1//==- SystemZInstrFP.td - Floating-point SystemZ instructions --*- tblgen-*-==//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9
10//===----------------------------------------------------------------------===//
11// Select instructions
12//===----------------------------------------------------------------------===//
13
14// C's ?: operator for floating-point operands.
15def SelectF32  : SelectWrapper<FP32>;
16def SelectF64  : SelectWrapper<FP64>;
17def SelectF128 : SelectWrapper<FP128>;
18
19defm CondStoreF32 : CondStores<FP32, nonvolatile_store,
20                               nonvolatile_load, bdxaddr20only>;
21defm CondStoreF64 : CondStores<FP64, nonvolatile_store,
22                               nonvolatile_load, bdxaddr20only>;
23
24//===----------------------------------------------------------------------===//
25// Move instructions
26//===----------------------------------------------------------------------===//
27
28// Load zero.
29let neverHasSideEffects = 1, isAsCheapAsAMove = 1, isMoveImm = 1 in {
30  def LZER : InherentRRE<"lzer", 0xB374, FP32,  (fpimm0)>;
31  def LZDR : InherentRRE<"lzdr", 0xB375, FP64,  (fpimm0)>;
32  def LZXR : InherentRRE<"lzxr", 0xB376, FP128, (fpimm0)>;
33}
34
35// Moves between two floating-point registers.
36let neverHasSideEffects = 1 in {
37  def LER : UnaryRR <"le", 0x38,   null_frag, FP32,  FP32>;
38  def LDR : UnaryRR <"ld", 0x28,   null_frag, FP64,  FP64>;
39  def LXR : UnaryRRE<"lx", 0xB365, null_frag, FP128, FP128>;
40}
41
42// Moves between two floating-point registers that also set the condition
43// codes.
44let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
45  defm LTEBR : LoadAndTestRRE<"lteb", 0xB302, FP32>;
46  defm LTDBR : LoadAndTestRRE<"ltdb", 0xB312, FP64>;
47  defm LTXBR : LoadAndTestRRE<"ltxb", 0xB342, FP128>;
48}
49def : CompareZeroFP<LTEBRCompare, FP32>;
50def : CompareZeroFP<LTDBRCompare, FP64>;
51def : CompareZeroFP<LTXBRCompare, FP128>;
52
53// Moves between 64-bit integer and floating-point registers.
54def LGDR : UnaryRRE<"lgd", 0xB3CD, bitconvert, GR64, FP64>;
55def LDGR : UnaryRRE<"ldg", 0xB3C1, bitconvert, FP64, GR64>;
56
57// fcopysign with an FP32 result.
58let isCodeGenOnly = 1 in {
59  def CPSDRss : BinaryRRF<"cpsd", 0xB372, fcopysign, FP32, FP32>;
60  def CPSDRsd : BinaryRRF<"cpsd", 0xB372, fcopysign, FP32, FP64>;
61}
62
63// The sign of an FP128 is in the high register.
64def : Pat<(fcopysign FP32:$src1, FP128:$src2),
65          (CPSDRsd FP32:$src1, (EXTRACT_SUBREG FP128:$src2, subreg_h64))>;
66
67// fcopysign with an FP64 result.
68let isCodeGenOnly = 1 in
69  def CPSDRds : BinaryRRF<"cpsd", 0xB372, fcopysign, FP64, FP32>;
70def CPSDRdd : BinaryRRF<"cpsd", 0xB372, fcopysign, FP64, FP64>;
71
72// The sign of an FP128 is in the high register.
73def : Pat<(fcopysign FP64:$src1, FP128:$src2),
74          (CPSDRdd FP64:$src1, (EXTRACT_SUBREG FP128:$src2, subreg_h64))>;
75
76// fcopysign with an FP128 result.  Use "upper" as the high half and leave
77// the low half as-is.
78class CopySign128<RegisterOperand cls, dag upper>
79  : Pat<(fcopysign FP128:$src1, cls:$src2),
80        (INSERT_SUBREG FP128:$src1, upper, subreg_h64)>;
81
82def : CopySign128<FP32,  (CPSDRds (EXTRACT_SUBREG FP128:$src1, subreg_h64),
83                                  FP32:$src2)>;
84def : CopySign128<FP64,  (CPSDRdd (EXTRACT_SUBREG FP128:$src1, subreg_h64),
85                                  FP64:$src2)>;
86def : CopySign128<FP128, (CPSDRdd (EXTRACT_SUBREG FP128:$src1, subreg_h64),
87                                  (EXTRACT_SUBREG FP128:$src2, subreg_h64))>;
88
89defm LoadStoreF32  : MVCLoadStore<load, f32,  MVCSequence, 4>;
90defm LoadStoreF64  : MVCLoadStore<load, f64,  MVCSequence, 8>;
91defm LoadStoreF128 : MVCLoadStore<load, f128, MVCSequence, 16>;
92
93//===----------------------------------------------------------------------===//
94// Load instructions
95//===----------------------------------------------------------------------===//
96
97let canFoldAsLoad = 1, SimpleBDXLoad = 1 in {
98  defm LE : UnaryRXPair<"le", 0x78, 0xED64, load, FP32, 4>;
99  defm LD : UnaryRXPair<"ld", 0x68, 0xED65, load, FP64, 8>;
100
101  // These instructions are split after register allocation, so we don't
102  // want a custom inserter.
103  let Has20BitOffset = 1, HasIndex = 1, Is128Bit = 1 in {
104    def LX : Pseudo<(outs FP128:$dst), (ins bdxaddr20only128:$src),
105                     [(set FP128:$dst, (load bdxaddr20only128:$src))]>;
106  }
107}
108
109//===----------------------------------------------------------------------===//
110// Store instructions
111//===----------------------------------------------------------------------===//
112
113let SimpleBDXStore = 1 in {
114  defm STE : StoreRXPair<"ste", 0x70, 0xED66, store, FP32, 4>;
115  defm STD : StoreRXPair<"std", 0x60, 0xED67, store, FP64, 8>;
116
117  // These instructions are split after register allocation, so we don't
118  // want a custom inserter.
119  let Has20BitOffset = 1, HasIndex = 1, Is128Bit = 1 in {
120    def STX : Pseudo<(outs), (ins FP128:$src, bdxaddr20only128:$dst),
121                     [(store FP128:$src, bdxaddr20only128:$dst)]>;
122  }
123}
124
125//===----------------------------------------------------------------------===//
126// Conversion instructions
127//===----------------------------------------------------------------------===//
128
129// Convert floating-point values to narrower representations, rounding
130// according to the current mode.  The destination of LEXBR and LDXBR
131// is a 128-bit value, but only the first register of the pair is used.
132def LEDBR : UnaryRRE<"ledb", 0xB344, fround,    FP32,  FP64>;
133def LEXBR : UnaryRRE<"lexb", 0xB346, null_frag, FP128, FP128>;
134def LDXBR : UnaryRRE<"ldxb", 0xB345, null_frag, FP128, FP128>;
135
136def : Pat<(f32 (fround FP128:$src)),
137          (EXTRACT_SUBREG (LEXBR FP128:$src), subreg_hh32)>;
138def : Pat<(f64 (fround FP128:$src)),
139          (EXTRACT_SUBREG (LDXBR FP128:$src), subreg_h64)>;
140
141// Extend register floating-point values to wider representations.
142def LDEBR : UnaryRRE<"ldeb", 0xB304, fextend, FP64,  FP32>;
143def LXEBR : UnaryRRE<"lxeb", 0xB306, fextend, FP128, FP32>;
144def LXDBR : UnaryRRE<"lxdb", 0xB305, fextend, FP128, FP64>;
145
146// Extend memory floating-point values to wider representations.
147def LDEB : UnaryRXE<"ldeb", 0xED04, extloadf32, FP64,  4>;
148def LXEB : UnaryRXE<"lxeb", 0xED06, extloadf32, FP128, 4>;
149def LXDB : UnaryRXE<"lxdb", 0xED05, extloadf64, FP128, 8>;
150
151// Convert a signed integer register value to a floating-point one.
152def CEFBR : UnaryRRE<"cefb", 0xB394, sint_to_fp, FP32,  GR32>;
153def CDFBR : UnaryRRE<"cdfb", 0xB395, sint_to_fp, FP64,  GR32>;
154def CXFBR : UnaryRRE<"cxfb", 0xB396, sint_to_fp, FP128, GR32>;
155
156def CEGBR : UnaryRRE<"cegb", 0xB3A4, sint_to_fp, FP32,  GR64>;
157def CDGBR : UnaryRRE<"cdgb", 0xB3A5, sint_to_fp, FP64,  GR64>;
158def CXGBR : UnaryRRE<"cxgb", 0xB3A6, sint_to_fp, FP128, GR64>;
159
160// Convert a floating-point register value to a signed integer value,
161// with the second operand (modifier M3) specifying the rounding mode.
162let Defs = [CC] in {
163  def CFEBR : UnaryRRF<"cfeb", 0xB398, GR32, FP32>;
164  def CFDBR : UnaryRRF<"cfdb", 0xB399, GR32, FP64>;
165  def CFXBR : UnaryRRF<"cfxb", 0xB39A, GR32, FP128>;
166
167  def CGEBR : UnaryRRF<"cgeb", 0xB3A8, GR64, FP32>;
168  def CGDBR : UnaryRRF<"cgdb", 0xB3A9, GR64, FP64>;
169  def CGXBR : UnaryRRF<"cgxb", 0xB3AA, GR64, FP128>;
170}
171
172// fp_to_sint always rounds towards zero, which is modifier value 5.
173def : Pat<(i32 (fp_to_sint FP32:$src)),  (CFEBR 5, FP32:$src)>;
174def : Pat<(i32 (fp_to_sint FP64:$src)),  (CFDBR 5, FP64:$src)>;
175def : Pat<(i32 (fp_to_sint FP128:$src)), (CFXBR 5, FP128:$src)>;
176
177def : Pat<(i64 (fp_to_sint FP32:$src)),  (CGEBR 5, FP32:$src)>;
178def : Pat<(i64 (fp_to_sint FP64:$src)),  (CGDBR 5, FP64:$src)>;
179def : Pat<(i64 (fp_to_sint FP128:$src)), (CGXBR 5, FP128:$src)>;
180
181//===----------------------------------------------------------------------===//
182// Unary arithmetic
183//===----------------------------------------------------------------------===//
184
185// Negation (Load Complement).
186let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
187  def LCEBR : UnaryRRE<"lceb", 0xB303, fneg, FP32,  FP32>;
188  def LCDBR : UnaryRRE<"lcdb", 0xB313, fneg, FP64,  FP64>;
189  def LCXBR : UnaryRRE<"lcxb", 0xB343, fneg, FP128, FP128>;
190}
191
192// Absolute value (Load Positive).
193let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
194  def LPEBR : UnaryRRE<"lpeb", 0xB300, fabs, FP32,  FP32>;
195  def LPDBR : UnaryRRE<"lpdb", 0xB310, fabs, FP64,  FP64>;
196  def LPXBR : UnaryRRE<"lpxb", 0xB340, fabs, FP128, FP128>;
197}
198
199// Negative absolute value (Load Negative).
200let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
201  def LNEBR : UnaryRRE<"lneb", 0xB301, fnabs, FP32,  FP32>;
202  def LNDBR : UnaryRRE<"lndb", 0xB311, fnabs, FP64,  FP64>;
203  def LNXBR : UnaryRRE<"lnxb", 0xB341, fnabs, FP128, FP128>;
204}
205
206// Square root.
207def SQEBR : UnaryRRE<"sqeb", 0xB314, fsqrt, FP32,  FP32>;
208def SQDBR : UnaryRRE<"sqdb", 0xB315, fsqrt, FP64,  FP64>;
209def SQXBR : UnaryRRE<"sqxb", 0xB316, fsqrt, FP128, FP128>;
210
211def SQEB : UnaryRXE<"sqeb", 0xED14, loadu<fsqrt>, FP32, 4>;
212def SQDB : UnaryRXE<"sqdb", 0xED15, loadu<fsqrt>, FP64, 8>;
213
214// Round to an integer, with the second operand (modifier M3) specifying
215// the rounding mode.  These forms always check for inexact conditions.
216def FIEBR : UnaryRRF<"fieb", 0xB357, FP32,  FP32>;
217def FIDBR : UnaryRRF<"fidb", 0xB35F, FP64,  FP64>;
218def FIXBR : UnaryRRF<"fixb", 0xB347, FP128, FP128>;
219
220// Extended forms of the previous three instructions.  M4 can be set to 4
221// to suppress detection of inexact conditions.
222def FIEBRA : UnaryRRF4<"fiebra", 0xB357, FP32,  FP32>,
223             Requires<[FeatureFPExtension]>;
224def FIDBRA : UnaryRRF4<"fidbra", 0xB35F, FP64,  FP64>,
225             Requires<[FeatureFPExtension]>;
226def FIXBRA : UnaryRRF4<"fixbra", 0xB347, FP128, FP128>,
227             Requires<[FeatureFPExtension]>;
228
229// frint rounds according to the current mode (modifier 0) and detects
230// inexact conditions.
231def : Pat<(frint FP32:$src),  (FIEBR 0, FP32:$src)>;
232def : Pat<(frint FP64:$src),  (FIDBR 0, FP64:$src)>;
233def : Pat<(frint FP128:$src), (FIXBR 0, FP128:$src)>;
234
235let Predicates = [FeatureFPExtension] in {
236  // fnearbyint is like frint but does not detect inexact conditions.
237  def : Pat<(fnearbyint FP32:$src),  (FIEBRA 0, FP32:$src,  4)>;
238  def : Pat<(fnearbyint FP64:$src),  (FIDBRA 0, FP64:$src,  4)>;
239  def : Pat<(fnearbyint FP128:$src), (FIXBRA 0, FP128:$src, 4)>;
240
241  // floor is no longer allowed to raise an inexact condition,
242  // so restrict it to the cases where the condition can be suppressed.
243  // Mode 7 is round towards -inf.
244  def : Pat<(ffloor FP32:$src),  (FIEBRA 7, FP32:$src,  4)>;
245  def : Pat<(ffloor FP64:$src),  (FIDBRA 7, FP64:$src,  4)>;
246  def : Pat<(ffloor FP128:$src), (FIXBRA 7, FP128:$src, 4)>;
247
248  // Same idea for ceil, where mode 6 is round towards +inf.
249  def : Pat<(fceil FP32:$src),  (FIEBRA 6, FP32:$src,  4)>;
250  def : Pat<(fceil FP64:$src),  (FIDBRA 6, FP64:$src,  4)>;
251  def : Pat<(fceil FP128:$src), (FIXBRA 6, FP128:$src, 4)>;
252
253  // Same idea for trunc, where mode 5 is round towards zero.
254  def : Pat<(ftrunc FP32:$src),  (FIEBRA 5, FP32:$src,  4)>;
255  def : Pat<(ftrunc FP64:$src),  (FIDBRA 5, FP64:$src,  4)>;
256  def : Pat<(ftrunc FP128:$src), (FIXBRA 5, FP128:$src, 4)>;
257
258  // Same idea for round, where mode 1 is round towards nearest with
259  // ties away from zero.
260  def : Pat<(frnd FP32:$src),  (FIEBRA 1, FP32:$src,  4)>;
261  def : Pat<(frnd FP64:$src),  (FIDBRA 1, FP64:$src,  4)>;
262  def : Pat<(frnd FP128:$src), (FIXBRA 1, FP128:$src, 4)>;
263}
264
265//===----------------------------------------------------------------------===//
266// Binary arithmetic
267//===----------------------------------------------------------------------===//
268
269// Addition.
270let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
271  let isCommutable = 1 in {
272    def AEBR : BinaryRRE<"aeb", 0xB30A, fadd, FP32,  FP32>;
273    def ADBR : BinaryRRE<"adb", 0xB31A, fadd, FP64,  FP64>;
274    def AXBR : BinaryRRE<"axb", 0xB34A, fadd, FP128, FP128>;
275  }
276  def AEB : BinaryRXE<"aeb", 0xED0A, fadd, FP32, load, 4>;
277  def ADB : BinaryRXE<"adb", 0xED1A, fadd, FP64, load, 8>;
278}
279
280// Subtraction.
281let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
282  def SEBR : BinaryRRE<"seb", 0xB30B, fsub, FP32,  FP32>;
283  def SDBR : BinaryRRE<"sdb", 0xB31B, fsub, FP64,  FP64>;
284  def SXBR : BinaryRRE<"sxb", 0xB34B, fsub, FP128, FP128>;
285
286  def SEB : BinaryRXE<"seb",  0xED0B, fsub, FP32, load, 4>;
287  def SDB : BinaryRXE<"sdb",  0xED1B, fsub, FP64, load, 8>;
288}
289
290// Multiplication.
291let isCommutable = 1 in {
292  def MEEBR : BinaryRRE<"meeb", 0xB317, fmul, FP32,  FP32>;
293  def MDBR  : BinaryRRE<"mdb",  0xB31C, fmul, FP64,  FP64>;
294  def MXBR  : BinaryRRE<"mxb",  0xB34C, fmul, FP128, FP128>;
295}
296def MEEB : BinaryRXE<"meeb", 0xED17, fmul, FP32, load, 4>;
297def MDB  : BinaryRXE<"mdb",  0xED1C, fmul, FP64, load, 8>;
298
299// f64 multiplication of two FP32 registers.
300def MDEBR : BinaryRRE<"mdeb", 0xB30C, null_frag, FP64, FP32>;
301def : Pat<(fmul (f64 (fextend FP32:$src1)), (f64 (fextend FP32:$src2))),
302          (MDEBR (INSERT_SUBREG (f64 (IMPLICIT_DEF)),
303                                FP32:$src1, subreg_h32), FP32:$src2)>;
304
305// f64 multiplication of an FP32 register and an f32 memory.
306def MDEB : BinaryRXE<"mdeb", 0xED0C, null_frag, FP64, load, 4>;
307def : Pat<(fmul (f64 (fextend FP32:$src1)),
308                (f64 (extloadf32 bdxaddr12only:$addr))),
309          (MDEB (INSERT_SUBREG (f64 (IMPLICIT_DEF)), FP32:$src1, subreg_h32),
310                bdxaddr12only:$addr)>;
311
312// f128 multiplication of two FP64 registers.
313def MXDBR : BinaryRRE<"mxdb", 0xB307, null_frag, FP128, FP64>;
314def : Pat<(fmul (f128 (fextend FP64:$src1)), (f128 (fextend FP64:$src2))),
315          (MXDBR (INSERT_SUBREG (f128 (IMPLICIT_DEF)),
316                                FP64:$src1, subreg_h64), FP64:$src2)>;
317
318// f128 multiplication of an FP64 register and an f64 memory.
319def MXDB : BinaryRXE<"mxdb", 0xED07, null_frag, FP128, load, 8>;
320def : Pat<(fmul (f128 (fextend FP64:$src1)),
321                (f128 (extloadf64 bdxaddr12only:$addr))),
322          (MXDB (INSERT_SUBREG (f128 (IMPLICIT_DEF)), FP64:$src1, subreg_h64),
323                bdxaddr12only:$addr)>;
324
325// Fused multiply-add.
326def MAEBR : TernaryRRD<"maeb", 0xB30E, z_fma, FP32>;
327def MADBR : TernaryRRD<"madb", 0xB31E, z_fma, FP64>;
328
329def MAEB : TernaryRXF<"maeb", 0xED0E, z_fma, FP32, load, 4>;
330def MADB : TernaryRXF<"madb", 0xED1E, z_fma, FP64, load, 8>;
331
332// Fused multiply-subtract.
333def MSEBR : TernaryRRD<"mseb", 0xB30F, z_fms, FP32>;
334def MSDBR : TernaryRRD<"msdb", 0xB31F, z_fms, FP64>;
335
336def MSEB : TernaryRXF<"mseb", 0xED0F, z_fms, FP32, load, 4>;
337def MSDB : TernaryRXF<"msdb", 0xED1F, z_fms, FP64, load, 8>;
338
339// Division.
340def DEBR : BinaryRRE<"deb", 0xB30D, fdiv, FP32,  FP32>;
341def DDBR : BinaryRRE<"ddb", 0xB31D, fdiv, FP64,  FP64>;
342def DXBR : BinaryRRE<"dxb", 0xB34D, fdiv, FP128, FP128>;
343
344def DEB : BinaryRXE<"deb", 0xED0D, fdiv, FP32, load, 4>;
345def DDB : BinaryRXE<"ddb", 0xED1D, fdiv, FP64, load, 8>;
346
347//===----------------------------------------------------------------------===//
348// Comparisons
349//===----------------------------------------------------------------------===//
350
351let Defs = [CC], CCValues = 0xF in {
352  def CEBR : CompareRRE<"ceb", 0xB309, z_fcmp, FP32,  FP32>;
353  def CDBR : CompareRRE<"cdb", 0xB319, z_fcmp, FP64,  FP64>;
354  def CXBR : CompareRRE<"cxb", 0xB349, z_fcmp, FP128, FP128>;
355
356  def CEB : CompareRXE<"ceb", 0xED09, z_fcmp, FP32, load, 4>;
357  def CDB : CompareRXE<"cdb", 0xED19, z_fcmp, FP64, load, 8>;
358}
359
360//===----------------------------------------------------------------------===//
361// Peepholes
362//===----------------------------------------------------------------------===//
363
364def : Pat<(f32  fpimmneg0), (LCEBR (LZER))>;
365def : Pat<(f64  fpimmneg0), (LCDBR (LZDR))>;
366def : Pat<(f128 fpimmneg0), (LCXBR (LZXR))>;
367