[InstCombine] Signed saturation patterns
[llvm-core.git] / lib / Target / ARM / MCTargetDesc / ARMAddressingModes.h
blob24a9fabf09798edaf5611c2e904b803377a51721
1 //===-- ARMAddressingModes.h - ARM Addressing Modes -------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains the ARM addressing mode implementation stuff.
11 //===----------------------------------------------------------------------===//
13 #ifndef LLVM_LIB_TARGET_ARM_MCTARGETDESC_ARMADDRESSINGMODES_H
14 #define LLVM_LIB_TARGET_ARM_MCTARGETDESC_ARMADDRESSINGMODES_H
16 #include "llvm/ADT/APFloat.h"
17 #include "llvm/ADT/APInt.h"
18 #include "llvm/ADT/bit.h"
19 #include "llvm/Support/ErrorHandling.h"
20 #include "llvm/Support/MathExtras.h"
21 #include <cassert>
23 namespace llvm {
25 /// ARM_AM - ARM Addressing Mode Stuff
26 namespace ARM_AM {
27 enum ShiftOpc {
28 no_shift = 0,
29 asr,
30 lsl,
31 lsr,
32 ror,
33 rrx,
34 uxtw
37 enum AddrOpc {
38 sub = 0,
39 add
42 inline const char *getAddrOpcStr(AddrOpc Op) { return Op == sub ? "-" : ""; }
44 inline const char *getShiftOpcStr(ShiftOpc Op) {
45 switch (Op) {
46 default: llvm_unreachable("Unknown shift opc!");
47 case ARM_AM::asr: return "asr";
48 case ARM_AM::lsl: return "lsl";
49 case ARM_AM::lsr: return "lsr";
50 case ARM_AM::ror: return "ror";
51 case ARM_AM::rrx: return "rrx";
52 case ARM_AM::uxtw: return "uxtw";
56 inline unsigned getShiftOpcEncoding(ShiftOpc Op) {
57 switch (Op) {
58 default: llvm_unreachable("Unknown shift opc!");
59 case ARM_AM::asr: return 2;
60 case ARM_AM::lsl: return 0;
61 case ARM_AM::lsr: return 1;
62 case ARM_AM::ror: return 3;
66 enum AMSubMode {
67 bad_am_submode = 0,
68 ia,
69 ib,
70 da,
74 inline const char *getAMSubModeStr(AMSubMode Mode) {
75 switch (Mode) {
76 default: llvm_unreachable("Unknown addressing sub-mode!");
77 case ARM_AM::ia: return "ia";
78 case ARM_AM::ib: return "ib";
79 case ARM_AM::da: return "da";
80 case ARM_AM::db: return "db";
84 /// rotr32 - Rotate a 32-bit unsigned value right by a specified # bits.
85 ///
86 inline unsigned rotr32(unsigned Val, unsigned Amt) {
87 assert(Amt < 32 && "Invalid rotate amount");
88 return (Val >> Amt) | (Val << ((32-Amt)&31));
91 /// rotl32 - Rotate a 32-bit unsigned value left by a specified # bits.
92 ///
93 inline unsigned rotl32(unsigned Val, unsigned Amt) {
94 assert(Amt < 32 && "Invalid rotate amount");
95 return (Val << Amt) | (Val >> ((32-Amt)&31));
98 //===--------------------------------------------------------------------===//
99 // Addressing Mode #1: shift_operand with registers
100 //===--------------------------------------------------------------------===//
102 // This 'addressing mode' is used for arithmetic instructions. It can
103 // represent things like:
104 // reg
105 // reg [asr|lsl|lsr|ror|rrx] reg
106 // reg [asr|lsl|lsr|ror|rrx] imm
108 // This is stored three operands [rega, regb, opc]. The first is the base
109 // reg, the second is the shift amount (or reg0 if not present or imm). The
110 // third operand encodes the shift opcode and the imm if a reg isn't present.
112 inline unsigned getSORegOpc(ShiftOpc ShOp, unsigned Imm) {
113 return ShOp | (Imm << 3);
115 inline unsigned getSORegOffset(unsigned Op) { return Op >> 3; }
116 inline ShiftOpc getSORegShOp(unsigned Op) { return (ShiftOpc)(Op & 7); }
118 /// getSOImmValImm - Given an encoded imm field for the reg/imm form, return
119 /// the 8-bit imm value.
120 inline unsigned getSOImmValImm(unsigned Imm) { return Imm & 0xFF; }
121 /// getSOImmValRot - Given an encoded imm field for the reg/imm form, return
122 /// the rotate amount.
123 inline unsigned getSOImmValRot(unsigned Imm) { return (Imm >> 8) * 2; }
125 /// getSOImmValRotate - Try to handle Imm with an immediate shifter operand,
126 /// computing the rotate amount to use. If this immediate value cannot be
127 /// handled with a single shifter-op, determine a good rotate amount that will
128 /// take a maximal chunk of bits out of the immediate.
129 inline unsigned getSOImmValRotate(unsigned Imm) {
130 // 8-bit (or less) immediates are trivially shifter_operands with a rotate
131 // of zero.
132 if ((Imm & ~255U) == 0) return 0;
134 // Use CTZ to compute the rotate amount.
135 unsigned TZ = countTrailingZeros(Imm);
137 // Rotate amount must be even. Something like 0x200 must be rotated 8 bits,
138 // not 9.
139 unsigned RotAmt = TZ & ~1;
141 // If we can handle this spread, return it.
142 if ((rotr32(Imm, RotAmt) & ~255U) == 0)
143 return (32-RotAmt)&31; // HW rotates right, not left.
145 // For values like 0xF000000F, we should ignore the low 6 bits, then
146 // retry the hunt.
147 if (Imm & 63U) {
148 unsigned TZ2 = countTrailingZeros(Imm & ~63U);
149 unsigned RotAmt2 = TZ2 & ~1;
150 if ((rotr32(Imm, RotAmt2) & ~255U) == 0)
151 return (32-RotAmt2)&31; // HW rotates right, not left.
154 // Otherwise, we have no way to cover this span of bits with a single
155 // shifter_op immediate. Return a chunk of bits that will be useful to
156 // handle.
157 return (32-RotAmt)&31; // HW rotates right, not left.
160 /// getSOImmVal - Given a 32-bit immediate, if it is something that can fit
161 /// into an shifter_operand immediate operand, return the 12-bit encoding for
162 /// it. If not, return -1.
163 inline int getSOImmVal(unsigned Arg) {
164 // 8-bit (or less) immediates are trivially shifter_operands with a rotate
165 // of zero.
166 if ((Arg & ~255U) == 0) return Arg;
168 unsigned RotAmt = getSOImmValRotate(Arg);
170 // If this cannot be handled with a single shifter_op, bail out.
171 if (rotr32(~255U, RotAmt) & Arg)
172 return -1;
174 // Encode this correctly.
175 return rotl32(Arg, RotAmt) | ((RotAmt>>1) << 8);
178 /// isSOImmTwoPartVal - Return true if the specified value can be obtained by
179 /// or'ing together two SOImmVal's.
180 inline bool isSOImmTwoPartVal(unsigned V) {
181 // If this can be handled with a single shifter_op, bail out.
182 V = rotr32(~255U, getSOImmValRotate(V)) & V;
183 if (V == 0)
184 return false;
186 // If this can be handled with two shifter_op's, accept.
187 V = rotr32(~255U, getSOImmValRotate(V)) & V;
188 return V == 0;
191 /// getSOImmTwoPartFirst - If V is a value that satisfies isSOImmTwoPartVal,
192 /// return the first chunk of it.
193 inline unsigned getSOImmTwoPartFirst(unsigned V) {
194 return rotr32(255U, getSOImmValRotate(V)) & V;
197 /// getSOImmTwoPartSecond - If V is a value that satisfies isSOImmTwoPartVal,
198 /// return the second chunk of it.
199 inline unsigned getSOImmTwoPartSecond(unsigned V) {
200 // Mask out the first hunk.
201 V = rotr32(~255U, getSOImmValRotate(V)) & V;
203 // Take what's left.
204 assert(V == (rotr32(255U, getSOImmValRotate(V)) & V));
205 return V;
208 /// getThumbImmValShift - Try to handle Imm with a 8-bit immediate followed
209 /// by a left shift. Returns the shift amount to use.
210 inline unsigned getThumbImmValShift(unsigned Imm) {
211 // 8-bit (or less) immediates are trivially immediate operand with a shift
212 // of zero.
213 if ((Imm & ~255U) == 0) return 0;
215 // Use CTZ to compute the shift amount.
216 return countTrailingZeros(Imm);
219 /// isThumbImmShiftedVal - Return true if the specified value can be obtained
220 /// by left shifting a 8-bit immediate.
221 inline bool isThumbImmShiftedVal(unsigned V) {
222 // If this can be handled with
223 V = (~255U << getThumbImmValShift(V)) & V;
224 return V == 0;
227 /// getThumbImm16ValShift - Try to handle Imm with a 16-bit immediate followed
228 /// by a left shift. Returns the shift amount to use.
229 inline unsigned getThumbImm16ValShift(unsigned Imm) {
230 // 16-bit (or less) immediates are trivially immediate operand with a shift
231 // of zero.
232 if ((Imm & ~65535U) == 0) return 0;
234 // Use CTZ to compute the shift amount.
235 return countTrailingZeros(Imm);
238 /// isThumbImm16ShiftedVal - Return true if the specified value can be
239 /// obtained by left shifting a 16-bit immediate.
240 inline bool isThumbImm16ShiftedVal(unsigned V) {
241 // If this can be handled with
242 V = (~65535U << getThumbImm16ValShift(V)) & V;
243 return V == 0;
246 /// getThumbImmNonShiftedVal - If V is a value that satisfies
247 /// isThumbImmShiftedVal, return the non-shiftd value.
248 inline unsigned getThumbImmNonShiftedVal(unsigned V) {
249 return V >> getThumbImmValShift(V);
253 /// getT2SOImmValSplat - Return the 12-bit encoded representation
254 /// if the specified value can be obtained by splatting the low 8 bits
255 /// into every other byte or every byte of a 32-bit value. i.e.,
256 /// 00000000 00000000 00000000 abcdefgh control = 0
257 /// 00000000 abcdefgh 00000000 abcdefgh control = 1
258 /// abcdefgh 00000000 abcdefgh 00000000 control = 2
259 /// abcdefgh abcdefgh abcdefgh abcdefgh control = 3
260 /// Return -1 if none of the above apply.
261 /// See ARM Reference Manual A6.3.2.
262 inline int getT2SOImmValSplatVal(unsigned V) {
263 unsigned u, Vs, Imm;
264 // control = 0
265 if ((V & 0xffffff00) == 0)
266 return V;
268 // If the value is zeroes in the first byte, just shift those off
269 Vs = ((V & 0xff) == 0) ? V >> 8 : V;
270 // Any passing value only has 8 bits of payload, splatted across the word
271 Imm = Vs & 0xff;
272 // Likewise, any passing values have the payload splatted into the 3rd byte
273 u = Imm | (Imm << 16);
275 // control = 1 or 2
276 if (Vs == u)
277 return (((Vs == V) ? 1 : 2) << 8) | Imm;
279 // control = 3
280 if (Vs == (u | (u << 8)))
281 return (3 << 8) | Imm;
283 return -1;
286 /// getT2SOImmValRotateVal - Return the 12-bit encoded representation if the
287 /// specified value is a rotated 8-bit value. Return -1 if no rotation
288 /// encoding is possible.
289 /// See ARM Reference Manual A6.3.2.
290 inline int getT2SOImmValRotateVal(unsigned V) {
291 unsigned RotAmt = countLeadingZeros(V);
292 if (RotAmt >= 24)
293 return -1;
295 // If 'Arg' can be handled with a single shifter_op return the value.
296 if ((rotr32(0xff000000U, RotAmt) & V) == V)
297 return (rotr32(V, 24 - RotAmt) & 0x7f) | ((RotAmt + 8) << 7);
299 return -1;
302 /// getT2SOImmVal - Given a 32-bit immediate, if it is something that can fit
303 /// into a Thumb-2 shifter_operand immediate operand, return the 12-bit
304 /// encoding for it. If not, return -1.
305 /// See ARM Reference Manual A6.3.2.
306 inline int getT2SOImmVal(unsigned Arg) {
307 // If 'Arg' is an 8-bit splat, then get the encoded value.
308 int Splat = getT2SOImmValSplatVal(Arg);
309 if (Splat != -1)
310 return Splat;
312 // If 'Arg' can be handled with a single shifter_op return the value.
313 int Rot = getT2SOImmValRotateVal(Arg);
314 if (Rot != -1)
315 return Rot;
317 return -1;
320 inline unsigned getT2SOImmValRotate(unsigned V) {
321 if ((V & ~255U) == 0) return 0;
322 // Use CTZ to compute the rotate amount.
323 unsigned RotAmt = countTrailingZeros(V);
324 return (32 - RotAmt) & 31;
327 inline bool isT2SOImmTwoPartVal(unsigned Imm) {
328 unsigned V = Imm;
329 // Passing values can be any combination of splat values and shifter
330 // values. If this can be handled with a single shifter or splat, bail
331 // out. Those should be handled directly, not with a two-part val.
332 if (getT2SOImmValSplatVal(V) != -1)
333 return false;
334 V = rotr32 (~255U, getT2SOImmValRotate(V)) & V;
335 if (V == 0)
336 return false;
338 // If this can be handled as an immediate, accept.
339 if (getT2SOImmVal(V) != -1) return true;
341 // Likewise, try masking out a splat value first.
342 V = Imm;
343 if (getT2SOImmValSplatVal(V & 0xff00ff00U) != -1)
344 V &= ~0xff00ff00U;
345 else if (getT2SOImmValSplatVal(V & 0x00ff00ffU) != -1)
346 V &= ~0x00ff00ffU;
347 // If what's left can be handled as an immediate, accept.
348 if (getT2SOImmVal(V) != -1) return true;
350 // Otherwise, do not accept.
351 return false;
354 inline unsigned getT2SOImmTwoPartFirst(unsigned Imm) {
355 assert (isT2SOImmTwoPartVal(Imm) &&
356 "Immedate cannot be encoded as two part immediate!");
357 // Try a shifter operand as one part
358 unsigned V = rotr32 (~255, getT2SOImmValRotate(Imm)) & Imm;
359 // If the rest is encodable as an immediate, then return it.
360 if (getT2SOImmVal(V) != -1) return V;
362 // Try masking out a splat value first.
363 if (getT2SOImmValSplatVal(Imm & 0xff00ff00U) != -1)
364 return Imm & 0xff00ff00U;
366 // The other splat is all that's left as an option.
367 assert (getT2SOImmValSplatVal(Imm & 0x00ff00ffU) != -1);
368 return Imm & 0x00ff00ffU;
371 inline unsigned getT2SOImmTwoPartSecond(unsigned Imm) {
372 // Mask out the first hunk
373 Imm ^= getT2SOImmTwoPartFirst(Imm);
374 // Return what's left
375 assert (getT2SOImmVal(Imm) != -1 &&
376 "Unable to encode second part of T2 two part SO immediate");
377 return Imm;
381 //===--------------------------------------------------------------------===//
382 // Addressing Mode #2
383 //===--------------------------------------------------------------------===//
385 // This is used for most simple load/store instructions.
387 // addrmode2 := reg +/- reg shop imm
388 // addrmode2 := reg +/- imm12
390 // The first operand is always a Reg. The second operand is a reg if in
391 // reg/reg form, otherwise it's reg#0. The third field encodes the operation
392 // in bit 12, the immediate in bits 0-11, and the shift op in 13-15. The
393 // fourth operand 16-17 encodes the index mode.
395 // If this addressing mode is a frame index (before prolog/epilog insertion
396 // and code rewriting), this operand will have the form: FI#, reg0, <offs>
397 // with no shift amount for the frame offset.
399 inline unsigned getAM2Opc(AddrOpc Opc, unsigned Imm12, ShiftOpc SO,
400 unsigned IdxMode = 0) {
401 assert(Imm12 < (1 << 12) && "Imm too large!");
402 bool isSub = Opc == sub;
403 return Imm12 | ((int)isSub << 12) | (SO << 13) | (IdxMode << 16) ;
405 inline unsigned getAM2Offset(unsigned AM2Opc) {
406 return AM2Opc & ((1 << 12)-1);
408 inline AddrOpc getAM2Op(unsigned AM2Opc) {
409 return ((AM2Opc >> 12) & 1) ? sub : add;
411 inline ShiftOpc getAM2ShiftOpc(unsigned AM2Opc) {
412 return (ShiftOpc)((AM2Opc >> 13) & 7);
414 inline unsigned getAM2IdxMode(unsigned AM2Opc) { return (AM2Opc >> 16); }
416 //===--------------------------------------------------------------------===//
417 // Addressing Mode #3
418 //===--------------------------------------------------------------------===//
420 // This is used for sign-extending loads, and load/store-pair instructions.
422 // addrmode3 := reg +/- reg
423 // addrmode3 := reg +/- imm8
425 // The first operand is always a Reg. The second operand is a reg if in
426 // reg/reg form, otherwise it's reg#0. The third field encodes the operation
427 // in bit 8, the immediate in bits 0-7. The fourth operand 9-10 encodes the
428 // index mode.
430 /// getAM3Opc - This function encodes the addrmode3 opc field.
431 inline unsigned getAM3Opc(AddrOpc Opc, unsigned char Offset,
432 unsigned IdxMode = 0) {
433 bool isSub = Opc == sub;
434 return ((int)isSub << 8) | Offset | (IdxMode << 9);
436 inline unsigned char getAM3Offset(unsigned AM3Opc) { return AM3Opc & 0xFF; }
437 inline AddrOpc getAM3Op(unsigned AM3Opc) {
438 return ((AM3Opc >> 8) & 1) ? sub : add;
440 inline unsigned getAM3IdxMode(unsigned AM3Opc) { return (AM3Opc >> 9); }
442 //===--------------------------------------------------------------------===//
443 // Addressing Mode #4
444 //===--------------------------------------------------------------------===//
446 // This is used for load / store multiple instructions.
448 // addrmode4 := reg, <mode>
450 // The four modes are:
451 // IA - Increment after
452 // IB - Increment before
453 // DA - Decrement after
454 // DB - Decrement before
455 // For VFP instructions, only the IA and DB modes are valid.
457 inline AMSubMode getAM4SubMode(unsigned Mode) {
458 return (AMSubMode)(Mode & 0x7);
461 inline unsigned getAM4ModeImm(AMSubMode SubMode) { return (int)SubMode; }
463 //===--------------------------------------------------------------------===//
464 // Addressing Mode #5
465 //===--------------------------------------------------------------------===//
467 // This is used for coprocessor instructions, such as FP load/stores.
469 // addrmode5 := reg +/- imm8*4
471 // The first operand is always a Reg. The second operand encodes the
472 // operation (add or subtract) in bit 8 and the immediate in bits 0-7.
474 /// getAM5Opc - This function encodes the addrmode5 opc field.
475 inline unsigned getAM5Opc(AddrOpc Opc, unsigned char Offset) {
476 bool isSub = Opc == sub;
477 return ((int)isSub << 8) | Offset;
479 inline unsigned char getAM5Offset(unsigned AM5Opc) { return AM5Opc & 0xFF; }
480 inline AddrOpc getAM5Op(unsigned AM5Opc) {
481 return ((AM5Opc >> 8) & 1) ? sub : add;
484 //===--------------------------------------------------------------------===//
485 // Addressing Mode #5 FP16
486 //===--------------------------------------------------------------------===//
488 // This is used for coprocessor instructions, such as 16-bit FP load/stores.
490 // addrmode5fp16 := reg +/- imm8*2
492 // The first operand is always a Reg. The second operand encodes the
493 // operation (add or subtract) in bit 8 and the immediate in bits 0-7.
495 /// getAM5FP16Opc - This function encodes the addrmode5fp16 opc field.
496 inline unsigned getAM5FP16Opc(AddrOpc Opc, unsigned char Offset) {
497 bool isSub = Opc == sub;
498 return ((int)isSub << 8) | Offset;
500 inline unsigned char getAM5FP16Offset(unsigned AM5Opc) {
501 return AM5Opc & 0xFF;
503 inline AddrOpc getAM5FP16Op(unsigned AM5Opc) {
504 return ((AM5Opc >> 8) & 1) ? sub : add;
507 //===--------------------------------------------------------------------===//
508 // Addressing Mode #6
509 //===--------------------------------------------------------------------===//
511 // This is used for NEON load / store instructions.
513 // addrmode6 := reg with optional alignment
515 // This is stored in two operands [regaddr, align]. The first is the
516 // address register. The second operand is the value of the alignment
517 // specifier in bytes or zero if no explicit alignment.
518 // Valid alignments depend on the specific instruction.
520 //===--------------------------------------------------------------------===//
521 // NEON/MVE Modified Immediates
522 //===--------------------------------------------------------------------===//
524 // Several NEON and MVE instructions (e.g., VMOV) take a "modified immediate"
525 // vector operand, where a small immediate encoded in the instruction
526 // specifies a full NEON vector value. These modified immediates are
527 // represented here as encoded integers. The low 8 bits hold the immediate
528 // value; bit 12 holds the "Op" field of the instruction, and bits 11-8 hold
529 // the "Cmode" field of the instruction. The interfaces below treat the
530 // Op and Cmode values as a single 5-bit value.
532 inline unsigned createVMOVModImm(unsigned OpCmode, unsigned Val) {
533 return (OpCmode << 8) | Val;
535 inline unsigned getVMOVModImmOpCmode(unsigned ModImm) {
536 return (ModImm >> 8) & 0x1f;
538 inline unsigned getVMOVModImmVal(unsigned ModImm) { return ModImm & 0xff; }
540 /// decodeVMOVModImm - Decode a NEON/MVE modified immediate value into the
541 /// element value and the element size in bits. (If the element size is
542 /// smaller than the vector, it is splatted into all the elements.)
543 inline uint64_t decodeVMOVModImm(unsigned ModImm, unsigned &EltBits) {
544 unsigned OpCmode = getVMOVModImmOpCmode(ModImm);
545 unsigned Imm8 = getVMOVModImmVal(ModImm);
546 uint64_t Val = 0;
548 if (OpCmode == 0xe) {
549 // 8-bit vector elements
550 Val = Imm8;
551 EltBits = 8;
552 } else if ((OpCmode & 0xc) == 0x8) {
553 // 16-bit vector elements
554 unsigned ByteNum = (OpCmode & 0x6) >> 1;
555 Val = Imm8 << (8 * ByteNum);
556 EltBits = 16;
557 } else if ((OpCmode & 0x8) == 0) {
558 // 32-bit vector elements, zero with one byte set
559 unsigned ByteNum = (OpCmode & 0x6) >> 1;
560 Val = Imm8 << (8 * ByteNum);
561 EltBits = 32;
562 } else if ((OpCmode & 0xe) == 0xc) {
563 // 32-bit vector elements, one byte with low bits set
564 unsigned ByteNum = 1 + (OpCmode & 0x1);
565 Val = (Imm8 << (8 * ByteNum)) | (0xffff >> (8 * (2 - ByteNum)));
566 EltBits = 32;
567 } else if (OpCmode == 0x1e) {
568 // 64-bit vector elements
569 for (unsigned ByteNum = 0; ByteNum < 8; ++ByteNum) {
570 if ((ModImm >> ByteNum) & 1)
571 Val |= (uint64_t)0xff << (8 * ByteNum);
573 EltBits = 64;
574 } else {
575 llvm_unreachable("Unsupported VMOV immediate");
577 return Val;
580 // Generic validation for single-byte immediate (0X00, 00X0, etc).
581 inline bool isNEONBytesplat(unsigned Value, unsigned Size) {
582 assert(Size >= 1 && Size <= 4 && "Invalid size");
583 unsigned count = 0;
584 for (unsigned i = 0; i < Size; ++i) {
585 if (Value & 0xff) count++;
586 Value >>= 8;
588 return count == 1;
591 /// Checks if Value is a correct immediate for instructions like VBIC/VORR.
592 inline bool isNEONi16splat(unsigned Value) {
593 if (Value > 0xffff)
594 return false;
595 // i16 value with set bits only in one byte X0 or 0X.
596 return Value == 0 || isNEONBytesplat(Value, 2);
599 // Encode NEON 16 bits Splat immediate for instructions like VBIC/VORR
600 inline unsigned encodeNEONi16splat(unsigned Value) {
601 assert(isNEONi16splat(Value) && "Invalid NEON splat value");
602 if (Value >= 0x100)
603 Value = (Value >> 8) | 0xa00;
604 else
605 Value |= 0x800;
606 return Value;
609 /// Checks if Value is a correct immediate for instructions like VBIC/VORR.
610 inline bool isNEONi32splat(unsigned Value) {
611 // i32 value with set bits only in one byte X000, 0X00, 00X0, or 000X.
612 return Value == 0 || isNEONBytesplat(Value, 4);
615 /// Encode NEON 32 bits Splat immediate for instructions like VBIC/VORR.
616 inline unsigned encodeNEONi32splat(unsigned Value) {
617 assert(isNEONi32splat(Value) && "Invalid NEON splat value");
618 if (Value >= 0x100 && Value <= 0xff00)
619 Value = (Value >> 8) | 0x200;
620 else if (Value > 0xffff && Value <= 0xff0000)
621 Value = (Value >> 16) | 0x400;
622 else if (Value > 0xffffff)
623 Value = (Value >> 24) | 0x600;
624 return Value;
627 //===--------------------------------------------------------------------===//
628 // Floating-point Immediates
630 inline float getFPImmFloat(unsigned Imm) {
631 // We expect an 8-bit binary encoding of a floating-point number here.
633 uint8_t Sign = (Imm >> 7) & 0x1;
634 uint8_t Exp = (Imm >> 4) & 0x7;
635 uint8_t Mantissa = Imm & 0xf;
637 // 8-bit FP IEEE Float Encoding
638 // abcd efgh aBbbbbbc defgh000 00000000 00000000
640 // where B = NOT(b);
641 uint32_t I = 0;
642 I |= Sign << 31;
643 I |= ((Exp & 0x4) != 0 ? 0 : 1) << 30;
644 I |= ((Exp & 0x4) != 0 ? 0x1f : 0) << 25;
645 I |= (Exp & 0x3) << 23;
646 I |= Mantissa << 19;
647 return bit_cast<float>(I);
650 /// getFP16Imm - Return an 8-bit floating-point version of the 16-bit
651 /// floating-point value. If the value cannot be represented as an 8-bit
652 /// floating-point value, then return -1.
653 inline int getFP16Imm(const APInt &Imm) {
654 uint32_t Sign = Imm.lshr(15).getZExtValue() & 1;
655 int32_t Exp = (Imm.lshr(10).getSExtValue() & 0x1f) - 15; // -14 to 15
656 int64_t Mantissa = Imm.getZExtValue() & 0x3ff; // 10 bits
658 // We can handle 4 bits of mantissa.
659 // mantissa = (16+UInt(e:f:g:h))/16.
660 if (Mantissa & 0x3f)
661 return -1;
662 Mantissa >>= 6;
664 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
665 if (Exp < -3 || Exp > 4)
666 return -1;
667 Exp = ((Exp+3) & 0x7) ^ 4;
669 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
672 inline int getFP16Imm(const APFloat &FPImm) {
673 return getFP16Imm(FPImm.bitcastToAPInt());
676 /// getFP32Imm - Return an 8-bit floating-point version of the 32-bit
677 /// floating-point value. If the value cannot be represented as an 8-bit
678 /// floating-point value, then return -1.
679 inline int getFP32Imm(const APInt &Imm) {
680 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
681 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
682 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
684 // We can handle 4 bits of mantissa.
685 // mantissa = (16+UInt(e:f:g:h))/16.
686 if (Mantissa & 0x7ffff)
687 return -1;
688 Mantissa >>= 19;
689 if ((Mantissa & 0xf) != Mantissa)
690 return -1;
692 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
693 if (Exp < -3 || Exp > 4)
694 return -1;
695 Exp = ((Exp+3) & 0x7) ^ 4;
697 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
700 inline int getFP32Imm(const APFloat &FPImm) {
701 return getFP32Imm(FPImm.bitcastToAPInt());
704 /// getFP64Imm - Return an 8-bit floating-point version of the 64-bit
705 /// floating-point value. If the value cannot be represented as an 8-bit
706 /// floating-point value, then return -1.
707 inline int getFP64Imm(const APInt &Imm) {
708 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
709 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
710 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffULL;
712 // We can handle 4 bits of mantissa.
713 // mantissa = (16+UInt(e:f:g:h))/16.
714 if (Mantissa & 0xffffffffffffULL)
715 return -1;
716 Mantissa >>= 48;
717 if ((Mantissa & 0xf) != Mantissa)
718 return -1;
720 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
721 if (Exp < -3 || Exp > 4)
722 return -1;
723 Exp = ((Exp+3) & 0x7) ^ 4;
725 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
728 inline int getFP64Imm(const APFloat &FPImm) {
729 return getFP64Imm(FPImm.bitcastToAPInt());
732 } // end namespace ARM_AM
733 } // end namespace llvm
735 #endif