synth: hash post-read_design_sources state to isolate slang non-determinism

flow/scripts/synth_canonicalize.tcl

@@ -1,5 +1,11 @@
 source $::env(SCRIPTS_DIR)/synth_preamble.tcl
 read_design_sources
+# Fingerprint the design state right after the HDL frontend
+# (slang / builtin Verilog) returns, before any other pass runs.
+# An unstable hash here means the frontend itself is non-idempotent
+# -- distinct from drift introduced later by hierarchy / opt_clean
+# / etc.  Surfaced as a literal warning-level rule via genRuleFile.py.
+write_state_hash synth__post_read_sources__hash
 
 hierarchy -check -top $::env(DESIGN_NAME)
 

flow/scripts/synth_preamble.tcl

@@ -3,6 +3,39 @@ yosys -import
 source $::env(SCRIPTS_DIR)/util.tcl
 erase_non_stage_variables synth
 
+# Fingerprint the current yosys design state to a `<metric>: <sha>`
+# line in the surrounding yosys log so `genMetrics.py` can pick it up.
+#
+# `setattr -unset src` strips file:line attribute lines from the
+# RTLIL before hashing so the hash is path-independent (bazel sandbox
+# paths differ from the classic-make build's relative paths; without
+# stripping, hashes always differ trivially).
+#
+# The strip is wrapped in `design -push` / `design -pop` so it does
+# not leak into the rest of the synth run -- src attributes are
+# preserved for back-annotation / debugging downstream.  This matters
+# specifically for builds without `SYNTH_REPEATABLE_BUILD=1`, where
+# synth_canonicalize.tcl deliberately leaves src attrs alone after
+# the canonical-RTLIL write.
+proc write_state_hash { metric } {
+  design -push
+  # `*/*` strips src attrs across objects in *all* modules.  Bare `*`
+  # only targets the current module's objects, which is fine after
+  # `hierarchy -check -top` collapses things to one design, but
+  # `write_state_hash` is called before `hierarchy` here, so the
+  # post-frontend state may have many separate modules whose src
+  # attrs would otherwise survive into the hashed RTLIL and break
+  # path-independence.
+  setattr -unset src */*
+  setattr -mod -unset src *
+  set tmp $::env(OBJECTS_DIR)/.${metric}.tmp.rtlil
+  write_rtlil $tmp
+  design -pop
+  set sha [lindex [split [exec sha1sum $tmp]] 0]
+  file delete $tmp
+  puts "${metric}: $sha"
+}
+
 # If using a cached, gate level netlist, then copy over to the results dir with
 # preserve timestamps flag set. If you don't, subsequent runs will cause the
 # floorplan step to be re-executed.

flow/util/genMetrics.py

@@ -262,10 +262,29 @@ def extract_metrics(
         rptPath + "/synth_stat.txt",
     )
 
-    # Netlist hashes: fingerprints of the canonical RTLIL (pre-ABC) and
-    # the final post-synthesis Verilog so the rules-base.json check
-    # (level=warning) flags when bazel-built vs make-built yosys
-    # disagree for the same RTL.
+    # Netlist hashes: fingerprints at three points in the yosys
+    # pipeline so the rules-base.json check (level=warning) can
+    # isolate frontend drift (yosys-slang isn't idempotent) from
+    # mid-synth drift from ABC drift.
+    #
+    #   post_read_sources = state right after `read_design_sources`
+    #                       (HDL frontend output only)
+    #   canonical_netlist = state after `opt_clean -purge`
+    #                       (= `1_1_yosys_canonicalize.rtlil`)
+    #   netlist           = `1_2_yosys.v`, post-ABC
+    #
+    # `post_read_sources` is emitted by synth_canonicalize.tcl via
+    # `write_state_hash` (synth_preamble.tcl) as a `<metric>: <sha>`
+    # line in 1_1_yosys_canonicalize.log; the other two come straight
+    # from `file_sha1` of the already-emitted RTLIL / Verilog.
+    extractTagFromFile(
+        "synth__post_read_sources__hash",
+        metrics_dict,
+        r"^\s*synth__post_read_sources__hash:\s+([0-9a-f]{40})\s*$",
+        logPath + "/1_1_yosys_canonicalize.log",
+        t=str,
+        required=False,
+    )
     metrics_dict["synth__canonical_netlist__hash"] = file_sha1(
         resultPath + "/1_1_yosys_canonicalize.rtlil"
     )

flow/util/genRuleFile.py

@@ -79,6 +79,11 @@ def gen_rule_file(
         # surfaces as a [WARN] diagnostic in checkMetadata.py without
         # failing the build, matching how rules-base.json already
         # treats warning counts.
+        "synth__post_read_sources__hash": {
+            "mode": "literal",
+            "compare": "==",
+            "level": "warning",
+        },
         "synth__canonical_netlist__hash": {
             "mode": "literal",
             "compare": "==",