3 Similar to VLE (but without immediate-prefixing) this encoding is designed
4 to fit on top of OpenPOWER ISA v3.0B when a "Modeswitch" bit is set (PCR
5 is recommended). Note that Compressed is *mutually exclusively incompatible*
6 with OpenPOWER v3.1B "prefixing" due to using (requiring) both EXT000
7 and EXT001. Hypothetically it could be made to use anything other than
8 EXT001, with some inconvenience (extra gates). The incompatibility is
9 "fixed" by swapping out of "Compressed" Mode and back into "Normal"
10 (v3.1B) Mode, at runtime, as needed.
12 Although initially intended to be augmented by Simple-V Prefixing (to
13 add Vector context, width overrides, e.g IEEE754 FP16, and predication) yet not put pressure on I-Cache power
14 or size, this Compressed Encoding is not critically dependent
15 *on* SV Prefixing, and may be used stand-alone.
19 * <https://bugs.libre-soc.org/show_bug.cgi?id=238>
20 * <https://ftp.libre-soc.org/VLE_314-68105.pdf> VLE Encoding
21 * <http://lists.mailinglist.openpowerfoundation.org/pipermail/openpower-hdl-cores/2020-November/000210.html>
23 This one is a conundrum. OpenPOWER ISA was never designed with 16
24 bit in mind. VLE was added 10 years ago but only by way of marking
25 an entire 64k page as "VLE". With VLE not maintained it is not
26 fully compatible with current PowerISA.
28 Here, in order to embed 16 bit into a predominantly 32 bit stream the
29 overhead of using an entire 16 bits just to switch into Compressed mode
30 is itself a significant overhead. The situation is made worse by
31 OpenPOWER ISA being fundamentally designed with 6 bits uniformly
32 taking up Major Opcode space, leaving only 10 bits to allocate
33 to actual instructions.
35 Contrast this with RVC which takes 3 out of 4
36 combinations of the first 2 bits for indicating 16-bit (anything with 0b00 to 0b10 in the LSBs), and uses the 4th (0b11) as a Huffman-style escape-sequence, easily allowing standard 32 bit and 16 bit to intermingle cleanly. To achieve the same thing on OpenPOWER would require a whopping 24 6-bit Major Opcodes which is clearly impractical: other schemes need to be devised.
38 In addition we would like to add SV-C32 which is a Vectorised version
39 of 16 bit Compressed, and ideally have a variant that adds the 27-bit
40 prefix format from SV-P64, as well.
42 Potential ways to reduce pressure on the 16 bit space are:
44 * To use more than one v3.0B Major Opcode, preferably an odd-even
46 * To provide "paging". This involves bank-switching to alternative optimised encodings for specific workloads
47 * To enter "16 bit mode" for durations specified at the start
48 * To reserve one bit of every 16 bit instruction to indicate that the 16 bit mode is to continue to be sustained
50 This latter would be useful in the Vector context to have an alternative
51 meaning: as the bit which determines whether the instruction is 11-bit
52 prefixed or 27-bit prefixed:
54 0 1 2 3 4 5 6 7 8 9 a b c d e f |
55 |major op | 11 bit vector prefix|
56 |16 bit opcode alt vec. mode ^ |
57 | extra vector prefix if alt set|
59 Using a major opcode to enter 16 bit mode, leaves 11 bits to find
60 something to use them for:
62 0 1 2 3 4 5 6 7 8 9 a b c d e f |
63 |major op | what to do here 1 |
64 |16 bit stay in 16bit mode 1 |
65 |16 bit stay in 16bit mode 1 |
66 |16 bit exit 16bit mode 0 |
68 One possibility is that the 11 bits are used for bank selection, with
69 some room for additional context such as altering the registers used
70 for the 16 bit operations (bank selection of which scalar regs).
71 However the downside is that short sequences of Compressed instructions
72 become penalised by the fixed overhead. Even a single 16 bit instruction requires a 16 bit overhead to "gain access" to 16 bit "mode", making the exercise pointless.
74 An alternative is to use the first 11 bits for only the utmost commonly used
75 instructions. That being the case then one of those 11 bits could
76 be dedicated to saying if 16 bit mode is to be continued, at which
77 point *all* 16 bits can be used for Compressed.
78 10 bits remain for actual opcodes, which is ridiculously tight,
79 however the opportunity to subsequently use all 16 bits is worth it.
81 The reason for picking 2 contiguous Major v3.0B opcodes is illustrated below:
83 |0 1 2 3 4 5 6 7 8 9 a b c d e f|
84 |major op..0| LO Half C space |
85 |major op..1| HI Half C space |
86 |N N N N N|<--11 bits C space-->|
88 If NNNNN is the same value (two contiguous Major v3.0B Opcodes) this saves gates at a critical part of the decode phase.
92 Unlike RVC, the above "context" encodings require state, to be stored in the PCR, MSR, or a dedicated SPR. These bits (just like LE/BE 32bit mode and the IEEE754 FPCSR mode) all require taking that context into consideration.
94 In particular it is critically important to recognise that context (in general) is an implicit part of the ABI implemented for example by glibc6. Therefore (in specific) Compressed Mode Context **must not** be permitted to cross into or out of a function call.
96 Thus it is the mandatory responsibility of the compiler to ensure that context returns to "v3.0B Standard" prior to entering a function call (responsibility of caller) and prior to exit from a function call (responsibility of callee).
98 Trap Handlers also take responsibility for saving and restoring of Compressed Mode state, just as they already take responsibility for other critical state. This makes traps transparent to functions as far as Compressed Mode Context is concerned, just as traps are already transparent to functions.
100 Note however that there are exceptions in a compiler to the otherwise hard rule that Compressed Mode context not be permitted to cross function boundaries: inline functions and static functions. static functions, if correctly identified as never to be called externally, may, as an optimisation, disregard standard ABIs, bearing in mind that this will be fraught (pointers to functions) and not easy to get right.
102 # Opcode Allocation Ideas
104 * one bit from the 16-bit mode is used to indicate that standard
105 (v3.0B) mode is to be dropped into for only one single instruction
106 <https://bugs.libre-soc.org/show_bug.cgi?id=238#c2>
108 ## Opcodes exploration (Attempt 1)
110 Switching between different encoding modes is controlled by M (alone)
111 in 10-bit mode, and M and N in 16-bit mode.
113 * M in 10-bit mode if zero indicates that following instructions are
114 standard OpenPOWER ISA 32-bit encoded (including, redundantly,
115 further 10/16-bit instructions)
116 * M in 10-bit mode if 1 indicates that following instructions are
117 in 16-bit encoding mode
121 * 0b01 (M=1, N=0): stay in 16-bit mode
122 * 0b00: leave 16-bit mode permanently (return to standard OpenPOWER ISA)
123 * 0b10: leave 16-bit mode for one cycle (return to standard OpenPOWER ISA)
124 * 0b11: free to be used for something completely different.
126 The current "top" idea for 0b11 is to use it for a new encoding format
127 of predominantly "immediates-based" 16-bit instructions (branch-conditional,
130 * The Compressed Major Opcode is in bits 5-7.
131 * Minor opcode in bit 8.
132 * In some cases bit 9 is taken as an additional sub-opcode, followed
133 by bits 0-4 (for CR operations)
134 * M+N mode-switching is not available for C-Major.minor 0b001.1
135 * 10 bit mode may be expanded by 16 bit mode, adding capabilities
136 that do not fit in the extreme limited space.
138 Mode-switching FSM showing relationship between v3.0B, C 10bit and C 16bit.
139 16-bit immediate mode remains in 16-bit.
141 | 0 | 1234 | 567 8 | 9abcde | f | explanation
142 | EXT000/1 | Cmaj.m | fields | 0 | 10bit then v3.0B
143 | EXT000/1 | Cmaj.m | fields | 1 | 10bit then 16bit
144 | 0 | flds | Cmaj.m | fields | 0 | 16bit then v3.0B
145 | 0 | flds | Cmaj.m | fields | 1 | 16bit then 16bit
146 | 1 | flds | Cmaj.m | fields | 0 | 16b then 1x v3.0B
147 | 1 | flds | Cmaj.m | fields | 1 | 16b/imm then 16bit
151 * Cmaj.m is the C major/minor opcode: 3 bits for major, 1 for minor
152 * EXT000 and EXT001 are v3.0B Major Opcodes. The first 5 bits
153 are zero, therefore the 6th bit is actually part of Cmaj.
154 * "10bit then 16bit" means "this instruction is encoded C 10bit
155 and the following one in C 16bit"
157 ### C Instruction Encoding types
159 10-bit Opcode formats (all start with v3.0B EXT000 or EXT001
162 | 01234 | 567 8 | 9 | a b | c | d e | f | enc
163 | E01 | Cmaj.m | fld1 | fld2 | M | 10b
164 | E01 | Cmaj.m | offset | M | 10b b
165 | E01 | 001.1 | S1 | fd1 | S2 | fd2 | M | 10b sub
166 | E01 | 111.m | fld1 | fld2 | M | 10b LDST
168 16-bit Opcode formats (including 10/16/v3.0B Switching)
170 | 0 | 1234 | 567 8 | 9 | a b | c | d e | f | enc
171 | N | immf | Cmaj.m | fld1 | fld2 | M | 16b
172 | 1 | immf | Cmaj.m | fld1 | imm | 1 | 16b imm
173 | fd3 | 001.1 | S1 | fd1 | S2 | fd2 | M | 16b sub
174 | N | fd4 | 111.m | fld1 | fld2 | M | 16b LDST
178 * fld1 and fld2 can contain reg numbers, immediates, or opcode
180 * S1 and S2 are further sub-selectors of C 001.1
182 ### Immediate Opcodes
184 only available in 16-bit mode, only available when M=1 and N=1
185 and when Cmaj.min is not 0b001.1.
187 instruction counts from objdump on /bin/bash:
203 | 0 | 1 | 2 | 3 4 | | 567.8 | 9ab | cde | f |
204 | 1 | 0 | 0 0 0 | | 001.0 | | 000 | 1 | TBD
205 | 1 | 0 | sh2 | | 001.0 | RA | sh | 1 | sradi.
206 | 1 | 1 | 0 0 0 | | 001.0 | | 000 | 1 | TBD
207 | 1 | 1 | 0 | sh2 | | 001.0 | RA | sh | 1 | srawi.
208 | 1 | 1 | 1 | | | 001.0 | | | 1 | TBD
209 | 1 | i2 | RT | | 010.0 | RA|0 | imm | 1 | addi
210 | 1 | 0 | i2 | | 010.1 | RA | imm | 1 | cmpdi
211 | 1 | 1 | i2 | | 010.1 | RA | imm | 1 | cmpwi
212 | 1 | 0 | i2 | | 011.0 | RT | imm | 1 | ldspi
213 | 1 | 1 | i2 | | 011.0 | RT | imm | 1 | lwspi
214 | 1 | 0 | i2 | | 011.1 | RT | imm | 1 | stwspi
215 | 1 | 1 | i2 | | 011.1 | RT | imm | 1 | stdspi
216 | 1 | i2 | RA | | 100.0 | RT | imm | 1 | stwi
217 | 1 | i2 | RA | | 100.1 | RT | imm | 1 | stdi
218 | 1 | i2 | RT | | 101.0 | RA | imm | 1 | ldi
219 | 1 | i2 | RT | | 101.1 | RA | imm | 1 | lwi
220 | 1 | i2 | RA | | 110.0 | RT | imm | 1 | fsti
221 | 1 | i2 | RA | | 110.1 | RT | imm | 1 | fstdi
222 | 1 | i2 | RT | | 111.0 | RA | imm | 1 | flwi
223 | 1 | i2 | RT | | 111.1 | RA | imm | 1 | fldi
225 Construction of immediate:
227 * LD/ST r1 (SP) variants should be offset by -256
228 see <https://bugs.libre-soc.org/show_bug.cgi?id=238#c43>
229 - SP variants map to e.g ld RT, imm(r1)
230 - SV Prefixing can be used to map r1 to alternate regs
231 * [1] not the same as v3.0B addis: the shift amount is smaller and actually
232 still maps to within the v3.0B addi immediate range.
233 * addi is EXTS(i2||imm) to give a 4-bit range -8 to +7
234 * addis is EXTS(i2||imm||000) to give a 11-bit range -1024 to +1023 in increments of 8
235 * all others are EXTS(i2||imm) to give a 7-bit range -128 to +127
236 (further for LD/ST due to word/dword-alignment)
240 * bc also has an immediate mode, listed separately below in Branch section
241 * for LD/ST, offset is aligned. 8-byte: i2||imm||0b000 4-byte: 0b00
242 * SV Prefix over-rides help provide alternative bitwidths for LD/ST
243 * RA|0 if RA is zero, addi. becomes "li"
244 - this only works if RT takes part of opcode
245 - mv is also possible by specifying an immediate of zero
249 Note that illeg is all zeros, including in the 16-bit mode.
250 Given that C is allocated to OpenPOWER ISA Major opcodes EXT000 and
251 EXT001 this ensures that in both 10-bit *and* 16-bit mode, a 16-bit
252 run of all zeros is considered "illegal" whilst 0b0000.0000.1000.0000
255 | 16-bit mode | | 10-bit mode |
256 | 0 | 1 | 234 | | 567.8 | 9 ab | c de | f |
257 | 0 | 0 000 | | 000.0 | 0 00 | 0 00 | 0 | illeg
258 | 0 | 0 000 | | 000.0 | 0 00 | 0 00 | 1 | nop
262 | 1 | 0 000 | | 000.0 | 0 00 | 0 00 | 0 | nop
263 | 1 | nonzero | | 000.0 | 0 00 | 0 00 | 0 | TBD
267 * All-zeros being an illegal instruction is normal for ISAs. Ensuring that
268 this remains true at all times i.e. for both 10 bit and 16 bit mode is
270 * The 10-bit nop (bit 15, M=1) is intended for circumstances
271 where alignment to 32-bit before returning to v3.0B is required.
272 M=1 being an indication "return to Standard v3.0B Encoding Mode".
273 * The 16-bit nop (bit 0, N=1) is intended for circumstances where a
274 return to Standard v3.0B Encoding is required for one cycle
275 but one cycle where alignment to a 32-bit boundary is needed.
276 Examples of this would be to return to "strict" (non-C) mode
277 where the PC may not be on a non-word-aligned boundary.
278 * If for any reason multiple 16 bit nops are needed in succession
279 the M=1 variant can be used, because each one returns to
280 Standard v3.0B Encoding Mode, each time.
282 In essence the 2 nops are needed due to there being 2 different C forms: 10 and 16 bit.
286 | 16-bit mode | | 10-bit mode |
287 | 0 | 1 | 234 | | 567.8 | 9 ab | c de | f |
288 | N | offs2 | | 000.LK | offs!=0 | M | b, bl
289 | 1 | offs2 | | 000.LK | BI | BO1 oo | 1 | bc, bcl
290 | N | BO3 BI3 | | 001.0 | LK BI | BO | M | bclr, bclrl
294 * bc only available when N,M=0b11
295 * offs2 extends offset in MSBs
296 * BI3 extends BI in MSBs to allow selection of full CR
298 * bc offset constructed from oo as LSBs and offs2 as MSBs
299 * bc BI allows selection of all bits from CR0 or CR1
300 * bc CR check is always active (as if BO0=1) therefore BO1 inverts
304 * illegal (all zeros) covers part of branch (offs=0,M=0,LK=0)
305 * nop also covers part of branch (offs=0,M=0,LK=1)
306 * bc **not available** in 10-bit mode
307 * BO[0] enables CR check, BO[1] inverts check
308 * BI refers to CR0 only (4 bits of)
309 * no Branch Conditional with immediate
310 * no Absolute Address
311 * CTR mode allowed with BO[2] for b only.
312 * offs is to 2 byte (signed) aligned
313 * all branches to 2 byte aligned
317 | 16-bit mode | | 10-bit mode |
318 | 0 | 1 | 2 3 4 | | 567.8 | 9 a b | c d e | f |
319 | RA2 | SZ | RB | | 001.1 | 1 RA | 0 RT | M | st
320 | RA2 | SZ | RB | | 001.1 | 1 RA | 1 RT | M | fst
321 | N | SZ | RT | | 111.0 | RA | RB | M | ld
322 | N | SZ | RT | | 111.1 | RA | RB | M | fld
324 * elwidth overrides can set different widths
328 * SZ=1 is 64 bit, SZ=0 is 32 bit
329 * RA2 extends RA to 3 bits (MSB)
330 * RT2 extends RT to 3 bits (MSB)
334 * RA and RB are only 2 bit (0-3)
335 * for LD, RT is implicitly RB: "ld RT=RB, RA(RB)"
336 * for ST, there is no offset: "st RT, RA(0)"
340 | 16-bit mode | | 10-bit mode |
341 | 0 | 1 | 234 | | 567.8 | 9ab | c d e | f |
342 | N | 0 | RT | | 010.0 | RB | RA!=0 | M | add
343 | N | 0 | RT | | 010.1 | RB | RA|0 | M | sub.
344 | N | 0 | BF | | 011.0 | RB | RA|0 | M | cmpl
348 * sub. and cmpl: default CR target is CR0
349 * for (RA|0) when RA=0 the input is a zero immediate,
350 meaning that sub. becomes neg. and cmp becomes cmpi against zero
351 * RT is implicitly RB: "add RT(=RB), RA, RB"
352 * Opcode 0b010.0 RA=0 is not missing from the above:
353 it is a system-wide instruction, "cbank" (section below)
357 | 0 | 1 | 234 | | 567.8 | 9ab | cde | f |
358 | N | 1 | RA | | 010.0 | RB | RS | 0 | sld.
359 | N | 1 | RA | | 010.1 | RB | RS!=0 | 0 | srd.
360 | N | 1 | RA | | 010.1 | RB | 000 | 0 | srad.
361 | N | 1 | BF | | 011.0 | RB | RA|0 | 0 | cmpw
365 * for srad, RS=RA: "srad. RA(=RS), RS, RB"
370 | 16-bit mode | | 10-bit mode |
371 | 0 | 1 | 2 3 4 | | 567.8 | 9ab | c d e | f |
372 | N | 0 | RT | | 100.0 | RB | RA!=0 | M | and
373 | N | 0 | RT | | 100.1 | RB | RA!=0 | M | nand
374 | N | 0 | RT | | 101.0 | RB | RA!=0 | M | or
375 | N | 0 | RT | | 101.1 | RB | RA!=0 | M | nor
376 | N | 0 | RT | | 100.0 | RB | 0 0 0 | M | extsw
377 | N | 0 | RT | | 100.1 | RB | 0 0 0 | M | cntlz
378 | N | 0 | RT | | 101.0 | RB | 0 0 0 | M | popcnt
379 | N | 0 | RT | | 101.1 | RB | 0 0 0 | M | not
383 | 0 | 1 | 2 3 4 | | 567.8 | 9ab | c d e | f |
384 | N | 1 | RT | | 100.0 | RB | RA!=0 | 0 | TBD
385 | N | 1 | RT | | 100.1 | RB | RA!=0 | 0 | TBD
386 | N | 1 | RT | | 101.0 | RB | RA!=0 | 0 | xor
387 | N | 1 | RT | | 101.1 | RB | RA!=0 | 0 | eqv (xnor)
388 | N | 1 | RT | | 100.0 | RB | 0 0 0 | 0 | extsb
389 | N | 1 | RT | | 100.1 | RB | 0 0 0 | 0 | cnttz
390 | N | 1 | RT | | 101.0 | RB | 0 0 0 | 0 | TBD
391 | N | 1 | RT | | 101.1 | RB | 0 0 0 | 0 | extsh
395 * for (RA|0) when RA=0 the input is a zero immediate,
396 meaning that nor becomes not
397 * cntlz, popcnt, exts **not available** in 10-bit mode
398 * RT is implicitly RB: "and RT(=RB), RA, RB"
402 Note here that elwidth overrides (SV Prefix) can be used to select FP16/32/64
404 | 16-bit mode | | 10-bit mode |
405 | 0 | 1 | 2 3 4 | | 567.8 | 9ab | c d e | f |
406 | N | | RT | | 011.1 | RB | RA!=0 | M | fsub.
407 | N | 0 | RT | | 110.0 | RB | RA!=0 | M | fadd
408 | N | 0 | RT | | 110.1 | RB | RA!=0 | M | fmul
409 | N | 0 | RT | | 011.1 | RB | 0 0 0 | M | fneg.
410 | N | 0 | RT | | 110.0 | RB | 0 0 0 | M |
411 | N | 0 | RT | | 110.1 | RB | 0 0 0 | M |
415 | 0 | 1 | 2 3 4 | | 567.8 | 9ab | c d e | f |
416 | N | 1 | RT | | 011.1 | RB | RA!=0 | 0 |
417 | N | 1 | RT | | 110.0 | RB | RA!=0 | 0 |
418 | N | 1 | RT | | 110.1 | RB | RA!=0 | 0 | fdiv
419 | N | 1 | RT | | 011.1 | RB | 0 0 0 | 0 | fabs.
420 | N | 1 | RT | | 110.0 | RB | 0 0 0 | 0 | fmr.
421 | N | 1 | RT | | 110.1 | RB | 0 0 0 | 0 |
423 16 bit only, FP to INT convert (using C 0b001.1 subencoding)
425 | 0123 | 4 | | 567.8 | 9 ab | cde | f |
426 | 0010 | X | | 001.1 | 0 RA | Y RT | M | fp2int
427 | 0011 | X | | 001.1 | 0 RA | Y RT | M | int2fp
429 * X: signed=1, unsigned=0
434 * fsub. fneg. and fmr. default target is CR1
435 * fmr. is **not available** in 10-bit mode
436 * fdiv is **not available** in 10-bit mode
440 * fmr. copies RB to RT (and sets CR1)
442 ### Condition Register
444 | 16-bit mode | | 10-bit mode |
445 | 0 1 2 3 | 4 | | 567.8 | 9 ab | cde | f |
446 | 0 0 0 0 | BF2 | | 001.1 | 0 BF | BFA | M | mcrf
447 | 0 0 0 1 | BA2 | | 001.1 | 0 BA | BB | M | crnor
448 | 0 1 0 0 | BA2 | | 001.1 | 0 BA | BB | M | crandc
449 | 0 1 1 0 | BA2 | | 001.1 | 0 BA | BB | M | crxor
450 | 0 1 1 1 | BA2 | | 001.1 | 0 BA | BB | M | crnand
451 | 1 0 0 0 | BA2 | | 001.1 | 0 BA | BB | M | crand
452 | 1 0 0 1 | BA2 | | 001.1 | 0 BA | BB | M | creqv
453 | 1 1 0 1 | BA2 | | 001.1 | 0 BA | BB | M | crorc
454 | 1 1 1 0 | BA2 | | 001.1 | 0 BA | BB | M | cror
458 * mcrf BF is only 2 bits which means the destination is only CR0-CR3
459 * CR operations: **not available** in 10-bit mode (but mcrf is)
463 * mcrf BF2 extends BF (in MSB) to 3 bits
464 * CR operations: destination register is same as BA.
465 * CR operations: only possible on CR0 and CR1
469 * CR operations: greatly extended reach/range (useful for predicates)
473 cbank: Selection of Compressed-encoding "Bank". Different "banks"
474 give different meanings to opcodes. Example: CBank=0b001 is heavily
475 optimised to A/Video Encode/Decode. cbank borrows from add's encoding
478 | 16-bit mode | | 10-bit mode |
479 | 0 | 1 2 3 4 | | 567.8 | 9ab | cde | f |
480 | N | 0 Bank2 | | 010.0 | CBank | 000 | M | cbank
482 **not available** in 10-bit mode:
484 | 0 1 2 3 | 4 | | 567.8 | 9 ab | cde | f |
485 | 1 1 1 1 | 0 | | 001.1 | 0 00 | RT | M | mtlr
486 | 1 1 1 1 | 0 | | 001.1 | 0 01 | RT | M | mtctr
487 | 1 1 1 1 | 0 | | 001.1 | 0 11 | RT | M | mtcr
488 | 1 1 1 1 | 1 | | 001.1 | 0 00 | RA | M | mflr
489 | 1 1 1 1 | 1 | | 001.1 | 0 01 | RA | M | mfctr
490 | 1 1 1 1 | 1 | | 001.1 | 0 11 | RA | M | mfcr
494 | 0 1 2 3 | 4 | | 567.8 | 9 ab | cde | f |
495 | 0 1 0 1 | | | 001.1 | 0 | | M |
496 | 1 0 1 0 | | | 001.1 | 0 | | M |
497 | 1 0 1 1 | | | 001.1 | 0 | | M |
498 | 1 1 0 0 | | | 001.1 | 0 | | M |
499 | 1 1 1 1 | | | 001.1 | 0 10 | | M |
501 ## Other ideas (Attempt 2)
503 ### 8-bit mode-switching instructions, odd addresses for C mode
505 Drop the complexity of the 16-bit encoding further reduced to 10-bit,
506 and use a single byte instead of two to switch between modes. This
507 would place compressed (C) mode instructions at odd bytes, so the LSB
508 of the PC can be used for the processor to tell which mode it is in.
510 To switch from traditional to compressed mode, the single-byte
511 instruction would be at the MSByte, that holds the EXT bits. (When we
512 break up a 32-bit instruction across words, the most significant half
513 should go in the word with the lower address.)
515 To switch from compressed mode to traditional mode, the single-byte
516 instruction would also be at the opcode/format portion, placed in the
517 lower-address word if split across words, so that the instruction can
518 be recognized as the mode-switching one without going for its second
521 The C-mode nop should be encoded so that its second byte encodes a
522 switch to compressed mode, if decoded in traditional mode. This
523 enables such a nop to straddle across a label:
525 8-bit first half of nop
527 8-bit second half of nop AKA switch to compressed mode
530 so that if traditional code jumps to the word-aligned label (because
531 traditional branches drop the 2 LSB), it immediately switches to
532 compressed mode; if we fall-through, we remain in 16-bit mode; and if
533 we branch to it from compressed mode, whether we jump to the odd or
534 the even address, we end up in compressed mode as desired.
536 ### Use 2- rather than 3-register opcodes
538 Successful compact ISAs have used 2- rather than 3-register insns, in
539 which the same register serves as input and output. Some 20% of
540 general-purpose 3-register insns already use either input register as
541 output, without any effort by the compiler to do so.
543 Repurposing the 3 bits used to encode one one of the input registers
544 in arithmetic, logical and floating-pointer registers, and the 2 bits
545 used to encode the mode of the next two insns, we could make the full
546 register files available to the opcodes already selected for
547 compressed mode, with one bit to spare to bring additional opcodes in.
549 An opcode could be assigned to an instruction that combines and
550 extends with the subsequent instruction, providing it with a separate
551 input operand to use rather than the output register, or with
552 additional range for immediate and offset operands, effectively
553 forming a 32-bit operation, enabling us to remain in compressed mode
557 # Analysis techniques and tools
559 objdump -d --no-show-raw-insn /bin/bash | sed 'y/\t/ /;
560 s/^[ x0-9A-F]*: *\([a-z.]\+\) *\(.*\)/\1 \2 /p; d' |
561 sed 's/\([, (]\)r[1-9][0-9]*/\1r1/g;
562 s/\([ ,]\)-*[0-9]\+\([^0-9]\)/\11\2/g' | sort | uniq --count |
565 # gcc register allocation
567 FTR, information extracted from gcc's gcc/config/rs6000/rs6000.h about fixed registers (assigned to special purposes) and register allocation order:
569 Special-purpose registers on ppc are:
571 r0: constant zero/throw-away
573 r2: thread-local storage pointer in 32-bit mode
574 r2: non-minimal TOC register
575 r10: EH return stack adjust register
576 r11: static chain pointer
577 r13: thread-local storage pointer in 64-bit mode
578 r30: minimal-TOC/-fPIC/-fpic base register
580 lr: return address register
582 the register allocation order in GCC (i.e., it takes the earliest available register that fits the constraints) is:
584 We allocate in the following order:
585 fp0 (not saved or used for anything)
586 fp13 - fp2 (not saved; incoming fp arg registers)
587 fp1 (not saved; return value)
588 fp31 - fp14 (saved; order given to save least number)
589 cr7, cr5 (not saved or special)
590 cr6 (not saved, but used for vector operations)
591 cr1 (not saved, but used for FP operations)
592 cr0 (not saved, but used for arithmetic operations)
593 cr4, cr3, cr2 (saved)
594 r9 (not saved; best for TImode)
595 r10, r8-r4 (not saved; highest first for less conflict with params)
596 r3 (not saved; return value register)
597 r11 (not saved; later alloc to help shrink-wrap)
598 r0 (not saved; cannot be base reg)
599 r31 - r13 (saved; order given to save least number)
600 r12 (not saved; if used for DImode or DFmode would use r13)
601 ctr (not saved; when we have the choice ctr is better)
603 r1, r2, ap, ca (fixed)
604 v0 - v1 (not saved or used for anything)
605 v13 - v3 (not saved; incoming vector arg registers)
606 v2 (not saved; incoming vector arg reg; return value)
607 v19 - v14 (not saved or used for anything)
608 v31 - v20 (saved; order given to save least number)