aboutsummaryrefslogtreecommitdiffstats
path: root/backends/cxxrtl/cxxrtl.h
diff options
context:
space:
mode:
Diffstat (limited to 'backends/cxxrtl/cxxrtl.h')
-rw-r--r--backends/cxxrtl/cxxrtl.h243
1 files changed, 191 insertions, 52 deletions
diff --git a/backends/cxxrtl/cxxrtl.h b/backends/cxxrtl/cxxrtl.h
index f0d7b9fc7..b4ffa87cd 100644
--- a/backends/cxxrtl/cxxrtl.h
+++ b/backends/cxxrtl/cxxrtl.h
@@ -36,22 +36,48 @@
#include <map>
#include <algorithm>
#include <memory>
+#include <functional>
#include <sstream>
#include <backends/cxxrtl/cxxrtl_capi.h>
+#ifndef __has_attribute
+# define __has_attribute(x) 0
+#endif
+
// CXXRTL essentially uses the C++ compiler as a hygienic macro engine that feeds an instruction selector.
// It generates a lot of specialized template functions with relatively large bodies that, when inlined
// into the caller and (for those with loops) unrolled, often expose many new optimization opportunities.
// Because of this, most of the CXXRTL runtime must be always inlined for best performance.
-#ifndef __has_attribute
-# define __has_attribute(x) 0
-#endif
#if __has_attribute(always_inline)
#define CXXRTL_ALWAYS_INLINE inline __attribute__((__always_inline__))
#else
#define CXXRTL_ALWAYS_INLINE inline
#endif
+// Conversely, some functions in the generated code are extremely large yet very cold, with both of these
+// properties being extreme enough to confuse C++ compilers into spending pathological amounts of time
+// on a futile (the code becomes worse) attempt to optimize the least important parts of code.
+#if __has_attribute(optnone)
+#define CXXRTL_EXTREMELY_COLD __attribute__((__optnone__))
+#elif __has_attribute(optimize)
+#define CXXRTL_EXTREMELY_COLD __attribute__((__optimize__(0)))
+#else
+#define CXXRTL_EXTREMELY_COLD
+#endif
+
+// CXXRTL uses assert() to check for C++ contract violations (which may result in e.g. undefined behavior
+// of the simulation code itself), and CXXRTL_ASSERT to check for RTL contract violations (which may at
+// most result in undefined simulation results).
+//
+// Though by default, CXXRTL_ASSERT() expands to assert(), it may be overridden e.g. when integrating
+// the simulation into another process that should survive violating RTL contracts.
+#ifndef CXXRTL_ASSERT
+#ifndef CXXRTL_NDEBUG
+#define CXXRTL_ASSERT(x) assert(x)
+#else
+#define CXXRTL_ASSERT(x)
+#endif
+#endif
namespace cxxrtl {
@@ -96,9 +122,11 @@ struct value : public expr_base<value<Bits>> {
explicit constexpr value(Init ...init) : data{init...} {}
value(const value<Bits> &) = default;
- value(value<Bits> &&) = default;
value<Bits> &operator=(const value<Bits> &) = default;
+ value(value<Bits> &&) = default;
+ value<Bits> &operator=(value<Bits> &&) = default;
+
// A (no-op) helper that forces the cast to value<>.
CXXRTL_ALWAYS_INLINE
const value<Bits> &val() const {
@@ -289,6 +317,14 @@ struct value : public expr_base<value<Bits>> {
return sext_cast<NewBits>()(*this);
}
+ // Bit replication is far more efficient than the equivalent concatenation.
+ template<size_t Count>
+ CXXRTL_ALWAYS_INLINE
+ value<Bits * Count> repeat() const {
+ static_assert(Bits == 1, "repeat() is implemented only for 1-bit values");
+ return *this ? value<Bits * Count>().bit_not() : value<Bits * Count>();
+ }
+
// Operations with run-time parameters (offsets, amounts, etc).
//
// These operations are used for computations.
@@ -421,6 +457,42 @@ struct value : public expr_base<value<Bits>> {
return shr<AmountBits, /*Signed=*/true>(amount);
}
+ template<size_t ResultBits, size_t SelBits>
+ value<ResultBits> bmux(const value<SelBits> &sel) const {
+ static_assert(ResultBits << SelBits == Bits, "invalid sizes used in bmux()");
+ size_t amount = sel.data[0] * ResultBits;
+ size_t shift_chunks = amount / chunk::bits;
+ size_t shift_bits = amount % chunk::bits;
+ value<ResultBits> result;
+ chunk::type carry = 0;
+ if (ResultBits % chunk::bits + shift_bits > chunk::bits)
+ carry = data[result.chunks + shift_chunks] << (chunk::bits - shift_bits);
+ for (size_t n = 0; n < result.chunks; n++) {
+ result.data[result.chunks - 1 - n] = carry | (data[result.chunks + shift_chunks - 1 - n] >> shift_bits);
+ carry = (shift_bits == 0) ? 0
+ : data[result.chunks + shift_chunks - 1 - n] << (chunk::bits - shift_bits);
+ }
+ return result;
+ }
+
+ template<size_t ResultBits, size_t SelBits>
+ value<ResultBits> demux(const value<SelBits> &sel) const {
+ static_assert(Bits << SelBits == ResultBits, "invalid sizes used in demux()");
+ size_t amount = sel.data[0] * Bits;
+ size_t shift_chunks = amount / chunk::bits;
+ size_t shift_bits = amount % chunk::bits;
+ value<ResultBits> result;
+ chunk::type carry = 0;
+ for (size_t n = 0; n < chunks; n++) {
+ result.data[shift_chunks + n] = (data[n] << shift_bits) | carry;
+ carry = (shift_bits == 0) ? 0
+ : data[n] >> (chunk::bits - shift_bits);
+ }
+ if (Bits % chunk::bits + shift_bits > chunk::bits)
+ result.data[shift_chunks + chunks] = carry;
+ return result;
+ }
+
size_t ctpop() const {
size_t count = 0;
for (size_t n = 0; n < chunks; n++) {
@@ -452,10 +524,11 @@ struct value : public expr_base<value<Bits>> {
bool carry = CarryIn;
for (size_t n = 0; n < result.chunks; n++) {
result.data[n] = data[n] + (Invert ? ~other.data[n] : other.data[n]) + carry;
+ if (result.chunks - 1 == n)
+ result.data[result.chunks - 1] &= result.msb_mask;
carry = (result.data[n] < data[n]) ||
(result.data[n] == data[n] && carry);
}
- result.data[result.chunks - 1] &= result.msb_mask;
return {result, carry};
}
@@ -642,14 +715,20 @@ struct wire {
value<Bits> next;
wire() = default;
- constexpr wire(const value<Bits> &init) : curr(init), next(init) {}
+ explicit constexpr wire(const value<Bits> &init) : curr(init), next(init) {}
template<typename... Init>
explicit constexpr wire(Init ...init) : curr{init...}, next{init...} {}
+ // Copying and copy-assigning values is natural. If, however, a value is replaced with a wire,
+ // e.g. because a module is built with a different optimization level, then existing code could
+ // unintentionally copy a wire instead, which would create a subtle but serious bug. To make sure
+ // this doesn't happen, prohibit copying and copy-assigning wires.
wire(const wire<Bits> &) = delete;
- wire(wire<Bits> &&) = default;
wire<Bits> &operator=(const wire<Bits> &) = delete;
+ wire(wire<Bits> &&) = default;
+ wire<Bits> &operator=(wire<Bits> &&) = default;
+
template<class IntegerT>
CXXRTL_ALWAYS_INLINE
IntegerT get() const {
@@ -679,47 +758,32 @@ std::ostream &operator<<(std::ostream &os, const wire<Bits> &val) {
template<size_t Width>
struct memory {
- std::vector<value<Width>> data;
-
- size_t depth() const {
- return data.size();
- }
+ const size_t depth;
+ std::unique_ptr<value<Width>[]> data;
- memory() = delete;
- explicit memory(size_t depth) : data(depth) {}
+ explicit memory(size_t depth) : depth(depth), data(new value<Width>[depth]) {}
memory(const memory<Width> &) = delete;
memory<Width> &operator=(const memory<Width> &) = delete;
- // The only way to get the compiler to put the initializer in .rodata and do not copy it on stack is to stuff it
- // into a plain array. You'd think an std::initializer_list would work here, but it doesn't, because you can't
- // construct an initializer_list in a constexpr (or something) and so if you try to do that the whole thing is
- // first copied on the stack (probably overflowing it) and then again into `data`.
- template<size_t Size>
- struct init {
- size_t offset;
- value<Width> data[Size];
- };
-
- template<size_t... InitSize>
- explicit memory(size_t depth, const init<InitSize> &...init) : data(depth) {
- data.resize(depth);
- // This utterly reprehensible construct is the most reasonable way to apply a function to every element
- // of a parameter pack, if the elements all have different types and so cannot be cast to an initializer list.
- auto _ = {std::move(std::begin(init.data), std::end(init.data), data.begin() + init.offset)...};
- (void)_;
+ memory(memory<Width> &&) = default;
+ memory<Width> &operator=(memory<Width> &&other) {
+ assert(depth == other.depth);
+ data = std::move(other.data);
+ write_queue = std::move(other.write_queue);
+ return *this;
}
// An operator for direct memory reads. May be used at any time during the simulation.
const value<Width> &operator [](size_t index) const {
- assert(index < data.size());
+ assert(index < depth);
return data[index];
}
// An operator for direct memory writes. May only be used before the simulation is started. If used
// after the simulation is started, the design may malfunction.
value<Width> &operator [](size_t index) {
- assert(index < data.size());
+ assert(index < depth);
return data[index];
}
@@ -744,7 +808,7 @@ struct memory {
std::vector<write> write_queue;
void update(size_t index, const value<Width> &val, const value<Width> &mask, int priority = 0) {
- assert(index < data.size());
+ assert(index < depth);
// Queue up the write while keeping the queue sorted by priority.
write_queue.insert(
std::upper_bound(write_queue.begin(), write_queue.end(), priority,
@@ -814,35 +878,52 @@ struct metadata {
typedef std::map<std::string, metadata> metadata_map;
-// Helper class to disambiguate values/wires and their aliases.
+// Tag class to disambiguate values/wires and their aliases.
struct debug_alias {};
+// Tag declaration to disambiguate values and debug outlines.
+using debug_outline = ::_cxxrtl_outline;
+
// This structure is intended for consumption via foreign function interfaces, like Python's ctypes.
// Because of this it uses a C-style layout that is easy to parse rather than more idiomatic C++.
//
// To avoid violating strict aliasing rules, this structure has to be a subclass of the one used
// in the C API, or it would not be possible to cast between the pointers to these.
struct debug_item : ::cxxrtl_object {
+ // Object types.
enum : uint32_t {
- VALUE = CXXRTL_VALUE,
- WIRE = CXXRTL_WIRE,
- MEMORY = CXXRTL_MEMORY,
- ALIAS = CXXRTL_ALIAS,
+ VALUE = CXXRTL_VALUE,
+ WIRE = CXXRTL_WIRE,
+ MEMORY = CXXRTL_MEMORY,
+ ALIAS = CXXRTL_ALIAS,
+ OUTLINE = CXXRTL_OUTLINE,
+ };
+
+ // Object flags.
+ enum : uint32_t {
+ INPUT = CXXRTL_INPUT,
+ OUTPUT = CXXRTL_OUTPUT,
+ INOUT = CXXRTL_INOUT,
+ DRIVEN_SYNC = CXXRTL_DRIVEN_SYNC,
+ DRIVEN_COMB = CXXRTL_DRIVEN_COMB,
+ UNDRIVEN = CXXRTL_UNDRIVEN,
};
debug_item(const ::cxxrtl_object &object) : cxxrtl_object(object) {}
template<size_t Bits>
- debug_item(value<Bits> &item, size_t lsb_offset = 0) {
+ debug_item(value<Bits> &item, size_t lsb_offset = 0, uint32_t flags_ = 0) {
static_assert(sizeof(item) == value<Bits>::chunks * sizeof(chunk_t),
"value<Bits> is not compatible with C layout");
type = VALUE;
+ flags = flags_;
width = Bits;
lsb_at = lsb_offset;
depth = 1;
zero_at = 0;
curr = item.data;
next = item.data;
+ outline = nullptr;
}
template<size_t Bits>
@@ -850,26 +931,30 @@ struct debug_item : ::cxxrtl_object {
static_assert(sizeof(item) == value<Bits>::chunks * sizeof(chunk_t),
"value<Bits> is not compatible with C layout");
type = VALUE;
+ flags = DRIVEN_COMB;
width = Bits;
lsb_at = lsb_offset;
depth = 1;
zero_at = 0;
curr = const_cast<chunk_t*>(item.data);
next = nullptr;
+ outline = nullptr;
}
template<size_t Bits>
- debug_item(wire<Bits> &item, size_t lsb_offset = 0) {
+ debug_item(wire<Bits> &item, size_t lsb_offset = 0, uint32_t flags_ = 0) {
static_assert(sizeof(item.curr) == value<Bits>::chunks * sizeof(chunk_t) &&
sizeof(item.next) == value<Bits>::chunks * sizeof(chunk_t),
"wire<Bits> is not compatible with C layout");
type = WIRE;
+ flags = flags_;
width = Bits;
lsb_at = lsb_offset;
depth = 1;
zero_at = 0;
curr = item.curr.data;
next = item.next.data;
+ outline = nullptr;
}
template<size_t Width>
@@ -877,12 +962,14 @@ struct debug_item : ::cxxrtl_object {
static_assert(sizeof(item.data[0]) == value<Width>::chunks * sizeof(chunk_t),
"memory<Width> is not compatible with C layout");
type = MEMORY;
+ flags = 0;
width = Width;
lsb_at = 0;
- depth = item.data.size();
+ depth = item.depth;
zero_at = zero_offset;
- curr = item.data.empty() ? nullptr : item.data[0].data;
+ curr = item.data ? item.data[0].data : nullptr;
next = nullptr;
+ outline = nullptr;
}
template<size_t Bits>
@@ -890,12 +977,14 @@ struct debug_item : ::cxxrtl_object {
static_assert(sizeof(item) == value<Bits>::chunks * sizeof(chunk_t),
"value<Bits> is not compatible with C layout");
type = ALIAS;
+ flags = DRIVEN_COMB;
width = Bits;
lsb_at = lsb_offset;
depth = 1;
zero_at = 0;
curr = const_cast<chunk_t*>(item.data);
next = nullptr;
+ outline = nullptr;
}
template<size_t Bits>
@@ -904,12 +993,45 @@ struct debug_item : ::cxxrtl_object {
sizeof(item.next) == value<Bits>::chunks * sizeof(chunk_t),
"wire<Bits> is not compatible with C layout");
type = ALIAS;
+ flags = DRIVEN_COMB;
width = Bits;
lsb_at = lsb_offset;
depth = 1;
zero_at = 0;
curr = const_cast<chunk_t*>(item.curr.data);
next = nullptr;
+ outline = nullptr;
+ }
+
+ template<size_t Bits>
+ debug_item(debug_outline &group, const value<Bits> &item, size_t lsb_offset = 0) {
+ static_assert(sizeof(item) == value<Bits>::chunks * sizeof(chunk_t),
+ "value<Bits> is not compatible with C layout");
+ type = OUTLINE;
+ flags = DRIVEN_COMB;
+ width = Bits;
+ lsb_at = lsb_offset;
+ depth = 1;
+ zero_at = 0;
+ curr = const_cast<chunk_t*>(item.data);
+ next = nullptr;
+ outline = &group;
+ }
+
+ template<size_t Bits, class IntegerT>
+ IntegerT get() const {
+ assert(width == Bits && depth == 1);
+ value<Bits> item;
+ std::copy(curr, curr + value<Bits>::chunks, item.data);
+ return item.template get<IntegerT>();
+ }
+
+ template<size_t Bits, class IntegerT>
+ void set(IntegerT other) const {
+ assert(width == Bits && depth == 1);
+ value<Bits> item;
+ item.template set<IntegerT>(other);
+ std::copy(item.data, item.data + value<Bits>::chunks, next);
}
};
static_assert(std::is_standard_layout<debug_item>::value, "debug_item is not compatible with C layout");
@@ -947,13 +1069,25 @@ struct debug_items {
}
};
+// Tag class to disambiguate the default constructor used by the toplevel module that calls reset(),
+// and the constructor of interior modules that should not call it.
+struct interior {};
+
struct module {
module() {}
virtual ~module() {}
+ // Modules with black boxes cannot be copied. Although not all designs include black boxes,
+ // delete the copy constructor and copy assignment operator to make sure that any downstream
+ // code that manipulates modules doesn't accidentally depend on their availability.
module(const module &) = delete;
module &operator=(const module &) = delete;
+ module(module &&) = default;
+ module &operator=(module &&) = default;
+
+ virtual void reset() = 0;
+
virtual bool eval() = 0;
virtual bool commit() = 0;
@@ -974,11 +1108,16 @@ struct module {
} // namespace cxxrtl
-// Internal structure used to communicate with the implementation of the C interface.
+// Internal structures used to communicate with the implementation of the C interface.
+
typedef struct _cxxrtl_toplevel {
std::unique_ptr<cxxrtl::module> module;
} *cxxrtl_toplevel;
+typedef struct _cxxrtl_outline {
+ std::function<void()> eval;
+} *cxxrtl_outline;
+
// Definitions of internal Yosys cells. Other than the functions in this namespace, CXXRTL is fully generic
// and indepenent of Yosys implementation details.
//
@@ -1112,49 +1251,49 @@ value<BitsY> xnor_ss(const value<BitsA> &a, const value<BitsB> &b) {
template<size_t BitsY, size_t BitsA, size_t BitsB>
CXXRTL_ALWAYS_INLINE
value<BitsY> shl_uu(const value<BitsA> &a, const value<BitsB> &b) {
- return a.template zcast<BitsY>().template shl(b);
+ return a.template zcast<BitsY>().shl(b);
}
template<size_t BitsY, size_t BitsA, size_t BitsB>
CXXRTL_ALWAYS_INLINE
value<BitsY> shl_su(const value<BitsA> &a, const value<BitsB> &b) {
- return a.template scast<BitsY>().template shl(b);
+ return a.template scast<BitsY>().shl(b);
}
template<size_t BitsY, size_t BitsA, size_t BitsB>
CXXRTL_ALWAYS_INLINE
value<BitsY> sshl_uu(const value<BitsA> &a, const value<BitsB> &b) {
- return a.template zcast<BitsY>().template shl(b);
+ return a.template zcast<BitsY>().shl(b);
}
template<size_t BitsY, size_t BitsA, size_t BitsB>
CXXRTL_ALWAYS_INLINE
value<BitsY> sshl_su(const value<BitsA> &a, const value<BitsB> &b) {
- return a.template scast<BitsY>().template shl(b);
+ return a.template scast<BitsY>().shl(b);
}
template<size_t BitsY, size_t BitsA, size_t BitsB>
CXXRTL_ALWAYS_INLINE
value<BitsY> shr_uu(const value<BitsA> &a, const value<BitsB> &b) {
- return a.template shr(b).template zcast<BitsY>();
+ return a.shr(b).template zcast<BitsY>();
}
template<size_t BitsY, size_t BitsA, size_t BitsB>
CXXRTL_ALWAYS_INLINE
value<BitsY> shr_su(const value<BitsA> &a, const value<BitsB> &b) {
- return a.template shr(b).template scast<BitsY>();
+ return a.shr(b).template scast<BitsY>();
}
template<size_t BitsY, size_t BitsA, size_t BitsB>
CXXRTL_ALWAYS_INLINE
value<BitsY> sshr_uu(const value<BitsA> &a, const value<BitsB> &b) {
- return a.template shr(b).template zcast<BitsY>();
+ return a.shr(b).template zcast<BitsY>();
}
template<size_t BitsY, size_t BitsA, size_t BitsB>
CXXRTL_ALWAYS_INLINE
value<BitsY> sshr_su(const value<BitsA> &a, const value<BitsB> &b) {
- return a.template sshr(b).template scast<BitsY>();
+ return a.sshr(b).template scast<BitsY>();
}
template<size_t BitsY, size_t BitsA, size_t BitsB>
generated by cgit v1.2.3 (git 2.25.1) at 2025-04-05 23:54:25 +0000