wasm32: add node-only TB wasm dump tooling for static analysis (E9)

tcg/tcg.c

@@ -6675,7 +6675,19 @@ int tcg_gen_code(TCGContext *s, TranslationBlock *tb, uint64_t pc_start)
     }
     *(uint32_t *)wasm_blob_ptr_base = s->code_ptr - wasm_blob_ptr_base - 4;
 
-    // record importing helper functions
+    {
+        static __thread int dumped_store, dumped_load;
+        uint32_t mod_len = *(uint32_t *)wasm_blob_ptr_base;
+        const uint8_t *mod = wasm_blob_ptr_base + 4;
+        if (wasm_tb_had_store && !dumped_store) {
+            dumped_store = 1;
+            wasm_dump_module(mod, mod_len, "store");
+        } else if (wasm_tb_had_load && !wasm_tb_had_store && !dumped_load) {
+            dumped_load = 1;
+            wasm_dump_module(mod, mod_len, "load");
+        }
+    }
+
     size_base = (uint32_t*)s->code_ptr;
     s->code_ptr += 4;
     memcpy(s->code_ptr, target_helper_funcs, num_helper_funcs * 4);

tcg/wasm32.c

@@ -39,6 +39,23 @@ EMSCRIPTEN_KEEPALIVE int wasm_instantiated_tb_count(void)
     return qatomic_read(&instantiated_wasm_count);
 }
 
+EM_JS(void, wasm_dump_module_js, (int ptr, int len, int name), {
+    try {
+        const fs = require('fs');
+        const bytes = Uint8Array.from(HEAP8.subarray(ptr, ptr + len));
+        const fname = '/tmp/tbdump-' + UTF8ToString(name) + '.wasm';
+        fs.writeFileSync(fname, bytes);
+        out('WASM_DUMP wrote ' + fname + ' (' + len + ' bytes)');
+    } catch (e) {
+        /* Browser has no require('fs'); dump is a node-only analysis aid. */
+    }
+});
+
+void wasm_dump_module(const uint8_t *p, int len, const char *name)
+{
+    wasm_dump_module_js((int)p, len, (int)name);
+}
+
 /* Disassemble TCI bytecode. */
 int print_insn_tci(bfd_vma addr, disassemble_info *info)
 {

tcg/wasm32.h

@@ -41,6 +41,8 @@ void flush_tb_instances(void);
 
 void init_wasm32();
 
+void wasm_dump_module(const uint8_t *p, int len, const char *name);
+
 extern __thread bool wasm_tci_only_tb;
 
 #define INSTANTIATE_NUM 1500

tcg/wasm32/tcg-target.c.inc

@@ -3362,13 +3362,19 @@ static uint8_t tcg_tci_out_qemu_ldst(TCGContext *s, TCGOpcode opc, const TCGArg
     insn = deposit32(insn, 0, 8, opc);
     tcg_tci_out32(s, insn);
 }
+/* Dump tags: instrumentation-only, set during codegen so tcg.c can dump one store-TB / load-TB
+ * module for static analysis. No effect on emitted code. */
+__thread bool wasm_tb_had_store;
+__thread bool wasm_tb_had_load;
 static void tcg_out_qemu_ld(TCGContext *s, TCGOpcode opc, const TCGArg *args, bool is_64)
 {
+    wasm_tb_had_load = true;
     tcg_tci_out_qemu_ldst(s, opc, args);
     tcg_wasm_out_qemu_ld(s, args, is_64);
 }
 static void tcg_out_qemu_st(TCGContext *s, TCGOpcode opc, const TCGArg *args, bool is_64)
 {
+    wasm_tb_had_store = true;
     tcg_tci_out_qemu_ldst(s, opc, args);
     tcg_wasm_out_qemu_st(s, args, is_64);
     /* ponytail: Store-Containing TBs Run on TCI - Upstream qemu-wasm sets this for every guest store.