3
0
Fork 0
mirror of https://github.com/YosysHQ/yosys synced 2026-07-14 19:25:40 +00:00

Merge pull request #209 from Silimate/opt_ffa_pt2

opt_first_fit_alloc updates + tests + some nice refactoring
This commit is contained in:
Akash Levy 2026-07-09 00:04:39 -07:00 committed by GitHub
commit 6e80c5348d
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 1758 additions and 148 deletions

View file

@ -17,21 +17,25 @@
*/
// Shared cut-region matching infrastructure for the functional rewrite
// passes (opt_argmax, opt_priority_onehot, opt_compact_prefix). These passes
// find combinational regions between "cut" signals (module ports, FF data
// pins, or internal buses), verify their function by ConstEval
// fingerprinting, and replace the region while leaving surrounding logic
// untouched.
// passes (opt_argmax, opt_priority_onehot, opt_compact_prefix,
// opt_first_fit_alloc). These passes find combinational regions between
// "cut" signals (module ports, FF data pins, or internal buses), verify
// their function by ConstEval fingerprinting, and replace the region while
// leaving surrounding logic untouched.
//
// This header is designed to be included INSIDE each pass's private
// namespace (after PRIVATE_NAMESPACE_BEGIN), so the shared code has a single
// source without introducing link-level coupling between the passes.
// Tiny shared helpers (is_sequential, clog2_int, const_u64, ...) live in
// rewrite_utils.h so non-cut-region passes can share them too.
//
// All graph walks and fingerprint evaluations are charged against
// per-module work budgets so that adversarial netlist shapes (deep shared
// cones with hundreds of same-width candidate buses) degrade into skipped
// candidates instead of multi-minute runtimes.
#include "passes/opt/rewrite_utils.h"
struct CutRegionWorker
{
struct RootCand {
@ -93,25 +97,6 @@ struct CutRegionWorker
build_indexes();
}
bool is_sequential(Cell *c)
{
return c->type.in(
ID($ff), ID($dff), ID($dffe), ID($adff), ID($adffe),
ID($sdff), ID($sdffe), ID($sdffce), ID($dffsr), ID($dffsre),
ID($_DFF_P_), ID($_DFF_N_),
ID($_DFFE_PP_), ID($_DFFE_PN_), ID($_DFFE_NP_), ID($_DFFE_NN_),
ID($_DFF_PP0_), ID($_DFF_PP1_), ID($_DFF_PN0_), ID($_DFF_PN1_),
ID($_DFF_NP0_), ID($_DFF_NP1_), ID($_DFF_NN0_), ID($_DFF_NN1_),
ID($dlatch), ID($adlatch), ID($dlatchsr),
ID($mem), ID($mem_v2), ID($meminit), ID($meminit_v2),
ID($memrd), ID($memrd_v2), ID($memwr), ID($memwr_v2),
ID($fsm),
ID($assert), ID($assume), ID($cover), ID($live), ID($fair),
ID($print), ID($check),
ID($anyconst), ID($anyseq), ID($allconst), ID($allseq),
ID($initstate));
}
void build_indexes()
{
for (auto c : module->cells()) {
@ -717,11 +702,7 @@ struct CutRegionWorker
}
for (auto c : module->cells()) {
if (!c->type.in(ID($ff), ID($dff), ID($dffe), ID($adff), ID($adffe),
ID($sdff), ID($sdffe), ID($sdffce), ID($dffsr), ID($dffsre),
ID($dlatch), ID($adlatch), ID($dlatchsr)))
continue;
if (!c->hasPort(ID::D))
if (!is_storage_ff(c) || !c->hasPort(ID::D))
continue;
consider_root(sigmap(c->getPort(ID::D)), stringf("%s.D", log_id(c->name)), true);
}
@ -796,26 +777,6 @@ struct CutRegionWorker
return roots;
}
// --- Small numeric helpers shared by the fingerprints. ---
static uint64_t lowmask_u64(int w)
{
if (w <= 0)
return 0;
if (w >= 64)
return ~0ULL;
return (1ULL << w) - 1;
}
static Const const_u64(uint64_t value, int width)
{
vector<State> bits(width, State::S0);
for (int i = 0; i < width && i < 64; i++)
if ((value >> i) & 1ULL)
bits[i] = State::S1;
return Const(bits);
}
SigSpec zext_sig(SigSpec sig, int width)
{
sig = sigmap(sig);

View file

@ -28,19 +28,6 @@
USING_YOSYS_NAMESPACE
PRIVATE_NAMESPACE_BEGIN
static int clog2_int(int x)
{
int r = 0;
while ((1 << r) < x)
r++;
return r;
}
static bool is_power_of_two(int x)
{
return x > 0 && (x & (x - 1)) == 0;
}
static Const packed_table_const(const vector<uint64_t> &values, int elem_width)
{
vector<State> bits(values.size() * elem_width, State::S0);

File diff suppressed because it is too large Load diff

View file

@ -25,6 +25,8 @@
USING_YOSYS_NAMESPACE
PRIVATE_NAMESPACE_BEGIN
#include "passes/opt/rewrite_utils.h"
// Priority-encoder variants the pass recognises.
enum class PEVariant { NONE, CLZ_FULL, CLZ_SHORT, CTZ_FULL, CTZ_SHORT };
@ -38,21 +40,6 @@ static const char* variant_name(PEVariant v) {
}
}
static int clog2_int(int x) {
int r = 0;
while ((1 << r) < x) r++;
return r;
}
// Build an N-bit Const from a uint64_t pattern. Bit i set in `pattern` -> bit i
// of the result. Bits beyond 64 are zero.
static Const u64_const(uint64_t pattern, int N) {
std::vector<State> bits(N, State::S0);
for (int i = 0; i < N && i < 64; i++)
if ((pattern >> i) & 1ULL) bits[i] = State::S1;
return Const(bits);
}
// Return the index of the highest set bit (MSB) of `c`, or -1 if all zero.
static int const_msb_set(const Const& c, int N) {
auto bits = c.to_bits();
@ -99,24 +86,6 @@ struct OptPriEncWorker {
OptPriEncWorker(Module* m) : module(m), sigmap(m) { build_indexes(); }
bool is_sequential(Cell* c) {
return c->type.in(
ID($ff), ID($dff), ID($dffe), ID($adff), ID($adffe),
ID($sdff), ID($sdffe), ID($sdffce), ID($dffsr), ID($dffsre),
ID($_DFF_P_), ID($_DFF_N_),
ID($_DFFE_PP_), ID($_DFFE_PN_), ID($_DFFE_NP_), ID($_DFFE_NN_),
ID($_DFF_PP0_), ID($_DFF_PP1_), ID($_DFF_PN0_), ID($_DFF_PN1_),
ID($_DFF_NP0_), ID($_DFF_NP1_), ID($_DFF_NN0_), ID($_DFF_NN1_),
ID($dlatch), ID($adlatch), ID($dlatchsr),
ID($mem), ID($mem_v2), ID($meminit), ID($meminit_v2),
ID($memrd), ID($memrd_v2), ID($memwr), ID($memwr_v2),
ID($fsm),
ID($assert), ID($assume), ID($cover), ID($live), ID($fair),
ID($print), ID($check),
ID($anyconst), ID($anyseq), ID($allconst), ID($allseq),
ID($initstate));
}
void build_indexes() {
for (auto cell : module->cells()) {
if (is_sequential(cell)) {
@ -213,7 +182,7 @@ struct OptPriEncWorker {
// Build the test-vector deck for an N-bit input.
vector<Const> gen_test_vectors(int N) {
vector<Const> vs;
vs.push_back(u64_const(0, N));
vs.push_back(const_u64(0, N));
for (int k = 0; k < N; k++) {
std::vector<State> bits(N, State::S0);
bits[k] = State::S1;

View file

@ -264,26 +264,21 @@ struct OptPrioKeyWorker {
range = 1ULL << cap;
}
ConstEval ce(module);
uint64_t lfsr = 0x9e3779b97f4a7c15ULL ^ (uintptr_t)read;
// Deterministic seed (not a Cell*): ASLR made uintptr_t seeding flake
// across runs, so -strict could miss the OOR counterexample.
uint64_t lfsr = 0x9e3779b97f4a7c15ULL ^ ((uint64_t)S << 1) ^
((uint64_t)kw << 17) ^ (uint64_t)GetSize(steps);
auto rnd = [&]() {
lfsr ^= lfsr << 13; lfsr ^= lfsr >> 7; lfsr ^= lfsr << 17;
return lfsr;
};
for (int t = 0; t < fp_trials; t++) {
auto trial = [&](int rk, const vector<int> &kv, const vector<int> &gv) -> bool {
ce.push();
int rk = (int)(rnd() % range);
ce.set(read_key, Const(rk, GetSize(read_key)));
vector<int> kv(GetSize(steps));
vector<int> gv(GetSize(steps));
for (int i = 0; i < GetSize(steps); i++) {
kv[i] = (int)(rnd() % range);
ce.set(steps[i].key, Const(kv[i], GetSize(steps[i].key)));
if (steps[i].guard == State::S1) {
gv[i] = 1;
} else {
gv[i] = (int)(rnd() & 1);
if (steps[i].guard != State::S1)
ce.set(SigSpec(steps[i].guard), Const(gv[i], 1));
}
}
SigSpec out(read->getPort(ID::Y));
SigSpec undef;
@ -293,7 +288,39 @@ struct OptPrioKeyWorker {
for (int i = 0; i < GetSize(steps); i++)
if (gv[i] && kv[i] == rk) { expect = 1; break; }
ce.pop();
if (!ok || actual != expect)
return ok && actual == expect;
};
// Strict + non-pow2 S: force OOR key collisions the rewrite would accept
// but the S-bit accumulator cannot store (taken[key>=S] is not a set bit).
if (strict && range > (uint64_t)S && !steps.empty()) {
int oor_n = (int)(range - (uint64_t)S);
int forced = oor_n < 16 ? oor_n : 16;
for (int f = 0; f < forced; f++) {
int rk = S + f;
vector<int> kv(GetSize(steps), 0);
vector<int> gv(GetSize(steps), 0);
kv[0] = rk;
gv[0] = 1;
for (int i = 1; i < GetSize(steps); i++) {
kv[i] = (int)(rnd() % range);
gv[i] = steps[i].guard == State::S1 ? 1 : (int)(rnd() & 1);
}
if (!trial(rk, kv, gv))
return false;
}
}
for (int t = 0; t < fp_trials; t++) {
int rk = (int)(rnd() % range);
vector<int> kv(GetSize(steps));
vector<int> gv(GetSize(steps));
for (int i = 0; i < GetSize(steps); i++) {
kv[i] = (int)(rnd() % range);
if (steps[i].guard == State::S1)
gv[i] = 1;
else
gv[i] = (int)(rnd() & 1);
}
if (!trial(rk, kv, gv))
return false;
}
return true;

View file

@ -29,19 +29,6 @@
USING_YOSYS_NAMESPACE
PRIVATE_NAMESPACE_BEGIN
static int clog2_int(int x)
{
int r = 0;
while ((1 << r) < x)
r++;
return r;
}
static bool is_power_of_two(int x)
{
return x > 0 && (x & (x - 1)) == 0;
}
// Pack a per-lane integer vector into a Const with elem_w bits per lane.
static Const pack_lanes(const vector<int> &vals, int elem_w)
{

105
passes/opt/rewrite_utils.h Normal file
View file

@ -0,0 +1,105 @@
/*
* yosys -- Yosys Open SYnthesis Suite
*
* Copyright (C) 2026 Silimate Inc. <akash@silimate.com>
*
* Permission to use, copy, modify, and/or distribute this software for any
* purpose with or without fee is hereby granted, provided that the above
* copyright notice and this permission notice appear in all copies.
*
* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
*/
// Small shared helpers for Silimate functional-rewrite passes. Include inside
// each pass's PRIVATE_NAMESPACE_BEGIN (same pattern as cut_region.h). Keep
// cone/cut infrastructure in cut_region.h; put only type lists and tiny
// numeric/Const helpers here so passes that do not use CutRegionWorker
// (e.g. opt_prienc) can share them without pulling that machinery in.
//
// Guarded so a pass may include this directly and also via cut_region.h.
#ifndef SILIMATE_REWRITE_UTILS_H
#define SILIMATE_REWRITE_UTILS_H
// Process-level / gate-level cells that terminate combinational cone walks.
static inline bool is_sequential(Cell *c)
{
// Include $aldff/$aldffe: proc emits async-load FFs for async-reset
// flops; opt later rewrites them to $adff. Omitting them lets get_cone
// walk through the launch flop into D/CLK/ALOAD (e.g. full wdata).
return c->type.in(
ID($ff), ID($dff), ID($dffe), ID($adff), ID($adffe),
ID($aldff), ID($aldffe),
ID($sdff), ID($sdffe), ID($sdffce), ID($dffsr), ID($dffsre),
ID($_DFF_P_), ID($_DFF_N_),
ID($_DFFE_PP_), ID($_DFFE_PN_), ID($_DFFE_NP_), ID($_DFFE_NN_),
ID($_DFF_PP0_), ID($_DFF_PP1_), ID($_DFF_PN0_), ID($_DFF_PN1_),
ID($_DFF_NP0_), ID($_DFF_NP1_), ID($_DFF_NN0_), ID($_DFF_NN1_),
ID($dlatch), ID($adlatch), ID($dlatchsr),
ID($mem), ID($mem_v2), ID($meminit), ID($meminit_v2),
ID($memrd), ID($memrd_v2), ID($memwr), ID($memwr_v2),
ID($fsm),
ID($assert), ID($assume), ID($cover), ID($live), ID($fair),
ID($print), ID($check),
ID($anyconst), ID($anyseq), ID($allconst), ID($allseq),
ID($initstate));
}
// Clocked process-level FFs (no $ff, latches, or gate $_DFF_*). Used for
// flop-D gather roots where Q/D semantics matter.
static inline bool is_clocked_ff(Cell *c)
{
return c->type.in(ID($dff), ID($dffe), ID($adff), ID($adffe),
ID($aldff), ID($aldffe),
ID($sdff), ID($sdffe), ID($sdffce), ID($dffsr), ID($dffsre));
}
// Storage cells whose D port is a useful rewrite root (clocked FFs + $ff +
// latches). Broader than is_clocked_ff; narrower than is_sequential.
static inline bool is_storage_ff(Cell *c)
{
return is_clocked_ff(c) ||
c->type.in(ID($ff), ID($dlatch), ID($adlatch), ID($dlatchsr));
}
// Ceiling of log2(x) for x >= 1; clog2_int(1) == 0. Same as Yosys $clog2 for
// positive widths used by these passes.
static inline int clog2_int(int x)
{
int r = 0;
while ((1 << r) < x)
r++;
return r;
}
static inline bool is_power_of_two(int x)
{
return x > 0 && (x & (x - 1)) == 0;
}
static inline uint64_t lowmask_u64(int w)
{
if (w <= 0)
return 0;
if (w >= 64)
return ~0ULL;
return (1ULL << w) - 1;
}
// Pack the low `width` bits of `value` into a Const (bit i <- value bit i).
static inline Const const_u64(uint64_t value, int width)
{
vector<State> bits(width, State::S0);
for (int i = 0; i < width && i < 64; i++)
if ((value >> i) & 1ULL)
bits[i] = State::S1;
return Const(bits);
}
#endif // SILIMATE_REWRITE_UTILS_H

View file

@ -26,17 +26,6 @@
USING_YOSYS_NAMESPACE
PRIVATE_NAMESPACE_BEGIN
static int ceil_log2_int(int v)
{
int r = 0;
int n = 1;
while (n < v) {
n <<= 1;
r++;
}
return r;
}
#include "passes/opt/cut_region.h"
struct OptCompactPrefixWorker : CutRegionWorker
@ -788,7 +777,7 @@ struct OptCompactPrefixWorker : CutRegionWorker
{
ref_cell = rw.anchor;
int width = GetSize(rw.root);
int count_width = ceil_log2_int(width + 1);
int count_width = clog2_int(width + 1);
SigSpec in_s = sigmap(rw.a);
vector<SigSpec> bits;
@ -815,7 +804,7 @@ struct OptCompactPrefixWorker : CutRegionWorker
int width = GetSize(rw.root);
int loop_width = rw.loop_width;
bool en_high = rw.msb_first; // polarity flag (see matching loop)
int count_width = ceil_log2_int(loop_width + 1);
int count_width = clog2_int(loop_width + 1);
SigSpec dis_s = sigmap(rw.a);
SigSpec data_s = sigmap(rw.b);
@ -863,7 +852,7 @@ struct OptCompactPrefixWorker : CutRegionWorker
SigSpec n_s = sigmap(rw.b);
bool msb_first = rw.msb_first;
int cnt_width = ceil_log2_int(width + 1);
int cnt_width = clog2_int(width + 1);
int table_size = 1 << cnt_width;
int cmp_width = std::max(GetSize(n_s), cnt_width);
@ -933,7 +922,7 @@ struct OptCompactPrefixWorker : CutRegionWorker
{
ref_cell = rw.anchor;
int width = GetSize(rw.root);
int cnt_width = ceil_log2_int(width + 1);
int cnt_width = clog2_int(width + 1);
SigSpec en_s = sigmap(rw.a);
SigSpec data_s = sigmap(rw.b);

View file

@ -857,3 +857,356 @@ design -load postopt
select -assert-min 1 w:*ffa_*
design -reset
log -pop
# ============================================================================
# Group I: exclusive saturating first-fit (no category / broadcast)
# ============================================================================
#
# Plain taken[]/done[] scan without same-cat coalescing: each enabled lane
# takes the next free slot until NB slots are exhausted. Later requesters get
# rank 0 / not done. This is the qor_vmw_slot_lane shape.
# I1: exclusive dsel-only, N=8 NB=4 (equiv + fires).
log -header "I1: exclusive saturating allocator N=8 NB=4 (equiv)"
log -push
design -reset
read_verilog -sv <<EOF
module top #(parameter N=8, NB=4, W=2) (
input logic mode,
input logic [N-1:0] req,
output logic [N*W-1:0] dsel_flat
);
logic [N-1:0] en = req & {N{mode}};
logic [W-1:0] dsel [0:N-1];
logic [NB-1:0] taken;
logic [N-1:0] done;
always_comb begin
for (int i=0;i<N;i++) dsel[i] = '0;
taken = '0; done = '0;
for (int i=0;i<N;i++)
if (en[i])
for (int j=0;j<NB;j++)
if (!taken[j] && !done[i]) begin
dsel[i] = W'(j); done[i] = 1'b1; taken[j] = 1'b1;
end
end
for (genvar g=0;g<N;g++) assign dsel_flat[g*W +: W] = dsel[g];
endmodule
EOF
hierarchy -top top
proc
opt
check -assert
equiv_opt -assert opt_first_fit_alloc
design -load postopt
select -assert-min 1 w:*ffa_*
design -reset
log -pop
# I2: exclusive with raw-input enable + sibling done bus, N=16 NB=4.
log -header "I2: exclusive N=16 NB=4 raw-input en + done (equiv)"
log -push
design -reset
read_verilog -sv <<EOF
module top #(parameter N=16, NB=4, W=2) (
input logic [N-1:0] lane_en,
output logic [N*W-1:0] dsel_flat,
output logic [N-1:0] done_o
);
logic [W-1:0] dsel [0:N-1];
logic [NB-1:0] taken;
logic [N-1:0] done;
always_comb begin
for (int i=0;i<N;i++) dsel[i] = '0;
taken = '0; done = '0;
for (int i=0;i<N;i++)
if (lane_en[i])
for (int j=0;j<NB;j++)
if (!taken[j] && !done[i]) begin
dsel[i] = W'(j); done[i] = 1'b1; taken[j] = 1'b1;
end
end
for (genvar g=0;g<N;g++) assign dsel_flat[g*W +: W] = dsel[g];
assign done_o = done;
endmodule
EOF
hierarchy -top top
proc
opt
check -assert
equiv_opt -assert opt_first_fit_alloc
design -load postopt
select -assert-min 1 w:*ffa_*
design -reset
log -pop
# I3: exclusive with non-power-of-two slot budget (NB=3, W=2) -- learn nb=3.
log -header "I3: exclusive NB=3 (non-power-of-two) N=8 (equiv)"
log -push
design -reset
read_verilog -sv <<EOF
module top #(parameter N=8, NB=3, W=2) (
input logic [N-1:0] lane_en,
output logic [N*W-1:0] dsel_flat
);
logic [W-1:0] dsel [0:N-1];
logic [NB-1:0] taken;
logic [N-1:0] done;
always_comb begin
for (int i=0;i<N;i++) dsel[i] = '0;
taken = '0; done = '0;
for (int i=0;i<N;i++)
if (lane_en[i])
for (int j=0;j<NB;j++)
if (!taken[j] && !done[i]) begin
dsel[i] = W'(j); done[i] = 1'b1; taken[j] = 1'b1;
end
end
for (genvar g=0;g<N;g++) assign dsel_flat[g*W +: W] = dsel[g];
endmodule
EOF
hierarchy -top top
proc
opt
check -assert
equiv_opt -assert opt_first_fit_alloc
design -load postopt
select -assert-min 1 w:*ffa_*
design -reset
log -pop
# I4: exclusive last-fit near-miss (scan slots NB-1..0) -> no rewrite.
log -header "I4: exclusive last-fit near-miss -> no rewrite"
log -push
design -reset
read_verilog -sv <<EOF
module top #(parameter N=8, NB=4, W=2) (
input logic [N-1:0] lane_en,
output logic [N*W-1:0] dsel_flat
);
logic [W-1:0] dsel [0:N-1];
logic [NB-1:0] taken;
logic [N-1:0] done;
always_comb begin
for (int i=0;i<N;i++) dsel[i] = '0;
taken = '0; done = '0;
for (int i=0;i<N;i++)
if (lane_en[i])
for (int j=NB-1;j>=0;j--)
if (!taken[j] && !done[i]) begin
dsel[i] = W'(j); done[i] = 1'b1; taken[j] = 1'b1;
end
end
for (genvar g=0;g<N;g++) assign dsel_flat[g*W +: W] = dsel[g];
endmodule
EOF
hierarchy -top top
proc
opt
opt_first_fit_alloc
select -assert-count 0 w:*ffa_*
design -reset
log -pop
# I5: exclusive MSB-first scan (lanes N-1..0), N=8 NB=4 (equiv).
log -header "I5: exclusive MSB-first scan N=8 NB=4 (equiv)"
log -push
design -reset
read_verilog -sv <<EOF
module top #(parameter N=8, NB=4, W=2) (
input logic [N-1:0] lane_en,
output logic [N*W-1:0] dsel_flat
);
logic [W-1:0] dsel [0:N-1];
logic [NB-1:0] taken;
logic [N-1:0] done;
always_comb begin
for (int i=0;i<N;i++) dsel[i] = '0;
taken = '0; done = '0;
for (int i=N-1;i>=0;i--)
if (lane_en[i])
for (int j=0;j<NB;j++)
if (!taken[j] && !done[i]) begin
dsel[i] = W'(j); done[i] = 1'b1; taken[j] = 1'b1;
end
end
for (genvar g=0;g<N;g++) assign dsel_flat[g*W +: W] = dsel[g];
endmodule
EOF
hierarchy -top top
proc
opt
check -assert
equiv_opt -assert opt_first_fit_alloc
design -load postopt
select -assert-min 1 w:*ffa_*
design -reset
log -pop
# I6: exclusive + identity gather of wide lane payloads into NB slots
# (qor_vmw_slot_lane shape: slot_data[j] = data[leader at j]).
# DW != N so per-lane data buses are not mistaken for width-N enables.
log -header "I6: exclusive + identity gather N=8 NB=4 DW=16 (equiv)"
log -push
design -reset
read_verilog -sv <<EOF
module top #(parameter N=8, NB=4, W=2, DW=16) (
input logic [N-1:0] lane_en,
input logic [N*DW-1:0] data_flat,
output logic [N*W-1:0] dsel_flat,
output logic [NB*DW-1:0] slot_flat
);
logic [W-1:0] dsel [0:N-1];
logic [DW-1:0] data [0:N-1];
logic [DW-1:0] slot_data [0:NB-1];
logic [NB-1:0] taken;
logic [N-1:0] done;
for (genvar g=0;g<N;g++) assign data[g] = data_flat[g*DW +: DW];
always_comb begin
for (int i=0;i<N;i++) dsel[i] = '0;
for (int j=0;j<NB;j++) slot_data[j] = '0;
taken = '0; done = '0;
for (int i=0;i<N;i++)
if (lane_en[i])
for (int j=0;j<NB;j++)
if (!taken[j] && !done[i]) begin
dsel[i] = W'(j); done[i] = 1'b1; taken[j] = 1'b1;
slot_data[j] = data[i];
end
end
for (genvar g=0;g<N;g++) assign dsel_flat[g*W +: W] = dsel[g];
for (genvar g=0;g<NB;g++) assign slot_flat[g*DW +: DW] = slot_data[g];
endmodule
EOF
hierarchy -top top
proc
opt
check -assert
equiv_opt -assert opt_first_fit_alloc
design -load postopt
select -assert-min 1 w:*ffa_*
design -reset
log -pop
# I7: exclusive and2 enable (req = a & b), with sibling done — the
# qor_vmw_slot_lane enable shape after opt folds the named `req` wire.
# Equiv covers the and2 fingerprint path; I6 already covers identity gather.
log -header "I7: exclusive and2 enable + done N=16 NB=4 (equiv)"
log -push
design -reset
read_verilog -sv <<EOF
module top #(parameter N=16, NB=4, W=2) (
input logic [N-1:0] a,
input logic [N-1:0] b,
output logic [N*W-1:0] dsel_flat,
output logic [N-1:0] done_o
);
logic [N-1:0] en = a & b;
logic [W-1:0] dsel [0:N-1];
logic [NB-1:0] taken;
logic [N-1:0] done;
always_comb begin
for (int i=0;i<N;i++) dsel[i] = '0;
taken = '0; done = '0;
for (int i=0;i<N;i++)
if (en[i])
for (int j=0;j<NB;j++)
if (!taken[j] && !done[i]) begin
dsel[i] = W'(j); done[i] = 1'b1; taken[j] = 1'b1;
end
end
for (genvar g=0;g<N;g++) assign dsel_flat[g*W +: W] = dsel[g];
assign done_o = done;
endmodule
EOF
hierarchy -top top
proc
opt
check -assert
equiv_opt -assert opt_first_fit_alloc
design -load postopt
select -assert-min 1 w:*ffa_*
design -reset
log -pop
# I8: exclusive and2 enable without a global `opt` — launch flop stays a
# plain $dff (sync reset) so leaf runs are data_q halves. Covers the pre-opt
# and2 path; $aldff sequential handling is exercised by qor_vmw_slot_lane.
log -header "I8: exclusive and2 enable pre-opt N=8 NB=4 (equiv)"
log -push
design -reset
read_verilog -sv <<EOF
module top #(parameter N=8, NB=4, W=2) (
input logic clk,
input logic [2*N-1:0] wdata,
output logic [N*W-1:0] dsel_flat
);
logic [2*N-1:0] data_q;
always_ff @(posedge clk)
data_q <= wdata;
wire [N-1:0] req = data_q[N-1:0] & data_q[2*N-1:N];
logic [W-1:0] dsel [0:N-1];
logic [NB-1:0] taken;
logic [N-1:0] done;
always_comb begin
for (int i=0;i<N;i++) dsel[i] = '0;
taken = '0;
done = '0;
for (int i=0;i<N;i++)
if (req[i])
for (int j=0;j<NB;j++)
if (!taken[j] && !done[i]) begin
dsel[i] = W'(j); done[i] = 1'b1; taken[j] = 1'b1;
end
end
for (genvar g=0;g<N;g++) assign dsel_flat[g*W +: W] = dsel[g];
endmodule
EOF
hierarchy -top top
proc
# Intentionally no `opt`.
check -assert
equiv_opt -assert opt_first_fit_alloc
design -load postopt
select -assert-min 1 w:*ffa_*
design -reset
log -pop
# I9: exclusive with NB > max_therm_nb (8) — forces the binary saturating
# HillisSteele fallback (emit_scan_exclusive_bin / emit_sat_add) instead of
# the thermometer scan used by I1I8.
log -header "I9: exclusive binary-scan fallback N=16 NB=12 (equiv)"
log -push
design -reset
read_verilog -sv <<EOF
module top #(parameter N=16, NB=12, W=4) (
input logic mode,
input logic [N-1:0] req,
output logic [N*W-1:0] dsel_flat
);
logic [N-1:0] en = req & {N{mode}};
logic [W-1:0] dsel [0:N-1];
logic [NB-1:0] taken;
logic [N-1:0] done;
always_comb begin
for (int i=0;i<N;i++) dsel[i] = '0;
taken = '0; done = '0;
for (int i=0;i<N;i++)
if (en[i])
for (int j=0;j<NB;j++)
if (!taken[j] && !done[i]) begin
dsel[i] = W'(j); done[i] = 1'b1; taken[j] = 1'b1;
end
end
for (genvar g=0;g<N;g++) assign dsel_flat[g*W +: W] = dsel[g];
endmodule
EOF
hierarchy -top top
proc
opt
check -assert
equiv_opt -assert opt_first_fit_alloc
design -load postopt
select -assert-min 1 w:*ffa_*
design -reset
log -pop