xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/EvergreenInstructions.td (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1//===-- EvergreenInstructions.td - EG Instruction defs  ----*- tablegen -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// TableGen definitions for instructions which are:
10// - Available to Evergreen and newer VLIW4/VLIW5 GPUs
11// - Available only on Evergreen family GPUs.
12//
13//===----------------------------------------------------------------------===//
14
15def isEG : Predicate<
16  "Subtarget->getGeneration() >= AMDGPUSubtarget::EVERGREEN && "
17  "!Subtarget->hasCaymanISA()"
18>;
19
20def isEGorCayman : Predicate<
21  "Subtarget->getGeneration() == AMDGPUSubtarget::EVERGREEN ||"
22  "Subtarget->getGeneration() == AMDGPUSubtarget::NORTHERN_ISLANDS"
23>;
24
25class EGPat<dag pattern, dag result> : AMDGPUPat<pattern, result> {
26  let SubtargetPredicate = isEG;
27}
28
29class EGOrCaymanPat<dag pattern, dag result> : AMDGPUPat<pattern, result> {
30  let SubtargetPredicate = isEGorCayman;
31}
32
33def IMMZeroBasedBitfieldMask : ImmLeaf <i32, [{
34  return isMask_32(Imm);
35}]>;
36
37def IMMPopCount : SDNodeXForm<imm, [{
38  return CurDAG->getTargetConstant(llvm::popcount(N->getZExtValue()), SDLoc(N),
39                                   MVT::i32);
40}]>;
41
42//===----------------------------------------------------------------------===//
43// Evergreen / Cayman store instructions
44//===----------------------------------------------------------------------===//
45
46let SubtargetPredicate = isEGorCayman in {
47
48class CF_MEM_RAT_CACHELESS <bits<6> rat_inst, bits<4> rat_id, bits<4> mask, dag ins,
49                           string name, list<dag> pattern>
50    : EG_CF_RAT <0x57, rat_inst, rat_id, mask, (outs), ins,
51                 "MEM_RAT_CACHELESS "#name, pattern>;
52
53class CF_MEM_RAT <bits<6> rat_inst, bits<4> rat_id, bits<4> mask, dag ins,
54                  dag outs, string name, list<dag> pattern>
55    : EG_CF_RAT <0x56, rat_inst, rat_id, mask, outs, ins,
56                 "MEM_RAT "#name, pattern>;
57
58class CF_MEM_RAT_STORE_TYPED<bits<1> has_eop>
59    : CF_MEM_RAT <0x1, ?, 0xf, (ins R600_Reg128:$rw_gpr, R600_Reg128:$index_gpr,
60                           i32imm:$rat_id, InstFlag:$eop), (outs),
61                  "STORE_TYPED RAT($rat_id) $rw_gpr, $index_gpr"
62                               #!if(has_eop, ", $eop", ""),
63                  [(int_r600_rat_store_typed R600_Reg128:$rw_gpr,
64                                             R600_Reg128:$index_gpr,
65                                             (i32 imm:$rat_id))]>;
66
67def RAT_MSKOR : CF_MEM_RAT <0x11, 0, 0xf,
68  (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr), (outs),
69  "MSKOR $rw_gpr.XW, $index_gpr",
70  [(mskor_global v4i32:$rw_gpr, i32:$index_gpr)]
71> {
72  let eop = 0;
73}
74
75
76multiclass RAT_ATOMIC<bits<6> op_ret, bits<6> op_noret, string name> {
77  let Constraints = "$rw_gpr = $out_gpr", eop = 0, mayStore = 1 in {
78  def  _RTN: CF_MEM_RAT <op_ret, 0, 0xf,
79             (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr),
80             (outs R600_Reg128:$out_gpr),
81             name # "_RTN $rw_gpr, $index_gpr", [] >;
82  def _NORET: CF_MEM_RAT <op_noret, 0, 0xf,
83              (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr),
84              (outs R600_Reg128:$out_gpr),
85              name # " $rw_gpr, $index_gpr", [] >;
86  }
87}
88
89// Swap no-ret is just store. Raw store to cached target
90// can only store on dword, which exactly matches swap_no_ret.
91defm RAT_ATOMIC_XCHG_INT : RAT_ATOMIC<1, 34, "ATOMIC_XCHG_INT">;
92defm RAT_ATOMIC_CMPXCHG_INT : RAT_ATOMIC<4, 36, "ATOMIC_CMPXCHG_INT">;
93defm RAT_ATOMIC_ADD : RAT_ATOMIC<7, 39, "ATOMIC_ADD">;
94defm RAT_ATOMIC_SUB : RAT_ATOMIC<8, 40, "ATOMIC_SUB">;
95defm RAT_ATOMIC_RSUB : RAT_ATOMIC<9, 41, "ATOMIC_RSUB">;
96defm RAT_ATOMIC_MIN_INT : RAT_ATOMIC<10, 42, "ATOMIC_MIN_INT">;
97defm RAT_ATOMIC_MIN_UINT : RAT_ATOMIC<11, 43, "ATOMIC_MIN_UINT">;
98defm RAT_ATOMIC_MAX_INT : RAT_ATOMIC<12, 44, "ATOMIC_MAX_INT">;
99defm RAT_ATOMIC_MAX_UINT : RAT_ATOMIC<13, 45, "ATOMIC_MAX_UINT">;
100defm RAT_ATOMIC_AND : RAT_ATOMIC<14, 46, "ATOMIC_AND">;
101defm RAT_ATOMIC_OR : RAT_ATOMIC<15, 47, "ATOMIC_OR">;
102defm RAT_ATOMIC_XOR : RAT_ATOMIC<16, 48, "ATOMIC_XOR">;
103defm RAT_ATOMIC_INC_UINT : RAT_ATOMIC<18, 50, "ATOMIC_INC_UINT">;
104defm RAT_ATOMIC_DEC_UINT : RAT_ATOMIC<19, 51, "ATOMIC_DEC_UINT">;
105
106} // End SubtargetPredicate = isEGorCayman
107
108//===----------------------------------------------------------------------===//
109// Evergreen Only instructions
110//===----------------------------------------------------------------------===//
111
112let SubtargetPredicate = isEG in {
113
114def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
115defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
116
117def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
118def MULHI_INT_eg : MULHI_INT_Common<0x90>;
119def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
120def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
121def MULHI_UINT24_eg : MULHI_UINT24_Common<0xb2>;
122
123def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
124def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
125def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
126def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
127def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
128def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
129def : SqrtPat<RECIPSQRT_IEEE_eg, RECIP_IEEE_eg>;
130
131def SIN_eg : SIN_Common<0x8D>;
132def COS_eg : COS_Common<0x8E>;
133
134def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
135} // End SubtargetPredicate = isEG
136
137//===----------------------------------------------------------------------===//
138// Memory read/write instructions
139//===----------------------------------------------------------------------===//
140
141let usesCustomInserter = 1 in {
142
143// 32-bit store
144def RAT_WRITE_CACHELESS_32_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x1,
145  (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
146  "STORE_RAW $rw_gpr, $index_gpr, $eop",
147  [(store_global i32:$rw_gpr, i32:$index_gpr)]
148>;
149
150// 64-bit store
151def RAT_WRITE_CACHELESS_64_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x3,
152  (ins R600_Reg64:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
153  "STORE_RAW $rw_gpr.XY, $index_gpr, $eop",
154  [(store_global v2i32:$rw_gpr, i32:$index_gpr)]
155>;
156
157//128-bit store
158def RAT_WRITE_CACHELESS_128_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0xf,
159  (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
160  "STORE_RAW $rw_gpr.XYZW, $index_gpr, $eop",
161  [(store_global v4i32:$rw_gpr, i32:$index_gpr)]
162>;
163
164def RAT_STORE_TYPED_eg: CF_MEM_RAT_STORE_TYPED<1>;
165
166} // End usesCustomInserter = 1
167
168class VTX_READ_eg <string name, dag outs>
169    : VTX_WORD0_eg, VTX_READ<name, outs, []> {
170
171  // Static fields
172  let VC_INST = 0;
173  let FETCH_TYPE = 2;
174  let FETCH_WHOLE_QUAD = 0;
175  let SRC_REL = 0;
176  // XXX: We can infer this field based on the SRC_GPR.  This would allow us
177  // to store vertex addresses in any channel, not just X.
178  let SRC_SEL_X = 0;
179
180  let Inst{31-0} = Word0;
181}
182
183def VTX_READ_8_eg
184    : VTX_READ_eg <"VTX_READ_8 $dst_gpr, $src_gpr",
185                   (outs R600_TReg32_X:$dst_gpr)> {
186
187  let MEGA_FETCH_COUNT = 1;
188  let DST_SEL_X = 0;
189  let DST_SEL_Y = 7;   // Masked
190  let DST_SEL_Z = 7;   // Masked
191  let DST_SEL_W = 7;   // Masked
192  let DATA_FORMAT = 1; // FMT_8
193}
194
195def VTX_READ_16_eg
196    : VTX_READ_eg <"VTX_READ_16 $dst_gpr, $src_gpr",
197                   (outs R600_TReg32_X:$dst_gpr)> {
198  let MEGA_FETCH_COUNT = 2;
199  let DST_SEL_X = 0;
200  let DST_SEL_Y = 7;   // Masked
201  let DST_SEL_Z = 7;   // Masked
202  let DST_SEL_W = 7;   // Masked
203  let DATA_FORMAT = 5; // FMT_16
204
205}
206
207def VTX_READ_32_eg
208    : VTX_READ_eg <"VTX_READ_32 $dst_gpr, $src_gpr",
209                   (outs R600_TReg32_X:$dst_gpr)> {
210
211  let MEGA_FETCH_COUNT = 4;
212  let DST_SEL_X        = 0;
213  let DST_SEL_Y        = 7;   // Masked
214  let DST_SEL_Z        = 7;   // Masked
215  let DST_SEL_W        = 7;   // Masked
216  let DATA_FORMAT      = 0xD; // COLOR_32
217
218  // This is not really necessary, but there were some GPU hangs that appeared
219  // to be caused by ALU instructions in the next instruction group that wrote
220  // to the $src_gpr registers of the VTX_READ.
221  // e.g.
222  // %t3_x = VTX_READ_PARAM_32_eg killed %t2_x, 24
223  // %t2_x = MOV %zero
224  //Adding this constraint prevents this from happening.
225  let Constraints = "$src_gpr.ptr = $dst_gpr";
226}
227
228def VTX_READ_64_eg
229    : VTX_READ_eg <"VTX_READ_64 $dst_gpr.XY, $src_gpr",
230                   (outs R600_Reg64:$dst_gpr)> {
231
232  let MEGA_FETCH_COUNT = 8;
233  let DST_SEL_X        = 0;
234  let DST_SEL_Y        = 1;
235  let DST_SEL_Z        = 7;
236  let DST_SEL_W        = 7;
237  let DATA_FORMAT      = 0x1D; // COLOR_32_32
238}
239
240def VTX_READ_128_eg
241    : VTX_READ_eg <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr",
242                   (outs R600_Reg128:$dst_gpr)> {
243
244  let MEGA_FETCH_COUNT = 16;
245  let DST_SEL_X        =  0;
246  let DST_SEL_Y        =  1;
247  let DST_SEL_Z        =  2;
248  let DST_SEL_W        =  3;
249  let DATA_FORMAT      =  0x22; // COLOR_32_32_32_32
250
251  // XXX: Need to force VTX_READ_128 instructions to write to the same register
252  // that holds its buffer address to avoid potential hangs.  We can't use
253  // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst
254  // registers are different sizes.
255}
256
257//===----------------------------------------------------------------------===//
258// VTX Read from parameter memory space
259//===----------------------------------------------------------------------===//
260def : EGPat<(i32:$dst_gpr (vtx_id3_az_extloadi8 ADDRVTX_READ:$src_gpr)),
261          (VTX_READ_8_eg MEMxi:$src_gpr, 3)>;
262def : EGPat<(i32:$dst_gpr (vtx_id3_az_extloadi16 ADDRVTX_READ:$src_gpr)),
263          (VTX_READ_16_eg MEMxi:$src_gpr, 3)>;
264def : EGPat<(i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)),
265          (VTX_READ_32_eg MEMxi:$src_gpr, 3)>;
266def : EGPat<(v2i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)),
267          (VTX_READ_64_eg MEMxi:$src_gpr, 3)>;
268def : EGPat<(v4i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)),
269          (VTX_READ_128_eg MEMxi:$src_gpr, 3)>;
270
271//===----------------------------------------------------------------------===//
272// VTX Read from constant memory space
273//===----------------------------------------------------------------------===//
274def : EGPat<(i32:$dst_gpr (vtx_id2_az_extloadi8 ADDRVTX_READ:$src_gpr)),
275          (VTX_READ_8_eg MEMxi:$src_gpr, 2)>;
276def : EGPat<(i32:$dst_gpr (vtx_id2_az_extloadi16 ADDRVTX_READ:$src_gpr)),
277          (VTX_READ_16_eg MEMxi:$src_gpr, 2)>;
278def : EGPat<(i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)),
279          (VTX_READ_32_eg MEMxi:$src_gpr, 2)>;
280def : EGPat<(v2i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)),
281          (VTX_READ_64_eg MEMxi:$src_gpr, 2)>;
282def : EGPat<(v4i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)),
283          (VTX_READ_128_eg MEMxi:$src_gpr, 2)>;
284
285//===----------------------------------------------------------------------===//
286// VTX Read from global memory space
287//===----------------------------------------------------------------------===//
288def : EGPat<(i32:$dst_gpr (vtx_id1_az_extloadi8 ADDRVTX_READ:$src_gpr)),
289          (VTX_READ_8_eg MEMxi:$src_gpr, 1)>;
290def : EGPat<(i32:$dst_gpr (vtx_id1_az_extloadi16 ADDRVTX_READ:$src_gpr)),
291          (VTX_READ_16_eg MEMxi:$src_gpr, 1)>;
292def : EGPat<(i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)),
293          (VTX_READ_32_eg MEMxi:$src_gpr, 1)>;
294def : EGPat<(v2i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)),
295          (VTX_READ_64_eg MEMxi:$src_gpr, 1)>;
296def : EGPat<(v4i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)),
297          (VTX_READ_128_eg MEMxi:$src_gpr, 1)>;
298
299//===----------------------------------------------------------------------===//
300// Evergreen / Cayman Instructions
301//===----------------------------------------------------------------------===//
302
303let SubtargetPredicate = isEGorCayman in {
304
305multiclass AtomicPat<Instruction inst_noret,
306                     SDPatternOperator node_noret> {
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:$ptr, i32:$data)),
310            (EXTRACT_SUBREG (inst_noret
311              (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)), $data, sub0), $ptr), sub1)>;
312}
313
314// CMPSWAP is pattern is special
315// EXTRACT_SUBREG here is dummy, we know the node has no uses
316// FIXME: Add _RTN version. We need per WI scratch location to store the old value
317def : EGOrCaymanPat<(i32 (atomic_cmp_swap_global_noret i32:$ptr, i32:$cmp, i32:$data)),
318          (EXTRACT_SUBREG (RAT_ATOMIC_CMPXCHG_INT_NORET
319            (INSERT_SUBREG
320              (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)), $cmp, sub3),
321            $data, sub0),
322          $ptr), sub1)>;
323
324defm AtomicSwapPat : AtomicPat <RAT_ATOMIC_XCHG_INT_NORET,
325                                atomic_swap_global_noret_i32>;
326defm AtomicAddPat : AtomicPat <RAT_ATOMIC_ADD_NORET,
327                               atomic_load_add_global_noret_i32>;
328defm AtomicSubPat : AtomicPat <RAT_ATOMIC_SUB_NORET,
329                               atomic_load_sub_global_noret_i32>;
330defm AtomicMinPat : AtomicPat <RAT_ATOMIC_MIN_INT_NORET,
331                               atomic_load_min_global_noret_i32>;
332defm AtomicUMinPat : AtomicPat <RAT_ATOMIC_MIN_UINT_NORET,
333                                atomic_load_umin_global_noret_i32>;
334defm AtomicMaxPat : AtomicPat <RAT_ATOMIC_MAX_INT_NORET,
335                               atomic_load_max_global_noret_i32>;
336defm AtomicUMaxPat : AtomicPat <RAT_ATOMIC_MAX_UINT_NORET,
337                                atomic_load_umax_global_noret_i32>;
338defm AtomicAndPat : AtomicPat <RAT_ATOMIC_AND_NORET,
339                               atomic_load_and_global_noret_i32>;
340defm AtomicOrPat : AtomicPat <RAT_ATOMIC_OR_NORET,
341                              atomic_load_or_global_noret_i32>;
342defm AtomicXorPat : AtomicPat <RAT_ATOMIC_XOR_NORET,
343                               atomic_load_xor_global_noret_i32>;
344
345// Should be predicated on FeatureFP64
346// def FMA_64 : R600_3OP <
347//   0xA, "FMA_64",
348//   [(set f64:$dst, (fma f64:$src0, f64:$src1, f64:$src2))]
349// >;
350
351// BFE_UINT - bit_extract, an optimization for mask and shift
352// Src0 = Input
353// Src1 = Offset
354// Src2 = Width
355//
356// bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
357//
358// Example Usage:
359// (Offset, Width)
360//
361// (0, 8)  = (Input << 24) >> 24 = (Input &  0xff)       >> 0
362// (8, 8)  = (Input << 16) >> 24 = (Input &  0xffff)     >> 8
363// (16, 8) = (Input <<  8) >> 24 = (Input &  0xffffff)   >> 16
364// (24, 8) = (Input <<  0) >> 24 = (Input &  0xffffffff) >> 24
365def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
366  [(set i32:$dst, (AMDGPUbfe_u32 i32:$src0, i32:$src1, i32:$src2))],
367  VecALU
368>;
369
370def BFE_INT_eg : R600_3OP <0x5, "BFE_INT",
371  [(set i32:$dst, (AMDGPUbfe_i32 i32:$src0, i32:$src1, i32:$src2))],
372  VecALU
373>;
374
375// Bitfield extract patterns
376
377def : AMDGPUPat <
378  (and (i32 (srl i32:$src, i32:$rshift)), IMMZeroBasedBitfieldMask:$mask),
379  (BFE_UINT_eg $src, $rshift, (MOV_IMM_I32 (i32 (IMMPopCount $mask))))
380>;
381
382// x & ((1 << y) - 1)
383def : AMDGPUPat <
384  (and i32:$src, (add_oneuse (shl_oneuse 1, i32:$width), -1)),
385  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
386>;
387
388// x & ~(-1 << y)
389def : AMDGPUPat <
390  (and i32:$src, (xor_oneuse (shl_oneuse -1, i32:$width), -1)),
391  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
392>;
393
394// x & (-1 >> (bitwidth - y))
395def : AMDGPUPat <
396  (and i32:$src, (srl_oneuse -1, (sub 32, i32:$width))),
397  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
398>;
399
400// x << (bitwidth - y) >> (bitwidth - y)
401def : AMDGPUPat <
402  (srl (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
403  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
404>;
405
406def : AMDGPUPat <
407  (sra (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
408  (BFE_INT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
409>;
410
411def BFI_INT_eg : R600_3OP <0x06, "BFI_INT",
412  [(set i32:$dst, (AMDGPUbfi i32:$src0, i32:$src1, i32:$src2))],
413  VecALU
414>;
415
416def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i1)),
417  (BFE_INT_eg i32:$src, (i32 ZERO), (i32 ONE_INT))>;
418def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i8)),
419  (BFE_INT_eg i32:$src, (i32 ZERO), (MOV_IMM_I32 8))>;
420def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i16)),
421  (BFE_INT_eg i32:$src, (i32 ZERO), (MOV_IMM_I32 16))>;
422
423// BFI patterns
424
425// Definition from ISA doc:
426// (y & x) | (z & ~x)
427def : AMDGPUPat <
428  (or (and i32:$y, i32:$x), (and i32:$z, (not i32:$x))),
429  (BFI_INT_eg $x, $y, $z)
430>;
431
432// 64-bit version
433def : AMDGPUPat <
434  (or (and i64:$y, i64:$x), (and i64:$z, (not i64:$x))),
435  (REG_SEQUENCE R600_Reg64,
436    (BFI_INT_eg (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub0)),
437                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub0)),
438                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub0))), sub0,
439    (BFI_INT_eg (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub1)),
440                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub1)),
441                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub1))), sub1)
442>;
443
444// SHA-256 Ch function
445// z ^ (x & (y ^ z))
446def : AMDGPUPat <
447  (xor i32:$z, (and i32:$x, (xor i32:$y, i32:$z))),
448  (BFI_INT_eg $x, $y, $z)
449>;
450
451// 64-bit version
452def : AMDGPUPat <
453  (xor i64:$z, (and i64:$x, (xor i64:$y, i64:$z))),
454  (REG_SEQUENCE R600_Reg64,
455    (BFI_INT_eg (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub0)),
456                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub0)),
457                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub0))), sub0,
458    (BFI_INT_eg (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub1)),
459                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub1)),
460                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub1))), sub1)
461>;
462
463def : AMDGPUPat <
464  (fcopysign f32:$src0, f32:$src1),
465  (BFI_INT_eg (MOV_IMM_I32 (i32 0x7fffffff)), $src0, $src1)
466>;
467
468def : AMDGPUPat <
469  (fcopysign f32:$src0, f64:$src1),
470  (BFI_INT_eg (MOV_IMM_I32 (i32 0x7fffffff)), $src0,
471              (i32 (EXTRACT_SUBREG R600_Reg64:$src1, sub1)))
472>;
473
474def : AMDGPUPat <
475  (fcopysign f64:$src0, f64:$src1),
476  (REG_SEQUENCE R600_Reg64,
477    (i32 (EXTRACT_SUBREG $src0, sub0)), sub0,
478    (BFI_INT_eg (MOV_IMM_I32 (i32 0x7fffffff)),
479                (i32 (EXTRACT_SUBREG R600_Reg64:$src0, sub1)),
480                (i32 (EXTRACT_SUBREG R600_Reg64:$src1, sub1))), sub1)
481>;
482
483def : AMDGPUPat <
484  (fcopysign f64:$src0, f32:$src1),
485  (REG_SEQUENCE R600_Reg64,
486    (i32 (EXTRACT_SUBREG $src0, sub0)), sub0,
487    (BFI_INT_eg (MOV_IMM_I32 (i32 0x7fffffff)),
488                (i32 (EXTRACT_SUBREG R600_Reg64:$src0, sub1)),
489                $src1), sub1)
490>;
491
492def BFM_INT_eg : R600_2OP <0xA0, "BFM_INT",
493  [(set i32:$dst, (AMDGPUbfm i32:$src0, i32:$src1))],
494  VecALU
495>;
496
497def MULADD_UINT24_eg : R600_3OP <0x10, "MULADD_UINT24",
498  [(set i32:$dst, (AMDGPUmad_u24 i32:$src0, i32:$src1, i32:$src2))], VecALU
499>;
500
501def : UMad24Pat<MULADD_UINT24_eg>;
502
503def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
504def : AMDGPUPat <
505  (fshr i32:$src0, i32:$src1, i32:$src2),
506  (BIT_ALIGN_INT_eg $src0, $src1, $src2)
507>;
508def : ROTRPattern <BIT_ALIGN_INT_eg>;
509def MULADD_eg : MULADD_Common<0x14>;
510def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
511def FMA_eg : FMA_Common<0x7>;
512def ASHR_eg : ASHR_Common<0x15>;
513def LSHR_eg : LSHR_Common<0x16>;
514def LSHL_eg : LSHL_Common<0x17>;
515def CNDE_eg : CNDE_Common<0x19>;
516def CNDGT_eg : CNDGT_Common<0x1A>;
517def CNDGE_eg : CNDGE_Common<0x1B>;
518def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
519def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
520def MUL_UINT24_eg : R600_2OP <0xB5, "MUL_UINT24",
521  [(set i32:$dst, (AMDGPUmul_u24 i32:$src0, i32:$src1))], VecALU
522>;
523def DOT4_eg : DOT4_Common<0xBE>;
524defm CUBE_eg : CUBE_Common<0xC0>;
525
526
527def ADDC_UINT : R600_2OP_Helper <0x52, "ADDC_UINT", AMDGPUcarry>;
528def SUBB_UINT : R600_2OP_Helper <0x53, "SUBB_UINT", AMDGPUborrow>;
529
530def FLT32_TO_FLT16 : R600_1OP_Helper <0xA2, "FLT32_TO_FLT16", AMDGPUfp_to_f16, VecALU>;
531def FLT16_TO_FLT32 : R600_1OP_Helper <0xA3, "FLT16_TO_FLT32", f16_to_fp, VecALU>;
532def BCNT_INT : R600_1OP_Helper <0xAA, "BCNT_INT", ctpop, VecALU>;
533def FFBH_UINT : R600_1OP_Helper <0xAB, "FFBH_UINT", AMDGPUffbh_u32, VecALU>;
534def FFBL_INT : R600_1OP_Helper <0xAC, "FFBL_INT", AMDGPUffbl_b32, VecALU>;
535
536let hasSideEffects = 1 in {
537  def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", [], VecALU>;
538}
539
540def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
541  let Pattern = [];
542  let Itinerary = AnyALU;
543}
544
545def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
546
547def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
548  let Pattern = [];
549}
550
551def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
552
553def GROUP_BARRIER : InstR600 <
554    (outs), (ins), "  GROUP_BARRIER", [(int_r600_group_barrier)], AnyALU>,
555    R600ALU_Word0,
556    R600ALU_Word1_OP2 <0x54> {
557
558  let dst = 0;
559  let dst_rel = 0;
560  let src0 = 0;
561  let src0_rel = 0;
562  let src0_neg = 0;
563  let src0_abs = 0;
564  let src1 = 0;
565  let src1_rel = 0;
566  let src1_neg = 0;
567  let src1_abs = 0;
568  let write = 0;
569  let omod = 0;
570  let clamp = 0;
571  let last = 1;
572  let bank_swizzle = 0;
573  let pred_sel = 0;
574  let update_exec_mask = 0;
575  let update_pred = 0;
576
577  let Inst{31-0}  = Word0;
578  let Inst{63-32} = Word1;
579
580  let ALUInst = 1;
581}
582
583//===----------------------------------------------------------------------===//
584// LDS Instructions
585//===----------------------------------------------------------------------===//
586class R600_LDS  <bits<6> op, dag outs, dag ins, string asm,
587                 list<dag> pattern = []> :
588
589    InstR600 <outs, ins, asm, pattern, XALU>,
590    R600_ALU_LDS_Word0,
591    R600LDS_Word1 {
592
593  bits<6>  offset = 0;
594  let lds_op = op;
595
596  let Word1{27} = offset{0};
597  let Word1{12} = offset{1};
598  let Word1{28} = offset{2};
599  let Word1{31} = offset{3};
600  let Word0{12} = offset{4};
601  let Word0{25} = offset{5};
602
603
604  let Inst{31-0}  = Word0;
605  let Inst{63-32} = Word1;
606
607  let ALUInst = 1;
608  let HasNativeOperands = 1;
609  let UseNamedOperandTable = 1;
610}
611
612class R600_LDS_1A <bits<6> lds_op, string name, list<dag> pattern> : R600_LDS <
613  lds_op,
614  (outs R600_Reg32:$dst),
615  (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
616       LAST:$last, R600_Pred:$pred_sel,
617       BANK_SWIZZLE:$bank_swizzle),
618  "  "#name#" $last OQAP, $src0$src0_rel $pred_sel",
619  pattern
620  > {
621
622  let src1 = 0;
623  let src1_rel = 0;
624  let src2 = 0;
625  let src2_rel = 0;
626
627  let usesCustomInserter = 1;
628  let LDS_1A = 1;
629  let DisableEncoding = "$dst";
630}
631
632class R600_LDS_1A1D <bits<6> lds_op, dag outs, string name, list<dag> pattern,
633                     string dst =""> :
634    R600_LDS <
635  lds_op, outs,
636  (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
637       R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel,
638       LAST:$last, R600_Pred:$pred_sel,
639       BANK_SWIZZLE:$bank_swizzle),
640  "  "#name#" $last "#dst#"$src0$src0_rel, $src1$src1_rel, $pred_sel",
641  pattern
642  > {
643
644  field string BaseOp;
645
646  let src2 = 0;
647  let src2_rel = 0;
648  let LDS_1A1D = 1;
649}
650
651class R600_LDS_1A1D_NORET <bits<6> lds_op, string name, list<dag> pattern> :
652    R600_LDS_1A1D <lds_op, (outs), name, pattern> {
653  let BaseOp = name;
654}
655
656class R600_LDS_1A1D_RET <bits<6> lds_op, string name, list<dag> pattern> :
657    R600_LDS_1A1D <lds_op,  (outs R600_Reg32:$dst), name#"_RET", pattern, "OQAP, "> {
658
659  let BaseOp = name;
660  let usesCustomInserter = 1;
661  let DisableEncoding = "$dst";
662}
663
664class R600_LDS_1A2D <bits<6> lds_op, dag outs, string name, list<dag> pattern,
665                     string dst =""> :
666    R600_LDS <
667  lds_op, outs,
668  (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
669       R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel,
670       R600_Reg32:$src2, REL:$src2_rel, SEL:$src2_sel,
671       LAST:$last, R600_Pred:$pred_sel, BANK_SWIZZLE:$bank_swizzle),
672  "  "#name# "$last "#dst#"$src0$src0_rel, $src1$src1_rel, $src2$src2_rel, $pred_sel",
673  pattern> {
674
675  field string BaseOp;
676
677  let LDS_1A1D = 0;
678  let LDS_1A2D = 1;
679}
680
681class R600_LDS_1A2D_NORET <bits<6> lds_op, string name, list<dag> pattern> :
682    R600_LDS_1A2D <lds_op, (outs), name, pattern> {
683  let BaseOp = name;
684}
685
686class R600_LDS_1A2D_RET <bits<6> lds_op, string name, list<dag> pattern> :
687    R600_LDS_1A2D <lds_op, (outs R600_Reg32:$dst), name, pattern> {
688
689  let BaseOp = name;
690  let usesCustomInserter = 1;
691  let DisableEncoding = "$dst";
692}
693
694def LDS_ADD : R600_LDS_1A1D_NORET <0x0, "LDS_ADD", [] >;
695def LDS_SUB : R600_LDS_1A1D_NORET <0x1, "LDS_SUB", [] >;
696def LDS_AND : R600_LDS_1A1D_NORET <0x9, "LDS_AND", [] >;
697def LDS_OR : R600_LDS_1A1D_NORET <0xa, "LDS_OR", [] >;
698def LDS_XOR : R600_LDS_1A1D_NORET <0xb, "LDS_XOR", [] >;
699def LDS_WRXCHG: R600_LDS_1A1D_NORET <0xd, "LDS_WRXCHG", [] >;
700def LDS_CMPST: R600_LDS_1A2D_NORET <0x10, "LDS_CMPST", [] >;
701def LDS_MIN_INT : R600_LDS_1A1D_NORET <0x5, "LDS_MIN_INT", [] >;
702def LDS_MAX_INT : R600_LDS_1A1D_NORET <0x6, "LDS_MAX_INT", [] >;
703def LDS_MIN_UINT : R600_LDS_1A1D_NORET <0x7, "LDS_MIN_UINT", [] >;
704def LDS_MAX_UINT : R600_LDS_1A1D_NORET <0x8, "LDS_MAX_UINT", [] >;
705def LDS_WRITE : R600_LDS_1A1D_NORET <0xD, "LDS_WRITE",
706  [(store_local (i32 R600_Reg32:$src1), R600_Reg32:$src0)]
707>;
708def LDS_BYTE_WRITE : R600_LDS_1A1D_NORET<0x12, "LDS_BYTE_WRITE",
709  [(truncstorei8_local i32:$src1, i32:$src0)]
710>;
711def LDS_SHORT_WRITE : R600_LDS_1A1D_NORET<0x13, "LDS_SHORT_WRITE",
712  [(truncstorei16_local i32:$src1, i32:$src0)]
713>;
714def LDS_ADD_RET : R600_LDS_1A1D_RET <0x20, "LDS_ADD",
715  [(set i32:$dst, (atomic_load_add_local_i32 i32:$src0, i32:$src1))]
716>;
717def LDS_SUB_RET : R600_LDS_1A1D_RET <0x21, "LDS_SUB",
718  [(set i32:$dst, (atomic_load_sub_local_i32 i32:$src0, i32:$src1))]
719>;
720def LDS_AND_RET : R600_LDS_1A1D_RET <0x29, "LDS_AND",
721  [(set i32:$dst, (atomic_load_and_local_i32 i32:$src0, i32:$src1))]
722>;
723def LDS_OR_RET : R600_LDS_1A1D_RET <0x2a, "LDS_OR",
724  [(set i32:$dst, (atomic_load_or_local_i32 i32:$src0, i32:$src1))]
725>;
726def LDS_XOR_RET : R600_LDS_1A1D_RET <0x2b, "LDS_XOR",
727  [(set i32:$dst, (atomic_load_xor_local_i32 i32:$src0, i32:$src1))]
728>;
729def LDS_MIN_INT_RET : R600_LDS_1A1D_RET <0x25, "LDS_MIN_INT",
730  [(set i32:$dst, (atomic_load_min_local_i32 i32:$src0, i32:$src1))]
731>;
732def LDS_MAX_INT_RET : R600_LDS_1A1D_RET <0x26, "LDS_MAX_INT",
733  [(set i32:$dst, (atomic_load_max_local_i32 i32:$src0, i32:$src1))]
734>;
735def LDS_MIN_UINT_RET : R600_LDS_1A1D_RET <0x27, "LDS_MIN_UINT",
736  [(set i32:$dst, (atomic_load_umin_local_i32 i32:$src0, i32:$src1))]
737>;
738def LDS_MAX_UINT_RET : R600_LDS_1A1D_RET <0x28, "LDS_MAX_UINT",
739  [(set i32:$dst, (atomic_load_umax_local_i32 i32:$src0, i32:$src1))]
740>;
741def LDS_WRXCHG_RET : R600_LDS_1A1D_RET <0x2d, "LDS_WRXCHG",
742  [(set i32:$dst, (atomic_swap_local_i32 i32:$src0, i32:$src1))]
743>;
744def LDS_CMPST_RET : R600_LDS_1A2D_RET <0x30, "LDS_CMPST",
745  [(set i32:$dst, (atomic_cmp_swap_local_i32 i32:$src0, i32:$src1, i32:$src2))]
746>;
747def LDS_READ_RET : R600_LDS_1A <0x32, "LDS_READ_RET",
748  [(set (i32 R600_Reg32:$dst), (load_local R600_Reg32:$src0))]
749>;
750def LDS_BYTE_READ_RET : R600_LDS_1A <0x36, "LDS_BYTE_READ_RET",
751  [(set i32:$dst, (sextloadi8_local i32:$src0))]
752>;
753def LDS_UBYTE_READ_RET : R600_LDS_1A <0x37, "LDS_UBYTE_READ_RET",
754  [(set i32:$dst, (az_extloadi8_local i32:$src0))]
755>;
756def LDS_SHORT_READ_RET : R600_LDS_1A <0x38, "LDS_SHORT_READ_RET",
757  [(set i32:$dst, (sextloadi16_local i32:$src0))]
758>;
759def LDS_USHORT_READ_RET : R600_LDS_1A <0x39, "LDS_USHORT_READ_RET",
760  [(set i32:$dst, (az_extloadi16_local i32:$src0))]
761>;
762
763// TRUNC is used for the FLT_TO_INT instructions to work around a
764// perceived problem where the rounding modes are applied differently
765// depending on the instruction and the slot they are in.
766// See:
767// https://bugs.freedesktop.org/show_bug.cgi?id=50232
768// Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
769//
770// XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
771// which do not need to be truncated since the fp values are 0.0f or 1.0f.
772// We should look into handling these cases separately.
773def : EGOrCaymanPat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
774
775def : EGOrCaymanPat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
776
777// SHA-256 Ma patterns
778
779// ((x & z) | (y & (x | z))) -> BFI (XOR x, y), z, y
780def : AMDGPUPat <
781  (or (and i32:$x, i32:$z), (and i32:$y, (or i32:$x, i32:$z))),
782  (BFI_INT_eg (XOR_INT i32:$x, i32:$y), i32:$z, i32:$y)
783>;
784
785def : AMDGPUPat <
786  (or (and i64:$x, i64:$z), (and i64:$y, (or i64:$x, i64:$z))),
787  (REG_SEQUENCE R600_Reg64,
788    (BFI_INT_eg (XOR_INT (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub0)),
789                     (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub0))),
790                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub0)),
791                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub0))), sub0,
792    (BFI_INT_eg (XOR_INT (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub1)),
793                     (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub1))),
794                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub1)),
795                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub1))), sub1)
796>;
797
798def EG_ExportSwz : ExportSwzInst {
799  let Word1{19-16} = 0; // BURST_COUNT
800  let Word1{20} = 0; // VALID_PIXEL_MODE
801  let Word1{21} = eop;
802  let Word1{29-22} = inst;
803  let Word1{30} = 0; // MARK
804  let Word1{31} = 1; // BARRIER
805}
806defm : ExportPattern<EG_ExportSwz, 83>;
807
808def EG_ExportBuf : ExportBufInst {
809  let Word1{19-16} = 0; // BURST_COUNT
810  let Word1{20} = 0; // VALID_PIXEL_MODE
811  let Word1{21} = eop;
812  let Word1{29-22} = inst;
813  let Word1{30} = 0; // MARK
814  let Word1{31} = 1; // BARRIER
815}
816defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
817
818def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
819  "TEX $COUNT @$ADDR"> {
820  let POP_COUNT = 0;
821}
822def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
823  "VTX $COUNT @$ADDR"> {
824  let POP_COUNT = 0;
825}
826def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
827  "LOOP_START_DX10 @$ADDR"> {
828  let POP_COUNT = 0;
829  let COUNT = 0;
830}
831def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
832  let POP_COUNT = 0;
833  let COUNT = 0;
834}
835def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
836  "LOOP_BREAK @$ADDR"> {
837  let POP_COUNT = 0;
838  let COUNT = 0;
839}
840def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
841  "CONTINUE @$ADDR"> {
842  let POP_COUNT = 0;
843  let COUNT = 0;
844}
845def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
846  "JUMP @$ADDR POP:$POP_COUNT"> {
847  let COUNT = 0;
848}
849def CF_PUSH_EG : CF_CLAUSE_EG<11, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
850                              "PUSH @$ADDR POP:$POP_COUNT"> {
851  let COUNT = 0;
852}
853def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
854  "ELSE @$ADDR POP:$POP_COUNT"> {
855  let COUNT = 0;
856}
857def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
858  let ADDR = 0;
859  let COUNT = 0;
860  let POP_COUNT = 0;
861}
862def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
863  "POP @$ADDR POP:$POP_COUNT"> {
864  let COUNT = 0;
865}
866def CF_END_EG :  CF_CLAUSE_EG<0, (ins), "CF_END"> {
867  let COUNT = 0;
868  let POP_COUNT = 0;
869  let ADDR = 0;
870  let END_OF_PROGRAM = 1;
871}
872
873} // End Predicates = [isEGorCayman]
874