build_noramlized_v3s4 mostly reviewed.

This commit is contained in:
ed
2026-08-18 16:10:29 -04:00
parent 290bb0e07a
commit bbda5efaea
12 changed files with 382 additions and 170 deletions
+3 -2
View File
@@ -37,6 +37,7 @@
// The 'Yield' sequence for Tape Atoms (mac_yield). // The 'Yield' sequence for Tape Atoms (mac_yield).
#define mac_yield(...) \ #define mac_yield(...) \
load_word(R_AtomJmp, R_TapePtr, 0) \ load_word(R_AtomJmp, R_TapePtr, 0) \
LdSlot_ \
, add_ui_self( R_TapePtr, S_(MipsCode)) \ , add_ui_self( R_TapePtr, S_(MipsCode)) \
, jump_reg( R_AtomJmp) \ , jump_reg( R_AtomJmp) \
, BdSlot_ nop , BdSlot_ nop
@@ -217,7 +218,7 @@ WORD_COUNT(mac_gte_ld_ir123_v3s4, 3)
, mac_gte_mv_from_mac123_v3s4(a) \ , mac_gte_mv_from_mac123_v3s4(a) \
GteDelay_ /* Read MAC1/2/3 → a.xyz (overwrites source-A's load targets) */ \ GteDelay_ /* Read MAC1/2/3 → a.xyz (overwrites source-A's load targets) */ \
, mac_shift_aright_v3s4_self(a, 12) /* Right-shift MAC by 12 (S12.20 → S12.0 OuterProduct12) */ , mac_shift_aright_v3s4_self(a, 12) /* Right-shift MAC by 12 (S12.20 → S12.0 OuterProduct12) */
WORD_COUNT(mac_gte_op_cross_v3s4, 16) WORD_COUNT(mac_gte_op_cross_v3s4, 13)
/* atom_dbg_skip */ /* atom_dbg_skip */
#define mac_gte_store_f3(r_primitive_cursor) \ #define mac_gte_store_f3(r_primitive_cursor) \
@@ -291,7 +292,7 @@ WORD_COUNT(mac_gte_sqr_v3s4, 5)
, shift_aright_var(r_dx, r_dx, r_shift) \ , shift_aright_var(r_dx, r_dx, r_shift) \
, shift_aright_var(r_dy, r_dy, r_shift) \ , shift_aright_var(r_dy, r_dy, r_shift) \
, shift_aright_var(r_dz, r_dz, r_shift) , shift_aright_var(r_dz, r_dz, r_shift)
WORD_COUNT(mac_gte_gpf_scale, 12) WORD_COUNT(mac_gte_gpf_scale, 13)
#define mac_trans_mt3s3s4(r_mtx, r_off, r_t0, r_t1, r_t2) \ #define mac_trans_mt3s3s4(r_mtx, r_off, r_t0, r_t1, r_t2) \
load_word( r_t0, r_off, O_(V3_S4,x)) \ load_word( r_t0, r_off, O_(V3_S4,x)) \
+1 -1
View File
@@ -33,7 +33,7 @@ enum {
atom_offset_example_atom_proc_skip = _atom_offset_example_atom_proc_skip, atom_offset_example_atom_proc_skip = _atom_offset_example_atom_proc_skip,
}; };
// --- atom: build_normalize_v3s4 (65 words) --- // --- atom: build_normalize_v3s4 (63 words) ---
#define _atom_offset_aligned_done_srav_path 3 #define _atom_offset_aligned_done_srav_path 3
#define _atom_offset_srav_path_aligned_done 4 #define _atom_offset_srav_path_aligned_done 4
+41 -58
View File
@@ -272,82 +272,65 @@ typedef Struct_(Binds_NormalizeV3S4) {
U2 dst_offset; /* offset of dst V3_S4 within the BIOS scratchpad */ U2 dst_offset; /* offset of dst V3_S4 within the BIOS scratchpad */
}; };
typedef Struct_(RegUse_build_normalize_v3s4) { typedef Struct_(RegUse_build_normalize_v3s4) {
Reg src_ptr; union { Reg_(V3_S4) res, src; };
Reg dst_ptr; union { Reg t0, src_ptr, mac2; };
Reg src_x; union { Reg t1, dst_ptr; };
Reg src_z; union { Reg t2, dst_offset, mac1, v_sqr_aligned; };
Reg recip_est; union { Reg t3, src_offset, btarget, shift_count, sqrtbl_index; };
Reg norm; union { Reg t4, mac3, v_sqr_sum, half_shift_tmp, inv_len; };
Reg shift; union { Reg t5, lzcr, half_shift; };
union {
Reg v_sqr_aligned, dst_offset;
} t3;
union { Reg mac2; } t4;
union { Reg btarget, shift_count, sqrtbl_index, src_offset; } t5;
}; };
/* ─── Full normalize (all 4 stages inline) ─── /* ─── Full normalize (all 4 stages inline) ───
* Generic 4-stage GTE normalize (SQR → sum+LZCR → align+sqrtbl → GPF+srav). */ * Generic 4-stage GTE normalize (SQR → sum+LZCR → align+sqrtbl → GPF+srav). */
internal MipsAtom* build_normalize_v3s4(AtomArena_R aa, RegUse_build_normalize_v3s4 r) internal MipsAtom* build_normalize_v3s4(AtomArena_R aa, RegUse_build_normalize_v3s4 r)
MipsAtom_Proc_(aa, { MipsAtom_Proc_(aa, {
load_half(r.t5.src_offset, R_TapePtr, O_(Binds_NormalizeV3S4, src_offset)), load_half(r.src_offset, R_TapePtr, O_(Binds_NormalizeV3S4, src_offset)),
load_half(r.t3.dst_offset, R_TapePtr, O_(Binds_NormalizeV3S4, dst_offset)), load_half(r.dst_offset, R_TapePtr, O_(Binds_NormalizeV3S4, dst_offset)),
LdSlot_ add_u(r.src_ptr, R_ScratchBase, r.t5.src_offset), LdSlot_ add_u(r.src_ptr, R_ScratchBase, r.src_offset),
LdSlot_ add_u(r.dst_ptr, R_ScratchBase, r.t3.dst_offset), LdSlot_ add_u(r.dst_ptr, R_ScratchBase, r.dst_offset),
LdSlot_ add_ui_self(R_TapePtr, S_(Binds_NormalizeV3S4)), LdSlot_ add_ui_self(R_TapePtr, S_(Binds_NormalizeV3S4)),
/* Load src.x/y/z from r_src_ptr (caller-determined address) into r_src_x / r_recip_est / r_src_z. mac_load_v3s4(r.src, r.src_ptr, 0),
* r_src_x holds src.x throughout stages 1-2; r_recip_est (which is src.y during this phase) is reused as MAC2 in stage 4. */
mac_load_word_v3(r.src_x, r.recip_est, r.src_z, r.src_ptr, 0),
/* Stage 1: mtc2 src → IR1/2/3, SQR fires. */ /* Stage 1: mtc2 src → IR1/2/3, SQR fires. */
LdSlot_ mac_gte_sqr_v3s4(r.src_x, r.recip_est, r.src_z, LdSlot_ nop), LdSlot_ mac_gte_sqr_v3s4(r.src.x, r.src.y, r.src.z, LdSlot_ nop),
/* Stage 2: mfc2 MAC1/2/3, sum, mtc2 LZCS. */ /* Stage 2: mfc2 MAC1/2/3, sum, mtc2 LZCS. src_ptr is dead; reuse as mac2. */
mac_gte_mv_from_data_r_mac123(r.t3.v_sqr_aligned, r.t4.mac2, r.norm), LdSlot_ nop, mac_gte_mv_from_data_r_mac123(r.mac1, r.mac2, r.mac3), LdSlot_ nop,
add_u_self( r.norm, r.t3.v_sqr_aligned), add_u_self( r.v_sqr_sum, r.mac1),
add_u_self( r.norm, r.t4.mac2), add_u_self( r.v_sqr_sum, r.mac2),
gte_mv_to_data_r( r.norm, C2_LZCS), GteDelay_ nop2, gte_mv_to_data_r( r.v_sqr_sum, C2_LZCS), GteDelay_ nop2,
gte_mv_from_data_r(r.shift, C2_LZCR), GteDelay_ nop, gte_mv_from_data_r(r.lzcr, C2_LZCR), GteDelay_ nop,
/* Stage 3: round LZCR to even, compute half-shift, align |v|² to bit 24. /* Stage 3: even(LZCR), half-shift, align |v|² to bit 24. */
* r_norm holds |v|² sum; r_shift holds the LZCR count from mfc2. mac_lzcr_round_even_half_shift(r.lzcr, r.v_sqr_sum, r.v_sqr_aligned),
* After the component: r_shift = even(LZCR), r_norm = half-shift, r_t3.v_sqr_aligned = |v|². */ add_si( r.btarget, r.lzcr, -24),
mac_lzcr_round_even_half_shift(r.shift, r.norm, r.t3.v_sqr_aligned), branch_lt_zero(r.btarget, atom_offset(aligned_done, srav_path)), BdSlot_ nop, /* bltz → srav_path (LZCR < 24 path) */
add_si( r.t5.btarget, r.shift, -24),
branch_lt_zero(r.t5.btarget, atom_offset(aligned_done, srav_path)), BdSlot_ nop, /* bltz → srav_path (LZCR < 24 path) */
jump_rel(atom_offset(srav_path, aligned_done)), /* b → aligned_done (LZCR >= 24 path) */ jump_rel(atom_offset(srav_path, aligned_done)), /* b → aligned_done (LZCR >= 24 path) */
BdSlot_ shift_lleft_var(r.t3.v_sqr_aligned, r.t3.v_sqr_aligned, r.t5.btarget), BdSlot_ shift_lleft_var(r.v_sqr_aligned, r.v_sqr_aligned, r.btarget),
atom_label(srav_path) atom_label(srav_path)
li_s( r.t5.shift_count, 24), li_s( r.shift_count, 24),
sub_s(r.t5.shift_count, r.t5.shift_count, r.shift), sub_s(r.shift_count, r.shift_count, r.lzcr),
shift_aright_var(r.t3.v_sqr_aligned, r.t3.v_sqr_aligned, r.t5.shift_count), shift_aright_var(r.v_sqr_aligned, r.v_sqr_aligned, r.shift_count),
atom_label(aligned_done) atom_label(aligned_done)
// Save the shift count to r_shift before the next 5 instructions overwrite r_norm (the sqrtbl lookup loads 1/|v| into r_norm, which becomes IR0 in stage 4). or_u(r.half_shift, r.half_shift_tmp, 0),
or_u(r.shift, r.norm, 0), /* r_shift ← shift count (preserved through stage 4) */ add_si( r.v_sqr_aligned, r.v_sqr_aligned, -64),
add_si( r.t3.v_sqr_aligned, r.t3.v_sqr_aligned, -64), /* r_t3.v_sqr_aligned holds |v|² aligned (top bit at bit 7). */ shift_lleft(r.v_sqr_aligned, r.v_sqr_aligned, 1),
shift_lleft(r.t3.v_sqr_aligned, r.t3.v_sqr_aligned, 1), mac_load_word_imm(r.sqrtbl_index, & gte_normalize_sqr_tbl), add_u_self(r.sqrtbl_index, r.v_sqr_aligned),
mac_load_word_imm(r.t5.sqrtbl_index, & gte_normalize_sqr_tbl), add_u_self(r.t5.sqrtbl_index, r.t3.v_sqr_aligned), load_half(r.inv_len, r.sqrtbl_index, 0),
load_half(r.norm, r.t5.sqrtbl_index, 0), /* r_norm = sqrtbl[aligned-64] = 1/|v| (IR0 in stage 4) */
LdSlot_ nop, LdSlot_ nop,
/* r.src_z holds src.z from the initial load (r.src_z is a dedicated slot, mac_gte_general_purpose_interopolation(r.inv_len,
* never clobbered between the load at body start and the stage-4 GPF use below). r.src.x, r.src.y, r.src.z,
* Stage 4: GPF + srav finalize (r_shift = shift count, r_norm = 1/|v|). */ r.res.x, r.res.y, r.res.z,
mac_gte_general_purpose_interopolation( GteDelay_ load_word(R_AtomJmp, R_TapePtr, 0), LdSlot_ // ac_yield: word 1
r.norm, GteDelay_ add_ui_self( R_TapePtr, S_(MipsCode)) // ac_yield: word 2
r.src_x, /* IR1 = src.x (preserved in r_tmp — r_mac2_scratch was clobbered to MAC2 in stage 1.5) */
r.recip_est,
r.src_z, /* IR3 = src.z (reloaded) */
r.t4.mac2, r.recip_est, r.src_z,
GteDelay_ nop,
GteDelay_ nop
), ),
/* sra by r_shift = (31-LZCR)/2 (saved before sqrtbl lookup) */ mac_shift_aright_var_v3s4_self(r.res, r.half_shift),
mac_shift_aright_var_v3_self(r.t4.mac2, r.recip_est, r.src_z, r.shift), mac_store_v3s4(r.res, r.dst_ptr, 0),
/* Store result.x/y/z to r_dst_ptr (caller-determined dst address). */
mac_store_word_v3(r.t4.mac2, r.recip_est, r.src_z, r.dst_ptr, 0),
mac_yield() jump_reg(R_AtomJmp), BdSlot_ nop // ac_yield: word 3-4
// mac_yield()
}) })
/* ─── GTE OP cross product (a × b → out) ─── /* ─── GTE OP cross product (a × b → out) ───
+1 -1
View File
@@ -263,7 +263,7 @@ FI_ void tb_scope_run_end(TapeBuilder* tb) { tb_emit(tb,tape_exit); tape_run(tb_
// The 'Yield' sequence for Tape Atoms (mac_yield). // The 'Yield' sequence for Tape Atoms (mac_yield).
atom_dbg_skip MipsAtomComp_(ac_yield) { atom_dbg_skip MipsAtomComp_(ac_yield) {
load_word(R_AtomJmp, R_TapePtr, 0), load_word(R_AtomJmp, R_TapePtr, 0), LdSlot_
add_ui_self( R_TapePtr, S_(MipsCode)), add_ui_self( R_TapePtr, S_(MipsCode)),
jump_reg( R_AtomJmp), BdSlot_ nop, jump_reg( R_AtomJmp), BdSlot_ nop,
}; };
+4 -8
View File
@@ -151,14 +151,10 @@ internal void compile_init_atoms(void) {
smem.normalize_v3s4 = build_normalize_v3s4(& ab, smem.normalize_v3s4 = build_normalize_v3s4(& ab,
RegUse_(build_normalize_v3s4) { RegUse_(build_normalize_v3s4) {
.src_ptr = ralloc(), .res = ralloc_v3(),
.dst_ptr = ralloc(), .t0 = ralloc(),
.recip_est = ralloc(), .t1 = ralloc(),
.norm = ralloc(), .t2 = ralloc(),
.shift = ralloc(),
.src_x = ralloc(),
.src_z = ralloc(),
// .shift_count = ralloc(), /* dedicated slot for stage-3 → stage-4 shift count */
.t3 = ralloc(), .t3 = ralloc(),
.t4 = ralloc(), .t4 = ralloc(),
.t5 = ralloc(), .t5 = ralloc(),
+8 -1
View File
@@ -550,7 +550,14 @@ local function _project_emission_inner(root_body_entry, ctx_table)
local tok_line = line_of(body_off + bt.rel) or 0 local tok_line = line_of(body_off + bt.rel) or 0
if M.DELAY_MARKERS[ident] then if M.DELAY_MARKERS[ident] then
emit_marker("delay", ident, nil, tok_line) emit_marker("delay", ident, nil, tok_line)
local rest = M.trim(tok:sub(after or (#tok + 1))) local rest = tok:sub(after or (#tok + 1))
while true do
rest = M.trim(rest)
if rest:sub(1, 2) ~= "/*" then break end
local close = rest:find("*/", 3, true)
if not close then rest = ""; break end
rest = rest:sub(close + 2)
end
if rest ~= "" then if rest ~= "" then
process_token({ tok = rest, rel = bt.rel }) process_token({ tok = rest, rel = bt.rel })
end end
+31 -1
View File
@@ -329,6 +329,24 @@ local function strip_mac_prefix(ident)
return ident return ident
end end
--- Strip a leading delay marker (`LdSlot_` / `BdSlot_` / `GteDelay_` / `DmaSlot_`)
--- plus following whitespace and block comments. Returns the remainder, or ""
--- when the token is only the marker.
--- `BdSlot_ nop` becomes `nop`. Bare `LdSlot_` becomes "".
--- @param tok string
--- @return string
local function strip_leading_delay_marker(tok)
local ident = duffle.read_ident(tok, 1)
if not ident or not duffle.DELAY_MARKERS[ident] then return tok end
local rest = tok:sub(#ident + 1):match("^%s*(.*)$") or ""
while rest:sub(1, 2) == "/*" do
local close = rest:find("*/", 3, true)
if not close then return "" end
rest = rest:sub(close + 2):match("^%s*(.*)$") or ""
end
return rest
end
--- (internal) Recursive word-count lookup. `cache` is the memoization table shared across all components --- (internal) Recursive word-count lookup. `cache` is the memoization table shared across all components
--- in a single source's `count_all_components` pass; the in-progress -1 sentinel detects cycles (A -> B -> A). --- in a single source's `count_all_components` pass; the in-progress -1 sentinel detects cycles (A -> B -> A).
--- @param name string -- the component name (without `mac_`) --- @param name string -- the component name (without `mac_`)
@@ -347,7 +365,18 @@ local function word_count_rec(name, comp_by_name, wc, cache)
for _, t in ipairs(tokens) do for _, t in ipairs(tokens) do
local trimmed = t.tok local trimmed = t.tok
if trimmed ~= "" then if trimmed ~= "" then
local lookup = strip_mac_prefix(duffle.read_ident(trimmed, 1)) local work = trimmed
while true do
local marker = duffle.read_ident(work, 1)
if marker and duffle.DELAY_MARKERS[marker] then
work = strip_leading_delay_marker(work)
if work == "" then break end
else
break
end
end
if work ~= "" then
local lookup = strip_mac_prefix(duffle.read_ident(work, 1))
if lookup == "atom_label" or lookup == "atom_offset" then if lookup == "atom_label" or lookup == "atom_offset" then
-- Pure metaprogram anchors; emit zero words. -- Pure metaprogram anchors; emit zero words.
elseif lookup and comp_by_name[lookup] then elseif lookup and comp_by_name[lookup] then
@@ -362,6 +391,7 @@ local function word_count_rec(name, comp_by_name, wc, cache)
end end
end end
end end
end
else else
-- Not a known component: assume 1 word (regular instruction). -- Not a known component: assume 1 word (regular instruction).
n = 1 n = 1
+1 -1
View File
@@ -7,7 +7,7 @@
--- Public boundary: --- Public boundary:
--- * `M.run(ctx)` is the only entry point. --- * `M.run(ctx)` is the only entry point.
--- * The pass returns `{outputs = {}, errors = ..., warnings = ...}`. --- * The pass returns `{outputs = {}, errors = ..., warnings = ...}`.
--- Pass kind = `validation` → `PASS_KIND_STOP_ON_ERROR.validation` preserves the existing build-stopping policy. --- Pass kind = `validation`. Findings record on the result; the orchestrator does not exit non-zero.
--- ---
--- Source-order discipline: --- Source-order discipline:
--- * `corpus.source_order` sets the source-record order. --- * `corpus.source_order` sets the source-record order.
+20 -13
View File
@@ -129,33 +129,39 @@ end
--- This preserves backward compatibility for any top-level marker that may exist outside a control-transfer instruction. --- This preserves backward compatibility for any top-level marker that may exist outside a control-transfer instruction.
--- @param labels table<string, integer> --- @param labels table<string, integer>
--- @param branches table[] --- @param branches table[]
--- @param errors table[]
--- @return BranchOffset[] --- @return BranchOffset[]
local function compute_offsets(labels, branches) local function compute_offsets(labels, branches, errors)
local results = {} local results = {}
for _, br in ipairs(branches) do for _, br in ipairs(branches) do
local target = labels[br.target] local target = labels[br.target]
if not target then if not target then
error("Branch target '" .. br.target .. "' has no atom_label (at word " .. br.branch_word .. ")") errors[#errors + 1] = {
end line = br.line or 0,
msg = "Branch target '" .. br.target .. "' has no atom_label (at word " .. br.branch_word .. ")",
}
else
local consuming = br.consuming_encoder local consuming = br.consuming_encoder
local offset
if consuming == "jump_reg" or consuming == "call_reg" or consuming == "jump_link" then if consuming == "jump_reg" or consuming == "call_reg" or consuming == "jump_link" then
-- Register-form jumps have no offset field. `atom_offset` cannot be used here. errors[#errors + 1] = {
error("atom_offset cannot be used with " .. consuming line = br.line or 0,
.. " (register-form jumps have no offset field); at word " .. br.branch_word) msg = "atom_offset cannot be used with " .. consuming
end .. " (register-form jumps have no offset field); at word " .. br.branch_word,
}
else
-- All other consuming instructions (including `branch_*`, `jump`, `call_addr`, and nil for top-level markers) use the same relative offset value. -- All other consuming instructions (including `branch_*`, `jump`, `call_addr`, and nil for top-level markers) use the same relative offset value.
-- The MIPS encoding differs per opcode but the duffle `enc_i` macro handles the truncation to the immediate-field width. -- The MIPS encoding differs per opcode but the duffle `enc_i` macro handles the truncation to the immediate-field width.
offset = target - br.branch_word - 1
results[#results + 1] = { results[#results + 1] = {
target = br.target, target = br.target,
tag = br.tag, tag = br.tag,
branch_word = br.branch_word, branch_word = br.branch_word,
offset = offset, offset = target - br.branch_word - 1,
consuming_encoder = br.consuming_encoder, consuming_encoder = br.consuming_encoder,
consuming_arg_pos = br.consuming_arg_pos, consuming_arg_pos = br.consuming_arg_pos,
} }
end end
end
end
return results return results
end end
@@ -237,8 +243,9 @@ local M = {}
--- @param ctx PassCtx --- @param ctx PassCtx
--- @param dir string -- the absolute source directory --- @param dir string -- the absolute source directory
--- @param sources SourceFile[] -- sources in this directory --- @param sources SourceFile[] -- sources in this directory
--- @param errors table[]
--- @return string|nil -- the offsets_h path --- @return string|nil -- the offsets_h path
local function process_directory(ctx, dir, sources) local function process_directory(ctx, dir, sources, errors)
local atoms_data = {} local atoms_data = {}
local function append_atom(atom) local function append_atom(atom)
@@ -248,7 +255,7 @@ local function process_directory(ctx, dir, sources)
atoms_data[#atoms_data + 1] = { atoms_data[#atoms_data + 1] = {
name = atom.raw_name or atom.name, name = atom.raw_name or atom.name,
total_words = #(paths.word_events or {}), total_words = #(paths.word_events or {}),
offsets = compute_offsets(labels, branches), offsets = compute_offsets(labels, branches, errors),
} }
end end
@@ -286,7 +293,7 @@ function M.run(ctx)
-- Per-directory aggregation: every source in the same directory contributes to one `gen/offsets.h`. -- Per-directory aggregation: every source in the same directory contributes to one `gen/offsets.h`.
local sources_by_dir = corpus.sources_by_dir or duffle.group_sources_by_dir(corpus.source_order) local sources_by_dir = corpus.sources_by_dir or duffle.group_sources_by_dir(corpus.source_order)
for dir, sources in pairs(sources_by_dir) do for dir, sources in pairs(sources_by_dir) do
local out_path = process_directory(ctx, dir, sources) local out_path = process_directory(ctx, dir, sources, errors)
if out_path then if out_path then
outputs[#outputs + 1] = { offsets_h = out_path } outputs[#outputs + 1] = { offsets_h = out_path }
end end
+197 -13
View File
@@ -1674,9 +1674,20 @@ parse_reg_use_schema_body = function(body, type_registry, opts)
errors[#errors + 1] = { kind = "reguse_malformed" } errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors return nil, errors
end end
local members = {} local views = {}
local union_readonly = nil local union_readonly = nil
local inner_pos = 1 local inner_pos = 1
local function note_readonly(flag)
if union_readonly == nil then
union_readonly = flag
elseif union_readonly ~= flag then
errors[#errors + 1] = { kind = "reguse_mixed_const" }
return false
end
return true
end
while inner_pos <= #inner do while inner_pos <= #inner do
inner_pos = duffle.skip_ws_and_cmt(inner, inner_pos) inner_pos = duffle.skip_ws_and_cmt(inner, inner_pos)
if inner_pos > #inner then break end if inner_pos > #inner then break end
@@ -1689,10 +1700,138 @@ parse_reg_use_schema_body = function(body, type_registry, opts)
errors[#errors + 1] = { kind = "reguse_const_reg_spelling" } errors[#errors + 1] = { kind = "reguse_const_reg_spelling" }
return nil, errors return nil, errors
end end
if m_type ~= "Reg" then
if m_type == "Reg_" then
local after_ty = duffle.skip_ws_and_cmt(inner, m_type_end)
if inner:sub(after_ty, after_ty) ~= "(" then
errors[#errors + 1] = { kind = "reguse_malformed" } errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors return nil, errors
end end
local type_inner, after_paren = duffle.read_parens(inner, after_ty)
if not type_inner then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
local typed_fields = fields_for_reg_type(duffle.trim(type_inner), type_registry)
after_paren = duffle.skip_ws_and_cmt(inner, after_paren)
local inst_names, new_inner = parse_reg_names(inner, after_paren)
if not inst_names or #inst_names == 0 then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
if not note_readonly(false) then return nil, errors end
if not typed_fields then
if require_types then
errors[#errors + 1] = { kind = "reguse_unknown_reg_type", type_name = duffle.trim(type_inner) }
return nil, errors
end
pending = true
else
for _, inst in ipairs(inst_names) do
local names = {}
for _, field in ipairs(typed_fields) do
names[#names + 1] = inst .. "." .. field
end
views[#views + 1] = { names = names, lanes = true }
end
end
inner_pos = new_inner
elseif m_type == "struct" then
local after_struct = duffle.skip_ws_and_cmt(inner, m_type_end)
if inner:sub(after_struct, after_struct) ~= "{" then
local _, tag_end = duffle.read_ident(inner, after_struct)
if not tag_end then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
after_struct = duffle.skip_ws_and_cmt(inner, tag_end)
end
if inner:sub(after_struct, after_struct) ~= "{" then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
local struct_inner, after_struct_braces = duffle.read_braces(inner, after_struct)
if not struct_inner then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
local lane_names = {}
local s_pos = 1
while s_pos <= #struct_inner do
s_pos = duffle.skip_ws_and_cmt(struct_inner, s_pos)
if s_pos > #struct_inner then break end
local s_ty, s_ty_end = duffle.read_ident(struct_inner, s_pos)
if not s_ty then
s_pos = s_pos + 1
goto continue_struct
end
if s_ty == "Reg_" then
local after_ty = duffle.skip_ws_and_cmt(struct_inner, s_ty_end)
if struct_inner:sub(after_ty, after_ty) ~= "(" then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
local type_inner, after_paren = duffle.read_parens(struct_inner, after_ty)
if not type_inner then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
local typed_fields = fields_for_reg_type(duffle.trim(type_inner), type_registry)
after_paren = duffle.skip_ws_and_cmt(struct_inner, after_paren)
local inst_names, new_s = parse_reg_names(struct_inner, after_paren)
if not inst_names or #inst_names == 0 then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
if not note_readonly(false) then return nil, errors end
if not typed_fields then
if require_types then
errors[#errors + 1] = { kind = "reguse_unknown_reg_type", type_name = duffle.trim(type_inner) }
return nil, errors
end
pending = true
else
for _, inst in ipairs(inst_names) do
for _, field in ipairs(typed_fields) do
lane_names[#lane_names + 1] = inst .. "." .. field
end
end
end
s_pos = new_s
elseif s_ty == "Reg" then
local s_after = duffle.skip_ws_and_cmt(struct_inner, s_ty_end)
local s_readonly = false
local maybe_const, maybe_end = duffle.read_ident(struct_inner, s_after)
if maybe_const == "const" then
s_readonly = true
s_after = duffle.skip_ws_and_cmt(struct_inner, maybe_end)
end
if not note_readonly(s_readonly) then return nil, errors end
local names, new_s = parse_reg_names(struct_inner, s_after)
if not names or #names == 0 then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
for _, n in ipairs(names) do lane_names[#lane_names + 1] = n end
s_pos = new_s
else
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
::continue_struct::
end
if #lane_names == 0 and not pending then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
if #lane_names > 0 then
views[#views + 1] = { names = lane_names, lanes = true }
end
inner_pos = duffle.skip_ws_and_cmt(inner, after_struct_braces)
if inner:sub(inner_pos, inner_pos) == ";" then inner_pos = inner_pos + 1 end
elseif m_type == "Reg" then
local m_after = duffle.skip_ws_and_cmt(inner, m_type_end) local m_after = duffle.skip_ws_and_cmt(inner, m_type_end)
local m_readonly = false local m_readonly = false
local maybe_const, maybe_end = duffle.read_ident(inner, m_after) local maybe_const, maybe_end = duffle.read_ident(inner, m_after)
@@ -1700,27 +1839,68 @@ parse_reg_use_schema_body = function(body, type_registry, opts)
m_readonly = true m_readonly = true
m_after = duffle.skip_ws_and_cmt(inner, maybe_end) m_after = duffle.skip_ws_and_cmt(inner, maybe_end)
end end
if union_readonly == nil then if not note_readonly(m_readonly) then return nil, errors end
union_readonly = m_readonly
elseif union_readonly ~= m_readonly then
errors[#errors + 1] = { kind = "reguse_mixed_const" }
return nil, errors
end
local names, new_inner = parse_reg_names(inner, m_after) local names, new_inner = parse_reg_names(inner, m_after)
if not names or #names == 0 then if not names or #names == 0 then
errors[#errors + 1] = { kind = "reguse_malformed" } errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors return nil, errors
end end
for _, n in ipairs(names) do members[#members + 1] = n end views[#views + 1] = { names = names, lanes = false }
inner_pos = new_inner inner_pos = new_inner
::continue_inner:: else
end
if #members == 0 then
errors[#errors + 1] = { kind = "reguse_malformed" } errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors return nil, errors
end end
::continue_inner::
end
local after_close = duffle.skip_ws_and_cmt(body, after_braces) local after_close = duffle.skip_ws_and_cmt(body, after_braces)
local inst_name, inst_end = duffle.read_ident(body, after_close) local inst_name, inst_end = duffle.read_ident(body, after_close)
if #views == 0 then
if not pending then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
else
local has_lanes = false
for _, v in ipairs(views) do
if v.lanes then has_lanes = true end
end
if has_lanes then
local width = nil
for _, v in ipairs(views) do
if not v.lanes then
errors[#errors + 1] = { kind = "reguse_malformed" }
return nil, errors
end
if width == nil then
width = #v.names
elseif #v.names ~= width then
errors[#errors + 1] = { kind = "reguse_union_width" }
return nil, errors
end
end
if width then
for i = 1, width do
local slot_name = views[1].names[i]
if inst_name then slot_name = inst_name .. "." .. slot_name end
local aliases = {}
for _, v in ipairs(views) do
local n = v.names[i]
if inst_name then n = inst_name .. "." .. n end
if not add_alias(n, slot_name) then return nil, errors end
aliases[#aliases + 1] = n
end
if not add_slot(slot_name, aliases, union_readonly) then return nil, errors end
end
end
else
local members = {}
for _, v in ipairs(views) do
for _, n in ipairs(v.names) do members[#members + 1] = n end
end
local aliases = {} local aliases = {}
local slot_name local slot_name
if inst_name then if inst_name then
@@ -1730,7 +1910,6 @@ parse_reg_use_schema_body = function(body, type_registry, opts)
if not add_alias(path, slot_name) then return nil, errors end if not add_alias(path, slot_name) then return nil, errors end
aliases[#aliases + 1] = path aliases[#aliases + 1] = path
end end
after_close = inst_end
else else
slot_name = members[1] slot_name = members[1]
for _, m in ipairs(members) do for _, m in ipairs(members) do
@@ -1739,6 +1918,10 @@ parse_reg_use_schema_body = function(body, type_registry, opts)
end end
end end
if not add_slot(slot_name, aliases, union_readonly) then return nil, errors end if not add_slot(slot_name, aliases, union_readonly) then return nil, errors end
end
end
if inst_name then after_close = inst_end end
after_close = duffle.skip_ws_and_cmt(body, after_close) after_close = duffle.skip_ws_and_cmt(body, after_close)
if body:sub(after_close, after_close) == ";" then after_close = after_close + 1 end if body:sub(after_close, after_close) == ";" then after_close = after_close + 1 end
pos = after_close pos = after_close
@@ -2801,6 +2984,7 @@ local SCHEMA_BODY_ERROR = {
reguse_duplicate_slot = true, reguse_duplicate_slot = true,
reguse_const_reg_spelling = true, reguse_const_reg_spelling = true,
reguse_mixed_const = true, reguse_mixed_const = true,
reguse_union_width = true,
} }
-- Re-parse every RegUse_* body against the merged type_name_registry. -- Re-parse every RegUse_* body against the merged type_name_registry.
+12 -4
View File
@@ -879,6 +879,14 @@ local function analyze_hardware_relations(atom)
end end
if is_match then if is_match then
local gap = ev_word - prod.word - 1 local gap = ev_word - prod.word - 1
local required = prod.required
-- A following COP2 command waits at the transfer boundary.
-- Software nops are not required for that consumer class.
if is_gte_command(ev)
and (semantic == "MTC2" or semantic == "CTC2")
and relation.id ~= "mtc2_irgb_visibility" then
required = 0
end
local unknown_visibility = relation.visibility and relation.visibility.kind == "unknown_consumer" local unknown_visibility = relation.visibility and relation.visibility.kind == "unknown_consumer"
if unknown_visibility then if unknown_visibility then
hazards[#hazards + 1] = { hazards[#hazards + 1] = {
@@ -906,7 +914,7 @@ local function analyze_hardware_relations(atom)
, (relation.evidence and relation.evidence.confidence or "unknown") , (relation.evidence and relation.evidence.confidence or "unknown")
), ),
} }
elseif prod.required ~= nil and gap < prod.required then elseif required ~= nil and gap < required then
local payload = { local payload = {
check = "transfer_hazards", check = "transfer_hazards",
kind = prod.violation_kind or "error", kind = prod.violation_kind or "error",
@@ -923,7 +931,7 @@ local function analyze_hardware_relations(atom)
consumer_word = ev_word, consumer_word = ev_word,
consumer_token = ev_ident, consumer_token = ev_ident,
gap = gap, gap = gap,
required = prod.required, required = required,
evidence_confidence = relation.evidence and relation.evidence.confidence or "unknown", evidence_confidence = relation.evidence and relation.evidence.confidence or "unknown",
evidence_source = relation.evidence and relation.evidence.source or "", evidence_source = relation.evidence and relation.evidence.source or "",
msg = string.format("%s at line %d: %s relation %s (producer %s at word %d, %s:%d) violated: consumer at word %d (gap=%d, required=%d) [%s]" msg = string.format("%s at line %d: %s relation %s (producer %s at word %d, %s:%d) violated: consumer at word %d (gap=%d, required=%d) [%s]"
@@ -941,7 +949,7 @@ local function analyze_hardware_relations(atom)
hazards[#hazards + 1] = payload hazards[#hazards + 1] = payload
end end
local satisfied = nil local satisfied = nil
if not unknown_visibility then satisfied = gap >= prod.required end if not unknown_visibility then satisfied = gap >= required end
-- Record the relation touch on `paths.relations` even when the gap is satisfied. -- Record the relation touch on `paths.relations` even when the gap is satisfied.
-- Unknown relations are informational, not numeric pass/fail measurements. -- Unknown relations are informational, not numeric pass/fail measurements.
relations[#relations + 1] = { relations[#relations + 1] = {
@@ -951,7 +959,7 @@ local function analyze_hardware_relations(atom)
consumer_word = ev_word, consumer_word = ev_word,
destination = prod.destination, destination = prod.destination,
gap = gap, gap = gap,
required = prod.required, required = required,
satisfied = satisfied, satisfied = satisfied,
} }
table.remove(pending, pending_idx) table.remove(pending, pending_idx)
+9 -13
View File
@@ -142,8 +142,7 @@ local PASSES = {
}, },
["static-analysis"] = { ["static-analysis"] = {
module = "passes.static_analysis", module = "passes.static_analysis",
-- "diagnostic" — every `error`/`warning` finding is written to the report file; -- "diagnostic" — every `error`/`warning` finding is written to the report file.
-- The orchestrator does NOT exit non-zero on these findings (see PASS_KIND_STOP_ON_ERROR).
-- Report severity is independent from process exit policy. -- Report severity is independent from process exit policy.
kind = "diagnostic", kind = "diagnostic",
deps = {"scan-source", "word-counts", "components", "emission-model"}, deps = {"scan-source", "word-counts", "components", "emission-model"},
@@ -206,16 +205,13 @@ local function request_roots_for_group(args, group_name)
end end
end end
-- Pass-kind taxonomy: Which kinds stop the build on errors? -- Pass-kind taxonomy: findings always print. No pass kind stops the build.
--
-- Report severity is independent from process exit policy. -- Report severity is independent from process exit policy.
-- A "diagnostic" pass still writes every `error`/`warning` finding into its report file, -- Adding a new pass kind requires listing it here explicitly; an unknown kind must not silently fall back to "true".
-- but `report_validation_errors` returns early for non-stopping kinds, so nothing is printed to stderr and the orchestrator does not exit non-zero.
-- Adding a new pass kind requires listing it here explicitly; An unknown kind must not silently fall back to "true".
local PASS_KIND_STOP_ON_ERROR = { local PASS_KIND_STOP_ON_ERROR = {
["shared"] = false, ["shared"] = false,
["header-output"] = true, ["header-output"] = false,
["validation"] = true, ["validation"] = false,
["diagnostic"] = false, ["diagnostic"] = false,
["report"] = false, ["report"] = false,
} }
@@ -693,19 +689,19 @@ end
-- Main Orchestrator -- Main Orchestrator
-- ════════════════════════════════════════════════════════════════════════════ -- ════════════════════════════════════════════════════════════════════════════
--- (internal) If the pass's kind is in PASS_KIND_STOP_ON_ERROR and it reported errors, write each error to stderr. --- (internal) Write every pass error to stderr.
--- Returns true if any validation errors were reported. --- Returns true only when the pass kind still stops the build.
--- @param pass_name string --- @param pass_name string
--- @param pass PassDescriptor --- @param pass PassDescriptor
--- @param result PassResult --- @param result PassResult
--- @return boolean --- @return boolean
local function report_validation_errors(pass_name, pass, result) local function report_validation_errors(pass_name, pass, result)
local has_errors = result.errors and #result.errors > 0 local has_errors = result.errors and #result.errors > 0
if not (has_errors and PASS_KIND_STOP_ON_ERROR[pass.kind]) then return false end if not has_errors then return false end
for _, e in ipairs(result.errors) do for _, e in ipairs(result.errors) do
io.stderr:write(string.format("[%s] line %d: %s\n", pass_name, e.line or 0, e.msg or "")) io.stderr:write(string.format("[%s] line %d: %s\n", pass_name, e.line or 0, e.msg or ""))
end end
return true return PASS_KIND_STOP_ON_ERROR[pass.kind] == true
end end
--- (internal) Run each pass in `order` in topological sequence. --- (internal) Run each pass in `order` in topological sequence.