2 * yosys -- Yosys Open SYnthesis Suite
4 * Copyright (C) 2019-2020 whitequark <whitequark@whitequark.org>
6 * Permission to use, copy, modify, and/or distribute this software for any
7 * purpose with or without fee is hereby granted.
9 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
10 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
11 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
12 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
13 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
14 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
15 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
19 // This file is included by the designs generated with `write_cxxrtl`. It is not used in Yosys itself.
28 #include <type_traits>
36 #include <backends/cxxrtl/cxxrtl_capi.h>
38 // The CXXRTL support library implements compile time specialized arbitrary width arithmetics, as well as provides
39 // composite lvalues made out of bit slices and concatenations of lvalues. This allows the `write_cxxrtl` pass
40 // to perform a straightforward translation of RTLIL structures to readable C++, relying on the C++ compiler
41 // to unwrap the abstraction and generate efficient code.
44 // All arbitrary-width values in CXXRTL are backed by arrays of unsigned integers called chunks. The chunk size
45 // is the same regardless of the value width to simplify manipulating values via FFI interfaces, e.g. driving
46 // and introspecting the simulation in Python.
48 // It is practical to use chunk sizes between 32 bits and platform register size because when arithmetics on
49 // narrower integer types is legalized by the C++ compiler, it inserts code to clear the high bits of the register.
50 // However, (a) most of our operations do not change those bits in the first place because of invariants that are
51 // invisible to the compiler, (b) we often operate on non-power-of-2 values and have to clear the high bits anyway.
52 // Therefore, using relatively wide chunks and clearing the high bits explicitly and only when we know they may be
53 // clobbered results in simpler generated code.
54 typedef uint32_t chunk_t
;
58 static_assert(std::is_integral
<T
>::value
&& std::is_unsigned
<T
>::value
,
59 "chunk type must be an unsigned integral type");
61 static constexpr size_t bits
= std::numeric_limits
<T
>::digits
;
62 static constexpr T mask
= std::numeric_limits
<T
>::max();
69 struct value
: public expr_base
<value
<Bits
>> {
70 static constexpr size_t bits
= Bits
;
72 using chunk
= chunk_traits
<chunk_t
>;
73 static constexpr chunk::type msb_mask
= (Bits
% chunk::bits
== 0) ? chunk::mask
74 : chunk::mask
>> (chunk::bits
- (Bits
% chunk::bits
));
76 static constexpr size_t chunks
= (Bits
+ chunk::bits
- 1) / chunk::bits
;
77 chunk::type data
[chunks
] = {};
80 template<typename
... Init
>
81 explicit constexpr value(Init
...init
) : data
{init
...} {}
83 value(const value
<Bits
> &) = default;
84 value(value
<Bits
> &&) = default;
85 value
<Bits
> &operator=(const value
<Bits
> &) = default;
87 // A (no-op) helper that forces the cast to value<>.
88 const value
<Bits
> &val() const {
92 std::string
str() const {
98 // Operations with compile-time parameters.
100 // These operations are used to implement slicing, concatenation, and blitting.
101 // The trunc, zext and sext operations add or remove most significant bits (i.e. on the left);
102 // the rtrunc and rzext operations add or remove least significant bits (i.e. on the right).
103 template<size_t NewBits
>
104 value
<NewBits
> trunc() const {
105 static_assert(NewBits
<= Bits
, "trunc() may not increase width");
106 value
<NewBits
> result
;
107 for (size_t n
= 0; n
< result
.chunks
; n
++)
108 result
.data
[n
] = data
[n
];
109 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
113 template<size_t NewBits
>
114 value
<NewBits
> zext() const {
115 static_assert(NewBits
>= Bits
, "zext() may not decrease width");
116 value
<NewBits
> result
;
117 for (size_t n
= 0; n
< chunks
; n
++)
118 result
.data
[n
] = data
[n
];
122 template<size_t NewBits
>
123 value
<NewBits
> sext() const {
124 static_assert(NewBits
>= Bits
, "sext() may not decrease width");
125 value
<NewBits
> result
;
126 for (size_t n
= 0; n
< chunks
; n
++)
127 result
.data
[n
] = data
[n
];
129 result
.data
[chunks
- 1] |= ~msb_mask
;
130 for (size_t n
= chunks
; n
< result
.chunks
; n
++)
131 result
.data
[n
] = chunk::mask
;
132 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
137 template<size_t NewBits
>
138 value
<NewBits
> rtrunc() const {
139 static_assert(NewBits
<= Bits
, "rtrunc() may not increase width");
140 value
<NewBits
> result
;
141 constexpr size_t shift_chunks
= (Bits
- NewBits
) / chunk::bits
;
142 constexpr size_t shift_bits
= (Bits
- NewBits
) % chunk::bits
;
143 chunk::type carry
= 0;
144 if (shift_chunks
+ result
.chunks
< chunks
) {
145 carry
= (shift_bits
== 0) ? 0
146 : data
[shift_chunks
+ result
.chunks
] << (chunk::bits
- shift_bits
);
148 for (size_t n
= result
.chunks
; n
> 0; n
--) {
149 result
.data
[n
- 1] = carry
| (data
[shift_chunks
+ n
- 1] >> shift_bits
);
150 carry
= (shift_bits
== 0) ? 0
151 : data
[shift_chunks
+ n
- 1] << (chunk::bits
- shift_bits
);
156 template<size_t NewBits
>
157 value
<NewBits
> rzext() const {
158 static_assert(NewBits
>= Bits
, "rzext() may not decrease width");
159 value
<NewBits
> result
;
160 constexpr size_t shift_chunks
= (NewBits
- Bits
) / chunk::bits
;
161 constexpr size_t shift_bits
= (NewBits
- Bits
) % chunk::bits
;
162 chunk::type carry
= 0;
163 for (size_t n
= 0; n
< chunks
; n
++) {
164 result
.data
[shift_chunks
+ n
] = (data
[n
] << shift_bits
) | carry
;
165 carry
= (shift_bits
== 0) ? 0
166 : data
[n
] >> (chunk::bits
- shift_bits
);
169 result
.data
[result
.chunks
- 1] = carry
;
173 // Bit blit operation, i.e. a partial read-modify-write.
174 template<size_t Stop
, size_t Start
>
175 value
<Bits
> blit(const value
<Stop
- Start
+ 1> &source
) const {
176 static_assert(Stop
>= Start
, "blit() may not reverse bit order");
177 constexpr chunk::type start_mask
= ~(chunk::mask
<< (Start
% chunk::bits
));
178 constexpr chunk::type stop_mask
= (Stop
% chunk::bits
+ 1 == chunk::bits
) ? 0
179 : (chunk::mask
<< (Stop
% chunk::bits
+ 1));
180 value
<Bits
> masked
= *this;
181 if (Start
/ chunk::bits
== Stop
/ chunk::bits
) {
182 masked
.data
[Start
/ chunk::bits
] &= stop_mask
| start_mask
;
184 masked
.data
[Start
/ chunk::bits
] &= start_mask
;
185 for (size_t n
= Start
/ chunk::bits
+ 1; n
< Stop
/ chunk::bits
; n
++)
187 masked
.data
[Stop
/ chunk::bits
] &= stop_mask
;
189 value
<Bits
> shifted
= source
190 .template rzext
<Stop
+ 1>()
191 .template zext
<Bits
>();
192 return masked
.bit_or(shifted
);
195 // Helpers for selecting extending or truncating operation depending on whether the result is wider or narrower
196 // than the operand. In C++17 these can be replaced with `if constexpr`.
197 template<size_t NewBits
, typename
= void>
199 value
<NewBits
> operator()(const value
<Bits
> &val
) {
200 return val
.template zext
<NewBits
>();
204 template<size_t NewBits
>
205 struct zext_cast
<NewBits
, typename
std::enable_if
<(NewBits
< Bits
)>::type
> {
206 value
<NewBits
> operator()(const value
<Bits
> &val
) {
207 return val
.template trunc
<NewBits
>();
211 template<size_t NewBits
, typename
= void>
213 value
<NewBits
> operator()(const value
<Bits
> &val
) {
214 return val
.template sext
<NewBits
>();
218 template<size_t NewBits
>
219 struct sext_cast
<NewBits
, typename
std::enable_if
<(NewBits
< Bits
)>::type
> {
220 value
<NewBits
> operator()(const value
<Bits
> &val
) {
221 return val
.template trunc
<NewBits
>();
225 template<size_t NewBits
>
226 value
<NewBits
> zcast() const {
227 return zext_cast
<NewBits
>()(*this);
230 template<size_t NewBits
>
231 value
<NewBits
> scast() const {
232 return sext_cast
<NewBits
>()(*this);
235 // Operations with run-time parameters (offsets, amounts, etc).
237 // These operations are used for computations.
238 bool bit(size_t offset
) const {
239 return data
[offset
/ chunk::bits
] & (1 << (offset
% chunk::bits
));
242 void set_bit(size_t offset
, bool value
= true) {
243 size_t offset_chunks
= offset
/ chunk::bits
;
244 size_t offset_bits
= offset
% chunk::bits
;
245 data
[offset_chunks
] &= ~(1 << offset_bits
);
246 data
[offset_chunks
] |= value
? 1 << offset_bits
: 0;
249 bool is_zero() const {
250 for (size_t n
= 0; n
< chunks
; n
++)
256 explicit operator bool() const {
260 bool is_neg() const {
261 return data
[chunks
- 1] & (1 << ((Bits
- 1) % chunk::bits
));
264 bool operator ==(const value
<Bits
> &other
) const {
265 for (size_t n
= 0; n
< chunks
; n
++)
266 if (data
[n
] != other
.data
[n
])
271 bool operator !=(const value
<Bits
> &other
) const {
272 return !(*this == other
);
275 value
<Bits
> bit_not() const {
277 for (size_t n
= 0; n
< chunks
; n
++)
278 result
.data
[n
] = ~data
[n
];
279 result
.data
[chunks
- 1] &= msb_mask
;
283 value
<Bits
> bit_and(const value
<Bits
> &other
) const {
285 for (size_t n
= 0; n
< chunks
; n
++)
286 result
.data
[n
] = data
[n
] & other
.data
[n
];
290 value
<Bits
> bit_or(const value
<Bits
> &other
) const {
292 for (size_t n
= 0; n
< chunks
; n
++)
293 result
.data
[n
] = data
[n
] | other
.data
[n
];
297 value
<Bits
> bit_xor(const value
<Bits
> &other
) const {
299 for (size_t n
= 0; n
< chunks
; n
++)
300 result
.data
[n
] = data
[n
] ^ other
.data
[n
];
304 value
<Bits
> update(const value
<Bits
> &val
, const value
<Bits
> &mask
) const {
305 return bit_and(mask
.bit_not()).bit_or(val
.bit_and(mask
));
308 template<size_t AmountBits
>
309 value
<Bits
> shl(const value
<AmountBits
> &amount
) const {
310 // Ensure our early return is correct by prohibiting values larger than 4 Gbit.
311 static_assert(Bits
<= chunk::mask
, "shl() of unreasonably large values is not supported");
312 // Detect shifts definitely large than Bits early.
313 for (size_t n
= 1; n
< amount
.chunks
; n
++)
314 if (amount
.data
[n
] != 0)
316 // Past this point we can use the least significant chunk as the shift size.
317 size_t shift_chunks
= amount
.data
[0] / chunk::bits
;
318 size_t shift_bits
= amount
.data
[0] % chunk::bits
;
319 if (shift_chunks
>= chunks
)
322 chunk::type carry
= 0;
323 for (size_t n
= 0; n
< chunks
- shift_chunks
; n
++) {
324 result
.data
[shift_chunks
+ n
] = (data
[n
] << shift_bits
) | carry
;
325 carry
= (shift_bits
== 0) ? 0
326 : data
[n
] >> (chunk::bits
- shift_bits
);
331 template<size_t AmountBits
, bool Signed
= false>
332 value
<Bits
> shr(const value
<AmountBits
> &amount
) const {
333 // Ensure our early return is correct by prohibiting values larger than 4 Gbit.
334 static_assert(Bits
<= chunk::mask
, "shr() of unreasonably large values is not supported");
335 // Detect shifts definitely large than Bits early.
336 for (size_t n
= 1; n
< amount
.chunks
; n
++)
337 if (amount
.data
[n
] != 0)
339 // Past this point we can use the least significant chunk as the shift size.
340 size_t shift_chunks
= amount
.data
[0] / chunk::bits
;
341 size_t shift_bits
= amount
.data
[0] % chunk::bits
;
342 if (shift_chunks
>= chunks
)
345 chunk::type carry
= 0;
346 for (size_t n
= 0; n
< chunks
- shift_chunks
; n
++) {
347 result
.data
[chunks
- shift_chunks
- 1 - n
] = carry
| (data
[chunks
- 1 - n
] >> shift_bits
);
348 carry
= (shift_bits
== 0) ? 0
349 : data
[chunks
- 1 - n
] << (chunk::bits
- shift_bits
);
351 if (Signed
&& is_neg()) {
352 for (size_t n
= chunks
- shift_chunks
; n
< chunks
; n
++)
353 result
.data
[n
] = chunk::mask
;
355 result
.data
[chunks
- shift_chunks
] |= chunk::mask
<< (chunk::bits
- shift_bits
);
360 template<size_t AmountBits
>
361 value
<Bits
> sshr(const value
<AmountBits
> &amount
) const {
362 return shr
<AmountBits
, /*Signed=*/true>(amount
);
365 size_t ctpop() const {
367 for (size_t n
= 0; n
< chunks
; n
++) {
368 // This loop implements the population count idiom as recognized by LLVM and GCC.
369 for (chunk::type x
= data
[n
]; x
!= 0; count
++)
375 size_t ctlz() const {
377 for (size_t n
= 0; n
< chunks
; n
++) {
378 chunk::type x
= data
[chunks
- 1 - n
];
380 count
+= (n
== 0 ? Bits
% chunk::bits
: chunk::bits
);
382 // This loop implements the find first set idiom as recognized by LLVM.
383 for (; x
!= 0; count
++)
390 template<bool Invert
, bool CarryIn
>
391 std::pair
<value
<Bits
>, bool /*CarryOut*/> alu(const value
<Bits
> &other
) const {
393 bool carry
= CarryIn
;
394 for (size_t n
= 0; n
< result
.chunks
; n
++) {
395 result
.data
[n
] = data
[n
] + (Invert
? ~other
.data
[n
] : other
.data
[n
]) + carry
;
396 carry
= (result
.data
[n
] < data
[n
]) ||
397 (result
.data
[n
] == data
[n
] && carry
);
399 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
400 return {result
, carry
};
403 value
<Bits
> add(const value
<Bits
> &other
) const {
404 return alu
</*Invert=*/false, /*CarryIn=*/false>(other
).first
;
407 value
<Bits
> sub(const value
<Bits
> &other
) const {
408 return alu
</*Invert=*/true, /*CarryIn=*/true>(other
).first
;
411 value
<Bits
> neg() const {
412 return value
<Bits
> { 0u }.sub(*this);
415 bool ucmp(const value
<Bits
> &other
) const {
417 std::tie(std::ignore
, carry
) = alu
</*Invert=*/true, /*CarryIn=*/true>(other
);
418 return !carry
; // a.ucmp(b) ≡ a u< b
421 bool scmp(const value
<Bits
> &other
) const {
424 std::tie(result
, carry
) = alu
</*Invert=*/true, /*CarryIn=*/true>(other
);
425 bool overflow
= (is_neg() == !other
.is_neg()) && (is_neg() != result
.is_neg());
426 return result
.is_neg() ^ overflow
; // a.scmp(b) ≡ a s< b
430 // Expression template for a slice, usable as lvalue or rvalue, and composable with other expression templates here.
431 template<class T
, size_t Stop
, size_t Start
>
432 struct slice_expr
: public expr_base
<slice_expr
<T
, Stop
, Start
>> {
433 static_assert(Stop
>= Start
, "slice_expr() may not reverse bit order");
434 static_assert(Start
< T::bits
&& Stop
< T::bits
, "slice_expr() must be within bounds");
435 static constexpr size_t bits
= Stop
- Start
+ 1;
439 slice_expr(T
&expr
) : expr(expr
) {}
440 slice_expr(const slice_expr
<T
, Stop
, Start
> &) = delete;
442 operator value
<bits
>() const {
443 return static_cast<const value
<T::bits
> &>(expr
)
444 .template rtrunc
<T::bits
- Start
>()
445 .template trunc
<bits
>();
448 slice_expr
<T
, Stop
, Start
> &operator=(const value
<bits
> &rhs
) {
449 // Generic partial assignment implemented using a read-modify-write operation on the sliced expression.
450 expr
= static_cast<const value
<T::bits
> &>(expr
)
451 .template blit
<Stop
, Start
>(rhs
);
455 // A helper that forces the cast to value<>, which allows deduction to work.
456 value
<bits
> val() const {
457 return static_cast<const value
<bits
> &>(*this);
461 // Expression template for a concatenation, usable as lvalue or rvalue, and composable with other expression templates here.
462 template<class T
, class U
>
463 struct concat_expr
: public expr_base
<concat_expr
<T
, U
>> {
464 static constexpr size_t bits
= T::bits
+ U::bits
;
469 concat_expr(T
&ms_expr
, U
&ls_expr
) : ms_expr(ms_expr
), ls_expr(ls_expr
) {}
470 concat_expr(const concat_expr
<T
, U
> &) = delete;
472 operator value
<bits
>() const {
473 value
<bits
> ms_shifted
= static_cast<const value
<T::bits
> &>(ms_expr
)
474 .template rzext
<bits
>();
475 value
<bits
> ls_extended
= static_cast<const value
<U::bits
> &>(ls_expr
)
476 .template zext
<bits
>();
477 return ms_shifted
.bit_or(ls_extended
);
480 concat_expr
<T
, U
> &operator=(const value
<bits
> &rhs
) {
481 ms_expr
= rhs
.template rtrunc
<T::bits
>();
482 ls_expr
= rhs
.template trunc
<U::bits
>();
486 // A helper that forces the cast to value<>, which allows deduction to work.
487 value
<bits
> val() const {
488 return static_cast<const value
<bits
> &>(*this);
492 // Base class for expression templates, providing helper methods for operations that are valid on both rvalues and lvalues.
494 // Note that expression objects (slices and concatenations) constructed in this way should NEVER be captured because
495 // they refer to temporaries that will, in general, only live until the end of the statement. For example, both of
496 // these snippets perform use-after-free:
498 // const auto &a = val.slice<7,0>().slice<1>();
501 // auto &&c = val.slice<7,0>().slice<1>();
504 // An easy way to write code using slices and concatenations safely is to follow two simple rules:
505 // * Never explicitly name any type except `value<W>` or `const value<W> &`.
506 // * Never use a `const auto &` or `auto &&` in any such expression.
507 // Then, any code that compiles will be well-defined.
510 template<size_t Stop
, size_t Start
= Stop
>
511 slice_expr
<const T
, Stop
, Start
> slice() const {
512 return {*static_cast<const T
*>(this)};
515 template<size_t Stop
, size_t Start
= Stop
>
516 slice_expr
<T
, Stop
, Start
> slice() {
517 return {*static_cast<T
*>(this)};
521 concat_expr
<const T
, typename
std::remove_reference
<const U
>::type
> concat(const U
&other
) const {
522 return {*static_cast<const T
*>(this), other
};
526 concat_expr
<T
, typename
std::remove_reference
<U
>::type
> concat(U
&&other
) {
527 return {*static_cast<T
*>(this), other
};
531 template<size_t Bits
>
532 std::ostream
&operator<<(std::ostream
&os
, const value
<Bits
> &val
) {
533 auto old_flags
= os
.flags(std::ios::right
);
534 auto old_width
= os
.width(0);
535 auto old_fill
= os
.fill('0');
536 os
<< val
.bits
<< '\'' << std::hex
;
537 for (size_t n
= val
.chunks
- 1; n
!= (size_t)-1; n
--) {
538 if (n
== val
.chunks
- 1 && Bits
% value
<Bits
>::chunk::bits
!= 0)
539 os
.width((Bits
% value
<Bits
>::chunk::bits
+ 3) / 4);
541 os
.width((value
<Bits
>::chunk::bits
+ 3) / 4);
550 template<size_t Bits
>
552 static constexpr size_t bits
= Bits
;
558 constexpr wire(const value
<Bits
> &init
) : curr(init
), next(init
) {}
559 template<typename
... Init
>
560 explicit constexpr wire(Init
...init
) : curr
{init
...}, next
{init
...} {}
562 wire(const wire
<Bits
> &) = delete;
563 wire(wire
<Bits
> &&) = default;
564 wire
<Bits
> &operator=(const wire
<Bits
> &) = delete;
575 template<size_t Bits
>
576 std::ostream
&operator<<(std::ostream
&os
, const wire
<Bits
> &val
) {
581 template<size_t Width
>
583 std::vector
<value
<Width
>> data
;
585 size_t depth() const {
590 explicit memory(size_t depth
) : data(depth
) {}
592 memory(const memory
<Width
> &) = delete;
593 memory
<Width
> &operator=(const memory
<Width
> &) = delete;
595 // The only way to get the compiler to put the initializer in .rodata and do not copy it on stack is to stuff it
596 // into a plain array. You'd think an std::initializer_list would work here, but it doesn't, because you can't
597 // construct an initializer_list in a constexpr (or something) and so if you try to do that the whole thing is
598 // first copied on the stack (probably overflowing it) and then again into `data`.
599 template<size_t Size
>
602 value
<Width
> data
[Size
];
605 template<size_t... InitSize
>
606 explicit memory(size_t depth
, const init
<InitSize
> &...init
) : data(depth
) {
608 // This utterly reprehensible construct is the most reasonable way to apply a function to every element
609 // of a parameter pack, if the elements all have different types and so cannot be cast to an initializer list.
610 auto _
= {std::move(std::begin(init
.data
), std::end(init
.data
), data
.begin() + init
.offset
)...};
613 // An operator for direct memory reads. May be used at any time during the simulation.
614 const value
<Width
> &operator [](size_t index
) const {
615 assert(index
< data
.size());
619 // An operator for direct memory writes. May only be used before the simulation is started. If used
620 // after the simulation is started, the design may malfunction.
621 value
<Width
> &operator [](size_t index
) {
622 assert(index
< data
.size());
626 // A simple way to make a writable memory would be to use an array of wires instead of an array of values.
627 // However, there are two significant downsides to this approach: first, it has large overhead (2× space
628 // overhead, and O(depth) time overhead during commit); second, it does not simplify handling write port
629 // priorities. Although in principle write ports could be ordered or conditionally enabled in generated
630 // code based on their priorities and selected addresses, the feedback arc set problem is computationally
631 // expensive, and the heuristic based algorithms are not easily modified to guarantee (rather than prefer)
632 // a particular write port evaluation order.
634 // The approach used here instead is to queue writes into a buffer during the eval phase, then perform
635 // the writes during the commit phase in the priority order. This approach has low overhead, with both space
636 // and time proportional to the amount of write ports. Because virtually every memory in a practical design
637 // has at most two write ports, linear search is used on every write, being the fastest and simplest approach.
644 std::vector
<write
> write_queue
;
646 void update(size_t index
, const value
<Width
> &val
, const value
<Width
> &mask
, int priority
= 0) {
647 assert(index
< data
.size());
648 // Queue up the write while keeping the queue sorted by priority.
650 std::upper_bound(write_queue
.begin(), write_queue
.end(), priority
,
651 [](const int a
, const write
& b
) { return a
< b
.priority
; }),
652 write
{ index
, val
, mask
, priority
});
656 bool changed
= false;
657 for (const write
&entry
: write_queue
) {
658 value
<Width
> elem
= data
[entry
.index
];
659 elem
= elem
.update(entry
.val
, entry
.mask
);
660 changed
|= (data
[entry
.index
] != elem
);
661 data
[entry
.index
] = elem
;
677 // In debug mode, using the wrong .as_*() function will assert.
678 // In release mode, using the wrong .as_*() function will safely return a default value.
680 const unsigned uint_value
= 0;
681 const signed sint_value
;
683 const std::string string_value
= "";
684 const double double_value
= 0.0;
686 metadata() : value_type(MISSING
) {}
687 metadata(unsigned value
) : value_type(UINT
), uint_value(value
) {}
688 metadata(signed value
) : value_type(SINT
), sint_value(value
) {}
689 metadata(const std::string
&value
) : value_type(STRING
), string_value(value
) {}
690 metadata(const char *value
) : value_type(STRING
), string_value(value
) {}
691 metadata(double value
) : value_type(DOUBLE
), double_value(value
) {}
693 metadata(const metadata
&) = default;
694 metadata
&operator=(const metadata
&) = delete;
696 unsigned as_uint() const {
697 assert(value_type
== UINT
);
701 signed as_sint() const {
702 assert(value_type
== SINT
);
706 const std::string
&as_string() const {
707 assert(value_type
== STRING
);
711 double as_double() const {
712 assert(value_type
== DOUBLE
);
717 typedef std::map
<std::string
, metadata
> metadata_map
;
719 // This structure is intended for consumption via foreign function interfaces, like Python's ctypes.
720 // Because of this it uses a C-style layout that is easy to parse rather than more idiomatic C++.
722 // To avoid violating strict aliasing rules, this structure has to be a subclass of the one used
723 // in the C API, or it would not be possible to cast between the pointers to these.
724 struct debug_item
: ::cxxrtl_object
{
726 VALUE
= CXXRTL_VALUE
,
728 MEMORY
= CXXRTL_MEMORY
,
731 debug_item(const ::cxxrtl_object
&object
) : cxxrtl_object(object
) {}
733 template<size_t Bits
>
734 debug_item(value
<Bits
> &item
) {
735 static_assert(sizeof(item
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
736 "value<Bits> is not compatible with C layout");
744 template<size_t Bits
>
745 debug_item(wire
<Bits
> &item
) {
746 static_assert(sizeof(item
.curr
) == value
<Bits
>::chunks
* sizeof(chunk_t
) &&
747 sizeof(item
.next
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
748 "wire<Bits> is not compatible with C layout");
752 curr
= item
.curr
.data
;
753 next
= item
.next
.data
;
756 template<size_t Width
>
757 debug_item(memory
<Width
> &item
) {
758 static_assert(sizeof(item
.data
[0]) == value
<Width
>::chunks
* sizeof(chunk_t
),
759 "memory<Width> is not compatible with C layout");
762 depth
= item
.data
.size();
763 curr
= item
.data
.empty() ? nullptr : item
.data
[0].data
;
767 static_assert(std::is_standard_layout
<debug_item
>::value
, "debug_item is not compatible with C layout");
769 typedef std::map
<std::string
, debug_item
> debug_items
;
775 module(const module
&) = delete;
776 module
&operator=(const module
&) = delete;
778 virtual bool eval() = 0;
779 virtual bool commit() = 0;
783 bool converged
= false;
787 } while (commit() && !converged
);
791 virtual void debug_info(debug_items
&items
, std::string path
= "") {}
794 } // namespace cxxrtl
796 // Internal structure used to communicate with the implementation of the C interface.
797 typedef struct _cxxrtl_toplevel
{
798 std::unique_ptr
<cxxrtl::module
> module
;
801 // Definitions of internal Yosys cells. Other than the functions in this namespace, CXXRTL is fully generic
802 // and indepenent of Yosys implementation details.
804 // The `write_cxxrtl` pass translates internal cells (cells with names that start with `$`) to calls of these
805 // functions. All of Yosys arithmetic and logical cells perform sign or zero extension on their operands,
806 // whereas basic operations on arbitrary width values require operands to be of the same width. These functions
807 // bridge the gap by performing the necessary casts. They are named similar to `cell_A[B]`, where A and B are `u`
808 // if the corresponding operand is unsigned, and `s` if it is signed.
809 namespace cxxrtl_yosys
{
811 using namespace cxxrtl
;
813 // std::max isn't constexpr until C++14 for no particular reason (it's an oversight), so we define our own.
815 constexpr T
max(const T
&a
, const T
&b
) {
816 return a
> b
? a
: b
;
820 template<size_t BitsY
, size_t BitsA
>
821 value
<BitsY
> not_u(const value
<BitsA
> &a
) {
822 return a
.template zcast
<BitsY
>().bit_not();
825 template<size_t BitsY
, size_t BitsA
>
826 value
<BitsY
> not_s(const value
<BitsA
> &a
) {
827 return a
.template scast
<BitsY
>().bit_not();
830 template<size_t BitsY
, size_t BitsA
>
831 value
<BitsY
> logic_not_u(const value
<BitsA
> &a
) {
832 return value
<BitsY
> { a
? 0u : 1u };
835 template<size_t BitsY
, size_t BitsA
>
836 value
<BitsY
> logic_not_s(const value
<BitsA
> &a
) {
837 return value
<BitsY
> { a
? 0u : 1u };
840 template<size_t BitsY
, size_t BitsA
>
841 value
<BitsY
> reduce_and_u(const value
<BitsA
> &a
) {
842 return value
<BitsY
> { a
.bit_not().is_zero() ? 1u : 0u };
845 template<size_t BitsY
, size_t BitsA
>
846 value
<BitsY
> reduce_and_s(const value
<BitsA
> &a
) {
847 return value
<BitsY
> { a
.bit_not().is_zero() ? 1u : 0u };
850 template<size_t BitsY
, size_t BitsA
>
851 value
<BitsY
> reduce_or_u(const value
<BitsA
> &a
) {
852 return value
<BitsY
> { a
? 1u : 0u };
855 template<size_t BitsY
, size_t BitsA
>
856 value
<BitsY
> reduce_or_s(const value
<BitsA
> &a
) {
857 return value
<BitsY
> { a
? 1u : 0u };
860 template<size_t BitsY
, size_t BitsA
>
861 value
<BitsY
> reduce_xor_u(const value
<BitsA
> &a
) {
862 return value
<BitsY
> { (a
.ctpop() % 2) ? 1u : 0u };
865 template<size_t BitsY
, size_t BitsA
>
866 value
<BitsY
> reduce_xor_s(const value
<BitsA
> &a
) {
867 return value
<BitsY
> { (a
.ctpop() % 2) ? 1u : 0u };
870 template<size_t BitsY
, size_t BitsA
>
871 value
<BitsY
> reduce_xnor_u(const value
<BitsA
> &a
) {
872 return value
<BitsY
> { (a
.ctpop() % 2) ? 0u : 1u };
875 template<size_t BitsY
, size_t BitsA
>
876 value
<BitsY
> reduce_xnor_s(const value
<BitsA
> &a
) {
877 return value
<BitsY
> { (a
.ctpop() % 2) ? 0u : 1u };
880 template<size_t BitsY
, size_t BitsA
>
881 value
<BitsY
> reduce_bool_u(const value
<BitsA
> &a
) {
882 return value
<BitsY
> { a
? 1u : 0u };
885 template<size_t BitsY
, size_t BitsA
>
886 value
<BitsY
> reduce_bool_s(const value
<BitsA
> &a
) {
887 return value
<BitsY
> { a
? 1u : 0u };
890 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
891 value
<BitsY
> and_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
892 return a
.template zcast
<BitsY
>().bit_and(b
.template zcast
<BitsY
>());
895 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
896 value
<BitsY
> and_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
897 return a
.template scast
<BitsY
>().bit_and(b
.template scast
<BitsY
>());
900 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
901 value
<BitsY
> or_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
902 return a
.template zcast
<BitsY
>().bit_or(b
.template zcast
<BitsY
>());
905 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
906 value
<BitsY
> or_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
907 return a
.template scast
<BitsY
>().bit_or(b
.template scast
<BitsY
>());
910 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
911 value
<BitsY
> xor_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
912 return a
.template zcast
<BitsY
>().bit_xor(b
.template zcast
<BitsY
>());
915 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
916 value
<BitsY
> xor_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
917 return a
.template scast
<BitsY
>().bit_xor(b
.template scast
<BitsY
>());
920 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
921 value
<BitsY
> xnor_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
922 return a
.template zcast
<BitsY
>().bit_xor(b
.template zcast
<BitsY
>()).bit_not();
925 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
926 value
<BitsY
> xnor_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
927 return a
.template scast
<BitsY
>().bit_xor(b
.template scast
<BitsY
>()).bit_not();
930 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
931 value
<BitsY
> logic_and_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
932 return value
<BitsY
> { (bool(a
) & bool(b
)) ? 1u : 0u };
935 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
936 value
<BitsY
> logic_and_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
937 return value
<BitsY
> { (bool(a
) & bool(b
)) ? 1u : 0u };
940 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
941 value
<BitsY
> logic_or_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
942 return value
<BitsY
> { (bool(a
) | bool(b
)) ? 1u : 0u };
945 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
946 value
<BitsY
> logic_or_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
947 return value
<BitsY
> { (bool(a
) | bool(b
)) ? 1u : 0u };
950 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
951 value
<BitsY
> shl_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
952 return a
.template zcast
<BitsY
>().template shl(b
);
955 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
956 value
<BitsY
> shl_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
957 return a
.template scast
<BitsY
>().template shl(b
);
960 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
961 value
<BitsY
> sshl_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
962 return a
.template zcast
<BitsY
>().template shl(b
);
965 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
966 value
<BitsY
> sshl_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
967 return a
.template scast
<BitsY
>().template shl(b
);
970 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
971 value
<BitsY
> shr_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
972 return a
.template shr(b
).template zcast
<BitsY
>();
975 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
976 value
<BitsY
> shr_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
977 return a
.template shr(b
).template scast
<BitsY
>();
980 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
981 value
<BitsY
> sshr_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
982 return a
.template shr(b
).template zcast
<BitsY
>();
985 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
986 value
<BitsY
> sshr_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
987 return a
.template sshr(b
).template scast
<BitsY
>();
990 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
991 value
<BitsY
> shift_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
992 return shr_uu
<BitsY
>(a
, b
);
995 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
996 value
<BitsY
> shift_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
997 return shr_su
<BitsY
>(a
, b
);
1000 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1001 value
<BitsY
> shift_us(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1002 return b
.is_neg() ? shl_uu
<BitsY
>(a
, b
.template sext
<BitsB
+ 1>().neg()) : shr_uu
<BitsY
>(a
, b
);
1005 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1006 value
<BitsY
> shift_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1007 return b
.is_neg() ? shl_su
<BitsY
>(a
, b
.template sext
<BitsB
+ 1>().neg()) : shr_su
<BitsY
>(a
, b
);
1010 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1011 value
<BitsY
> shiftx_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1012 return shift_uu
<BitsY
>(a
, b
);
1015 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1016 value
<BitsY
> shiftx_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1017 return shift_su
<BitsY
>(a
, b
);
1020 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1021 value
<BitsY
> shiftx_us(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1022 return shift_us
<BitsY
>(a
, b
);
1025 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1026 value
<BitsY
> shiftx_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1027 return shift_ss
<BitsY
>(a
, b
);
1030 // Comparison operations
1031 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1032 value
<BitsY
> eq_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1033 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1034 return value
<BitsY
>{ a
.template zext
<BitsExt
>() == b
.template zext
<BitsExt
>() ? 1u : 0u };
1037 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1038 value
<BitsY
> eq_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1039 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1040 return value
<BitsY
>{ a
.template sext
<BitsExt
>() == b
.template sext
<BitsExt
>() ? 1u : 0u };
1043 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1044 value
<BitsY
> ne_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1045 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1046 return value
<BitsY
>{ a
.template zext
<BitsExt
>() != b
.template zext
<BitsExt
>() ? 1u : 0u };
1049 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1050 value
<BitsY
> ne_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1051 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1052 return value
<BitsY
>{ a
.template sext
<BitsExt
>() != b
.template sext
<BitsExt
>() ? 1u : 0u };
1055 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1056 value
<BitsY
> eqx_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1057 return eq_uu
<BitsY
>(a
, b
);
1060 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1061 value
<BitsY
> eqx_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1062 return eq_ss
<BitsY
>(a
, b
);
1065 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1066 value
<BitsY
> nex_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1067 return ne_uu
<BitsY
>(a
, b
);
1070 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1071 value
<BitsY
> nex_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1072 return ne_ss
<BitsY
>(a
, b
);
1075 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1076 value
<BitsY
> gt_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1077 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1078 return value
<BitsY
> { b
.template zext
<BitsExt
>().ucmp(a
.template zext
<BitsExt
>()) ? 1u : 0u };
1081 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1082 value
<BitsY
> gt_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1083 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1084 return value
<BitsY
> { b
.template sext
<BitsExt
>().scmp(a
.template sext
<BitsExt
>()) ? 1u : 0u };
1087 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1088 value
<BitsY
> ge_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1089 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1090 return value
<BitsY
> { !a
.template zext
<BitsExt
>().ucmp(b
.template zext
<BitsExt
>()) ? 1u : 0u };
1093 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1094 value
<BitsY
> ge_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1095 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1096 return value
<BitsY
> { !a
.template sext
<BitsExt
>().scmp(b
.template sext
<BitsExt
>()) ? 1u : 0u };
1099 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1100 value
<BitsY
> lt_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1101 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1102 return value
<BitsY
> { a
.template zext
<BitsExt
>().ucmp(b
.template zext
<BitsExt
>()) ? 1u : 0u };
1105 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1106 value
<BitsY
> lt_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1107 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1108 return value
<BitsY
> { a
.template sext
<BitsExt
>().scmp(b
.template sext
<BitsExt
>()) ? 1u : 0u };
1111 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1112 value
<BitsY
> le_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1113 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1114 return value
<BitsY
> { !b
.template zext
<BitsExt
>().ucmp(a
.template zext
<BitsExt
>()) ? 1u : 0u };
1117 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1118 value
<BitsY
> le_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1119 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1120 return value
<BitsY
> { !b
.template sext
<BitsExt
>().scmp(a
.template sext
<BitsExt
>()) ? 1u : 0u };
1123 // Arithmetic operations
1124 template<size_t BitsY
, size_t BitsA
>
1125 value
<BitsY
> pos_u(const value
<BitsA
> &a
) {
1126 return a
.template zcast
<BitsY
>();
1129 template<size_t BitsY
, size_t BitsA
>
1130 value
<BitsY
> pos_s(const value
<BitsA
> &a
) {
1131 return a
.template scast
<BitsY
>();
1134 template<size_t BitsY
, size_t BitsA
>
1135 value
<BitsY
> neg_u(const value
<BitsA
> &a
) {
1136 return a
.template zcast
<BitsY
>().neg();
1139 template<size_t BitsY
, size_t BitsA
>
1140 value
<BitsY
> neg_s(const value
<BitsA
> &a
) {
1141 return a
.template scast
<BitsY
>().neg();
1144 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1145 value
<BitsY
> add_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1146 return a
.template zcast
<BitsY
>().add(b
.template zcast
<BitsY
>());
1149 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1150 value
<BitsY
> add_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1151 return a
.template scast
<BitsY
>().add(b
.template scast
<BitsY
>());
1154 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1155 value
<BitsY
> sub_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1156 return a
.template zcast
<BitsY
>().sub(b
.template zcast
<BitsY
>());
1159 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1160 value
<BitsY
> sub_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1161 return a
.template scast
<BitsY
>().sub(b
.template scast
<BitsY
>());
1164 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1165 value
<BitsY
> mul_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1166 value
<BitsY
> product
;
1167 value
<BitsY
> multiplicand
= a
.template zcast
<BitsY
>();
1168 const value
<BitsB
> &multiplier
= b
;
1169 uint32_t multiplicand_shift
= 0;
1170 for (size_t step
= 0; step
< BitsB
; step
++) {
1171 if (multiplier
.bit(step
)) {
1172 multiplicand
= multiplicand
.shl(value
<32> { multiplicand_shift
});
1173 product
= product
.add(multiplicand
);
1174 multiplicand_shift
= 0;
1176 multiplicand_shift
++;
1181 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1182 value
<BitsY
> mul_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1183 value
<BitsB
+ 1> ub
= b
.template sext
<BitsB
+ 1>();
1184 if (ub
.is_neg()) ub
= ub
.neg();
1185 value
<BitsY
> y
= mul_uu
<BitsY
>(a
.template scast
<BitsY
>(), ub
);
1186 return b
.is_neg() ? y
.neg() : y
;
1189 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1190 std::pair
<value
<BitsY
>, value
<BitsY
>> divmod_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1191 constexpr size_t Bits
= max(BitsY
, max(BitsA
, BitsB
));
1192 value
<Bits
> quotient
;
1193 value
<Bits
> dividend
= a
.template zext
<Bits
>();
1194 value
<Bits
> divisor
= b
.template zext
<Bits
>();
1195 if (dividend
.ucmp(divisor
))
1196 return {/*quotient=*/value
<BitsY
> { 0u }, /*remainder=*/dividend
.template trunc
<BitsY
>()};
1197 uint32_t divisor_shift
= dividend
.ctlz() - divisor
.ctlz();
1198 divisor
= divisor
.shl(value
<32> { divisor_shift
});
1199 for (size_t step
= 0; step
<= divisor_shift
; step
++) {
1200 quotient
= quotient
.shl(value
<1> { 1u });
1201 if (!dividend
.ucmp(divisor
)) {
1202 dividend
= dividend
.sub(divisor
);
1203 quotient
.set_bit(0, true);
1205 divisor
= divisor
.shr(value
<1> { 1u });
1207 return {quotient
.template trunc
<BitsY
>(), /*remainder=*/dividend
.template trunc
<BitsY
>()};
1210 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1211 std::pair
<value
<BitsY
>, value
<BitsY
>> divmod_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1212 value
<BitsA
+ 1> ua
= a
.template sext
<BitsA
+ 1>();
1213 value
<BitsB
+ 1> ub
= b
.template sext
<BitsB
+ 1>();
1214 if (ua
.is_neg()) ua
= ua
.neg();
1215 if (ub
.is_neg()) ub
= ub
.neg();
1217 std::tie(y
, r
) = divmod_uu
<BitsY
>(ua
, ub
);
1218 if (a
.is_neg() != b
.is_neg()) y
= y
.neg();
1219 if (a
.is_neg()) r
= r
.neg();
1223 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1224 value
<BitsY
> div_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1225 return divmod_uu
<BitsY
>(a
, b
).first
;
1228 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1229 value
<BitsY
> div_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1230 return divmod_ss
<BitsY
>(a
, b
).first
;
1233 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1234 value
<BitsY
> mod_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1235 return divmod_uu
<BitsY
>(a
, b
).second
;
1238 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1239 value
<BitsY
> mod_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1240 return divmod_ss
<BitsY
>(a
, b
).second
;
1244 struct memory_index
{
1248 template<size_t BitsAddr
>
1249 memory_index(const value
<BitsAddr
> &addr
, size_t offset
, size_t depth
) {
1250 static_assert(value
<BitsAddr
>::chunks
<= 1, "memory address is too wide");
1251 size_t offset_index
= addr
.data
[0];
1253 valid
= (offset_index
>= offset
&& offset_index
< offset
+ depth
);
1254 index
= offset_index
- offset
;
1258 } // namespace cxxrtl_yosys