1 //===- NVPTXInstrInfo.td - NVPTX Instruction defs -------------*- tblgen-*-===//
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 // This file describes the PTX instructions in TableGen format.
12 //===----------------------------------------------------------------------===//
14 include "NVPTXInstrFormats.td"
17 def NOP : NVPTXInst<(outs), (ins), "", []>;
19 // List of vector specific properties
20 def isVecLD : VecInstTypeEnum<1>;
21 def isVecST : VecInstTypeEnum<2>;
22 def isVecBuild : VecInstTypeEnum<3>;
23 def isVecShuffle : VecInstTypeEnum<4>;
24 def isVecExtract : VecInstTypeEnum<5>;
25 def isVecInsert : VecInstTypeEnum<6>;
26 def isVecDest : VecInstTypeEnum<7>;
27 def isVecOther : VecInstTypeEnum<15>;
29 //===----------------------------------------------------------------------===//
30 // NVPTX Operand Definitions.
31 //===----------------------------------------------------------------------===//
33 def brtarget : Operand<OtherVT>;
35 // CVT conversion modes
36 // These must match the enum in NVPTX.h
37 def CvtNONE : PatLeaf<(i32 0x0)>;
38 def CvtRNI : PatLeaf<(i32 0x1)>;
39 def CvtRZI : PatLeaf<(i32 0x2)>;
40 def CvtRMI : PatLeaf<(i32 0x3)>;
41 def CvtRPI : PatLeaf<(i32 0x4)>;
42 def CvtRN : PatLeaf<(i32 0x5)>;
43 def CvtRZ : PatLeaf<(i32 0x6)>;
44 def CvtRM : PatLeaf<(i32 0x7)>;
45 def CvtRP : PatLeaf<(i32 0x8)>;
47 def CvtNONE_FTZ : PatLeaf<(i32 0x10)>;
48 def CvtRNI_FTZ : PatLeaf<(i32 0x11)>;
49 def CvtRZI_FTZ : PatLeaf<(i32 0x12)>;
50 def CvtRMI_FTZ : PatLeaf<(i32 0x13)>;
51 def CvtRPI_FTZ : PatLeaf<(i32 0x14)>;
52 def CvtRN_FTZ : PatLeaf<(i32 0x15)>;
53 def CvtRZ_FTZ : PatLeaf<(i32 0x16)>;
54 def CvtRM_FTZ : PatLeaf<(i32 0x17)>;
55 def CvtRP_FTZ : PatLeaf<(i32 0x18)>;
57 def CvtSAT : PatLeaf<(i32 0x20)>;
58 def CvtSAT_FTZ : PatLeaf<(i32 0x30)>;
60 def CvtMode : Operand<i32> {
61 let PrintMethod = "printCvtMode";
65 // These must match the enum in NVPTX.h
66 def CmpEQ : PatLeaf<(i32 0)>;
67 def CmpNE : PatLeaf<(i32 1)>;
68 def CmpLT : PatLeaf<(i32 2)>;
69 def CmpLE : PatLeaf<(i32 3)>;
70 def CmpGT : PatLeaf<(i32 4)>;
71 def CmpGE : PatLeaf<(i32 5)>;
72 def CmpLO : PatLeaf<(i32 6)>;
73 def CmpLS : PatLeaf<(i32 7)>;
74 def CmpHI : PatLeaf<(i32 8)>;
75 def CmpHS : PatLeaf<(i32 9)>;
76 def CmpEQU : PatLeaf<(i32 10)>;
77 def CmpNEU : PatLeaf<(i32 11)>;
78 def CmpLTU : PatLeaf<(i32 12)>;
79 def CmpLEU : PatLeaf<(i32 13)>;
80 def CmpGTU : PatLeaf<(i32 14)>;
81 def CmpGEU : PatLeaf<(i32 15)>;
82 def CmpNUM : PatLeaf<(i32 16)>;
83 def CmpNAN : PatLeaf<(i32 17)>;
85 def CmpEQ_FTZ : PatLeaf<(i32 0x100)>;
86 def CmpNE_FTZ : PatLeaf<(i32 0x101)>;
87 def CmpLT_FTZ : PatLeaf<(i32 0x102)>;
88 def CmpLE_FTZ : PatLeaf<(i32 0x103)>;
89 def CmpGT_FTZ : PatLeaf<(i32 0x104)>;
90 def CmpGE_FTZ : PatLeaf<(i32 0x105)>;
91 def CmpLO_FTZ : PatLeaf<(i32 0x106)>;
92 def CmpLS_FTZ : PatLeaf<(i32 0x107)>;
93 def CmpHI_FTZ : PatLeaf<(i32 0x108)>;
94 def CmpHS_FTZ : PatLeaf<(i32 0x109)>;
95 def CmpEQU_FTZ : PatLeaf<(i32 0x10A)>;
96 def CmpNEU_FTZ : PatLeaf<(i32 0x10B)>;
97 def CmpLTU_FTZ : PatLeaf<(i32 0x10C)>;
98 def CmpLEU_FTZ : PatLeaf<(i32 0x10D)>;
99 def CmpGTU_FTZ : PatLeaf<(i32 0x10E)>;
100 def CmpGEU_FTZ : PatLeaf<(i32 0x10F)>;
101 def CmpNUM_FTZ : PatLeaf<(i32 0x110)>;
102 def CmpNAN_FTZ : PatLeaf<(i32 0x111)>;
104 def CmpMode : Operand<i32> {
105 let PrintMethod = "printCmpMode";
108 def F32ConstZero : Operand<f32>, PatLeaf<(f32 fpimm)>, SDNodeXForm<fpimm, [{
109 return CurDAG->getTargetConstantFP(0.0, MVT::f32);
111 def F32ConstOne : Operand<f32>, PatLeaf<(f32 fpimm)>, SDNodeXForm<fpimm, [{
112 return CurDAG->getTargetConstantFP(1.0, MVT::f32);
115 //===----------------------------------------------------------------------===//
116 // NVPTX Instruction Predicate Definitions
117 //===----------------------------------------------------------------------===//
120 def hasAtomRedG32 : Predicate<"Subtarget.hasAtomRedG32()">;
121 def hasAtomRedS32 : Predicate<"Subtarget.hasAtomRedS32()">;
122 def hasAtomRedGen32 : Predicate<"Subtarget.hasAtomRedGen32()">;
123 def useAtomRedG32forGen32 :
124 Predicate<"!Subtarget.hasAtomRedGen32() && Subtarget.hasAtomRedG32()">;
125 def hasBrkPt : Predicate<"Subtarget.hasBrkPt()">;
126 def hasAtomRedG64 : Predicate<"Subtarget.hasAtomRedG64()">;
127 def hasAtomRedS64 : Predicate<"Subtarget.hasAtomRedS64()">;
128 def hasAtomRedGen64 : Predicate<"Subtarget.hasAtomRedGen64()">;
129 def useAtomRedG64forGen64 :
130 Predicate<"!Subtarget.hasAtomRedGen64() && Subtarget.hasAtomRedG64()">;
131 def hasAtomAddF32 : Predicate<"Subtarget.hasAtomAddF32()">;
132 def hasVote : Predicate<"Subtarget.hasVote()">;
133 def hasDouble : Predicate<"Subtarget.hasDouble()">;
134 def reqPTX20 : Predicate<"Subtarget.reqPTX20()">;
135 def hasLDG : Predicate<"Subtarget.hasLDG()">;
136 def hasLDU : Predicate<"Subtarget.hasLDU()">;
137 def hasGenericLdSt : Predicate<"Subtarget.hasGenericLdSt()">;
139 def doF32FTZ : Predicate<"useF32FTZ()">;
140 def doNoF32FTZ : Predicate<"!useF32FTZ()">;
142 def doFMAF32 : Predicate<"doFMAF32">;
143 def doFMAF32_ftz : Predicate<"(doFMAF32 && useF32FTZ())">;
144 def doFMAF32AGG : Predicate<"doFMAF32AGG">;
145 def doFMAF32AGG_ftz : Predicate<"(doFMAF32AGG && useF32FTZ())">;
146 def doFMAF64 : Predicate<"doFMAF64">;
147 def doFMAF64AGG : Predicate<"doFMAF64AGG">;
149 def doMulWide : Predicate<"doMulWide">;
151 def allowFMA : Predicate<"allowFMA">;
152 def allowFMA_ftz : Predicate<"(allowFMA && useF32FTZ())">;
154 def do_DIVF32_APPROX : Predicate<"getDivF32Level()==0">;
155 def do_DIVF32_FULL : Predicate<"getDivF32Level()==1">;
157 def do_SQRTF32_APPROX : Predicate<"!usePrecSqrtF32()">;
158 def do_SQRTF32_RN : Predicate<"usePrecSqrtF32()">;
160 def hasHWROT32 : Predicate<"Subtarget.hasHWROT32()">;
162 def true : Predicate<"1">;
164 def hasPTX31 : Predicate<"Subtarget.getPTXVersion() >= 31">;
167 //===----------------------------------------------------------------------===//
168 // Some Common Instruction Class Templates
169 //===----------------------------------------------------------------------===//
171 multiclass I3<string OpcStr, SDNode OpNode> {
172 def i64rr : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, Int64Regs:$b),
173 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
174 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
176 def i64ri : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i64imm:$b),
177 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
178 [(set Int64Regs:$dst, (OpNode Int64Regs:$a, imm:$b))]>;
179 def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, Int32Regs:$b),
180 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
181 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
183 def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
184 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
185 [(set Int32Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
186 def i16rr : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
187 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
188 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
190 def i16ri : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
191 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
192 [(set Int16Regs:$dst, (OpNode Int16Regs:$a, (imm):$b))]>;
195 multiclass ADD_SUB_INT_32<string OpcStr, SDNode OpNode> {
196 def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a,
198 !strconcat(OpcStr, ".s32 \t$dst, $a, $b;"),
199 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
201 def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
202 !strconcat(OpcStr, ".s32 \t$dst, $a, $b;"),
203 [(set Int32Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
206 multiclass F3<string OpcStr, SDNode OpNode> {
207 def f64rr : NVPTXInst<(outs Float64Regs:$dst),
208 (ins Float64Regs:$a, Float64Regs:$b),
209 !strconcat(OpcStr, ".f64 \t$dst, $a, $b;"),
210 [(set Float64Regs:$dst,
211 (OpNode Float64Regs:$a, Float64Regs:$b))]>,
212 Requires<[allowFMA]>;
213 def f64ri : NVPTXInst<(outs Float64Regs:$dst),
214 (ins Float64Regs:$a, f64imm:$b),
215 !strconcat(OpcStr, ".f64 \t$dst, $a, $b;"),
216 [(set Float64Regs:$dst,
217 (OpNode Float64Regs:$a, fpimm:$b))]>,
218 Requires<[allowFMA]>;
219 def f32rr_ftz : NVPTXInst<(outs Float32Regs:$dst),
220 (ins Float32Regs:$a, Float32Regs:$b),
221 !strconcat(OpcStr, ".ftz.f32 \t$dst, $a, $b;"),
222 [(set Float32Regs:$dst,
223 (OpNode Float32Regs:$a, Float32Regs:$b))]>,
224 Requires<[allowFMA_ftz]>;
225 def f32ri_ftz : NVPTXInst<(outs Float32Regs:$dst),
226 (ins Float32Regs:$a, f32imm:$b),
227 !strconcat(OpcStr, ".ftz.f32 \t$dst, $a, $b;"),
228 [(set Float32Regs:$dst,
229 (OpNode Float32Regs:$a, fpimm:$b))]>,
230 Requires<[allowFMA_ftz]>;
231 def f32rr : NVPTXInst<(outs Float32Regs:$dst),
232 (ins Float32Regs:$a, Float32Regs:$b),
233 !strconcat(OpcStr, ".f32 \t$dst, $a, $b;"),
234 [(set Float32Regs:$dst,
235 (OpNode Float32Regs:$a, Float32Regs:$b))]>,
236 Requires<[allowFMA]>;
237 def f32ri : NVPTXInst<(outs Float32Regs:$dst),
238 (ins Float32Regs:$a, f32imm:$b),
239 !strconcat(OpcStr, ".f32 \t$dst, $a, $b;"),
240 [(set Float32Regs:$dst,
241 (OpNode Float32Regs:$a, fpimm:$b))]>,
242 Requires<[allowFMA]>;
245 multiclass F3_rn<string OpcStr, SDNode OpNode> {
246 def f64rr : NVPTXInst<(outs Float64Regs:$dst),
247 (ins Float64Regs:$a, Float64Regs:$b),
248 !strconcat(OpcStr, ".rn.f64 \t$dst, $a, $b;"),
249 [(set Float64Regs:$dst,
250 (OpNode Float64Regs:$a, Float64Regs:$b))]>;
251 def f64ri : NVPTXInst<(outs Float64Regs:$dst),
252 (ins Float64Regs:$a, f64imm:$b),
253 !strconcat(OpcStr, ".rn.f64 \t$dst, $a, $b;"),
254 [(set Float64Regs:$dst,
255 (OpNode Float64Regs:$a, fpimm:$b))]>;
256 def f32rr_ftz : NVPTXInst<(outs Float32Regs:$dst),
257 (ins Float32Regs:$a, Float32Regs:$b),
258 !strconcat(OpcStr, ".rn.ftz.f32 \t$dst, $a, $b;"),
259 [(set Float32Regs:$dst,
260 (OpNode Float32Regs:$a, Float32Regs:$b))]>,
261 Requires<[doF32FTZ]>;
262 def f32ri_ftz : NVPTXInst<(outs Float32Regs:$dst),
263 (ins Float32Regs:$a, f32imm:$b),
264 !strconcat(OpcStr, ".rn.ftz.f32 \t$dst, $a, $b;"),
265 [(set Float32Regs:$dst,
266 (OpNode Float32Regs:$a, fpimm:$b))]>,
267 Requires<[doF32FTZ]>;
268 def f32rr : NVPTXInst<(outs Float32Regs:$dst),
269 (ins Float32Regs:$a, Float32Regs:$b),
270 !strconcat(OpcStr, ".rn.f32 \t$dst, $a, $b;"),
271 [(set Float32Regs:$dst,
272 (OpNode Float32Regs:$a, Float32Regs:$b))]>;
273 def f32ri : NVPTXInst<(outs Float32Regs:$dst),
274 (ins Float32Regs:$a, f32imm:$b),
275 !strconcat(OpcStr, ".rn.f32 \t$dst, $a, $b;"),
276 [(set Float32Regs:$dst,
277 (OpNode Float32Regs:$a, fpimm:$b))]>;
280 multiclass F2<string OpcStr, SDNode OpNode> {
281 def f64 : NVPTXInst<(outs Float64Regs:$dst), (ins Float64Regs:$a),
282 !strconcat(OpcStr, ".f64 \t$dst, $a;"),
283 [(set Float64Regs:$dst, (OpNode Float64Regs:$a))]>;
284 def f32_ftz : NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$a),
285 !strconcat(OpcStr, ".ftz.f32 \t$dst, $a;"),
286 [(set Float32Regs:$dst, (OpNode Float32Regs:$a))]>,
287 Requires<[doF32FTZ]>;
288 def f32 : NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$a),
289 !strconcat(OpcStr, ".f32 \t$dst, $a;"),
290 [(set Float32Regs:$dst, (OpNode Float32Regs:$a))]>;
293 //===----------------------------------------------------------------------===//
294 // NVPTX Instructions.
295 //===----------------------------------------------------------------------===//
297 //-----------------------------------
298 // General Type Conversion
299 //-----------------------------------
301 let neverHasSideEffects = 1 in {
302 // Generate a cvt to the given type from all possible types.
303 // Each instance takes a CvtMode immediate that defines the conversion mode to
304 // use. It can be CvtNONE to omit a conversion mode.
305 multiclass CVT_FROM_ALL<string FromName, RegisterClass RC> {
306 def _s16 : NVPTXInst<(outs RC:$dst),
307 (ins Int16Regs:$src, CvtMode:$mode),
308 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
309 FromName, ".s16\t$dst, $src;"),
311 def _u16 : NVPTXInst<(outs RC:$dst),
312 (ins Int16Regs:$src, CvtMode:$mode),
313 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
314 FromName, ".u16\t$dst, $src;"),
316 def _f16 : NVPTXInst<(outs RC:$dst),
317 (ins Int16Regs:$src, CvtMode:$mode),
318 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
319 FromName, ".f16\t$dst, $src;"),
321 def _s32 : NVPTXInst<(outs RC:$dst),
322 (ins Int32Regs:$src, CvtMode:$mode),
323 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
324 FromName, ".s32\t$dst, $src;"),
326 def _u32 : NVPTXInst<(outs RC:$dst),
327 (ins Int32Regs:$src, CvtMode:$mode),
328 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
329 FromName, ".u32\t$dst, $src;"),
331 def _s64 : NVPTXInst<(outs RC:$dst),
332 (ins Int64Regs:$src, CvtMode:$mode),
333 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
334 FromName, ".s64\t$dst, $src;"),
336 def _u64 : NVPTXInst<(outs RC:$dst),
337 (ins Int64Regs:$src, CvtMode:$mode),
338 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
339 FromName, ".u64\t$dst, $src;"),
341 def _f32 : NVPTXInst<(outs RC:$dst),
342 (ins Float32Regs:$src, CvtMode:$mode),
343 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
344 FromName, ".f32\t$dst, $src;"),
346 def _f64 : NVPTXInst<(outs RC:$dst),
347 (ins Float64Regs:$src, CvtMode:$mode),
348 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
349 FromName, ".f64\t$dst, $src;"),
353 // Generate a cvt to all possible types.
354 defm CVT_s16 : CVT_FROM_ALL<"s16", Int16Regs>;
355 defm CVT_u16 : CVT_FROM_ALL<"u16", Int16Regs>;
356 defm CVT_f16 : CVT_FROM_ALL<"f16", Int16Regs>;
357 defm CVT_s32 : CVT_FROM_ALL<"s32", Int32Regs>;
358 defm CVT_u32 : CVT_FROM_ALL<"u32", Int32Regs>;
359 defm CVT_s64 : CVT_FROM_ALL<"s64", Int64Regs>;
360 defm CVT_u64 : CVT_FROM_ALL<"u64", Int64Regs>;
361 defm CVT_f32 : CVT_FROM_ALL<"f32", Float32Regs>;
362 defm CVT_f64 : CVT_FROM_ALL<"f64", Float64Regs>;
364 // This set of cvt is different from the above. The type of the source
365 // and target are the same.
367 def CVT_INREG_s16_s8 : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
368 "cvt.s16.s8 \t$dst, $src;", []>;
369 def CVT_INREG_s32_s8 : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
370 "cvt.s32.s8 \t$dst, $src;", []>;
371 def CVT_INREG_s32_s16 : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
372 "cvt.s32.s16 \t$dst, $src;", []>;
373 def CVT_INREG_s64_s8 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
374 "cvt.s64.s8 \t$dst, $src;", []>;
375 def CVT_INREG_s64_s16 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
376 "cvt.s64.s16 \t$dst, $src;", []>;
377 def CVT_INREG_s64_s32 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
378 "cvt.s64.s32 \t$dst, $src;", []>;
381 //-----------------------------------
382 // Integer Arithmetic
383 //-----------------------------------
385 multiclass ADD_SUB_i1<SDNode OpNode> {
386 def _rr: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$a, Int1Regs:$b),
387 "xor.pred \t$dst, $a, $b;",
388 [(set Int1Regs:$dst, (OpNode Int1Regs:$a, Int1Regs:$b))]>;
389 def _ri: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$a, i1imm:$b),
390 "xor.pred \t$dst, $a, $b;",
391 [(set Int1Regs:$dst, (OpNode Int1Regs:$a, (imm):$b))]>;
394 defm ADD_i1 : ADD_SUB_i1<add>;
395 defm SUB_i1 : ADD_SUB_i1<sub>;
398 defm ADD : I3<"add.s", add>;
399 defm SUB : I3<"sub.s", sub>;
401 defm ADDCC : ADD_SUB_INT_32<"add.cc", addc>;
402 defm SUBCC : ADD_SUB_INT_32<"sub.cc", subc>;
404 defm ADDCCC : ADD_SUB_INT_32<"addc.cc", adde>;
405 defm SUBCCC : ADD_SUB_INT_32<"subc.cc", sube>;
407 //mul.wide PTX instruction
408 def SInt32Const : PatLeaf<(imm), [{
409 const APInt &v = N->getAPIntValue();
410 if (v.isSignedIntN(32))
415 def UInt32Const : PatLeaf<(imm), [{
416 const APInt &v = N->getAPIntValue();
422 def SInt16Const : PatLeaf<(imm), [{
423 const APInt &v = N->getAPIntValue();
424 if (v.isSignedIntN(16))
429 def UInt16Const : PatLeaf<(imm), [{
430 const APInt &v = N->getAPIntValue();
436 def Int5Const : PatLeaf<(imm), [{
437 const APInt &v = N->getAPIntValue();
438 // Check if 0 <= v < 32
439 // Only then the result from (x << v) will be i32
440 if (v.sge(0) && v.slt(32))
445 def Int4Const : PatLeaf<(imm), [{
446 const APInt &v = N->getAPIntValue();
447 // Check if 0 <= v < 16
448 // Only then the result from (x << v) will be i16
449 if (v.sge(0) && v.slt(16))
454 def SHL2MUL32 : SDNodeXForm<imm, [{
455 const APInt &v = N->getAPIntValue();
457 return CurDAG->getTargetConstant(temp.shl(v), MVT::i32);
460 def SHL2MUL16 : SDNodeXForm<imm, [{
461 const APInt &v = N->getAPIntValue();
463 return CurDAG->getTargetConstant(temp.shl(v), MVT::i16);
466 def MULWIDES64 : NVPTXInst<(outs Int64Regs:$dst),
467 (ins Int32Regs:$a, Int32Regs:$b),
468 "mul.wide.s32 \t$dst, $a, $b;", []>;
469 def MULWIDES64Imm : NVPTXInst<(outs Int64Regs:$dst),
470 (ins Int32Regs:$a, i64imm:$b),
471 "mul.wide.s32 \t$dst, $a, $b;", []>;
473 def MULWIDEU64 : NVPTXInst<(outs Int64Regs:$dst),
474 (ins Int32Regs:$a, Int32Regs:$b),
475 "mul.wide.u32 \t$dst, $a, $b;", []>;
476 def MULWIDEU64Imm : NVPTXInst<(outs Int64Regs:$dst),
477 (ins Int32Regs:$a, i64imm:$b),
478 "mul.wide.u32 \t$dst, $a, $b;", []>;
480 def MULWIDES32 : NVPTXInst<(outs Int32Regs:$dst),
481 (ins Int16Regs:$a, Int16Regs:$b),
482 "mul.wide.s16 \t$dst, $a, $b;", []>;
483 def MULWIDES32Imm : NVPTXInst<(outs Int32Regs:$dst),
484 (ins Int16Regs:$a, i32imm:$b),
485 "mul.wide.s16 \t$dst, $a, $b;", []>;
487 def MULWIDEU32 : NVPTXInst<(outs Int32Regs:$dst),
488 (ins Int16Regs:$a, Int16Regs:$b),
489 "mul.wide.u16 \t$dst, $a, $b;", []>;
490 def MULWIDEU32Imm : NVPTXInst<(outs Int32Regs:$dst),
491 (ins Int16Regs:$a, i32imm:$b),
492 "mul.wide.u16 \t$dst, $a, $b;", []>;
494 def : Pat<(shl (sext Int32Regs:$a), (i32 Int5Const:$b)),
495 (MULWIDES64Imm Int32Regs:$a, (SHL2MUL32 node:$b))>,
496 Requires<[doMulWide]>;
497 def : Pat<(shl (zext Int32Regs:$a), (i32 Int5Const:$b)),
498 (MULWIDEU64Imm Int32Regs:$a, (SHL2MUL32 node:$b))>,
499 Requires<[doMulWide]>;
501 def : Pat<(shl (sext Int16Regs:$a), (i16 Int4Const:$b)),
502 (MULWIDES32Imm Int16Regs:$a, (SHL2MUL16 node:$b))>,
503 Requires<[doMulWide]>;
504 def : Pat<(shl (zext Int16Regs:$a), (i16 Int4Const:$b)),
505 (MULWIDEU32Imm Int16Regs:$a, (SHL2MUL16 node:$b))>,
506 Requires<[doMulWide]>;
508 def : Pat<(mul (sext Int32Regs:$a), (sext Int32Regs:$b)),
509 (MULWIDES64 Int32Regs:$a, Int32Regs:$b)>,
510 Requires<[doMulWide]>;
511 def : Pat<(mul (sext Int32Regs:$a), (i64 SInt32Const:$b)),
512 (MULWIDES64Imm Int32Regs:$a, (i64 SInt32Const:$b))>,
513 Requires<[doMulWide]>;
515 def : Pat<(mul (zext Int32Regs:$a), (zext Int32Regs:$b)),
516 (MULWIDEU64 Int32Regs:$a, Int32Regs:$b)>, Requires<[doMulWide]>;
517 def : Pat<(mul (zext Int32Regs:$a), (i64 UInt32Const:$b)),
518 (MULWIDEU64Imm Int32Regs:$a, (i64 UInt32Const:$b))>,
519 Requires<[doMulWide]>;
521 def : Pat<(mul (sext Int16Regs:$a), (sext Int16Regs:$b)),
522 (MULWIDES32 Int16Regs:$a, Int16Regs:$b)>, Requires<[doMulWide]>;
523 def : Pat<(mul (sext Int16Regs:$a), (i32 SInt16Const:$b)),
524 (MULWIDES32Imm Int16Regs:$a, (i32 SInt16Const:$b))>,
525 Requires<[doMulWide]>;
527 def : Pat<(mul (zext Int16Regs:$a), (zext Int16Regs:$b)),
528 (MULWIDEU32 Int16Regs:$a, Int16Regs:$b)>, Requires<[doMulWide]>;
529 def : Pat<(mul (zext Int16Regs:$a), (i32 UInt16Const:$b)),
530 (MULWIDEU32Imm Int16Regs:$a, (i32 UInt16Const:$b))>,
531 Requires<[doMulWide]>;
533 defm MULT : I3<"mul.lo.s", mul>;
535 defm MULTHS : I3<"mul.hi.s", mulhs>;
536 defm MULTHU : I3<"mul.hi.u", mulhu>;
538 defm SDIV : I3<"div.s", sdiv>;
539 defm UDIV : I3<"div.u", udiv>;
541 defm SREM : I3<"rem.s", srem>;
542 // The ri version will not be selected as DAGCombiner::visitSREM will lower it.
543 defm UREM : I3<"rem.u", urem>;
544 // The ri version will not be selected as DAGCombiner::visitUREM will lower it.
546 def MAD16rrr : NVPTXInst<(outs Int16Regs:$dst),
547 (ins Int16Regs:$a, Int16Regs:$b, Int16Regs:$c),
548 "mad.lo.s16 \t$dst, $a, $b, $c;",
549 [(set Int16Regs:$dst, (add
550 (mul Int16Regs:$a, Int16Regs:$b), Int16Regs:$c))]>;
551 def MAD16rri : NVPTXInst<(outs Int16Regs:$dst),
552 (ins Int16Regs:$a, Int16Regs:$b, i16imm:$c),
553 "mad.lo.s16 \t$dst, $a, $b, $c;",
554 [(set Int16Regs:$dst, (add
555 (mul Int16Regs:$a, Int16Regs:$b), imm:$c))]>;
556 def MAD16rir : NVPTXInst<(outs Int16Regs:$dst),
557 (ins Int16Regs:$a, i16imm:$b, Int16Regs:$c),
558 "mad.lo.s16 \t$dst, $a, $b, $c;",
559 [(set Int16Regs:$dst, (add
560 (mul Int16Regs:$a, imm:$b), Int16Regs:$c))]>;
561 def MAD16rii : NVPTXInst<(outs Int16Regs:$dst),
562 (ins Int16Regs:$a, i16imm:$b, i16imm:$c),
563 "mad.lo.s16 \t$dst, $a, $b, $c;",
564 [(set Int16Regs:$dst, (add (mul Int16Regs:$a, imm:$b),
567 def MAD32rrr : NVPTXInst<(outs Int32Regs:$dst),
568 (ins Int32Regs:$a, Int32Regs:$b, Int32Regs:$c),
569 "mad.lo.s32 \t$dst, $a, $b, $c;",
570 [(set Int32Regs:$dst, (add
571 (mul Int32Regs:$a, Int32Regs:$b), Int32Regs:$c))]>;
572 def MAD32rri : NVPTXInst<(outs Int32Regs:$dst),
573 (ins Int32Regs:$a, Int32Regs:$b, i32imm:$c),
574 "mad.lo.s32 \t$dst, $a, $b, $c;",
575 [(set Int32Regs:$dst, (add
576 (mul Int32Regs:$a, Int32Regs:$b), imm:$c))]>;
577 def MAD32rir : NVPTXInst<(outs Int32Regs:$dst),
578 (ins Int32Regs:$a, i32imm:$b, Int32Regs:$c),
579 "mad.lo.s32 \t$dst, $a, $b, $c;",
580 [(set Int32Regs:$dst, (add
581 (mul Int32Regs:$a, imm:$b), Int32Regs:$c))]>;
582 def MAD32rii : NVPTXInst<(outs Int32Regs:$dst),
583 (ins Int32Regs:$a, i32imm:$b, i32imm:$c),
584 "mad.lo.s32 \t$dst, $a, $b, $c;",
585 [(set Int32Regs:$dst, (add
586 (mul Int32Regs:$a, imm:$b), imm:$c))]>;
588 def MAD64rrr : NVPTXInst<(outs Int64Regs:$dst),
589 (ins Int64Regs:$a, Int64Regs:$b, Int64Regs:$c),
590 "mad.lo.s64 \t$dst, $a, $b, $c;",
591 [(set Int64Regs:$dst, (add
592 (mul Int64Regs:$a, Int64Regs:$b), Int64Regs:$c))]>;
593 def MAD64rri : NVPTXInst<(outs Int64Regs:$dst),
594 (ins Int64Regs:$a, Int64Regs:$b, i64imm:$c),
595 "mad.lo.s64 \t$dst, $a, $b, $c;",
596 [(set Int64Regs:$dst, (add
597 (mul Int64Regs:$a, Int64Regs:$b), imm:$c))]>;
598 def MAD64rir : NVPTXInst<(outs Int64Regs:$dst),
599 (ins Int64Regs:$a, i64imm:$b, Int64Regs:$c),
600 "mad.lo.s64 \t$dst, $a, $b, $c;",
601 [(set Int64Regs:$dst, (add
602 (mul Int64Regs:$a, imm:$b), Int64Regs:$c))]>;
603 def MAD64rii : NVPTXInst<(outs Int64Regs:$dst),
604 (ins Int64Regs:$a, i64imm:$b, i64imm:$c),
605 "mad.lo.s64 \t$dst, $a, $b, $c;",
606 [(set Int64Regs:$dst, (add
607 (mul Int64Regs:$a, imm:$b), imm:$c))]>;
610 def INEG16 : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
611 "neg.s16 \t$dst, $src;",
612 [(set Int16Regs:$dst, (ineg Int16Regs:$src))]>;
613 def INEG32 : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
614 "neg.s32 \t$dst, $src;",
615 [(set Int32Regs:$dst, (ineg Int32Regs:$src))]>;
616 def INEG64 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
617 "neg.s64 \t$dst, $src;",
618 [(set Int64Regs:$dst, (ineg Int64Regs:$src))]>;
620 //-----------------------------------
621 // Floating Point Arithmetic
622 //-----------------------------------
625 def FloatConst1 : PatLeaf<(fpimm), [{
626 if (&(N->getValueAPF().getSemantics()) != &llvm::APFloat::IEEEsingle)
628 float f = (float)N->getValueAPF().convertToFloat();
631 // Constand (double)1.0
632 def DoubleConst1 : PatLeaf<(fpimm), [{
633 if (&(N->getValueAPF().getSemantics()) != &llvm::APFloat::IEEEdouble)
635 double d = (double)N->getValueAPF().convertToDouble();
639 defm FADD : F3<"add", fadd>;
640 defm FSUB : F3<"sub", fsub>;
641 defm FMUL : F3<"mul", fmul>;
643 defm FADD_rn : F3_rn<"add", fadd>;
644 defm FSUB_rn : F3_rn<"sub", fsub>;
645 defm FMUL_rn : F3_rn<"mul", fmul>;
647 defm FABS : F2<"abs", fabs>;
648 defm FNEG : F2<"neg", fneg>;
649 defm FSQRT : F2<"sqrt.rn", fsqrt>;
654 def FDIV641r : NVPTXInst<(outs Float64Regs:$dst),
655 (ins f64imm:$a, Float64Regs:$b),
656 "rcp.rn.f64 \t$dst, $b;",
657 [(set Float64Regs:$dst,
658 (fdiv DoubleConst1:$a, Float64Regs:$b))]>;
659 def FDIV64rr : NVPTXInst<(outs Float64Regs:$dst),
660 (ins Float64Regs:$a, Float64Regs:$b),
661 "div.rn.f64 \t$dst, $a, $b;",
662 [(set Float64Regs:$dst,
663 (fdiv Float64Regs:$a, Float64Regs:$b))]>;
664 def FDIV64ri : NVPTXInst<(outs Float64Regs:$dst),
665 (ins Float64Regs:$a, f64imm:$b),
666 "div.rn.f64 \t$dst, $a, $b;",
667 [(set Float64Regs:$dst,
668 (fdiv Float64Regs:$a, fpimm:$b))]>;
671 // F32 Approximate reciprocal
673 def FDIV321r_ftz : NVPTXInst<(outs Float32Regs:$dst),
674 (ins f32imm:$a, Float32Regs:$b),
675 "rcp.approx.ftz.f32 \t$dst, $b;",
676 [(set Float32Regs:$dst,
677 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
678 Requires<[do_DIVF32_APPROX, doF32FTZ]>;
679 def FDIV321r : NVPTXInst<(outs Float32Regs:$dst),
680 (ins f32imm:$a, Float32Regs:$b),
681 "rcp.approx.f32 \t$dst, $b;",
682 [(set Float32Regs:$dst,
683 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
684 Requires<[do_DIVF32_APPROX]>;
686 // F32 Approximate division
688 def FDIV32approxrr_ftz : NVPTXInst<(outs Float32Regs:$dst),
689 (ins Float32Regs:$a, Float32Regs:$b),
690 "div.approx.ftz.f32 \t$dst, $a, $b;",
691 [(set Float32Regs:$dst,
692 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
693 Requires<[do_DIVF32_APPROX, doF32FTZ]>;
694 def FDIV32approxri_ftz : NVPTXInst<(outs Float32Regs:$dst),
695 (ins Float32Regs:$a, f32imm:$b),
696 "div.approx.ftz.f32 \t$dst, $a, $b;",
697 [(set Float32Regs:$dst,
698 (fdiv Float32Regs:$a, fpimm:$b))]>,
699 Requires<[do_DIVF32_APPROX, doF32FTZ]>;
700 def FDIV32approxrr : NVPTXInst<(outs Float32Regs:$dst),
701 (ins Float32Regs:$a, Float32Regs:$b),
702 "div.approx.f32 \t$dst, $a, $b;",
703 [(set Float32Regs:$dst,
704 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
705 Requires<[do_DIVF32_APPROX]>;
706 def FDIV32approxri : NVPTXInst<(outs Float32Regs:$dst),
707 (ins Float32Regs:$a, f32imm:$b),
708 "div.approx.f32 \t$dst, $a, $b;",
709 [(set Float32Regs:$dst,
710 (fdiv Float32Regs:$a, fpimm:$b))]>,
711 Requires<[do_DIVF32_APPROX]>;
713 // F32 Semi-accurate reciprocal
715 // rcp.approx gives the same result as div.full(1.0f, a) and is faster.
717 def FDIV321r_approx_ftz : NVPTXInst<(outs Float32Regs:$dst),
718 (ins f32imm:$a, Float32Regs:$b),
719 "rcp.approx.ftz.f32 \t$dst, $b;",
720 [(set Float32Regs:$dst,
721 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
722 Requires<[do_DIVF32_FULL, doF32FTZ]>;
723 def FDIV321r_approx : NVPTXInst<(outs Float32Regs:$dst),
724 (ins f32imm:$a, Float32Regs:$b),
725 "rcp.approx.f32 \t$dst, $b;",
726 [(set Float32Regs:$dst,
727 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
728 Requires<[do_DIVF32_FULL]>;
730 // F32 Semi-accurate division
732 def FDIV32rr_ftz : NVPTXInst<(outs Float32Regs:$dst),
733 (ins Float32Regs:$a, Float32Regs:$b),
734 "div.full.ftz.f32 \t$dst, $a, $b;",
735 [(set Float32Regs:$dst,
736 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
737 Requires<[do_DIVF32_FULL, doF32FTZ]>;
738 def FDIV32ri_ftz : NVPTXInst<(outs Float32Regs:$dst),
739 (ins Float32Regs:$a, f32imm:$b),
740 "div.full.ftz.f32 \t$dst, $a, $b;",
741 [(set Float32Regs:$dst,
742 (fdiv Float32Regs:$a, fpimm:$b))]>,
743 Requires<[do_DIVF32_FULL, doF32FTZ]>;
744 def FDIV32rr : NVPTXInst<(outs Float32Regs:$dst),
745 (ins Float32Regs:$a, Float32Regs:$b),
746 "div.full.f32 \t$dst, $a, $b;",
747 [(set Float32Regs:$dst,
748 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
749 Requires<[do_DIVF32_FULL]>;
750 def FDIV32ri : NVPTXInst<(outs Float32Regs:$dst),
751 (ins Float32Regs:$a, f32imm:$b),
752 "div.full.f32 \t$dst, $a, $b;",
753 [(set Float32Regs:$dst,
754 (fdiv Float32Regs:$a, fpimm:$b))]>,
755 Requires<[do_DIVF32_FULL]>;
757 // F32 Accurate reciprocal
759 def FDIV321r_prec_ftz : NVPTXInst<(outs Float32Regs:$dst),
760 (ins f32imm:$a, Float32Regs:$b),
761 "rcp.rn.ftz.f32 \t$dst, $b;",
762 [(set Float32Regs:$dst,
763 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
764 Requires<[reqPTX20, doF32FTZ]>;
765 def FDIV321r_prec : NVPTXInst<(outs Float32Regs:$dst),
766 (ins f32imm:$a, Float32Regs:$b),
767 "rcp.rn.f32 \t$dst, $b;",
768 [(set Float32Regs:$dst,
769 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
770 Requires<[reqPTX20]>;
772 // F32 Accurate division
774 def FDIV32rr_prec_ftz : NVPTXInst<(outs Float32Regs:$dst),
775 (ins Float32Regs:$a, Float32Regs:$b),
776 "div.rn.ftz.f32 \t$dst, $a, $b;",
777 [(set Float32Regs:$dst,
778 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
779 Requires<[doF32FTZ, reqPTX20]>;
780 def FDIV32ri_prec_ftz : NVPTXInst<(outs Float32Regs:$dst),
781 (ins Float32Regs:$a, f32imm:$b),
782 "div.rn.ftz.f32 \t$dst, $a, $b;",
783 [(set Float32Regs:$dst,
784 (fdiv Float32Regs:$a, fpimm:$b))]>,
785 Requires<[doF32FTZ, reqPTX20]>;
786 def FDIV32rr_prec : NVPTXInst<(outs Float32Regs:$dst),
787 (ins Float32Regs:$a, Float32Regs:$b),
788 "div.rn.f32 \t$dst, $a, $b;",
789 [(set Float32Regs:$dst,
790 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
791 Requires<[reqPTX20]>;
792 def FDIV32ri_prec : NVPTXInst<(outs Float32Regs:$dst),
793 (ins Float32Regs:$a, f32imm:$b),
794 "div.rn.f32 \t$dst, $a, $b;",
795 [(set Float32Regs:$dst,
796 (fdiv Float32Regs:$a, fpimm:$b))]>,
797 Requires<[reqPTX20]>;
803 def RSQRTF32approx1r : NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$b),
804 "rsqrt.approx.f32 \t$dst, $b;", []>;
806 def: Pat<(fdiv FloatConst1, (int_nvvm_sqrt_f Float32Regs:$b)),
807 (RSQRTF32approx1r Float32Regs:$b)>,
808 Requires<[do_DIVF32_FULL, do_SQRTF32_APPROX, doNoF32FTZ]>;
810 multiclass FPCONTRACT32<string OpcStr, Predicate Pred> {
811 def rrr : NVPTXInst<(outs Float32Regs:$dst),
812 (ins Float32Regs:$a, Float32Regs:$b, Float32Regs:$c),
813 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
814 [(set Float32Regs:$dst, (fadd
815 (fmul Float32Regs:$a, Float32Regs:$b),
816 Float32Regs:$c))]>, Requires<[Pred]>;
817 // This is to WAR a weird bug in Tablegen that does not automatically
818 // generate the following permutated rule rrr2 from the above rrr.
819 // So we explicitly add it here. This happens to FMA32 only.
820 // See the comments at FMAD32 and FMA32 for more information.
821 def rrr2 : NVPTXInst<(outs Float32Regs:$dst),
822 (ins Float32Regs:$a, Float32Regs:$b, Float32Regs:$c),
823 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
824 [(set Float32Regs:$dst, (fadd Float32Regs:$c,
825 (fmul Float32Regs:$a, Float32Regs:$b)))]>,
827 def rri : NVPTXInst<(outs Float32Regs:$dst),
828 (ins Float32Regs:$a, Float32Regs:$b, f32imm:$c),
829 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
830 [(set Float32Regs:$dst, (fadd
831 (fmul Float32Regs:$a, Float32Regs:$b), fpimm:$c))]>,
833 def rir : NVPTXInst<(outs Float32Regs:$dst),
834 (ins Float32Regs:$a, f32imm:$b, Float32Regs:$c),
835 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
836 [(set Float32Regs:$dst, (fadd
837 (fmul Float32Regs:$a, fpimm:$b), Float32Regs:$c))]>,
839 def rii : NVPTXInst<(outs Float32Regs:$dst),
840 (ins Float32Regs:$a, f32imm:$b, f32imm:$c),
841 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
842 [(set Float32Regs:$dst, (fadd
843 (fmul Float32Regs:$a, fpimm:$b), fpimm:$c))]>,
847 multiclass FPCONTRACT64<string OpcStr, Predicate Pred> {
848 def rrr : NVPTXInst<(outs Float64Regs:$dst),
849 (ins Float64Regs:$a, Float64Regs:$b, Float64Regs:$c),
850 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
851 [(set Float64Regs:$dst, (fadd
852 (fmul Float64Regs:$a, Float64Regs:$b),
853 Float64Regs:$c))]>, Requires<[Pred]>;
854 def rri : NVPTXInst<(outs Float64Regs:$dst),
855 (ins Float64Regs:$a, Float64Regs:$b, f64imm:$c),
856 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
857 [(set Float64Regs:$dst, (fadd (fmul Float64Regs:$a,
858 Float64Regs:$b), fpimm:$c))]>, Requires<[Pred]>;
859 def rir : NVPTXInst<(outs Float64Regs:$dst),
860 (ins Float64Regs:$a, f64imm:$b, Float64Regs:$c),
861 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
862 [(set Float64Regs:$dst, (fadd
863 (fmul Float64Regs:$a, fpimm:$b), Float64Regs:$c))]>,
865 def rii : NVPTXInst<(outs Float64Regs:$dst),
866 (ins Float64Regs:$a, f64imm:$b, f64imm:$c),
867 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
868 [(set Float64Regs:$dst, (fadd
869 (fmul Float64Regs:$a, fpimm:$b), fpimm:$c))]>,
873 // Due to a unknown reason (most likely a bug in tablegen), tablegen does not
874 // automatically generate the rrr2 rule from
875 // the rrr rule (see FPCONTRACT32) for FMA32, though it does for FMAD32.
876 // If we reverse the order of the following two lines, then rrr2 rule will be
877 // generated for FMA32, but not for rrr.
878 // Therefore, we manually write the rrr2 rule in FPCONTRACT32.
879 defm FMA32_ftz : FPCONTRACT32<"fma.rn.ftz.f32", doFMAF32_ftz>;
880 defm FMA32 : FPCONTRACT32<"fma.rn.f32", doFMAF32>;
881 defm FMA64 : FPCONTRACT64<"fma.rn.f64", doFMAF64>;
883 // b*c-a => fmad(b, c, -a)
884 multiclass FPCONTRACT32_SUB_PAT_MAD<NVPTXInst Inst, Predicate Pred> {
885 def : Pat<(fsub (fmul Float32Regs:$b, Float32Regs:$c), Float32Regs:$a),
886 (Inst Float32Regs:$b, Float32Regs:$c, (FNEGf32 Float32Regs:$a))>,
890 // a-b*c => fmad(-b,c, a)
891 // - legal because a-b*c <=> a+(-b*c) <=> a+(-b)*c
892 // b*c-a => fmad(b, c, -a)
893 // - legal because b*c-a <=> b*c+(-a)
894 multiclass FPCONTRACT32_SUB_PAT<NVPTXInst Inst, Predicate Pred> {
895 def : Pat<(fsub Float32Regs:$a, (fmul Float32Regs:$b, Float32Regs:$c)),
896 (Inst (FNEGf32 Float32Regs:$b), Float32Regs:$c, Float32Regs:$a)>,
898 def : Pat<(fsub (fmul Float32Regs:$b, Float32Regs:$c), Float32Regs:$a),
899 (Inst Float32Regs:$b, Float32Regs:$c, (FNEGf32 Float32Regs:$a))>,
903 // a-b*c => fmad(-b,c, a)
904 // b*c-a => fmad(b, c, -a)
905 multiclass FPCONTRACT64_SUB_PAT<NVPTXInst Inst, Predicate Pred> {
906 def : Pat<(fsub Float64Regs:$a, (fmul Float64Regs:$b, Float64Regs:$c)),
907 (Inst (FNEGf64 Float64Regs:$b), Float64Regs:$c, Float64Regs:$a)>,
910 def : Pat<(fsub (fmul Float64Regs:$b, Float64Regs:$c), Float64Regs:$a),
911 (Inst Float64Regs:$b, Float64Regs:$c, (FNEGf64 Float64Regs:$a))>,
915 defm FMAF32ext_ftz : FPCONTRACT32_SUB_PAT<FMA32_ftzrrr, doFMAF32AGG_ftz>;
916 defm FMAF32ext : FPCONTRACT32_SUB_PAT<FMA32rrr, doFMAF32AGG>;
917 defm FMAF64ext : FPCONTRACT64_SUB_PAT<FMA64rrr, doFMAF64AGG>;
919 def SINF: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$src),
920 "sin.approx.f32 \t$dst, $src;",
921 [(set Float32Regs:$dst, (fsin Float32Regs:$src))]>;
922 def COSF: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$src),
923 "cos.approx.f32 \t$dst, $src;",
924 [(set Float32Regs:$dst, (fcos Float32Regs:$src))]>;
926 // Lower (frem x, y) into (sub x, (mul (floor (div x, y)) y))
927 // e.g. "poor man's fmod()"
930 def : Pat<(frem Float32Regs:$x, Float32Regs:$y),
931 (FSUBf32rr_ftz Float32Regs:$x, (FMULf32rr_ftz (CVT_f32_f32
932 (FDIV32rr_prec_ftz Float32Regs:$x, Float32Regs:$y), CvtRMI_FTZ),
934 Requires<[doF32FTZ]>;
935 def : Pat<(frem Float32Regs:$x, fpimm:$y),
936 (FSUBf32rr_ftz Float32Regs:$x, (FMULf32ri_ftz (CVT_f32_f32
937 (FDIV32ri_prec_ftz Float32Regs:$x, fpimm:$y), CvtRMI_FTZ),
939 Requires<[doF32FTZ]>;
942 def : Pat<(frem Float32Regs:$x, Float32Regs:$y),
943 (FSUBf32rr Float32Regs:$x, (FMULf32rr (CVT_f32_f32
944 (FDIV32rr_prec Float32Regs:$x, Float32Regs:$y), CvtRMI),
946 def : Pat<(frem Float32Regs:$x, fpimm:$y),
947 (FSUBf32rr Float32Regs:$x, (FMULf32ri (CVT_f32_f32
948 (FDIV32ri_prec Float32Regs:$x, fpimm:$y), CvtRMI),
952 def : Pat<(frem Float64Regs:$x, Float64Regs:$y),
953 (FSUBf64rr Float64Regs:$x, (FMULf64rr (CVT_f64_f64
954 (FDIV64rr Float64Regs:$x, Float64Regs:$y), CvtRMI),
956 def : Pat<(frem Float64Regs:$x, fpimm:$y),
957 (FSUBf64rr Float64Regs:$x, (FMULf64ri (CVT_f64_f64
958 (FDIV64ri Float64Regs:$x, fpimm:$y), CvtRMI),
961 //-----------------------------------
962 // Logical Arithmetic
963 //-----------------------------------
965 multiclass LOG_FORMAT<string OpcStr, SDNode OpNode> {
966 def b1rr: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$a, Int1Regs:$b),
967 !strconcat(OpcStr, ".pred \t$dst, $a, $b;"),
968 [(set Int1Regs:$dst, (OpNode Int1Regs:$a, Int1Regs:$b))]>;
969 def b1ri: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$a, i1imm:$b),
970 !strconcat(OpcStr, ".pred \t$dst, $a, $b;"),
971 [(set Int1Regs:$dst, (OpNode Int1Regs:$a, imm:$b))]>;
972 def b16rr: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
973 !strconcat(OpcStr, ".b16 \t$dst, $a, $b;"),
974 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
976 def b16ri: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
977 !strconcat(OpcStr, ".b16 \t$dst, $a, $b;"),
978 [(set Int16Regs:$dst, (OpNode Int16Regs:$a, imm:$b))]>;
979 def b32rr: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, Int32Regs:$b),
980 !strconcat(OpcStr, ".b32 \t$dst, $a, $b;"),
981 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
983 def b32ri: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
984 !strconcat(OpcStr, ".b32 \t$dst, $a, $b;"),
985 [(set Int32Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
986 def b64rr: NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, Int64Regs:$b),
987 !strconcat(OpcStr, ".b64 \t$dst, $a, $b;"),
988 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
990 def b64ri: NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i64imm:$b),
991 !strconcat(OpcStr, ".b64 \t$dst, $a, $b;"),
992 [(set Int64Regs:$dst, (OpNode Int64Regs:$a, imm:$b))]>;
995 defm OR : LOG_FORMAT<"or", or>;
996 defm AND : LOG_FORMAT<"and", and>;
997 defm XOR : LOG_FORMAT<"xor", xor>;
999 def NOT1: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$src),
1000 "not.pred \t$dst, $src;",
1001 [(set Int1Regs:$dst, (not Int1Regs:$src))]>;
1002 def NOT16: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
1003 "not.b16 \t$dst, $src;",
1004 [(set Int16Regs:$dst, (not Int16Regs:$src))]>;
1005 def NOT32: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
1006 "not.b32 \t$dst, $src;",
1007 [(set Int32Regs:$dst, (not Int32Regs:$src))]>;
1008 def NOT64: NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
1009 "not.b64 \t$dst, $src;",
1010 [(set Int64Regs:$dst, (not Int64Regs:$src))]>;
1012 // For shifts, the second src operand must be 32-bit value
1013 multiclass LSHIFT_FORMAT<string OpcStr, SDNode OpNode> {
1014 def i64rr : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a,
1016 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
1017 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
1019 def i64ri : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i32imm:$b),
1020 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
1021 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
1023 def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a,
1025 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1026 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
1028 def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
1029 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1030 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
1032 def i32ii : NVPTXInst<(outs Int32Regs:$dst), (ins i32imm:$a, i32imm:$b),
1033 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1034 [(set Int32Regs:$dst, (OpNode (i32 imm:$a),
1036 def i16rr : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a,
1038 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
1039 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
1041 def i16ri : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i32imm:$b),
1042 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
1043 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
1047 defm SHL : LSHIFT_FORMAT<"shl.b", shl>;
1049 // For shifts, the second src operand must be 32-bit value
1050 // Need to add cvt for the 8-bits.
1051 multiclass RSHIFT_FORMAT<string OpcStr, SDNode OpNode> {
1052 def i64rr : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a,
1054 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
1055 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
1057 def i64ri : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i32imm:$b),
1058 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
1059 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
1061 def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a,
1063 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1064 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
1066 def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
1067 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1068 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
1070 def i32ii : NVPTXInst<(outs Int32Regs:$dst), (ins i32imm:$a, i32imm:$b),
1071 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1072 [(set Int32Regs:$dst, (OpNode (i32 imm:$a),
1074 def i16rr : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a,
1076 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
1077 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
1079 def i16ri : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i32imm:$b),
1080 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
1081 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
1085 defm SRA : RSHIFT_FORMAT<"shr.s", sra>;
1086 defm SRL : RSHIFT_FORMAT<"shr.u", srl>;
1089 def ROT32imm_sw : NVPTXInst<(outs Int32Regs:$dst),
1090 (ins Int32Regs:$src, i32imm:$amt1, i32imm:$amt2),
1091 !strconcat("{{\n\t",
1092 !strconcat(".reg .b32 %lhs;\n\t",
1093 !strconcat(".reg .b32 %rhs;\n\t",
1094 !strconcat("shl.b32 \t%lhs, $src, $amt1;\n\t",
1095 !strconcat("shr.b32 \t%rhs, $src, $amt2;\n\t",
1096 !strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t",
1097 !strconcat("}}", ""))))))),
1100 def SUB_FRM_32 : SDNodeXForm<imm, [{
1101 return CurDAG->getTargetConstant(32-N->getZExtValue(), MVT::i32);
1104 def : Pat<(rotl Int32Regs:$src, (i32 imm:$amt)),
1105 (ROT32imm_sw Int32Regs:$src, imm:$amt, (SUB_FRM_32 node:$amt))>;
1106 def : Pat<(rotr Int32Regs:$src, (i32 imm:$amt)),
1107 (ROT32imm_sw Int32Regs:$src, (SUB_FRM_32 node:$amt), imm:$amt)>;
1109 def ROTL32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src,
1111 !strconcat("{{\n\t",
1112 !strconcat(".reg .b32 %lhs;\n\t",
1113 !strconcat(".reg .b32 %rhs;\n\t",
1114 !strconcat(".reg .b32 %amt2;\n\t",
1115 !strconcat("shl.b32 \t%lhs, $src, $amt;\n\t",
1116 !strconcat("sub.s32 \t%amt2, 32, $amt;\n\t",
1117 !strconcat("shr.b32 \t%rhs, $src, %amt2;\n\t",
1118 !strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t",
1119 !strconcat("}}", ""))))))))),
1120 [(set Int32Regs:$dst, (rotl Int32Regs:$src, Int32Regs:$amt))]>;
1122 def ROTR32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src,
1124 !strconcat("{{\n\t",
1125 !strconcat(".reg .b32 %lhs;\n\t",
1126 !strconcat(".reg .b32 %rhs;\n\t",
1127 !strconcat(".reg .b32 %amt2;\n\t",
1128 !strconcat("shr.b32 \t%lhs, $src, $amt;\n\t",
1129 !strconcat("sub.s32 \t%amt2, 32, $amt;\n\t",
1130 !strconcat("shl.b32 \t%rhs, $src, %amt2;\n\t",
1131 !strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t",
1132 !strconcat("}}", ""))))))))),
1133 [(set Int32Regs:$dst, (rotr Int32Regs:$src, Int32Regs:$amt))]>;
1136 def ROT64imm_sw : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src,
1137 i32imm:$amt1, i32imm:$amt2),
1138 !strconcat("{{\n\t",
1139 !strconcat(".reg .b64 %lhs;\n\t",
1140 !strconcat(".reg .b64 %rhs;\n\t",
1141 !strconcat("shl.b64 \t%lhs, $src, $amt1;\n\t",
1142 !strconcat("shr.b64 \t%rhs, $src, $amt2;\n\t",
1143 !strconcat("add.u64 \t$dst, %lhs, %rhs;\n\t",
1144 !strconcat("}}", ""))))))),
1147 def SUB_FRM_64 : SDNodeXForm<imm, [{
1148 return CurDAG->getTargetConstant(64-N->getZExtValue(), MVT::i32);
1151 def : Pat<(rotl Int64Regs:$src, (i32 imm:$amt)),
1152 (ROT64imm_sw Int64Regs:$src, imm:$amt, (SUB_FRM_64 node:$amt))>;
1153 def : Pat<(rotr Int64Regs:$src, (i32 imm:$amt)),
1154 (ROT64imm_sw Int64Regs:$src, (SUB_FRM_64 node:$amt), imm:$amt)>;
1156 def ROTL64reg_sw : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src,
1158 !strconcat("{{\n\t",
1159 !strconcat(".reg .b64 %lhs;\n\t",
1160 !strconcat(".reg .b64 %rhs;\n\t",
1161 !strconcat(".reg .u32 %amt2;\n\t",
1162 !strconcat("shl.b64 \t%lhs, $src, $amt;\n\t",
1163 !strconcat("sub.u32 \t%amt2, 64, $amt;\n\t",
1164 !strconcat("shr.b64 \t%rhs, $src, %amt2;\n\t",
1165 !strconcat("add.u64 \t$dst, %lhs, %rhs;\n\t",
1166 !strconcat("}}", ""))))))))),
1167 [(set Int64Regs:$dst, (rotl Int64Regs:$src, Int32Regs:$amt))]>;
1169 def ROTR64reg_sw : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src,
1171 !strconcat("{{\n\t",
1172 !strconcat(".reg .b64 %lhs;\n\t",
1173 !strconcat(".reg .b64 %rhs;\n\t",
1174 !strconcat(".reg .u32 %amt2;\n\t",
1175 !strconcat("shr.b64 \t%lhs, $src, $amt;\n\t",
1176 !strconcat("sub.u32 \t%amt2, 64, $amt;\n\t",
1177 !strconcat("shl.b64 \t%rhs, $src, %amt2;\n\t",
1178 !strconcat("add.u64 \t$dst, %lhs, %rhs;\n\t",
1179 !strconcat("}}", ""))))))))),
1180 [(set Int64Regs:$dst, (rotr Int64Regs:$src, Int32Regs:$amt))]>;
1182 // BFE - bit-field extract
1184 multiclass BFE<string TyStr, RegisterClass RC> {
1185 // BFE supports both 32-bit and 64-bit values, but the start and length
1186 // operands are always 32-bit
1188 : NVPTXInst<(outs RC:$d),
1189 (ins RC:$a, Int32Regs:$b, Int32Regs:$c),
1190 !strconcat("bfe.", TyStr, " \t$d, $a, $b, $c;"), []>;
1192 : NVPTXInst<(outs RC:$d),
1193 (ins RC:$a, Int32Regs:$b, i32imm:$c),
1194 !strconcat("bfe.", TyStr, " \t$d, $a, $b, $c;"), []>;
1196 : NVPTXInst<(outs RC:$d),
1197 (ins RC:$a, i32imm:$b, i32imm:$c),
1198 !strconcat("bfe.", TyStr, " \t$d, $a, $b, $c;"), []>;
1201 defm BFE_S32 : BFE<"s32", Int32Regs>;
1202 defm BFE_U32 : BFE<"u32", Int32Regs>;
1203 defm BFE_S64 : BFE<"s64", Int64Regs>;
1204 defm BFE_U64 : BFE<"u64", Int64Regs>;
1206 //-----------------------------------
1207 // General Comparison
1208 //-----------------------------------
1210 // General setp instructions
1211 multiclass SETP<string TypeStr, RegisterClass RC, Operand ImmCls> {
1212 def rr : NVPTXInst<(outs Int1Regs:$dst),
1213 (ins RC:$a, RC:$b, CmpMode:$cmp),
1214 !strconcat("setp${cmp:base}${cmp:ftz}.", TypeStr, "\t$dst, $a, $b;"),
1216 def ri : NVPTXInst<(outs Int1Regs:$dst),
1217 (ins RC:$a, ImmCls:$b, CmpMode:$cmp),
1218 !strconcat("setp${cmp:base}${cmp:ftz}.", TypeStr, "\t$dst, $a, $b;"),
1220 def ir : NVPTXInst<(outs Int1Regs:$dst),
1221 (ins ImmCls:$a, RC:$b, CmpMode:$cmp),
1222 !strconcat("setp${cmp:base}${cmp:ftz}.", TypeStr, "\t$dst, $a, $b;"),
1226 defm SETP_b16 : SETP<"b16", Int16Regs, i16imm>;
1227 defm SETP_s16 : SETP<"s16", Int16Regs, i16imm>;
1228 defm SETP_u16 : SETP<"u16", Int16Regs, i16imm>;
1229 defm SETP_b32 : SETP<"b32", Int32Regs, i32imm>;
1230 defm SETP_s32 : SETP<"s32", Int32Regs, i32imm>;
1231 defm SETP_u32 : SETP<"u32", Int32Regs, i32imm>;
1232 defm SETP_b64 : SETP<"b64", Int64Regs, i64imm>;
1233 defm SETP_s64 : SETP<"s64", Int64Regs, i64imm>;
1234 defm SETP_u64 : SETP<"u64", Int64Regs, i64imm>;
1235 defm SETP_f32 : SETP<"f32", Float32Regs, f32imm>;
1236 defm SETP_f64 : SETP<"f64", Float64Regs, f64imm>;
1238 // General set instructions
1239 multiclass SET<string TypeStr, RegisterClass RC, Operand ImmCls> {
1240 def rr : NVPTXInst<(outs Int32Regs:$dst),
1241 (ins RC:$a, RC:$b, CmpMode:$cmp),
1242 !strconcat("set$cmp.", TypeStr, "\t$dst, $a, $b;"), []>;
1243 def ri : NVPTXInst<(outs Int32Regs:$dst),
1244 (ins RC:$a, ImmCls:$b, CmpMode:$cmp),
1245 !strconcat("set$cmp.", TypeStr, "\t$dst, $a, $b;"), []>;
1246 def ir : NVPTXInst<(outs Int32Regs:$dst),
1247 (ins ImmCls:$a, RC:$b, CmpMode:$cmp),
1248 !strconcat("set$cmp.", TypeStr, "\t$dst, $a, $b;"), []>;
1251 defm SET_b16 : SET<"b16", Int16Regs, i16imm>;
1252 defm SET_s16 : SET<"s16", Int16Regs, i16imm>;
1253 defm SET_u16 : SET<"u16", Int16Regs, i16imm>;
1254 defm SET_b32 : SET<"b32", Int32Regs, i32imm>;
1255 defm SET_s32 : SET<"s32", Int32Regs, i32imm>;
1256 defm SET_u32 : SET<"u32", Int32Regs, i32imm>;
1257 defm SET_b64 : SET<"b64", Int64Regs, i64imm>;
1258 defm SET_s64 : SET<"s64", Int64Regs, i64imm>;
1259 defm SET_u64 : SET<"u64", Int64Regs, i64imm>;
1260 defm SET_f32 : SET<"f32", Float32Regs, f32imm>;
1261 defm SET_f64 : SET<"f64", Float64Regs, f64imm>;
1263 //-----------------------------------
1264 // General Selection
1265 //-----------------------------------
1267 // General selp instructions
1268 multiclass SELP<string TypeStr, RegisterClass RC, Operand ImmCls> {
1269 def rr : NVPTXInst<(outs RC:$dst),
1270 (ins RC:$a, RC:$b, Int1Regs:$p),
1271 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
1272 def ri : NVPTXInst<(outs RC:$dst),
1273 (ins RC:$a, ImmCls:$b, Int1Regs:$p),
1274 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
1275 def ir : NVPTXInst<(outs RC:$dst),
1276 (ins ImmCls:$a, RC:$b, Int1Regs:$p),
1277 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
1278 def ii : NVPTXInst<(outs RC:$dst),
1279 (ins ImmCls:$a, ImmCls:$b, Int1Regs:$p),
1280 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
1283 multiclass SELP_PATTERN<string TypeStr, RegisterClass RC, Operand ImmCls,
1285 def rr : NVPTXInst<(outs RC:$dst),
1286 (ins RC:$a, RC:$b, Int1Regs:$p),
1287 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
1288 [(set RC:$dst, (select Int1Regs:$p, RC:$a, RC:$b))]>;
1289 def ri : NVPTXInst<(outs RC:$dst),
1290 (ins RC:$a, ImmCls:$b, Int1Regs:$p),
1291 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
1292 [(set RC:$dst, (select Int1Regs:$p, RC:$a, ImmNode:$b))]>;
1293 def ir : NVPTXInst<(outs RC:$dst),
1294 (ins ImmCls:$a, RC:$b, Int1Regs:$p),
1295 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
1296 [(set RC:$dst, (select Int1Regs:$p, ImmNode:$a, RC:$b))]>;
1297 def ii : NVPTXInst<(outs RC:$dst),
1298 (ins ImmCls:$a, ImmCls:$b, Int1Regs:$p),
1299 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
1300 [(set RC:$dst, (select Int1Regs:$p, ImmNode:$a, ImmNode:$b))]>;
1303 defm SELP_b16 : SELP_PATTERN<"b16", Int16Regs, i16imm, imm>;
1304 defm SELP_s16 : SELP<"s16", Int16Regs, i16imm>;
1305 defm SELP_u16 : SELP<"u16", Int16Regs, i16imm>;
1306 defm SELP_b32 : SELP_PATTERN<"b32", Int32Regs, i32imm, imm>;
1307 defm SELP_s32 : SELP<"s32", Int32Regs, i32imm>;
1308 defm SELP_u32 : SELP<"u32", Int32Regs, i32imm>;
1309 defm SELP_b64 : SELP_PATTERN<"b64", Int64Regs, i64imm, imm>;
1310 defm SELP_s64 : SELP<"s64", Int64Regs, i64imm>;
1311 defm SELP_u64 : SELP<"u64", Int64Regs, i64imm>;
1312 defm SELP_f32 : SELP_PATTERN<"f32", Float32Regs, f32imm, fpimm>;
1313 defm SELP_f64 : SELP_PATTERN<"f64", Float64Regs, f64imm, fpimm>;
1315 // Special select for predicate operands
1316 def : Pat<(i1 (select Int1Regs:$p, Int1Regs:$a, Int1Regs:$b)),
1317 (ORb1rr (ANDb1rr Int1Regs:$p, Int1Regs:$a),
1318 (ANDb1rr (NOT1 Int1Regs:$p), Int1Regs:$b))>;
1320 //-----------------------------------
1321 // Data Movement (Load / Store, Move)
1322 //-----------------------------------
1324 def ADDRri : ComplexPattern<i32, 2, "SelectADDRri", [frameindex],
1326 def ADDRri64 : ComplexPattern<i64, 2, "SelectADDRri64", [frameindex],
1329 def MEMri : Operand<i32> {
1330 let PrintMethod = "printMemOperand";
1331 let MIOperandInfo = (ops Int32Regs, i32imm);
1333 def MEMri64 : Operand<i64> {
1334 let PrintMethod = "printMemOperand";
1335 let MIOperandInfo = (ops Int64Regs, i64imm);
1338 def imem : Operand<iPTR> {
1339 let PrintMethod = "printOperand";
1342 def imemAny : Operand<iPTRAny> {
1343 let PrintMethod = "printOperand";
1346 def LdStCode : Operand<i32> {
1347 let PrintMethod = "printLdStCode";
1350 def SDTWrapper : SDTypeProfile<1, 1, [SDTCisSameAs<0, 1>, SDTCisPtrTy<0>]>;
1351 def Wrapper : SDNode<"NVPTXISD::Wrapper", SDTWrapper>;
1353 def MOV_ADDR : NVPTXInst<(outs Int32Regs:$dst), (ins imem:$a),
1354 "mov.u32 \t$dst, $a;",
1355 [(set Int32Regs:$dst, (Wrapper tglobaladdr:$a))]>;
1357 def MOV_ADDR64 : NVPTXInst<(outs Int64Regs:$dst), (ins imem:$a),
1358 "mov.u64 \t$dst, $a;",
1359 [(set Int64Regs:$dst, (Wrapper tglobaladdr:$a))]>;
1361 // Get pointer to local stack
1363 : NVPTXInst<(outs Int32Regs:$d), (ins i32imm:$num),
1364 "mov.u32 \t$d, __local_depot$num;", []>;
1365 def MOV_DEPOT_ADDR_64
1366 : NVPTXInst<(outs Int64Regs:$d), (ins i32imm:$num),
1367 "mov.u64 \t$d, __local_depot$num;", []>;
1370 // copyPhysreg is hard-coded in NVPTXInstrInfo.cpp
1371 let IsSimpleMove=1 in {
1372 def IMOV1rr: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$sss),
1373 "mov.pred \t$dst, $sss;", []>;
1374 def IMOV16rr: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$sss),
1375 "mov.u16 \t$dst, $sss;", []>;
1376 def IMOV32rr: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$sss),
1377 "mov.u32 \t$dst, $sss;", []>;
1378 def IMOV64rr: NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$sss),
1379 "mov.u64 \t$dst, $sss;", []>;
1381 def FMOV32rr: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$src),
1382 "mov.f32 \t$dst, $src;", []>;
1383 def FMOV64rr: NVPTXInst<(outs Float64Regs:$dst), (ins Float64Regs:$src),
1384 "mov.f64 \t$dst, $src;", []>;
1386 def IMOV1ri: NVPTXInst<(outs Int1Regs:$dst), (ins i1imm:$src),
1387 "mov.pred \t$dst, $src;",
1388 [(set Int1Regs:$dst, imm:$src)]>;
1389 def IMOV16ri: NVPTXInst<(outs Int16Regs:$dst), (ins i16imm:$src),
1390 "mov.u16 \t$dst, $src;",
1391 [(set Int16Regs:$dst, imm:$src)]>;
1392 def IMOV32ri: NVPTXInst<(outs Int32Regs:$dst), (ins i32imm:$src),
1393 "mov.u32 \t$dst, $src;",
1394 [(set Int32Regs:$dst, imm:$src)]>;
1395 def IMOV64i: NVPTXInst<(outs Int64Regs:$dst), (ins i64imm:$src),
1396 "mov.u64 \t$dst, $src;",
1397 [(set Int64Regs:$dst, imm:$src)]>;
1399 def FMOV32ri: NVPTXInst<(outs Float32Regs:$dst), (ins f32imm:$src),
1400 "mov.f32 \t$dst, $src;",
1401 [(set Float32Regs:$dst, fpimm:$src)]>;
1402 def FMOV64ri: NVPTXInst<(outs Float64Regs:$dst), (ins f64imm:$src),
1403 "mov.f64 \t$dst, $src;",
1404 [(set Float64Regs:$dst, fpimm:$src)]>;
1406 def : Pat<(i32 (Wrapper texternalsym:$dst)), (IMOV32ri texternalsym:$dst)>;
1408 //---- Copy Frame Index ----
1409 def LEA_ADDRi : NVPTXInst<(outs Int32Regs:$dst), (ins MEMri:$addr),
1410 "add.u32 \t$dst, ${addr:add};",
1411 [(set Int32Regs:$dst, ADDRri:$addr)]>;
1412 def LEA_ADDRi64 : NVPTXInst<(outs Int64Regs:$dst), (ins MEMri64:$addr),
1413 "add.u64 \t$dst, ${addr:add};",
1414 [(set Int64Regs:$dst, ADDRri64:$addr)]>;
1416 //-----------------------------------
1417 // Comparison and Selection
1418 //-----------------------------------
1420 multiclass ISET_FORMAT<PatFrag OpNode, PatLeaf Mode,
1421 Instruction setp_16rr,
1422 Instruction setp_16ri,
1423 Instruction setp_16ir,
1424 Instruction setp_32rr,
1425 Instruction setp_32ri,
1426 Instruction setp_32ir,
1427 Instruction setp_64rr,
1428 Instruction setp_64ri,
1429 Instruction setp_64ir,
1430 Instruction set_16rr,
1431 Instruction set_16ri,
1432 Instruction set_16ir,
1433 Instruction set_32rr,
1434 Instruction set_32ri,
1435 Instruction set_32ir,
1436 Instruction set_64rr,
1437 Instruction set_64ri,
1438 Instruction set_64ir> {
1440 def : Pat<(i1 (OpNode Int16Regs:$a, Int16Regs:$b)),
1441 (setp_16rr Int16Regs:$a, Int16Regs:$b, Mode)>;
1442 def : Pat<(i1 (OpNode Int16Regs:$a, imm:$b)),
1443 (setp_16ri Int16Regs:$a, imm:$b, Mode)>;
1444 def : Pat<(i1 (OpNode imm:$a, Int16Regs:$b)),
1445 (setp_16ir imm:$a, Int16Regs:$b, Mode)>;
1447 def : Pat<(i1 (OpNode Int32Regs:$a, Int32Regs:$b)),
1448 (setp_32rr Int32Regs:$a, Int32Regs:$b, Mode)>;
1449 def : Pat<(i1 (OpNode Int32Regs:$a, imm:$b)),
1450 (setp_32ri Int32Regs:$a, imm:$b, Mode)>;
1451 def : Pat<(i1 (OpNode imm:$a, Int32Regs:$b)),
1452 (setp_32ir imm:$a, Int32Regs:$b, Mode)>;
1454 def : Pat<(i1 (OpNode Int64Regs:$a, Int64Regs:$b)),
1455 (setp_64rr Int64Regs:$a, Int64Regs:$b, Mode)>;
1456 def : Pat<(i1 (OpNode Int64Regs:$a, imm:$b)),
1457 (setp_64ri Int64Regs:$a, imm:$b, Mode)>;
1458 def : Pat<(i1 (OpNode imm:$a, Int64Regs:$b)),
1459 (setp_64ir imm:$a, Int64Regs:$b, Mode)>;
1462 def : Pat<(i32 (OpNode Int16Regs:$a, Int16Regs:$b)),
1463 (set_16rr Int16Regs:$a, Int16Regs:$b, Mode)>;
1464 def : Pat<(i32 (OpNode Int16Regs:$a, imm:$b)),
1465 (set_16ri Int16Regs:$a, imm:$b, Mode)>;
1466 def : Pat<(i32 (OpNode imm:$a, Int16Regs:$b)),
1467 (set_16ir imm:$a, Int16Regs:$b, Mode)>;
1469 def : Pat<(i32 (OpNode Int32Regs:$a, Int32Regs:$b)),
1470 (set_32rr Int32Regs:$a, Int32Regs:$b, Mode)>;
1471 def : Pat<(i32 (OpNode Int32Regs:$a, imm:$b)),
1472 (set_32ri Int32Regs:$a, imm:$b, Mode)>;
1473 def : Pat<(i32 (OpNode imm:$a, Int32Regs:$b)),
1474 (set_32ir imm:$a, Int32Regs:$b, Mode)>;
1476 def : Pat<(i32 (OpNode Int64Regs:$a, Int64Regs:$b)),
1477 (set_64rr Int64Regs:$a, Int64Regs:$b, Mode)>;
1478 def : Pat<(i32 (OpNode Int64Regs:$a, imm:$b)),
1479 (set_64ri Int64Regs:$a, imm:$b, Mode)>;
1480 def : Pat<(i32 (OpNode imm:$a, Int64Regs:$b)),
1481 (set_64ir imm:$a, Int64Regs:$b, Mode)>;
1484 multiclass ISET_FORMAT_SIGNED<PatFrag OpNode, PatLeaf Mode>
1485 : ISET_FORMAT<OpNode, Mode,
1486 SETP_s16rr, SETP_s16ri, SETP_s16ir,
1487 SETP_s32rr, SETP_s32ri, SETP_s32ir,
1488 SETP_s64rr, SETP_s64ri, SETP_s64ir,
1489 SET_s16rr, SET_s16ri, SET_s16ir,
1490 SET_s32rr, SET_s32ri, SET_s32ir,
1491 SET_s64rr, SET_s64ri, SET_s64ir> {
1492 // TableGen doesn't like empty multiclasses
1493 def : PatLeaf<(i32 0)>;
1496 multiclass ISET_FORMAT_UNSIGNED<PatFrag OpNode, PatLeaf Mode>
1497 : ISET_FORMAT<OpNode, Mode,
1498 SETP_u16rr, SETP_u16ri, SETP_u16ir,
1499 SETP_u32rr, SETP_u32ri, SETP_u32ir,
1500 SETP_u64rr, SETP_u64ri, SETP_u64ir,
1501 SET_u16rr, SET_u16ri, SET_u16ir,
1502 SET_u32rr, SET_u32ri, SET_u32ir,
1503 SET_u64rr, SET_u64ri, SET_u64ir> {
1504 // TableGen doesn't like empty multiclasses
1505 def : PatLeaf<(i32 0)>;
1508 defm : ISET_FORMAT_SIGNED<setgt, CmpGT>;
1509 defm : ISET_FORMAT_UNSIGNED<setugt, CmpGT>;
1510 defm : ISET_FORMAT_SIGNED<setlt, CmpLT>;
1511 defm : ISET_FORMAT_UNSIGNED<setult, CmpLT>;
1512 defm : ISET_FORMAT_SIGNED<setge, CmpGE>;
1513 defm : ISET_FORMAT_UNSIGNED<setuge, CmpGE>;
1514 defm : ISET_FORMAT_SIGNED<setle, CmpLE>;
1515 defm : ISET_FORMAT_UNSIGNED<setule, CmpLE>;
1516 defm : ISET_FORMAT_SIGNED<seteq, CmpEQ>;
1517 defm : ISET_FORMAT_UNSIGNED<setueq, CmpEQ>;
1518 defm : ISET_FORMAT_SIGNED<setne, CmpNE>;
1519 defm : ISET_FORMAT_UNSIGNED<setune, CmpNE>;
1522 def : Pat<(setne Int1Regs:$a, Int1Regs:$b),
1523 (XORb1rr Int1Regs:$a, Int1Regs:$b)>;
1524 def : Pat<(setune Int1Regs:$a, Int1Regs:$b),
1525 (XORb1rr Int1Regs:$a, Int1Regs:$b)>;
1527 def : Pat<(seteq Int1Regs:$a, Int1Regs:$b),
1528 (NOT1 (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
1529 def : Pat<(setueq Int1Regs:$a, Int1Regs:$b),
1530 (NOT1 (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
1532 // i1 compare -> i32
1533 def : Pat<(i32 (setne Int1Regs:$a, Int1Regs:$b)),
1534 (SELP_u32ii -1, 0, (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
1535 def : Pat<(i32 (setne Int1Regs:$a, Int1Regs:$b)),
1536 (SELP_u32ii 0, -1, (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
1540 multiclass FSET_FORMAT<PatFrag OpNode, PatLeaf Mode, PatLeaf ModeFTZ> {
1542 def : Pat<(i1 (OpNode Float32Regs:$a, Float32Regs:$b)),
1543 (SETP_f32rr Float32Regs:$a, Float32Regs:$b, ModeFTZ)>,
1544 Requires<[doF32FTZ]>;
1545 def : Pat<(i1 (OpNode Float32Regs:$a, Float32Regs:$b)),
1546 (SETP_f32rr Float32Regs:$a, Float32Regs:$b, Mode)>;
1547 def : Pat<(i1 (OpNode Float32Regs:$a, fpimm:$b)),
1548 (SETP_f32ri Float32Regs:$a, fpimm:$b, ModeFTZ)>,
1549 Requires<[doF32FTZ]>;
1550 def : Pat<(i1 (OpNode Float32Regs:$a, fpimm:$b)),
1551 (SETP_f32ri Float32Regs:$a, fpimm:$b, Mode)>;
1552 def : Pat<(i1 (OpNode fpimm:$a, Float32Regs:$b)),
1553 (SETP_f32ir fpimm:$a, Float32Regs:$b, ModeFTZ)>,
1554 Requires<[doF32FTZ]>;
1555 def : Pat<(i1 (OpNode fpimm:$a, Float32Regs:$b)),
1556 (SETP_f32ir fpimm:$a, Float32Regs:$b, Mode)>;
1559 def : Pat<(i1 (OpNode Float64Regs:$a, Float64Regs:$b)),
1560 (SETP_f64rr Float64Regs:$a, Float64Regs:$b, Mode)>;
1561 def : Pat<(i1 (OpNode Float64Regs:$a, fpimm:$b)),
1562 (SETP_f64ri Float64Regs:$a, fpimm:$b, Mode)>;
1563 def : Pat<(i1 (OpNode fpimm:$a, Float64Regs:$b)),
1564 (SETP_f64ir fpimm:$a, Float64Regs:$b, Mode)>;
1567 def : Pat<(i32 (OpNode Float32Regs:$a, Float32Regs:$b)),
1568 (SET_f32rr Float32Regs:$a, Float32Regs:$b, ModeFTZ)>,
1569 Requires<[doF32FTZ]>;
1570 def : Pat<(i32 (OpNode Float32Regs:$a, Float32Regs:$b)),
1571 (SET_f32rr Float32Regs:$a, Float32Regs:$b, Mode)>;
1572 def : Pat<(i32 (OpNode Float32Regs:$a, fpimm:$b)),
1573 (SET_f32ri Float32Regs:$a, fpimm:$b, ModeFTZ)>,
1574 Requires<[doF32FTZ]>;
1575 def : Pat<(i32 (OpNode Float32Regs:$a, fpimm:$b)),
1576 (SET_f32ri Float32Regs:$a, fpimm:$b, Mode)>;
1577 def : Pat<(i32 (OpNode fpimm:$a, Float32Regs:$b)),
1578 (SET_f32ir fpimm:$a, Float32Regs:$b, ModeFTZ)>,
1579 Requires<[doF32FTZ]>;
1580 def : Pat<(i32 (OpNode fpimm:$a, Float32Regs:$b)),
1581 (SET_f32ir fpimm:$a, Float32Regs:$b, Mode)>;
1584 def : Pat<(i32 (OpNode Float64Regs:$a, Float64Regs:$b)),
1585 (SET_f64rr Float64Regs:$a, Float64Regs:$b, Mode)>;
1586 def : Pat<(i32 (OpNode Float64Regs:$a, fpimm:$b)),
1587 (SET_f64ri Float64Regs:$a, fpimm:$b, Mode)>;
1588 def : Pat<(i32 (OpNode fpimm:$a, Float64Regs:$b)),
1589 (SET_f64ir fpimm:$a, Float64Regs:$b, Mode)>;
1592 defm FSetGT : FSET_FORMAT<setogt, CmpGT, CmpGT_FTZ>;
1593 defm FSetLT : FSET_FORMAT<setolt, CmpLT, CmpLT_FTZ>;
1594 defm FSetGE : FSET_FORMAT<setoge, CmpGE, CmpGE_FTZ>;
1595 defm FSetLE : FSET_FORMAT<setole, CmpLE, CmpLE_FTZ>;
1596 defm FSetEQ : FSET_FORMAT<setoeq, CmpEQ, CmpEQ_FTZ>;
1597 defm FSetNE : FSET_FORMAT<setone, CmpNE, CmpNE_FTZ>;
1599 defm FSetUGT : FSET_FORMAT<setugt, CmpGTU, CmpGTU_FTZ>;
1600 defm FSetULT : FSET_FORMAT<setult, CmpLTU, CmpLTU_FTZ>;
1601 defm FSetUGE : FSET_FORMAT<setuge, CmpGEU, CmpGEU_FTZ>;
1602 defm FSetULE : FSET_FORMAT<setule, CmpLEU, CmpLEU_FTZ>;
1603 defm FSetUEQ : FSET_FORMAT<setueq, CmpEQU, CmpEQU_FTZ>;
1604 defm FSetUNE : FSET_FORMAT<setune, CmpNEU, CmpNEU_FTZ>;
1606 defm FSetNUM : FSET_FORMAT<seto, CmpNUM, CmpNUM_FTZ>;
1607 defm FSetNAN : FSET_FORMAT<setuo, CmpNAN, CmpNAN_FTZ>;
1609 //def ld_param : SDNode<"NVPTXISD::LOAD_PARAM", SDTLoad,
1610 // [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
1612 def SDTDeclareParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>,
1614 def SDTDeclareScalarParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>,
1615 SDTCisInt<1>, SDTCisInt<2>]>;
1616 def SDTLoadParamProfile : SDTypeProfile<1, 2, [SDTCisInt<1>, SDTCisInt<2>]>;
1617 def SDTLoadParamV2Profile : SDTypeProfile<2, 2, [SDTCisSameAs<0, 1>, SDTCisInt<2>, SDTCisInt<3>]>;
1618 def SDTLoadParamV4Profile : SDTypeProfile<4, 2, [SDTCisInt<4>, SDTCisInt<5>]>;
1619 def SDTPrintCallProfile : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
1620 def SDTPrintCallUniProfile : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
1621 def SDTStoreParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>]>;
1622 def SDTStoreParamV2Profile : SDTypeProfile<0, 4, [SDTCisInt<0>, SDTCisInt<1>]>;
1623 def SDTStoreParamV4Profile : SDTypeProfile<0, 6, [SDTCisInt<0>, SDTCisInt<1>]>;
1624 def SDTStoreParam32Profile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>]>;
1625 def SDTCallArgProfile : SDTypeProfile<0, 2, [SDTCisInt<0>]>;
1626 def SDTCallArgMarkProfile : SDTypeProfile<0, 0, []>;
1627 def SDTCallVoidProfile : SDTypeProfile<0, 1, []>;
1628 def SDTCallValProfile : SDTypeProfile<1, 0, []>;
1629 def SDTMoveParamProfile : SDTypeProfile<1, 1, []>;
1630 def SDTStoreRetvalProfile : SDTypeProfile<0, 2, [SDTCisInt<0>]>;
1631 def SDTStoreRetvalV2Profile : SDTypeProfile<0, 3, [SDTCisInt<0>]>;
1632 def SDTStoreRetvalV4Profile : SDTypeProfile<0, 5, [SDTCisInt<0>]>;
1633 def SDTPseudoUseParamProfile : SDTypeProfile<0, 1, []>;
1635 def DeclareParam : SDNode<"NVPTXISD::DeclareParam", SDTDeclareParamProfile,
1636 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1637 def DeclareScalarParam : SDNode<"NVPTXISD::DeclareScalarParam",
1638 SDTDeclareScalarParamProfile,
1639 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1640 def DeclareRetParam : SDNode<"NVPTXISD::DeclareRetParam",
1641 SDTDeclareParamProfile,
1642 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1643 def DeclareRet : SDNode<"NVPTXISD::DeclareRet", SDTDeclareScalarParamProfile,
1644 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1645 def LoadParam : SDNode<"NVPTXISD::LoadParam", SDTLoadParamProfile,
1646 [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
1647 def LoadParamV2 : SDNode<"NVPTXISD::LoadParamV2", SDTLoadParamV2Profile,
1648 [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
1649 def LoadParamV4 : SDNode<"NVPTXISD::LoadParamV4", SDTLoadParamV4Profile,
1650 [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
1651 def PrintCall : SDNode<"NVPTXISD::PrintCall", SDTPrintCallProfile,
1652 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1653 def PrintCallUni : SDNode<"NVPTXISD::PrintCallUni", SDTPrintCallUniProfile,
1654 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1655 def StoreParam : SDNode<"NVPTXISD::StoreParam", SDTStoreParamProfile,
1656 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1657 def StoreParamV2 : SDNode<"NVPTXISD::StoreParamV2", SDTStoreParamV2Profile,
1658 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1659 def StoreParamV4 : SDNode<"NVPTXISD::StoreParamV4", SDTStoreParamV4Profile,
1660 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1661 def StoreParamU32 : SDNode<"NVPTXISD::StoreParamU32", SDTStoreParam32Profile,
1662 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1663 def StoreParamS32 : SDNode<"NVPTXISD::StoreParamS32", SDTStoreParam32Profile,
1664 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1665 def CallArgBegin : SDNode<"NVPTXISD::CallArgBegin", SDTCallArgMarkProfile,
1666 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1667 def CallArg : SDNode<"NVPTXISD::CallArg", SDTCallArgProfile,
1668 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1669 def LastCallArg : SDNode<"NVPTXISD::LastCallArg", SDTCallArgProfile,
1670 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1671 def CallArgEnd : SDNode<"NVPTXISD::CallArgEnd", SDTCallVoidProfile,
1672 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1673 def CallVoid : SDNode<"NVPTXISD::CallVoid", SDTCallVoidProfile,
1674 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1675 def Prototype : SDNode<"NVPTXISD::Prototype", SDTCallVoidProfile,
1676 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1677 def CallVal : SDNode<"NVPTXISD::CallVal", SDTCallValProfile,
1678 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1679 def MoveParam : SDNode<"NVPTXISD::MoveParam", SDTMoveParamProfile,
1681 def StoreRetval : SDNode<"NVPTXISD::StoreRetval", SDTStoreRetvalProfile,
1682 [SDNPHasChain, SDNPSideEffect]>;
1683 def StoreRetvalV2 : SDNode<"NVPTXISD::StoreRetvalV2", SDTStoreRetvalV2Profile,
1684 [SDNPHasChain, SDNPSideEffect]>;
1685 def StoreRetvalV4 : SDNode<"NVPTXISD::StoreRetvalV4", SDTStoreRetvalV4Profile,
1686 [SDNPHasChain, SDNPSideEffect]>;
1687 def PseudoUseParam : SDNode<"NVPTXISD::PseudoUseParam",
1688 SDTPseudoUseParamProfile,
1689 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1690 def RETURNNode : SDNode<"NVPTXISD::RETURN", SDTCallArgMarkProfile,
1691 [SDNPHasChain, SDNPSideEffect]>;
1693 class LoadParamMemInst<NVPTXRegClass regclass, string opstr> :
1694 NVPTXInst<(outs regclass:$dst), (ins i32imm:$b),
1695 !strconcat(!strconcat("ld.param", opstr),
1696 "\t$dst, [retval0+$b];"),
1699 class LoadParamRegInst<NVPTXRegClass regclass, string opstr> :
1700 NVPTXInst<(outs regclass:$dst), (ins i32imm:$b),
1701 !strconcat(!strconcat("mov", opstr),
1702 "\t$dst, retval$b;"),
1703 [(set regclass:$dst, (LoadParam (i32 0), (i32 imm:$b)))]>;
1705 class LoadParamV2MemInst<NVPTXRegClass regclass, string opstr> :
1706 NVPTXInst<(outs regclass:$dst, regclass:$dst2), (ins i32imm:$b),
1707 !strconcat(!strconcat("ld.param.v2", opstr),
1708 "\t{{$dst, $dst2}}, [retval0+$b];"), []>;
1710 class LoadParamV4MemInst<NVPTXRegClass regclass, string opstr> :
1711 NVPTXInst<(outs regclass:$dst, regclass:$dst2, regclass:$dst3,
1714 !strconcat(!strconcat("ld.param.v4", opstr),
1715 "\t{{$dst, $dst2, $dst3, $dst4}}, [retval0+$b];"), []>;
1717 class StoreParamInst<NVPTXRegClass regclass, string opstr> :
1718 NVPTXInst<(outs), (ins regclass:$val, i32imm:$a, i32imm:$b),
1719 !strconcat(!strconcat("st.param", opstr),
1720 "\t[param$a+$b], $val;"),
1723 class StoreParamV2Inst<NVPTXRegClass regclass, string opstr> :
1724 NVPTXInst<(outs), (ins regclass:$val, regclass:$val2,
1725 i32imm:$a, i32imm:$b),
1726 !strconcat(!strconcat("st.param.v2", opstr),
1727 "\t[param$a+$b], {{$val, $val2}};"),
1730 class StoreParamV4Inst<NVPTXRegClass regclass, string opstr> :
1731 NVPTXInst<(outs), (ins regclass:$val, regclass:$val1, regclass:$val2,
1732 regclass:$val3, i32imm:$a, i32imm:$b),
1733 !strconcat(!strconcat("st.param.v4", opstr),
1734 "\t[param$a+$b], {{$val, $val2, $val3, $val4}};"),
1737 class StoreRetvalInst<NVPTXRegClass regclass, string opstr> :
1738 NVPTXInst<(outs), (ins regclass:$val, i32imm:$a),
1739 !strconcat(!strconcat("st.param", opstr),
1740 "\t[func_retval0+$a], $val;"),
1743 class StoreRetvalV2Inst<NVPTXRegClass regclass, string opstr> :
1744 NVPTXInst<(outs), (ins regclass:$val, regclass:$val2, i32imm:$a),
1745 !strconcat(!strconcat("st.param.v2", opstr),
1746 "\t[func_retval0+$a], {{$val, $val2}};"),
1749 class StoreRetvalV4Inst<NVPTXRegClass regclass, string opstr> :
1751 (ins regclass:$val, regclass:$val2, regclass:$val3,
1752 regclass:$val4, i32imm:$a),
1753 !strconcat(!strconcat("st.param.v4", opstr),
1754 "\t[func_retval0+$a], {{$val, $val2, $val3, $val4}};"),
1757 def PrintCallRetInst1 : NVPTXInst<(outs), (ins),
1759 [(PrintCall (i32 1))]>;
1760 def PrintCallRetInst2 : NVPTXInst<(outs), (ins),
1761 "call (retval0, retval1), ",
1762 [(PrintCall (i32 2))]>;
1763 def PrintCallRetInst3 : NVPTXInst<(outs), (ins),
1764 "call (retval0, retval1, retval2), ",
1765 [(PrintCall (i32 3))]>;
1766 def PrintCallRetInst4 : NVPTXInst<(outs), (ins),
1767 "call (retval0, retval1, retval2, retval3), ",
1768 [(PrintCall (i32 4))]>;
1769 def PrintCallRetInst5 : NVPTXInst<(outs), (ins),
1770 "call (retval0, retval1, retval2, retval3, retval4), ",
1771 [(PrintCall (i32 5))]>;
1772 def PrintCallRetInst6 : NVPTXInst<(outs), (ins),
1773 "call (retval0, retval1, retval2, retval3, retval4, retval5), ",
1774 [(PrintCall (i32 6))]>;
1775 def PrintCallRetInst7 : NVPTXInst<(outs), (ins),
1776 "call (retval0, retval1, retval2, retval3, retval4, retval5, retval6), ",
1777 [(PrintCall (i32 7))]>;
1778 def PrintCallRetInst8 : NVPTXInst<(outs), (ins),
1779 !strconcat("call (retval0, retval1, retval2, retval3, retval4",
1780 ", retval5, retval6, retval7), "),
1781 [(PrintCall (i32 8))]>;
1783 def PrintCallNoRetInst : NVPTXInst<(outs), (ins), "call ",
1784 [(PrintCall (i32 0))]>;
1786 def PrintCallUniRetInst1 : NVPTXInst<(outs), (ins),
1787 "call.uni (retval0), ",
1788 [(PrintCallUni (i32 1))]>;
1789 def PrintCallUniRetInst2 : NVPTXInst<(outs), (ins),
1790 "call.uni (retval0, retval1), ",
1791 [(PrintCallUni (i32 2))]>;
1792 def PrintCallUniRetInst3 : NVPTXInst<(outs), (ins),
1793 "call.uni (retval0, retval1, retval2), ",
1794 [(PrintCallUni (i32 3))]>;
1795 def PrintCallUniRetInst4 : NVPTXInst<(outs), (ins),
1796 "call.uni (retval0, retval1, retval2, retval3), ",
1797 [(PrintCallUni (i32 4))]>;
1798 def PrintCallUniRetInst5 : NVPTXInst<(outs), (ins),
1799 "call.uni (retval0, retval1, retval2, retval3, retval4), ",
1800 [(PrintCallUni (i32 5))]>;
1801 def PrintCallUniRetInst6 : NVPTXInst<(outs), (ins),
1802 "call.uni (retval0, retval1, retval2, retval3, retval4, retval5), ",
1803 [(PrintCallUni (i32 6))]>;
1804 def PrintCallUniRetInst7 : NVPTXInst<(outs), (ins),
1805 "call.uni (retval0, retval1, retval2, retval3, retval4, retval5, retval6), ",
1806 [(PrintCallUni (i32 7))]>;
1807 def PrintCallUniRetInst8 : NVPTXInst<(outs), (ins),
1808 !strconcat("call.uni (retval0, retval1, retval2, retval3, retval4",
1809 ", retval5, retval6, retval7), "),
1810 [(PrintCallUni (i32 8))]>;
1812 def PrintCallUniNoRetInst : NVPTXInst<(outs), (ins), "call.uni ",
1813 [(PrintCallUni (i32 0))]>;
1815 def LoadParamMemI64 : LoadParamMemInst<Int64Regs, ".b64">;
1816 def LoadParamMemI32 : LoadParamMemInst<Int32Regs, ".b32">;
1817 def LoadParamMemI16 : LoadParamMemInst<Int16Regs, ".b16">;
1818 def LoadParamMemI8 : LoadParamMemInst<Int16Regs, ".b8">;
1819 def LoadParamMemV2I64 : LoadParamV2MemInst<Int64Regs, ".b64">;
1820 def LoadParamMemV2I32 : LoadParamV2MemInst<Int32Regs, ".b32">;
1821 def LoadParamMemV2I16 : LoadParamV2MemInst<Int16Regs, ".b16">;
1822 def LoadParamMemV2I8 : LoadParamV2MemInst<Int16Regs, ".b8">;
1823 def LoadParamMemV4I32 : LoadParamV4MemInst<Int32Regs, ".b32">;
1824 def LoadParamMemV4I16 : LoadParamV4MemInst<Int16Regs, ".b16">;
1825 def LoadParamMemV4I8 : LoadParamV4MemInst<Int16Regs, ".b8">;
1826 def LoadParamMemF32 : LoadParamMemInst<Float32Regs, ".f32">;
1827 def LoadParamMemF64 : LoadParamMemInst<Float64Regs, ".f64">;
1828 def LoadParamMemV2F32 : LoadParamV2MemInst<Float32Regs, ".f32">;
1829 def LoadParamMemV2F64 : LoadParamV2MemInst<Float64Regs, ".f64">;
1830 def LoadParamMemV4F32 : LoadParamV4MemInst<Float32Regs, ".f32">;
1832 def StoreParamI64 : StoreParamInst<Int64Regs, ".b64">;
1833 def StoreParamI32 : StoreParamInst<Int32Regs, ".b32">;
1835 def StoreParamI16 : StoreParamInst<Int16Regs, ".b16">;
1836 def StoreParamI8 : StoreParamInst<Int16Regs, ".b8">;
1837 def StoreParamV2I64 : StoreParamV2Inst<Int64Regs, ".b64">;
1838 def StoreParamV2I32 : StoreParamV2Inst<Int32Regs, ".b32">;
1839 def StoreParamV2I16 : StoreParamV2Inst<Int16Regs, ".b16">;
1840 def StoreParamV2I8 : StoreParamV2Inst<Int16Regs, ".b8">;
1842 // FIXME: StoreParamV4Inst crashes llvm-tblgen :(
1843 //def StoreParamV4I32 : StoreParamV4Inst<Int32Regs, ".b32">;
1844 def StoreParamV4I32 : NVPTXInst<(outs), (ins Int32Regs:$val, Int32Regs:$val2,
1845 Int32Regs:$val3, Int32Regs:$val4,
1846 i32imm:$a, i32imm:$b),
1847 "st.param.b32\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
1850 def StoreParamV4I16 : NVPTXInst<(outs), (ins Int16Regs:$val, Int16Regs:$val2,
1851 Int16Regs:$val3, Int16Regs:$val4,
1852 i32imm:$a, i32imm:$b),
1853 "st.param.v4.b16\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
1856 def StoreParamV4I8 : NVPTXInst<(outs), (ins Int16Regs:$val, Int16Regs:$val2,
1857 Int16Regs:$val3, Int16Regs:$val4,
1858 i32imm:$a, i32imm:$b),
1859 "st.param.v4.b8\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
1862 def StoreParamF32 : StoreParamInst<Float32Regs, ".f32">;
1863 def StoreParamF64 : StoreParamInst<Float64Regs, ".f64">;
1864 def StoreParamV2F32 : StoreParamV2Inst<Float32Regs, ".f32">;
1865 def StoreParamV2F64 : StoreParamV2Inst<Float64Regs, ".f64">;
1866 // FIXME: StoreParamV4Inst crashes llvm-tblgen :(
1867 //def StoreParamV4F32 : StoreParamV4Inst<Float32Regs, ".f32">;
1868 def StoreParamV4F32 : NVPTXInst<(outs),
1869 (ins Float32Regs:$val, Float32Regs:$val2,
1870 Float32Regs:$val3, Float32Regs:$val4,
1871 i32imm:$a, i32imm:$b),
1872 "st.param.v4.f32\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
1876 def StoreRetvalI64 : StoreRetvalInst<Int64Regs, ".b64">;
1877 def StoreRetvalI32 : StoreRetvalInst<Int32Regs, ".b32">;
1878 def StoreRetvalI16 : StoreRetvalInst<Int16Regs, ".b16">;
1879 def StoreRetvalI8 : StoreRetvalInst<Int16Regs, ".b8">;
1880 def StoreRetvalV2I64 : StoreRetvalV2Inst<Int64Regs, ".b64">;
1881 def StoreRetvalV2I32 : StoreRetvalV2Inst<Int32Regs, ".b32">;
1882 def StoreRetvalV2I16 : StoreRetvalV2Inst<Int16Regs, ".b16">;
1883 def StoreRetvalV2I8 : StoreRetvalV2Inst<Int16Regs, ".b8">;
1884 def StoreRetvalV4I32 : StoreRetvalV4Inst<Int32Regs, ".b32">;
1885 def StoreRetvalV4I16 : StoreRetvalV4Inst<Int16Regs, ".b16">;
1886 def StoreRetvalV4I8 : StoreRetvalV4Inst<Int16Regs, ".b8">;
1888 def StoreRetvalF64 : StoreRetvalInst<Float64Regs, ".f64">;
1889 def StoreRetvalF32 : StoreRetvalInst<Float32Regs, ".f32">;
1890 def StoreRetvalV2F64 : StoreRetvalV2Inst<Float64Regs, ".f64">;
1891 def StoreRetvalV2F32 : StoreRetvalV2Inst<Float32Regs, ".f32">;
1892 def StoreRetvalV4F32 : StoreRetvalV4Inst<Float32Regs, ".f32">;
1894 def CallArgBeginInst : NVPTXInst<(outs), (ins), "(", [(CallArgBegin)]>;
1895 def CallArgEndInst1 : NVPTXInst<(outs), (ins), ");", [(CallArgEnd (i32 1))]>;
1896 def CallArgEndInst0 : NVPTXInst<(outs), (ins), ")", [(CallArgEnd (i32 0))]>;
1897 def RETURNInst : NVPTXInst<(outs), (ins), "ret;", [(RETURNNode)]>;
1899 class CallArgInst<NVPTXRegClass regclass> :
1900 NVPTXInst<(outs), (ins regclass:$a), "$a, ",
1901 [(CallArg (i32 0), regclass:$a)]>;
1903 class LastCallArgInst<NVPTXRegClass regclass> :
1904 NVPTXInst<(outs), (ins regclass:$a), "$a",
1905 [(LastCallArg (i32 0), regclass:$a)]>;
1907 def CallArgI64 : CallArgInst<Int64Regs>;
1908 def CallArgI32 : CallArgInst<Int32Regs>;
1909 def CallArgI16 : CallArgInst<Int16Regs>;
1911 def CallArgF64 : CallArgInst<Float64Regs>;
1912 def CallArgF32 : CallArgInst<Float32Regs>;
1914 def LastCallArgI64 : LastCallArgInst<Int64Regs>;
1915 def LastCallArgI32 : LastCallArgInst<Int32Regs>;
1916 def LastCallArgI16 : LastCallArgInst<Int16Regs>;
1918 def LastCallArgF64 : LastCallArgInst<Float64Regs>;
1919 def LastCallArgF32 : LastCallArgInst<Float32Regs>;
1921 def CallArgI32imm : NVPTXInst<(outs), (ins i32imm:$a), "$a, ",
1922 [(CallArg (i32 0), (i32 imm:$a))]>;
1923 def LastCallArgI32imm : NVPTXInst<(outs), (ins i32imm:$a), "$a",
1924 [(LastCallArg (i32 0), (i32 imm:$a))]>;
1926 def CallArgParam : NVPTXInst<(outs), (ins i32imm:$a), "param$a, ",
1927 [(CallArg (i32 1), (i32 imm:$a))]>;
1928 def LastCallArgParam : NVPTXInst<(outs), (ins i32imm:$a), "param$a",
1929 [(LastCallArg (i32 1), (i32 imm:$a))]>;
1931 def CallVoidInst : NVPTXInst<(outs), (ins imem:$addr),
1933 [(CallVoid (Wrapper tglobaladdr:$addr))]>;
1934 def CallVoidInstReg : NVPTXInst<(outs), (ins Int32Regs:$addr),
1936 [(CallVoid Int32Regs:$addr)]>;
1937 def CallVoidInstReg64 : NVPTXInst<(outs), (ins Int64Regs:$addr),
1939 [(CallVoid Int64Regs:$addr)]>;
1940 def PrototypeInst : NVPTXInst<(outs), (ins i32imm:$val),
1941 ", prototype_$val;",
1942 [(Prototype (i32 imm:$val))]>;
1944 def DeclareRetMemInst : NVPTXInst<(outs),
1945 (ins i32imm:$align, i32imm:$size, i32imm:$num),
1946 ".param .align $align .b8 retval$num[$size];",
1947 [(DeclareRetParam (i32 imm:$align), (i32 imm:$size), (i32 imm:$num))]>;
1948 def DeclareRetScalarInst : NVPTXInst<(outs), (ins i32imm:$size, i32imm:$num),
1949 ".param .b$size retval$num;",
1950 [(DeclareRet (i32 1), (i32 imm:$size), (i32 imm:$num))]>;
1951 def DeclareRetRegInst : NVPTXInst<(outs), (ins i32imm:$size, i32imm:$num),
1952 ".reg .b$size retval$num;",
1953 [(DeclareRet (i32 2), (i32 imm:$size), (i32 imm:$num))]>;
1955 def DeclareParamInst : NVPTXInst<(outs),
1956 (ins i32imm:$align, i32imm:$a, i32imm:$size),
1957 ".param .align $align .b8 param$a[$size];",
1958 [(DeclareParam (i32 imm:$align), (i32 imm:$a), (i32 imm:$size))]>;
1959 def DeclareScalarParamInst : NVPTXInst<(outs), (ins i32imm:$a, i32imm:$size),
1960 ".param .b$size param$a;",
1961 [(DeclareScalarParam (i32 imm:$a), (i32 imm:$size), (i32 0))]>;
1962 def DeclareScalarRegInst : NVPTXInst<(outs), (ins i32imm:$a, i32imm:$size),
1963 ".reg .b$size param$a;",
1964 [(DeclareScalarParam (i32 imm:$a), (i32 imm:$size), (i32 1))]>;
1966 class MoveParamInst<NVPTXRegClass regclass, string asmstr> :
1967 NVPTXInst<(outs regclass:$dst), (ins regclass:$src),
1968 !strconcat(!strconcat("mov", asmstr), "\t$dst, $src;"),
1969 [(set regclass:$dst, (MoveParam regclass:$src))]>;
1971 def MoveParamI64 : MoveParamInst<Int64Regs, ".b64">;
1972 def MoveParamI32 : MoveParamInst<Int32Regs, ".b32">;
1973 def MoveParamI16 : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
1974 "cvt.u16.u32\t$dst, $src;",
1975 [(set Int16Regs:$dst, (MoveParam Int16Regs:$src))]>;
1976 def MoveParamF64 : MoveParamInst<Float64Regs, ".f64">;
1977 def MoveParamF32 : MoveParamInst<Float32Regs, ".f32">;
1979 class PseudoUseParamInst<NVPTXRegClass regclass> :
1980 NVPTXInst<(outs), (ins regclass:$src),
1981 "// Pseudo use of $src",
1982 [(PseudoUseParam regclass:$src)]>;
1984 def PseudoUseParamI64 : PseudoUseParamInst<Int64Regs>;
1985 def PseudoUseParamI32 : PseudoUseParamInst<Int32Regs>;
1986 def PseudoUseParamI16 : PseudoUseParamInst<Int16Regs>;
1987 def PseudoUseParamF64 : PseudoUseParamInst<Float64Regs>;
1988 def PseudoUseParamF32 : PseudoUseParamInst<Float32Regs>;
1992 // Load / Store Handling
1994 multiclass LD<NVPTXRegClass regclass> {
1995 def _avar : NVPTXInst<(outs regclass:$dst),
1996 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
1997 i32imm:$fromWidth, imem:$addr),
1998 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
1999 "$fromWidth \t$dst, [$addr];"), []>;
2000 def _areg : NVPTXInst<(outs regclass:$dst),
2001 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2002 i32imm:$fromWidth, Int32Regs:$addr),
2003 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2004 "$fromWidth \t$dst, [$addr];"), []>;
2005 def _areg_64 : NVPTXInst<(outs regclass:$dst),
2006 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2007 i32imm:$fromWidth, Int64Regs:$addr),
2008 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$fromWidth",
2009 " \t$dst, [$addr];"), []>;
2010 def _ari : NVPTXInst<(outs regclass:$dst),
2011 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2012 i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
2013 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2014 "$fromWidth \t$dst, [$addr+$offset];"), []>;
2015 def _ari_64 : NVPTXInst<(outs regclass:$dst),
2016 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2017 i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
2018 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$fromWidth",
2019 " \t$dst, [$addr+$offset];"), []>;
2020 def _asi : NVPTXInst<(outs regclass:$dst),
2021 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2022 i32imm:$fromWidth, imem:$addr, i32imm:$offset),
2023 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2024 "$fromWidth \t$dst, [$addr+$offset];"), []>;
2027 let mayLoad=1, neverHasSideEffects=1 in {
2028 defm LD_i8 : LD<Int16Regs>;
2029 defm LD_i16 : LD<Int16Regs>;
2030 defm LD_i32 : LD<Int32Regs>;
2031 defm LD_i64 : LD<Int64Regs>;
2032 defm LD_f32 : LD<Float32Regs>;
2033 defm LD_f64 : LD<Float64Regs>;
2036 multiclass ST<NVPTXRegClass regclass> {
2037 def _avar : NVPTXInst<(outs),
2038 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2039 LdStCode:$Sign, i32imm:$toWidth, imem:$addr),
2040 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
2041 " \t[$addr], $src;"), []>;
2042 def _areg : NVPTXInst<(outs),
2043 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2044 LdStCode:$Sign, i32imm:$toWidth, Int32Regs:$addr),
2045 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
2046 " \t[$addr], $src;"), []>;
2047 def _areg_64 : NVPTXInst<(outs),
2048 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2049 LdStCode:$Sign, i32imm:$toWidth, Int64Regs:$addr),
2050 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth ",
2051 "\t[$addr], $src;"), []>;
2052 def _ari : NVPTXInst<(outs),
2053 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2054 LdStCode:$Sign, i32imm:$toWidth, Int32Regs:$addr, i32imm:$offset),
2055 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
2056 " \t[$addr+$offset], $src;"), []>;
2057 def _ari_64 : NVPTXInst<(outs),
2058 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2059 LdStCode:$Sign, i32imm:$toWidth, Int64Regs:$addr, i32imm:$offset),
2060 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth ",
2061 "\t[$addr+$offset], $src;"), []>;
2062 def _asi : NVPTXInst<(outs),
2063 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2064 LdStCode:$Sign, i32imm:$toWidth, imem:$addr, i32imm:$offset),
2065 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
2066 " \t[$addr+$offset], $src;"), []>;
2069 let mayStore=1, neverHasSideEffects=1 in {
2070 defm ST_i8 : ST<Int16Regs>;
2071 defm ST_i16 : ST<Int16Regs>;
2072 defm ST_i32 : ST<Int32Regs>;
2073 defm ST_i64 : ST<Int64Regs>;
2074 defm ST_f32 : ST<Float32Regs>;
2075 defm ST_f64 : ST<Float64Regs>;
2078 // The following is used only in and after vector elementizations.
2079 // Vector elementization happens at the machine instruction level, so the
2080 // following instruction
2081 // never appears in the DAG.
2082 multiclass LD_VEC<NVPTXRegClass regclass> {
2083 def _v2_avar : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2084 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2085 i32imm:$fromWidth, imem:$addr),
2086 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2087 "$fromWidth \t{{$dst1, $dst2}}, [$addr];"), []>;
2088 def _v2_areg : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2089 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2090 i32imm:$fromWidth, Int32Regs:$addr),
2091 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2092 "$fromWidth \t{{$dst1, $dst2}}, [$addr];"), []>;
2093 def _v2_areg_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2094 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2095 i32imm:$fromWidth, Int64Regs:$addr),
2096 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2097 "$fromWidth \t{{$dst1, $dst2}}, [$addr];"), []>;
2098 def _v2_ari : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2099 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2100 i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
2101 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2102 "$fromWidth \t{{$dst1, $dst2}}, [$addr+$offset];"), []>;
2103 def _v2_ari_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2104 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2105 i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
2106 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2107 "$fromWidth \t{{$dst1, $dst2}}, [$addr+$offset];"), []>;
2108 def _v2_asi : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2109 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2110 i32imm:$fromWidth, imem:$addr, i32imm:$offset),
2111 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2112 "$fromWidth \t{{$dst1, $dst2}}, [$addr+$offset];"), []>;
2113 def _v4_avar : NVPTXInst<(outs regclass:$dst1, regclass:$dst2,
2114 regclass:$dst3, regclass:$dst4),
2115 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2116 i32imm:$fromWidth, imem:$addr),
2117 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2118 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr];"), []>;
2119 def _v4_areg : NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
2121 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2122 i32imm:$fromWidth, Int32Regs:$addr),
2123 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2124 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr];"), []>;
2125 def _v4_areg_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2,
2126 regclass:$dst3, regclass:$dst4),
2127 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2128 i32imm:$fromWidth, Int64Regs:$addr),
2129 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2130 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr];"), []>;
2131 def _v4_ari : NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
2133 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2134 i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
2135 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2136 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr+$offset];"),
2138 def _v4_ari_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2,
2139 regclass:$dst3, regclass:$dst4),
2140 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2141 i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
2142 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2143 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr+$offset];"),
2145 def _v4_asi : NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
2147 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2148 i32imm:$fromWidth, imem:$addr, i32imm:$offset),
2149 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2150 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr+$offset];"),
2153 let mayLoad=1, neverHasSideEffects=1 in {
2154 defm LDV_i8 : LD_VEC<Int16Regs>;
2155 defm LDV_i16 : LD_VEC<Int16Regs>;
2156 defm LDV_i32 : LD_VEC<Int32Regs>;
2157 defm LDV_i64 : LD_VEC<Int64Regs>;
2158 defm LDV_f32 : LD_VEC<Float32Regs>;
2159 defm LDV_f64 : LD_VEC<Float64Regs>;
2162 multiclass ST_VEC<NVPTXRegClass regclass> {
2163 def _v2_avar : NVPTXInst<(outs),
2164 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2165 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, imem:$addr),
2166 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2167 "$fromWidth \t[$addr], {{$src1, $src2}};"), []>;
2168 def _v2_areg : NVPTXInst<(outs),
2169 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2170 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int32Regs:$addr),
2171 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2172 "$fromWidth \t[$addr], {{$src1, $src2}};"), []>;
2173 def _v2_areg_64 : NVPTXInst<(outs),
2174 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2175 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int64Regs:$addr),
2176 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2177 "$fromWidth \t[$addr], {{$src1, $src2}};"), []>;
2178 def _v2_ari : NVPTXInst<(outs),
2179 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2180 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int32Regs:$addr,
2182 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2183 "$fromWidth \t[$addr+$offset], {{$src1, $src2}};"), []>;
2184 def _v2_ari_64 : NVPTXInst<(outs),
2185 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2186 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int64Regs:$addr,
2188 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2189 "$fromWidth \t[$addr+$offset], {{$src1, $src2}};"), []>;
2190 def _v2_asi : NVPTXInst<(outs),
2191 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2192 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, imem:$addr,
2194 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2195 "$fromWidth \t[$addr+$offset], {{$src1, $src2}};"), []>;
2196 def _v4_avar : NVPTXInst<(outs),
2197 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2198 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2199 i32imm:$fromWidth, imem:$addr),
2200 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2201 "$fromWidth \t[$addr], {{$src1, $src2, $src3, $src4}};"), []>;
2202 def _v4_areg : NVPTXInst<(outs),
2203 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2204 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2205 i32imm:$fromWidth, Int32Regs:$addr),
2206 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2207 "$fromWidth \t[$addr], {{$src1, $src2, $src3, $src4}};"), []>;
2208 def _v4_areg_64 : NVPTXInst<(outs),
2209 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2210 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2211 i32imm:$fromWidth, Int64Regs:$addr),
2212 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2213 "$fromWidth \t[$addr], {{$src1, $src2, $src3, $src4}};"), []>;
2214 def _v4_ari : NVPTXInst<(outs),
2215 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2216 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2217 i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
2218 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2219 "$fromWidth \t[$addr+$offset], {{$src1, $src2, $src3, $src4}};"),
2221 def _v4_ari_64 : NVPTXInst<(outs),
2222 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2223 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2224 i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
2225 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2226 "$fromWidth \t[$addr+$offset], {{$src1, $src2, $src3, $src4}};"),
2228 def _v4_asi : NVPTXInst<(outs),
2229 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2230 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2231 i32imm:$fromWidth, imem:$addr, i32imm:$offset),
2232 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2233 "$fromWidth \t[$addr+$offset], {{$src1, $src2, $src3, $src4}};"),
2236 let mayStore=1, neverHasSideEffects=1 in {
2237 defm STV_i8 : ST_VEC<Int16Regs>;
2238 defm STV_i16 : ST_VEC<Int16Regs>;
2239 defm STV_i32 : ST_VEC<Int32Regs>;
2240 defm STV_i64 : ST_VEC<Int64Regs>;
2241 defm STV_f32 : ST_VEC<Float32Regs>;
2242 defm STV_f64 : ST_VEC<Float64Regs>;
2246 //---- Conversion ----
2248 class F_BITCONVERT<string SzStr, NVPTXRegClass regclassIn,
2249 NVPTXRegClass regclassOut> :
2250 NVPTXInst<(outs regclassOut:$d), (ins regclassIn:$a),
2251 !strconcat("mov.b", !strconcat(SzStr, " \t $d, $a;")),
2252 [(set regclassOut:$d, (bitconvert regclassIn:$a))]>;
2254 def BITCONVERT_32_I2F : F_BITCONVERT<"32", Int32Regs, Float32Regs>;
2255 def BITCONVERT_32_F2I : F_BITCONVERT<"32", Float32Regs, Int32Regs>;
2256 def BITCONVERT_64_I2F : F_BITCONVERT<"64", Int64Regs, Float64Regs>;
2257 def BITCONVERT_64_F2I : F_BITCONVERT<"64", Float64Regs, Int64Regs>;
2259 // NOTE: pred->fp are currently sub-optimal due to an issue in TableGen where
2260 // we cannot specify floating-point literals in isel patterns. Therefore, we
2261 // use an integer selp to select either 1 or 0 and then cvt to floating-point.
2264 def : Pat<(f32 (sint_to_fp Int1Regs:$a)),
2265 (CVT_f32_s32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
2266 def : Pat<(f32 (sint_to_fp Int16Regs:$a)),
2267 (CVT_f32_s16 Int16Regs:$a, CvtRN)>;
2268 def : Pat<(f32 (sint_to_fp Int32Regs:$a)),
2269 (CVT_f32_s32 Int32Regs:$a, CvtRN)>;
2270 def : Pat<(f32 (sint_to_fp Int64Regs:$a)),
2271 (CVT_f32_s64 Int64Regs:$a, CvtRN)>;
2274 def : Pat<(f32 (uint_to_fp Int1Regs:$a)),
2275 (CVT_f32_u32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
2276 def : Pat<(f32 (uint_to_fp Int16Regs:$a)),
2277 (CVT_f32_u16 Int16Regs:$a, CvtRN)>;
2278 def : Pat<(f32 (uint_to_fp Int32Regs:$a)),
2279 (CVT_f32_u32 Int32Regs:$a, CvtRN)>;
2280 def : Pat<(f32 (uint_to_fp Int64Regs:$a)),
2281 (CVT_f32_u64 Int64Regs:$a, CvtRN)>;
2284 def : Pat<(f64 (sint_to_fp Int1Regs:$a)),
2285 (CVT_f64_s32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
2286 def : Pat<(f64 (sint_to_fp Int16Regs:$a)),
2287 (CVT_f64_s16 Int16Regs:$a, CvtRN)>;
2288 def : Pat<(f64 (sint_to_fp Int32Regs:$a)),
2289 (CVT_f64_s32 Int32Regs:$a, CvtRN)>;
2290 def : Pat<(f64 (sint_to_fp Int64Regs:$a)),
2291 (CVT_f64_s64 Int64Regs:$a, CvtRN)>;
2294 def : Pat<(f64 (uint_to_fp Int1Regs:$a)),
2295 (CVT_f64_u32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
2296 def : Pat<(f64 (uint_to_fp Int16Regs:$a)),
2297 (CVT_f64_u16 Int16Regs:$a, CvtRN)>;
2298 def : Pat<(f64 (uint_to_fp Int32Regs:$a)),
2299 (CVT_f64_u32 Int32Regs:$a, CvtRN)>;
2300 def : Pat<(f64 (uint_to_fp Int64Regs:$a)),
2301 (CVT_f64_u64 Int64Regs:$a, CvtRN)>;
2305 def : Pat<(i1 (fp_to_sint Float32Regs:$a)),
2306 (SETP_b32ri (BITCONVERT_32_F2I Float32Regs:$a), 0, CmpEQ)>;
2307 def : Pat<(i16 (fp_to_sint Float32Regs:$a)),
2308 (CVT_s16_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2309 def : Pat<(i16 (fp_to_sint Float32Regs:$a)),
2310 (CVT_s16_f32 Float32Regs:$a, CvtRZI)>;
2311 def : Pat<(i32 (fp_to_sint Float32Regs:$a)),
2312 (CVT_s32_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2313 def : Pat<(i32 (fp_to_sint Float32Regs:$a)),
2314 (CVT_s32_f32 Float32Regs:$a, CvtRZI)>;
2315 def : Pat<(i64 (fp_to_sint Float32Regs:$a)),
2316 (CVT_s64_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2317 def : Pat<(i64 (fp_to_sint Float32Regs:$a)),
2318 (CVT_s64_f32 Float32Regs:$a, CvtRZI)>;
2321 def : Pat<(i1 (fp_to_uint Float32Regs:$a)),
2322 (SETP_b32ri (BITCONVERT_32_F2I Float32Regs:$a), 0, CmpEQ)>;
2323 def : Pat<(i16 (fp_to_uint Float32Regs:$a)),
2324 (CVT_u16_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2325 def : Pat<(i16 (fp_to_uint Float32Regs:$a)),
2326 (CVT_u16_f32 Float32Regs:$a, CvtRZI)>;
2327 def : Pat<(i32 (fp_to_uint Float32Regs:$a)),
2328 (CVT_u32_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2329 def : Pat<(i32 (fp_to_uint Float32Regs:$a)),
2330 (CVT_u32_f32 Float32Regs:$a, CvtRZI)>;
2331 def : Pat<(i64 (fp_to_uint Float32Regs:$a)),
2332 (CVT_u64_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2333 def : Pat<(i64 (fp_to_uint Float32Regs:$a)),
2334 (CVT_u64_f32 Float32Regs:$a, CvtRZI)>;
2337 def : Pat<(i1 (fp_to_sint Float64Regs:$a)),
2338 (SETP_b64ri (BITCONVERT_64_F2I Float64Regs:$a), 0, CmpEQ)>;
2339 def : Pat<(i16 (fp_to_sint Float64Regs:$a)),
2340 (CVT_s16_f64 Float64Regs:$a, CvtRZI)>;
2341 def : Pat<(i32 (fp_to_sint Float64Regs:$a)),
2342 (CVT_s32_f64 Float64Regs:$a, CvtRZI)>;
2343 def : Pat<(i64 (fp_to_sint Float64Regs:$a)),
2344 (CVT_s64_f64 Float64Regs:$a, CvtRZI)>;
2347 def : Pat<(i1 (fp_to_uint Float64Regs:$a)),
2348 (SETP_b64ri (BITCONVERT_64_F2I Float64Regs:$a), 0, CmpEQ)>;
2349 def : Pat<(i16 (fp_to_uint Float64Regs:$a)),
2350 (CVT_u16_f64 Float64Regs:$a, CvtRZI)>;
2351 def : Pat<(i32 (fp_to_uint Float64Regs:$a)),
2352 (CVT_u32_f64 Float64Regs:$a, CvtRZI)>;
2353 def : Pat<(i64 (fp_to_uint Float64Regs:$a)),
2354 (CVT_u64_f64 Float64Regs:$a, CvtRZI)>;
2357 def : Pat<(i16 (sext Int1Regs:$a)),
2358 (SELP_s16ii -1, 0, Int1Regs:$a)>;
2359 def : Pat<(i32 (sext Int1Regs:$a)),
2360 (SELP_s32ii -1, 0, Int1Regs:$a)>;
2361 def : Pat<(i64 (sext Int1Regs:$a)),
2362 (SELP_s64ii -1, 0, Int1Regs:$a)>;
2365 def : Pat<(i16 (zext Int1Regs:$a)),
2366 (SELP_u16ii 1, 0, Int1Regs:$a)>;
2367 def : Pat<(i32 (zext Int1Regs:$a)),
2368 (SELP_u32ii 1, 0, Int1Regs:$a)>;
2369 def : Pat<(i64 (zext Int1Regs:$a)),
2370 (SELP_u64ii 1, 0, Int1Regs:$a)>;
2373 def : Pat<(i16 (anyext Int1Regs:$a)),
2374 (SELP_u16ii -1, 0, Int1Regs:$a)>;
2375 def : Pat<(i32 (anyext Int1Regs:$a)),
2376 (SELP_u32ii -1, 0, Int1Regs:$a)>;
2377 def : Pat<(i64 (anyext Int1Regs:$a)),
2378 (SELP_u64ii -1, 0, Int1Regs:$a)>;
2381 def : Pat<(i32 (sext Int16Regs:$a)),
2382 (CVT_s32_s16 Int16Regs:$a, CvtNONE)>;
2383 def : Pat<(i64 (sext Int16Regs:$a)),
2384 (CVT_s64_s16 Int16Regs:$a, CvtNONE)>;
2387 def : Pat<(i32 (zext Int16Regs:$a)),
2388 (CVT_u32_u16 Int16Regs:$a, CvtNONE)>;
2389 def : Pat<(i64 (zext Int16Regs:$a)),
2390 (CVT_u64_u16 Int16Regs:$a, CvtNONE)>;
2393 def : Pat<(i32 (anyext Int16Regs:$a)),
2394 (CVT_u32_u16 Int16Regs:$a, CvtNONE)>;
2395 def : Pat<(i64 (anyext Int16Regs:$a)),
2396 (CVT_u64_u16 Int16Regs:$a, CvtNONE)>;
2399 def : Pat<(i64 (sext Int32Regs:$a)),
2400 (CVT_s64_s32 Int32Regs:$a, CvtNONE)>;
2403 def : Pat<(i64 (zext Int32Regs:$a)),
2404 (CVT_u64_u32 Int32Regs:$a, CvtNONE)>;
2407 def : Pat<(i64 (anyext Int32Regs:$a)),
2408 (CVT_u64_u32 Int32Regs:$a, CvtNONE)>;
2412 def : Pat<(i32 (trunc Int64Regs:$a)),
2413 (CVT_u32_u64 Int64Regs:$a, CvtNONE)>;
2414 def : Pat<(i16 (trunc Int64Regs:$a)),
2415 (CVT_u16_u64 Int64Regs:$a, CvtNONE)>;
2416 def : Pat<(i1 (trunc Int64Regs:$a)),
2417 (SETP_b64ri (ANDb64ri Int64Regs:$a, 1), 1, CmpEQ)>;
2420 def : Pat<(i16 (trunc Int32Regs:$a)),
2421 (CVT_u16_u32 Int32Regs:$a, CvtNONE)>;
2422 def : Pat<(i1 (trunc Int32Regs:$a)),
2423 (SETP_b32ri (ANDb32ri Int32Regs:$a, 1), 1, CmpEQ)>;
2426 def : Pat<(i1 (trunc Int16Regs:$a)),
2427 (SETP_b16ri (ANDb16ri Int16Regs:$a, 1), 1, CmpEQ)>;
2430 def : Pat<(sext_inreg Int16Regs:$a, i8), (CVT_INREG_s16_s8 Int16Regs:$a)>;
2431 def : Pat<(sext_inreg Int32Regs:$a, i8), (CVT_INREG_s32_s8 Int32Regs:$a)>;
2432 def : Pat<(sext_inreg Int32Regs:$a, i16), (CVT_INREG_s32_s16 Int32Regs:$a)>;
2433 def : Pat<(sext_inreg Int64Regs:$a, i8), (CVT_INREG_s64_s8 Int64Regs:$a)>;
2434 def : Pat<(sext_inreg Int64Regs:$a, i16), (CVT_INREG_s64_s16 Int64Regs:$a)>;
2435 def : Pat<(sext_inreg Int64Regs:$a, i32), (CVT_INREG_s64_s32 Int64Regs:$a)>;
2438 // Select instructions with 32-bit predicates
2439 def : Pat<(select Int32Regs:$pred, Int16Regs:$a, Int16Regs:$b),
2440 (SELP_b16rr Int16Regs:$a, Int16Regs:$b,
2441 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2442 def : Pat<(select Int32Regs:$pred, Int32Regs:$a, Int32Regs:$b),
2443 (SELP_b32rr Int32Regs:$a, Int32Regs:$b,
2444 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2445 def : Pat<(select Int32Regs:$pred, Int64Regs:$a, Int64Regs:$b),
2446 (SELP_b64rr Int64Regs:$a, Int64Regs:$b,
2447 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2448 def : Pat<(select Int32Regs:$pred, Float32Regs:$a, Float32Regs:$b),
2449 (SELP_f32rr Float32Regs:$a, Float32Regs:$b,
2450 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2451 def : Pat<(select Int32Regs:$pred, Float64Regs:$a, Float64Regs:$b),
2452 (SELP_f64rr Float64Regs:$a, Float64Regs:$b,
2453 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2456 // pack a set of smaller int registers to a larger int register
2457 def V4I16toI64 : NVPTXInst<(outs Int64Regs:$d),
2458 (ins Int16Regs:$s1, Int16Regs:$s2,
2459 Int16Regs:$s3, Int16Regs:$s4),
2460 "mov.b64\t$d, {{$s1, $s2, $s3, $s4}};",
2462 def V2I16toI32 : NVPTXInst<(outs Int32Regs:$d),
2463 (ins Int16Regs:$s1, Int16Regs:$s2),
2464 "mov.b32\t$d, {{$s1, $s2}};",
2466 def V2I32toI64 : NVPTXInst<(outs Int64Regs:$d),
2467 (ins Int32Regs:$s1, Int32Regs:$s2),
2468 "mov.b64\t$d, {{$s1, $s2}};",
2470 def V2F32toF64 : NVPTXInst<(outs Float64Regs:$d),
2471 (ins Float32Regs:$s1, Float32Regs:$s2),
2472 "mov.b64\t$d, {{$s1, $s2}};",
2475 // unpack a larger int register to a set of smaller int registers
2476 def I64toV4I16 : NVPTXInst<(outs Int16Regs:$d1, Int16Regs:$d2,
2477 Int16Regs:$d3, Int16Regs:$d4),
2479 "mov.b64\t{{$d1, $d2, $d3, $d4}}, $s;",
2481 def I32toV2I16 : NVPTXInst<(outs Int16Regs:$d1, Int16Regs:$d2),
2483 "mov.b32\t{{$d1, $d2}}, $s;",
2485 def I64toV2I32 : NVPTXInst<(outs Int32Regs:$d1, Int32Regs:$d2),
2487 "mov.b64\t{{$d1, $d2}}, $s;",
2489 def F64toV2F32 : NVPTXInst<(outs Float32Regs:$d1, Float32Regs:$d2),
2490 (ins Float64Regs:$s),
2491 "mov.b64\t{{$d1, $d2}}, $s;",
2494 // Count leading zeros
2495 def CLZr32 : NVPTXInst<(outs Int32Regs:$d), (ins Int32Regs:$a),
2498 def CLZr64 : NVPTXInst<(outs Int32Regs:$d), (ins Int64Regs:$a),
2502 // 32-bit has a direct PTX instruction
2503 def : Pat<(ctlz Int32Regs:$a),
2504 (CLZr32 Int32Regs:$a)>;
2505 def : Pat<(ctlz_zero_undef Int32Regs:$a),
2506 (CLZr32 Int32Regs:$a)>;
2508 // For 64-bit, the result in PTX is actually 32-bit so we zero-extend
2509 // to 64-bit to match the LLVM semantics
2510 def : Pat<(ctlz Int64Regs:$a),
2511 (CVT_u64_u32 (CLZr64 Int64Regs:$a), CvtNONE)>;
2512 def : Pat<(ctlz_zero_undef Int64Regs:$a),
2513 (CVT_u64_u32 (CLZr64 Int64Regs:$a), CvtNONE)>;
2515 // For 16-bit, we zero-extend to 32-bit, then trunc the result back
2516 // to 16-bits (ctlz of a 16-bit value is guaranteed to require less
2517 // than 16 bits to store). We also need to subtract 16 because the
2518 // high-order 16 zeros were counted.
2519 def : Pat<(ctlz Int16Regs:$a),
2520 (SUBi16ri (CVT_u16_u32 (CLZr32
2521 (CVT_u32_u16 Int16Regs:$a, CvtNONE)),
2523 def : Pat<(ctlz_zero_undef Int16Regs:$a),
2524 (SUBi16ri (CVT_u16_u32 (CLZr32
2525 (CVT_u32_u16 Int16Regs:$a, CvtNONE)),
2529 def POPCr32 : NVPTXInst<(outs Int32Regs:$d), (ins Int32Regs:$a),
2530 "popc.b32\t$d, $a;",
2532 def POPCr64 : NVPTXInst<(outs Int32Regs:$d), (ins Int64Regs:$a),
2533 "popc.b64\t$d, $a;",
2536 // 32-bit has a direct PTX instruction
2537 def : Pat<(ctpop Int32Regs:$a),
2538 (POPCr32 Int32Regs:$a)>;
2540 // For 64-bit, the result in PTX is actually 32-bit so we zero-extend
2541 // to 64-bit to match the LLVM semantics
2542 def : Pat<(ctpop Int64Regs:$a),
2543 (CVT_u64_u32 (POPCr64 Int64Regs:$a), CvtNONE)>;
2545 // For 16-bit, we zero-extend to 32-bit, then trunc the result back
2546 // to 16-bits (ctpop of a 16-bit value is guaranteed to require less
2547 // than 16 bits to store)
2548 def : Pat<(ctpop Int16Regs:$a),
2549 (CVT_u16_u32 (POPCr32 (CVT_u32_u16 Int16Regs:$a, CvtNONE)),
2552 // fround f64 -> f32
2553 def : Pat<(f32 (fround Float64Regs:$a)),
2554 (CVT_f32_f64 Float64Regs:$a, CvtRN_FTZ)>, Requires<[doF32FTZ]>;
2555 def : Pat<(f32 (fround Float64Regs:$a)),
2556 (CVT_f32_f64 Float64Regs:$a, CvtRN)>;
2558 // fextend f32 -> f64
2559 def : Pat<(f64 (fextend Float32Regs:$a)),
2560 (CVT_f64_f32 Float32Regs:$a, CvtNONE_FTZ)>, Requires<[doF32FTZ]>;
2561 def : Pat<(f64 (fextend Float32Regs:$a)),
2562 (CVT_f64_f32 Float32Regs:$a, CvtNONE)>;
2564 def retflag : SDNode<"NVPTXISD::RET_FLAG", SDTNone,
2565 [SDNPHasChain, SDNPOptInGlue]>;
2567 //-----------------------------------
2569 //-----------------------------------
2571 let isTerminator=1 in {
2572 let isReturn=1, isBarrier=1 in
2573 def Return : NVPTXInst<(outs), (ins), "ret;", [(retflag)]>;
2576 def CBranch : NVPTXInst<(outs), (ins Int1Regs:$a, brtarget:$target),
2577 "@$a bra \t$target;",
2578 [(brcond Int1Regs:$a, bb:$target)]>;
2580 def CBranchOther : NVPTXInst<(outs), (ins Int1Regs:$a, brtarget:$target),
2581 "@!$a bra \t$target;",
2584 let isBranch=1, isBarrier=1 in
2585 def GOTO : NVPTXInst<(outs), (ins brtarget:$target),
2586 "bra.uni \t$target;",
2590 def : Pat<(brcond Int32Regs:$a, bb:$target),
2591 (CBranch (SETP_u32ri Int32Regs:$a, 0, CmpNE), bb:$target)>;
2593 // SelectionDAGBuilder::visitSWitchCase() will invert the condition of a
2594 // conditional branch if
2595 // the target block is the next block so that the code can fall through to the
2597 // The invertion is done by 'xor condition, 1', which will be translated to
2598 // (setne condition, -1).
2599 // Since ptx supports '@!pred bra target', we should use it.
2600 def : Pat<(brcond (i1 (setne Int1Regs:$a, -1)), bb:$target),
2601 (CBranchOther Int1Regs:$a, bb:$target)>;
2604 def SDT_NVPTXCallSeqStart : SDCallSeqStart<[ SDTCisVT<0, i32> ]>;
2605 def SDT_NVPTXCallSeqEnd : SDCallSeqEnd<[ SDTCisVT<0, i32>,
2606 SDTCisVT<1, i32> ]>;
2608 def callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_NVPTXCallSeqStart,
2609 [SDNPHasChain, SDNPOutGlue, SDNPSideEffect]>;
2610 def callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_NVPTXCallSeqEnd,
2611 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue,
2614 def SDT_NVPTXCall : SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
2615 def call : SDNode<"NVPTXISD::CALL", SDT_NVPTXCall,
2616 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
2617 def calltarget : Operand<i32>;
2619 def CALL : NVPTXInst<(outs), (ins calltarget:$dst),
2620 "call \t$dst, (1);", []>;
2623 def : Pat<(call tglobaladdr:$dst),
2624 (CALL tglobaladdr:$dst)>;
2625 def : Pat<(call texternalsym:$dst),
2626 (CALL texternalsym:$dst)>;
2628 // Pseudo instructions.
2629 class Pseudo<dag outs, dag ins, string asmstr, list<dag> pattern>
2630 : NVPTXInst<outs, ins, asmstr, pattern>;
2632 // @TODO: We use some tricks here to emit curly braces. Can we clean this up
2633 // a bit without TableGen modifications?
2634 def Callseq_Start : NVPTXInst<(outs), (ins i32imm:$amt),
2635 "// Callseq Start $amt\n\t{{\n\t.reg .b32 temp_param_reg;\n\t// <end>}}",
2636 [(callseq_start timm:$amt)]>;
2637 def Callseq_End : NVPTXInst<(outs), (ins i32imm:$amt1, i32imm:$amt2),
2638 "\n\t//{{\n\t}}// Callseq End $amt1",
2639 [(callseq_end timm:$amt1, timm:$amt2)]>;
2643 def trapinst : NVPTXInst<(outs), (ins),
2647 // Call prototype wrapper
2648 def SDTCallPrototype : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
2650 : SDNode<"NVPTXISD::CallPrototype", SDTCallPrototype,
2651 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
2652 def ProtoIdent : Operand<i32> {
2653 let PrintMethod = "printProtoIdent";
2656 : NVPTXInst<(outs), (ins ProtoIdent:$ident),
2657 "$ident", [(CallPrototype (i32 texternalsym:$ident))]>;
2661 include "NVPTXIntrinsics.td"
2664 //-----------------------------------
2666 //-----------------------------------
2667 // BSWAP is currently expanded. The following is a more efficient
2668 // - for < sm_20, use vector scalar mov, as tesla support native 16-bit register
2669 // - for sm_20, use pmpt (use vector scalar mov to get the pack and
2670 // unpack). sm_20 supports native 32-bit register, but not native 16-bit