gpr, 3)>;
258 def : EGPat<(v2i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)),
259 (VTX_READ_64_eg MEMxi:$src_gpr, 3)>;
260 def : EGPat<(v4i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)),
261 (VTX_READ_128_eg MEMxi:$src_gpr, 3)>;
263 //===----------------------------------------------------------------------===//
264 // VTX Read from constant memory space
265 //===----------------------------------------------------------------------===//
266 def : EGPat<(i32:$dst_gpr (vtx_id2_az_extloadi8 ADDRVTX_READ:$src_gpr)),
267 (VTX_READ_8_eg MEMxi:$src_gpr, 2)>;
268 def : EGPat<(i32:$dst_gpr (vtx_id2_az_extloadi16 ADDRVTX_READ:$src_gpr)),
269 (VTX_READ_16_eg MEMxi:$src_gpr, 2)>;
270 def : EGPat<(i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)),
271 (VTX_READ_32_eg MEMxi:$src_gpr, 2)>;
272 def : EGPat<(v2i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)),
273 (VTX_READ_64_eg MEMxi:$src_gpr, 2)>;
274 def : EGPat<(v4i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)),
275 (VTX_READ_128_eg MEMxi:$src_gpr, 2)>;
277 //===----------------------------------------------------------------------===//
278 // VTX Read from global memory space
279 //===----------------------------------------------------------------------===//
280 def : EGPat<(i32:$dst_gpr (vtx_id1_az_extloadi8 ADDRVTX_READ:$src_gpr)),
281 (VTX_READ_8_eg MEMxi:$src_gpr, 1)>;
282 def : EGPat<(i32:$dst_gpr (vtx_id1_az_extloadi16 ADDRVTX_READ:$src_gpr)),
283 (VTX_READ_16_eg MEMxi:$src_gpr, 1)>;
284 def : EGPat<(i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)),
285 (VTX_READ_32_eg MEMxi:$src_gpr, 1)>;
286 def : EGPat<(v2i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)),
287 (VTX_READ_64_eg MEMxi:$src_gpr, 1)>;
288 def : EGPat<(v4i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)),
289 (VTX_READ_128_eg MEMxi:$src_gpr, 1)>;
291 //===----------------------------------------------------------------------===//
292 // Evergreen / Cayman Instructions
293 //===----------------------------------------------------------------------===//
295 let SubtargetPredicate = isEGorCayman in {
297 multiclass AtomicPat<Instruction inst_ret, Instruction inst_noret,
298 SDPatternOperator node_ret, SDPatternOperator node_noret> {
299 // FIXME: Add _RTN version. We need per WI scratch location to store the old value
300 // EXTRACT_SUBREG here is dummy, we know the node has no uses
301 def : EGOrCaymanPat<(i32 (node_noret i32:$ptr, i32:$data)),
302 (EXTRACT_SUBREG (inst_noret
303 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)), $data, sub0), $ptr), sub1)>;
305 multiclass AtomicIncDecPat<Instruction inst_ret, Instruction inst_noret,
306 SDPatternOperator node_ret, SDPatternOperator node_noret, int C> {
307 // FIXME: Add _RTN version. We need per WI scratch location to store the old value
308 // EXTRACT_SUBREG here is dummy, we know the node has no uses
309 def : EGOrCaymanPat<(i32 (node_noret i32:$ptr, C)),
310 (EXTRACT_SUBREG (inst_noret
311 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)), (MOV_IMM_I32 -1), sub0), $ptr), sub1)>;
314 // CMPSWAP is pattern is special
315 // EXTRACT_SUBREG here is dummy, we know the node has no uses
316 // FIXME: Add _RTN version. We need per WI scratch location to store the old value
317 def : EGOrCaymanPat<(i32 (atomic_cmp_swap_global_noret i32:$ptr, i32:$cmp, i32:$data)),
318 (EXTRACT_SUBREG (RAT_ATOMIC_CMPXCHG_INT_NORET
320 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)), $cmp, sub3),
324 defm AtomicSwapPat : AtomicPat <RAT_ATOMIC_XCHG_INT_RTN,
325 RAT_ATOMIC_XCHG_INT_NORET,
326 atomic_swap_global_ret,
327 atomic_swap_global_noret>;
328 defm AtomicAddPat : AtomicPat <RAT_ATOMIC_ADD_RTN, RAT_ATOMIC_ADD_NORET,
329 atomic_add_global_ret, atomic_add_global_noret>;
330 defm AtomicSubPat : AtomicPat <RAT_ATOMIC_SUB_RTN, RAT_ATOMIC_SUB_NORET,
331 atomic_sub_global_ret, atomic_sub_global_noret>;
332 defm AtomicMinPat : AtomicPat <RAT_ATOMIC_MIN_INT_RTN,
333 RAT_ATOMIC_MIN_INT_NORET,
334 atomic_min_global_ret, atomic_min_global_noret>;
335 defm AtomicUMinPat : AtomicPat <RAT_ATOMIC_MIN_UINT_RTN,
336 RAT_ATOMIC_MIN_UINT_NORET,
337 atomic_umin_global_ret, atomic_umin_global_noret>;
338 defm AtomicMaxPat : AtomicPat <RAT_ATOMIC_MAX_INT_RTN,
339 RAT_ATOMIC_MAX_INT_NORET,
340 atomic_max_global_ret, atomic_max_global_noret>;
341 defm AtomicUMaxPat : AtomicPat <RAT_ATOMIC_MAX_UINT_RTN,
342 RAT_ATOMIC_MAX_UINT_NORET,
343 atomic_umax_global_ret, atomic_umax_global_noret>;
344 defm AtomicAndPat : AtomicPat <RAT_ATOMIC_AND_RTN, RAT_ATOMIC_AND_NORET,
345 atomic_and_global_ret, atomic_and_global_noret>;
346 defm AtomicOrPat : AtomicPat <RAT_ATOMIC_OR_RTN, RAT_ATOMIC_OR_NORET,
347 atomic_or_global_ret, atomic_or_global_noret>;
348 defm AtomicXorPat : AtomicPat <RAT_ATOMIC_XOR_RTN, RAT_ATOMIC_XOR_NORET,
349 atomic_xor_global_ret, atomic_xor_global_noret>;
350 defm AtomicIncAddPat : AtomicIncDecPat <RAT_ATOMIC_INC_UINT_RTN,
351 RAT_ATOMIC_INC_UINT_NORET,
352 atomic_add_global_ret,
353 atomic_add_global_noret, 1>;
354 defm AtomicIncSubPat : AtomicIncDecPat <RAT_ATOMIC_INC_UINT_RTN,
355 RAT_ATOMIC_INC_UINT_NORET,
356 atomic_sub_global_ret,
357 atomic_sub_global_noret, -1>;
358 defm AtomicDecAddPat : AtomicIncDecPat <RAT_ATOMIC_DEC_UINT_RTN,
359 RAT_ATOMIC_DEC_UINT_NORET,
360 atomic_add_global_ret,
361 atomic_add_global_noret, -1>;
362 defm AtomicDecSubPat : AtomicIncDecPat <RAT_ATOMIC_DEC_UINT_RTN,
363 RAT_ATOMIC_DEC_UINT_NORET,
364 atomic_sub_global_ret,
365 atomic_sub_global_noret, 1>;
367 // Should be predicated on FeatureFP64
368 // def FMA_64 : R600_3OP <
370 // [(set f64:$dst, (fma f64:$src0, f64:$src1, f64:$src2))]
373 // BFE_UINT - bit_extract, an optimization for mask and shift
378 // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
383 // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
384 // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
385 // (16, 8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
386 // (24, 8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
387 def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
388 [(set i32:$dst, (AMDGPUbfe_u32 i32:$src0, i32:$src1, i32:$src2))],
392 def BFE_INT_eg : R600_3OP <0x5, "BFE_INT",
393 [(set i32:$dst, (AMDGPUbfe_i32 i32:$src0, i32:$src1, i32:$src2))],
397 defm : BFEPattern <BFE_UINT_eg, BFE_INT_eg, MOV_IMM_I32>;
399 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT",
400 [(set i32:$dst, (AMDGPUbfi i32:$src0, i32:$src1, i32:$src2))],
404 def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i1)),
405 (BFE_INT_eg i32:$src, (i32 ZERO), (i32 ONE_INT))>;
406 def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i8)),
407 (BFE_INT_eg i32:$src, (i32 ZERO), (MOV_IMM_I32 8))>;
408 def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i16)),
409 (BFE_INT_eg i32:$src, (i32 ZERO), (MOV_IMM_I32 16))>;
411 defm : BFIPatterns <BFI_INT_eg, MOV_IMM_I32, R600_Reg64>;
413 def BFM_INT_eg : R600_2OP <0xA0, "BFM_INT",
414 [(set i32:$dst, (AMDGPUbfm i32:$src0, i32:$src1))],
418 def MULADD_UINT24_eg : R600_3OP <0x10, "MULADD_UINT24",
419 [(set i32:$dst, (AMDGPUmad_u24 i32:$src0, i32:$src1, i32:$src2))], VecALU
422 def : UMad24Pat<MULADD_UINT24_eg>;
424 def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
425 def : ROTRPattern <BIT_ALIGN_INT_eg>;
426 def MULADD_eg : MULADD_Common<0x14>;
427 def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
428 def FMA_eg : FMA_Common<0x7>;
429 def ASHR_eg : ASHR_Common<0x15>;
430 def LSHR_eg : LSHR_Common<0x16>;
431 def LSHL_eg : LSHL_Common<0x17>;
432 def CNDE_eg : CNDE_Common<0x19>;
433 def CNDGT_eg : CNDGT_Common<0x1A>;
434 def CNDGE_eg : CNDGE_Common<0x1B>;
435 def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
436 def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
437 def MUL_UINT24_eg : R600_2OP <0xB5, "MUL_UINT24",
438 [(set i32:$dst, (AMDGPUmul_u24 i32:$src0, i32:$src1))], VecALU
440 def DOT4_eg : DOT4_Common<0xBE>;
441 defm CUBE_eg : CUBE_Common<0xC0>;
444 def ADDC_UINT : R600_2OP_Helper <0x52, "ADDC_UINT", AMDGPUcarry>;
445 def SUBB_UINT : R600_2OP_Helper <0x53, "SUBB_UINT", AMDGPUborrow>;
447 def FLT32_TO_FLT16 : R600_1OP_Helper <0xA2, "FLT32_TO_FLT16", AMDGPUfp_to_f16, VecALU>;
448 def FLT16_TO_FLT32 : R600_1OP_Helper <0xA3, "FLT16_TO_FLT32", f16_to_fp, VecALU>;
449 def BCNT_INT : R600_1OP_Helper <0xAA, "BCNT_INT", ctpop, VecALU>;
450 def FFBH_UINT : R600_1OP_Helper <0xAB, "FFBH_UINT", AMDGPUffbh_u32, VecALU>;
451 def FFBL_INT : R600_1OP_Helper <0xAC, "FFBL_INT", AMDGPUffbl_b32, VecALU>;
453 let hasSideEffects = 1 in {
454 def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", [], VecALU>;
457 def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
459 let Itinerary = AnyALU;
462 def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
464 def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
468 def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
470 def GROUP_BARRIER : InstR600 <
471 (outs), (ins), " GROUP_BARRIER", [(int_r600_group_barrier)], AnyALU>,
473 R600ALU_Word1_OP2 <0x54> {
489 let bank_swizzle = 0;
491 let update_exec_mask = 0;
494 let Inst{31-0} = Word0;
495 let Inst{63-32} = Word1;
500 //===----------------------------------------------------------------------===//
502 //===----------------------------------------------------------------------===//
503 class R600_LDS <bits<6> op, dag outs, dag ins, string asm,
504 list<dag> pattern = []> :
506 InstR600 <outs, ins, asm, pattern, XALU>,
513 let Word1{27} = offset{0};
514 let Word1{12} = offset{1};
515 let Word1{28} = offset{2};
516 let Word1{31} = offset{3};
517 let Word0{12} = offset{4};
518 let Word0{25} = offset{5};
521 let Inst{31-0} = Word0;
522 let Inst{63-32} = Word1;
525 let HasNativeOperands = 1;
526 let UseNamedOperandTable = 1;
529 class R600_LDS_1A <bits<6> lds_op, string name, list<dag> pattern> : R600_LDS <
531 (outs R600_Reg32:$dst),
532 (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
533 LAST:$last, R600_Pred:$pred_sel,
534 BANK_SWIZZLE:$bank_swizzle),
535 " "#name#" $last OQAP, $src0$src0_rel $pred_sel",
544 let usesCustomInserter = 1;
546 let DisableEncoding = "$dst";
549 class R600_LDS_1A1D <bits<6> lds_op, dag outs, string name, list<dag> pattern,
553 (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
554 R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel,
555 LAST:$last, R600_Pred:$pred_sel,
556 BANK_SWIZZLE:$bank_swizzle),
557 " "#name#" $last "#dst#"$src0$src0_rel, $src1$src1_rel, $pred_sel",
568 class R600_LDS_1A1D_NORET <bits<6> lds_op, string name, list<dag> pattern> :
569 R600_LDS_1A1D <lds_op, (outs), name, pattern> {
573 class R600_LDS_1A1D_RET <bits<6> lds_op, string name, list<dag> pattern> :
574 R600_LDS_1A1D <lds_op, (outs R600_Reg32:$dst), name##"_RET", pattern, "OQAP, "> {
577 let usesCustomInserter = 1;
578 let DisableEncoding = "$dst";
581 class R600_LDS_1A2D <bits<6> lds_op, dag outs, string name, list<dag> pattern,
585 (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
586 R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel,
587 R600_Reg32:$src2, REL:$src2_rel, SEL:$src2_sel,
588 LAST:$last, R600_Pred:$pred_sel, BANK_SWIZZLE:$bank_swizzle),
589 " "#name# "$last "#dst#"$src0$src0_rel, $src1$src1_rel, $src2$src2_rel, $pred_sel",
598 class R600_LDS_1A2D_NORET <bits<6> lds_op, string name, list<dag> pattern> :
599 R600_LDS_1A2D <lds_op, (outs), name, pattern> {
603 class R600_LDS_1A2D_RET <bits<6> lds_op, string name, list<dag> pattern> :
604 R600_LDS_1A2D <lds_op, (outs R600_Reg32:$dst), name, pattern> {
607 let usesCustomInserter = 1;
608 let DisableEncoding = "$dst";
611 def LDS_ADD : R600_LDS_1A1D_NORET <0x0, "LDS_ADD", [] >;
612 def LDS_SUB : R600_LDS_1A1D_NORET <0x1, "LDS_SUB", [] >;
613 def LDS_AND : R600_LDS_1A1D_NORET <0x9, "LDS_AND", [] >;
614 def LDS_OR : R600_LDS_1A1D_NORET <0xa, "LDS_OR", [] >;
615 def LDS_XOR : R600_LDS_1A1D_NORET <0xb, "LDS_XOR", [] >;
616 def LDS_WRXCHG: R600_LDS_1A1D_NORET <0xd, "LDS_WRXCHG", [] >;
617 def LDS_CMPST: R600_LDS_1A2D_NORET <0x10, "LDS_CMPST", [] >;
618 def LDS_MIN_INT : R600_LDS_1A1D_NORET <0x5, "LDS_MIN_INT", [] >;
619 def LDS_MAX_INT : R600_LDS_1A1D_NORET <0x6, "LDS_MAX_INT", [] >;
620 def LDS_MIN_UINT : R600_LDS_1A1D_NORET <0x7, "LDS_MIN_UINT", [] >;
621 def LDS_MAX_UINT : R600_LDS_1A1D_NORET <0x8, "LDS_MAX_UINT", [] >;
622 def LDS_WRITE : R600_LDS_1A1D_NORET <0xD, "LDS_WRITE",
623 [(store_local (i32 R600_Reg32:$src1), R600_Reg32:$src0)]
625 def LDS_BYTE_WRITE : R600_LDS_1A1D_NORET<0x12, "LDS_BYTE_WRITE",
626 [(truncstorei8_local i32:$src1, i32:$src0)]
628 def LDS_SHORT_WRITE : R600_LDS_1A1D_NORET<0x13, "LDS_SHORT_WRITE",
629 [(truncstorei16_local i32:$src1, i32:$src0)]
631 def LDS_ADD_RET : R600_LDS_1A1D_RET <0x20, "LDS_ADD",
632 [(set i32:$dst, (atomic_load_add_local i32:$src0, i32:$src1))]
634 def LDS_SUB_RET : R600_LDS_1A1D_RET <0x21, "LDS_SUB",
635 [(set i32:$dst, (atomic_load_sub_local i32:$src0, i32:$src1))]
637 def LDS_AND_RET : R600_LDS_1A1D_RET <0x29, "LDS_AND",
638 [(set i32:$dst, (atomic_load_and_local i32:$src0, i32:$src1))]
640 def LDS_OR_RET : R600_LDS_1A1D_RET <0x2a, "LDS_OR",
641 [(set i32:$dst, (atomic_load_or_local i32:$src0, i32:$src1))]
643 def LDS_XOR_RET : R600_LDS_1A1D_RET <0x2b, "LDS_XOR",
644 [(set i32:$dst, (atomic_load_xor_local i32:$src0, i32:$src1))]
646 def LDS_MIN_INT_RET : R600_LDS_1A1D_RET <0x25, "LDS_MIN_INT",
647 [(set i32:$dst, (atomic_load_min_local i32:$src0, i32:$src1))]
649 def LDS_MAX_INT_RET : R600_LDS_1A1D_RET <0x26, "LDS_MAX_INT",
650 [(set i32:$dst, (atomic_load_max_local i32:$src0, i32:$src1))]
652 def LDS_MIN_UINT_RET : R600_LDS_1A1D_RET <0x27, "LDS_MIN_UINT",
653 [(set i32:$dst, (atomic_load_umin_local i32:$src0, i32:$src1))]
655 def LDS_MAX_UINT_RET : R600_LDS_1A1D_RET <0x28, "LDS_MAX_UINT",
656 [(set i32:$dst, (atomic_load_umax_local i32:$src0, i32:$src1))]
658 def LDS_WRXCHG_RET : R600_LDS_1A1D_RET <0x2d, "LDS_WRXCHG",
659 [(set i32:$dst, (atomic_swap_local i32:$src0, i32:$src1))]
661 def LDS_CMPST_RET : R600_LDS_1A2D_RET <0x30, "LDS_CMPST",
662 [(set i32:$dst, (atomic_cmp_swap_local i32:$src0, i32:$src1, i32:$src2))]
664 def LDS_READ_RET : R600_LDS_1A <0x32, "LDS_READ_RET",
665 [(set (i32 R600_Reg32:$dst), (load_local R600_Reg32:$src0))]
667 def LDS_BYTE_READ_RET : R600_LDS_1A <0x36, "LDS_BYTE_READ_RET",
668 [(set i32:$dst, (sextloadi8_local i32:$src0))]
670 def LDS_UBYTE_READ_RET : R600_LDS_1A <0x37, "LDS_UBYTE_READ_RET",
671 [(set i32:$dst, (az_extloadi8_local i32:$src0))]
673 def LDS_SHORT_READ_RET : R600_LDS_1A <0x38, "LDS_SHORT_READ_RET",
674 [(set i32:$dst, (sextloadi16_local i32:$src0))]
676 def LDS_USHORT_READ_RET : R600_LDS_1A <0x39, "LDS_USHORT_READ_RET",
677 [(set i32:$dst, (az_extloadi16_local i32:$src0))]
680 // TRUNC is used for the FLT_TO_INT instructions to work around a
681 // perceived problem where the rounding modes are applied differently
682 // depending on the instruction and the slot they are in.
684 // https://bugs.freedesktop.org/show_bug.cgi?id=50232
685 // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
687 // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
688 // which do not need to be truncated since the fp values are 0.0f or 1.0f.
689 // We should look into handling these cases separately.
690 def : EGOrCaymanPat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
692 def : EGOrCaymanPat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
695 defm : SHA256MaPattern <BFI_INT_eg, XOR_INT, R600_Reg64>;
697 def EG_ExportSwz : ExportSwzInst {
698 let Word1{19-16} = 0; // BURST_COUNT
699 let Word1{20} = 0; // VALID_PIXEL_MODE
701 let Word1{29-22} = inst;
702 let Word1{30} = 0; // MARK
703 let Word1{31} = 1; // BARRIER
705 defm : ExportPattern<EG_ExportSwz, 83>;
707 def EG_ExportBuf : ExportBufInst {
708 let Word1{19-16} = 0; // BURST_COUNT
709 let Word1{20} = 0; // VALID_PIXEL_MODE
711 let Word1{29-22} = inst;
712 let Word1{30} = 0; // MARK
713 let Word1{31} = 1; // BARRIER
715 defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
717 def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
718 "TEX $COUNT @$ADDR"> {
721 def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
722 "VTX $COUNT @$ADDR"> {
725 def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
726 "LOOP_START_DX10 @$ADDR"> {
730 def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
734 def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
735 "LOOP_BREAK @$ADDR"> {
739 def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
744 def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
745 "JUMP @$ADDR POP:$POP_COUNT"> {
748 def CF_PUSH_EG : CF_CLAUSE_EG<11, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
749 "PUSH @$ADDR POP:$POP_COUNT"> {
752 def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
753 "ELSE @$ADDR POP:$POP_COUNT"> {
756 def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
761 def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
762 "POP @$ADDR POP:$POP_COUNT"> {
765 def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
769 let END_OF_PROGRAM = 1;
772 } // End Predicates = [isEGorCayman]