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
;
69 typedef uint64_t wide_chunk_t
;
73 static_assert(std::is_integral
<T
>::value
&& std::is_unsigned
<T
>::value
,
74 "chunk type must be an unsigned integral type");
76 static constexpr size_t bits
= std::numeric_limits
<T
>::digits
;
77 static constexpr T mask
= std::numeric_limits
<T
>::max();
84 struct value
: public expr_base
<value
<Bits
>> {
85 static constexpr size_t bits
= Bits
;
87 using chunk
= chunk_traits
<chunk_t
>;
88 static constexpr chunk::type msb_mask
= (Bits
% chunk::bits
== 0) ? chunk::mask
89 : chunk::mask
>> (chunk::bits
- (Bits
% chunk::bits
));
91 static constexpr size_t chunks
= (Bits
+ chunk::bits
- 1) / chunk::bits
;
92 chunk::type data
[chunks
] = {};
95 template<typename
... Init
>
96 explicit constexpr value(Init
...init
) : data
{init
...} {}
98 value(const value
<Bits
> &) = default;
99 value(value
<Bits
> &&) = default;
100 value
<Bits
> &operator=(const value
<Bits
> &) = default;
102 // A (no-op) helper that forces the cast to value<>.
104 const value
<Bits
> &val() const {
108 std::string
str() const {
109 std::stringstream ss
;
114 // Operations with compile-time parameters.
116 // These operations are used to implement slicing, concatenation, and blitting.
117 // The trunc, zext and sext operations add or remove most significant bits (i.e. on the left);
118 // the rtrunc and rzext operations add or remove least significant bits (i.e. on the right).
119 template<size_t NewBits
>
121 value
<NewBits
> trunc() const {
122 static_assert(NewBits
<= Bits
, "trunc() may not increase width");
123 value
<NewBits
> result
;
124 for (size_t n
= 0; n
< result
.chunks
; n
++)
125 result
.data
[n
] = data
[n
];
126 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
130 template<size_t NewBits
>
132 value
<NewBits
> zext() const {
133 static_assert(NewBits
>= Bits
, "zext() may not decrease width");
134 value
<NewBits
> result
;
135 for (size_t n
= 0; n
< chunks
; n
++)
136 result
.data
[n
] = data
[n
];
140 template<size_t NewBits
>
142 value
<NewBits
> sext() const {
143 static_assert(NewBits
>= Bits
, "sext() may not decrease width");
144 value
<NewBits
> result
;
145 for (size_t n
= 0; n
< chunks
; n
++)
146 result
.data
[n
] = data
[n
];
148 result
.data
[chunks
- 1] |= ~msb_mask
;
149 for (size_t n
= chunks
; n
< result
.chunks
; n
++)
150 result
.data
[n
] = chunk::mask
;
151 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
156 template<size_t NewBits
>
158 value
<NewBits
> rtrunc() const {
159 static_assert(NewBits
<= Bits
, "rtrunc() may not increase width");
160 value
<NewBits
> result
;
161 constexpr size_t shift_chunks
= (Bits
- NewBits
) / chunk::bits
;
162 constexpr size_t shift_bits
= (Bits
- NewBits
) % chunk::bits
;
163 chunk::type carry
= 0;
164 if (shift_chunks
+ result
.chunks
< chunks
) {
165 carry
= (shift_bits
== 0) ? 0
166 : data
[shift_chunks
+ result
.chunks
] << (chunk::bits
- shift_bits
);
168 for (size_t n
= result
.chunks
; n
> 0; n
--) {
169 result
.data
[n
- 1] = carry
| (data
[shift_chunks
+ n
- 1] >> shift_bits
);
170 carry
= (shift_bits
== 0) ? 0
171 : data
[shift_chunks
+ n
- 1] << (chunk::bits
- shift_bits
);
176 template<size_t NewBits
>
178 value
<NewBits
> rzext() const {
179 static_assert(NewBits
>= Bits
, "rzext() may not decrease width");
180 value
<NewBits
> result
;
181 constexpr size_t shift_chunks
= (NewBits
- Bits
) / chunk::bits
;
182 constexpr size_t shift_bits
= (NewBits
- Bits
) % chunk::bits
;
183 chunk::type carry
= 0;
184 for (size_t n
= 0; n
< chunks
; n
++) {
185 result
.data
[shift_chunks
+ n
] = (data
[n
] << shift_bits
) | carry
;
186 carry
= (shift_bits
== 0) ? 0
187 : data
[n
] >> (chunk::bits
- shift_bits
);
189 if (shift_chunks
+ chunks
< result
.chunks
)
190 result
.data
[shift_chunks
+ chunks
] = carry
;
194 // Bit blit operation, i.e. a partial read-modify-write.
195 template<size_t Stop
, size_t Start
>
197 value
<Bits
> blit(const value
<Stop
- Start
+ 1> &source
) const {
198 static_assert(Stop
>= Start
, "blit() may not reverse bit order");
199 constexpr chunk::type start_mask
= ~(chunk::mask
<< (Start
% chunk::bits
));
200 constexpr chunk::type stop_mask
= (Stop
% chunk::bits
+ 1 == chunk::bits
) ? 0
201 : (chunk::mask
<< (Stop
% chunk::bits
+ 1));
202 value
<Bits
> masked
= *this;
203 if (Start
/ chunk::bits
== Stop
/ chunk::bits
) {
204 masked
.data
[Start
/ chunk::bits
] &= stop_mask
| start_mask
;
206 masked
.data
[Start
/ chunk::bits
] &= start_mask
;
207 for (size_t n
= Start
/ chunk::bits
+ 1; n
< Stop
/ chunk::bits
; n
++)
209 masked
.data
[Stop
/ chunk::bits
] &= stop_mask
;
211 value
<Bits
> shifted
= source
212 .template rzext
<Stop
+ 1>()
213 .template zext
<Bits
>();
214 return masked
.bit_or(shifted
);
217 // Helpers for selecting extending or truncating operation depending on whether the result is wider or narrower
218 // than the operand. In C++17 these can be replaced with `if constexpr`.
219 template<size_t NewBits
, typename
= void>
222 value
<NewBits
> operator()(const value
<Bits
> &val
) {
223 return val
.template zext
<NewBits
>();
227 template<size_t NewBits
>
228 struct zext_cast
<NewBits
, typename
std::enable_if
<(NewBits
< Bits
)>::type
> {
230 value
<NewBits
> operator()(const value
<Bits
> &val
) {
231 return val
.template trunc
<NewBits
>();
235 template<size_t NewBits
, typename
= void>
238 value
<NewBits
> operator()(const value
<Bits
> &val
) {
239 return val
.template sext
<NewBits
>();
243 template<size_t NewBits
>
244 struct sext_cast
<NewBits
, typename
std::enable_if
<(NewBits
< Bits
)>::type
> {
246 value
<NewBits
> operator()(const value
<Bits
> &val
) {
247 return val
.template trunc
<NewBits
>();
251 template<size_t NewBits
>
253 value
<NewBits
> zcast() const {
254 return zext_cast
<NewBits
>()(*this);
257 template<size_t NewBits
>
259 value
<NewBits
> scast() const {
260 return sext_cast
<NewBits
>()(*this);
263 // Operations with run-time parameters (offsets, amounts, etc).
265 // These operations are used for computations.
266 bool bit(size_t offset
) const {
267 return data
[offset
/ chunk::bits
] & (1 << (offset
% chunk::bits
));
270 void set_bit(size_t offset
, bool value
= true) {
271 size_t offset_chunks
= offset
/ chunk::bits
;
272 size_t offset_bits
= offset
% chunk::bits
;
273 data
[offset_chunks
] &= ~(1 << offset_bits
);
274 data
[offset_chunks
] |= value
? 1 << offset_bits
: 0;
277 bool is_zero() const {
278 for (size_t n
= 0; n
< chunks
; n
++)
284 explicit operator bool() const {
288 bool is_neg() const {
289 return data
[chunks
- 1] & (1 << ((Bits
- 1) % chunk::bits
));
292 bool operator ==(const value
<Bits
> &other
) const {
293 for (size_t n
= 0; n
< chunks
; n
++)
294 if (data
[n
] != other
.data
[n
])
299 bool operator !=(const value
<Bits
> &other
) const {
300 return !(*this == other
);
303 value
<Bits
> bit_not() const {
305 for (size_t n
= 0; n
< chunks
; n
++)
306 result
.data
[n
] = ~data
[n
];
307 result
.data
[chunks
- 1] &= msb_mask
;
311 value
<Bits
> bit_and(const value
<Bits
> &other
) const {
313 for (size_t n
= 0; n
< chunks
; n
++)
314 result
.data
[n
] = data
[n
] & other
.data
[n
];
318 value
<Bits
> bit_or(const value
<Bits
> &other
) const {
320 for (size_t n
= 0; n
< chunks
; n
++)
321 result
.data
[n
] = data
[n
] | other
.data
[n
];
325 value
<Bits
> bit_xor(const value
<Bits
> &other
) const {
327 for (size_t n
= 0; n
< chunks
; n
++)
328 result
.data
[n
] = data
[n
] ^ other
.data
[n
];
332 value
<Bits
> update(const value
<Bits
> &val
, const value
<Bits
> &mask
) const {
333 return bit_and(mask
.bit_not()).bit_or(val
.bit_and(mask
));
336 template<size_t AmountBits
>
337 value
<Bits
> shl(const value
<AmountBits
> &amount
) const {
338 // Ensure our early return is correct by prohibiting values larger than 4 Gbit.
339 static_assert(Bits
<= chunk::mask
, "shl() of unreasonably large values is not supported");
340 // Detect shifts definitely large than Bits early.
341 for (size_t n
= 1; n
< amount
.chunks
; n
++)
342 if (amount
.data
[n
] != 0)
344 // Past this point we can use the least significant chunk as the shift size.
345 size_t shift_chunks
= amount
.data
[0] / chunk::bits
;
346 size_t shift_bits
= amount
.data
[0] % chunk::bits
;
347 if (shift_chunks
>= chunks
)
350 chunk::type carry
= 0;
351 for (size_t n
= 0; n
< chunks
- shift_chunks
; n
++) {
352 result
.data
[shift_chunks
+ n
] = (data
[n
] << shift_bits
) | carry
;
353 carry
= (shift_bits
== 0) ? 0
354 : data
[n
] >> (chunk::bits
- shift_bits
);
359 template<size_t AmountBits
, bool Signed
= false>
360 value
<Bits
> shr(const value
<AmountBits
> &amount
) const {
361 // Ensure our early return is correct by prohibiting values larger than 4 Gbit.
362 static_assert(Bits
<= chunk::mask
, "shr() of unreasonably large values is not supported");
363 // Detect shifts definitely large than Bits early.
364 for (size_t n
= 1; n
< amount
.chunks
; n
++)
365 if (amount
.data
[n
] != 0)
367 // Past this point we can use the least significant chunk as the shift size.
368 size_t shift_chunks
= amount
.data
[0] / chunk::bits
;
369 size_t shift_bits
= amount
.data
[0] % chunk::bits
;
370 if (shift_chunks
>= chunks
)
373 chunk::type carry
= 0;
374 for (size_t n
= 0; n
< chunks
- shift_chunks
; n
++) {
375 result
.data
[chunks
- shift_chunks
- 1 - n
] = carry
| (data
[chunks
- 1 - n
] >> shift_bits
);
376 carry
= (shift_bits
== 0) ? 0
377 : data
[chunks
- 1 - n
] << (chunk::bits
- shift_bits
);
379 if (Signed
&& is_neg()) {
380 size_t top_chunk_idx
= (Bits
- shift_bits
) / chunk::bits
;
381 size_t top_chunk_bits
= (Bits
- shift_bits
) % chunk::bits
;
382 for (size_t n
= top_chunk_idx
+ 1; n
< chunks
; n
++)
383 result
.data
[n
] = chunk::mask
;
385 result
.data
[top_chunk_idx
] |= chunk::mask
<< top_chunk_bits
;
390 template<size_t AmountBits
>
391 value
<Bits
> sshr(const value
<AmountBits
> &amount
) const {
392 return shr
<AmountBits
, /*Signed=*/true>(amount
);
395 size_t ctpop() const {
397 for (size_t n
= 0; n
< chunks
; n
++) {
398 // This loop implements the population count idiom as recognized by LLVM and GCC.
399 for (chunk::type x
= data
[n
]; x
!= 0; count
++)
405 size_t ctlz() const {
407 for (size_t n
= 0; n
< chunks
; n
++) {
408 chunk::type x
= data
[chunks
- 1 - n
];
410 count
+= (n
== 0 ? Bits
% chunk::bits
: chunk::bits
);
412 // This loop implements the find first set idiom as recognized by LLVM.
413 for (; x
!= 0; count
++)
420 template<bool Invert
, bool CarryIn
>
421 std::pair
<value
<Bits
>, bool /*CarryOut*/> alu(const value
<Bits
> &other
) const {
423 bool carry
= CarryIn
;
424 for (size_t n
= 0; n
< result
.chunks
; n
++) {
425 result
.data
[n
] = data
[n
] + (Invert
? ~other
.data
[n
] : other
.data
[n
]) + carry
;
426 carry
= (result
.data
[n
] < data
[n
]) ||
427 (result
.data
[n
] == data
[n
] && carry
);
429 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
430 return {result
, carry
};
433 value
<Bits
> add(const value
<Bits
> &other
) const {
434 return alu
</*Invert=*/false, /*CarryIn=*/false>(other
).first
;
437 value
<Bits
> sub(const value
<Bits
> &other
) const {
438 return alu
</*Invert=*/true, /*CarryIn=*/true>(other
).first
;
441 value
<Bits
> neg() const {
442 return value
<Bits
> { 0u }.sub(*this);
445 bool ucmp(const value
<Bits
> &other
) const {
447 std::tie(std::ignore
, carry
) = alu
</*Invert=*/true, /*CarryIn=*/true>(other
);
448 return !carry
; // a.ucmp(b) ≡ a u< b
451 bool scmp(const value
<Bits
> &other
) const {
454 std::tie(result
, carry
) = alu
</*Invert=*/true, /*CarryIn=*/true>(other
);
455 bool overflow
= (is_neg() == !other
.is_neg()) && (is_neg() != result
.is_neg());
456 return result
.is_neg() ^ overflow
; // a.scmp(b) ≡ a s< b
459 template<size_t ResultBits
>
460 value
<ResultBits
> mul(const value
<Bits
> &other
) const {
461 value
<ResultBits
> result
;
462 wide_chunk_t wide_result
[result
.chunks
+ 1] = {};
463 for (size_t n
= 0; n
< chunks
; n
++) {
464 for (size_t m
= 0; m
< chunks
&& n
+ m
< result
.chunks
; m
++) {
465 wide_result
[n
+ m
] += wide_chunk_t(data
[n
]) * wide_chunk_t(other
.data
[m
]);
466 wide_result
[n
+ m
+ 1] += wide_result
[n
+ m
] >> chunk::bits
;
467 wide_result
[n
+ m
] &= chunk::mask
;
470 for (size_t n
= 0; n
< result
.chunks
; n
++) {
471 result
.data
[n
] = wide_result
[n
];
473 result
.data
[result
.chunks
- 1] &= result
.msb_mask
;
478 // Expression template for a slice, usable as lvalue or rvalue, and composable with other expression templates here.
479 template<class T
, size_t Stop
, size_t Start
>
480 struct slice_expr
: public expr_base
<slice_expr
<T
, Stop
, Start
>> {
481 static_assert(Stop
>= Start
, "slice_expr() may not reverse bit order");
482 static_assert(Start
< T::bits
&& Stop
< T::bits
, "slice_expr() must be within bounds");
483 static constexpr size_t bits
= Stop
- Start
+ 1;
487 slice_expr(T
&expr
) : expr(expr
) {}
488 slice_expr(const slice_expr
<T
, Stop
, Start
> &) = delete;
491 operator value
<bits
>() const {
492 return static_cast<const value
<T::bits
> &>(expr
)
493 .template rtrunc
<T::bits
- Start
>()
494 .template trunc
<bits
>();
498 slice_expr
<T
, Stop
, Start
> &operator=(const value
<bits
> &rhs
) {
499 // Generic partial assignment implemented using a read-modify-write operation on the sliced expression.
500 expr
= static_cast<const value
<T::bits
> &>(expr
)
501 .template blit
<Stop
, Start
>(rhs
);
505 // A helper that forces the cast to value<>, which allows deduction to work.
507 value
<bits
> val() const {
508 return static_cast<const value
<bits
> &>(*this);
512 // Expression template for a concatenation, usable as lvalue or rvalue, and composable with other expression templates here.
513 template<class T
, class U
>
514 struct concat_expr
: public expr_base
<concat_expr
<T
, U
>> {
515 static constexpr size_t bits
= T::bits
+ U::bits
;
520 concat_expr(T
&ms_expr
, U
&ls_expr
) : ms_expr(ms_expr
), ls_expr(ls_expr
) {}
521 concat_expr(const concat_expr
<T
, U
> &) = delete;
524 operator value
<bits
>() const {
525 value
<bits
> ms_shifted
= static_cast<const value
<T::bits
> &>(ms_expr
)
526 .template rzext
<bits
>();
527 value
<bits
> ls_extended
= static_cast<const value
<U::bits
> &>(ls_expr
)
528 .template zext
<bits
>();
529 return ms_shifted
.bit_or(ls_extended
);
533 concat_expr
<T
, U
> &operator=(const value
<bits
> &rhs
) {
534 ms_expr
= rhs
.template rtrunc
<T::bits
>();
535 ls_expr
= rhs
.template trunc
<U::bits
>();
539 // A helper that forces the cast to value<>, which allows deduction to work.
541 value
<bits
> val() const {
542 return static_cast<const value
<bits
> &>(*this);
546 // Base class for expression templates, providing helper methods for operations that are valid on both rvalues and lvalues.
548 // Note that expression objects (slices and concatenations) constructed in this way should NEVER be captured because
549 // they refer to temporaries that will, in general, only live until the end of the statement. For example, both of
550 // these snippets perform use-after-free:
552 // const auto &a = val.slice<7,0>().slice<1>();
555 // auto &&c = val.slice<7,0>().slice<1>();
558 // An easy way to write code using slices and concatenations safely is to follow two simple rules:
559 // * Never explicitly name any type except `value<W>` or `const value<W> &`.
560 // * Never use a `const auto &` or `auto &&` in any such expression.
561 // Then, any code that compiles will be well-defined.
564 template<size_t Stop
, size_t Start
= Stop
>
566 slice_expr
<const T
, Stop
, Start
> slice() const {
567 return {*static_cast<const T
*>(this)};
570 template<size_t Stop
, size_t Start
= Stop
>
572 slice_expr
<T
, Stop
, Start
> slice() {
573 return {*static_cast<T
*>(this)};
578 concat_expr
<const T
, typename
std::remove_reference
<const U
>::type
> concat(const U
&other
) const {
579 return {*static_cast<const T
*>(this), other
};
584 concat_expr
<T
, typename
std::remove_reference
<U
>::type
> concat(U
&&other
) {
585 return {*static_cast<T
*>(this), other
};
589 template<size_t Bits
>
590 std::ostream
&operator<<(std::ostream
&os
, const value
<Bits
> &val
) {
591 auto old_flags
= os
.flags(std::ios::right
);
592 auto old_width
= os
.width(0);
593 auto old_fill
= os
.fill('0');
594 os
<< val
.bits
<< '\'' << std::hex
;
595 for (size_t n
= val
.chunks
- 1; n
!= (size_t)-1; n
--) {
596 if (n
== val
.chunks
- 1 && Bits
% value
<Bits
>::chunk::bits
!= 0)
597 os
.width((Bits
% value
<Bits
>::chunk::bits
+ 3) / 4);
599 os
.width((value
<Bits
>::chunk::bits
+ 3) / 4);
608 template<size_t Bits
>
610 static constexpr size_t bits
= Bits
;
616 constexpr wire(const value
<Bits
> &init
) : curr(init
), next(init
) {}
617 template<typename
... Init
>
618 explicit constexpr wire(Init
...init
) : curr
{init
...}, next
{init
...} {}
620 wire(const wire
<Bits
> &) = delete;
621 wire(wire
<Bits
> &&) = default;
622 wire
<Bits
> &operator=(const wire
<Bits
> &) = delete;
633 template<size_t Bits
>
634 std::ostream
&operator<<(std::ostream
&os
, const wire
<Bits
> &val
) {
639 template<size_t Width
>
641 std::vector
<value
<Width
>> data
;
643 size_t depth() const {
648 explicit memory(size_t depth
) : data(depth
) {}
650 memory(const memory
<Width
> &) = delete;
651 memory
<Width
> &operator=(const memory
<Width
> &) = delete;
653 // The only way to get the compiler to put the initializer in .rodata and do not copy it on stack is to stuff it
654 // into a plain array. You'd think an std::initializer_list would work here, but it doesn't, because you can't
655 // construct an initializer_list in a constexpr (or something) and so if you try to do that the whole thing is
656 // first copied on the stack (probably overflowing it) and then again into `data`.
657 template<size_t Size
>
660 value
<Width
> data
[Size
];
663 template<size_t... InitSize
>
664 explicit memory(size_t depth
, const init
<InitSize
> &...init
) : data(depth
) {
666 // This utterly reprehensible construct is the most reasonable way to apply a function to every element
667 // of a parameter pack, if the elements all have different types and so cannot be cast to an initializer list.
668 auto _
= {std::move(std::begin(init
.data
), std::end(init
.data
), data
.begin() + init
.offset
)...};
672 // An operator for direct memory reads. May be used at any time during the simulation.
673 const value
<Width
> &operator [](size_t index
) const {
674 assert(index
< data
.size());
678 // An operator for direct memory writes. May only be used before the simulation is started. If used
679 // after the simulation is started, the design may malfunction.
680 value
<Width
> &operator [](size_t index
) {
681 assert(index
< data
.size());
685 // A simple way to make a writable memory would be to use an array of wires instead of an array of values.
686 // However, there are two significant downsides to this approach: first, it has large overhead (2× space
687 // overhead, and O(depth) time overhead during commit); second, it does not simplify handling write port
688 // priorities. Although in principle write ports could be ordered or conditionally enabled in generated
689 // code based on their priorities and selected addresses, the feedback arc set problem is computationally
690 // expensive, and the heuristic based algorithms are not easily modified to guarantee (rather than prefer)
691 // a particular write port evaluation order.
693 // The approach used here instead is to queue writes into a buffer during the eval phase, then perform
694 // the writes during the commit phase in the priority order. This approach has low overhead, with both space
695 // and time proportional to the amount of write ports. Because virtually every memory in a practical design
696 // has at most two write ports, linear search is used on every write, being the fastest and simplest approach.
703 std::vector
<write
> write_queue
;
705 void update(size_t index
, const value
<Width
> &val
, const value
<Width
> &mask
, int priority
= 0) {
706 assert(index
< data
.size());
707 // Queue up the write while keeping the queue sorted by priority.
709 std::upper_bound(write_queue
.begin(), write_queue
.end(), priority
,
710 [](const int a
, const write
& b
) { return a
< b
.priority
; }),
711 write
{ index
, val
, mask
, priority
});
715 bool changed
= false;
716 for (const write
&entry
: write_queue
) {
717 value
<Width
> elem
= data
[entry
.index
];
718 elem
= elem
.update(entry
.val
, entry
.mask
);
719 changed
|= (data
[entry
.index
] != elem
);
720 data
[entry
.index
] = elem
;
736 // In debug mode, using the wrong .as_*() function will assert.
737 // In release mode, using the wrong .as_*() function will safely return a default value.
738 const unsigned uint_value
= 0;
739 const signed sint_value
= 0;
740 const std::string string_value
= "";
741 const double double_value
= 0.0;
743 metadata() : value_type(MISSING
) {}
744 metadata(unsigned value
) : value_type(UINT
), uint_value(value
) {}
745 metadata(signed value
) : value_type(SINT
), sint_value(value
) {}
746 metadata(const std::string
&value
) : value_type(STRING
), string_value(value
) {}
747 metadata(const char *value
) : value_type(STRING
), string_value(value
) {}
748 metadata(double value
) : value_type(DOUBLE
), double_value(value
) {}
750 metadata(const metadata
&) = default;
751 metadata
&operator=(const metadata
&) = delete;
753 unsigned as_uint() const {
754 assert(value_type
== UINT
);
758 signed as_sint() const {
759 assert(value_type
== SINT
);
763 const std::string
&as_string() const {
764 assert(value_type
== STRING
);
768 double as_double() const {
769 assert(value_type
== DOUBLE
);
774 typedef std::map
<std::string
, metadata
> metadata_map
;
776 // Helper class to disambiguate values/wires and their aliases.
777 struct debug_alias
{};
779 // This structure is intended for consumption via foreign function interfaces, like Python's ctypes.
780 // Because of this it uses a C-style layout that is easy to parse rather than more idiomatic C++.
782 // To avoid violating strict aliasing rules, this structure has to be a subclass of the one used
783 // in the C API, or it would not be possible to cast between the pointers to these.
784 struct debug_item
: ::cxxrtl_object
{
786 VALUE
= CXXRTL_VALUE
,
788 MEMORY
= CXXRTL_MEMORY
,
789 ALIAS
= CXXRTL_ALIAS
,
792 debug_item(const ::cxxrtl_object
&object
) : cxxrtl_object(object
) {}
794 template<size_t Bits
>
795 debug_item(value
<Bits
> &item
, size_t lsb_offset
= 0) {
796 static_assert(sizeof(item
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
797 "value<Bits> is not compatible with C layout");
807 template<size_t Bits
>
808 debug_item(const value
<Bits
> &item
, size_t lsb_offset
= 0) {
809 static_assert(sizeof(item
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
810 "value<Bits> is not compatible with C layout");
816 curr
= const_cast<chunk_t
*>(item
.data
);
820 template<size_t Bits
>
821 debug_item(wire
<Bits
> &item
, size_t lsb_offset
= 0) {
822 static_assert(sizeof(item
.curr
) == value
<Bits
>::chunks
* sizeof(chunk_t
) &&
823 sizeof(item
.next
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
824 "wire<Bits> is not compatible with C layout");
830 curr
= item
.curr
.data
;
831 next
= item
.next
.data
;
834 template<size_t Width
>
835 debug_item(memory
<Width
> &item
, size_t zero_offset
= 0) {
836 static_assert(sizeof(item
.data
[0]) == value
<Width
>::chunks
* sizeof(chunk_t
),
837 "memory<Width> is not compatible with C layout");
841 depth
= item
.data
.size();
842 zero_at
= zero_offset
;
843 curr
= item
.data
.empty() ? nullptr : item
.data
[0].data
;
847 template<size_t Bits
>
848 debug_item(debug_alias
, const value
<Bits
> &item
, size_t lsb_offset
= 0) {
849 static_assert(sizeof(item
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
850 "value<Bits> is not compatible with C layout");
856 curr
= const_cast<chunk_t
*>(item
.data
);
860 template<size_t Bits
>
861 debug_item(debug_alias
, const wire
<Bits
> &item
, size_t lsb_offset
= 0) {
862 static_assert(sizeof(item
.curr
) == value
<Bits
>::chunks
* sizeof(chunk_t
) &&
863 sizeof(item
.next
) == value
<Bits
>::chunks
* sizeof(chunk_t
),
864 "wire<Bits> is not compatible with C layout");
870 curr
= const_cast<chunk_t
*>(item
.curr
.data
);
874 static_assert(std::is_standard_layout
<debug_item
>::value
, "debug_item is not compatible with C layout");
877 std::map
<std::string
, std::vector
<debug_item
>> table
;
879 void add(const std::string
&name
, debug_item
&&item
) {
880 std::vector
<debug_item
> &parts
= table
[name
];
881 parts
.emplace_back(item
);
882 std::sort(parts
.begin(), parts
.end(),
883 [](const debug_item
&a
, const debug_item
&b
) {
884 return a
.lsb_at
< b
.lsb_at
;
888 size_t count(const std::string
&name
) const {
889 if (table
.count(name
) == 0)
891 return table
.at(name
).size();
894 const std::vector
<debug_item
> &parts_at(const std::string
&name
) const {
895 return table
.at(name
);
898 const debug_item
&at(const std::string
&name
) const {
899 const std::vector
<debug_item
> &parts
= table
.at(name
);
900 assert(parts
.size() == 1);
904 const debug_item
&operator [](const std::string
&name
) const {
913 module(const module
&) = delete;
914 module
&operator=(const module
&) = delete;
916 virtual bool eval() = 0;
917 virtual bool commit() = 0;
921 bool converged
= false;
925 } while (commit() && !converged
);
929 virtual void debug_info(debug_items
&items
, std::string path
= "") {
930 (void)items
, (void)path
;
934 } // namespace cxxrtl
936 // Internal structure used to communicate with the implementation of the C interface.
937 typedef struct _cxxrtl_toplevel
{
938 std::unique_ptr
<cxxrtl::module
> module
;
941 // Definitions of internal Yosys cells. Other than the functions in this namespace, CXXRTL is fully generic
942 // and indepenent of Yosys implementation details.
944 // The `write_cxxrtl` pass translates internal cells (cells with names that start with `$`) to calls of these
945 // functions. All of Yosys arithmetic and logical cells perform sign or zero extension on their operands,
946 // whereas basic operations on arbitrary width values require operands to be of the same width. These functions
947 // bridge the gap by performing the necessary casts. They are named similar to `cell_A[B]`, where A and B are `u`
948 // if the corresponding operand is unsigned, and `s` if it is signed.
949 namespace cxxrtl_yosys
{
951 using namespace cxxrtl
;
953 // std::max isn't constexpr until C++14 for no particular reason (it's an oversight), so we define our own.
956 constexpr T
max(const T
&a
, const T
&b
) {
957 return a
> b
? a
: b
;
961 template<size_t BitsY
, size_t BitsA
>
963 value
<BitsY
> logic_not(const value
<BitsA
> &a
) {
964 return value
<BitsY
> { a
? 0u : 1u };
967 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
969 value
<BitsY
> logic_and(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
970 return value
<BitsY
> { (bool(a
) & bool(b
)) ? 1u : 0u };
973 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
975 value
<BitsY
> logic_or(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
976 return value
<BitsY
> { (bool(a
) | bool(b
)) ? 1u : 0u };
979 // Reduction operations
980 template<size_t BitsY
, size_t BitsA
>
982 value
<BitsY
> reduce_and(const value
<BitsA
> &a
) {
983 return value
<BitsY
> { a
.bit_not().is_zero() ? 1u : 0u };
986 template<size_t BitsY
, size_t BitsA
>
988 value
<BitsY
> reduce_or(const value
<BitsA
> &a
) {
989 return value
<BitsY
> { a
? 1u : 0u };
992 template<size_t BitsY
, size_t BitsA
>
994 value
<BitsY
> reduce_xor(const value
<BitsA
> &a
) {
995 return value
<BitsY
> { (a
.ctpop() % 2) ? 1u : 0u };
998 template<size_t BitsY
, size_t BitsA
>
1000 value
<BitsY
> reduce_xnor(const value
<BitsA
> &a
) {
1001 return value
<BitsY
> { (a
.ctpop() % 2) ? 0u : 1u };
1004 template<size_t BitsY
, size_t BitsA
>
1005 CXXRTL_ALWAYS_INLINE
1006 value
<BitsY
> reduce_bool(const value
<BitsA
> &a
) {
1007 return value
<BitsY
> { a
? 1u : 0u };
1010 // Bitwise operations
1011 template<size_t BitsY
, size_t BitsA
>
1012 CXXRTL_ALWAYS_INLINE
1013 value
<BitsY
> not_u(const value
<BitsA
> &a
) {
1014 return a
.template zcast
<BitsY
>().bit_not();
1017 template<size_t BitsY
, size_t BitsA
>
1018 CXXRTL_ALWAYS_INLINE
1019 value
<BitsY
> not_s(const value
<BitsA
> &a
) {
1020 return a
.template scast
<BitsY
>().bit_not();
1023 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1024 CXXRTL_ALWAYS_INLINE
1025 value
<BitsY
> and_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1026 return a
.template zcast
<BitsY
>().bit_and(b
.template zcast
<BitsY
>());
1029 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1030 CXXRTL_ALWAYS_INLINE
1031 value
<BitsY
> and_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1032 return a
.template scast
<BitsY
>().bit_and(b
.template scast
<BitsY
>());
1035 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1036 CXXRTL_ALWAYS_INLINE
1037 value
<BitsY
> or_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1038 return a
.template zcast
<BitsY
>().bit_or(b
.template zcast
<BitsY
>());
1041 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1042 CXXRTL_ALWAYS_INLINE
1043 value
<BitsY
> or_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1044 return a
.template scast
<BitsY
>().bit_or(b
.template scast
<BitsY
>());
1047 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1048 CXXRTL_ALWAYS_INLINE
1049 value
<BitsY
> xor_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1050 return a
.template zcast
<BitsY
>().bit_xor(b
.template zcast
<BitsY
>());
1053 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1054 CXXRTL_ALWAYS_INLINE
1055 value
<BitsY
> xor_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1056 return a
.template scast
<BitsY
>().bit_xor(b
.template scast
<BitsY
>());
1059 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1060 CXXRTL_ALWAYS_INLINE
1061 value
<BitsY
> xnor_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1062 return a
.template zcast
<BitsY
>().bit_xor(b
.template zcast
<BitsY
>()).bit_not();
1065 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1066 CXXRTL_ALWAYS_INLINE
1067 value
<BitsY
> xnor_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1068 return a
.template scast
<BitsY
>().bit_xor(b
.template scast
<BitsY
>()).bit_not();
1071 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1072 CXXRTL_ALWAYS_INLINE
1073 value
<BitsY
> shl_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1074 return a
.template zcast
<BitsY
>().template shl(b
);
1077 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1078 CXXRTL_ALWAYS_INLINE
1079 value
<BitsY
> shl_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1080 return a
.template scast
<BitsY
>().template shl(b
);
1083 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1084 CXXRTL_ALWAYS_INLINE
1085 value
<BitsY
> sshl_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1086 return a
.template zcast
<BitsY
>().template shl(b
);
1089 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1090 CXXRTL_ALWAYS_INLINE
1091 value
<BitsY
> sshl_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1092 return a
.template scast
<BitsY
>().template shl(b
);
1095 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1096 CXXRTL_ALWAYS_INLINE
1097 value
<BitsY
> shr_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1098 return a
.template shr(b
).template zcast
<BitsY
>();
1101 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1102 CXXRTL_ALWAYS_INLINE
1103 value
<BitsY
> shr_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1104 return a
.template shr(b
).template scast
<BitsY
>();
1107 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1108 CXXRTL_ALWAYS_INLINE
1109 value
<BitsY
> sshr_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1110 return a
.template shr(b
).template zcast
<BitsY
>();
1113 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1114 CXXRTL_ALWAYS_INLINE
1115 value
<BitsY
> sshr_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1116 return a
.template sshr(b
).template scast
<BitsY
>();
1119 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1120 CXXRTL_ALWAYS_INLINE
1121 value
<BitsY
> shift_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1122 return shr_uu
<BitsY
>(a
, b
);
1125 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1126 CXXRTL_ALWAYS_INLINE
1127 value
<BitsY
> shift_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1128 return shr_su
<BitsY
>(a
, b
);
1131 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1132 CXXRTL_ALWAYS_INLINE
1133 value
<BitsY
> shift_us(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1134 return b
.is_neg() ? shl_uu
<BitsY
>(a
, b
.template sext
<BitsB
+ 1>().neg()) : shr_uu
<BitsY
>(a
, b
);
1137 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1138 CXXRTL_ALWAYS_INLINE
1139 value
<BitsY
> shift_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1140 return b
.is_neg() ? shl_su
<BitsY
>(a
, b
.template sext
<BitsB
+ 1>().neg()) : shr_su
<BitsY
>(a
, b
);
1143 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1144 CXXRTL_ALWAYS_INLINE
1145 value
<BitsY
> shiftx_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1146 return shift_uu
<BitsY
>(a
, b
);
1149 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1150 CXXRTL_ALWAYS_INLINE
1151 value
<BitsY
> shiftx_su(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1152 return shift_su
<BitsY
>(a
, b
);
1155 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1156 CXXRTL_ALWAYS_INLINE
1157 value
<BitsY
> shiftx_us(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1158 return shift_us
<BitsY
>(a
, b
);
1161 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1162 CXXRTL_ALWAYS_INLINE
1163 value
<BitsY
> shiftx_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1164 return shift_ss
<BitsY
>(a
, b
);
1167 // Comparison operations
1168 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1169 CXXRTL_ALWAYS_INLINE
1170 value
<BitsY
> eq_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1171 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1172 return value
<BitsY
>{ a
.template zext
<BitsExt
>() == b
.template zext
<BitsExt
>() ? 1u : 0u };
1175 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1176 CXXRTL_ALWAYS_INLINE
1177 value
<BitsY
> eq_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1178 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1179 return value
<BitsY
>{ a
.template sext
<BitsExt
>() == b
.template sext
<BitsExt
>() ? 1u : 0u };
1182 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1183 CXXRTL_ALWAYS_INLINE
1184 value
<BitsY
> ne_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1185 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1186 return value
<BitsY
>{ a
.template zext
<BitsExt
>() != b
.template zext
<BitsExt
>() ? 1u : 0u };
1189 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1190 CXXRTL_ALWAYS_INLINE
1191 value
<BitsY
> ne_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1192 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1193 return value
<BitsY
>{ a
.template sext
<BitsExt
>() != b
.template sext
<BitsExt
>() ? 1u : 0u };
1196 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1197 CXXRTL_ALWAYS_INLINE
1198 value
<BitsY
> eqx_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1199 return eq_uu
<BitsY
>(a
, b
);
1202 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1203 CXXRTL_ALWAYS_INLINE
1204 value
<BitsY
> eqx_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1205 return eq_ss
<BitsY
>(a
, b
);
1208 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1209 CXXRTL_ALWAYS_INLINE
1210 value
<BitsY
> nex_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1211 return ne_uu
<BitsY
>(a
, b
);
1214 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1215 CXXRTL_ALWAYS_INLINE
1216 value
<BitsY
> nex_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1217 return ne_ss
<BitsY
>(a
, b
);
1220 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1221 CXXRTL_ALWAYS_INLINE
1222 value
<BitsY
> gt_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1223 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1224 return value
<BitsY
> { b
.template zext
<BitsExt
>().ucmp(a
.template zext
<BitsExt
>()) ? 1u : 0u };
1227 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1228 CXXRTL_ALWAYS_INLINE
1229 value
<BitsY
> gt_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1230 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1231 return value
<BitsY
> { b
.template sext
<BitsExt
>().scmp(a
.template sext
<BitsExt
>()) ? 1u : 0u };
1234 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1235 CXXRTL_ALWAYS_INLINE
1236 value
<BitsY
> ge_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1237 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1238 return value
<BitsY
> { !a
.template zext
<BitsExt
>().ucmp(b
.template zext
<BitsExt
>()) ? 1u : 0u };
1241 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1242 CXXRTL_ALWAYS_INLINE
1243 value
<BitsY
> ge_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1244 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1245 return value
<BitsY
> { !a
.template sext
<BitsExt
>().scmp(b
.template sext
<BitsExt
>()) ? 1u : 0u };
1248 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1249 CXXRTL_ALWAYS_INLINE
1250 value
<BitsY
> lt_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1251 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1252 return value
<BitsY
> { a
.template zext
<BitsExt
>().ucmp(b
.template zext
<BitsExt
>()) ? 1u : 0u };
1255 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1256 CXXRTL_ALWAYS_INLINE
1257 value
<BitsY
> lt_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1258 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1259 return value
<BitsY
> { a
.template sext
<BitsExt
>().scmp(b
.template sext
<BitsExt
>()) ? 1u : 0u };
1262 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1263 CXXRTL_ALWAYS_INLINE
1264 value
<BitsY
> le_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1265 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1266 return value
<BitsY
> { !b
.template zext
<BitsExt
>().ucmp(a
.template zext
<BitsExt
>()) ? 1u : 0u };
1269 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1270 CXXRTL_ALWAYS_INLINE
1271 value
<BitsY
> le_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1272 constexpr size_t BitsExt
= max(BitsA
, BitsB
);
1273 return value
<BitsY
> { !b
.template sext
<BitsExt
>().scmp(a
.template sext
<BitsExt
>()) ? 1u : 0u };
1276 // Arithmetic operations
1277 template<size_t BitsY
, size_t BitsA
>
1278 CXXRTL_ALWAYS_INLINE
1279 value
<BitsY
> pos_u(const value
<BitsA
> &a
) {
1280 return a
.template zcast
<BitsY
>();
1283 template<size_t BitsY
, size_t BitsA
>
1284 CXXRTL_ALWAYS_INLINE
1285 value
<BitsY
> pos_s(const value
<BitsA
> &a
) {
1286 return a
.template scast
<BitsY
>();
1289 template<size_t BitsY
, size_t BitsA
>
1290 CXXRTL_ALWAYS_INLINE
1291 value
<BitsY
> neg_u(const value
<BitsA
> &a
) {
1292 return a
.template zcast
<BitsY
>().neg();
1295 template<size_t BitsY
, size_t BitsA
>
1296 CXXRTL_ALWAYS_INLINE
1297 value
<BitsY
> neg_s(const value
<BitsA
> &a
) {
1298 return a
.template scast
<BitsY
>().neg();
1301 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1302 CXXRTL_ALWAYS_INLINE
1303 value
<BitsY
> add_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1304 return a
.template zcast
<BitsY
>().add(b
.template zcast
<BitsY
>());
1307 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1308 CXXRTL_ALWAYS_INLINE
1309 value
<BitsY
> add_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1310 return a
.template scast
<BitsY
>().add(b
.template scast
<BitsY
>());
1313 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1314 CXXRTL_ALWAYS_INLINE
1315 value
<BitsY
> sub_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1316 return a
.template zcast
<BitsY
>().sub(b
.template zcast
<BitsY
>());
1319 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1320 CXXRTL_ALWAYS_INLINE
1321 value
<BitsY
> sub_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1322 return a
.template scast
<BitsY
>().sub(b
.template scast
<BitsY
>());
1325 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1326 CXXRTL_ALWAYS_INLINE
1327 value
<BitsY
> mul_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1328 constexpr size_t BitsM
= BitsA
>= BitsB
? BitsA
: BitsB
;
1329 return a
.template zcast
<BitsM
>().template mul
<BitsY
>(b
.template zcast
<BitsM
>());
1332 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1333 CXXRTL_ALWAYS_INLINE
1334 value
<BitsY
> mul_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1335 return a
.template scast
<BitsY
>().template mul
<BitsY
>(b
.template scast
<BitsY
>());
1338 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1339 CXXRTL_ALWAYS_INLINE
1340 std::pair
<value
<BitsY
>, value
<BitsY
>> divmod_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1341 constexpr size_t Bits
= max(BitsY
, max(BitsA
, BitsB
));
1342 value
<Bits
> quotient
;
1343 value
<Bits
> dividend
= a
.template zext
<Bits
>();
1344 value
<Bits
> divisor
= b
.template zext
<Bits
>();
1345 if (dividend
.ucmp(divisor
))
1346 return {/*quotient=*/value
<BitsY
> { 0u }, /*remainder=*/dividend
.template trunc
<BitsY
>()};
1347 uint32_t divisor_shift
= dividend
.ctlz() - divisor
.ctlz();
1348 divisor
= divisor
.shl(value
<32> { divisor_shift
});
1349 for (size_t step
= 0; step
<= divisor_shift
; step
++) {
1350 quotient
= quotient
.shl(value
<1> { 1u });
1351 if (!dividend
.ucmp(divisor
)) {
1352 dividend
= dividend
.sub(divisor
);
1353 quotient
.set_bit(0, true);
1355 divisor
= divisor
.shr(value
<1> { 1u });
1357 return {quotient
.template trunc
<BitsY
>(), /*remainder=*/dividend
.template trunc
<BitsY
>()};
1360 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1361 CXXRTL_ALWAYS_INLINE
1362 std::pair
<value
<BitsY
>, value
<BitsY
>> divmod_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1363 value
<BitsA
+ 1> ua
= a
.template sext
<BitsA
+ 1>();
1364 value
<BitsB
+ 1> ub
= b
.template sext
<BitsB
+ 1>();
1365 if (ua
.is_neg()) ua
= ua
.neg();
1366 if (ub
.is_neg()) ub
= ub
.neg();
1368 std::tie(y
, r
) = divmod_uu
<BitsY
>(ua
, ub
);
1369 if (a
.is_neg() != b
.is_neg()) y
= y
.neg();
1370 if (a
.is_neg()) r
= r
.neg();
1374 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1375 CXXRTL_ALWAYS_INLINE
1376 value
<BitsY
> div_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1377 return divmod_uu
<BitsY
>(a
, b
).first
;
1380 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1381 CXXRTL_ALWAYS_INLINE
1382 value
<BitsY
> div_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1383 return divmod_ss
<BitsY
>(a
, b
).first
;
1386 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1387 CXXRTL_ALWAYS_INLINE
1388 value
<BitsY
> mod_uu(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1389 return divmod_uu
<BitsY
>(a
, b
).second
;
1392 template<size_t BitsY
, size_t BitsA
, size_t BitsB
>
1393 CXXRTL_ALWAYS_INLINE
1394 value
<BitsY
> mod_ss(const value
<BitsA
> &a
, const value
<BitsB
> &b
) {
1395 return divmod_ss
<BitsY
>(a
, b
).second
;
1399 struct memory_index
{
1403 template<size_t BitsAddr
>
1404 memory_index(const value
<BitsAddr
> &addr
, size_t offset
, size_t depth
) {
1405 static_assert(value
<BitsAddr
>::chunks
<= 1, "memory address is too wide");
1406 size_t offset_index
= addr
.data
[0];
1408 valid
= (offset_index
>= offset
&& offset_index
< offset
+ depth
);
1409 index
= offset_index
- offset
;
1413 } // namespace cxxrtl_yosys