Limit number of function parameters in MergeSimilarFunctions (#7341)

MergeSimilarFunctions can merge functions whose bodies differ only the
constants they use. These different constants become parameters to the
merged function. In extreme cases, the optimization could produce an
arbitrary number of new parameters, exceeding implementation limits.

Cap the total number of parameters to a merged function at 255. The Web
limitation on the number of parameters is 1000, but other
implementations, including those on the JVM, cannot handle more than
255. It would be possible to make this limit configurable in the future.

Author: tlively

src/passes/MergeSimilarFunctions.cpp

@@ -84,6 +84,7 @@
 #include "pass.h"
 #include "support/hash.h"
 #include "support/utilities.h"
+#include "wasm-limits.h"
 #include "wasm.h"
 #include <algorithm>
 #include <cassert>
@@ -490,6 +491,15 @@ void EquivalentClass::merge(Module* module,
 // than the reduced size.
 bool EquivalentClass::hasMergeBenefit(Module* module,
                                       const std::vector<ParamInfo>& params) {
+  if (params.size() + primaryFunction->getNumParams() >
+      MaxSyntheticFunctionParams) {
+    // It requires too many parameters to merge this equivalence class. In
+    // principle, we could try splitting the class into smaller classes of
+    // functions that share more constants with each other, but that could
+    // be expensive. TODO: investigate splitting the class.
+    return false;
+  }
+
   size_t funcCount = functions.size();
   Index exprSize = Measurer::measure(primaryFunction->body);
   size_t thunkCount = funcCount;

src/wasm-limits.h

@@ -31,6 +31,11 @@ enum WebLimitations : uint32_t {
   MaxFunctionParams = 1000
 };
 
+// Web limits allow up to 1000 function parameters, but other implementations,
+// such as those on the JVM, only allow up to 255. Do not allow optimizations to
+// introduce functions with more than this many parameters.
+static constexpr int MaxSyntheticFunctionParams = 255;
+
 } // namespace wasm
 
 #endif // wasm_wasm_limits_h