]>
Commit | Line | Data |
---|---|---|
8410904a NC |
1 | ; SF format is: |
2 | ; | |
3 | ; [sign] 1.[23bits] E[8bits(n-127)] | |
4 | ; | |
5 | ; SEEEEEEE Emmmmmmm mmmmmmmm mmmmmmmm | |
6 | ; | |
7 | ; [A+0] mmmmmmmm | |
8 | ; [A+1] mmmmmmmm | |
9 | ; [A+2] Emmmmmmm | |
10 | ; [A+3] SEEEEEEE | |
11 | ; | |
12 | ; Special values (xxx != 0): | |
13 | ; | |
14 | ; s1111111 10000000 00000000 00000000 infinity | |
15 | ; s1111111 1xxxxxxx xxxxxxxx xxxxxxxx NaN | |
16 | ; s0000000 00000000 00000000 00000000 zero | |
17 | ; s0000000 0xxxxxxx xxxxxxxx xxxxxxxx denormals | |
18 | ; | |
19 | ; Note that CMPtype is "signed char" for rl78 | |
20 | ; | |
21 | ||
22 | #include "vregs.h" | |
23 | ||
24 | #define Z PSW.6 | |
25 | ||
26 | START_FUNC ___negsf2 | |
27 | ||
28 | ;; Negate the floating point value. | |
29 | ;; Input at [SP+4]..[SP+7]. | |
30 | ;; Output to R8..R11. | |
31 | ||
32 | movw ax, [SP+4] | |
33 | movw r8, ax | |
34 | movw ax, [SP+6] | |
35 | xor a, #0x80 | |
36 | movw r10, ax | |
37 | ret | |
38 | ||
39 | END_FUNC ___negsf2 | |
40 | ||
41 | ;; ------------------internal functions used by later code -------------- | |
42 | ||
43 | START_FUNC __int_isnan | |
44 | ||
45 | ;; [HL] points to value, returns Z if it's a NaN | |
46 | ||
47 | mov a, [hl+2] | |
48 | and a, #0x80 | |
49 | mov x, a | |
50 | mov a, [hl+3] | |
51 | and a, #0x7f | |
52 | cmpw ax, #0x7f80 | |
53 | skz | |
54 | ret ; return NZ if not NaN | |
55 | mov a, [hl+2] | |
56 | and a, #0x7f | |
57 | or a, [hl+1] | |
58 | or a, [hl] | |
59 | bnz $1f | |
60 | clr1 Z ; Z, normal | |
61 | ret | |
62 | 1: | |
63 | set1 Z ; nan | |
64 | ret | |
65 | ||
66 | END_FUNC __int_isnan | |
67 | ||
68 | START_FUNC __int_eithernan | |
69 | ||
70 | ;; call from toplevel functions, returns Z if either number is a NaN, | |
71 | ;; or NZ if both are OK. | |
72 | ||
73 | movw ax, sp | |
74 | addw ax, #8 | |
75 | movw hl, ax | |
76 | call $!__int_isnan | |
77 | bz $1f | |
78 | ||
79 | movw ax, sp | |
80 | addw ax, #12 | |
81 | movw hl, ax | |
82 | call $!__int_isnan | |
83 | 1: | |
84 | ret | |
85 | ||
86 | END_FUNC __int_eithernan | |
87 | ||
88 | START_FUNC __int_iszero | |
89 | ||
90 | ;; [HL] points to value, returns Z if it's zero | |
91 | ||
92 | mov a, [hl+3] | |
93 | and a, #0x7f | |
94 | or a, [hl+2] | |
95 | or a, [hl+1] | |
96 | or a, [hl] | |
97 | ret | |
98 | ||
99 | END_FUNC __int_iszero | |
100 | ||
101 | START_FUNC __int_cmpsf | |
102 | ||
103 | ;; This is always called from some other function here, | |
104 | ;; so the stack offsets are adjusted accordingly. | |
105 | ||
106 | ;; X [SP+8] <=> Y [SP+12] : <a> <=> 0 | |
107 | ||
108 | movw ax, sp | |
109 | addw ax, #8 | |
110 | movw hl, ax | |
111 | call $!__int_iszero | |
112 | bnz $1f | |
113 | ||
114 | movw ax, sp | |
115 | addw ax, #12 | |
116 | movw hl, ax | |
117 | call $!__int_iszero | |
118 | bnz $2f | |
119 | ;; At this point, both args are zero. | |
120 | mov a, #0 | |
121 | ret | |
122 | ||
123 | 2: | |
124 | movw ax, sp | |
125 | addw ax, #8 | |
126 | movw hl, ax | |
127 | 1: | |
128 | ;; At least one arg is non-zero so we can just compare magnitudes. | |
129 | ;; Args are [HL] and [HL+4]. | |
130 | ||
131 | mov a, [HL+3] | |
132 | xor a, [HL+7] | |
133 | mov1 cy, a.7 | |
134 | bnc $1f | |
135 | ||
136 | mov a, [HL+3] | |
137 | sar a, 7 | |
138 | or a, #1 | |
139 | ret | |
140 | ||
141 | 1: ;; Signs the same, compare magnitude. It's safe to lump | |
142 | ;; the sign bits, exponent, and mantissa together here, since they're | |
143 | ;; stored in the right sequence. | |
144 | movw ax, [HL+2] | |
145 | cmpw ax, [HL+6] | |
146 | bc $ybig_cmpsf ; branch if X < Y | |
147 | bnz $xbig_cmpsf ; branch if X > Y | |
148 | ||
149 | movw ax, [HL] | |
150 | cmpw ax, [HL+4] | |
151 | bc $ybig_cmpsf ; branch if X < Y | |
152 | bnz $xbig_cmpsf ; branch if X > Y | |
153 | ||
154 | mov a, #0 | |
155 | ret | |
156 | ||
157 | xbig_cmpsf: ; |X| > |Y| so return A = 1 if pos, 0xff if neg | |
158 | mov a, [HL+3] | |
159 | sar a, 7 | |
160 | or a, #1 | |
161 | ret | |
162 | ybig_cmpsf: ; |X| < |Y| so return A = 0xff if pos, 1 if neg | |
163 | mov a, [HL+3] | |
164 | xor a, #0x80 | |
165 | sar a, 7 | |
166 | or a, #1 | |
167 | ret | |
168 | ||
169 | END_FUNC __int_cmpsf | |
170 | ||
171 | ;; ---------------------------------------------------------- | |
172 | ||
173 | START_FUNC ___cmpsf2 | |
174 | ;; This functions calculates "A <=> B". That is, if A is less than B | |
175 | ;; they return -1, if A is greater than B, they return 1, and if A | |
176 | ;; and B are equal they return 0. If either argument is NaN the | |
177 | ;; behaviour is undefined. | |
178 | ||
179 | ;; Input at [SP+4]..[SP+7]. | |
180 | ;; Output to R8..R9. | |
181 | ||
182 | call $!__int_eithernan | |
183 | bnz $1f | |
184 | movw r8, #1 | |
185 | ret | |
186 | 1: | |
187 | call $!__int_cmpsf | |
188 | mov r8, a | |
189 | sar a, 7 | |
190 | mov r9, a | |
191 | ret | |
192 | ||
193 | END_FUNC ___cmpsf2 | |
194 | ||
195 | ;; ---------------------------------------------------------- | |
196 | ||
197 | ;; These functions are all basically the same as ___cmpsf2 | |
198 | ;; except that they define how they handle NaNs. | |
199 | ||
200 | START_FUNC ___eqsf2 | |
201 | ;; Returns zero iff neither argument is NaN | |
202 | ;; and both arguments are equal. | |
203 | START_ANOTHER_FUNC ___nesf2 | |
204 | ;; Returns non-zero iff either argument is NaN or the arguments are | |
205 | ;; unequal. Effectively __nesf2 is the same as __eqsf2 | |
206 | START_ANOTHER_FUNC ___lesf2 | |
207 | ;; Returns a value less than or equal to zero if neither | |
208 | ;; argument is NaN, and the first is less than or equal to the second. | |
209 | START_ANOTHER_FUNC ___ltsf2 | |
210 | ;; Returns a value less than zero if neither argument is | |
211 | ;; NaN, and the first is strictly less than the second. | |
212 | ||
213 | ;; Input at [SP+4]..[SP+7]. | |
214 | ;; Output to R8. | |
215 | ||
216 | mov r8, #1 | |
217 | ||
218 | ;;; Fall through | |
219 | ||
220 | START_ANOTHER_FUNC __int_cmp_common | |
221 | ||
222 | call $!__int_eithernan | |
223 | sknz | |
224 | ;; return value (pre-filled-in below) for "either is nan" | |
225 | ret | |
226 | ||
227 | call $!__int_cmpsf | |
228 | mov r8, a | |
229 | ret | |
230 | ||
231 | END_ANOTHER_FUNC __int_cmp_common | |
232 | END_ANOTHER_FUNC ___ltsf2 | |
233 | END_ANOTHER_FUNC ___lesf2 | |
234 | END_ANOTHER_FUNC ___nesf2 | |
235 | END_FUNC ___eqsf2 | |
236 | ||
237 | START_FUNC ___gesf2 | |
238 | ;; Returns a value greater than or equal to zero if neither argument | |
239 | ;; is a NaN and the first is greater than or equal to the second. | |
240 | START_ANOTHER_FUNC ___gtsf2 | |
241 | ;; Returns a value greater than zero if neither argument | |
242 | ;; is NaN, and the first is strictly greater than the second. | |
243 | ||
244 | mov r8, #0xffff | |
245 | br $__int_cmp_common | |
246 | ||
247 | END_ANOTHER_FUNC ___gtsf2 | |
248 | END_FUNC ___gesf2 | |
249 | ||
250 | ;; ---------------------------------------------------------- | |
251 | ||
252 | START_FUNC ___unordsf2 | |
253 | ;; Returns a nonzero value if either argument is NaN, otherwise 0. | |
254 | ||
255 | call $!__int_eithernan | |
256 | movw r8, #0 | |
257 | sknz ; this is from the call, not the movw | |
258 | movw r8, #1 | |
259 | ret | |
260 | ||
261 | END_FUNC ___unordsf2 | |
262 | ||
263 | ;; ---------------------------------------------------------- | |
264 | ||
265 | START_FUNC ___fixsfsi | |
266 | ;; Converts its floating point argument into a signed long, | |
267 | ;; rounding toward zero. | |
268 | ;; The behaviour with NaNs and Infinities is not well defined. | |
269 | ;; We choose to return 0 for NaNs, -INTMAX for -inf and INTMAX for +inf. | |
270 | ;; This matches the behaviour of the C function in libgcc2.c. | |
271 | ||
272 | ;; Input at [SP+4]..[SP+7], result is in (lsb) R8..R11 (msb). | |
273 | ||
274 | ;; Special case handling for infinities as __fixunssfsi | |
275 | ;; will not give us the values that we want. | |
276 | movw ax, sp | |
277 | addw ax, #4 | |
278 | movw hl, ax | |
279 | call !!__int_isinf | |
280 | bnz $1f | |
281 | mov a, [SP+7] | |
282 | bt a.7, $2f | |
283 | ;; +inf | |
284 | movw r8, #-1 | |
285 | movw r10, #0x7fff | |
286 | ret | |
287 | ;; -inf | |
288 | 2: mov r8, #0 | |
289 | mov r10, #0x8000 | |
290 | ret | |
291 | ||
292 | ;; Load the value into r10:r11:X:A | |
293 | 1: movw ax, [SP+4] | |
294 | movw r10, ax | |
295 | movw ax, [SP+6] | |
296 | ||
297 | ;; If the value is positive we can just use __fixunssfsi | |
298 | bf a.7, $__int_fixunssfsi | |
299 | ||
300 | ;; Otherwise we negate the value, call __fixunssfsi and | |
301 | ;; then negate its result. | |
302 | clr1 a.7 | |
303 | call $!__int_fixunssfsi | |
304 | ||
305 | movw ax, #0 | |
306 | subw ax, r8 | |
307 | movw r8, ax | |
308 | movw ax, #0 | |
309 | sknc | |
310 | decw ax | |
311 | subw ax, r10 | |
312 | movw r10, ax | |
313 | ||
314 | ;; Check for a positive result (which should only happen when | |
315 | ;; __fixunssfsi returns UINTMAX or 0). In such cases just return 0. | |
316 | mov a, r11 | |
317 | bt a.7, $1f | |
318 | movw r10,#0x0 | |
319 | movw r8, #0x0 | |
320 | ||
321 | 1: ret | |
322 | ||
323 | END_FUNC ___fixsfsi | |
324 | ||
325 | START_FUNC ___fixunssfsi | |
326 | ;; Converts its floating point argument into an unsigned long | |
327 | ;; rounding towards zero. Negative arguments all become zero. | |
328 | ;; We choose to return 0 for NaNs and -inf, but UINTMAX for +inf. | |
329 | ;; This matches the behaviour of the C function in libgcc2.c. | |
330 | ||
331 | ;; Input at [SP+4]..[SP+7], result is in (lsb) R8..R11 (msb) | |
332 | ||
333 | ;; Get the input value. | |
334 | movw ax, [SP+4] | |
335 | movw r10, ax | |
336 | movw ax, [SP+6] | |
337 | ||
338 | ;; Fall through into the internal function. | |
339 | ||
340 | .global __int_fixunssfsi | |
341 | __int_fixunssfsi: | |
342 | ;; Input in (lsb) r10.r11.x.a (msb). | |
343 | ||
344 | ;; Test for a negative input. We shift the other bits at the | |
345 | ;; same time so that A ends up holding the whole exponent: | |
346 | ;; | |
347 | ;; before: | |
348 | ;; SEEEEEEE EMMMMMMM MMMMMMMM MMMMMMMM | |
349 | ;; A X R11 R10 | |
350 | ;; | |
351 | ;; after: | |
352 | ;; EEEEEEEE MMMMMMM0 MMMMMMMM MMMMMMMM | |
353 | ;; A X R11 R10 | |
354 | shlw ax, 1 | |
355 | bnc $1f | |
356 | ||
357 | ;; Return zero. | |
358 | 2: movw r8, #0 | |
359 | movw r10, #0 | |
360 | ret | |
361 | ||
362 | ;; An exponent of -1 is either a NaN or infinity. | |
363 | 1: cmp a, #-1 | |
364 | bnz $3f | |
365 | ;; For NaN we return 0. For infinity we return UINTMAX. | |
366 | mov a, x | |
367 | or a, r10 | |
368 | or a, r11 | |
369 | cmp0 a | |
370 | bnz $2b | |
371 | ||
372 | 6: movw r8, #-1 ; -1 => UINT_MAX | |
373 | movw r10, #-1 | |
374 | ret | |
375 | ||
376 | ;; If the exponent is negative the value is < 1 and so the | |
377 | ;; converted value is 0. Note we must allow for the bias | |
378 | ;; applied to the exponent. Thus a value of 127 in the | |
379 | ;; EEEEEEEE bits actually represents an exponent of 0, whilst | |
380 | ;; a value less than 127 actually represents a negative exponent. | |
381 | ;; Also if the EEEEEEEE bits are all zero then this represents | |
382 | ;; either a denormal value or 0.0. Either way for these values | |
383 | ;; we return 0. | |
384 | 3: sub a, #127 | |
385 | bc $2b | |
386 | ||
387 | ;; A now holds the bias adjusted exponent, which is known to be >= 0. | |
388 | ;; If the exponent is > 31 then the conversion will overflow. | |
389 | cmp a, #32 | |
390 | bnc $6b | |
391 | 4: | |
392 | ;; Save the exponent in H. We increment it by one because we want | |
393 | ;; to be sure that the loop below will always execute at least once. | |
394 | inc a | |
395 | mov h, a | |
396 | ||
397 | ;; Get the top 24 bits of the mantissa into A:X:R10 | |
398 | ;; Include the implicit 1-bit that is inherent in the IEEE fp format. | |
399 | ;; | |
400 | ;; before: | |
401 | ;; EEEEEEEE MMMMMMM0 MMMMMMMM MMMMMMMM | |
402 | ;; H X R11 R10 | |
403 | ;; after: | |
404 | ;; EEEEEEEE 1MMMMMMM MMMMMMMM MMMMMMMM | |
405 | ;; H A X R10 | |
406 | ||
407 | mov a, r11 | |
408 | xch a, x | |
409 | shr a, 1 | |
410 | set1 a.7 | |
411 | ||
412 | ;; Clear B:C:R12:R13 | |
413 | movw bc, #0 | |
414 | movw r12, #0 | |
415 | ||
416 | ;; Shift bits from the mantissa (A:X:R10) into (B:C:R12:R13), | |
417 | ;; decrementing the exponent as we go. | |
418 | ||
419 | ;; before: | |
420 | ;; MMMMMMMM MMMMMMMM MMMMMMMM xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx | |
421 | ;; A X R10 B C R12 R13 | |
422 | ;; first iter: | |
423 | ;; MMMMMMMM MMMMMMMM MMMMMMM0 xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxM | |
424 | ;; A X R10 B C R12 R13 | |
425 | ;; second iter: | |
426 | ;; MMMMMMMM MMMMMMMM MMMMMM00 xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxMM | |
427 | ;; A X R10 B C R12 R13 | |
428 | ;; etc. | |
429 | 5: | |
430 | xch a, r10 | |
431 | shl a, 1 | |
432 | xch a, r10 | |
433 | ||
434 | rolwc ax, 1 | |
435 | ||
436 | xch a, r13 | |
437 | rolc a, 1 | |
438 | xch a, r13 | |
439 | ||
440 | xch a, r12 | |
441 | rolc a, 1 | |
442 | xch a, r12 | |
443 | ||
444 | rolwc bc, 1 | |
445 | ||
446 | dec h | |
447 | bnz $5b | |
448 | ||
449 | ;; Result is currently in (lsb) r13.r12. c. b. (msb), | |
450 | ;; Move it into (lsb) r8. r9. r10. r11 (msb). | |
451 | ||
452 | mov a, r13 | |
453 | mov r8, a | |
454 | ||
455 | mov a, r12 | |
456 | mov r9, a | |
457 | ||
458 | mov a, c | |
459 | mov r10, a | |
460 | ||
461 | mov a, b | |
462 | mov r11, a | |
463 | ||
464 | ret | |
465 | ||
466 | END_FUNC ___fixunssfsi | |
467 | ||
468 | ;; ------------------------------------------------------------------------ | |
469 | ||
470 | START_FUNC ___floatsisf | |
471 | ;; Converts its signed long argument into a floating point. | |
472 | ;; Argument in [SP+4]..[SP+7]. Result in R8..R11. | |
473 | ||
474 | ;; Get the argument. | |
475 | movw ax, [SP+4] | |
476 | movw bc, ax | |
477 | movw ax, [SP+6] | |
478 | ||
479 | ;; Test the sign bit. If the value is positive then drop into | |
480 | ;; the unsigned conversion routine. | |
481 | bf a.7, $2f | |
482 | ||
483 | ;; If negative convert to positive ... | |
484 | movw hl, ax | |
485 | movw ax, #0 | |
486 | subw ax, bc | |
487 | movw bc, ax | |
488 | movw ax, #0 | |
489 | sknc | |
490 | decw ax | |
491 | subw ax, hl | |
492 | ||
493 | ;; If the result is negative then the input was 0x80000000 and | |
494 | ;; we want to return -0.0, which will not happen if we call | |
495 | ;; __int_floatunsisf. | |
496 | bt a.7, $1f | |
497 | ||
498 | ;; Call the unsigned conversion routine. | |
499 | call $!__int_floatunsisf | |
500 | ||
501 | ;; Negate the result. | |
502 | set1 r11.7 | |
503 | ||
504 | ;; Done. | |
505 | ret | |
506 | ||
507 | 1: ;; Return -0.0 aka 0xcf000000 | |
508 | ||
509 | clrb a | |
510 | mov r8, a | |
511 | mov r9, a | |
512 | mov r10, a | |
513 | mov a, #0xcf | |
514 | mov r11, a | |
515 | ret | |
516 | ||
517 | START_ANOTHER_FUNC ___floatunsisf | |
518 | ;; Converts its unsigned long argument into a floating point. | |
519 | ;; Argument in [SP+4]..[SP+7]. Result in R8..R11. | |
520 | ||
521 | ;; Get the argument. | |
522 | movw ax, [SP+4] | |
523 | movw bc, ax | |
524 | movw ax, [SP+6] | |
525 | ||
526 | 2: ;; Internal entry point from __floatsisf | |
527 | ;; Input in AX (high) and BC (low) | |
528 | .global __int_floatunsisf | |
529 | __int_floatunsisf: | |
530 | ||
531 | ;; Special case handling for zero. | |
532 | cmpw ax, #0 | |
533 | bnz $1f | |
534 | movw ax, bc | |
535 | cmpw ax, #0 | |
536 | movw ax, #0 | |
537 | bnz $1f | |
538 | ||
539 | ;; Return 0.0 | |
540 | movw r8, ax | |
541 | movw r10, ax | |
542 | ret | |
543 | ||
544 | 1: ;; Pre-load the loop count/exponent. | |
545 | ;; Exponents are biased by 0x80 and we start the loop knowing that | |
546 | ;; we are going to skip the highest set bit. Hence the highest value | |
547 | ;; that we can get for the exponent is 0x1e (bits from input) + 0x80 = 0x9e. | |
548 | mov h, #0x9e | |
549 | ||
550 | ;; Move bits off the top of AX:BC until we hit a 1 bit. | |
551 | ;; Decrement the count of remaining bits as we go. | |
552 | ||
553 | 2: shlw bc, 1 | |
554 | rolwc ax, 1 | |
555 | bc $3f | |
556 | dec h | |
557 | br $2b | |
558 | ||
559 | ;; Ignore the first one bit - it is implicit in the IEEE format. | |
560 | ;; The count of remaining bits is the exponent. | |
561 | ||
562 | ;; Assemble the final floating point value. We have... | |
563 | ;; before: | |
564 | ;; EEEEEEEE MMMMMMMM MMMMMMMM MMMMMMMM xxxxxxxx | |
565 | ;; H A X B C | |
566 | ;; after: | |
567 | ;; 0EEEEEEE EMMMMMMM MMMMMMMM MMMMMMMM | |
568 | ;; R11 R10 R9 R8 | |
569 | ||
570 | ||
571 | 3: shrw ax, 1 | |
572 | mov r10, a | |
573 | mov a, x | |
574 | mov r9, a | |
575 | ||
576 | mov a, b | |
577 | rorc a, 1 | |
578 | ||
579 | ;; If the bottom bit of B was set before we shifted it out then we | |
580 | ;; need to round the result up. Unless none of the bits in C are set. | |
581 | ;; In this case we are exactly half-way between two values, and we | |
582 | ;; round towards an even value. We round up by increasing the | |
583 | ;; mantissa by 1. If this results in a zero mantissa we have to | |
584 | ;; increment the exponent. We round down by ignoring the dropped bits. | |
585 | ||
586 | bnc $4f | |
587 | cmp0 c | |
588 | sknz | |
589 | bf a.0, $4f | |
590 | ||
591 | 5: ;; Round the mantissa up by 1. | |
592 | add a, #1 | |
593 | addc r9, #0 | |
594 | addc r10, #0 | |
595 | bf r10.7, $4f | |
596 | inc h | |
597 | clr1 r10.7 | |
598 | ||
599 | 4: mov r8, a | |
600 | mov a, h | |
601 | shr a, 1 | |
602 | mov r11, a | |
603 | sknc | |
604 | set1 r10.7 | |
605 | ret | |
606 | ||
607 | END_ANOTHER_FUNC ___floatunsisf | |
608 | END_FUNC ___floatsisf |