1 //===- NVPTXVector.td - NVPTX Vector Specific 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 //-----------------------------------
12 //-----------------------------------
15 // All vector instructions derive from NVPTXVecInst
18 class NVPTXVecInst<dag outs, dag ins, string asmstr, list<dag> pattern,
20 : NVPTXInst<outs, ins, asmstr, pattern> {
21 NVPTXInst scalarInst=sInst;
24 let isAsCheapAsAMove=1, VecInstType=isVecExtract.Value in {
26 def V2i16Extract : NVPTXVecInst<(outs Int16Regs:$dst),
27 (ins V2I16Regs:$src, i8imm:$c),
28 "mov.u16 \t$dst, $src${c:vecelem};",
29 [(set Int16Regs:$dst, (extractelt
30 (v2i16 V2I16Regs:$src), imm:$c))],
34 def V4i16Extract : NVPTXVecInst<(outs Int16Regs:$dst),
35 (ins V4I16Regs:$src, i8imm:$c),
36 "mov.u16 \t$dst, $src${c:vecelem};",
37 [(set Int16Regs:$dst, (extractelt
38 (v4i16 V4I16Regs:$src), imm:$c))],
42 def V2i8Extract : NVPTXVecInst<(outs Int8Regs:$dst),
43 (ins V2I8Regs:$src, i8imm:$c),
44 "mov.u16 \t$dst, $src${c:vecelem};",
45 [(set Int8Regs:$dst, (extractelt
46 (v2i8 V2I8Regs:$src), imm:$c))],
50 def V4i8Extract : NVPTXVecInst<(outs Int8Regs:$dst),
51 (ins V4I8Regs:$src, i8imm:$c),
52 "mov.u16 \t$dst, $src${c:vecelem};",
53 [(set Int8Regs:$dst, (extractelt
54 (v4i8 V4I8Regs:$src), imm:$c))],
58 def V2i32Extract : NVPTXVecInst<(outs Int32Regs:$dst),
59 (ins V2I32Regs:$src, i8imm:$c),
60 "mov.u32 \t$dst, $src${c:vecelem};",
61 [(set Int32Regs:$dst, (extractelt
62 (v2i32 V2I32Regs:$src), imm:$c))],
66 def V2f32Extract : NVPTXVecInst<(outs Float32Regs:$dst),
67 (ins V2F32Regs:$src, i8imm:$c),
68 "mov.f32 \t$dst, $src${c:vecelem};",
69 [(set Float32Regs:$dst, (extractelt
70 (v2f32 V2F32Regs:$src), imm:$c))],
74 def V2i64Extract : NVPTXVecInst<(outs Int64Regs:$dst),
75 (ins V2I64Regs:$src, i8imm:$c),
76 "mov.u64 \t$dst, $src${c:vecelem};",
77 [(set Int64Regs:$dst, (extractelt
78 (v2i64 V2I64Regs:$src), imm:$c))],
82 def V2f64Extract : NVPTXVecInst<(outs Float64Regs:$dst),
83 (ins V2F64Regs:$src, i8imm:$c),
84 "mov.f64 \t$dst, $src${c:vecelem};",
85 [(set Float64Regs:$dst, (extractelt
86 (v2f64 V2F64Regs:$src), imm:$c))],
90 def V4i32Extract : NVPTXVecInst<(outs Int32Regs:$dst),
91 (ins V4I32Regs:$src, i8imm:$c),
92 "mov.u32 \t$dst, $src${c:vecelem};",
93 [(set Int32Regs:$dst, (extractelt
94 (v4i32 V4I32Regs:$src), imm:$c))],
98 def V4f32Extract : NVPTXVecInst<(outs Float32Regs:$dst),
99 (ins V4F32Regs:$src, i8imm:$c),
100 "mov.f32 \t$dst, $src${c:vecelem};",
101 [(set Float32Regs:$dst, (extractelt
102 (v4f32 V4F32Regs:$src), imm:$c))],
106 let isAsCheapAsAMove=1, VecInstType=isVecInsert.Value in {
108 def V2i8Insert : NVPTXVecInst<(outs V2I8Regs:$dst),
109 (ins V2I8Regs:$src, Int8Regs:$val, i8imm:$c),
110 "mov.v2.u16 \t${dst:vecfull}, ${src:vecfull};"
111 "\n\tmov.u16 \t$dst${c:vecelem}, $val;",
113 (insertelt V2I8Regs:$src, Int8Regs:$val, imm:$c))], IMOV8rr>;
116 def V4i8Insert : NVPTXVecInst<(outs V4I8Regs:$dst),
117 (ins V4I8Regs:$src, Int8Regs:$val, i8imm:$c),
118 "mov.v4.u16 \t${dst:vecfull}, ${src:vecfull};"
119 "\n\tmov.u16 \t$dst${c:vecelem}, $val;",
121 (insertelt V4I8Regs:$src, Int8Regs:$val, imm:$c))], IMOV8rr>;
124 def V2i16Insert : NVPTXVecInst<(outs V2I16Regs:$dst),
125 (ins V2I16Regs:$src, Int16Regs:$val, i8imm:$c),
126 "mov.v2.u16 \t${dst:vecfull}, ${src:vecfull};"
127 "\n\tmov.u16 \t$dst${c:vecelem}, $val;",
128 [(set V2I16Regs:$dst,
129 (insertelt V2I16Regs:$src, Int16Regs:$val, imm:$c))],
133 def V4i16Insert : NVPTXVecInst<(outs V4I16Regs:$dst),
134 (ins V4I16Regs:$src, Int16Regs:$val, i8imm:$c),
135 "mov.v4.u16 \t${dst:vecfull}, ${src:vecfull};"
136 "\n\tmov.u16 \t$dst${c:vecelem}, $val;",
137 [(set V4I16Regs:$dst,
138 (insertelt V4I16Regs:$src, Int16Regs:$val, imm:$c))],
142 def V2i32Insert : NVPTXVecInst<(outs V2I32Regs:$dst),
143 (ins V2I32Regs:$src, Int32Regs:$val, i8imm:$c),
144 "mov.v2.u32 \t${dst:vecfull}, ${src:vecfull};"
145 "\n\tmov.u32 \t$dst${c:vecelem}, $val;",
146 [(set V2I32Regs:$dst,
147 (insertelt V2I32Regs:$src, Int32Regs:$val, imm:$c))],
151 def V2f32Insert : NVPTXVecInst<(outs V2F32Regs:$dst),
152 (ins V2F32Regs:$src, Float32Regs:$val, i8imm:$c),
153 "mov.v2.f32 \t${dst:vecfull}, ${src:vecfull};"
154 "\n\tmov.f32 \t$dst${c:vecelem}, $val;",
155 [(set V2F32Regs:$dst,
156 (insertelt V2F32Regs:$src, Float32Regs:$val, imm:$c))],
160 def V2i64Insert : NVPTXVecInst<(outs V2I64Regs:$dst),
161 (ins V2I64Regs:$src, Int64Regs:$val, i8imm:$c),
162 "mov.v2.u64 \t${dst:vecfull}, ${src:vecfull};"
163 "\n\tmov.u64 \t$dst${c:vecelem}, $val;",
164 [(set V2I64Regs:$dst,
165 (insertelt V2I64Regs:$src, Int64Regs:$val, imm:$c))],
169 def V2f64Insert : NVPTXVecInst<(outs V2F64Regs:$dst),
170 (ins V2F64Regs:$src, Float64Regs:$val, i8imm:$c),
171 "mov.v2.f64 \t${dst:vecfull}, ${src:vecfull};"
172 "\n\tmov.f64 \t$dst${c:vecelem}, $val;",
173 [(set V2F64Regs:$dst,
174 (insertelt V2F64Regs:$src, Float64Regs:$val, imm:$c))],
178 def V4i32Insert : NVPTXVecInst<(outs V4I32Regs:$dst),
179 (ins V4I32Regs:$src, Int32Regs:$val, i8imm:$c),
180 "mov.v4.u32 \t${dst:vecfull}, ${src:vecfull};"
181 "\n\tmov.u32 \t$dst${c:vecelem}, $val;",
182 [(set V4I32Regs:$dst,
183 (insertelt V4I32Regs:$src, Int32Regs:$val, imm:$c))],
187 def V4f32Insert : NVPTXVecInst<(outs V4F32Regs:$dst),
188 (ins V4F32Regs:$src, Float32Regs:$val, i8imm:$c),
189 "mov.v4.f32 \t${dst:vecfull}, ${src:vecfull};"
190 "\n\tmov.f32 \t$dst${c:vecelem}, $val;",
191 [(set V4F32Regs:$dst,
192 (insertelt V4F32Regs:$src, Float32Regs:$val, imm:$c))],
196 class BinOpAsmString<string c> {
200 class V4AsmStr<string opcode> : BinOpAsmString<
201 !strconcat(!strconcat(!strconcat(!strconcat(
202 !strconcat(!strconcat(!strconcat(
203 opcode, " \t${dst}_0, ${a}_0, ${b}_0;\n\t"),
204 opcode), " \t${dst}_1, ${a}_1, ${b}_1;\n\t"),
205 opcode), " \t${dst}_2, ${a}_2, ${b}_2;\n\t"),
206 opcode), " \t${dst}_3, ${a}_3, ${b}_3;")>;
208 class V2AsmStr<string opcode> : BinOpAsmString<
209 !strconcat(!strconcat(!strconcat(
210 opcode, " \t${dst}_0, ${a}_0, ${b}_0;\n\t"),
211 opcode), " \t${dst}_1, ${a}_1, ${b}_1;")>;
213 class V4MADStr<string opcode> : BinOpAsmString<
214 !strconcat(!strconcat(!strconcat(!strconcat(
215 !strconcat(!strconcat(!strconcat(
216 opcode, " \t${dst}_0, ${a}_0, ${b}_0, ${c}_0;\n\t"),
217 opcode), " \t${dst}_1, ${a}_1, ${b}_1, ${c}_1;\n\t"),
218 opcode), " \t${dst}_2, ${a}_2, ${b}_2, ${c}_2;\n\t"),
219 opcode), " \t${dst}_3, ${a}_3, ${b}_3, ${c}_3;")>;
221 class V2MADStr<string opcode> : BinOpAsmString<
222 !strconcat(!strconcat(!strconcat(
223 opcode, " \t${dst}_0, ${a}_0, ${b}_0, ${c}_0;\n\t"),
224 opcode), " \t${dst}_1, ${a}_1, ${b}_1, ${c}_1;")>;
226 class V4UnaryStr<string opcode> : BinOpAsmString<
227 !strconcat(!strconcat(!strconcat(!strconcat(
228 !strconcat(!strconcat(!strconcat(
229 opcode, " \t${dst}_0, ${a}_0;\n\t"),
230 opcode), " \t${dst}_1, ${a}_1;\n\t"),
231 opcode), " \t${dst}_2, ${a}_2;\n\t"),
232 opcode), " \t${dst}_3, ${a}_3;")>;
234 class V2UnaryStr<string opcode> : BinOpAsmString<
235 !strconcat(!strconcat(!strconcat(
236 opcode, " \t${dst}_0, ${a}_0;\n\t"),
237 opcode), " \t${dst}_1, ${a}_1;")>;
239 class VecBinaryOp<BinOpAsmString asmstr, SDNode OpNode, NVPTXRegClass regclass,
240 NVPTXInst sInst=NOP> :
241 NVPTXVecInst<(outs regclass:$dst), (ins regclass:$a, regclass:$b),
243 [(set regclass:$dst, (OpNode regclass:$a, regclass:$b))],
246 class VecShiftOp<BinOpAsmString asmstr, SDNode OpNode, NVPTXRegClass regclass1,
247 NVPTXRegClass regclass2, NVPTXInst sInst=NOP> :
248 NVPTXVecInst<(outs regclass1:$dst), (ins regclass1:$a, regclass2:$b),
250 [(set regclass1:$dst, (OpNode regclass1:$a, regclass2:$b))],
253 class VecUnaryOp<BinOpAsmString asmstr, PatFrag OpNode, NVPTXRegClass regclass,
254 NVPTXInst sInst=NOP> :
255 NVPTXVecInst<(outs regclass:$dst), (ins regclass:$a),
257 [(set regclass:$dst, (OpNode regclass:$a))], sInst>;
259 multiclass IntBinVOp<string asmstr, SDNode OpNode,
260 NVPTXInst i64op=NOP, NVPTXInst i32op=NOP, NVPTXInst
261 i16op=NOP, NVPTXInst i8op=NOP> {
262 def V2I64 : VecBinaryOp<V2AsmStr<!strconcat(asmstr, "64")>, OpNode, V2I64Regs,
264 def V4I32 : VecBinaryOp<V4AsmStr<!strconcat(asmstr, "32")>, OpNode, V4I32Regs,
266 def V2I32 : VecBinaryOp<V2AsmStr<!strconcat(asmstr, "32")>, OpNode, V2I32Regs,
268 def V4I16 : VecBinaryOp<V4AsmStr<!strconcat(asmstr, "16")>, OpNode, V4I16Regs,
270 def V2I16 : VecBinaryOp<V2AsmStr<!strconcat(asmstr, "16")>, OpNode, V2I16Regs,
272 def V4I8 : VecBinaryOp<V4AsmStr<!strconcat(asmstr, "16")>, OpNode, V4I8Regs,
274 def V2I8 : VecBinaryOp<V2AsmStr<!strconcat(asmstr, "16")>, OpNode, V2I8Regs,
278 multiclass FloatBinVOp<string asmstr, SDNode OpNode,
279 NVPTXInst f64=NOP, NVPTXInst f32=NOP,
280 NVPTXInst f32_ftz=NOP> {
281 def V2F64 : VecBinaryOp<V2AsmStr<!strconcat(asmstr, "f64")>, OpNode,
283 def V4F32_ftz : VecBinaryOp<V4AsmStr<!strconcat(asmstr, "ftz.f32")>, OpNode,
284 V4F32Regs, f32_ftz>, Requires<[doF32FTZ]>;
285 def V2F32_ftz : VecBinaryOp<V2AsmStr<!strconcat(asmstr, "ftz.f32")>, OpNode,
286 V2F32Regs, f32_ftz>, Requires<[doF32FTZ]>;
287 def V4F32 : VecBinaryOp<V4AsmStr<!strconcat(asmstr, "f32")>, OpNode,
289 def V2F32 : VecBinaryOp<V2AsmStr<!strconcat(asmstr, "f32")>, OpNode,
293 multiclass IntUnaryVOp<string asmstr, PatFrag OpNode,
294 NVPTXInst i64op=NOP, NVPTXInst i32op=NOP,
295 NVPTXInst i16op=NOP, NVPTXInst i8op=NOP> {
296 def V2I64 : VecUnaryOp<V2UnaryStr<!strconcat(asmstr, "64")>, OpNode,
298 def V4I32 : VecUnaryOp<V4UnaryStr<!strconcat(asmstr, "32")>, OpNode,
300 def V2I32 : VecUnaryOp<V2UnaryStr<!strconcat(asmstr, "32")>, OpNode,
302 def V4I16 : VecUnaryOp<V4UnaryStr<!strconcat(asmstr, "16")>, OpNode,
304 def V2I16 : VecUnaryOp<V2UnaryStr<!strconcat(asmstr, "16")>, OpNode,
306 def V4I8 : VecUnaryOp<V4UnaryStr<!strconcat(asmstr, "16")>, OpNode,
308 def V2I8 : VecUnaryOp<V2UnaryStr<!strconcat(asmstr, "16")>, OpNode,
313 // Integer Arithmetic
314 let VecInstType=isVecOther.Value in {
315 defm VAdd : IntBinVOp<"add.s", add, ADDi64rr, ADDi32rr, ADDi16rr, ADDi8rr>;
316 defm VSub : IntBinVOp<"sub.s", sub, SUBi64rr, SUBi32rr, SUBi16rr, SUBi8rr>;
318 def AddCCV4I32 : VecBinaryOp<V4AsmStr<"add.cc.s32">, addc, V4I32Regs,
320 def AddCCV2I32 : VecBinaryOp<V2AsmStr<"add.cc.s32">, addc, V2I32Regs,
322 def SubCCV4I32 : VecBinaryOp<V4AsmStr<"sub.cc.s32">, subc, V4I32Regs,
324 def SubCCV2I32 : VecBinaryOp<V2AsmStr<"sub.cc.s32">, subc, V2I32Regs,
326 def AddCCCV4I32 : VecBinaryOp<V4AsmStr<"addc.cc.s32">, adde, V4I32Regs,
328 def AddCCCV2I32 : VecBinaryOp<V2AsmStr<"addc.cc.s32">, adde, V2I32Regs,
330 def SubCCCV4I32 : VecBinaryOp<V4AsmStr<"subc.cc.s32">, sube, V4I32Regs,
332 def SubCCCV2I32 : VecBinaryOp<V2AsmStr<"subc.cc.s32">, sube, V2I32Regs,
335 def ShiftLV2I64 : VecShiftOp<V2AsmStr<"shl.b64">, shl, V2I64Regs, V2I32Regs,
337 def ShiftLV2I32 : VecShiftOp<V2AsmStr<"shl.b32">, shl, V2I32Regs, V2I32Regs,
339 def ShiftLV4I32 : VecShiftOp<V4AsmStr<"shl.b32">, shl, V4I32Regs, V4I32Regs,
341 def ShiftLV2I16 : VecShiftOp<V2AsmStr<"shl.b16">, shl, V2I16Regs, V2I32Regs,
343 def ShiftLV4I16 : VecShiftOp<V4AsmStr<"shl.b16">, shl, V4I16Regs, V4I32Regs,
345 def ShiftLV2I8 : VecShiftOp<V2AsmStr<"shl.b16">, shl, V2I8Regs, V2I32Regs,
347 def ShiftLV4I8 : VecShiftOp<V4AsmStr<"shl.b16">, shl, V4I8Regs, V4I32Regs,
351 // cvt to v*i32, helpers for shift
352 class CVTtoVeci32<NVPTXRegClass inclass, NVPTXRegClass outclass, string asmstr,
353 NVPTXInst sInst=NOP> :
354 NVPTXVecInst<(outs outclass:$d), (ins inclass:$s), asmstr, [], sInst>;
356 class VecCVTStrHelper<string op, string dest, string src> {
357 string s=!strconcat(op, !strconcat("\t",
358 !strconcat(dest, !strconcat(", ", !strconcat(src, ";")))));
361 class Vec2CVTStr<string op> {
362 string s=!strconcat(VecCVTStrHelper<op, "${d}_0", "${s}_0">.s,
363 !strconcat("\n\t", VecCVTStrHelper<op, "${d}_1", "${s}_1">.s));
366 class Vec4CVTStr<string op> {
367 string s=!strconcat(VecCVTStrHelper<op, "${d}_0", "${s}_0">.s,
369 !strconcat(VecCVTStrHelper<op, "${d}_1", "${s}_1">.s,
371 !strconcat(VecCVTStrHelper<op, "${d}_2", "${s}_2">.s,
372 !strconcat("\n\t", VecCVTStrHelper<op, "${d}_3", "${s}_3">.s))))));
375 let VecInstType=isVecOther.Value in {
376 def CVTv2i8tov2i32 : CVTtoVeci32<V2I8Regs, V2I32Regs,
377 Vec2CVTStr<"cvt.u32.u16">.s, Zint_extendext8to32>;
378 def CVTv2i16tov2i32 : CVTtoVeci32<V2I16Regs, V2I32Regs,
379 Vec2CVTStr<"cvt.u32.u16">.s, Zint_extendext16to32>;
380 def CVTv4i8tov4i32 : CVTtoVeci32<V4I8Regs, V4I32Regs,
381 Vec4CVTStr<"cvt.u32.u16">.s, Zint_extendext8to32>;
382 def CVTv4i16tov4i32 : CVTtoVeci32<V4I16Regs, V4I32Regs,
383 Vec4CVTStr<"cvt.u32.u16">.s, Zint_extendext16to32>;
384 def CVTv2i64tov2i32 : CVTtoVeci32<V2I64Regs, V2I32Regs,
385 Vec2CVTStr<"cvt.u32.u64">.s, TRUNC_64to32>;
388 def : Pat<(shl V2I16Regs:$src1, V2I16Regs:$src2),
389 (ShiftLV2I16 V2I16Regs:$src1, (CVTv2i16tov2i32 V2I16Regs:$src2))>;
390 def : Pat<(shl V2I8Regs:$src1, V2I8Regs:$src2),
391 (ShiftLV2I8 V2I8Regs:$src1, (CVTv2i8tov2i32 V2I8Regs:$src2))>;
392 def : Pat<(shl V2I64Regs:$src1, V2I64Regs:$src2),
393 (ShiftLV2I64 V2I64Regs:$src1, (CVTv2i64tov2i32 V2I64Regs:$src2))>;
395 def : Pat<(shl V4I16Regs:$src1, V4I16Regs:$src2),
396 (ShiftLV4I16 V4I16Regs:$src1, (CVTv4i16tov4i32 V4I16Regs:$src2))>;
397 def : Pat<(shl V4I8Regs:$src1, V4I8Regs:$src2),
398 (ShiftLV4I8 V4I8Regs:$src1, (CVTv4i8tov4i32 V4I8Regs:$src2))>;
400 let VecInstType=isVecOther.Value in {
401 def ShiftRAV2I64 : VecShiftOp<V2AsmStr<"shr.s64">, sra, V2I64Regs, V2I32Regs,
403 def ShiftRAV2I32 : VecShiftOp<V2AsmStr<"shr.s32">, sra, V2I32Regs, V2I32Regs,
405 def ShiftRAV4I32 : VecShiftOp<V4AsmStr<"shr.s32">, sra, V4I32Regs, V4I32Regs,
407 def ShiftRAV2I16 : VecShiftOp<V2AsmStr<"shr.s16">, sra, V2I16Regs, V2I32Regs,
409 def ShiftRAV4I16 : VecShiftOp<V4AsmStr<"shr.s16">, sra, V4I16Regs, V4I32Regs,
411 def ShiftRAV2I8 : VecShiftOp<V2AsmStr<"shr.s16">, sra, V2I8Regs, V2I32Regs,
413 def ShiftRAV4I8 : VecShiftOp<V4AsmStr<"shr.s16">, sra, V4I8Regs, V4I32Regs,
416 def ShiftRLV2I64 : VecShiftOp<V2AsmStr<"shr.u64">, srl, V2I64Regs, V2I32Regs,
418 def ShiftRLV2I32 : VecShiftOp<V2AsmStr<"shr.u32">, srl, V2I32Regs, V2I32Regs,
420 def ShiftRLV4I32 : VecShiftOp<V4AsmStr<"shr.u32">, srl, V4I32Regs, V4I32Regs,
422 def ShiftRLV2I16 : VecShiftOp<V2AsmStr<"shr.u16">, srl, V2I16Regs, V2I32Regs,
424 def ShiftRLV4I16 : VecShiftOp<V4AsmStr<"shr.u16">, srl, V4I16Regs, V4I32Regs,
426 def ShiftRLV2I8 : VecShiftOp<V2AsmStr<"shr.u16">, srl, V2I8Regs, V2I32Regs,
428 def ShiftRLV4I8 : VecShiftOp<V4AsmStr<"shr.u16">, srl, V4I8Regs, V4I32Regs,
431 defm VMult : IntBinVOp<"mul.lo.s", mul, MULTi64rr, MULTi32rr, MULTi16rr,
433 defm VMultHS : IntBinVOp<"mul.hi.s", mulhs, MULTHSi64rr, MULTHSi32rr,
436 defm VMultHU : IntBinVOp<"mul.hi.u", mulhu, MULTHUi64rr, MULTHUi32rr,
439 defm VSDiv : IntBinVOp<"div.s", sdiv, SDIVi64rr, SDIVi32rr, SDIVi16rr,
441 defm VUDiv : IntBinVOp<"div.u", udiv, UDIVi64rr, UDIVi32rr, UDIVi16rr,
443 defm VSRem : IntBinVOp<"rem.s", srem, SREMi64rr, SREMi32rr, SREMi16rr,
445 defm VURem : IntBinVOp<"rem.u", urem, UREMi64rr, UREMi32rr, UREMi16rr,
449 def : Pat<(sra V2I16Regs:$src1, V2I16Regs:$src2),
450 (ShiftRAV2I16 V2I16Regs:$src1, (CVTv2i16tov2i32 V2I16Regs:$src2))>;
451 def : Pat<(sra V2I8Regs:$src1, V2I8Regs:$src2),
452 (ShiftRAV2I8 V2I8Regs:$src1, (CVTv2i8tov2i32 V2I8Regs:$src2))>;
453 def : Pat<(sra V2I64Regs:$src1, V2I64Regs:$src2),
454 (ShiftRAV2I64 V2I64Regs:$src1, (CVTv2i64tov2i32 V2I64Regs:$src2))>;
456 def : Pat<(sra V4I16Regs:$src1, V4I16Regs:$src2),
457 (ShiftRAV4I16 V4I16Regs:$src1, (CVTv4i16tov4i32 V4I16Regs:$src2))>;
458 def : Pat<(sra V4I8Regs:$src1, V4I8Regs:$src2),
459 (ShiftRAV4I8 V4I8Regs:$src1, (CVTv4i8tov4i32 V4I8Regs:$src2))>;
461 def : Pat<(srl V2I16Regs:$src1, V2I16Regs:$src2),
462 (ShiftRLV2I16 V2I16Regs:$src1, (CVTv2i16tov2i32 V2I16Regs:$src2))>;
463 def : Pat<(srl V2I8Regs:$src1, V2I8Regs:$src2),
464 (ShiftRLV2I8 V2I8Regs:$src1, (CVTv2i8tov2i32 V2I8Regs:$src2))>;
465 def : Pat<(srl V2I64Regs:$src1, V2I64Regs:$src2),
466 (ShiftRLV2I64 V2I64Regs:$src1, (CVTv2i64tov2i32 V2I64Regs:$src2))>;
468 def : Pat<(srl V4I16Regs:$src1, V4I16Regs:$src2),
469 (ShiftRLV4I16 V4I16Regs:$src1, (CVTv4i16tov4i32 V4I16Regs:$src2))>;
470 def : Pat<(srl V4I8Regs:$src1, V4I8Regs:$src2),
471 (ShiftRLV4I8 V4I8Regs:$src1, (CVTv4i8tov4i32 V4I8Regs:$src2))>;
473 multiclass VMAD<string asmstr, NVPTXRegClass regclassv4,
474 NVPTXRegClass regclassv2,
475 SDNode an=add, SDNode mn=mul, NVPTXInst sop=NOP,
477 def V4 : NVPTXVecInst<(outs regclassv4:$dst),
478 (ins regclassv4:$a, regclassv4:$b, regclassv4:$c),
480 [(set regclassv4:$dst,
481 (an (mn regclassv4:$a, regclassv4:$b), regclassv4:$c))],
484 def V2 : NVPTXVecInst<(outs regclassv2:$dst),
485 (ins regclassv2:$a, regclassv2:$b, regclassv2:$c),
487 [(set regclassv2:$dst,
488 (an (mn regclassv2:$a, regclassv2:$b), regclassv2:$c))],
493 multiclass VMADV2Only<string asmstr, NVPTXRegClass regclass, NVPTXInst sop=NOP,
495 def V2 : NVPTXVecInst<(outs regclass:$dst),
496 (ins regclass:$a, regclass:$b, regclass:$c),
498 [(set regclass:$dst, (add
499 (mul regclass:$a, regclass:$b), regclass:$c))], sop>,
502 multiclass VFMADV2Only<string asmstr, NVPTXRegClass regclass, NVPTXInst sop=NOP,
504 def V2 : NVPTXVecInst<(outs regclass:$dst),
505 (ins regclass:$a, regclass:$b, regclass:$c),
507 [(set regclass:$dst, (fadd
508 (fmul regclass:$a, regclass:$b), regclass:$c))], sop>,
512 let VecInstType=isVecOther.Value in {
513 defm I8MAD : VMAD<"mad.lo.s16", V4I8Regs, V2I8Regs, add, mul, MAD8rrr, true>;
514 defm I16MAD : VMAD<"mad.lo.s16", V4I16Regs, V2I16Regs, add, mul, MAD16rrr,
516 defm I32MAD : VMAD<"mad.lo.s32", V4I32Regs, V2I32Regs, add, mul, MAD32rrr,
518 defm I64MAD : VMADV2Only<"mad.lo.s64", V2I64Regs, MAD64rrr, true>;
520 defm VNeg : IntUnaryVOp<"neg.s", ineg, INEG64, INEG32, INEG16, INEG8>;
522 defm VAddf : FloatBinVOp<"add.", fadd, FADDf64rr, FADDf32rr, FADDf32rr_ftz>;
523 defm VSubf : FloatBinVOp<"sub.", fsub, FSUBf64rr, FSUBf32rr, FSUBf32rr_ftz>;
524 defm VMulf : FloatBinVOp<"mul.", fmul, FMULf64rr, FMULf32rr, FMULf32rr_ftz>;
526 defm F32MAD_ftz : VMAD<"mad.ftz.f32", V4F32Regs, V2F32Regs, fadd, fmul,
527 FMAD32_ftzrrr, doFMADF32_ftz>;
528 defm F32FMA_ftz : VMAD<"fma.rn.ftz.f32", V4F32Regs, V2F32Regs, fadd, fmul,
529 FMA32_ftzrrr, doFMAF32_ftz>;
530 defm F32MAD : VMAD<"mad.f32", V4F32Regs, V2F32Regs, fadd, fmul, FMAD32rrr,
532 defm F32FMA : VMAD<"fma.rn.f32", V4F32Regs, V2F32Regs, fadd, fmul, FMA32rrr,
534 defm F64FMA : VFMADV2Only<"fma.rn.f64", V2F64Regs, FMA64rrr, doFMAF64>;
537 let VecInstType=isVecOther.Value in {
538 def V4F32Div_prec_ftz : VecBinaryOp<V4AsmStr<"div.rn.ftz.f32">, fdiv, V4F32Regs,
539 FDIV32rr_prec_ftz>, Requires<[doF32FTZ, reqPTX20]>;
540 def V2F32Div_prec_ftz : VecBinaryOp<V2AsmStr<"div.rn.ftz.f32">, fdiv, V2F32Regs,
541 FDIV32rr_prec_ftz>, Requires<[doF32FTZ, reqPTX20]>;
542 def V4F32Div_prec : VecBinaryOp<V4AsmStr<"div.rn.f32">, fdiv, V4F32Regs,
543 FDIV32rr_prec>, Requires<[reqPTX20]>;
544 def V2F32Div_prec : VecBinaryOp<V2AsmStr<"div.rn.f32">, fdiv, V2F32Regs,
545 FDIV32rr_prec>, Requires<[reqPTX20]>;
546 def V2F32Div_ftz : VecBinaryOp<V2AsmStr<"div.full.ftz.f32">, fdiv, V2F32Regs,
547 FDIV32rr_ftz>, Requires<[doF32FTZ]>;
548 def V4F32Div_ftz : VecBinaryOp<V4AsmStr<"div.full.ftz.f32">, fdiv, V4F32Regs,
549 FDIV32rr_ftz>, Requires<[doF32FTZ]>;
550 def V2F32Div : VecBinaryOp<V2AsmStr<"div.full.f32">, fdiv, V2F32Regs, FDIV32rr>;
551 def V4F32Div : VecBinaryOp<V4AsmStr<"div.full.f32">, fdiv, V4F32Regs, FDIV32rr>;
552 def V2F64Div : VecBinaryOp<V2AsmStr<"div.rn.f64">, fdiv, V2F64Regs, FDIV64rr>;
555 def fnegpat : PatFrag<(ops node:$in), (fneg node:$in)>;
557 let VecInstType=isVecOther.Value in {
558 def VNegv2f32_ftz : VecUnaryOp<V2UnaryStr<"neg.ftz.f32">, fnegpat, V2F32Regs,
559 FNEGf32_ftz>, Requires<[doF32FTZ]>;
560 def VNegv4f32_ftz : VecUnaryOp<V4UnaryStr<"neg.ftz.f32">, fnegpat, V4F32Regs,
561 FNEGf32_ftz>, Requires<[doF32FTZ]>;
562 def VNegv2f32 : VecUnaryOp<V2UnaryStr<"neg.f32">, fnegpat, V2F32Regs, FNEGf32>;
563 def VNegv4f32 : VecUnaryOp<V4UnaryStr<"neg.f32">, fnegpat, V4F32Regs, FNEGf32>;
564 def VNegv2f64 : VecUnaryOp<V2UnaryStr<"neg.f64">, fnegpat, V2F64Regs, FNEGf64>;
566 // Logical Arithmetic
567 defm VAnd : IntBinVOp<"and.b", and, ANDb64rr, ANDb32rr, ANDb16rr, ANDb8rr>;
568 defm VOr : IntBinVOp<"or.b", or, ORb64rr, ORb32rr, ORb16rr, ORb8rr>;
569 defm VXor : IntBinVOp<"xor.b", xor, XORb64rr, XORb32rr, XORb16rr, XORb8rr>;
571 defm VNot : IntUnaryVOp<"not.b", not, NOT64, NOT32, NOT16, NOT8>;
575 multiclass V2FPCONTRACT32_SUB_PAT<NVPTXInst Inst, Predicate Pred> {
576 def : Pat<(fsub V2F32Regs:$a, (fmul V2F32Regs:$b, V2F32Regs:$c)),
577 (Inst (VNegv2f32 V2F32Regs:$b), V2F32Regs:$c, V2F32Regs:$a)>,
580 def : Pat<(fsub (fmul V2F32Regs:$a, V2F32Regs:$b), V2F32Regs:$c),
581 (Inst V2F32Regs:$a, V2F32Regs:$b, (VNegv2f32 V2F32Regs:$c))>,
585 defm V2FMAF32ext_ftz : V2FPCONTRACT32_SUB_PAT<F32FMA_ftzV2, doFMAF32AGG_ftz>;
586 defm V2FMADF32ext_ftz : V2FPCONTRACT32_SUB_PAT<F32MAD_ftzV2, doFMADF32_ftz>;
587 defm V2FMAF32ext : V2FPCONTRACT32_SUB_PAT<F32FMAV2, doFMAF32AGG>;
588 defm V2FMADF32ext : V2FPCONTRACT32_SUB_PAT<F32MADV2, doFMADF32>;
590 multiclass V4FPCONTRACT32_SUB_PAT<NVPTXInst Inst, Predicate Pred> {
591 def : Pat<(fsub V4F32Regs:$a, (fmul V4F32Regs:$b, V4F32Regs:$c)),
592 (Inst (VNegv4f32 V4F32Regs:$b), V4F32Regs:$c, V4F32Regs:$a)>,
595 def : Pat<(fsub (fmul V4F32Regs:$a, V4F32Regs:$b), V4F32Regs:$c),
596 (Inst V4F32Regs:$a, V4F32Regs:$b, (VNegv4f32 V4F32Regs:$c))>,
600 defm V4FMAF32ext_ftz : V4FPCONTRACT32_SUB_PAT<F32FMA_ftzV4, doFMAF32AGG_ftz>;
601 defm V4FMADF32ext_ftz : V4FPCONTRACT32_SUB_PAT<F32MAD_ftzV4, doFMADF32_ftz>;
602 defm V4FMAF32ext : V4FPCONTRACT32_SUB_PAT<F32FMAV4, doFMAF32AGG>;
603 defm V4FMADF32ext : V4FPCONTRACT32_SUB_PAT<F32MADV4, doFMADF32>;
605 multiclass V2FPCONTRACT64_SUB_PAT<NVPTXInst Inst, Predicate Pred> {
606 def : Pat<(fsub V2F64Regs:$a, (fmul V2F64Regs:$b, V2F64Regs:$c)),
607 (Inst (VNegv2f64 V2F64Regs:$b), V2F64Regs:$c, V2F64Regs:$a)>,
610 def : Pat<(fsub (fmul V2F64Regs:$a, V2F64Regs:$b), V2F64Regs:$c),
611 (Inst V2F64Regs:$a, V2F64Regs:$b, (VNegv2f64 V2F64Regs:$c))>,
615 defm V2FMAF64ext : V2FPCONTRACT64_SUB_PAT<F64FMAV2, doFMAF64AGG>;
617 class VecModStr<string vecsize, string elem, string extra, string l="">
619 string t1 = !strconcat("${c", elem);
620 string t2 = !strconcat(t1, ":vecv");
621 string t3 = !strconcat(t2, vecsize);
622 string t4 = !strconcat(t3, extra);
623 string t5 = !strconcat(t4, l);
624 string s = !strconcat(t5, "}");
626 class ShuffleOneLine<string vecsize, string elem, string type>
628 string t1 = VecModStr<vecsize, elem, "comm", "1">.s;
629 string t2 = !strconcat(t1, "mov.");
630 string t3 = !strconcat(t2, type);
631 string t4 = !strconcat(t3, " \t${dst}_");
632 string t5 = !strconcat(t4, elem);
633 string t6 = !strconcat(t5, ", $src1");
634 string t7 = !strconcat(t6, VecModStr<vecsize, elem, "pos">.s);
635 string t8 = !strconcat(t7, ";\n\t");
636 string t9 = !strconcat(t8, VecModStr<vecsize, elem, "comm", "2">.s);
637 string t10 = !strconcat(t9, "mov.");
638 string t11 = !strconcat(t10, type);
639 string t12 = !strconcat(t11, " \t${dst}_");
640 string t13 = !strconcat(t12, elem);
641 string t14 = !strconcat(t13, ", $src2");
642 string t15 = !strconcat(t14, VecModStr<vecsize, elem, "pos">.s);
643 string s = !strconcat(t15, ";");
645 class ShuffleAsmStr2<string type>
647 string t1 = ShuffleOneLine<"2", "0", type>.s;
648 string t2 = !strconcat(t1, "\n\t");
649 string s = !strconcat(t2, ShuffleOneLine<"2", "1", type>.s);
651 class ShuffleAsmStr4<string type>
653 string t1 = ShuffleOneLine<"4", "0", type>.s;
654 string t2 = !strconcat(t1, "\n\t");
655 string t3 = !strconcat(t2, ShuffleOneLine<"4", "1", type>.s);
656 string t4 = !strconcat(t3, "\n\t");
657 string t5 = !strconcat(t4, ShuffleOneLine<"4", "2", type>.s);
658 string t6 = !strconcat(t5, "\n\t");
659 string s = !strconcat(t6, ShuffleOneLine<"4", "3", type>.s);
662 let hasSideEffects=0, VecInstType=isVecShuffle.Value in {
663 def VecShuffle_v4f32 : NVPTXVecInst<(outs V4F32Regs:$dst),
664 (ins V4F32Regs:$src1, V4F32Regs:$src2,
665 i8imm:$c0, i8imm:$c1, i8imm:$c2, i8imm:$c3),
666 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1, $c2, $c3;\n\t",
667 ShuffleAsmStr4<"f32">.s),
670 def VecShuffle_v4i32 : NVPTXVecInst<(outs V4I32Regs:$dst),
671 (ins V4I32Regs:$src1, V4I32Regs:$src2,
672 i8imm:$c0, i8imm:$c1, i8imm:$c2, i8imm:$c3),
673 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1, $c2, $c3;\n\t",
674 ShuffleAsmStr4<"u32">.s),
677 def VecShuffle_v4i16 : NVPTXVecInst<(outs V4I16Regs:$dst),
678 (ins V4I16Regs:$src1, V4I16Regs:$src2,
679 i8imm:$c0, i8imm:$c1, i8imm:$c2, i8imm:$c3),
680 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1, $c2, $c3;\n\t",
681 ShuffleAsmStr4<"u16">.s),
684 def VecShuffle_v4i8 : NVPTXVecInst<(outs V4I8Regs:$dst),
685 (ins V4I8Regs:$src1, V4I8Regs:$src2,
686 i8imm:$c0, i8imm:$c1, i8imm:$c2, i8imm:$c3),
687 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1, $c2, $c3;\n\t",
688 ShuffleAsmStr4<"u16">.s),
691 def VecShuffle_v2f32 : NVPTXVecInst<(outs V2F32Regs:$dst),
692 (ins V2F32Regs:$src1, V2F32Regs:$src2,
693 i8imm:$c0, i8imm:$c1),
694 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1;\n\t",
695 ShuffleAsmStr2<"f32">.s),
698 def VecShuffle_v2i32 : NVPTXVecInst<(outs V2I32Regs:$dst),
699 (ins V2I32Regs:$src1, V2I32Regs:$src2,
700 i8imm:$c0, i8imm:$c1),
701 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1;\n\t",
702 ShuffleAsmStr2<"u32">.s),
705 def VecShuffle_v2i8 : NVPTXVecInst<(outs V2I8Regs:$dst),
706 (ins V2I8Regs:$src1, V2I8Regs:$src2,
707 i8imm:$c0, i8imm:$c1),
708 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1;\n\t",
709 ShuffleAsmStr2<"u16">.s),
712 def VecShuffle_v2i16 : NVPTXVecInst<(outs V2I16Regs:$dst),
713 (ins V2I16Regs:$src1, V2I16Regs:$src2,
714 i8imm:$c0, i8imm:$c1),
715 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1;\n\t",
716 ShuffleAsmStr2<"u16">.s),
719 def VecShuffle_v2f64 : NVPTXVecInst<(outs V2F64Regs:$dst),
720 (ins V2F64Regs:$src1, V2F64Regs:$src2,
721 i8imm:$c0, i8imm:$c1),
722 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1;\n\t",
723 ShuffleAsmStr2<"f64">.s),
726 def VecShuffle_v2i64 : NVPTXVecInst<(outs V2I64Regs:$dst),
727 (ins V2I64Regs:$src1, V2I64Regs:$src2,
728 i8imm:$c0, i8imm:$c1),
729 !strconcat("//Mov $dst, $src1, $src2, $c0, $c1;\n\t",
730 ShuffleAsmStr2<"u64">.s),
734 def ShuffleMask0 : SDNodeXForm<vector_shuffle, [{
735 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
736 return CurDAG->getTargetConstant(SVOp->getMaskElt(0), SDLoc(N), MVT::i32);
738 def ShuffleMask1 : SDNodeXForm<vector_shuffle, [{
739 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
740 return CurDAG->getTargetConstant(SVOp->getMaskElt(1), SDLoc(N), MVT::i32);
742 def ShuffleMask2 : SDNodeXForm<vector_shuffle, [{
743 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
744 return CurDAG->getTargetConstant(SVOp->getMaskElt(2), SDLoc(N), MVT::i32);
746 def ShuffleMask3 : SDNodeXForm<vector_shuffle, [{
747 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
748 return CurDAG->getTargetConstant(SVOp->getMaskElt(3), SDLoc(N), MVT::i32);
751 // The spurious call is here to silence a compiler warning about N being
753 def vec_shuf : PatFrag<(ops node:$lhs, node:$rhs),
754 (vector_shuffle node:$lhs, node:$rhs),
755 [{ N->getGluedNode(); return true; }]>;
757 def : Pat<(v2f64 (vec_shuf:$op V2F64Regs:$src1, V2F64Regs:$src2)),
758 (VecShuffle_v2f64 V2F64Regs:$src1, V2F64Regs:$src2,
759 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op))>;
761 def : Pat<(v4f32 (vec_shuf:$op V4F32Regs:$src1, V4F32Regs:$src2)),
762 (VecShuffle_v4f32 V4F32Regs:$src1, V4F32Regs:$src2,
763 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op),
764 (ShuffleMask2 node:$op), (ShuffleMask3 node:$op))>;
766 def : Pat<(v2f32 (vec_shuf:$op V2F32Regs:$src1, V2F32Regs:$src2)),
767 (VecShuffle_v2f32 V2F32Regs:$src1, V2F32Regs:$src2,
768 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op))>;
770 def : Pat<(v2i64 (vec_shuf:$op V2I64Regs:$src1, V2I64Regs:$src2)),
771 (VecShuffle_v2i64 V2I64Regs:$src1, V2I64Regs:$src2,
772 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op))>;
774 def : Pat<(v4i32 (vec_shuf:$op V4I32Regs:$src1, V4I32Regs:$src2)),
775 (VecShuffle_v4i32 V4I32Regs:$src1, V4I32Regs:$src2,
776 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op),
777 (ShuffleMask2 node:$op), (ShuffleMask3 node:$op))>;
779 def : Pat<(v2i32 (vec_shuf:$op V2I32Regs:$src1, V2I32Regs:$src2)),
780 (VecShuffle_v2i32 V2I32Regs:$src1, V2I32Regs:$src2,
781 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op))>;
783 def : Pat<(v4i16 (vec_shuf:$op V4I16Regs:$src1, V4I16Regs:$src2)),
784 (VecShuffle_v4i16 V4I16Regs:$src1, V4I16Regs:$src2,
785 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op),
786 (ShuffleMask2 node:$op), (ShuffleMask3 node:$op))>;
788 def : Pat<(v2i16 (vec_shuf:$op V2I16Regs:$src1, V2I16Regs:$src2)),
789 (VecShuffle_v2i16 V2I16Regs:$src1, V2I16Regs:$src2,
790 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op))>;
792 def : Pat<(v4i8 (vec_shuf:$op V4I8Regs:$src1, V4I8Regs:$src2)),
793 (VecShuffle_v4i8 V4I8Regs:$src1, V4I8Regs:$src2,
794 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op),
795 (ShuffleMask2 node:$op), (ShuffleMask3 node:$op))>;
797 def : Pat<(v2i8 (vec_shuf:$op V2I8Regs:$src1, V2I8Regs:$src2)),
798 (VecShuffle_v2i8 V2I8Regs:$src1, V2I8Regs:$src2,
799 (ShuffleMask0 node:$op), (ShuffleMask1 node:$op))>;
801 class Build_Vector2<string asmstr, NVPTXRegClass vclass, NVPTXRegClass sclass,
803 : NVPTXVecInst<(outs vclass:$dst),
804 (ins sclass:$a1, sclass:$a2),
805 !strconcat(asmstr, "\t${dst:vecfull}, {{$a1, $a2}};"),
806 [(set vclass:$dst, (build_vector sclass:$a1, sclass:$a2))],
808 class Build_Vector4<string asmstr, NVPTXRegClass vclass, NVPTXRegClass sclass,
810 : NVPTXVecInst<(outs vclass:$dst),
811 (ins sclass:$a1, sclass:$a2, sclass:$a3, sclass:$a4),
812 !strconcat(asmstr, "\t${dst:vecfull}, {{$a1, $a2, $a3, $a4}};"),
814 (build_vector sclass:$a1, sclass:$a2,
815 sclass:$a3, sclass:$a4))], si>;
817 let isAsCheapAsAMove=1, VecInstType=isVecBuild.Value in {
818 def Build_Vector2_f32 : Build_Vector2<"mov.v2.f32", V2F32Regs, Float32Regs,
820 def Build_Vector2_f64 : Build_Vector2<"mov.v2.f64", V2F64Regs, Float64Regs,
823 def Build_Vector2_i32 : Build_Vector2<"mov.v2.u32", V2I32Regs, Int32Regs,
825 def Build_Vector2_i64 : Build_Vector2<"mov.v2.u64", V2I64Regs, Int64Regs,
827 def Build_Vector2_i16 : Build_Vector2<"mov.v2.u16", V2I16Regs, Int16Regs,
829 def Build_Vector2_i8 : Build_Vector2<"mov.v2.u16", V2I8Regs, Int8Regs,
832 def Build_Vector4_f32 : Build_Vector4<"mov.v4.f32", V4F32Regs, Float32Regs,
835 def Build_Vector4_i32 : Build_Vector4<"mov.v4.u32", V4I32Regs, Int32Regs,
837 def Build_Vector4_i16 : Build_Vector4<"mov.v4.u16", V4I16Regs, Int16Regs,
839 def Build_Vector4_i8 : Build_Vector4<"mov.v4.u16", V4I8Regs, Int8Regs,
843 class Vec_Move<string asmstr, NVPTXRegClass vclass, NVPTXInst sop=NOP>
844 : NVPTXVecInst<(outs vclass:$dst), (ins vclass:$src),
845 !strconcat(asmstr, "\t${dst:vecfull}, ${src:vecfull};"),
848 let isAsCheapAsAMove=1, hasSideEffects=0, IsSimpleMove=1,
849 VecInstType=isVecOther.Value in {
850 def V4f32Mov : Vec_Move<"mov.v4.f32", V4F32Regs, FMOV32rr>;
851 def V2f32Mov : Vec_Move<"mov.v2.f32", V2F32Regs, FMOV32rr>;
853 def V4i32Mov : Vec_Move<"mov.v4.u32", V4I32Regs, IMOV32rr>;
854 def V2i32Mov : Vec_Move<"mov.v2.u32", V2I32Regs, IMOV32rr>;
856 def V4i16Mov : Vec_Move<"mov.v4.u16", V4I16Regs, IMOV16rr>;
857 def V2i16Mov : Vec_Move<"mov.v2.u16", V2I16Regs, IMOV16rr>;
859 def V4i8Mov : Vec_Move<"mov.v4.u16", V4I8Regs, IMOV8rr>;
860 def V2i8Mov : Vec_Move<"mov.v2.u16", V2I8Regs, IMOV8rr>;
862 def V2f64Mov : Vec_Move<"mov.v2.f64", V2F64Regs, FMOV64rr>;
863 def V2i64Mov : Vec_Move<"mov.v2.u64", V2I64Regs, IMOV64rr>;
866 // extract subvector patterns
867 def extract_subvec : SDNode<"ISD::EXTRACT_SUBVECTOR",
868 SDTypeProfile<1, 2, [SDTCisPtrTy<2>]>>;
870 def : Pat<(v2f32 (extract_subvec V4F32Regs:$src, 0)),
871 (Build_Vector2_f32 (V4f32Extract V4F32Regs:$src, 0),
872 (V4f32Extract V4F32Regs:$src, 1))>;
873 def : Pat<(v2f32 (extract_subvec V4F32Regs:$src, 2)),
874 (Build_Vector2_f32 (V4f32Extract V4F32Regs:$src, 2),
875 (V4f32Extract V4F32Regs:$src, 3))>;
876 def : Pat<(v2i32 (extract_subvec V4I32Regs:$src, 0)),
877 (Build_Vector2_i32 (V4i32Extract V4I32Regs:$src, 0),
878 (V4i32Extract V4I32Regs:$src, 1))>;
879 def : Pat<(v2i32 (extract_subvec V4I32Regs:$src, 2)),
880 (Build_Vector2_i32 (V4i32Extract V4I32Regs:$src, 2),
881 (V4i32Extract V4I32Regs:$src, 3))>;
882 def : Pat<(v2i16 (extract_subvec V4I16Regs:$src, 0)),
883 (Build_Vector2_i16 (V4i16Extract V4I16Regs:$src, 0),
884 (V4i16Extract V4I16Regs:$src, 1))>;
885 def : Pat<(v2i16 (extract_subvec V4I16Regs:$src, 2)),
886 (Build_Vector2_i16 (V4i16Extract V4I16Regs:$src, 2),
887 (V4i16Extract V4I16Regs:$src, 3))>;
888 def : Pat<(v2i8 (extract_subvec V4I8Regs:$src, 0)),
889 (Build_Vector2_i8 (V4i8Extract V4I8Regs:$src, 0),
890 (V4i8Extract V4I8Regs:$src, 1))>;
891 def : Pat<(v2i8 (extract_subvec V4I8Regs:$src, 2)),
892 (Build_Vector2_i8 (V4i8Extract V4I8Regs:$src, 2),
893 (V4i8Extract V4I8Regs:$src, 3))>;
895 // Select instructions
896 class Select_OneLine<string type, string pos> {
897 string t1 = !strconcat("selp.", type);
898 string t2 = !strconcat(t1, " \t${dst}_");
899 string t3 = !strconcat(t2, pos);
900 string t4 = !strconcat(t3, ", ${src1}_");
901 string t5 = !strconcat(t4, pos);
902 string t6 = !strconcat(t5, ", ${src2}_");
903 string t7 = !strconcat(t6, pos);
904 string s = !strconcat(t7, ", $p;");
907 class Select_Str2<string type> {
908 string t1 = Select_OneLine<type, "0">.s;
909 string t2 = !strconcat(t1, "\n\t");
910 string s = !strconcat(t2, Select_OneLine<type, "1">.s);
913 class Select_Str4<string type> {
914 string t1 = Select_OneLine<type, "0">.s;
915 string t2 = !strconcat(t1, "\n\t");
916 string t3 = !strconcat(t2, Select_OneLine<type, "1">.s);
917 string t4 = !strconcat(t3, "\n\t");
918 string t5 = !strconcat(t4, Select_OneLine<type, "2">.s);
919 string t6 = !strconcat(t5, "\n\t");
920 string s = !strconcat(t6, Select_OneLine<type, "3">.s);
924 class Vec_Select<NVPTXRegClass vclass, string asmstr, NVPTXInst sop>
925 : NVPTXVecInst<(outs vclass:$dst),
926 (ins vclass:$src1, vclass:$src2, Int1Regs:$p),
928 [(set vclass:$dst, (select Int1Regs:$p, vclass:$src1,
932 let VecInstType=isVecOther.Value in {
933 def V2I64_Select : Vec_Select<V2I64Regs, Select_Str2<"b64">.s, SELECTi64rr>;
934 def V4I32_Select : Vec_Select<V4I32Regs, Select_Str4<"b32">.s, SELECTi32rr>;
935 def V2I32_Select : Vec_Select<V2I32Regs, Select_Str2<"b32">.s, SELECTi32rr>;
936 def V4I16_Select : Vec_Select<V4I16Regs, Select_Str4<"b16">.s, SELECTi16rr>;
937 def V2I16_Select : Vec_Select<V2I16Regs, Select_Str2<"b16">.s, SELECTi16rr>;
938 def V4I8_Select : Vec_Select<V4I8Regs, Select_Str4<"b16">.s, SELECTi8rr>;
939 def V2I8_Select : Vec_Select<V2I8Regs, Select_Str2<"b16">.s, SELECTi8rr>;
941 def V2F64_Select : Vec_Select<V2F64Regs, Select_Str2<"f64">.s, SELECTf64rr>;
942 def V4F32_Select : Vec_Select<V4F32Regs, Select_Str4<"f32">.s, SELECTf32rr>;
943 def V2F32_Select : Vec_Select<V2F32Regs, Select_Str2<"f32">.s, SELECTf32rr>;
946 // Comparison instructions
948 // setcc convenience fragments.
949 def vsetoeq : PatFrag<(ops node:$lhs, node:$rhs),
950 (setcc node:$lhs, node:$rhs, SETOEQ)>;
951 def vsetogt : PatFrag<(ops node:$lhs, node:$rhs),
952 (setcc node:$lhs, node:$rhs, SETOGT)>;
953 def vsetoge : PatFrag<(ops node:$lhs, node:$rhs),
954 (setcc node:$lhs, node:$rhs, SETOGE)>;
955 def vsetolt : PatFrag<(ops node:$lhs, node:$rhs),
956 (setcc node:$lhs, node:$rhs, SETOLT)>;
957 def vsetole : PatFrag<(ops node:$lhs, node:$rhs),
958 (setcc node:$lhs, node:$rhs, SETOLE)>;
959 def vsetone : PatFrag<(ops node:$lhs, node:$rhs),
960 (setcc node:$lhs, node:$rhs, SETONE)>;
961 def vseto : PatFrag<(ops node:$lhs, node:$rhs),
962 (setcc node:$lhs, node:$rhs, SETO)>;
963 def vsetuo : PatFrag<(ops node:$lhs, node:$rhs),
964 (setcc node:$lhs, node:$rhs, SETUO)>;
965 def vsetueq : PatFrag<(ops node:$lhs, node:$rhs),
966 (setcc node:$lhs, node:$rhs, SETUEQ)>;
967 def vsetugt : PatFrag<(ops node:$lhs, node:$rhs),
968 (setcc node:$lhs, node:$rhs, SETUGT)>;
969 def vsetuge : PatFrag<(ops node:$lhs, node:$rhs),
970 (setcc node:$lhs, node:$rhs, SETUGE)>;
971 def vsetult : PatFrag<(ops node:$lhs, node:$rhs),
972 (setcc node:$lhs, node:$rhs, SETULT)>;
973 def vsetule : PatFrag<(ops node:$lhs, node:$rhs),
974 (setcc node:$lhs, node:$rhs, SETULE)>;
975 def vsetune : PatFrag<(ops node:$lhs, node:$rhs),
976 (setcc node:$lhs, node:$rhs, SETUNE)>;
977 def vseteq : PatFrag<(ops node:$lhs, node:$rhs),
978 (setcc node:$lhs, node:$rhs, SETEQ)>;
979 def vsetgt : PatFrag<(ops node:$lhs, node:$rhs),
980 (setcc node:$lhs, node:$rhs, SETGT)>;
981 def vsetge : PatFrag<(ops node:$lhs, node:$rhs),
982 (setcc node:$lhs, node:$rhs, SETGE)>;
983 def vsetlt : PatFrag<(ops node:$lhs, node:$rhs),
984 (setcc node:$lhs, node:$rhs, SETLT)>;
985 def vsetle : PatFrag<(ops node:$lhs, node:$rhs),
986 (setcc node:$lhs, node:$rhs, SETLE)>;
987 def vsetne : PatFrag<(ops node:$lhs, node:$rhs),
988 (setcc node:$lhs, node:$rhs, SETNE)>;
990 class Vec_Compare<PatFrag op, NVPTXRegClass outrclass, NVPTXRegClass inrclass,
992 : NVPTXVecInst<(outs outrclass:$dst),
993 (ins inrclass:$a, inrclass:$b),
995 [(set outrclass:$dst, (op inrclass:$a, inrclass:$b))],
998 multiclass Vec_Compare_All<PatFrag op,
1004 def V2I8 : Vec_Compare<op, V2I8Regs, V2I8Regs, inst8>;
1005 def V4I8 : Vec_Compare<op, V4I8Regs, V4I8Regs, inst8>;
1006 def V2I16 : Vec_Compare<op, V2I16Regs, V2I16Regs, inst16>;
1007 def V4I16 : Vec_Compare<op, V4I16Regs, V4I16Regs, inst16>;
1008 def V2I32 : Vec_Compare<op, V2I32Regs, V2I32Regs, inst32>;
1009 def V4I32 : Vec_Compare<op, V4I32Regs, V4I32Regs, inst32>;
1010 def V2I64 : Vec_Compare<op, V2I64Regs, V2I64Regs, inst64>;
1013 let VecInstType=isVecOther.Value in {
1014 defm VecSGT : Vec_Compare_All<vsetgt, ISetSGTi8rr_toi8, ISetSGTi16rr_toi16,
1015 ISetSGTi32rr_toi32, ISetSGTi64rr_toi64>;
1016 defm VecUGT : Vec_Compare_All<vsetugt, ISetUGTi8rr_toi8, ISetUGTi16rr_toi16,
1017 ISetUGTi32rr_toi32, ISetUGTi64rr_toi64>;
1018 defm VecSLT : Vec_Compare_All<vsetlt, ISetSLTi8rr_toi8, ISetSLTi16rr_toi16,
1019 ISetSLTi32rr_toi32, ISetSLTi64rr_toi64>;
1020 defm VecULT : Vec_Compare_All<vsetult, ISetULTi8rr_toi8, ISetULTi16rr_toi16,
1021 ISetULTi32rr_toi32, ISetULTi64rr_toi64>;
1022 defm VecSGE : Vec_Compare_All<vsetge, ISetSGEi8rr_toi8, ISetSGEi16rr_toi16,
1023 ISetSGEi32rr_toi32, ISetSGEi64rr_toi64>;
1024 defm VecUGE : Vec_Compare_All<vsetuge, ISetUGEi8rr_toi8, ISetUGEi16rr_toi16,
1025 ISetUGEi32rr_toi32, ISetUGEi64rr_toi64>;
1026 defm VecSLE : Vec_Compare_All<vsetle, ISetSLEi8rr_toi8, ISetSLEi16rr_toi16,
1027 ISetSLEi32rr_toi32, ISetSLEi64rr_toi64>;
1028 defm VecULE : Vec_Compare_All<vsetule, ISetULEi8rr_toi8, ISetULEi16rr_toi16,
1029 ISetULEi32rr_toi32, ISetULEi64rr_toi64>;
1030 defm VecSEQ : Vec_Compare_All<vseteq, ISetSEQi8rr_toi8, ISetSEQi16rr_toi16,
1031 ISetSEQi32rr_toi32, ISetSEQi64rr_toi64>;
1032 defm VecUEQ : Vec_Compare_All<vsetueq, ISetUEQi8rr_toi8, ISetUEQi16rr_toi16,
1033 ISetUEQi32rr_toi32, ISetUEQi64rr_toi64>;
1034 defm VecSNE : Vec_Compare_All<vsetne, ISetSNEi8rr_toi8, ISetSNEi16rr_toi16,
1035 ISetSNEi32rr_toi32, ISetSNEi64rr_toi64>;
1036 defm VecUNE : Vec_Compare_All<vsetune, ISetUNEi8rr_toi8, ISetUNEi16rr_toi16,
1037 ISetUNEi32rr_toi32, ISetUNEi64rr_toi64>;
1040 multiclass FVec_Compare_All<PatFrag op,
1044 def V2F32 : Vec_Compare<op, V2I32Regs, V2F32Regs, instf32>;
1045 def V4F32 : Vec_Compare<op, V4I32Regs, V4F32Regs, instf32>;
1046 def V2F64 : Vec_Compare<op, V2I64Regs, V2F64Regs, instf64>;
1049 let VecInstType=isVecOther.Value in {
1050 defm FVecGT : FVec_Compare_All<vsetogt, FSetGTf32rr_toi32,
1052 defm FVecLT : FVec_Compare_All<vsetolt, FSetLTf32rr_toi32,
1054 defm FVecGE : FVec_Compare_All<vsetoge, FSetGEf32rr_toi32,
1056 defm FVecLE : FVec_Compare_All<vsetole, FSetLEf32rr_toi32,
1058 defm FVecEQ : FVec_Compare_All<vsetoeq, FSetEQf32rr_toi32,
1060 defm FVecNE : FVec_Compare_All<vsetone, FSetNEf32rr_toi32,
1063 defm FVecUGT : FVec_Compare_All<vsetugt, FSetUGTf32rr_toi32,
1064 FSetUGTf64rr_toi64>;
1065 defm FVecULT : FVec_Compare_All<vsetult, FSetULTf32rr_toi32,
1066 FSetULTf64rr_toi64>;
1067 defm FVecUGE : FVec_Compare_All<vsetuge, FSetUGEf32rr_toi32,
1068 FSetUGEf64rr_toi64>;
1069 defm FVecULE : FVec_Compare_All<vsetule, FSetULEf32rr_toi32,
1070 FSetULEf64rr_toi64>;
1071 defm FVecUEQ : FVec_Compare_All<vsetueq, FSetUEQf32rr_toi32,
1072 FSetUEQf64rr_toi64>;
1073 defm FVecUNE : FVec_Compare_All<vsetune, FSetUNEf32rr_toi32,
1074 FSetUNEf64rr_toi64>;
1076 defm FVecNUM : FVec_Compare_All<vseto, FSetNUMf32rr_toi32,
1077 FSetNUMf64rr_toi64>;
1078 defm FVecNAN : FVec_Compare_All<vsetuo, FSetNANf32rr_toi32,
1079 FSetNANf64rr_toi64>;
1082 class LoadParamScalar4Inst<NVPTXRegClass regclass, string opstr> :
1083 NVPTXInst<(outs regclass:$d1, regclass:$d2, regclass:$d3, regclass:$d4),
1084 (ins i32imm:$a, i32imm:$b),
1085 !strconcat(!strconcat("ld.param", opstr),
1086 "\t{{$d1, $d2, $d3, $d4}}, [retval0+$b];"), []>;
1088 class LoadParamScalar2Inst<NVPTXRegClass regclass, string opstr> :
1089 NVPTXInst<(outs regclass:$d1, regclass:$d2),
1090 (ins i32imm:$a, i32imm:$b),
1091 !strconcat(!strconcat("ld.param", opstr),
1092 "\t{{$d1, $d2}}, [retval0+$b];"), []>;
1095 class StoreParamScalar4Inst<NVPTXRegClass regclass, string opstr> :
1097 (ins regclass:$s1, regclass:$s2, regclass:$s3, regclass:$s4,
1098 i32imm:$a, i32imm:$b),
1099 !strconcat(!strconcat("st.param", opstr),
1100 "\t[param$a+$b], {{$s1, $s2, $s3, $s4}};"), []>;
1102 class StoreParamScalar2Inst<NVPTXRegClass regclass, string opstr> :
1104 (ins regclass:$s1, regclass:$s2, i32imm:$a, i32imm:$b),
1105 !strconcat(!strconcat("st.param", opstr),
1106 "\t[param$a+$b], {{$s1, $s2}};"), []>;
1108 class StoreRetvalScalar4Inst<NVPTXRegClass regclass, string opstr> :
1110 (ins regclass:$s1, regclass:$s2, regclass:$s3, regclass:$s4,
1112 !strconcat(!strconcat("st.param", opstr),
1113 "\t[func_retval+$a], {{$s1, $s2, $s3, $s4}};"), []>;
1115 class StoreRetvalScalar2Inst<NVPTXRegClass regclass, string opstr> :
1117 (ins regclass:$s1, regclass:$s2, i32imm:$a),
1118 !strconcat(!strconcat("st.param", opstr),
1119 "\t[func_retval+$a], {{$s1, $s2}};"), []>;
1121 def LoadParamScalar4I32 : LoadParamScalar4Inst<Int32Regs, ".v4.b32">;
1122 def LoadParamScalar4I16 : LoadParamScalar4Inst<Int16Regs, ".v4.b16">;
1123 def LoadParamScalar4I8 : LoadParamScalar4Inst<Int8Regs, ".v4.b8">;
1125 def LoadParamScalar2I64 : LoadParamScalar2Inst<Int32Regs, ".v2.b64">;
1126 def LoadParamScalar2I32 : LoadParamScalar2Inst<Int32Regs, ".v2.b32">;
1127 def LoadParamScalar2I16 : LoadParamScalar2Inst<Int32Regs, ".v2.b16">;
1128 def LoadParamScalar2I8 : LoadParamScalar2Inst<Int32Regs, ".v2.b8">;
1130 def LoadParamScalar4F32 : LoadParamScalar4Inst<Float32Regs, ".v4.f32">;
1131 def LoadParamScalar2F32 : LoadParamScalar2Inst<Float32Regs, ".v2.f32">;
1132 def LoadParamScalar2F64 : LoadParamScalar2Inst<Float64Regs, ".v2.f64">;
1134 def StoreParamScalar4I32 : StoreParamScalar4Inst<Int32Regs, ".v4.b32">;
1135 def StoreParamScalar4I16 : StoreParamScalar4Inst<Int16Regs, ".v4.b16">;
1136 def StoreParamScalar4I8 : StoreParamScalar4Inst<Int8Regs, ".v4.b8">;
1138 def StoreParamScalar2I64 : StoreParamScalar2Inst<Int64Regs, ".v2.b64">;
1139 def StoreParamScalar2I32 : StoreParamScalar2Inst<Int32Regs, ".v2.b32">;
1140 def StoreParamScalar2I16 : StoreParamScalar2Inst<Int16Regs, ".v2.b16">;
1141 def StoreParamScalar2I8 : StoreParamScalar2Inst<Int8Regs, ".v2.b8">;
1143 def StoreParamScalar4F32 : StoreParamScalar4Inst<Float32Regs, ".v4.f32">;
1144 def StoreParamScalar2F32 : StoreParamScalar2Inst<Float32Regs, ".v2.f32">;
1145 def StoreParamScalar2F64 : StoreParamScalar2Inst<Float64Regs, ".v2.f64">;
1147 def StoreRetvalScalar4I32 : StoreRetvalScalar4Inst<Int32Regs, ".v4.b32">;
1148 def StoreRetvalScalar4I16 : StoreRetvalScalar4Inst<Int16Regs, ".v4.b16">;
1149 def StoreRetvalScalar4I8 : StoreRetvalScalar4Inst<Int8Regs, ".v4.b8">;
1151 def StoreRetvalScalar2I64 : StoreRetvalScalar2Inst<Int64Regs, ".v2.b64">;
1152 def StoreRetvalScalar2I32 : StoreRetvalScalar2Inst<Int32Regs, ".v2.b32">;
1153 def StoreRetvalScalar2I16 : StoreRetvalScalar2Inst<Int16Regs, ".v2.b16">;
1154 def StoreRetvalScalar2I8 : StoreRetvalScalar2Inst<Int8Regs, ".v2.b8">;
1156 def StoreRetvalScalar4F32 : StoreRetvalScalar4Inst<Float32Regs, ".v4.f32">;
1157 def StoreRetvalScalar2F32 : StoreRetvalScalar2Inst<Float32Regs, ".v2.f32">;
1158 def StoreRetvalScalar2F64 : StoreRetvalScalar2Inst<Float64Regs, ".v2.f64">;
1160 class LoadParamVecInst<NVPTXRegClass regclass, string opstr, NVPTXInst sop=NOP>:
1161 NVPTXVecInst<(outs regclass:$dst), (ins i32imm:$a, i32imm:$b),
1162 "loadparam : $dst <- [$a, $b]",
1163 [(set regclass:$dst, (LoadParam (i32 imm:$a), (i32 imm:$b)))],
1166 class StoreParamVecInst<NVPTXRegClass regclass, string opstr, NVPTXInst sop=NOP>
1167 : NVPTXVecInst<(outs), (ins regclass:$val, i32imm:$a, i32imm:$b),
1168 "storeparam : [$a, $b] <- $val",
1169 [(StoreParam (i32 imm:$a), (i32 imm:$b), regclass:$val)], sop>;
1171 class StoreRetvalVecInst<NVPTXRegClass regclass, string opstr,
1173 : NVPTXVecInst<(outs), (ins regclass:$val, i32imm:$a),
1174 "storeretval : retval[$a] <- $val",
1175 [(StoreRetval (i32 imm:$a), regclass:$val)], sop>;
1177 let VecInstType=isVecLD.Value in {
1178 def LoadParamV4I32 : LoadParamVecInst<V4I32Regs, ".v4.b32",
1179 LoadParamScalar4I32>;
1180 def LoadParamV4I16 : LoadParamVecInst<V4I16Regs, ".v4.b16",
1181 LoadParamScalar4I16>;
1182 def LoadParamV4I8 : LoadParamVecInst<V4I8Regs, ".v4.b8",
1183 LoadParamScalar4I8>;
1185 def LoadParamV2I64 : LoadParamVecInst<V2I64Regs, ".v2.b64",
1186 LoadParamScalar2I64>;
1187 def LoadParamV2I32 : LoadParamVecInst<V2I32Regs, ".v2.b32",
1188 LoadParamScalar2I32>;
1189 def LoadParamV2I16 : LoadParamVecInst<V2I16Regs, ".v2.b16",
1190 LoadParamScalar2I16>;
1191 def LoadParamV2I8 : LoadParamVecInst<V2I8Regs, ".v2.b8",
1192 LoadParamScalar2I8>;
1194 def LoadParamV4F32 : LoadParamVecInst<V4F32Regs, ".v4.f32",
1195 LoadParamScalar4F32>;
1196 def LoadParamV2F32 : LoadParamVecInst<V2F32Regs, ".v2.f32",
1197 LoadParamScalar2F32>;
1198 def LoadParamV2F64 : LoadParamVecInst<V2F64Regs, ".v2.f64",
1199 LoadParamScalar2F64>;
1202 let VecInstType=isVecST.Value in {
1203 def StoreParamV4I32 : StoreParamVecInst<V4I32Regs, ".v4.b32",
1204 StoreParamScalar4I32>;
1205 def StoreParamV4I16 : StoreParamVecInst<V4I16Regs, ".v4.b16",
1206 StoreParamScalar4I16>;
1207 def StoreParamV4I8 : StoreParamVecInst<V4I8Regs, ".v4.b8",
1208 StoreParamScalar4I8>;
1210 def StoreParamV2I64 : StoreParamVecInst<V2I64Regs, ".v2.b64",
1211 StoreParamScalar2I64>;
1212 def StoreParamV2I32 : StoreParamVecInst<V2I32Regs, ".v2.b32",
1213 StoreParamScalar2I32>;
1214 def StoreParamV2I16 : StoreParamVecInst<V2I16Regs, ".v2.b16",
1215 StoreParamScalar2I16>;
1216 def StoreParamV2I8 : StoreParamVecInst<V2I8Regs, ".v2.b8",
1217 StoreParamScalar2I8>;
1219 def StoreParamV4F32 : StoreParamVecInst<V4F32Regs, ".v4.f32",
1220 StoreParamScalar4F32>;
1221 def StoreParamV2F32 : StoreParamVecInst<V2F32Regs, ".v2.f32",
1222 StoreParamScalar2F32>;
1223 def StoreParamV2F64 : StoreParamVecInst<V2F64Regs, ".v2.f64",
1224 StoreParamScalar2F64>;
1226 def StoreRetvalV4I32 : StoreRetvalVecInst<V4I32Regs, ".v4.b32",
1227 StoreRetvalScalar4I32>;
1228 def StoreRetvalV4I16 : StoreRetvalVecInst<V4I16Regs, ".v4.b16",
1229 StoreRetvalScalar4I16>;
1230 def StoreRetvalV4I8 : StoreRetvalVecInst<V4I8Regs, ".v4.b8",
1231 StoreRetvalScalar4I8>;
1233 def StoreRetvalV2I64 : StoreRetvalVecInst<V2I64Regs, ".v2.b64",
1234 StoreRetvalScalar2I64>;
1235 def StoreRetvalV2I32 : StoreRetvalVecInst<V2I32Regs, ".v2.b32",
1236 StoreRetvalScalar2I32>;
1237 def StoreRetvalV2I16 : StoreRetvalVecInst<V2I16Regs, ".v2.b16",
1238 StoreRetvalScalar2I16>;
1239 def StoreRetvalV2I8 : StoreRetvalVecInst<V2I8Regs, ".v2.b8",
1240 StoreRetvalScalar2I8>;
1242 def StoreRetvalV4F32 : StoreRetvalVecInst<V4F32Regs, ".v4.f32",
1243 StoreRetvalScalar4F32>;
1244 def StoreRetvalV2F32 : StoreRetvalVecInst<V2F32Regs, ".v2.f32",
1245 StoreRetvalScalar2F32>;
1246 def StoreRetvalV2F64 : StoreRetvalVecInst<V2F64Regs, ".v2.f64",
1247 StoreRetvalScalar2F64>;
1252 // Int vector to int scalar bit convert
1254 def : Pat<(i32 (bitconvert V4I8Regs:$s)),
1255 (V4I8toI32 (V4i8Extract V4I8Regs:$s,0), (V4i8Extract V4I8Regs:$s,1),
1256 (V4i8Extract V4I8Regs:$s,2), (V4i8Extract V4I8Regs:$s,3))>;
1258 def : Pat<(i64 (bitconvert V4I16Regs:$s)),
1259 (V4I16toI64 (V4i16Extract V4I16Regs:$s,0),
1260 (V4i16Extract V4I16Regs:$s,1),
1261 (V4i16Extract V4I16Regs:$s,2),
1262 (V4i16Extract V4I16Regs:$s,3))>;
1264 def : Pat<(i16 (bitconvert V2I8Regs:$s)),
1265 (V2I8toI16 (V2i8Extract V2I8Regs:$s,0), (V2i8Extract V2I8Regs:$s,1))>;
1267 def : Pat<(i32 (bitconvert V2I16Regs:$s)),
1268 (V2I16toI32 (V2i16Extract V2I16Regs:$s,0),
1269 (V2i16Extract V2I16Regs:$s,1))>;
1271 def : Pat<(i64 (bitconvert V2I32Regs:$s)),
1272 (V2I32toI64 (V2i32Extract V2I32Regs:$s,0),
1273 (V2i32Extract V2I32Regs:$s,1))>;
1275 // Int scalar to int vector bit convert
1276 let VecInstType=isVecDest.Value in {
1278 def VecI32toV4I8 : NVPTXVecInst<(outs V4I8Regs:$d), (ins Int32Regs:$s),
1280 [(set V4I8Regs:$d, (bitconvert Int32Regs:$s))],
1283 def VecI64toV4I16 : NVPTXVecInst<(outs V4I16Regs:$d), (ins Int64Regs:$s),
1285 [(set V4I16Regs:$d, (bitconvert Int64Regs:$s))],
1288 def VecI16toV2I8 : NVPTXVecInst<(outs V2I8Regs:$d), (ins Int16Regs:$s),
1290 [(set V2I8Regs:$d, (bitconvert Int16Regs:$s))],
1293 def VecI32toV2I16 : NVPTXVecInst<(outs V2I16Regs:$d), (ins Int32Regs:$s),
1295 [(set V2I16Regs:$d, (bitconvert Int32Regs:$s))],
1298 def VecI64toV2I32 : NVPTXVecInst<(outs V2I32Regs:$d), (ins Int64Regs:$s),
1300 [(set V2I32Regs:$d, (bitconvert Int64Regs:$s))],
1304 // Int vector to int vector bit convert
1306 def : Pat<(v2i16 (bitconvert V4I8Regs:$s)),
1308 (V4I8toI32 (V4i8Extract V4I8Regs:$s,0), (V4i8Extract V4I8Regs:$s,1),
1309 (V4i8Extract V4I8Regs:$s,2), (V4i8Extract V4I8Regs:$s,3)))>;
1311 def : Pat<(v2i32 (bitconvert V4I16Regs:$s)),
1313 (V4I16toI64 (V4i16Extract V4I16Regs:$s,0), (V4i16Extract V4I16Regs:$s,1),
1314 (V4i16Extract V4I16Regs:$s,2), (V4i16Extract V4I16Regs:$s,3)))>;
1316 def : Pat<(v4i8 (bitconvert V2I16Regs:$s)),
1318 (V2I16toI32 (V2i16Extract V2I16Regs:$s,0), (V2i16Extract V2I16Regs:$s,1)))>;
1320 def : Pat<(v4i16 (bitconvert V2I32Regs:$s)),
1322 (V2I32toI64 (V2i32Extract V2I32Regs:$s,0), (V2i32Extract V2I32Regs:$s,1)))>;
1324 def : Pat<(v4i32 (bitconvert V2I64Regs:$s)),
1326 (V2i32Extract (VecI64toV2I32 (V2i64Extract V2I64Regs:$s, 0)), 0),
1327 (V2i32Extract (VecI64toV2I32 (V2i64Extract V2I64Regs:$s, 0)), 1),
1328 (V2i32Extract (VecI64toV2I32 (V2i64Extract V2I64Regs:$s, 1)), 0),
1329 (V2i32Extract (VecI64toV2I32 (V2i64Extract V2I64Regs:$s, 1)), 1))>;
1331 def : Pat<(v2i64 (bitconvert V4I32Regs:$s)),
1333 (V2I32toI64 (V4i32Extract V4I32Regs:$s,0), (V4i32Extract V4I32Regs:$s,1)),
1334 (V2I32toI64 (V4i32Extract V4I32Regs:$s,2), (V4i32Extract V4I32Regs:$s,3)))>;
1336 // Fp scalar to fp vector convert
1338 let VecInstType=isVecDest.Value in {
1339 def VecF64toV2F32 : NVPTXVecInst<(outs V2F32Regs:$d), (ins Float64Regs:$s),
1341 [(set V2F32Regs:$d, (bitconvert Float64Regs:$s))],
1345 // Fp vector to fp scalar convert
1347 def : Pat<(f64 (bitconvert V2F32Regs:$s)),
1348 (V2F32toF64 (V2f32Extract V2F32Regs:$s,0), (V2f32Extract V2F32Regs:$s,1))>;
1350 // Fp scalar to int vector convert
1352 def : Pat<(v4i8 (bitconvert Float32Regs:$s)),
1353 (VecI32toV4I8 (BITCONVERT_32_F2I Float32Regs:$s))>;
1355 def : Pat<(v2i16 (bitconvert Float32Regs:$s)),
1356 (VecI32toV2I16 (BITCONVERT_32_F2I Float32Regs:$s))>;
1358 def : Pat<(v4i16 (bitconvert Float64Regs:$s)),
1359 (VecI64toV4I16 (BITCONVERT_64_F2I Float64Regs:$s))>;
1361 def : Pat<(v2i32 (bitconvert Float64Regs:$s)),
1362 (VecI64toV2I32 (BITCONVERT_64_F2I Float64Regs:$s))>;
1364 // Int vector to fp scalar convert
1366 def : Pat<(f32 (bitconvert V4I8Regs:$s)),
1368 (V4I8toI32 (V4i8Extract V4I8Regs:$s,0), (V4i8Extract V4I8Regs:$s,1),
1369 (V4i8Extract V4I8Regs:$s,2), (V4i8Extract V4I8Regs:$s,3)))>;
1371 def : Pat<(f64 (bitconvert V4I16Regs:$s)),
1373 (V4I16toI64 (V4i16Extract V4I16Regs:$s,0), (V4i16Extract V4I16Regs:$s,1),
1374 (V4i16Extract V4I16Regs:$s,2), (V4i16Extract V4I16Regs:$s,3)))>;
1376 def : Pat<(f32 (bitconvert V2I16Regs:$s)),
1378 (V2I16toI32 (V2i16Extract V2I16Regs:$s,0), (V2i16Extract V2I16Regs:$s,1)))>;
1380 def : Pat<(f64 (bitconvert V2I32Regs:$s)),
1382 (V2I32toI64 (V2i32Extract V2I32Regs:$s,0), (V2i32Extract V2I32Regs:$s,1)))>;
1384 // Int scalar to fp vector convert
1386 def : Pat<(v2f32 (bitconvert Int64Regs:$s)),
1387 (VecF64toV2F32 (BITCONVERT_64_I2F Int64Regs:$s))>;
1389 // Fp vector to int scalar convert
1391 def : Pat<(i64 (bitconvert V2F32Regs:$s)),
1393 (V2F32toF64 (V2f32Extract V2F32Regs:$s,0), (V2f32Extract V2F32Regs:$s,1)))>;
1395 // Int vector to fp vector convert
1397 def : Pat<(v4f32 (bitconvert V2I64Regs:$s)),
1399 (BITCONVERT_32_I2F (V2i32Extract (VecI64toV2I32
1400 (V2i64Extract V2I64Regs:$s, 0)), 0)),
1401 (BITCONVERT_32_I2F (V2i32Extract (VecI64toV2I32
1402 (V2i64Extract V2I64Regs:$s, 0)), 1)),
1403 (BITCONVERT_32_I2F (V2i32Extract (VecI64toV2I32
1404 (V2i64Extract V2I64Regs:$s, 1)), 0)),
1405 (BITCONVERT_32_I2F (V2i32Extract (VecI64toV2I32
1406 (V2i64Extract V2I64Regs:$s, 1)), 1)))>;
1408 def : Pat<(v2f64 (bitconvert V2I64Regs:$s)),
1410 (BITCONVERT_64_I2F (V2i64Extract V2I64Regs:$s,0)),
1411 (BITCONVERT_64_I2F (V2i64Extract V2I64Regs:$s,1)))>;
1413 def : Pat<(v2f32 (bitconvert V2I32Regs:$s)),
1415 (BITCONVERT_32_I2F (V2i32Extract V2I32Regs:$s,0)),
1416 (BITCONVERT_32_I2F (V2i32Extract V2I32Regs:$s,1)))>;
1418 def : Pat<(v2f64 (bitconvert V4I32Regs:$s)),
1420 (BITCONVERT_64_I2F (V2I32toI64 (V4i32Extract V4I32Regs:$s,0),
1421 (V4i32Extract V4I32Regs:$s,1))),
1422 (BITCONVERT_64_I2F (V2I32toI64 (V4i32Extract V4I32Regs:$s,2),
1423 (V4i32Extract V4I32Regs:$s,3))))>;
1425 def : Pat<(v4f32 (bitconvert V4I32Regs:$s)),
1427 (BITCONVERT_32_I2F (V4i32Extract V4I32Regs:$s,0)),
1428 (BITCONVERT_32_I2F (V4i32Extract V4I32Regs:$s,1)),
1429 (BITCONVERT_32_I2F (V4i32Extract V4I32Regs:$s,2)),
1430 (BITCONVERT_32_I2F (V4i32Extract V4I32Regs:$s,3)))>;
1432 def : Pat<(v2f32 (bitconvert V4I16Regs:$s)),
1433 (VecF64toV2F32 (BITCONVERT_64_I2F
1434 (V4I16toI64 (V4i16Extract V4I16Regs:$s,0),
1435 (V4i16Extract V4I16Regs:$s,1),
1436 (V4i16Extract V4I16Regs:$s,2),
1437 (V4i16Extract V4I16Regs:$s,3))))>;
1439 // Fp vector to int vector convert
1441 def : Pat<(v2i64 (bitconvert V4F32Regs:$s)),
1443 (BITCONVERT_64_F2I (V2F32toF64 (V4f32Extract V4F32Regs:$s,0),
1444 (V4f32Extract V4F32Regs:$s,1))),
1445 (BITCONVERT_64_F2I (V2F32toF64 (V4f32Extract V4F32Regs:$s,2),
1446 (V4f32Extract V4F32Regs:$s,3))))>;
1448 def : Pat<(v2i64 (bitconvert V2F64Regs:$s)),
1450 (BITCONVERT_64_F2I (V2f64Extract V2F64Regs:$s,0)),
1451 (BITCONVERT_64_F2I (V2f64Extract V2F64Regs:$s,1)))>;
1453 def : Pat<(v2i32 (bitconvert V2F32Regs:$s)),
1455 (BITCONVERT_32_F2I (V2f32Extract V2F32Regs:$s,0)),
1456 (BITCONVERT_32_F2I (V2f32Extract V2F32Regs:$s,1)))>;
1458 def : Pat<(v4i32 (bitconvert V2F64Regs:$s)),
1460 (BITCONVERT_32_F2I (V2f32Extract (VecF64toV2F32
1461 (V2f64Extract V2F64Regs:$s, 0)), 0)),
1462 (BITCONVERT_32_F2I (V2f32Extract (VecF64toV2F32
1463 (V2f64Extract V2F64Regs:$s, 0)), 1)),
1464 (BITCONVERT_32_F2I (V2f32Extract (VecF64toV2F32
1465 (V2f64Extract V2F64Regs:$s, 1)), 0)),
1466 (BITCONVERT_32_F2I (V2f32Extract (VecF64toV2F32
1467 (V2f64Extract V2F64Regs:$s, 1)), 1)))>;
1469 def : Pat<(v4i32 (bitconvert V4F32Regs:$s)),
1471 (BITCONVERT_32_F2I (V4f32Extract V4F32Regs:$s,0)),
1472 (BITCONVERT_32_F2I (V4f32Extract V4F32Regs:$s,1)),
1473 (BITCONVERT_32_F2I (V4f32Extract V4F32Regs:$s,2)),
1474 (BITCONVERT_32_F2I (V4f32Extract V4F32Regs:$s,3)))>;
1476 def : Pat<(v4i16 (bitconvert V2F32Regs:$s)),
1477 (VecI64toV4I16 (BITCONVERT_64_F2I
1478 (V2F32toF64 (V2f32Extract V2F32Regs:$s,0),
1479 (V2f32Extract V2F32Regs:$s,1))))>;