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.
21 // The CXXRTL support library implements compile time specialized arbitrary width arithmetics, as well as provides
22 // composite lvalues made out of bit slices and concatenations of lvalues. This allows the `write_cxxrtl` pass
23 // to perform a straightforward translation of RTLIL structures to readable C++, relying on the C++ compiler
24 // to unwrap the abstraction and generate efficient code.
33 #include <type_traits>
41 #include <backends/cxxrtl/cxxrtl_capi.h>
43 // CXXRTL essentially uses the C++ compiler as a hygienic macro engine that feeds an instruction selector.
44 // It generates a lot of specialized template functions with relatively large bodies that, when inlined
45 // into the caller and (for those with loops) unrolled, often expose many new optimization opportunities.
46 // Because of this, most of the CXXRTL runtime must be always inlined for best performance.
47 #ifndef __has_attribute
48 # define __has_attribute(x) 0
50 #if __has_attribute(always_inline)
51 #define CXXRTL_ALWAYS_INLINE inline __attribute__((__always_inline__))
53 #define CXXRTL_ALWAYS_INLINE inline
58 // All arbitrary-width values in CXXRTL are backed by arrays of unsigned integers called chunks. The chunk size
59 // is the same regardless of the value width to simplify manipulating values via FFI interfaces, e.g. driving
60 // and introspecting the simulation in Python.
62 // It is practical to use chunk sizes between 32 bits and platform register size because when arithmetics on
63 // narrower integer types is legalized by the C++ compiler, it inserts code to clear the high bits of the register.
64 // However, (a) most of our operations do not change those bits in the first place because of invariants that are
65 // invisible to the compiler, (b) we often operate on non-power-of-2 values and have to clear the high bits anyway.
66 // Therefore, using relatively wide chunks and clearing the high bits explicitly and only when we know they may be
67 // clobbered results in simpler generated code.
68 typedef uint32_t chunk_t
;
72 static_assert(std::is_integral
<T
>::value
&& std::is_unsigned
<T
>::value
,
73 "chunk type must be an unsigned integral type");
75 static constexpr size_t bits
= std::numeric_limits
<T
>::digits
;
76 static constexpr T mask
= std::numeric_limits
<T
>::max();
83 struct value
: public expr_base
<value
<Bits
>> {
84 static constexpr size_t bits
= Bits
;
86 using chunk
= chunk_traits
<chunk_t
>;
87 static constexpr chunk::type msb_mask
= (Bits
% chunk::bits
== 0) ? chunk::mask
88 : chunk::mask
>> (chunk::bits
- (Bits
% chunk::bits
));
90 static constexpr size_t chunks
= (Bits
+ chunk::bits
- 1) / chunk::bits
;
91 chunk::type data
[chunks
] = {};
94 template<typename
... Init
>
95 explicit constexpr value(Init
...init
) : data
{init
...} {}
97 value(const value
<Bits
> &) = default;
98 value(value
<Bits
> &&) = default;
99 value
<Bits
> &operator=(const value
<Bits
> &) = default;
101 // A (no-op) helper that forces the cast to value<>.
103 const value
<Bits
> &val() const {
107 std::string
str() const {
108 std::stringstream ss
;
113 // Operations with compile-time parameters.
115 // These operations are used to implement slicing, concatenation, and blitting.
116 // The trunc, zext and sext operations add or remove most significant bits (i.e. on the left);
117 // the rtrunc and rzext operations add or remove least significant bits (i.e. on the right).
118 template<size_t NewBits
>
120 value
<NewBits
> trunc() const {
121 static_assert(NewBits
<= Bits
, "trunc() may not increase width");
122 value
<NewBits
> result
;
123 for (size_t n
= 0; n
< result
.chunks
; n
++)
124 result
.data
[n
] = data
[n
];
125 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
129 template<size_t NewBits
>
131 value
<NewBits
> zext() const {
132 static_assert(NewBits
>= Bits
, "zext() may not decrease width");
133 value
<NewBits
> result
;
134 for (size_t n
= 0; n
< chunks
; n
++)
135 result
.data
[n
] = data
[n
];
139 template<size_t NewBits
>
141 value
<NewBits
> sext() const {
142 static_assert(NewBits
>= Bits
, "sext() may not decrease width");
143 value
<NewBits
> result
;
144 for (size_t n
= 0; n
< chunks
; n
++)
145 result
.data
[n
] = data
[n
];
147 result
.data
[chunks
- 1] |= ~msb_mask
;
148 for (size_t n
= chunks
; n
< result
.chunks
; n
++)
149 result
.data
[n
] = chunk::mask
;
150 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
155 template<size_t NewBits
>
157 value
<NewBits
> rtrunc() const {
158 static_assert(NewBits
<= Bits
, "rtrunc() may not increase width");
159 value
<NewBits
> result
;
160 constexpr size_t shift_chunks
= (Bits
- NewBits
) / chunk::bits
;
161 constexpr size_t shift_bits
= (Bits
- NewBits
) % chunk::bits
;
162 chunk::type carry
= 0;
163 if (shift_chunks
+ result
.chunks
< chunks
) {
164 carry
= (shift_bits
== 0) ? 0
165 : data
[shift_chunks
+ result
.chunks
] << (chunk::bits
- shift_bits
);
167 for (size_t n
= result
.chunks
; n
> 0; n
--) {
168 result
.data
[n
- 1] = carry
| (data
[shift_chunks
+ n
- 1] >> shift_bits
);
169 carry
= (shift_bits
== 0) ? 0
170 : data
[shift_chunks
+ n
- 1] << (chunk::bits
- shift_bits
);
175 template<size_t NewBits
>
177 value
<NewBits
> rzext() const {
178 static_assert(NewBits
>= Bits
, "rzext() may not decrease width");
179 value
<NewBits
> result
;
180 constexpr size_t shift_chunks
= (NewBits
- Bits
) / chunk::bits
;
181 constexpr size_t shift_bits
= (NewBits
- Bits
) % chunk::bits
;
182 chunk::type carry
= 0;
183 for (size_t n
= 0; n
< chunks
; n
++) {
184 result
.data
[shift_chunks
+ n
] = (data
[n
] << shift_bits
) | carry
;
185 carry
= (shift_bits
== 0) ? 0
186 : data
[n
] >> (chunk::bits
- shift_bits
);
189 result
.data
[result
.chunks
- 1] = carry
;
193 // Bit blit operation, i.e. a partial read-modify-write.
194 template<size_t Stop
, size_t Start
>
196 value
<Bits
> blit(const value
<Stop
- Start
+ 1> &source
) const {
197 static_assert(Stop
>= Start
, "blit() may not reverse bit order");
198 constexpr chunk::type start_mask
= ~(chunk::mask
<< (Start
% chunk::bits
));
199 constexpr chunk::type stop_mask
= (Stop
% chunk::bits
+ 1 == chunk::bits
) ? 0
200 : (chunk::mask
<< (Stop
% chunk::bits
+ 1));
201 value
<Bits
> masked
= *this;
202 if (Start
/ chunk::bits
== Stop
/ chunk::bits
) {
203 masked
.data
[Start
/ chunk::bits
] &= stop_mask
| start_mask
;
205 masked
.data
[Start
/ chunk::bits
] &= start_mask
;
206 for (size_t n
= Start
/ chunk::bits
+ 1; n
< Stop
/ chunk::bits
; n
++)
208 masked
.data
[Stop
/ chunk::bits
] &= stop_mask
;
210 value
<Bits
> shifted
= source
211 .template rzext
<Stop
+ 1>()
212 .template zext
<Bits
>();
213 return masked
.bit_or(shifted
);
216 // Helpers for selecting extending or truncating operation depending on whether the result is wider or narrower
217 // than the operand. In C++17 these can be replaced with `if constexpr`.
218 template<size_t NewBits
, typename
= void>
221 value
<NewBits
> operator()(const value
<Bits
> &val
) {
222 return val
.template zext
<NewBits
>();
226 template<size_t NewBits
>
227 struct zext_cast
<NewBits
, typename
std::enable_if
<(NewBits
< Bits
)>::type
> {
229 value
<NewBits
> operator()(const value
<Bits
> &val
) {
230 return val
.template trunc
<NewBits
>();
234 template<size_t NewBits
, typename
= void>
237 value
<NewBits
> operator()(const value
<Bits
> &val
) {
238 return val
.template sext
<NewBits
>();
242 template<size_t NewBits
>
243 struct sext_cast
<NewBits
, typename
std::enable_if
<(NewBits
< Bits
)>::type
> {
245 value
<NewBits
> operator()(const value
<Bits
> &val
) {
246 return val
.template trunc
<NewBits
>();
250 template<size_t NewBits
>
252 value
<NewBits
> zcast() const {
253 return zext_cast
<NewBits
>()(*this);
256 template<size_t NewBits
>
258 value
<NewBits
> scast() const {
259 return sext_cast
<NewBits
>()(*this);
262 // Operations with run-time parameters (offsets, amounts, etc).
264 // These operations are used for computations.
265 bool bit(size_t offset
) const {
266 return data
[offset
/ chunk::bits
] & (1 << (offset
% chunk::bits
));
269 void set_bit(size_t offset
, bool value
= true) {
270 size_t offset_chunks
= offset
/ chunk::bits
;
271 size_t offset_bits
= offset
% chunk::bits
;
272 data
[offset_chunks
] &= ~(1 << offset_bits
);
273 data
[offset_chunks
] |= value
? 1 << offset_bits
: 0;
276 bool is_zero() const {
277 for (size_t n
= 0; n
< chunks
; n
++)
283 explicit operator bool() const {
287 bool is_neg() const {
288 return data
[chunks
- 1] & (1 << ((Bits
- 1) % chunk::bits
));
291 bool operator ==(const value
<Bits
> &other
) const {
292 for (size_t n
= 0; n
< chunks
; n
++)
293 if (data
[n
] != other
.data
[n
])
298 bool operator !=(const value
<Bits
> &other
) const {
299 return !(*this == other
);
302 value
<Bits
> bit_not() const {
304 for (size_t n
= 0; n
< chunks
; n
++)
305 result
.data
[n
] = ~data
[n
];
306 result
.data
[chunks
- 1] &= msb_mask
;
310 value
<Bits
> bit_and(const value
<Bits
> &other
) const {
312 for (size_t n
= 0; n
< chunks
; n
++)
313 result
.data
[n
] = data
[n
] & other
.data
[n
];
317 value
<Bits
> bit_or(const value
<Bits
> &other
) const {
319 for (size_t n
= 0; n
< chunks
; n
++)
320 result
.data
[n
] = data
[n
] | other
.data
[n
];
324 value
<Bits
> bit_xor(const value
<Bits
> &other
) const {
326 for (size_t n
= 0; n
< chunks
; n
++)
327 result
.data
[n
] = data
[n
] ^ other
.data
[n
];
331 value
<Bits
> update(const value
<Bits
> &val
, const value
<Bits
> &mask
) const {
332 return bit_and(mask
.bit_not()).bit_or(val
.bit_and(mask
));
335 template<size_t AmountBits
>
336 value
<Bits
> shl(const value
<AmountBits
> &amount
) const {
337 // Ensure our early return is correct by prohibiting values larger than 4 Gbit.
338 static_assert(Bits
<= chunk::mask
, "shl() of unreasonably large values is not supported");
339 // Detect shifts definitely large than Bits early.
340 for (size_t n
= 1; n
< amount
.chunks
; n
++)
341 if (amount
.data
[n
] != 0)
343 // Past this point we can use the least significant chunk as the shift size.
344 size_t shift_chunks
= amount
.data
[0] / chunk::bits
;
345 size_t shift_bits
= amount
.data
[0] % chunk::bits
;
346 if (shift_chunks
>= chunks
)
349 chunk::type carry
= 0;
350 for (size_t n
= 0; n
< chunks
- shift_chunks
; n
++) {
351 result
.data
[shift_chunks
+ n
] = (data
[n
] << shift_bits
) | carry
;
352 carry
= (shift_bits
== 0) ? 0
353 : data
[n
] >> (chunk::bits
- shift_bits
);
358 template<size_t AmountBits
, bool Signed
= false>
359 value
<Bits
> shr(const value
<AmountBits
> &amount
) const {
360 // Ensure our early return is correct by prohibiting values larger than 4 Gbit.
361 static_assert(Bits
<= chunk::mask
, "shr() of unreasonably large values is not supported");
362 // Detect shifts definitely large than Bits early.
363 for (size_t n
= 1; n
< amount
.chunks
; n
++)
364 if (amount
.data
[n
] != 0)
366 // Past this point we can use the least significant chunk as the shift size.
367 size_t shift_chunks
= amount
.data
[0] / chunk::bits
;
368 size_t shift_bits
= amount
.data
[0] % chunk::bits
;
369 if (shift_chunks
>= chunks
)
372 chunk::type carry
= 0;
373 for (size_t n
= 0; n
< chunks
- shift_chunks
; n
++) {
374 result
.data
[chunks
- shift_chunks
- 1 - n
] = carry
| (data
[chunks
- 1 - n
] >> shift_bits
);
375 carry
= (shift_bits
== 0) ? 0
376 : data
[chunks
- 1 - n
] << (chunk::bits
- shift_bits
);
378 if (Signed
&& is_neg()) {
379 for (size_t n
= chunks
- shift_chunks
; n
< chunks
; n
++)
380 result
.data
[n
] = chunk::mask
;
382 result
.data
[chunks
- shift_chunks
] |= chunk::mask
<< (chunk::bits
- shift_bits
);
387 template<size_t AmountBits
>
388 value
<Bits
> sshr(const value
<AmountBits
> &amount
) const {
389 return shr
<AmountBits
, /*Signed=*/true>(amount
);
392 size_t ctpop() const {
394 for (size_t n
= 0; n
< chunks
; n
++) {
395 // This loop implements the population count idiom as recognized by LLVM and GCC.
396 for (chunk::type x
= data
[n
]; x
!= 0; count
++)
402 size_t ctlz() const {
404 for (size_t n
= 0; n
< chunks
; n
++) {
405 chunk::type x
= data
[chunks
- 1 - n
];
407 count
+= (n
== 0 ? Bits
% chunk::bits
: chunk::bits
);
409 // This loop implements the find first set idiom as recognized by LLVM.
410 for (; x
!= 0; count
++)
417 template<bool Invert
, bool CarryIn
>
418 std::pair
<value
<Bits
>, bool /*CarryOut*/> alu(const value
<Bits
> &other
) const {
420 bool carry
= CarryIn
;
421 for (size_t n
= 0; n
< result
.chunks
; n
++) {
422 result
.data
[n
] = data
[n
] + (Invert
? ~other
.data
[n
] : other
.data
[n
]) + carry
;
423 carry
= (result
.data
[n
] < data
[n
]) ||
424 (result
.data
[n
] == data
[n
] && carry
);
426 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
427 return {result
, carry
};
430 value
<Bits
> add(const value
<Bits
> &other
) const {
431 return alu
</*Invert=*/false, /*CarryIn=*/false>(other
).first
;
434 value
<Bits
> sub(const value
<Bits
> &other
) const {
435 return alu
</*Invert=*/true, /*CarryIn=*/true>(other
).first
;
438 value
<Bits
> neg() const {
439 return value
<Bits
> { 0u }.sub(*this);
442 bool ucmp(const value
<Bits
> &other
) const {
444 std::tie(std::ignore
, carry
) = alu
</*Invert=*/true, /*CarryIn=*/true>(other
);
445 return !carry
; // a.ucmp(b) ≡ a u< b
448 bool scmp(const value
<Bits
> &other
) const {
451 std::tie(result
, carry
) = alu
</*Invert=*/true, /*CarryIn=*/true>(other
);
452 bool overflow
= (is_neg() == !other
.is_neg()) && (is_neg() != result
.is_neg());
453 return result
.is_neg() ^ overflow
; // a.scmp(b) ≡ a s< b
457 // Expression template for a slice, usable as lvalue or rvalue, and composable with other expression templates here.
458 template<class T
, size_t Stop
, size_t Start
>
459 struct slice_expr
: public expr_base
<slice_expr
<T
, Stop
, Start
>> {
460 static_assert(Stop
>= Start
, "slice_expr() may not reverse bit order");
461 static_assert(Start
< T::bits
&& Stop
< T::bits
, "slice_expr() must be within bounds");
462 static constexpr size_t bits
= Stop
- Start
+ 1;
466 slice_expr(T
&expr
) : expr(expr
) {}
467 slice_expr(const slice_expr
<T
, Stop
, Start
> &) = delete;
470 operator value
<bits
>() const {
471 return static_cast<const value
<T::bits
> &>(expr
)
472 .template rtrunc
<T::bits
- Start
>()
473 .template trunc
<bits
>();
477 slice_expr
<T
, Stop
, Start
> &operator=(const value
<bits
> &rhs
) {
478 // Generic partial assignment implemented using a read-modify-write operation on the sliced expression.
479 expr
= static_cast<const value
<T::bits
> &>(expr
)
480 .template blit
<Stop
, Start
>(rhs
);
484 // A helper that forces the cast to value<>, which allows deduction to work.
486 value
<bits
> val() const {
487 return static_cast<const value
<bits
> &>(*this);
491 // Expression template for a concatenation, usable as lvalue or rvalue, and composable with other expression templates here.
492 template<class T
, class U
>
493 struct concat_expr
: public expr_base
<concat_expr
<T
, U
>> {
494 static constexpr size_t bits
= T::bits
+ U::bits
;
499 concat_expr(T
&ms_expr
, U
&ls_expr
) : ms_expr(ms_expr
), ls_expr(ls_expr
) {}
500 concat_expr(const concat_expr
<T
, U
> &) = delete;
503 operator value
<bits
>() const {
504 value
<bits
> ms_shifted
= static_cast<const value
<T::bits
> &>(ms_expr
)
505 .template rzext
<bits
>();
506 value
<bits
> ls_extended
= static_cast<const value
<U::bits
> &>(ls_expr
)
507 .template zext
<bits
>();
508 return ms_shifted
.bit_or(ls_extended
);
512 concat_expr
<T
, U
> &operator=(const value
<bits
> &rhs
) {
513 ms_expr
= rhs
.template rtrunc
<T::bits
>();
514 ls_expr
= rhs
.template trunc
<U::bits
>();
518 // A helper that forces the cast to value<>, which allows deduction to work.
520 value
<bits
> val() const {
521 return static_cast<const value
<bits
> &>(*this);
525 // Base class for expression templates, providing helper methods for operations that are valid on both rvalues and lvalues.
527 // Note that expression objects (slices and concatenations) constructed in this way should NEVER be captured because
528 // they refer to temporaries that will, in general, only live until the end of the statement. For example, both of
529 // these snippets perform use-after-free:
531 // const auto &a = val.slice<7,0>().slice<1>();
534 // auto &&c = val.slice<7,0>().slice<1>();
537 // An easy way to write code using slices and concatenations safely is to follow two simple rules:
538 // * Never explicitly name any type except `value<W>` or `const value<W> &`.
539 // * Never use a `const auto &` or `auto &&` in any such expression.
540 // Then, any code that compiles will be well-defined.
543 template<size_t Stop
, size_t Start
= Stop
>
545 slice_expr
<const T
, Stop
, Start
> slice() const {
546 return {*static_cast<const T
*>(this)};
549 template<size_t Stop
, size_t Start
= Stop
>
551 slice_expr
<T
, Stop
, Start
> slice() {
552 return {*static_cast<T
*>(this)};
557 concat_expr
<const T
, typename
std::remove_reference
<const U
>::type
> concat(const U
&other
) const {
558 return {*static_cast<const T
*>(this), other
};
563 concat_expr
<T
, typename
std::remove_reference
<U
>::type
> concat(U
&&other
) {
564 return {*static_cast<T
*>(this), other
};
568 template<size_t Bits
>
569 std::ostream
&operator<<(std::ostream
&os
, const value
<Bits
> &val
) {
570 auto old_flags
= os
.flags(std::ios::right
);
571 auto old_width
= os
.width(0);
572 auto old_fill
= os
.fill('0');
573 os
<< val
.bits
<< '\'' << std::hex
;
574 for (size_t n
= val
.chunks
- 1; n
!= (size_t)-1; n
--) {
575 if (n
== val
.chunks
- 1 && Bits
% value
<Bits
>::chunk::bits
!= 0)
576 os
.width((Bits
% value
<Bits
>::chunk::bits
+ 3) / 4);
578 os
.width((value
<Bits
>::chunk::bits
+ 3) / 4);
587 template<size_t Bits
>
589 static constexpr size_t bits
= Bits
;
595 constexpr wire(const value
<Bits
> &init
) : curr(init
), next(init
) {}
596 template<typename
... Init
>
597 explicit constexpr wire(Init
...init
) : curr
{init
...}, next
{init
...} {}
599 wire(const wire
<Bits
> &) = delete;
600 wire(wire
<Bits
> &&) = default;
601 wire
<Bits
> &operator=(const wire
<Bits
> &) = delete;
612 template<size_t Bits
>
613 std::ostream
&operator<<(std::ostream
&os
, const wire
<Bits
> &val
) {
618 template<size_t Width
>
620 std::vector
<value
<Width
>> data
;
622 size_t depth() const {
627 explicit memory(size_t depth
) : data(depth
) {}
629 memory(const memory
<Width
> &) = delete;
630 memory
<Width
> &operator=(const memory
<Width
> &) = delete;
632 // The only way to get the compiler to put the initializer in .rodata and do not copy it on stack is to stuff it
633 // into a plain array. You'd think an std::initializer_list would work here, but it doesn't, because you can't
634 // construct an initializer_list in a constexpr (or something) and so if you try to do that the whole thing is
635 // first copied on the stack (probably overflowing it) and then again into `data`.
636 template<size_t Size
>
639 value
<Width
> data
[Size
];
642 template<size_t... InitSize
>
643 explicit memory(size_t depth
, const init
<InitSize
> &...init
) : data(depth
) {
645 // This utterly reprehensible construct is the most reasonable way to apply a function to every element
646 // of a parameter pack, if the elements all have different types and so cannot be cast to an initializer list.
647 auto _
= {std::move(std::begin(init
.data
), std::end(init
.data
), data
.begin() + init
.offset
)...};
651 // An operator for direct memory reads. May be used at any time during the simulation.
652 const value
<Width
> &operator [](size_t index
) const {
653 assert(index
< data
.size());
657 // An operator for direct memory writes. May only be used before the simulation is started. If used
658 // after the simulation is started, the design may malfunction.
659 value
<Width
> &operator [](size_t index
) {
660 assert(index
< data
.size());
664 // A simple way to make a writable memory would be to use an array of wires instead of an array of values.
665 // However, there are two significant downsides to this approach: first, it has large overhead (2× space
666 // overhead, and O(depth) time overhead during commit); second, it does not simplify handling write port
667 // priorities. Although in principle write ports could be ordered or conditionally enabled in generated
668 // code based on their priorities and selected addresses, the feedback arc set problem is computationally
669 // expensive, and the heuristic based algorithms are not easily modified to guarantee (rather than prefer)
670 // a particular write port evaluation order.
672 // The approach used here instead is to queue writes into a buffer during the eval phase, then perform
673 // the writes during the commit phase in the priority order. This approach has low overhead, with both space
674 // and time proportional to the amount of write ports. Because virtually every memory in a practical design
675 // has at most two write ports, linear search is used on every write, being the fastest and simplest approach.
682 std::vector
<write
> write_queue
;
684 void update(size_t index
, const value
<Width
> &val
, const value
<Width
> &mask
, int priority
= 0) {
685 assert(index
< data
.size());
686 // Queue up the write while keeping the queue sorted by priority.
688 std::upper_bound(write_queue
.begin(), write_queue
.end(), priority
,
689 [](const int a
, const write
& b
) { return a
< b
.priority
; }),
690 write
{ index
, val
, mask
, priority
});
694 bool changed
= false;
695 for (const write
&entry
: write_queue
) {
696 value
<Width
> elem
= data
[entry
.index
];
697 elem
= elem
.update(entry
.val
, entry
.mask
);
698 changed
|= (data
[entry
.index
] != elem
);
699 data
[entry
.index
] = elem
;
715 // In debug mode, using the wrong .as_*() function will assert.
716 // In release mode, using the wrong .as_*() function will safely return a default value.
717 const unsigned uint_value
= 0;
718 const signed sint_value
= 0;
719 const std::string string_value
= "";
720 const double double_value
= 0.0;
722 metadata() : value_type(MISSING
) {}
723 metadata(unsigned value
) : value_type(UINT
), uint_value(value
) {}
724 metadata(signed value
) : value_type(SINT
), sint_value(value
) {}
725 metadata(const std::string
&value
) : value_type(STRING
), string_value(value
) {}
726 metadata(const char *value
) : value_type(STRING
), string_value(value
) {}
727 metadata(double value
) : value_type(DOUBLE
), double_value(value
) {}
729 metadata(const metadata
&) = default;
730 metadata
&operator=(const metadata
&) = delete;
732 unsigned as_uint() const {
733 assert(value_type
== UINT
);
737 signed as_sint() const {
738 assert(value_type
== SINT
);
742 const std::string
&as_string() const {
743 assert(value_type
== STRING
);
747 double as_double() const {
748 assert(value_type
== DOUBLE
);
753 typedef std::map
<std::string
, metadata
> metadata_map
;
755 // Helper class to disambiguate values/wires and their aliases.
756 struct debug_alias
{};
758 // This structure is intended for consumption via foreign function interfaces, like Python's ctypes.
759 // Because of this it uses a C-style layout that is easy to parse rather than more idiomatic C++.
761 // To avoid violating strict aliasing rules, this structure has to be a subclass of the one used
762 // in the C API, or it would not be possible to cast between the pointers to these.
763 struct debug_item
: ::cxxrtl_object
{
765 VALUE
= CXXRTL_VALUE
,
767 MEMORY
= CXXRTL_MEMORY
,
768 ALIAS
= CXXRTL_ALIAS
,
771 debug_item(const ::cxxrtl_object
&object
) : cxxrtl_object(object
) {}
773 template<size_t Bits
>
774 debug_item(value
<Bits
> &item
) {
775 static_assert(sizeof(item
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
776 "value<Bits> is not compatible with C layout");
784 template<size_t Bits
>
785 debug_item(const value
<Bits
> &item
) {
786 static_assert(sizeof(item
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
787 "value<Bits> is not compatible with C layout");
791 curr
= const_cast<chunk_t
*>(item
.data
);
795 template<size_t Bits
>
796 debug_item(wire
<Bits
> &item
) {
797 static_assert(sizeof(item
.curr
) == value
<Bits
>::chunks
* sizeof(chunk_t
) &&
798 sizeof(item
.next
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
799 "wire<Bits> is not compatible with C layout");
803 curr
= item
.curr
.data
;
804 next
= item
.next
.data
;
807 template<size_t Width
>
808 debug_item(memory
<Width
> &item
) {
809 static_assert(sizeof(item
.data
[0]) == value
<Width
>::chunks
* sizeof(chunk_t
),
810 "memory<Width> is not compatible with C layout");
813 depth
= item
.data
.size();
814 curr
= item
.data
.empty() ? nullptr : item
.data
[0].data
;
818 template<size_t Bits
>
819 debug_item(debug_alias
, const value
<Bits
> &item
) {
820 static_assert(sizeof(item
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
821 "value<Bits> is not compatible with C layout");
825 curr
= const_cast<chunk_t
*>(item
.data
);
829 template<size_t Bits
>
830 debug_item(debug_alias
, const wire
<Bits
> &item
) {
831 static_assert(sizeof(item
.curr
) == value
<Bits
>::chunks
* sizeof(chunk_t
) &&
832 sizeof(item
.next
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
833 "wire<Bits> is not compatible with C layout");
837 curr
= const_cast<chunk_t
*>(item
.curr
.data
);
841 static_assert(std::is_standard_layout
<debug_item
>::value
, "debug_item is not compatible with C layout");
843 typedef std::map
<std::string
, debug_item
> debug_items
;
849 module(const module
&) = delete;
850 module
&operator=(const module
&) = delete;
852 virtual bool eval() = 0;
853 virtual bool commit() = 0;
857 bool converged
= false;
861 } while (commit() && !converged
);
865 virtual void debug_info(debug_items
&items
, std::string path
= "") {
866 (void)items
, (void)path
;
870 } // namespace cxxrtl
872 // Internal structure used to communicate with the implementation of the C interface.
873 typedef struct _cxxrtl_toplevel
{
874 std::unique_ptr
<cxxrtl::module
> module
;
877 // Definitions of internal Yosys cells. Other than the functions in this namespace, CXXRTL is fully generic
878 // and indepenent of Yosys implementation details.
880 // The `write_cxxrtl` pass translates internal cells (cells with names that start with `$`) to calls of these
881 // functions. All of Yosys arithmetic and logical cells perform sign or zero extension on their operands,
882 // whereas basic operations on arbitrary width values require operands to be of the same width. These functions
883 // bridge the gap by performing the necessary casts. They are named similar to `cell_A[B]`, where A and B are `u`
884 // if the corresponding operand is unsigned, and `s` if it is signed.
885 namespace cxxrtl_yosys
{
887 using namespace cxxrtl
;
889 // std::max isn't constexpr until C++14 for no particular reason (it's an oversight), so we define our own.
892 constexpr T
max(const T
&a
, const T
&b
) {
893 return a
> b
? a
: b
;
897 template<size_t BitsY
, size_t BitsA
>
899 value
<BitsY
> logic_not(const value
<BitsA
> &a
) {
900 return value
<BitsY
> { a
? 0u : 1u };
903 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
905 value
<BitsY
> logic_and(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
906 return value
<BitsY
> { (bool(a
) & bool(b
)) ? 1u : 0u };
909 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
911 value
<BitsY
> logic_or(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
912 return value
<BitsY
> { (bool(a
) | bool(b
)) ? 1u : 0u };
915 // Reduction operations
916 template<size_t BitsY
, size_t BitsA
>
918 value
<BitsY
> reduce_and(const value
<BitsA
> &a
) {
919 return value
<BitsY
> { a
.bit_not().is_zero() ? 1u : 0u };
922 template<size_t BitsY
, size_t BitsA
>
924 value
<BitsY
> reduce_or(const value
<BitsA
> &a
) {
925 return value
<BitsY
> { a
? 1u : 0u };
928 template<size_t BitsY
, size_t BitsA
>
930 value
<BitsY
> reduce_xor(const value
<BitsA
> &a
) {
931 return value
<BitsY
> { (a
.ctpop() % 2) ? 1u : 0u };
934 template<size_t BitsY
, size_t BitsA
>
936 value
<BitsY
> reduce_xnor(const value
<BitsA
> &a
) {
937 return value
<BitsY
> { (a
.ctpop() % 2) ? 0u : 1u };
940 template<size_t BitsY
, size_t BitsA
>
942 value
<BitsY
> reduce_bool(const value
<BitsA
> &a
) {
943 return value
<BitsY
> { a
? 1u : 0u };
946 // Bitwise operations
947 template<size_t BitsY
, size_t BitsA
>
949 value
<BitsY
> not_u(const value
<BitsA
> &a
) {
950 return a
.template zcast
<BitsY
>().bit_not();
953 template<size_t BitsY
, size_t BitsA
>
955 value
<BitsY
> not_s(const value
<BitsA
> &a
) {
956 return a
.template scast
<BitsY
>().bit_not();
959 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
961 value
<BitsY
> and_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
962 return a
.template zcast
<BitsY
>().bit_and(b
.template zcast
<BitsY
>());
965 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
967 value
<BitsY
> and_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
968 return a
.template scast
<BitsY
>().bit_and(b
.template scast
<BitsY
>());
971 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
973 value
<BitsY
> or_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
974 return a
.template zcast
<BitsY
>().bit_or(b
.template zcast
<BitsY
>());
977 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
979 value
<BitsY
> or_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
980 return a
.template scast
<BitsY
>().bit_or(b
.template scast
<BitsY
>());
983 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
985 value
<BitsY
> xor_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
986 return a
.template zcast
<BitsY
>().bit_xor(b
.template zcast
<BitsY
>());
989 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
991 value
<BitsY
> xor_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
992 return a
.template scast
<BitsY
>().bit_xor(b
.template scast
<BitsY
>());
995 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
997 value
<BitsY
> xnor_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
998 return a
.template zcast
<BitsY
>().bit_xor(b
.template zcast
<BitsY
>()).bit_not();
1001 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1002 CXXRTL_ALWAYS_INLINE
1003 value
<BitsY
> xnor_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1004 return a
.template scast
<BitsY
>().bit_xor(b
.template scast
<BitsY
>()).bit_not();
1007 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1008 CXXRTL_ALWAYS_INLINE
1009 value
<BitsY
> shl_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1010 return a
.template zcast
<BitsY
>().template shl(b
);
1013 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1014 CXXRTL_ALWAYS_INLINE
1015 value
<BitsY
> shl_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1016 return a
.template scast
<BitsY
>().template shl(b
);
1019 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1020 CXXRTL_ALWAYS_INLINE
1021 value
<BitsY
> sshl_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1022 return a
.template zcast
<BitsY
>().template shl(b
);
1025 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1026 CXXRTL_ALWAYS_INLINE
1027 value
<BitsY
> sshl_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1028 return a
.template scast
<BitsY
>().template shl(b
);
1031 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1032 CXXRTL_ALWAYS_INLINE
1033 value
<BitsY
> shr_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1034 return a
.template shr(b
).template zcast
<BitsY
>();
1037 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1038 CXXRTL_ALWAYS_INLINE
1039 value
<BitsY
> shr_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1040 return a
.template shr(b
).template scast
<BitsY
>();
1043 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1044 CXXRTL_ALWAYS_INLINE
1045 value
<BitsY
> sshr_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1046 return a
.template shr(b
).template zcast
<BitsY
>();
1049 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1050 CXXRTL_ALWAYS_INLINE
1051 value
<BitsY
> sshr_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1052 return a
.template sshr(b
).template scast
<BitsY
>();
1055 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1056 CXXRTL_ALWAYS_INLINE
1057 value
<BitsY
> shift_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1058 return shr_uu
<BitsY
>(a
, b
);
1061 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1062 CXXRTL_ALWAYS_INLINE
1063 value
<BitsY
> shift_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1064 return shr_su
<BitsY
>(a
, b
);
1067 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1068 CXXRTL_ALWAYS_INLINE
1069 value
<BitsY
> shift_us(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1070 return b
.is_neg() ? shl_uu
<BitsY
>(a
, b
.template sext
<BitsB
+ 1>().neg()) : shr_uu
<BitsY
>(a
, b
);
1073 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1074 CXXRTL_ALWAYS_INLINE
1075 value
<BitsY
> shift_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1076 return b
.is_neg() ? shl_su
<BitsY
>(a
, b
.template sext
<BitsB
+ 1>().neg()) : shr_su
<BitsY
>(a
, b
);
1079 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1080 CXXRTL_ALWAYS_INLINE
1081 value
<BitsY
> shiftx_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1082 return shift_uu
<BitsY
>(a
, b
);
1085 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1086 CXXRTL_ALWAYS_INLINE
1087 value
<BitsY
> shiftx_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1088 return shift_su
<BitsY
>(a
, b
);
1091 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1092 CXXRTL_ALWAYS_INLINE
1093 value
<BitsY
> shiftx_us(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1094 return shift_us
<BitsY
>(a
, b
);
1097 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1098 CXXRTL_ALWAYS_INLINE
1099 value
<BitsY
> shiftx_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1100 return shift_ss
<BitsY
>(a
, b
);
1103 // Comparison operations
1104 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1105 CXXRTL_ALWAYS_INLINE
1106 value
<BitsY
> eq_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1107 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1108 return value
<BitsY
>{ a
.template zext
<BitsExt
>() == b
.template zext
<BitsExt
>() ? 1u : 0u };
1111 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1112 CXXRTL_ALWAYS_INLINE
1113 value
<BitsY
> eq_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1114 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1115 return value
<BitsY
>{ a
.template sext
<BitsExt
>() == b
.template sext
<BitsExt
>() ? 1u : 0u };
1118 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1119 CXXRTL_ALWAYS_INLINE
1120 value
<BitsY
> ne_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1121 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1122 return value
<BitsY
>{ a
.template zext
<BitsExt
>() != b
.template zext
<BitsExt
>() ? 1u : 0u };
1125 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1126 CXXRTL_ALWAYS_INLINE
1127 value
<BitsY
> ne_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1128 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1129 return value
<BitsY
>{ a
.template sext
<BitsExt
>() != b
.template sext
<BitsExt
>() ? 1u : 0u };
1132 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1133 CXXRTL_ALWAYS_INLINE
1134 value
<BitsY
> eqx_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1135 return eq_uu
<BitsY
>(a
, b
);
1138 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1139 CXXRTL_ALWAYS_INLINE
1140 value
<BitsY
> eqx_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1141 return eq_ss
<BitsY
>(a
, b
);
1144 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1145 CXXRTL_ALWAYS_INLINE
1146 value
<BitsY
> nex_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1147 return ne_uu
<BitsY
>(a
, b
);
1150 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1151 CXXRTL_ALWAYS_INLINE
1152 value
<BitsY
> nex_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1153 return ne_ss
<BitsY
>(a
, b
);
1156 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1157 CXXRTL_ALWAYS_INLINE
1158 value
<BitsY
> gt_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1159 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1160 return value
<BitsY
> { b
.template zext
<BitsExt
>().ucmp(a
.template zext
<BitsExt
>()) ? 1u : 0u };
1163 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1164 CXXRTL_ALWAYS_INLINE
1165 value
<BitsY
> gt_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1166 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1167 return value
<BitsY
> { b
.template sext
<BitsExt
>().scmp(a
.template sext
<BitsExt
>()) ? 1u : 0u };
1170 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1171 CXXRTL_ALWAYS_INLINE
1172 value
<BitsY
> ge_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1173 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1174 return value
<BitsY
> { !a
.template zext
<BitsExt
>().ucmp(b
.template zext
<BitsExt
>()) ? 1u : 0u };
1177 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1178 CXXRTL_ALWAYS_INLINE
1179 value
<BitsY
> ge_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1180 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1181 return value
<BitsY
> { !a
.template sext
<BitsExt
>().scmp(b
.template sext
<BitsExt
>()) ? 1u : 0u };
1184 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1185 CXXRTL_ALWAYS_INLINE
1186 value
<BitsY
> lt_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1187 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1188 return value
<BitsY
> { a
.template zext
<BitsExt
>().ucmp(b
.template zext
<BitsExt
>()) ? 1u : 0u };
1191 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1192 CXXRTL_ALWAYS_INLINE
1193 value
<BitsY
> lt_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1194 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1195 return value
<BitsY
> { a
.template sext
<BitsExt
>().scmp(b
.template sext
<BitsExt
>()) ? 1u : 0u };
1198 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1199 CXXRTL_ALWAYS_INLINE
1200 value
<BitsY
> le_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1201 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1202 return value
<BitsY
> { !b
.template zext
<BitsExt
>().ucmp(a
.template zext
<BitsExt
>()) ? 1u : 0u };
1205 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1206 CXXRTL_ALWAYS_INLINE
1207 value
<BitsY
> le_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1208 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1209 return value
<BitsY
> { !b
.template sext
<BitsExt
>().scmp(a
.template sext
<BitsExt
>()) ? 1u : 0u };
1212 // Arithmetic operations
1213 template<size_t BitsY
, size_t BitsA
>
1214 CXXRTL_ALWAYS_INLINE
1215 value
<BitsY
> pos_u(const value
<BitsA
> &a
) {
1216 return a
.template zcast
<BitsY
>();
1219 template<size_t BitsY
, size_t BitsA
>
1220 CXXRTL_ALWAYS_INLINE
1221 value
<BitsY
> pos_s(const value
<BitsA
> &a
) {
1222 return a
.template scast
<BitsY
>();
1225 template<size_t BitsY
, size_t BitsA
>
1226 CXXRTL_ALWAYS_INLINE
1227 value
<BitsY
> neg_u(const value
<BitsA
> &a
) {
1228 return a
.template zcast
<BitsY
>().neg();
1231 template<size_t BitsY
, size_t BitsA
>
1232 CXXRTL_ALWAYS_INLINE
1233 value
<BitsY
> neg_s(const value
<BitsA
> &a
) {
1234 return a
.template scast
<BitsY
>().neg();
1237 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1238 CXXRTL_ALWAYS_INLINE
1239 value
<BitsY
> add_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1240 return a
.template zcast
<BitsY
>().add(b
.template zcast
<BitsY
>());
1243 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1244 CXXRTL_ALWAYS_INLINE
1245 value
<BitsY
> add_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1246 return a
.template scast
<BitsY
>().add(b
.template scast
<BitsY
>());
1249 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1250 CXXRTL_ALWAYS_INLINE
1251 value
<BitsY
> sub_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1252 return a
.template zcast
<BitsY
>().sub(b
.template zcast
<BitsY
>());
1255 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1256 CXXRTL_ALWAYS_INLINE
1257 value
<BitsY
> sub_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1258 return a
.template scast
<BitsY
>().sub(b
.template scast
<BitsY
>());
1261 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1262 CXXRTL_ALWAYS_INLINE
1263 value
<BitsY
> mul_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1264 value
<BitsY
> product
;
1265 value
<BitsY
> multiplicand
= a
.template zcast
<BitsY
>();
1266 const value
<BitsB
> &multiplier
= b
;
1267 uint32_t multiplicand_shift
= 0;
1268 for (size_t step
= 0; step
< BitsB
; step
++) {
1269 if (multiplier
.bit(step
)) {
1270 multiplicand
= multiplicand
.shl(value
<32> { multiplicand_shift
});
1271 product
= product
.add(multiplicand
);
1272 multiplicand_shift
= 0;
1274 multiplicand_shift
++;
1279 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1280 CXXRTL_ALWAYS_INLINE
1281 value
<BitsY
> mul_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1282 value
<BitsB
+ 1> ub
= b
.template sext
<BitsB
+ 1>();
1283 if (ub
.is_neg()) ub
= ub
.neg();
1284 value
<BitsY
> y
= mul_uu
<BitsY
>(a
.template scast
<BitsY
>(), ub
);
1285 return b
.is_neg() ? y
.neg() : y
;
1288 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1289 CXXRTL_ALWAYS_INLINE
1290 std::pair
<value
<BitsY
>, value
<BitsY
>> divmod_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1291 constexpr size_t Bits
= max(BitsY
, max(BitsA
, BitsB
));
1292 value
<Bits
> quotient
;
1293 value
<Bits
> dividend
= a
.template zext
<Bits
>();
1294 value
<Bits
> divisor
= b
.template zext
<Bits
>();
1295 if (dividend
.ucmp(divisor
))
1296 return {/*quotient=*/value
<BitsY
> { 0u }, /*remainder=*/dividend
.template trunc
<BitsY
>()};
1297 uint32_t divisor_shift
= dividend
.ctlz() - divisor
.ctlz();
1298 divisor
= divisor
.shl(value
<32> { divisor_shift
});
1299 for (size_t step
= 0; step
<= divisor_shift
; step
++) {
1300 quotient
= quotient
.shl(value
<1> { 1u });
1301 if (!dividend
.ucmp(divisor
)) {
1302 dividend
= dividend
.sub(divisor
);
1303 quotient
.set_bit(0, true);
1305 divisor
= divisor
.shr(value
<1> { 1u });
1307 return {quotient
.template trunc
<BitsY
>(), /*remainder=*/dividend
.template trunc
<BitsY
>()};
1310 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1311 CXXRTL_ALWAYS_INLINE
1312 std::pair
<value
<BitsY
>, value
<BitsY
>> divmod_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1313 value
<BitsA
+ 1> ua
= a
.template sext
<BitsA
+ 1>();
1314 value
<BitsB
+ 1> ub
= b
.template sext
<BitsB
+ 1>();
1315 if (ua
.is_neg()) ua
= ua
.neg();
1316 if (ub
.is_neg()) ub
= ub
.neg();
1318 std::tie(y
, r
) = divmod_uu
<BitsY
>(ua
, ub
);
1319 if (a
.is_neg() != b
.is_neg()) y
= y
.neg();
1320 if (a
.is_neg()) r
= r
.neg();
1324 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1325 CXXRTL_ALWAYS_INLINE
1326 value
<BitsY
> div_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1327 return divmod_uu
<BitsY
>(a
, b
).first
;
1330 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1331 CXXRTL_ALWAYS_INLINE
1332 value
<BitsY
> div_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1333 return divmod_ss
<BitsY
>(a
, b
).first
;
1336 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1337 CXXRTL_ALWAYS_INLINE
1338 value
<BitsY
> mod_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1339 return divmod_uu
<BitsY
>(a
, b
).second
;
1342 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1343 CXXRTL_ALWAYS_INLINE
1344 value
<BitsY
> mod_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1345 return divmod_ss
<BitsY
>(a
, b
).second
;
1349 struct memory_index
{
1353 template<size_t BitsAddr
>
1354 memory_index(const value
<BitsAddr
> &addr
, size_t offset
, size_t depth
) {
1355 static_assert(value
<BitsAddr
>::chunks
<= 1, "memory address is too wide");
1356 size_t offset_index
= addr
.data
[0];
1358 valid
= (offset_index
>= offset
&& offset_index
< offset
+ depth
);
1359 index
= offset_index
- offset
;
1363 } // namespace cxxrtl_yosys