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()">;
161 def noHWROT32 : Predicate<"!Subtarget.hasHWROT32()">;
163 def true : Predicate<"1">;
165 def hasPTX31 : Predicate<"Subtarget.getPTXVersion() >= 31">;
168 //===----------------------------------------------------------------------===//
169 // Some Common Instruction Class Templates
170 //===----------------------------------------------------------------------===//
172 multiclass I3<string OpcStr, SDNode OpNode> {
173 def i64rr : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, Int64Regs:$b),
174 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
175 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
177 def i64ri : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i64imm:$b),
178 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
179 [(set Int64Regs:$dst, (OpNode Int64Regs:$a, imm:$b))]>;
180 def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, Int32Regs:$b),
181 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
182 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
184 def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
185 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
186 [(set Int32Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
187 def i16rr : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
188 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
189 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
191 def i16ri : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
192 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
193 [(set Int16Regs:$dst, (OpNode Int16Regs:$a, (imm):$b))]>;
196 multiclass ADD_SUB_INT_32<string OpcStr, SDNode OpNode> {
197 def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a,
199 !strconcat(OpcStr, ".s32 \t$dst, $a, $b;"),
200 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
202 def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
203 !strconcat(OpcStr, ".s32 \t$dst, $a, $b;"),
204 [(set Int32Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
207 multiclass F3<string OpcStr, SDNode OpNode> {
208 def f64rr : NVPTXInst<(outs Float64Regs:$dst),
209 (ins Float64Regs:$a, Float64Regs:$b),
210 !strconcat(OpcStr, ".f64 \t$dst, $a, $b;"),
211 [(set Float64Regs:$dst,
212 (OpNode Float64Regs:$a, Float64Regs:$b))]>,
213 Requires<[allowFMA]>;
214 def f64ri : NVPTXInst<(outs Float64Regs:$dst),
215 (ins Float64Regs:$a, f64imm:$b),
216 !strconcat(OpcStr, ".f64 \t$dst, $a, $b;"),
217 [(set Float64Regs:$dst,
218 (OpNode Float64Regs:$a, fpimm:$b))]>,
219 Requires<[allowFMA]>;
220 def f32rr_ftz : NVPTXInst<(outs Float32Regs:$dst),
221 (ins Float32Regs:$a, Float32Regs:$b),
222 !strconcat(OpcStr, ".ftz.f32 \t$dst, $a, $b;"),
223 [(set Float32Regs:$dst,
224 (OpNode Float32Regs:$a, Float32Regs:$b))]>,
225 Requires<[allowFMA_ftz]>;
226 def f32ri_ftz : NVPTXInst<(outs Float32Regs:$dst),
227 (ins Float32Regs:$a, f32imm:$b),
228 !strconcat(OpcStr, ".ftz.f32 \t$dst, $a, $b;"),
229 [(set Float32Regs:$dst,
230 (OpNode Float32Regs:$a, fpimm:$b))]>,
231 Requires<[allowFMA_ftz]>;
232 def f32rr : NVPTXInst<(outs Float32Regs:$dst),
233 (ins Float32Regs:$a, Float32Regs:$b),
234 !strconcat(OpcStr, ".f32 \t$dst, $a, $b;"),
235 [(set Float32Regs:$dst,
236 (OpNode Float32Regs:$a, Float32Regs:$b))]>,
237 Requires<[allowFMA]>;
238 def f32ri : NVPTXInst<(outs Float32Regs:$dst),
239 (ins Float32Regs:$a, f32imm:$b),
240 !strconcat(OpcStr, ".f32 \t$dst, $a, $b;"),
241 [(set Float32Regs:$dst,
242 (OpNode Float32Regs:$a, fpimm:$b))]>,
243 Requires<[allowFMA]>;
246 multiclass F3_rn<string OpcStr, SDNode OpNode> {
247 def f64rr : NVPTXInst<(outs Float64Regs:$dst),
248 (ins Float64Regs:$a, Float64Regs:$b),
249 !strconcat(OpcStr, ".rn.f64 \t$dst, $a, $b;"),
250 [(set Float64Regs:$dst,
251 (OpNode Float64Regs:$a, Float64Regs:$b))]>;
252 def f64ri : NVPTXInst<(outs Float64Regs:$dst),
253 (ins Float64Regs:$a, f64imm:$b),
254 !strconcat(OpcStr, ".rn.f64 \t$dst, $a, $b;"),
255 [(set Float64Regs:$dst,
256 (OpNode Float64Regs:$a, fpimm:$b))]>;
257 def f32rr_ftz : NVPTXInst<(outs Float32Regs:$dst),
258 (ins Float32Regs:$a, Float32Regs:$b),
259 !strconcat(OpcStr, ".rn.ftz.f32 \t$dst, $a, $b;"),
260 [(set Float32Regs:$dst,
261 (OpNode Float32Regs:$a, Float32Regs:$b))]>,
262 Requires<[doF32FTZ]>;
263 def f32ri_ftz : NVPTXInst<(outs Float32Regs:$dst),
264 (ins Float32Regs:$a, f32imm:$b),
265 !strconcat(OpcStr, ".rn.ftz.f32 \t$dst, $a, $b;"),
266 [(set Float32Regs:$dst,
267 (OpNode Float32Regs:$a, fpimm:$b))]>,
268 Requires<[doF32FTZ]>;
269 def f32rr : NVPTXInst<(outs Float32Regs:$dst),
270 (ins Float32Regs:$a, Float32Regs:$b),
271 !strconcat(OpcStr, ".rn.f32 \t$dst, $a, $b;"),
272 [(set Float32Regs:$dst,
273 (OpNode Float32Regs:$a, Float32Regs:$b))]>;
274 def f32ri : NVPTXInst<(outs Float32Regs:$dst),
275 (ins Float32Regs:$a, f32imm:$b),
276 !strconcat(OpcStr, ".rn.f32 \t$dst, $a, $b;"),
277 [(set Float32Regs:$dst,
278 (OpNode Float32Regs:$a, fpimm:$b))]>;
281 multiclass F2<string OpcStr, SDNode OpNode> {
282 def f64 : NVPTXInst<(outs Float64Regs:$dst), (ins Float64Regs:$a),
283 !strconcat(OpcStr, ".f64 \t$dst, $a;"),
284 [(set Float64Regs:$dst, (OpNode Float64Regs:$a))]>;
285 def f32_ftz : NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$a),
286 !strconcat(OpcStr, ".ftz.f32 \t$dst, $a;"),
287 [(set Float32Regs:$dst, (OpNode Float32Regs:$a))]>,
288 Requires<[doF32FTZ]>;
289 def f32 : NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$a),
290 !strconcat(OpcStr, ".f32 \t$dst, $a;"),
291 [(set Float32Regs:$dst, (OpNode Float32Regs:$a))]>;
294 //===----------------------------------------------------------------------===//
295 // NVPTX Instructions.
296 //===----------------------------------------------------------------------===//
298 //-----------------------------------
299 // General Type Conversion
300 //-----------------------------------
302 let neverHasSideEffects = 1 in {
303 // Generate a cvt to the given type from all possible types.
304 // Each instance takes a CvtMode immediate that defines the conversion mode to
305 // use. It can be CvtNONE to omit a conversion mode.
306 multiclass CVT_FROM_ALL<string FromName, RegisterClass RC> {
307 def _s16 : NVPTXInst<(outs RC:$dst),
308 (ins Int16Regs:$src, CvtMode:$mode),
309 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
310 FromName, ".s16\t$dst, $src;"),
312 def _u16 : NVPTXInst<(outs RC:$dst),
313 (ins Int16Regs:$src, CvtMode:$mode),
314 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
315 FromName, ".u16\t$dst, $src;"),
317 def _f16 : NVPTXInst<(outs RC:$dst),
318 (ins Int16Regs:$src, CvtMode:$mode),
319 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
320 FromName, ".f16\t$dst, $src;"),
322 def _s32 : NVPTXInst<(outs RC:$dst),
323 (ins Int32Regs:$src, CvtMode:$mode),
324 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
325 FromName, ".s32\t$dst, $src;"),
327 def _u32 : NVPTXInst<(outs RC:$dst),
328 (ins Int32Regs:$src, CvtMode:$mode),
329 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
330 FromName, ".u32\t$dst, $src;"),
332 def _s64 : NVPTXInst<(outs RC:$dst),
333 (ins Int64Regs:$src, CvtMode:$mode),
334 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
335 FromName, ".s64\t$dst, $src;"),
337 def _u64 : NVPTXInst<(outs RC:$dst),
338 (ins Int64Regs:$src, CvtMode:$mode),
339 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
340 FromName, ".u64\t$dst, $src;"),
342 def _f32 : NVPTXInst<(outs RC:$dst),
343 (ins Float32Regs:$src, CvtMode:$mode),
344 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
345 FromName, ".f32\t$dst, $src;"),
347 def _f64 : NVPTXInst<(outs RC:$dst),
348 (ins Float64Regs:$src, CvtMode:$mode),
349 !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
350 FromName, ".f64\t$dst, $src;"),
354 // Generate a cvt to all possible types.
355 defm CVT_s16 : CVT_FROM_ALL<"s16", Int16Regs>;
356 defm CVT_u16 : CVT_FROM_ALL<"u16", Int16Regs>;
357 defm CVT_f16 : CVT_FROM_ALL<"f16", Int16Regs>;
358 defm CVT_s32 : CVT_FROM_ALL<"s32", Int32Regs>;
359 defm CVT_u32 : CVT_FROM_ALL<"u32", Int32Regs>;
360 defm CVT_s64 : CVT_FROM_ALL<"s64", Int64Regs>;
361 defm CVT_u64 : CVT_FROM_ALL<"u64", Int64Regs>;
362 defm CVT_f32 : CVT_FROM_ALL<"f32", Float32Regs>;
363 defm CVT_f64 : CVT_FROM_ALL<"f64", Float64Regs>;
365 // This set of cvt is different from the above. The type of the source
366 // and target are the same.
368 def CVT_INREG_s16_s8 : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
369 "cvt.s16.s8 \t$dst, $src;", []>;
370 def CVT_INREG_s32_s8 : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
371 "cvt.s32.s8 \t$dst, $src;", []>;
372 def CVT_INREG_s32_s16 : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
373 "cvt.s32.s16 \t$dst, $src;", []>;
374 def CVT_INREG_s64_s8 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
375 "cvt.s64.s8 \t$dst, $src;", []>;
376 def CVT_INREG_s64_s16 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
377 "cvt.s64.s16 \t$dst, $src;", []>;
378 def CVT_INREG_s64_s32 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
379 "cvt.s64.s32 \t$dst, $src;", []>;
382 //-----------------------------------
383 // Integer Arithmetic
384 //-----------------------------------
386 multiclass ADD_SUB_i1<SDNode OpNode> {
387 def _rr: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$a, Int1Regs:$b),
388 "xor.pred \t$dst, $a, $b;",
389 [(set Int1Regs:$dst, (OpNode Int1Regs:$a, Int1Regs:$b))]>;
390 def _ri: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$a, i1imm:$b),
391 "xor.pred \t$dst, $a, $b;",
392 [(set Int1Regs:$dst, (OpNode Int1Regs:$a, (imm):$b))]>;
395 defm ADD_i1 : ADD_SUB_i1<add>;
396 defm SUB_i1 : ADD_SUB_i1<sub>;
399 defm ADD : I3<"add.s", add>;
400 defm SUB : I3<"sub.s", sub>;
402 defm ADDCC : ADD_SUB_INT_32<"add.cc", addc>;
403 defm SUBCC : ADD_SUB_INT_32<"sub.cc", subc>;
405 defm ADDCCC : ADD_SUB_INT_32<"addc.cc", adde>;
406 defm SUBCCC : ADD_SUB_INT_32<"subc.cc", sube>;
408 //mul.wide PTX instruction
409 def SInt32Const : PatLeaf<(imm), [{
410 const APInt &v = N->getAPIntValue();
411 if (v.isSignedIntN(32))
416 def UInt32Const : PatLeaf<(imm), [{
417 const APInt &v = N->getAPIntValue();
423 def SInt16Const : PatLeaf<(imm), [{
424 const APInt &v = N->getAPIntValue();
425 if (v.isSignedIntN(16))
430 def UInt16Const : PatLeaf<(imm), [{
431 const APInt &v = N->getAPIntValue();
437 def Int5Const : PatLeaf<(imm), [{
438 const APInt &v = N->getAPIntValue();
439 // Check if 0 <= v < 32
440 // Only then the result from (x << v) will be i32
441 if (v.sge(0) && v.slt(32))
446 def Int4Const : PatLeaf<(imm), [{
447 const APInt &v = N->getAPIntValue();
448 // Check if 0 <= v < 16
449 // Only then the result from (x << v) will be i16
450 if (v.sge(0) && v.slt(16))
455 def SHL2MUL32 : SDNodeXForm<imm, [{
456 const APInt &v = N->getAPIntValue();
458 return CurDAG->getTargetConstant(temp.shl(v), MVT::i32);
461 def SHL2MUL16 : SDNodeXForm<imm, [{
462 const APInt &v = N->getAPIntValue();
464 return CurDAG->getTargetConstant(temp.shl(v), MVT::i16);
467 def MULWIDES64 : NVPTXInst<(outs Int64Regs:$dst),
468 (ins Int32Regs:$a, Int32Regs:$b),
469 "mul.wide.s32 \t$dst, $a, $b;", []>;
470 def MULWIDES64Imm : NVPTXInst<(outs Int64Regs:$dst),
471 (ins Int32Regs:$a, i64imm:$b),
472 "mul.wide.s32 \t$dst, $a, $b;", []>;
474 def MULWIDEU64 : NVPTXInst<(outs Int64Regs:$dst),
475 (ins Int32Regs:$a, Int32Regs:$b),
476 "mul.wide.u32 \t$dst, $a, $b;", []>;
477 def MULWIDEU64Imm : NVPTXInst<(outs Int64Regs:$dst),
478 (ins Int32Regs:$a, i64imm:$b),
479 "mul.wide.u32 \t$dst, $a, $b;", []>;
481 def MULWIDES32 : NVPTXInst<(outs Int32Regs:$dst),
482 (ins Int16Regs:$a, Int16Regs:$b),
483 "mul.wide.s16 \t$dst, $a, $b;", []>;
484 def MULWIDES32Imm : NVPTXInst<(outs Int32Regs:$dst),
485 (ins Int16Regs:$a, i32imm:$b),
486 "mul.wide.s16 \t$dst, $a, $b;", []>;
488 def MULWIDEU32 : NVPTXInst<(outs Int32Regs:$dst),
489 (ins Int16Regs:$a, Int16Regs:$b),
490 "mul.wide.u16 \t$dst, $a, $b;", []>;
491 def MULWIDEU32Imm : NVPTXInst<(outs Int32Regs:$dst),
492 (ins Int16Regs:$a, i32imm:$b),
493 "mul.wide.u16 \t$dst, $a, $b;", []>;
495 def : Pat<(shl (sext Int32Regs:$a), (i32 Int5Const:$b)),
496 (MULWIDES64Imm Int32Regs:$a, (SHL2MUL32 node:$b))>,
497 Requires<[doMulWide]>;
498 def : Pat<(shl (zext Int32Regs:$a), (i32 Int5Const:$b)),
499 (MULWIDEU64Imm Int32Regs:$a, (SHL2MUL32 node:$b))>,
500 Requires<[doMulWide]>;
502 def : Pat<(shl (sext Int16Regs:$a), (i16 Int4Const:$b)),
503 (MULWIDES32Imm Int16Regs:$a, (SHL2MUL16 node:$b))>,
504 Requires<[doMulWide]>;
505 def : Pat<(shl (zext Int16Regs:$a), (i16 Int4Const:$b)),
506 (MULWIDEU32Imm Int16Regs:$a, (SHL2MUL16 node:$b))>,
507 Requires<[doMulWide]>;
509 def : Pat<(mul (sext Int32Regs:$a), (sext Int32Regs:$b)),
510 (MULWIDES64 Int32Regs:$a, Int32Regs:$b)>,
511 Requires<[doMulWide]>;
512 def : Pat<(mul (sext Int32Regs:$a), (i64 SInt32Const:$b)),
513 (MULWIDES64Imm Int32Regs:$a, (i64 SInt32Const:$b))>,
514 Requires<[doMulWide]>;
516 def : Pat<(mul (zext Int32Regs:$a), (zext Int32Regs:$b)),
517 (MULWIDEU64 Int32Regs:$a, Int32Regs:$b)>, Requires<[doMulWide]>;
518 def : Pat<(mul (zext Int32Regs:$a), (i64 UInt32Const:$b)),
519 (MULWIDEU64Imm Int32Regs:$a, (i64 UInt32Const:$b))>,
520 Requires<[doMulWide]>;
522 def : Pat<(mul (sext Int16Regs:$a), (sext Int16Regs:$b)),
523 (MULWIDES32 Int16Regs:$a, Int16Regs:$b)>, Requires<[doMulWide]>;
524 def : Pat<(mul (sext Int16Regs:$a), (i32 SInt16Const:$b)),
525 (MULWIDES32Imm Int16Regs:$a, (i32 SInt16Const:$b))>,
526 Requires<[doMulWide]>;
528 def : Pat<(mul (zext Int16Regs:$a), (zext Int16Regs:$b)),
529 (MULWIDEU32 Int16Regs:$a, Int16Regs:$b)>, Requires<[doMulWide]>;
530 def : Pat<(mul (zext Int16Regs:$a), (i32 UInt16Const:$b)),
531 (MULWIDEU32Imm Int16Regs:$a, (i32 UInt16Const:$b))>,
532 Requires<[doMulWide]>;
534 defm MULT : I3<"mul.lo.s", mul>;
536 defm MULTHS : I3<"mul.hi.s", mulhs>;
537 defm MULTHU : I3<"mul.hi.u", mulhu>;
539 defm SDIV : I3<"div.s", sdiv>;
540 defm UDIV : I3<"div.u", udiv>;
542 defm SREM : I3<"rem.s", srem>;
543 // The ri version will not be selected as DAGCombiner::visitSREM will lower it.
544 defm UREM : I3<"rem.u", urem>;
545 // The ri version will not be selected as DAGCombiner::visitUREM will lower it.
547 def MAD16rrr : NVPTXInst<(outs Int16Regs:$dst),
548 (ins Int16Regs:$a, Int16Regs:$b, Int16Regs:$c),
549 "mad.lo.s16 \t$dst, $a, $b, $c;",
550 [(set Int16Regs:$dst, (add
551 (mul Int16Regs:$a, Int16Regs:$b), Int16Regs:$c))]>;
552 def MAD16rri : NVPTXInst<(outs Int16Regs:$dst),
553 (ins Int16Regs:$a, Int16Regs:$b, i16imm:$c),
554 "mad.lo.s16 \t$dst, $a, $b, $c;",
555 [(set Int16Regs:$dst, (add
556 (mul Int16Regs:$a, Int16Regs:$b), imm:$c))]>;
557 def MAD16rir : NVPTXInst<(outs Int16Regs:$dst),
558 (ins Int16Regs:$a, i16imm:$b, Int16Regs:$c),
559 "mad.lo.s16 \t$dst, $a, $b, $c;",
560 [(set Int16Regs:$dst, (add
561 (mul Int16Regs:$a, imm:$b), Int16Regs:$c))]>;
562 def MAD16rii : NVPTXInst<(outs Int16Regs:$dst),
563 (ins Int16Regs:$a, i16imm:$b, i16imm:$c),
564 "mad.lo.s16 \t$dst, $a, $b, $c;",
565 [(set Int16Regs:$dst, (add (mul Int16Regs:$a, imm:$b),
568 def MAD32rrr : NVPTXInst<(outs Int32Regs:$dst),
569 (ins Int32Regs:$a, Int32Regs:$b, Int32Regs:$c),
570 "mad.lo.s32 \t$dst, $a, $b, $c;",
571 [(set Int32Regs:$dst, (add
572 (mul Int32Regs:$a, Int32Regs:$b), Int32Regs:$c))]>;
573 def MAD32rri : NVPTXInst<(outs Int32Regs:$dst),
574 (ins Int32Regs:$a, Int32Regs:$b, i32imm:$c),
575 "mad.lo.s32 \t$dst, $a, $b, $c;",
576 [(set Int32Regs:$dst, (add
577 (mul Int32Regs:$a, Int32Regs:$b), imm:$c))]>;
578 def MAD32rir : NVPTXInst<(outs Int32Regs:$dst),
579 (ins Int32Regs:$a, i32imm:$b, Int32Regs:$c),
580 "mad.lo.s32 \t$dst, $a, $b, $c;",
581 [(set Int32Regs:$dst, (add
582 (mul Int32Regs:$a, imm:$b), Int32Regs:$c))]>;
583 def MAD32rii : NVPTXInst<(outs Int32Regs:$dst),
584 (ins Int32Regs:$a, i32imm:$b, i32imm:$c),
585 "mad.lo.s32 \t$dst, $a, $b, $c;",
586 [(set Int32Regs:$dst, (add
587 (mul Int32Regs:$a, imm:$b), imm:$c))]>;
589 def MAD64rrr : NVPTXInst<(outs Int64Regs:$dst),
590 (ins Int64Regs:$a, Int64Regs:$b, Int64Regs:$c),
591 "mad.lo.s64 \t$dst, $a, $b, $c;",
592 [(set Int64Regs:$dst, (add
593 (mul Int64Regs:$a, Int64Regs:$b), Int64Regs:$c))]>;
594 def MAD64rri : NVPTXInst<(outs Int64Regs:$dst),
595 (ins Int64Regs:$a, Int64Regs:$b, i64imm:$c),
596 "mad.lo.s64 \t$dst, $a, $b, $c;",
597 [(set Int64Regs:$dst, (add
598 (mul Int64Regs:$a, Int64Regs:$b), imm:$c))]>;
599 def MAD64rir : NVPTXInst<(outs Int64Regs:$dst),
600 (ins Int64Regs:$a, i64imm:$b, Int64Regs:$c),
601 "mad.lo.s64 \t$dst, $a, $b, $c;",
602 [(set Int64Regs:$dst, (add
603 (mul Int64Regs:$a, imm:$b), Int64Regs:$c))]>;
604 def MAD64rii : NVPTXInst<(outs Int64Regs:$dst),
605 (ins Int64Regs:$a, i64imm:$b, i64imm:$c),
606 "mad.lo.s64 \t$dst, $a, $b, $c;",
607 [(set Int64Regs:$dst, (add
608 (mul Int64Regs:$a, imm:$b), imm:$c))]>;
611 def INEG16 : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
612 "neg.s16 \t$dst, $src;",
613 [(set Int16Regs:$dst, (ineg Int16Regs:$src))]>;
614 def INEG32 : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
615 "neg.s32 \t$dst, $src;",
616 [(set Int32Regs:$dst, (ineg Int32Regs:$src))]>;
617 def INEG64 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
618 "neg.s64 \t$dst, $src;",
619 [(set Int64Regs:$dst, (ineg Int64Regs:$src))]>;
621 //-----------------------------------
622 // Floating Point Arithmetic
623 //-----------------------------------
626 def FloatConst1 : PatLeaf<(fpimm), [{
627 if (&(N->getValueAPF().getSemantics()) != &llvm::APFloat::IEEEsingle)
629 float f = (float)N->getValueAPF().convertToFloat();
632 // Constand (double)1.0
633 def DoubleConst1 : PatLeaf<(fpimm), [{
634 if (&(N->getValueAPF().getSemantics()) != &llvm::APFloat::IEEEdouble)
636 double d = (double)N->getValueAPF().convertToDouble();
640 defm FADD : F3<"add", fadd>;
641 defm FSUB : F3<"sub", fsub>;
642 defm FMUL : F3<"mul", fmul>;
644 defm FADD_rn : F3_rn<"add", fadd>;
645 defm FSUB_rn : F3_rn<"sub", fsub>;
646 defm FMUL_rn : F3_rn<"mul", fmul>;
648 defm FABS : F2<"abs", fabs>;
649 defm FNEG : F2<"neg", fneg>;
650 defm FSQRT : F2<"sqrt.rn", fsqrt>;
655 def FDIV641r : NVPTXInst<(outs Float64Regs:$dst),
656 (ins f64imm:$a, Float64Regs:$b),
657 "rcp.rn.f64 \t$dst, $b;",
658 [(set Float64Regs:$dst,
659 (fdiv DoubleConst1:$a, Float64Regs:$b))]>;
660 def FDIV64rr : NVPTXInst<(outs Float64Regs:$dst),
661 (ins Float64Regs:$a, Float64Regs:$b),
662 "div.rn.f64 \t$dst, $a, $b;",
663 [(set Float64Regs:$dst,
664 (fdiv Float64Regs:$a, Float64Regs:$b))]>;
665 def FDIV64ri : NVPTXInst<(outs Float64Regs:$dst),
666 (ins Float64Regs:$a, f64imm:$b),
667 "div.rn.f64 \t$dst, $a, $b;",
668 [(set Float64Regs:$dst,
669 (fdiv Float64Regs:$a, fpimm:$b))]>;
672 // F32 Approximate reciprocal
674 def FDIV321r_ftz : NVPTXInst<(outs Float32Regs:$dst),
675 (ins f32imm:$a, Float32Regs:$b),
676 "rcp.approx.ftz.f32 \t$dst, $b;",
677 [(set Float32Regs:$dst,
678 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
679 Requires<[do_DIVF32_APPROX, doF32FTZ]>;
680 def FDIV321r : NVPTXInst<(outs Float32Regs:$dst),
681 (ins f32imm:$a, Float32Regs:$b),
682 "rcp.approx.f32 \t$dst, $b;",
683 [(set Float32Regs:$dst,
684 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
685 Requires<[do_DIVF32_APPROX]>;
687 // F32 Approximate division
689 def FDIV32approxrr_ftz : NVPTXInst<(outs Float32Regs:$dst),
690 (ins Float32Regs:$a, Float32Regs:$b),
691 "div.approx.ftz.f32 \t$dst, $a, $b;",
692 [(set Float32Regs:$dst,
693 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
694 Requires<[do_DIVF32_APPROX, doF32FTZ]>;
695 def FDIV32approxri_ftz : NVPTXInst<(outs Float32Regs:$dst),
696 (ins Float32Regs:$a, f32imm:$b),
697 "div.approx.ftz.f32 \t$dst, $a, $b;",
698 [(set Float32Regs:$dst,
699 (fdiv Float32Regs:$a, fpimm:$b))]>,
700 Requires<[do_DIVF32_APPROX, doF32FTZ]>;
701 def FDIV32approxrr : NVPTXInst<(outs Float32Regs:$dst),
702 (ins Float32Regs:$a, Float32Regs:$b),
703 "div.approx.f32 \t$dst, $a, $b;",
704 [(set Float32Regs:$dst,
705 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
706 Requires<[do_DIVF32_APPROX]>;
707 def FDIV32approxri : NVPTXInst<(outs Float32Regs:$dst),
708 (ins Float32Regs:$a, f32imm:$b),
709 "div.approx.f32 \t$dst, $a, $b;",
710 [(set Float32Regs:$dst,
711 (fdiv Float32Regs:$a, fpimm:$b))]>,
712 Requires<[do_DIVF32_APPROX]>;
714 // F32 Semi-accurate reciprocal
716 // rcp.approx gives the same result as div.full(1.0f, a) and is faster.
718 def FDIV321r_approx_ftz : NVPTXInst<(outs Float32Regs:$dst),
719 (ins f32imm:$a, Float32Regs:$b),
720 "rcp.approx.ftz.f32 \t$dst, $b;",
721 [(set Float32Regs:$dst,
722 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
723 Requires<[do_DIVF32_FULL, doF32FTZ]>;
724 def FDIV321r_approx : NVPTXInst<(outs Float32Regs:$dst),
725 (ins f32imm:$a, Float32Regs:$b),
726 "rcp.approx.f32 \t$dst, $b;",
727 [(set Float32Regs:$dst,
728 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
729 Requires<[do_DIVF32_FULL]>;
731 // F32 Semi-accurate division
733 def FDIV32rr_ftz : NVPTXInst<(outs Float32Regs:$dst),
734 (ins Float32Regs:$a, Float32Regs:$b),
735 "div.full.ftz.f32 \t$dst, $a, $b;",
736 [(set Float32Regs:$dst,
737 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
738 Requires<[do_DIVF32_FULL, doF32FTZ]>;
739 def FDIV32ri_ftz : NVPTXInst<(outs Float32Regs:$dst),
740 (ins Float32Regs:$a, f32imm:$b),
741 "div.full.ftz.f32 \t$dst, $a, $b;",
742 [(set Float32Regs:$dst,
743 (fdiv Float32Regs:$a, fpimm:$b))]>,
744 Requires<[do_DIVF32_FULL, doF32FTZ]>;
745 def FDIV32rr : NVPTXInst<(outs Float32Regs:$dst),
746 (ins Float32Regs:$a, Float32Regs:$b),
747 "div.full.f32 \t$dst, $a, $b;",
748 [(set Float32Regs:$dst,
749 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
750 Requires<[do_DIVF32_FULL]>;
751 def FDIV32ri : NVPTXInst<(outs Float32Regs:$dst),
752 (ins Float32Regs:$a, f32imm:$b),
753 "div.full.f32 \t$dst, $a, $b;",
754 [(set Float32Regs:$dst,
755 (fdiv Float32Regs:$a, fpimm:$b))]>,
756 Requires<[do_DIVF32_FULL]>;
758 // F32 Accurate reciprocal
760 def FDIV321r_prec_ftz : NVPTXInst<(outs Float32Regs:$dst),
761 (ins f32imm:$a, Float32Regs:$b),
762 "rcp.rn.ftz.f32 \t$dst, $b;",
763 [(set Float32Regs:$dst,
764 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
765 Requires<[reqPTX20, doF32FTZ]>;
766 def FDIV321r_prec : NVPTXInst<(outs Float32Regs:$dst),
767 (ins f32imm:$a, Float32Regs:$b),
768 "rcp.rn.f32 \t$dst, $b;",
769 [(set Float32Regs:$dst,
770 (fdiv FloatConst1:$a, Float32Regs:$b))]>,
771 Requires<[reqPTX20]>;
773 // F32 Accurate division
775 def FDIV32rr_prec_ftz : NVPTXInst<(outs Float32Regs:$dst),
776 (ins Float32Regs:$a, Float32Regs:$b),
777 "div.rn.ftz.f32 \t$dst, $a, $b;",
778 [(set Float32Regs:$dst,
779 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
780 Requires<[doF32FTZ, reqPTX20]>;
781 def FDIV32ri_prec_ftz : NVPTXInst<(outs Float32Regs:$dst),
782 (ins Float32Regs:$a, f32imm:$b),
783 "div.rn.ftz.f32 \t$dst, $a, $b;",
784 [(set Float32Regs:$dst,
785 (fdiv Float32Regs:$a, fpimm:$b))]>,
786 Requires<[doF32FTZ, reqPTX20]>;
787 def FDIV32rr_prec : NVPTXInst<(outs Float32Regs:$dst),
788 (ins Float32Regs:$a, Float32Regs:$b),
789 "div.rn.f32 \t$dst, $a, $b;",
790 [(set Float32Regs:$dst,
791 (fdiv Float32Regs:$a, Float32Regs:$b))]>,
792 Requires<[reqPTX20]>;
793 def FDIV32ri_prec : NVPTXInst<(outs Float32Regs:$dst),
794 (ins Float32Regs:$a, f32imm:$b),
795 "div.rn.f32 \t$dst, $a, $b;",
796 [(set Float32Regs:$dst,
797 (fdiv Float32Regs:$a, fpimm:$b))]>,
798 Requires<[reqPTX20]>;
804 def RSQRTF32approx1r : NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$b),
805 "rsqrt.approx.f32 \t$dst, $b;", []>;
807 def: Pat<(fdiv FloatConst1, (int_nvvm_sqrt_f Float32Regs:$b)),
808 (RSQRTF32approx1r Float32Regs:$b)>,
809 Requires<[do_DIVF32_FULL, do_SQRTF32_APPROX, doNoF32FTZ]>;
811 multiclass FPCONTRACT32<string OpcStr, Predicate Pred> {
812 def rrr : NVPTXInst<(outs Float32Regs:$dst),
813 (ins Float32Regs:$a, Float32Regs:$b, Float32Regs:$c),
814 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
815 [(set Float32Regs:$dst, (fadd
816 (fmul Float32Regs:$a, Float32Regs:$b),
817 Float32Regs:$c))]>, Requires<[Pred]>;
818 // This is to WAR a weird bug in Tablegen that does not automatically
819 // generate the following permutated rule rrr2 from the above rrr.
820 // So we explicitly add it here. This happens to FMA32 only.
821 // See the comments at FMAD32 and FMA32 for more information.
822 def rrr2 : NVPTXInst<(outs Float32Regs:$dst),
823 (ins Float32Regs:$a, Float32Regs:$b, Float32Regs:$c),
824 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
825 [(set Float32Regs:$dst, (fadd Float32Regs:$c,
826 (fmul Float32Regs:$a, Float32Regs:$b)))]>,
828 def rri : NVPTXInst<(outs Float32Regs:$dst),
829 (ins Float32Regs:$a, Float32Regs:$b, f32imm:$c),
830 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
831 [(set Float32Regs:$dst, (fadd
832 (fmul Float32Regs:$a, Float32Regs:$b), fpimm:$c))]>,
834 def rir : NVPTXInst<(outs Float32Regs:$dst),
835 (ins Float32Regs:$a, f32imm:$b, Float32Regs:$c),
836 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
837 [(set Float32Regs:$dst, (fadd
838 (fmul Float32Regs:$a, fpimm:$b), Float32Regs:$c))]>,
840 def rii : NVPTXInst<(outs Float32Regs:$dst),
841 (ins Float32Regs:$a, f32imm:$b, f32imm:$c),
842 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
843 [(set Float32Regs:$dst, (fadd
844 (fmul Float32Regs:$a, fpimm:$b), fpimm:$c))]>,
848 multiclass FPCONTRACT64<string OpcStr, Predicate Pred> {
849 def rrr : NVPTXInst<(outs Float64Regs:$dst),
850 (ins Float64Regs:$a, Float64Regs:$b, Float64Regs:$c),
851 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
852 [(set Float64Regs:$dst, (fadd
853 (fmul Float64Regs:$a, Float64Regs:$b),
854 Float64Regs:$c))]>, Requires<[Pred]>;
855 def rri : NVPTXInst<(outs Float64Regs:$dst),
856 (ins Float64Regs:$a, Float64Regs:$b, f64imm:$c),
857 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
858 [(set Float64Regs:$dst, (fadd (fmul Float64Regs:$a,
859 Float64Regs:$b), fpimm:$c))]>, Requires<[Pred]>;
860 def rir : NVPTXInst<(outs Float64Regs:$dst),
861 (ins Float64Regs:$a, f64imm:$b, Float64Regs:$c),
862 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
863 [(set Float64Regs:$dst, (fadd
864 (fmul Float64Regs:$a, fpimm:$b), Float64Regs:$c))]>,
866 def rii : NVPTXInst<(outs Float64Regs:$dst),
867 (ins Float64Regs:$a, f64imm:$b, f64imm:$c),
868 !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
869 [(set Float64Regs:$dst, (fadd
870 (fmul Float64Regs:$a, fpimm:$b), fpimm:$c))]>,
874 // Due to a unknown reason (most likely a bug in tablegen), tablegen does not
875 // automatically generate the rrr2 rule from
876 // the rrr rule (see FPCONTRACT32) for FMA32, though it does for FMAD32.
877 // If we reverse the order of the following two lines, then rrr2 rule will be
878 // generated for FMA32, but not for rrr.
879 // Therefore, we manually write the rrr2 rule in FPCONTRACT32.
880 defm FMA32_ftz : FPCONTRACT32<"fma.rn.ftz.f32", doFMAF32_ftz>;
881 defm FMA32 : FPCONTRACT32<"fma.rn.f32", doFMAF32>;
882 defm FMA64 : FPCONTRACT64<"fma.rn.f64", doFMAF64>;
884 // b*c-a => fmad(b, c, -a)
885 multiclass FPCONTRACT32_SUB_PAT_MAD<NVPTXInst Inst, Predicate Pred> {
886 def : Pat<(fsub (fmul Float32Regs:$b, Float32Regs:$c), Float32Regs:$a),
887 (Inst Float32Regs:$b, Float32Regs:$c, (FNEGf32 Float32Regs:$a))>,
891 // a-b*c => fmad(-b,c, a)
892 // - legal because a-b*c <=> a+(-b*c) <=> a+(-b)*c
893 // b*c-a => fmad(b, c, -a)
894 // - legal because b*c-a <=> b*c+(-a)
895 multiclass FPCONTRACT32_SUB_PAT<NVPTXInst Inst, Predicate Pred> {
896 def : Pat<(fsub Float32Regs:$a, (fmul Float32Regs:$b, Float32Regs:$c)),
897 (Inst (FNEGf32 Float32Regs:$b), Float32Regs:$c, Float32Regs:$a)>,
899 def : Pat<(fsub (fmul Float32Regs:$b, Float32Regs:$c), Float32Regs:$a),
900 (Inst Float32Regs:$b, Float32Regs:$c, (FNEGf32 Float32Regs:$a))>,
904 // a-b*c => fmad(-b,c, a)
905 // b*c-a => fmad(b, c, -a)
906 multiclass FPCONTRACT64_SUB_PAT<NVPTXInst Inst, Predicate Pred> {
907 def : Pat<(fsub Float64Regs:$a, (fmul Float64Regs:$b, Float64Regs:$c)),
908 (Inst (FNEGf64 Float64Regs:$b), Float64Regs:$c, Float64Regs:$a)>,
911 def : Pat<(fsub (fmul Float64Regs:$b, Float64Regs:$c), Float64Regs:$a),
912 (Inst Float64Regs:$b, Float64Regs:$c, (FNEGf64 Float64Regs:$a))>,
916 defm FMAF32ext_ftz : FPCONTRACT32_SUB_PAT<FMA32_ftzrrr, doFMAF32AGG_ftz>;
917 defm FMAF32ext : FPCONTRACT32_SUB_PAT<FMA32rrr, doFMAF32AGG>;
918 defm FMAF64ext : FPCONTRACT64_SUB_PAT<FMA64rrr, doFMAF64AGG>;
920 def SINF: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$src),
921 "sin.approx.f32 \t$dst, $src;",
922 [(set Float32Regs:$dst, (fsin Float32Regs:$src))]>;
923 def COSF: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$src),
924 "cos.approx.f32 \t$dst, $src;",
925 [(set Float32Regs:$dst, (fcos Float32Regs:$src))]>;
927 // Lower (frem x, y) into (sub x, (mul (floor (div x, y)) y))
928 // e.g. "poor man's fmod()"
931 def : Pat<(frem Float32Regs:$x, Float32Regs:$y),
932 (FSUBf32rr_ftz Float32Regs:$x, (FMULf32rr_ftz (CVT_f32_f32
933 (FDIV32rr_prec_ftz Float32Regs:$x, Float32Regs:$y), CvtRMI_FTZ),
935 Requires<[doF32FTZ]>;
936 def : Pat<(frem Float32Regs:$x, fpimm:$y),
937 (FSUBf32rr_ftz Float32Regs:$x, (FMULf32ri_ftz (CVT_f32_f32
938 (FDIV32ri_prec_ftz Float32Regs:$x, fpimm:$y), CvtRMI_FTZ),
940 Requires<[doF32FTZ]>;
943 def : Pat<(frem Float32Regs:$x, Float32Regs:$y),
944 (FSUBf32rr Float32Regs:$x, (FMULf32rr (CVT_f32_f32
945 (FDIV32rr_prec Float32Regs:$x, Float32Regs:$y), CvtRMI),
947 def : Pat<(frem Float32Regs:$x, fpimm:$y),
948 (FSUBf32rr Float32Regs:$x, (FMULf32ri (CVT_f32_f32
949 (FDIV32ri_prec Float32Regs:$x, fpimm:$y), CvtRMI),
953 def : Pat<(frem Float64Regs:$x, Float64Regs:$y),
954 (FSUBf64rr Float64Regs:$x, (FMULf64rr (CVT_f64_f64
955 (FDIV64rr Float64Regs:$x, Float64Regs:$y), CvtRMI),
957 def : Pat<(frem Float64Regs:$x, fpimm:$y),
958 (FSUBf64rr Float64Regs:$x, (FMULf64ri (CVT_f64_f64
959 (FDIV64ri Float64Regs:$x, fpimm:$y), CvtRMI),
962 //-----------------------------------
963 // Logical Arithmetic
964 //-----------------------------------
966 multiclass LOG_FORMAT<string OpcStr, SDNode OpNode> {
967 def b1rr: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$a, Int1Regs:$b),
968 !strconcat(OpcStr, ".pred \t$dst, $a, $b;"),
969 [(set Int1Regs:$dst, (OpNode Int1Regs:$a, Int1Regs:$b))]>;
970 def b1ri: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$a, i1imm:$b),
971 !strconcat(OpcStr, ".pred \t$dst, $a, $b;"),
972 [(set Int1Regs:$dst, (OpNode Int1Regs:$a, imm:$b))]>;
973 def b16rr: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
974 !strconcat(OpcStr, ".b16 \t$dst, $a, $b;"),
975 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
977 def b16ri: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
978 !strconcat(OpcStr, ".b16 \t$dst, $a, $b;"),
979 [(set Int16Regs:$dst, (OpNode Int16Regs:$a, imm:$b))]>;
980 def b32rr: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, Int32Regs:$b),
981 !strconcat(OpcStr, ".b32 \t$dst, $a, $b;"),
982 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
984 def b32ri: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
985 !strconcat(OpcStr, ".b32 \t$dst, $a, $b;"),
986 [(set Int32Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
987 def b64rr: NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, Int64Regs:$b),
988 !strconcat(OpcStr, ".b64 \t$dst, $a, $b;"),
989 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
991 def b64ri: NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i64imm:$b),
992 !strconcat(OpcStr, ".b64 \t$dst, $a, $b;"),
993 [(set Int64Regs:$dst, (OpNode Int64Regs:$a, imm:$b))]>;
996 defm OR : LOG_FORMAT<"or", or>;
997 defm AND : LOG_FORMAT<"and", and>;
998 defm XOR : LOG_FORMAT<"xor", xor>;
1000 def NOT1: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$src),
1001 "not.pred \t$dst, $src;",
1002 [(set Int1Regs:$dst, (not Int1Regs:$src))]>;
1003 def NOT16: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
1004 "not.b16 \t$dst, $src;",
1005 [(set Int16Regs:$dst, (not Int16Regs:$src))]>;
1006 def NOT32: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
1007 "not.b32 \t$dst, $src;",
1008 [(set Int32Regs:$dst, (not Int32Regs:$src))]>;
1009 def NOT64: NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
1010 "not.b64 \t$dst, $src;",
1011 [(set Int64Regs:$dst, (not Int64Regs:$src))]>;
1013 // For shifts, the second src operand must be 32-bit value
1014 multiclass LSHIFT_FORMAT<string OpcStr, SDNode OpNode> {
1015 def i64rr : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a,
1017 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
1018 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
1020 def i64ri : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i32imm:$b),
1021 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
1022 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
1024 def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a,
1026 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1027 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
1029 def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
1030 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1031 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
1033 def i32ii : NVPTXInst<(outs Int32Regs:$dst), (ins i32imm:$a, i32imm:$b),
1034 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1035 [(set Int32Regs:$dst, (OpNode (i32 imm:$a),
1037 def i16rr : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a,
1039 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
1040 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
1042 def i16ri : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i32imm:$b),
1043 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
1044 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
1048 defm SHL : LSHIFT_FORMAT<"shl.b", shl>;
1050 // For shifts, the second src operand must be 32-bit value
1051 // Need to add cvt for the 8-bits.
1052 multiclass RSHIFT_FORMAT<string OpcStr, SDNode OpNode> {
1053 def i64rr : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a,
1055 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
1056 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
1058 def i64ri : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i32imm:$b),
1059 !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
1060 [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
1062 def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a,
1064 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1065 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
1067 def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
1068 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1069 [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
1071 def i32ii : NVPTXInst<(outs Int32Regs:$dst), (ins i32imm:$a, i32imm:$b),
1072 !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
1073 [(set Int32Regs:$dst, (OpNode (i32 imm:$a),
1075 def i16rr : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a,
1077 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
1078 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
1080 def i16ri : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i32imm:$b),
1081 !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
1082 [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
1086 defm SRA : RSHIFT_FORMAT<"shr.s", sra>;
1087 defm SRL : RSHIFT_FORMAT<"shr.u", srl>;
1090 // Rotate: use ptx shf instruction if available.
1093 // 32 bit r2 = rotl r1, n
1095 // r2 = shf.l r1, r1, n
1096 def ROTL32imm_hw : NVPTXInst<(outs Int32Regs:$dst),
1097 (ins Int32Regs:$src, i32imm:$amt),
1098 "shf.l.wrap.b32 \t$dst, $src, $src, $amt;",
1099 [(set Int32Regs:$dst, (rotl Int32Regs:$src, (i32 imm:$amt)))]>,
1100 Requires<[hasHWROT32]> ;
1102 def ROTL32reg_hw : NVPTXInst<(outs Int32Regs:$dst),
1103 (ins Int32Regs:$src, Int32Regs:$amt),
1104 "shf.l.wrap.b32 \t$dst, $src, $src, $amt;",
1105 [(set Int32Regs:$dst, (rotl Int32Regs:$src, Int32Regs:$amt))]>,
1106 Requires<[hasHWROT32]>;
1108 // 32 bit r2 = rotr r1, n
1110 // r2 = shf.r r1, r1, n
1111 def ROTR32imm_hw : NVPTXInst<(outs Int32Regs:$dst),
1112 (ins Int32Regs:$src, i32imm:$amt),
1113 "shf.r.wrap.b32 \t$dst, $src, $src, $amt;",
1114 [(set Int32Regs:$dst, (rotr Int32Regs:$src, (i32 imm:$amt)))]>,
1115 Requires<[hasHWROT32]>;
1117 def ROTR32reg_hw : NVPTXInst<(outs Int32Regs:$dst),
1118 (ins Int32Regs:$src, Int32Regs:$amt),
1119 "shf.r.wrap.b32 \t$dst, $src, $src, $amt;",
1120 [(set Int32Regs:$dst, (rotr Int32Regs:$src, Int32Regs:$amt))]>,
1121 Requires<[hasHWROT32]>;
1124 // Rotate: if ptx shf instruction is not available, then use shift+add
1127 def ROT32imm_sw : NVPTXInst<(outs Int32Regs:$dst),
1128 (ins Int32Regs:$src, i32imm:$amt1, i32imm:$amt2),
1129 !strconcat("{{\n\t",
1130 !strconcat(".reg .b32 %lhs;\n\t",
1131 !strconcat(".reg .b32 %rhs;\n\t",
1132 !strconcat("shl.b32 \t%lhs, $src, $amt1;\n\t",
1133 !strconcat("shr.b32 \t%rhs, $src, $amt2;\n\t",
1134 !strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t",
1135 !strconcat("}}", ""))))))),
1138 def SUB_FRM_32 : SDNodeXForm<imm, [{
1139 return CurDAG->getTargetConstant(32-N->getZExtValue(), MVT::i32);
1142 def : Pat<(rotl Int32Regs:$src, (i32 imm:$amt)),
1143 (ROT32imm_sw Int32Regs:$src, imm:$amt, (SUB_FRM_32 node:$amt))>,
1144 Requires<[noHWROT32]>;
1145 def : Pat<(rotr Int32Regs:$src, (i32 imm:$amt)),
1146 (ROT32imm_sw Int32Regs:$src, (SUB_FRM_32 node:$amt), imm:$amt)>,
1147 Requires<[noHWROT32]>;
1149 def ROTL32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src,
1151 !strconcat("{{\n\t",
1152 !strconcat(".reg .b32 %lhs;\n\t",
1153 !strconcat(".reg .b32 %rhs;\n\t",
1154 !strconcat(".reg .b32 %amt2;\n\t",
1155 !strconcat("shl.b32 \t%lhs, $src, $amt;\n\t",
1156 !strconcat("sub.s32 \t%amt2, 32, $amt;\n\t",
1157 !strconcat("shr.b32 \t%rhs, $src, %amt2;\n\t",
1158 !strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t",
1159 !strconcat("}}", ""))))))))),
1160 [(set Int32Regs:$dst, (rotl Int32Regs:$src, Int32Regs:$amt))]>,
1161 Requires<[noHWROT32]>;
1163 def ROTR32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src,
1165 !strconcat("{{\n\t",
1166 !strconcat(".reg .b32 %lhs;\n\t",
1167 !strconcat(".reg .b32 %rhs;\n\t",
1168 !strconcat(".reg .b32 %amt2;\n\t",
1169 !strconcat("shr.b32 \t%lhs, $src, $amt;\n\t",
1170 !strconcat("sub.s32 \t%amt2, 32, $amt;\n\t",
1171 !strconcat("shl.b32 \t%rhs, $src, %amt2;\n\t",
1172 !strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t",
1173 !strconcat("}}", ""))))))))),
1174 [(set Int32Regs:$dst, (rotr Int32Regs:$src, Int32Regs:$amt))]>,
1175 Requires<[noHWROT32]>;
1178 def ROT64imm_sw : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src,
1179 i32imm:$amt1, i32imm:$amt2),
1180 !strconcat("{{\n\t",
1181 !strconcat(".reg .b64 %lhs;\n\t",
1182 !strconcat(".reg .b64 %rhs;\n\t",
1183 !strconcat("shl.b64 \t%lhs, $src, $amt1;\n\t",
1184 !strconcat("shr.b64 \t%rhs, $src, $amt2;\n\t",
1185 !strconcat("add.u64 \t$dst, %lhs, %rhs;\n\t",
1186 !strconcat("}}", ""))))))),
1189 def SUB_FRM_64 : SDNodeXForm<imm, [{
1190 return CurDAG->getTargetConstant(64-N->getZExtValue(), MVT::i32);
1193 def : Pat<(rotl Int64Regs:$src, (i32 imm:$amt)),
1194 (ROT64imm_sw Int64Regs:$src, imm:$amt, (SUB_FRM_64 node:$amt))>;
1195 def : Pat<(rotr Int64Regs:$src, (i32 imm:$amt)),
1196 (ROT64imm_sw Int64Regs:$src, (SUB_FRM_64 node:$amt), imm:$amt)>;
1198 def ROTL64reg_sw : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src,
1200 !strconcat("{{\n\t",
1201 !strconcat(".reg .b64 %lhs;\n\t",
1202 !strconcat(".reg .b64 %rhs;\n\t",
1203 !strconcat(".reg .u32 %amt2;\n\t",
1204 !strconcat("shl.b64 \t%lhs, $src, $amt;\n\t",
1205 !strconcat("sub.u32 \t%amt2, 64, $amt;\n\t",
1206 !strconcat("shr.b64 \t%rhs, $src, %amt2;\n\t",
1207 !strconcat("add.u64 \t$dst, %lhs, %rhs;\n\t",
1208 !strconcat("}}", ""))))))))),
1209 [(set Int64Regs:$dst, (rotl Int64Regs:$src, Int32Regs:$amt))]>;
1211 def ROTR64reg_sw : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src,
1213 !strconcat("{{\n\t",
1214 !strconcat(".reg .b64 %lhs;\n\t",
1215 !strconcat(".reg .b64 %rhs;\n\t",
1216 !strconcat(".reg .u32 %amt2;\n\t",
1217 !strconcat("shr.b64 \t%lhs, $src, $amt;\n\t",
1218 !strconcat("sub.u32 \t%amt2, 64, $amt;\n\t",
1219 !strconcat("shl.b64 \t%rhs, $src, %amt2;\n\t",
1220 !strconcat("add.u64 \t$dst, %lhs, %rhs;\n\t",
1221 !strconcat("}}", ""))))))))),
1222 [(set Int64Regs:$dst, (rotr Int64Regs:$src, Int32Regs:$amt))]>;
1224 // BFE - bit-field extract
1226 multiclass BFE<string TyStr, RegisterClass RC> {
1227 // BFE supports both 32-bit and 64-bit values, but the start and length
1228 // operands are always 32-bit
1230 : NVPTXInst<(outs RC:$d),
1231 (ins RC:$a, Int32Regs:$b, Int32Regs:$c),
1232 !strconcat("bfe.", TyStr, " \t$d, $a, $b, $c;"), []>;
1234 : NVPTXInst<(outs RC:$d),
1235 (ins RC:$a, Int32Regs:$b, i32imm:$c),
1236 !strconcat("bfe.", TyStr, " \t$d, $a, $b, $c;"), []>;
1238 : NVPTXInst<(outs RC:$d),
1239 (ins RC:$a, i32imm:$b, i32imm:$c),
1240 !strconcat("bfe.", TyStr, " \t$d, $a, $b, $c;"), []>;
1243 defm BFE_S32 : BFE<"s32", Int32Regs>;
1244 defm BFE_U32 : BFE<"u32", Int32Regs>;
1245 defm BFE_S64 : BFE<"s64", Int64Regs>;
1246 defm BFE_U64 : BFE<"u64", Int64Regs>;
1248 //-----------------------------------
1249 // General Comparison
1250 //-----------------------------------
1252 // General setp instructions
1253 multiclass SETP<string TypeStr, RegisterClass RC, Operand ImmCls> {
1254 def rr : NVPTXInst<(outs Int1Regs:$dst),
1255 (ins RC:$a, RC:$b, CmpMode:$cmp),
1256 !strconcat("setp${cmp:base}${cmp:ftz}.", TypeStr, "\t$dst, $a, $b;"),
1258 def ri : NVPTXInst<(outs Int1Regs:$dst),
1259 (ins RC:$a, ImmCls:$b, CmpMode:$cmp),
1260 !strconcat("setp${cmp:base}${cmp:ftz}.", TypeStr, "\t$dst, $a, $b;"),
1262 def ir : NVPTXInst<(outs Int1Regs:$dst),
1263 (ins ImmCls:$a, RC:$b, CmpMode:$cmp),
1264 !strconcat("setp${cmp:base}${cmp:ftz}.", TypeStr, "\t$dst, $a, $b;"),
1268 defm SETP_b16 : SETP<"b16", Int16Regs, i16imm>;
1269 defm SETP_s16 : SETP<"s16", Int16Regs, i16imm>;
1270 defm SETP_u16 : SETP<"u16", Int16Regs, i16imm>;
1271 defm SETP_b32 : SETP<"b32", Int32Regs, i32imm>;
1272 defm SETP_s32 : SETP<"s32", Int32Regs, i32imm>;
1273 defm SETP_u32 : SETP<"u32", Int32Regs, i32imm>;
1274 defm SETP_b64 : SETP<"b64", Int64Regs, i64imm>;
1275 defm SETP_s64 : SETP<"s64", Int64Regs, i64imm>;
1276 defm SETP_u64 : SETP<"u64", Int64Regs, i64imm>;
1277 defm SETP_f32 : SETP<"f32", Float32Regs, f32imm>;
1278 defm SETP_f64 : SETP<"f64", Float64Regs, f64imm>;
1280 // General set instructions
1281 multiclass SET<string TypeStr, RegisterClass RC, Operand ImmCls> {
1282 def rr : NVPTXInst<(outs Int32Regs:$dst),
1283 (ins RC:$a, RC:$b, CmpMode:$cmp),
1284 !strconcat("set$cmp.", TypeStr, "\t$dst, $a, $b;"), []>;
1285 def ri : NVPTXInst<(outs Int32Regs:$dst),
1286 (ins RC:$a, ImmCls:$b, CmpMode:$cmp),
1287 !strconcat("set$cmp.", TypeStr, "\t$dst, $a, $b;"), []>;
1288 def ir : NVPTXInst<(outs Int32Regs:$dst),
1289 (ins ImmCls:$a, RC:$b, CmpMode:$cmp),
1290 !strconcat("set$cmp.", TypeStr, "\t$dst, $a, $b;"), []>;
1293 defm SET_b16 : SET<"b16", Int16Regs, i16imm>;
1294 defm SET_s16 : SET<"s16", Int16Regs, i16imm>;
1295 defm SET_u16 : SET<"u16", Int16Regs, i16imm>;
1296 defm SET_b32 : SET<"b32", Int32Regs, i32imm>;
1297 defm SET_s32 : SET<"s32", Int32Regs, i32imm>;
1298 defm SET_u32 : SET<"u32", Int32Regs, i32imm>;
1299 defm SET_b64 : SET<"b64", Int64Regs, i64imm>;
1300 defm SET_s64 : SET<"s64", Int64Regs, i64imm>;
1301 defm SET_u64 : SET<"u64", Int64Regs, i64imm>;
1302 defm SET_f32 : SET<"f32", Float32Regs, f32imm>;
1303 defm SET_f64 : SET<"f64", Float64Regs, f64imm>;
1305 //-----------------------------------
1306 // General Selection
1307 //-----------------------------------
1309 // General selp instructions
1310 multiclass SELP<string TypeStr, RegisterClass RC, Operand ImmCls> {
1311 def rr : NVPTXInst<(outs RC:$dst),
1312 (ins RC:$a, RC:$b, Int1Regs:$p),
1313 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
1314 def ri : NVPTXInst<(outs RC:$dst),
1315 (ins RC:$a, ImmCls:$b, Int1Regs:$p),
1316 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
1317 def ir : NVPTXInst<(outs RC:$dst),
1318 (ins ImmCls:$a, RC:$b, Int1Regs:$p),
1319 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
1320 def ii : NVPTXInst<(outs RC:$dst),
1321 (ins ImmCls:$a, ImmCls:$b, Int1Regs:$p),
1322 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
1325 multiclass SELP_PATTERN<string TypeStr, RegisterClass RC, Operand ImmCls,
1327 def rr : NVPTXInst<(outs RC:$dst),
1328 (ins RC:$a, RC:$b, Int1Regs:$p),
1329 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
1330 [(set RC:$dst, (select Int1Regs:$p, RC:$a, RC:$b))]>;
1331 def ri : NVPTXInst<(outs RC:$dst),
1332 (ins RC:$a, ImmCls:$b, Int1Regs:$p),
1333 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
1334 [(set RC:$dst, (select Int1Regs:$p, RC:$a, ImmNode:$b))]>;
1335 def ir : NVPTXInst<(outs RC:$dst),
1336 (ins ImmCls:$a, RC:$b, Int1Regs:$p),
1337 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
1338 [(set RC:$dst, (select Int1Regs:$p, ImmNode:$a, RC:$b))]>;
1339 def ii : NVPTXInst<(outs RC:$dst),
1340 (ins ImmCls:$a, ImmCls:$b, Int1Regs:$p),
1341 !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
1342 [(set RC:$dst, (select Int1Regs:$p, ImmNode:$a, ImmNode:$b))]>;
1345 defm SELP_b16 : SELP_PATTERN<"b16", Int16Regs, i16imm, imm>;
1346 defm SELP_s16 : SELP<"s16", Int16Regs, i16imm>;
1347 defm SELP_u16 : SELP<"u16", Int16Regs, i16imm>;
1348 defm SELP_b32 : SELP_PATTERN<"b32", Int32Regs, i32imm, imm>;
1349 defm SELP_s32 : SELP<"s32", Int32Regs, i32imm>;
1350 defm SELP_u32 : SELP<"u32", Int32Regs, i32imm>;
1351 defm SELP_b64 : SELP_PATTERN<"b64", Int64Regs, i64imm, imm>;
1352 defm SELP_s64 : SELP<"s64", Int64Regs, i64imm>;
1353 defm SELP_u64 : SELP<"u64", Int64Regs, i64imm>;
1354 defm SELP_f32 : SELP_PATTERN<"f32", Float32Regs, f32imm, fpimm>;
1355 defm SELP_f64 : SELP_PATTERN<"f64", Float64Regs, f64imm, fpimm>;
1357 // Special select for predicate operands
1358 def : Pat<(i1 (select Int1Regs:$p, Int1Regs:$a, Int1Regs:$b)),
1359 (ORb1rr (ANDb1rr Int1Regs:$p, Int1Regs:$a),
1360 (ANDb1rr (NOT1 Int1Regs:$p), Int1Regs:$b))>;
1362 //-----------------------------------
1363 // Data Movement (Load / Store, Move)
1364 //-----------------------------------
1366 def ADDRri : ComplexPattern<i32, 2, "SelectADDRri", [frameindex],
1368 def ADDRri64 : ComplexPattern<i64, 2, "SelectADDRri64", [frameindex],
1371 def MEMri : Operand<i32> {
1372 let PrintMethod = "printMemOperand";
1373 let MIOperandInfo = (ops Int32Regs, i32imm);
1375 def MEMri64 : Operand<i64> {
1376 let PrintMethod = "printMemOperand";
1377 let MIOperandInfo = (ops Int64Regs, i64imm);
1380 def imem : Operand<iPTR> {
1381 let PrintMethod = "printOperand";
1384 def imemAny : Operand<iPTRAny> {
1385 let PrintMethod = "printOperand";
1388 def LdStCode : Operand<i32> {
1389 let PrintMethod = "printLdStCode";
1392 def SDTWrapper : SDTypeProfile<1, 1, [SDTCisSameAs<0, 1>, SDTCisPtrTy<0>]>;
1393 def Wrapper : SDNode<"NVPTXISD::Wrapper", SDTWrapper>;
1395 def MOV_ADDR : NVPTXInst<(outs Int32Regs:$dst), (ins imem:$a),
1396 "mov.u32 \t$dst, $a;",
1397 [(set Int32Regs:$dst, (Wrapper tglobaladdr:$a))]>;
1399 def MOV_ADDR64 : NVPTXInst<(outs Int64Regs:$dst), (ins imem:$a),
1400 "mov.u64 \t$dst, $a;",
1401 [(set Int64Regs:$dst, (Wrapper tglobaladdr:$a))]>;
1403 // Get pointer to local stack
1405 : NVPTXInst<(outs Int32Regs:$d), (ins i32imm:$num),
1406 "mov.u32 \t$d, __local_depot$num;", []>;
1407 def MOV_DEPOT_ADDR_64
1408 : NVPTXInst<(outs Int64Regs:$d), (ins i32imm:$num),
1409 "mov.u64 \t$d, __local_depot$num;", []>;
1412 // copyPhysreg is hard-coded in NVPTXInstrInfo.cpp
1413 let IsSimpleMove=1 in {
1414 def IMOV1rr: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$sss),
1415 "mov.pred \t$dst, $sss;", []>;
1416 def IMOV16rr: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$sss),
1417 "mov.u16 \t$dst, $sss;", []>;
1418 def IMOV32rr: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$sss),
1419 "mov.u32 \t$dst, $sss;", []>;
1420 def IMOV64rr: NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$sss),
1421 "mov.u64 \t$dst, $sss;", []>;
1423 def FMOV32rr: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$src),
1424 "mov.f32 \t$dst, $src;", []>;
1425 def FMOV64rr: NVPTXInst<(outs Float64Regs:$dst), (ins Float64Regs:$src),
1426 "mov.f64 \t$dst, $src;", []>;
1428 def IMOV1ri: NVPTXInst<(outs Int1Regs:$dst), (ins i1imm:$src),
1429 "mov.pred \t$dst, $src;",
1430 [(set Int1Regs:$dst, imm:$src)]>;
1431 def IMOV16ri: NVPTXInst<(outs Int16Regs:$dst), (ins i16imm:$src),
1432 "mov.u16 \t$dst, $src;",
1433 [(set Int16Regs:$dst, imm:$src)]>;
1434 def IMOV32ri: NVPTXInst<(outs Int32Regs:$dst), (ins i32imm:$src),
1435 "mov.u32 \t$dst, $src;",
1436 [(set Int32Regs:$dst, imm:$src)]>;
1437 def IMOV64i: NVPTXInst<(outs Int64Regs:$dst), (ins i64imm:$src),
1438 "mov.u64 \t$dst, $src;",
1439 [(set Int64Regs:$dst, imm:$src)]>;
1441 def FMOV32ri: NVPTXInst<(outs Float32Regs:$dst), (ins f32imm:$src),
1442 "mov.f32 \t$dst, $src;",
1443 [(set Float32Regs:$dst, fpimm:$src)]>;
1444 def FMOV64ri: NVPTXInst<(outs Float64Regs:$dst), (ins f64imm:$src),
1445 "mov.f64 \t$dst, $src;",
1446 [(set Float64Regs:$dst, fpimm:$src)]>;
1448 def : Pat<(i32 (Wrapper texternalsym:$dst)), (IMOV32ri texternalsym:$dst)>;
1450 //---- Copy Frame Index ----
1451 def LEA_ADDRi : NVPTXInst<(outs Int32Regs:$dst), (ins MEMri:$addr),
1452 "add.u32 \t$dst, ${addr:add};",
1453 [(set Int32Regs:$dst, ADDRri:$addr)]>;
1454 def LEA_ADDRi64 : NVPTXInst<(outs Int64Regs:$dst), (ins MEMri64:$addr),
1455 "add.u64 \t$dst, ${addr:add};",
1456 [(set Int64Regs:$dst, ADDRri64:$addr)]>;
1458 //-----------------------------------
1459 // Comparison and Selection
1460 //-----------------------------------
1462 multiclass ISET_FORMAT<PatFrag OpNode, PatLeaf Mode,
1463 Instruction setp_16rr,
1464 Instruction setp_16ri,
1465 Instruction setp_16ir,
1466 Instruction setp_32rr,
1467 Instruction setp_32ri,
1468 Instruction setp_32ir,
1469 Instruction setp_64rr,
1470 Instruction setp_64ri,
1471 Instruction setp_64ir,
1472 Instruction set_16rr,
1473 Instruction set_16ri,
1474 Instruction set_16ir,
1475 Instruction set_32rr,
1476 Instruction set_32ri,
1477 Instruction set_32ir,
1478 Instruction set_64rr,
1479 Instruction set_64ri,
1480 Instruction set_64ir> {
1482 def : Pat<(i1 (OpNode Int16Regs:$a, Int16Regs:$b)),
1483 (setp_16rr Int16Regs:$a, Int16Regs:$b, Mode)>;
1484 def : Pat<(i1 (OpNode Int16Regs:$a, imm:$b)),
1485 (setp_16ri Int16Regs:$a, imm:$b, Mode)>;
1486 def : Pat<(i1 (OpNode imm:$a, Int16Regs:$b)),
1487 (setp_16ir imm:$a, Int16Regs:$b, Mode)>;
1489 def : Pat<(i1 (OpNode Int32Regs:$a, Int32Regs:$b)),
1490 (setp_32rr Int32Regs:$a, Int32Regs:$b, Mode)>;
1491 def : Pat<(i1 (OpNode Int32Regs:$a, imm:$b)),
1492 (setp_32ri Int32Regs:$a, imm:$b, Mode)>;
1493 def : Pat<(i1 (OpNode imm:$a, Int32Regs:$b)),
1494 (setp_32ir imm:$a, Int32Regs:$b, Mode)>;
1496 def : Pat<(i1 (OpNode Int64Regs:$a, Int64Regs:$b)),
1497 (setp_64rr Int64Regs:$a, Int64Regs:$b, Mode)>;
1498 def : Pat<(i1 (OpNode Int64Regs:$a, imm:$b)),
1499 (setp_64ri Int64Regs:$a, imm:$b, Mode)>;
1500 def : Pat<(i1 (OpNode imm:$a, Int64Regs:$b)),
1501 (setp_64ir imm:$a, Int64Regs:$b, Mode)>;
1504 def : Pat<(i32 (OpNode Int16Regs:$a, Int16Regs:$b)),
1505 (set_16rr Int16Regs:$a, Int16Regs:$b, Mode)>;
1506 def : Pat<(i32 (OpNode Int16Regs:$a, imm:$b)),
1507 (set_16ri Int16Regs:$a, imm:$b, Mode)>;
1508 def : Pat<(i32 (OpNode imm:$a, Int16Regs:$b)),
1509 (set_16ir imm:$a, Int16Regs:$b, Mode)>;
1511 def : Pat<(i32 (OpNode Int32Regs:$a, Int32Regs:$b)),
1512 (set_32rr Int32Regs:$a, Int32Regs:$b, Mode)>;
1513 def : Pat<(i32 (OpNode Int32Regs:$a, imm:$b)),
1514 (set_32ri Int32Regs:$a, imm:$b, Mode)>;
1515 def : Pat<(i32 (OpNode imm:$a, Int32Regs:$b)),
1516 (set_32ir imm:$a, Int32Regs:$b, Mode)>;
1518 def : Pat<(i32 (OpNode Int64Regs:$a, Int64Regs:$b)),
1519 (set_64rr Int64Regs:$a, Int64Regs:$b, Mode)>;
1520 def : Pat<(i32 (OpNode Int64Regs:$a, imm:$b)),
1521 (set_64ri Int64Regs:$a, imm:$b, Mode)>;
1522 def : Pat<(i32 (OpNode imm:$a, Int64Regs:$b)),
1523 (set_64ir imm:$a, Int64Regs:$b, Mode)>;
1526 multiclass ISET_FORMAT_SIGNED<PatFrag OpNode, PatLeaf Mode>
1527 : ISET_FORMAT<OpNode, Mode,
1528 SETP_s16rr, SETP_s16ri, SETP_s16ir,
1529 SETP_s32rr, SETP_s32ri, SETP_s32ir,
1530 SETP_s64rr, SETP_s64ri, SETP_s64ir,
1531 SET_s16rr, SET_s16ri, SET_s16ir,
1532 SET_s32rr, SET_s32ri, SET_s32ir,
1533 SET_s64rr, SET_s64ri, SET_s64ir> {
1534 // TableGen doesn't like empty multiclasses
1535 def : PatLeaf<(i32 0)>;
1538 multiclass ISET_FORMAT_UNSIGNED<PatFrag OpNode, PatLeaf Mode>
1539 : ISET_FORMAT<OpNode, Mode,
1540 SETP_u16rr, SETP_u16ri, SETP_u16ir,
1541 SETP_u32rr, SETP_u32ri, SETP_u32ir,
1542 SETP_u64rr, SETP_u64ri, SETP_u64ir,
1543 SET_u16rr, SET_u16ri, SET_u16ir,
1544 SET_u32rr, SET_u32ri, SET_u32ir,
1545 SET_u64rr, SET_u64ri, SET_u64ir> {
1546 // TableGen doesn't like empty multiclasses
1547 def : PatLeaf<(i32 0)>;
1550 defm : ISET_FORMAT_SIGNED<setgt, CmpGT>;
1551 defm : ISET_FORMAT_UNSIGNED<setugt, CmpGT>;
1552 defm : ISET_FORMAT_SIGNED<setlt, CmpLT>;
1553 defm : ISET_FORMAT_UNSIGNED<setult, CmpLT>;
1554 defm : ISET_FORMAT_SIGNED<setge, CmpGE>;
1555 defm : ISET_FORMAT_UNSIGNED<setuge, CmpGE>;
1556 defm : ISET_FORMAT_SIGNED<setle, CmpLE>;
1557 defm : ISET_FORMAT_UNSIGNED<setule, CmpLE>;
1558 defm : ISET_FORMAT_SIGNED<seteq, CmpEQ>;
1559 defm : ISET_FORMAT_UNSIGNED<setueq, CmpEQ>;
1560 defm : ISET_FORMAT_SIGNED<setne, CmpNE>;
1561 defm : ISET_FORMAT_UNSIGNED<setune, CmpNE>;
1564 def : Pat<(setne Int1Regs:$a, Int1Regs:$b),
1565 (XORb1rr Int1Regs:$a, Int1Regs:$b)>;
1566 def : Pat<(setune Int1Regs:$a, Int1Regs:$b),
1567 (XORb1rr Int1Regs:$a, Int1Regs:$b)>;
1569 def : Pat<(seteq Int1Regs:$a, Int1Regs:$b),
1570 (NOT1 (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
1571 def : Pat<(setueq Int1Regs:$a, Int1Regs:$b),
1572 (NOT1 (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
1574 // i1 compare -> i32
1575 def : Pat<(i32 (setne Int1Regs:$a, Int1Regs:$b)),
1576 (SELP_u32ii -1, 0, (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
1577 def : Pat<(i32 (setne Int1Regs:$a, Int1Regs:$b)),
1578 (SELP_u32ii 0, -1, (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
1582 multiclass FSET_FORMAT<PatFrag OpNode, PatLeaf Mode, PatLeaf ModeFTZ> {
1584 def : Pat<(i1 (OpNode Float32Regs:$a, Float32Regs:$b)),
1585 (SETP_f32rr Float32Regs:$a, Float32Regs:$b, ModeFTZ)>,
1586 Requires<[doF32FTZ]>;
1587 def : Pat<(i1 (OpNode Float32Regs:$a, Float32Regs:$b)),
1588 (SETP_f32rr Float32Regs:$a, Float32Regs:$b, Mode)>;
1589 def : Pat<(i1 (OpNode Float32Regs:$a, fpimm:$b)),
1590 (SETP_f32ri Float32Regs:$a, fpimm:$b, ModeFTZ)>,
1591 Requires<[doF32FTZ]>;
1592 def : Pat<(i1 (OpNode Float32Regs:$a, fpimm:$b)),
1593 (SETP_f32ri Float32Regs:$a, fpimm:$b, Mode)>;
1594 def : Pat<(i1 (OpNode fpimm:$a, Float32Regs:$b)),
1595 (SETP_f32ir fpimm:$a, Float32Regs:$b, ModeFTZ)>,
1596 Requires<[doF32FTZ]>;
1597 def : Pat<(i1 (OpNode fpimm:$a, Float32Regs:$b)),
1598 (SETP_f32ir fpimm:$a, Float32Regs:$b, Mode)>;
1601 def : Pat<(i1 (OpNode Float64Regs:$a, Float64Regs:$b)),
1602 (SETP_f64rr Float64Regs:$a, Float64Regs:$b, Mode)>;
1603 def : Pat<(i1 (OpNode Float64Regs:$a, fpimm:$b)),
1604 (SETP_f64ri Float64Regs:$a, fpimm:$b, Mode)>;
1605 def : Pat<(i1 (OpNode fpimm:$a, Float64Regs:$b)),
1606 (SETP_f64ir fpimm:$a, Float64Regs:$b, Mode)>;
1609 def : Pat<(i32 (OpNode Float32Regs:$a, Float32Regs:$b)),
1610 (SET_f32rr Float32Regs:$a, Float32Regs:$b, ModeFTZ)>,
1611 Requires<[doF32FTZ]>;
1612 def : Pat<(i32 (OpNode Float32Regs:$a, Float32Regs:$b)),
1613 (SET_f32rr Float32Regs:$a, Float32Regs:$b, Mode)>;
1614 def : Pat<(i32 (OpNode Float32Regs:$a, fpimm:$b)),
1615 (SET_f32ri Float32Regs:$a, fpimm:$b, ModeFTZ)>,
1616 Requires<[doF32FTZ]>;
1617 def : Pat<(i32 (OpNode Float32Regs:$a, fpimm:$b)),
1618 (SET_f32ri Float32Regs:$a, fpimm:$b, Mode)>;
1619 def : Pat<(i32 (OpNode fpimm:$a, Float32Regs:$b)),
1620 (SET_f32ir fpimm:$a, Float32Regs:$b, ModeFTZ)>,
1621 Requires<[doF32FTZ]>;
1622 def : Pat<(i32 (OpNode fpimm:$a, Float32Regs:$b)),
1623 (SET_f32ir fpimm:$a, Float32Regs:$b, Mode)>;
1626 def : Pat<(i32 (OpNode Float64Regs:$a, Float64Regs:$b)),
1627 (SET_f64rr Float64Regs:$a, Float64Regs:$b, Mode)>;
1628 def : Pat<(i32 (OpNode Float64Regs:$a, fpimm:$b)),
1629 (SET_f64ri Float64Regs:$a, fpimm:$b, Mode)>;
1630 def : Pat<(i32 (OpNode fpimm:$a, Float64Regs:$b)),
1631 (SET_f64ir fpimm:$a, Float64Regs:$b, Mode)>;
1634 defm FSetGT : FSET_FORMAT<setogt, CmpGT, CmpGT_FTZ>;
1635 defm FSetLT : FSET_FORMAT<setolt, CmpLT, CmpLT_FTZ>;
1636 defm FSetGE : FSET_FORMAT<setoge, CmpGE, CmpGE_FTZ>;
1637 defm FSetLE : FSET_FORMAT<setole, CmpLE, CmpLE_FTZ>;
1638 defm FSetEQ : FSET_FORMAT<setoeq, CmpEQ, CmpEQ_FTZ>;
1639 defm FSetNE : FSET_FORMAT<setone, CmpNE, CmpNE_FTZ>;
1641 defm FSetUGT : FSET_FORMAT<setugt, CmpGTU, CmpGTU_FTZ>;
1642 defm FSetULT : FSET_FORMAT<setult, CmpLTU, CmpLTU_FTZ>;
1643 defm FSetUGE : FSET_FORMAT<setuge, CmpGEU, CmpGEU_FTZ>;
1644 defm FSetULE : FSET_FORMAT<setule, CmpLEU, CmpLEU_FTZ>;
1645 defm FSetUEQ : FSET_FORMAT<setueq, CmpEQU, CmpEQU_FTZ>;
1646 defm FSetUNE : FSET_FORMAT<setune, CmpNEU, CmpNEU_FTZ>;
1648 defm FSetNUM : FSET_FORMAT<seto, CmpNUM, CmpNUM_FTZ>;
1649 defm FSetNAN : FSET_FORMAT<setuo, CmpNAN, CmpNAN_FTZ>;
1651 //def ld_param : SDNode<"NVPTXISD::LOAD_PARAM", SDTLoad,
1652 // [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
1654 def SDTDeclareParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>,
1656 def SDTDeclareScalarParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>,
1657 SDTCisInt<1>, SDTCisInt<2>]>;
1658 def SDTLoadParamProfile : SDTypeProfile<1, 2, [SDTCisInt<1>, SDTCisInt<2>]>;
1659 def SDTLoadParamV2Profile : SDTypeProfile<2, 2, [SDTCisSameAs<0, 1>, SDTCisInt<2>, SDTCisInt<3>]>;
1660 def SDTLoadParamV4Profile : SDTypeProfile<4, 2, [SDTCisInt<4>, SDTCisInt<5>]>;
1661 def SDTPrintCallProfile : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
1662 def SDTPrintCallUniProfile : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
1663 def SDTStoreParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>]>;
1664 def SDTStoreParamV2Profile : SDTypeProfile<0, 4, [SDTCisInt<0>, SDTCisInt<1>]>;
1665 def SDTStoreParamV4Profile : SDTypeProfile<0, 6, [SDTCisInt<0>, SDTCisInt<1>]>;
1666 def SDTStoreParam32Profile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>]>;
1667 def SDTCallArgProfile : SDTypeProfile<0, 2, [SDTCisInt<0>]>;
1668 def SDTCallArgMarkProfile : SDTypeProfile<0, 0, []>;
1669 def SDTCallVoidProfile : SDTypeProfile<0, 1, []>;
1670 def SDTCallValProfile : SDTypeProfile<1, 0, []>;
1671 def SDTMoveParamProfile : SDTypeProfile<1, 1, []>;
1672 def SDTStoreRetvalProfile : SDTypeProfile<0, 2, [SDTCisInt<0>]>;
1673 def SDTStoreRetvalV2Profile : SDTypeProfile<0, 3, [SDTCisInt<0>]>;
1674 def SDTStoreRetvalV4Profile : SDTypeProfile<0, 5, [SDTCisInt<0>]>;
1675 def SDTPseudoUseParamProfile : SDTypeProfile<0, 1, []>;
1677 def DeclareParam : SDNode<"NVPTXISD::DeclareParam", SDTDeclareParamProfile,
1678 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1679 def DeclareScalarParam : SDNode<"NVPTXISD::DeclareScalarParam",
1680 SDTDeclareScalarParamProfile,
1681 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1682 def DeclareRetParam : SDNode<"NVPTXISD::DeclareRetParam",
1683 SDTDeclareParamProfile,
1684 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1685 def DeclareRet : SDNode<"NVPTXISD::DeclareRet", SDTDeclareScalarParamProfile,
1686 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1687 def LoadParam : SDNode<"NVPTXISD::LoadParam", SDTLoadParamProfile,
1688 [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
1689 def LoadParamV2 : SDNode<"NVPTXISD::LoadParamV2", SDTLoadParamV2Profile,
1690 [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
1691 def LoadParamV4 : SDNode<"NVPTXISD::LoadParamV4", SDTLoadParamV4Profile,
1692 [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
1693 def PrintCall : SDNode<"NVPTXISD::PrintCall", SDTPrintCallProfile,
1694 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1695 def PrintCallUni : SDNode<"NVPTXISD::PrintCallUni", SDTPrintCallUniProfile,
1696 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1697 def StoreParam : SDNode<"NVPTXISD::StoreParam", SDTStoreParamProfile,
1698 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1699 def StoreParamV2 : SDNode<"NVPTXISD::StoreParamV2", SDTStoreParamV2Profile,
1700 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1701 def StoreParamV4 : SDNode<"NVPTXISD::StoreParamV4", SDTStoreParamV4Profile,
1702 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1703 def StoreParamU32 : SDNode<"NVPTXISD::StoreParamU32", SDTStoreParam32Profile,
1704 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1705 def StoreParamS32 : SDNode<"NVPTXISD::StoreParamS32", SDTStoreParam32Profile,
1706 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1707 def CallArgBegin : SDNode<"NVPTXISD::CallArgBegin", SDTCallArgMarkProfile,
1708 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1709 def CallArg : SDNode<"NVPTXISD::CallArg", SDTCallArgProfile,
1710 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1711 def LastCallArg : SDNode<"NVPTXISD::LastCallArg", SDTCallArgProfile,
1712 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1713 def CallArgEnd : SDNode<"NVPTXISD::CallArgEnd", SDTCallVoidProfile,
1714 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1715 def CallVoid : SDNode<"NVPTXISD::CallVoid", SDTCallVoidProfile,
1716 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1717 def Prototype : SDNode<"NVPTXISD::Prototype", SDTCallVoidProfile,
1718 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1719 def CallVal : SDNode<"NVPTXISD::CallVal", SDTCallValProfile,
1720 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1721 def MoveParam : SDNode<"NVPTXISD::MoveParam", SDTMoveParamProfile,
1723 def StoreRetval : SDNode<"NVPTXISD::StoreRetval", SDTStoreRetvalProfile,
1724 [SDNPHasChain, SDNPSideEffect]>;
1725 def StoreRetvalV2 : SDNode<"NVPTXISD::StoreRetvalV2", SDTStoreRetvalV2Profile,
1726 [SDNPHasChain, SDNPSideEffect]>;
1727 def StoreRetvalV4 : SDNode<"NVPTXISD::StoreRetvalV4", SDTStoreRetvalV4Profile,
1728 [SDNPHasChain, SDNPSideEffect]>;
1729 def PseudoUseParam : SDNode<"NVPTXISD::PseudoUseParam",
1730 SDTPseudoUseParamProfile,
1731 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
1732 def RETURNNode : SDNode<"NVPTXISD::RETURN", SDTCallArgMarkProfile,
1733 [SDNPHasChain, SDNPSideEffect]>;
1735 class LoadParamMemInst<NVPTXRegClass regclass, string opstr> :
1736 NVPTXInst<(outs regclass:$dst), (ins i32imm:$b),
1737 !strconcat(!strconcat("ld.param", opstr),
1738 "\t$dst, [retval0+$b];"),
1741 class LoadParamRegInst<NVPTXRegClass regclass, string opstr> :
1742 NVPTXInst<(outs regclass:$dst), (ins i32imm:$b),
1743 !strconcat(!strconcat("mov", opstr),
1744 "\t$dst, retval$b;"),
1745 [(set regclass:$dst, (LoadParam (i32 0), (i32 imm:$b)))]>;
1747 class LoadParamV2MemInst<NVPTXRegClass regclass, string opstr> :
1748 NVPTXInst<(outs regclass:$dst, regclass:$dst2), (ins i32imm:$b),
1749 !strconcat(!strconcat("ld.param.v2", opstr),
1750 "\t{{$dst, $dst2}}, [retval0+$b];"), []>;
1752 class LoadParamV4MemInst<NVPTXRegClass regclass, string opstr> :
1753 NVPTXInst<(outs regclass:$dst, regclass:$dst2, regclass:$dst3,
1756 !strconcat(!strconcat("ld.param.v4", opstr),
1757 "\t{{$dst, $dst2, $dst3, $dst4}}, [retval0+$b];"), []>;
1759 class StoreParamInst<NVPTXRegClass regclass, string opstr> :
1760 NVPTXInst<(outs), (ins regclass:$val, i32imm:$a, i32imm:$b),
1761 !strconcat(!strconcat("st.param", opstr),
1762 "\t[param$a+$b], $val;"),
1765 class StoreParamV2Inst<NVPTXRegClass regclass, string opstr> :
1766 NVPTXInst<(outs), (ins regclass:$val, regclass:$val2,
1767 i32imm:$a, i32imm:$b),
1768 !strconcat(!strconcat("st.param.v2", opstr),
1769 "\t[param$a+$b], {{$val, $val2}};"),
1772 class StoreParamV4Inst<NVPTXRegClass regclass, string opstr> :
1773 NVPTXInst<(outs), (ins regclass:$val, regclass:$val1, regclass:$val2,
1774 regclass:$val3, i32imm:$a, i32imm:$b),
1775 !strconcat(!strconcat("st.param.v4", opstr),
1776 "\t[param$a+$b], {{$val, $val2, $val3, $val4}};"),
1779 class StoreRetvalInst<NVPTXRegClass regclass, string opstr> :
1780 NVPTXInst<(outs), (ins regclass:$val, i32imm:$a),
1781 !strconcat(!strconcat("st.param", opstr),
1782 "\t[func_retval0+$a], $val;"),
1785 class StoreRetvalV2Inst<NVPTXRegClass regclass, string opstr> :
1786 NVPTXInst<(outs), (ins regclass:$val, regclass:$val2, i32imm:$a),
1787 !strconcat(!strconcat("st.param.v2", opstr),
1788 "\t[func_retval0+$a], {{$val, $val2}};"),
1791 class StoreRetvalV4Inst<NVPTXRegClass regclass, string opstr> :
1793 (ins regclass:$val, regclass:$val2, regclass:$val3,
1794 regclass:$val4, i32imm:$a),
1795 !strconcat(!strconcat("st.param.v4", opstr),
1796 "\t[func_retval0+$a], {{$val, $val2, $val3, $val4}};"),
1799 def PrintCallRetInst1 : NVPTXInst<(outs), (ins),
1801 [(PrintCall (i32 1))]>;
1802 def PrintCallRetInst2 : NVPTXInst<(outs), (ins),
1803 "call (retval0, retval1), ",
1804 [(PrintCall (i32 2))]>;
1805 def PrintCallRetInst3 : NVPTXInst<(outs), (ins),
1806 "call (retval0, retval1, retval2), ",
1807 [(PrintCall (i32 3))]>;
1808 def PrintCallRetInst4 : NVPTXInst<(outs), (ins),
1809 "call (retval0, retval1, retval2, retval3), ",
1810 [(PrintCall (i32 4))]>;
1811 def PrintCallRetInst5 : NVPTXInst<(outs), (ins),
1812 "call (retval0, retval1, retval2, retval3, retval4), ",
1813 [(PrintCall (i32 5))]>;
1814 def PrintCallRetInst6 : NVPTXInst<(outs), (ins),
1815 "call (retval0, retval1, retval2, retval3, retval4, retval5), ",
1816 [(PrintCall (i32 6))]>;
1817 def PrintCallRetInst7 : NVPTXInst<(outs), (ins),
1818 "call (retval0, retval1, retval2, retval3, retval4, retval5, retval6), ",
1819 [(PrintCall (i32 7))]>;
1820 def PrintCallRetInst8 : NVPTXInst<(outs), (ins),
1821 !strconcat("call (retval0, retval1, retval2, retval3, retval4",
1822 ", retval5, retval6, retval7), "),
1823 [(PrintCall (i32 8))]>;
1825 def PrintCallNoRetInst : NVPTXInst<(outs), (ins), "call ",
1826 [(PrintCall (i32 0))]>;
1828 def PrintCallUniRetInst1 : NVPTXInst<(outs), (ins),
1829 "call.uni (retval0), ",
1830 [(PrintCallUni (i32 1))]>;
1831 def PrintCallUniRetInst2 : NVPTXInst<(outs), (ins),
1832 "call.uni (retval0, retval1), ",
1833 [(PrintCallUni (i32 2))]>;
1834 def PrintCallUniRetInst3 : NVPTXInst<(outs), (ins),
1835 "call.uni (retval0, retval1, retval2), ",
1836 [(PrintCallUni (i32 3))]>;
1837 def PrintCallUniRetInst4 : NVPTXInst<(outs), (ins),
1838 "call.uni (retval0, retval1, retval2, retval3), ",
1839 [(PrintCallUni (i32 4))]>;
1840 def PrintCallUniRetInst5 : NVPTXInst<(outs), (ins),
1841 "call.uni (retval0, retval1, retval2, retval3, retval4), ",
1842 [(PrintCallUni (i32 5))]>;
1843 def PrintCallUniRetInst6 : NVPTXInst<(outs), (ins),
1844 "call.uni (retval0, retval1, retval2, retval3, retval4, retval5), ",
1845 [(PrintCallUni (i32 6))]>;
1846 def PrintCallUniRetInst7 : NVPTXInst<(outs), (ins),
1847 "call.uni (retval0, retval1, retval2, retval3, retval4, retval5, retval6), ",
1848 [(PrintCallUni (i32 7))]>;
1849 def PrintCallUniRetInst8 : NVPTXInst<(outs), (ins),
1850 !strconcat("call.uni (retval0, retval1, retval2, retval3, retval4",
1851 ", retval5, retval6, retval7), "),
1852 [(PrintCallUni (i32 8))]>;
1854 def PrintCallUniNoRetInst : NVPTXInst<(outs), (ins), "call.uni ",
1855 [(PrintCallUni (i32 0))]>;
1857 def LoadParamMemI64 : LoadParamMemInst<Int64Regs, ".b64">;
1858 def LoadParamMemI32 : LoadParamMemInst<Int32Regs, ".b32">;
1859 def LoadParamMemI16 : LoadParamMemInst<Int16Regs, ".b16">;
1860 def LoadParamMemI8 : LoadParamMemInst<Int16Regs, ".b8">;
1861 def LoadParamMemV2I64 : LoadParamV2MemInst<Int64Regs, ".b64">;
1862 def LoadParamMemV2I32 : LoadParamV2MemInst<Int32Regs, ".b32">;
1863 def LoadParamMemV2I16 : LoadParamV2MemInst<Int16Regs, ".b16">;
1864 def LoadParamMemV2I8 : LoadParamV2MemInst<Int16Regs, ".b8">;
1865 def LoadParamMemV4I32 : LoadParamV4MemInst<Int32Regs, ".b32">;
1866 def LoadParamMemV4I16 : LoadParamV4MemInst<Int16Regs, ".b16">;
1867 def LoadParamMemV4I8 : LoadParamV4MemInst<Int16Regs, ".b8">;
1868 def LoadParamMemF32 : LoadParamMemInst<Float32Regs, ".f32">;
1869 def LoadParamMemF64 : LoadParamMemInst<Float64Regs, ".f64">;
1870 def LoadParamMemV2F32 : LoadParamV2MemInst<Float32Regs, ".f32">;
1871 def LoadParamMemV2F64 : LoadParamV2MemInst<Float64Regs, ".f64">;
1872 def LoadParamMemV4F32 : LoadParamV4MemInst<Float32Regs, ".f32">;
1874 def StoreParamI64 : StoreParamInst<Int64Regs, ".b64">;
1875 def StoreParamI32 : StoreParamInst<Int32Regs, ".b32">;
1877 def StoreParamI16 : StoreParamInst<Int16Regs, ".b16">;
1878 def StoreParamI8 : StoreParamInst<Int16Regs, ".b8">;
1879 def StoreParamV2I64 : StoreParamV2Inst<Int64Regs, ".b64">;
1880 def StoreParamV2I32 : StoreParamV2Inst<Int32Regs, ".b32">;
1881 def StoreParamV2I16 : StoreParamV2Inst<Int16Regs, ".b16">;
1882 def StoreParamV2I8 : StoreParamV2Inst<Int16Regs, ".b8">;
1884 // FIXME: StoreParamV4Inst crashes llvm-tblgen :(
1885 //def StoreParamV4I32 : StoreParamV4Inst<Int32Regs, ".b32">;
1886 def StoreParamV4I32 : NVPTXInst<(outs), (ins Int32Regs:$val, Int32Regs:$val2,
1887 Int32Regs:$val3, Int32Regs:$val4,
1888 i32imm:$a, i32imm:$b),
1889 "st.param.b32\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
1892 def StoreParamV4I16 : NVPTXInst<(outs), (ins Int16Regs:$val, Int16Regs:$val2,
1893 Int16Regs:$val3, Int16Regs:$val4,
1894 i32imm:$a, i32imm:$b),
1895 "st.param.v4.b16\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
1898 def StoreParamV4I8 : NVPTXInst<(outs), (ins Int16Regs:$val, Int16Regs:$val2,
1899 Int16Regs:$val3, Int16Regs:$val4,
1900 i32imm:$a, i32imm:$b),
1901 "st.param.v4.b8\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
1904 def StoreParamF32 : StoreParamInst<Float32Regs, ".f32">;
1905 def StoreParamF64 : StoreParamInst<Float64Regs, ".f64">;
1906 def StoreParamV2F32 : StoreParamV2Inst<Float32Regs, ".f32">;
1907 def StoreParamV2F64 : StoreParamV2Inst<Float64Regs, ".f64">;
1908 // FIXME: StoreParamV4Inst crashes llvm-tblgen :(
1909 //def StoreParamV4F32 : StoreParamV4Inst<Float32Regs, ".f32">;
1910 def StoreParamV4F32 : NVPTXInst<(outs),
1911 (ins Float32Regs:$val, Float32Regs:$val2,
1912 Float32Regs:$val3, Float32Regs:$val4,
1913 i32imm:$a, i32imm:$b),
1914 "st.param.v4.f32\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
1918 def StoreRetvalI64 : StoreRetvalInst<Int64Regs, ".b64">;
1919 def StoreRetvalI32 : StoreRetvalInst<Int32Regs, ".b32">;
1920 def StoreRetvalI16 : StoreRetvalInst<Int16Regs, ".b16">;
1921 def StoreRetvalI8 : StoreRetvalInst<Int16Regs, ".b8">;
1922 def StoreRetvalV2I64 : StoreRetvalV2Inst<Int64Regs, ".b64">;
1923 def StoreRetvalV2I32 : StoreRetvalV2Inst<Int32Regs, ".b32">;
1924 def StoreRetvalV2I16 : StoreRetvalV2Inst<Int16Regs, ".b16">;
1925 def StoreRetvalV2I8 : StoreRetvalV2Inst<Int16Regs, ".b8">;
1926 def StoreRetvalV4I32 : StoreRetvalV4Inst<Int32Regs, ".b32">;
1927 def StoreRetvalV4I16 : StoreRetvalV4Inst<Int16Regs, ".b16">;
1928 def StoreRetvalV4I8 : StoreRetvalV4Inst<Int16Regs, ".b8">;
1930 def StoreRetvalF64 : StoreRetvalInst<Float64Regs, ".f64">;
1931 def StoreRetvalF32 : StoreRetvalInst<Float32Regs, ".f32">;
1932 def StoreRetvalV2F64 : StoreRetvalV2Inst<Float64Regs, ".f64">;
1933 def StoreRetvalV2F32 : StoreRetvalV2Inst<Float32Regs, ".f32">;
1934 def StoreRetvalV4F32 : StoreRetvalV4Inst<Float32Regs, ".f32">;
1936 def CallArgBeginInst : NVPTXInst<(outs), (ins), "(", [(CallArgBegin)]>;
1937 def CallArgEndInst1 : NVPTXInst<(outs), (ins), ");", [(CallArgEnd (i32 1))]>;
1938 def CallArgEndInst0 : NVPTXInst<(outs), (ins), ")", [(CallArgEnd (i32 0))]>;
1939 def RETURNInst : NVPTXInst<(outs), (ins), "ret;", [(RETURNNode)]>;
1941 class CallArgInst<NVPTXRegClass regclass> :
1942 NVPTXInst<(outs), (ins regclass:$a), "$a, ",
1943 [(CallArg (i32 0), regclass:$a)]>;
1945 class LastCallArgInst<NVPTXRegClass regclass> :
1946 NVPTXInst<(outs), (ins regclass:$a), "$a",
1947 [(LastCallArg (i32 0), regclass:$a)]>;
1949 def CallArgI64 : CallArgInst<Int64Regs>;
1950 def CallArgI32 : CallArgInst<Int32Regs>;
1951 def CallArgI16 : CallArgInst<Int16Regs>;
1953 def CallArgF64 : CallArgInst<Float64Regs>;
1954 def CallArgF32 : CallArgInst<Float32Regs>;
1956 def LastCallArgI64 : LastCallArgInst<Int64Regs>;
1957 def LastCallArgI32 : LastCallArgInst<Int32Regs>;
1958 def LastCallArgI16 : LastCallArgInst<Int16Regs>;
1960 def LastCallArgF64 : LastCallArgInst<Float64Regs>;
1961 def LastCallArgF32 : LastCallArgInst<Float32Regs>;
1963 def CallArgI32imm : NVPTXInst<(outs), (ins i32imm:$a), "$a, ",
1964 [(CallArg (i32 0), (i32 imm:$a))]>;
1965 def LastCallArgI32imm : NVPTXInst<(outs), (ins i32imm:$a), "$a",
1966 [(LastCallArg (i32 0), (i32 imm:$a))]>;
1968 def CallArgParam : NVPTXInst<(outs), (ins i32imm:$a), "param$a, ",
1969 [(CallArg (i32 1), (i32 imm:$a))]>;
1970 def LastCallArgParam : NVPTXInst<(outs), (ins i32imm:$a), "param$a",
1971 [(LastCallArg (i32 1), (i32 imm:$a))]>;
1973 def CallVoidInst : NVPTXInst<(outs), (ins imem:$addr),
1975 [(CallVoid (Wrapper tglobaladdr:$addr))]>;
1976 def CallVoidInstReg : NVPTXInst<(outs), (ins Int32Regs:$addr),
1978 [(CallVoid Int32Regs:$addr)]>;
1979 def CallVoidInstReg64 : NVPTXInst<(outs), (ins Int64Regs:$addr),
1981 [(CallVoid Int64Regs:$addr)]>;
1982 def PrototypeInst : NVPTXInst<(outs), (ins i32imm:$val),
1983 ", prototype_$val;",
1984 [(Prototype (i32 imm:$val))]>;
1986 def DeclareRetMemInst : NVPTXInst<(outs),
1987 (ins i32imm:$align, i32imm:$size, i32imm:$num),
1988 ".param .align $align .b8 retval$num[$size];",
1989 [(DeclareRetParam (i32 imm:$align), (i32 imm:$size), (i32 imm:$num))]>;
1990 def DeclareRetScalarInst : NVPTXInst<(outs), (ins i32imm:$size, i32imm:$num),
1991 ".param .b$size retval$num;",
1992 [(DeclareRet (i32 1), (i32 imm:$size), (i32 imm:$num))]>;
1993 def DeclareRetRegInst : NVPTXInst<(outs), (ins i32imm:$size, i32imm:$num),
1994 ".reg .b$size retval$num;",
1995 [(DeclareRet (i32 2), (i32 imm:$size), (i32 imm:$num))]>;
1997 def DeclareParamInst : NVPTXInst<(outs),
1998 (ins i32imm:$align, i32imm:$a, i32imm:$size),
1999 ".param .align $align .b8 param$a[$size];",
2000 [(DeclareParam (i32 imm:$align), (i32 imm:$a), (i32 imm:$size))]>;
2001 def DeclareScalarParamInst : NVPTXInst<(outs), (ins i32imm:$a, i32imm:$size),
2002 ".param .b$size param$a;",
2003 [(DeclareScalarParam (i32 imm:$a), (i32 imm:$size), (i32 0))]>;
2004 def DeclareScalarRegInst : NVPTXInst<(outs), (ins i32imm:$a, i32imm:$size),
2005 ".reg .b$size param$a;",
2006 [(DeclareScalarParam (i32 imm:$a), (i32 imm:$size), (i32 1))]>;
2008 class MoveParamInst<NVPTXRegClass regclass, string asmstr> :
2009 NVPTXInst<(outs regclass:$dst), (ins regclass:$src),
2010 !strconcat(!strconcat("mov", asmstr), "\t$dst, $src;"),
2011 [(set regclass:$dst, (MoveParam regclass:$src))]>;
2013 def MoveParamI64 : MoveParamInst<Int64Regs, ".b64">;
2014 def MoveParamI32 : MoveParamInst<Int32Regs, ".b32">;
2015 def MoveParamI16 : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
2016 "cvt.u16.u32\t$dst, $src;",
2017 [(set Int16Regs:$dst, (MoveParam Int16Regs:$src))]>;
2018 def MoveParamF64 : MoveParamInst<Float64Regs, ".f64">;
2019 def MoveParamF32 : MoveParamInst<Float32Regs, ".f32">;
2021 class PseudoUseParamInst<NVPTXRegClass regclass> :
2022 NVPTXInst<(outs), (ins regclass:$src),
2023 "// Pseudo use of $src",
2024 [(PseudoUseParam regclass:$src)]>;
2026 def PseudoUseParamI64 : PseudoUseParamInst<Int64Regs>;
2027 def PseudoUseParamI32 : PseudoUseParamInst<Int32Regs>;
2028 def PseudoUseParamI16 : PseudoUseParamInst<Int16Regs>;
2029 def PseudoUseParamF64 : PseudoUseParamInst<Float64Regs>;
2030 def PseudoUseParamF32 : PseudoUseParamInst<Float32Regs>;
2034 // Load / Store Handling
2036 multiclass LD<NVPTXRegClass regclass> {
2037 def _avar : NVPTXInst<(outs regclass:$dst),
2038 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2039 i32imm:$fromWidth, imem:$addr),
2040 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2041 "$fromWidth \t$dst, [$addr];"), []>;
2042 def _areg : NVPTXInst<(outs regclass:$dst),
2043 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2044 i32imm:$fromWidth, Int32Regs:$addr),
2045 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2046 "$fromWidth \t$dst, [$addr];"), []>;
2047 def _areg_64 : NVPTXInst<(outs regclass:$dst),
2048 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2049 i32imm:$fromWidth, Int64Regs:$addr),
2050 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$fromWidth",
2051 " \t$dst, [$addr];"), []>;
2052 def _ari : NVPTXInst<(outs regclass:$dst),
2053 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2054 i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
2055 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2056 "$fromWidth \t$dst, [$addr+$offset];"), []>;
2057 def _ari_64 : NVPTXInst<(outs regclass:$dst),
2058 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2059 i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
2060 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$fromWidth",
2061 " \t$dst, [$addr+$offset];"), []>;
2062 def _asi : NVPTXInst<(outs regclass:$dst),
2063 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2064 i32imm:$fromWidth, imem:$addr, i32imm:$offset),
2065 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2066 "$fromWidth \t$dst, [$addr+$offset];"), []>;
2069 let mayLoad=1, neverHasSideEffects=1 in {
2070 defm LD_i8 : LD<Int16Regs>;
2071 defm LD_i16 : LD<Int16Regs>;
2072 defm LD_i32 : LD<Int32Regs>;
2073 defm LD_i64 : LD<Int64Regs>;
2074 defm LD_f32 : LD<Float32Regs>;
2075 defm LD_f64 : LD<Float64Regs>;
2078 multiclass ST<NVPTXRegClass regclass> {
2079 def _avar : NVPTXInst<(outs),
2080 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2081 LdStCode:$Sign, i32imm:$toWidth, imem:$addr),
2082 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
2083 " \t[$addr], $src;"), []>;
2084 def _areg : NVPTXInst<(outs),
2085 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2086 LdStCode:$Sign, i32imm:$toWidth, Int32Regs:$addr),
2087 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
2088 " \t[$addr], $src;"), []>;
2089 def _areg_64 : NVPTXInst<(outs),
2090 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2091 LdStCode:$Sign, i32imm:$toWidth, Int64Regs:$addr),
2092 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth ",
2093 "\t[$addr], $src;"), []>;
2094 def _ari : NVPTXInst<(outs),
2095 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2096 LdStCode:$Sign, i32imm:$toWidth, Int32Regs:$addr, i32imm:$offset),
2097 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
2098 " \t[$addr+$offset], $src;"), []>;
2099 def _ari_64 : NVPTXInst<(outs),
2100 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2101 LdStCode:$Sign, i32imm:$toWidth, Int64Regs:$addr, i32imm:$offset),
2102 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth ",
2103 "\t[$addr+$offset], $src;"), []>;
2104 def _asi : NVPTXInst<(outs),
2105 (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
2106 LdStCode:$Sign, i32imm:$toWidth, imem:$addr, i32imm:$offset),
2107 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
2108 " \t[$addr+$offset], $src;"), []>;
2111 let mayStore=1, neverHasSideEffects=1 in {
2112 defm ST_i8 : ST<Int16Regs>;
2113 defm ST_i16 : ST<Int16Regs>;
2114 defm ST_i32 : ST<Int32Regs>;
2115 defm ST_i64 : ST<Int64Regs>;
2116 defm ST_f32 : ST<Float32Regs>;
2117 defm ST_f64 : ST<Float64Regs>;
2120 // The following is used only in and after vector elementizations.
2121 // Vector elementization happens at the machine instruction level, so the
2122 // following instruction
2123 // never appears in the DAG.
2124 multiclass LD_VEC<NVPTXRegClass regclass> {
2125 def _v2_avar : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2126 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2127 i32imm:$fromWidth, imem:$addr),
2128 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2129 "$fromWidth \t{{$dst1, $dst2}}, [$addr];"), []>;
2130 def _v2_areg : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2131 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2132 i32imm:$fromWidth, Int32Regs:$addr),
2133 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2134 "$fromWidth \t{{$dst1, $dst2}}, [$addr];"), []>;
2135 def _v2_areg_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2136 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2137 i32imm:$fromWidth, Int64Regs:$addr),
2138 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2139 "$fromWidth \t{{$dst1, $dst2}}, [$addr];"), []>;
2140 def _v2_ari : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2141 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2142 i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
2143 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2144 "$fromWidth \t{{$dst1, $dst2}}, [$addr+$offset];"), []>;
2145 def _v2_ari_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2146 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2147 i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
2148 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2149 "$fromWidth \t{{$dst1, $dst2}}, [$addr+$offset];"), []>;
2150 def _v2_asi : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
2151 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2152 i32imm:$fromWidth, imem:$addr, i32imm:$offset),
2153 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2154 "$fromWidth \t{{$dst1, $dst2}}, [$addr+$offset];"), []>;
2155 def _v4_avar : NVPTXInst<(outs regclass:$dst1, regclass:$dst2,
2156 regclass:$dst3, regclass:$dst4),
2157 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2158 i32imm:$fromWidth, imem:$addr),
2159 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2160 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr];"), []>;
2161 def _v4_areg : NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
2163 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2164 i32imm:$fromWidth, Int32Regs:$addr),
2165 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2166 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr];"), []>;
2167 def _v4_areg_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2,
2168 regclass:$dst3, regclass:$dst4),
2169 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2170 i32imm:$fromWidth, Int64Regs:$addr),
2171 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2172 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr];"), []>;
2173 def _v4_ari : NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
2175 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2176 i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
2177 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2178 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr+$offset];"),
2180 def _v4_ari_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2,
2181 regclass:$dst3, regclass:$dst4),
2182 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2183 i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
2184 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2185 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr+$offset];"),
2187 def _v4_asi : NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
2189 (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2190 i32imm:$fromWidth, imem:$addr, i32imm:$offset),
2191 !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2192 "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr+$offset];"),
2195 let mayLoad=1, neverHasSideEffects=1 in {
2196 defm LDV_i8 : LD_VEC<Int16Regs>;
2197 defm LDV_i16 : LD_VEC<Int16Regs>;
2198 defm LDV_i32 : LD_VEC<Int32Regs>;
2199 defm LDV_i64 : LD_VEC<Int64Regs>;
2200 defm LDV_f32 : LD_VEC<Float32Regs>;
2201 defm LDV_f64 : LD_VEC<Float64Regs>;
2204 multiclass ST_VEC<NVPTXRegClass regclass> {
2205 def _v2_avar : NVPTXInst<(outs),
2206 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2207 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, imem:$addr),
2208 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2209 "$fromWidth \t[$addr], {{$src1, $src2}};"), []>;
2210 def _v2_areg : NVPTXInst<(outs),
2211 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2212 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int32Regs:$addr),
2213 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2214 "$fromWidth \t[$addr], {{$src1, $src2}};"), []>;
2215 def _v2_areg_64 : NVPTXInst<(outs),
2216 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2217 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int64Regs:$addr),
2218 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2219 "$fromWidth \t[$addr], {{$src1, $src2}};"), []>;
2220 def _v2_ari : NVPTXInst<(outs),
2221 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2222 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int32Regs:$addr,
2224 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2225 "$fromWidth \t[$addr+$offset], {{$src1, $src2}};"), []>;
2226 def _v2_ari_64 : NVPTXInst<(outs),
2227 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2228 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int64Regs:$addr,
2230 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2231 "$fromWidth \t[$addr+$offset], {{$src1, $src2}};"), []>;
2232 def _v2_asi : NVPTXInst<(outs),
2233 (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
2234 LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, imem:$addr,
2236 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2237 "$fromWidth \t[$addr+$offset], {{$src1, $src2}};"), []>;
2238 def _v4_avar : NVPTXInst<(outs),
2239 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2240 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2241 i32imm:$fromWidth, imem:$addr),
2242 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2243 "$fromWidth \t[$addr], {{$src1, $src2, $src3, $src4}};"), []>;
2244 def _v4_areg : NVPTXInst<(outs),
2245 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2246 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2247 i32imm:$fromWidth, Int32Regs:$addr),
2248 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2249 "$fromWidth \t[$addr], {{$src1, $src2, $src3, $src4}};"), []>;
2250 def _v4_areg_64 : NVPTXInst<(outs),
2251 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2252 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2253 i32imm:$fromWidth, Int64Regs:$addr),
2254 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2255 "$fromWidth \t[$addr], {{$src1, $src2, $src3, $src4}};"), []>;
2256 def _v4_ari : NVPTXInst<(outs),
2257 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2258 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2259 i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
2260 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2261 "$fromWidth \t[$addr+$offset], {{$src1, $src2, $src3, $src4}};"),
2263 def _v4_ari_64 : NVPTXInst<(outs),
2264 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2265 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2266 i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
2267 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2268 "$fromWidth \t[$addr+$offset], {{$src1, $src2, $src3, $src4}};"),
2270 def _v4_asi : NVPTXInst<(outs),
2271 (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
2272 LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
2273 i32imm:$fromWidth, imem:$addr, i32imm:$offset),
2274 !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
2275 "$fromWidth \t[$addr+$offset], {{$src1, $src2, $src3, $src4}};"),
2278 let mayStore=1, neverHasSideEffects=1 in {
2279 defm STV_i8 : ST_VEC<Int16Regs>;
2280 defm STV_i16 : ST_VEC<Int16Regs>;
2281 defm STV_i32 : ST_VEC<Int32Regs>;
2282 defm STV_i64 : ST_VEC<Int64Regs>;
2283 defm STV_f32 : ST_VEC<Float32Regs>;
2284 defm STV_f64 : ST_VEC<Float64Regs>;
2288 //---- Conversion ----
2290 class F_BITCONVERT<string SzStr, NVPTXRegClass regclassIn,
2291 NVPTXRegClass regclassOut> :
2292 NVPTXInst<(outs regclassOut:$d), (ins regclassIn:$a),
2293 !strconcat("mov.b", !strconcat(SzStr, " \t $d, $a;")),
2294 [(set regclassOut:$d, (bitconvert regclassIn:$a))]>;
2296 def BITCONVERT_32_I2F : F_BITCONVERT<"32", Int32Regs, Float32Regs>;
2297 def BITCONVERT_32_F2I : F_BITCONVERT<"32", Float32Regs, Int32Regs>;
2298 def BITCONVERT_64_I2F : F_BITCONVERT<"64", Int64Regs, Float64Regs>;
2299 def BITCONVERT_64_F2I : F_BITCONVERT<"64", Float64Regs, Int64Regs>;
2301 // NOTE: pred->fp are currently sub-optimal due to an issue in TableGen where
2302 // we cannot specify floating-point literals in isel patterns. Therefore, we
2303 // use an integer selp to select either 1 or 0 and then cvt to floating-point.
2306 def : Pat<(f32 (sint_to_fp Int1Regs:$a)),
2307 (CVT_f32_s32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
2308 def : Pat<(f32 (sint_to_fp Int16Regs:$a)),
2309 (CVT_f32_s16 Int16Regs:$a, CvtRN)>;
2310 def : Pat<(f32 (sint_to_fp Int32Regs:$a)),
2311 (CVT_f32_s32 Int32Regs:$a, CvtRN)>;
2312 def : Pat<(f32 (sint_to_fp Int64Regs:$a)),
2313 (CVT_f32_s64 Int64Regs:$a, CvtRN)>;
2316 def : Pat<(f32 (uint_to_fp Int1Regs:$a)),
2317 (CVT_f32_u32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
2318 def : Pat<(f32 (uint_to_fp Int16Regs:$a)),
2319 (CVT_f32_u16 Int16Regs:$a, CvtRN)>;
2320 def : Pat<(f32 (uint_to_fp Int32Regs:$a)),
2321 (CVT_f32_u32 Int32Regs:$a, CvtRN)>;
2322 def : Pat<(f32 (uint_to_fp Int64Regs:$a)),
2323 (CVT_f32_u64 Int64Regs:$a, CvtRN)>;
2326 def : Pat<(f64 (sint_to_fp Int1Regs:$a)),
2327 (CVT_f64_s32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
2328 def : Pat<(f64 (sint_to_fp Int16Regs:$a)),
2329 (CVT_f64_s16 Int16Regs:$a, CvtRN)>;
2330 def : Pat<(f64 (sint_to_fp Int32Regs:$a)),
2331 (CVT_f64_s32 Int32Regs:$a, CvtRN)>;
2332 def : Pat<(f64 (sint_to_fp Int64Regs:$a)),
2333 (CVT_f64_s64 Int64Regs:$a, CvtRN)>;
2336 def : Pat<(f64 (uint_to_fp Int1Regs:$a)),
2337 (CVT_f64_u32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
2338 def : Pat<(f64 (uint_to_fp Int16Regs:$a)),
2339 (CVT_f64_u16 Int16Regs:$a, CvtRN)>;
2340 def : Pat<(f64 (uint_to_fp Int32Regs:$a)),
2341 (CVT_f64_u32 Int32Regs:$a, CvtRN)>;
2342 def : Pat<(f64 (uint_to_fp Int64Regs:$a)),
2343 (CVT_f64_u64 Int64Regs:$a, CvtRN)>;
2347 def : Pat<(i1 (fp_to_sint Float32Regs:$a)),
2348 (SETP_b32ri (BITCONVERT_32_F2I Float32Regs:$a), 0, CmpEQ)>;
2349 def : Pat<(i16 (fp_to_sint Float32Regs:$a)),
2350 (CVT_s16_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2351 def : Pat<(i16 (fp_to_sint Float32Regs:$a)),
2352 (CVT_s16_f32 Float32Regs:$a, CvtRZI)>;
2353 def : Pat<(i32 (fp_to_sint Float32Regs:$a)),
2354 (CVT_s32_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2355 def : Pat<(i32 (fp_to_sint Float32Regs:$a)),
2356 (CVT_s32_f32 Float32Regs:$a, CvtRZI)>;
2357 def : Pat<(i64 (fp_to_sint Float32Regs:$a)),
2358 (CVT_s64_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2359 def : Pat<(i64 (fp_to_sint Float32Regs:$a)),
2360 (CVT_s64_f32 Float32Regs:$a, CvtRZI)>;
2363 def : Pat<(i1 (fp_to_uint Float32Regs:$a)),
2364 (SETP_b32ri (BITCONVERT_32_F2I Float32Regs:$a), 0, CmpEQ)>;
2365 def : Pat<(i16 (fp_to_uint Float32Regs:$a)),
2366 (CVT_u16_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2367 def : Pat<(i16 (fp_to_uint Float32Regs:$a)),
2368 (CVT_u16_f32 Float32Regs:$a, CvtRZI)>;
2369 def : Pat<(i32 (fp_to_uint Float32Regs:$a)),
2370 (CVT_u32_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2371 def : Pat<(i32 (fp_to_uint Float32Regs:$a)),
2372 (CVT_u32_f32 Float32Regs:$a, CvtRZI)>;
2373 def : Pat<(i64 (fp_to_uint Float32Regs:$a)),
2374 (CVT_u64_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
2375 def : Pat<(i64 (fp_to_uint Float32Regs:$a)),
2376 (CVT_u64_f32 Float32Regs:$a, CvtRZI)>;
2379 def : Pat<(i1 (fp_to_sint Float64Regs:$a)),
2380 (SETP_b64ri (BITCONVERT_64_F2I Float64Regs:$a), 0, CmpEQ)>;
2381 def : Pat<(i16 (fp_to_sint Float64Regs:$a)),
2382 (CVT_s16_f64 Float64Regs:$a, CvtRZI)>;
2383 def : Pat<(i32 (fp_to_sint Float64Regs:$a)),
2384 (CVT_s32_f64 Float64Regs:$a, CvtRZI)>;
2385 def : Pat<(i64 (fp_to_sint Float64Regs:$a)),
2386 (CVT_s64_f64 Float64Regs:$a, CvtRZI)>;
2389 def : Pat<(i1 (fp_to_uint Float64Regs:$a)),
2390 (SETP_b64ri (BITCONVERT_64_F2I Float64Regs:$a), 0, CmpEQ)>;
2391 def : Pat<(i16 (fp_to_uint Float64Regs:$a)),
2392 (CVT_u16_f64 Float64Regs:$a, CvtRZI)>;
2393 def : Pat<(i32 (fp_to_uint Float64Regs:$a)),
2394 (CVT_u32_f64 Float64Regs:$a, CvtRZI)>;
2395 def : Pat<(i64 (fp_to_uint Float64Regs:$a)),
2396 (CVT_u64_f64 Float64Regs:$a, CvtRZI)>;
2399 def : Pat<(i16 (sext Int1Regs:$a)),
2400 (SELP_s16ii -1, 0, Int1Regs:$a)>;
2401 def : Pat<(i32 (sext Int1Regs:$a)),
2402 (SELP_s32ii -1, 0, Int1Regs:$a)>;
2403 def : Pat<(i64 (sext Int1Regs:$a)),
2404 (SELP_s64ii -1, 0, Int1Regs:$a)>;
2407 def : Pat<(i16 (zext Int1Regs:$a)),
2408 (SELP_u16ii 1, 0, Int1Regs:$a)>;
2409 def : Pat<(i32 (zext Int1Regs:$a)),
2410 (SELP_u32ii 1, 0, Int1Regs:$a)>;
2411 def : Pat<(i64 (zext Int1Regs:$a)),
2412 (SELP_u64ii 1, 0, Int1Regs:$a)>;
2415 def : Pat<(i16 (anyext Int1Regs:$a)),
2416 (SELP_u16ii -1, 0, Int1Regs:$a)>;
2417 def : Pat<(i32 (anyext Int1Regs:$a)),
2418 (SELP_u32ii -1, 0, Int1Regs:$a)>;
2419 def : Pat<(i64 (anyext Int1Regs:$a)),
2420 (SELP_u64ii -1, 0, Int1Regs:$a)>;
2423 def : Pat<(i32 (sext Int16Regs:$a)),
2424 (CVT_s32_s16 Int16Regs:$a, CvtNONE)>;
2425 def : Pat<(i64 (sext Int16Regs:$a)),
2426 (CVT_s64_s16 Int16Regs:$a, CvtNONE)>;
2429 def : Pat<(i32 (zext Int16Regs:$a)),
2430 (CVT_u32_u16 Int16Regs:$a, CvtNONE)>;
2431 def : Pat<(i64 (zext Int16Regs:$a)),
2432 (CVT_u64_u16 Int16Regs:$a, CvtNONE)>;
2435 def : Pat<(i32 (anyext Int16Regs:$a)),
2436 (CVT_u32_u16 Int16Regs:$a, CvtNONE)>;
2437 def : Pat<(i64 (anyext Int16Regs:$a)),
2438 (CVT_u64_u16 Int16Regs:$a, CvtNONE)>;
2441 def : Pat<(i64 (sext Int32Regs:$a)),
2442 (CVT_s64_s32 Int32Regs:$a, CvtNONE)>;
2445 def : Pat<(i64 (zext Int32Regs:$a)),
2446 (CVT_u64_u32 Int32Regs:$a, CvtNONE)>;
2449 def : Pat<(i64 (anyext Int32Regs:$a)),
2450 (CVT_u64_u32 Int32Regs:$a, CvtNONE)>;
2454 def : Pat<(i32 (trunc Int64Regs:$a)),
2455 (CVT_u32_u64 Int64Regs:$a, CvtNONE)>;
2456 def : Pat<(i16 (trunc Int64Regs:$a)),
2457 (CVT_u16_u64 Int64Regs:$a, CvtNONE)>;
2458 def : Pat<(i1 (trunc Int64Regs:$a)),
2459 (SETP_b64ri (ANDb64ri Int64Regs:$a, 1), 1, CmpEQ)>;
2462 def : Pat<(i16 (trunc Int32Regs:$a)),
2463 (CVT_u16_u32 Int32Regs:$a, CvtNONE)>;
2464 def : Pat<(i1 (trunc Int32Regs:$a)),
2465 (SETP_b32ri (ANDb32ri Int32Regs:$a, 1), 1, CmpEQ)>;
2468 def : Pat<(i1 (trunc Int16Regs:$a)),
2469 (SETP_b16ri (ANDb16ri Int16Regs:$a, 1), 1, CmpEQ)>;
2472 def : Pat<(sext_inreg Int16Regs:$a, i8), (CVT_INREG_s16_s8 Int16Regs:$a)>;
2473 def : Pat<(sext_inreg Int32Regs:$a, i8), (CVT_INREG_s32_s8 Int32Regs:$a)>;
2474 def : Pat<(sext_inreg Int32Regs:$a, i16), (CVT_INREG_s32_s16 Int32Regs:$a)>;
2475 def : Pat<(sext_inreg Int64Regs:$a, i8), (CVT_INREG_s64_s8 Int64Regs:$a)>;
2476 def : Pat<(sext_inreg Int64Regs:$a, i16), (CVT_INREG_s64_s16 Int64Regs:$a)>;
2477 def : Pat<(sext_inreg Int64Regs:$a, i32), (CVT_INREG_s64_s32 Int64Regs:$a)>;
2480 // Select instructions with 32-bit predicates
2481 def : Pat<(select Int32Regs:$pred, Int16Regs:$a, Int16Regs:$b),
2482 (SELP_b16rr Int16Regs:$a, Int16Regs:$b,
2483 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2484 def : Pat<(select Int32Regs:$pred, Int32Regs:$a, Int32Regs:$b),
2485 (SELP_b32rr Int32Regs:$a, Int32Regs:$b,
2486 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2487 def : Pat<(select Int32Regs:$pred, Int64Regs:$a, Int64Regs:$b),
2488 (SELP_b64rr Int64Regs:$a, Int64Regs:$b,
2489 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2490 def : Pat<(select Int32Regs:$pred, Float32Regs:$a, Float32Regs:$b),
2491 (SELP_f32rr Float32Regs:$a, Float32Regs:$b,
2492 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2493 def : Pat<(select Int32Regs:$pred, Float64Regs:$a, Float64Regs:$b),
2494 (SELP_f64rr Float64Regs:$a, Float64Regs:$b,
2495 (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
2498 // pack a set of smaller int registers to a larger int register
2499 def V4I16toI64 : NVPTXInst<(outs Int64Regs:$d),
2500 (ins Int16Regs:$s1, Int16Regs:$s2,
2501 Int16Regs:$s3, Int16Regs:$s4),
2502 "mov.b64\t$d, {{$s1, $s2, $s3, $s4}};",
2504 def V2I16toI32 : NVPTXInst<(outs Int32Regs:$d),
2505 (ins Int16Regs:$s1, Int16Regs:$s2),
2506 "mov.b32\t$d, {{$s1, $s2}};",
2508 def V2I32toI64 : NVPTXInst<(outs Int64Regs:$d),
2509 (ins Int32Regs:$s1, Int32Regs:$s2),
2510 "mov.b64\t$d, {{$s1, $s2}};",
2512 def V2F32toF64 : NVPTXInst<(outs Float64Regs:$d),
2513 (ins Float32Regs:$s1, Float32Regs:$s2),
2514 "mov.b64\t$d, {{$s1, $s2}};",
2517 // unpack a larger int register to a set of smaller int registers
2518 def I64toV4I16 : NVPTXInst<(outs Int16Regs:$d1, Int16Regs:$d2,
2519 Int16Regs:$d3, Int16Regs:$d4),
2521 "mov.b64\t{{$d1, $d2, $d3, $d4}}, $s;",
2523 def I32toV2I16 : NVPTXInst<(outs Int16Regs:$d1, Int16Regs:$d2),
2525 "mov.b32\t{{$d1, $d2}}, $s;",
2527 def I64toV2I32 : NVPTXInst<(outs Int32Regs:$d1, Int32Regs:$d2),
2529 "mov.b64\t{{$d1, $d2}}, $s;",
2531 def F64toV2F32 : NVPTXInst<(outs Float32Regs:$d1, Float32Regs:$d2),
2532 (ins Float64Regs:$s),
2533 "mov.b64\t{{$d1, $d2}}, $s;",
2536 // Count leading zeros
2537 def CLZr32 : NVPTXInst<(outs Int32Regs:$d), (ins Int32Regs:$a),
2540 def CLZr64 : NVPTXInst<(outs Int32Regs:$d), (ins Int64Regs:$a),
2544 // 32-bit has a direct PTX instruction
2545 def : Pat<(ctlz Int32Regs:$a),
2546 (CLZr32 Int32Regs:$a)>;
2547 def : Pat<(ctlz_zero_undef Int32Regs:$a),
2548 (CLZr32 Int32Regs:$a)>;
2550 // For 64-bit, the result in PTX is actually 32-bit so we zero-extend
2551 // to 64-bit to match the LLVM semantics
2552 def : Pat<(ctlz Int64Regs:$a),
2553 (CVT_u64_u32 (CLZr64 Int64Regs:$a), CvtNONE)>;
2554 def : Pat<(ctlz_zero_undef Int64Regs:$a),
2555 (CVT_u64_u32 (CLZr64 Int64Regs:$a), CvtNONE)>;
2557 // For 16-bit, we zero-extend to 32-bit, then trunc the result back
2558 // to 16-bits (ctlz of a 16-bit value is guaranteed to require less
2559 // than 16 bits to store). We also need to subtract 16 because the
2560 // high-order 16 zeros were counted.
2561 def : Pat<(ctlz Int16Regs:$a),
2562 (SUBi16ri (CVT_u16_u32 (CLZr32
2563 (CVT_u32_u16 Int16Regs:$a, CvtNONE)),
2565 def : Pat<(ctlz_zero_undef Int16Regs:$a),
2566 (SUBi16ri (CVT_u16_u32 (CLZr32
2567 (CVT_u32_u16 Int16Regs:$a, CvtNONE)),
2571 def POPCr32 : NVPTXInst<(outs Int32Regs:$d), (ins Int32Regs:$a),
2572 "popc.b32\t$d, $a;",
2574 def POPCr64 : NVPTXInst<(outs Int32Regs:$d), (ins Int64Regs:$a),
2575 "popc.b64\t$d, $a;",
2578 // 32-bit has a direct PTX instruction
2579 def : Pat<(ctpop Int32Regs:$a),
2580 (POPCr32 Int32Regs:$a)>;
2582 // For 64-bit, the result in PTX is actually 32-bit so we zero-extend
2583 // to 64-bit to match the LLVM semantics
2584 def : Pat<(ctpop Int64Regs:$a),
2585 (CVT_u64_u32 (POPCr64 Int64Regs:$a), CvtNONE)>;
2587 // For 16-bit, we zero-extend to 32-bit, then trunc the result back
2588 // to 16-bits (ctpop of a 16-bit value is guaranteed to require less
2589 // than 16 bits to store)
2590 def : Pat<(ctpop Int16Regs:$a),
2591 (CVT_u16_u32 (POPCr32 (CVT_u32_u16 Int16Regs:$a, CvtNONE)),
2594 // fround f64 -> f32
2595 def : Pat<(f32 (fround Float64Regs:$a)),
2596 (CVT_f32_f64 Float64Regs:$a, CvtRN_FTZ)>, Requires<[doF32FTZ]>;
2597 def : Pat<(f32 (fround Float64Regs:$a)),
2598 (CVT_f32_f64 Float64Regs:$a, CvtRN)>;
2600 // fextend f32 -> f64
2601 def : Pat<(f64 (fextend Float32Regs:$a)),
2602 (CVT_f64_f32 Float32Regs:$a, CvtNONE_FTZ)>, Requires<[doF32FTZ]>;
2603 def : Pat<(f64 (fextend Float32Regs:$a)),
2604 (CVT_f64_f32 Float32Regs:$a, CvtNONE)>;
2606 def retflag : SDNode<"NVPTXISD::RET_FLAG", SDTNone,
2607 [SDNPHasChain, SDNPOptInGlue]>;
2609 //-----------------------------------
2611 //-----------------------------------
2613 let isTerminator=1 in {
2614 let isReturn=1, isBarrier=1 in
2615 def Return : NVPTXInst<(outs), (ins), "ret;", [(retflag)]>;
2618 def CBranch : NVPTXInst<(outs), (ins Int1Regs:$a, brtarget:$target),
2619 "@$a bra \t$target;",
2620 [(brcond Int1Regs:$a, bb:$target)]>;
2622 def CBranchOther : NVPTXInst<(outs), (ins Int1Regs:$a, brtarget:$target),
2623 "@!$a bra \t$target;",
2626 let isBranch=1, isBarrier=1 in
2627 def GOTO : NVPTXInst<(outs), (ins brtarget:$target),
2628 "bra.uni \t$target;",
2632 def : Pat<(brcond Int32Regs:$a, bb:$target),
2633 (CBranch (SETP_u32ri Int32Regs:$a, 0, CmpNE), bb:$target)>;
2635 // SelectionDAGBuilder::visitSWitchCase() will invert the condition of a
2636 // conditional branch if
2637 // the target block is the next block so that the code can fall through to the
2639 // The invertion is done by 'xor condition, 1', which will be translated to
2640 // (setne condition, -1).
2641 // Since ptx supports '@!pred bra target', we should use it.
2642 def : Pat<(brcond (i1 (setne Int1Regs:$a, -1)), bb:$target),
2643 (CBranchOther Int1Regs:$a, bb:$target)>;
2646 def SDT_NVPTXCallSeqStart : SDCallSeqStart<[ SDTCisVT<0, i32> ]>;
2647 def SDT_NVPTXCallSeqEnd : SDCallSeqEnd<[ SDTCisVT<0, i32>,
2648 SDTCisVT<1, i32> ]>;
2650 def callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_NVPTXCallSeqStart,
2651 [SDNPHasChain, SDNPOutGlue, SDNPSideEffect]>;
2652 def callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_NVPTXCallSeqEnd,
2653 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue,
2656 def SDT_NVPTXCall : SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
2657 def call : SDNode<"NVPTXISD::CALL", SDT_NVPTXCall,
2658 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
2659 def calltarget : Operand<i32>;
2661 def CALL : NVPTXInst<(outs), (ins calltarget:$dst),
2662 "call \t$dst, (1);", []>;
2665 def : Pat<(call tglobaladdr:$dst),
2666 (CALL tglobaladdr:$dst)>;
2667 def : Pat<(call texternalsym:$dst),
2668 (CALL texternalsym:$dst)>;
2670 // Pseudo instructions.
2671 class Pseudo<dag outs, dag ins, string asmstr, list<dag> pattern>
2672 : NVPTXInst<outs, ins, asmstr, pattern>;
2674 // @TODO: We use some tricks here to emit curly braces. Can we clean this up
2675 // a bit without TableGen modifications?
2676 def Callseq_Start : NVPTXInst<(outs), (ins i32imm:$amt),
2677 "// Callseq Start $amt\n\t{{\n\t.reg .b32 temp_param_reg;\n\t// <end>}}",
2678 [(callseq_start timm:$amt)]>;
2679 def Callseq_End : NVPTXInst<(outs), (ins i32imm:$amt1, i32imm:$amt2),
2680 "\n\t//{{\n\t}}// Callseq End $amt1",
2681 [(callseq_end timm:$amt1, timm:$amt2)]>;
2685 def trapinst : NVPTXInst<(outs), (ins),
2689 // Call prototype wrapper
2690 def SDTCallPrototype : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
2692 : SDNode<"NVPTXISD::CallPrototype", SDTCallPrototype,
2693 [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
2694 def ProtoIdent : Operand<i32> {
2695 let PrintMethod = "printProtoIdent";
2698 : NVPTXInst<(outs), (ins ProtoIdent:$ident),
2699 "$ident", [(CallPrototype (i32 texternalsym:$ident))]>;
2703 include "NVPTXIntrinsics.td"
2706 //-----------------------------------
2708 //-----------------------------------
2709 // BSWAP is currently expanded. The following is a more efficient
2710 // - for < sm_20, use vector scalar mov, as tesla support native 16-bit register
2711 // - for sm_20, use pmpt (use vector scalar mov to get the pack and
2712 // unpack). sm_20 supports native 32-bit register, but not native 16-bit