Make MaxCombinedBinarySize configurable

Use case: I want to avoid generating "huge" (eg. 50KB) functions,
which can involve certain kind of text relocations in aot-compiled
binaries for my target.

Author: yamt

src/binaryen-c.cpp

@@ -5754,6 +5754,14 @@ void BinaryenSetFlexibleInlineMaxSize(BinaryenIndex size) {
   globalPassOptions.inlining.flexibleInlineMaxSize = size;
 }
 
+BinaryenIndex BinaryenGetMaxCombinedBinarySize(void) {
+  return globalPassOptions.inlining.maxCombinedBinarySize;
+}
+
+void BinaryenSetMaxCombinedBinarySize(BinaryenIndex size) {
+  globalPassOptions.inlining.maxCombinedBinarySize = size;
+}
+
 BinaryenIndex BinaryenGetOneCallerInlineMaxSize(void) {
   return globalPassOptions.inlining.oneCallerInlineMaxSize;
 }

src/binaryen-c.h

@@ -3148,6 +3148,14 @@ BINARYEN_API BinaryenIndex BinaryenGetFlexibleInlineMaxSize(void);
 // Applies to all modules, globally.
 BINARYEN_API void BinaryenSetFlexibleInlineMaxSize(BinaryenIndex size);
 
+// Gets the limit for the combined size of the code after inlining.
+// Applies to all modules, globally.
+BINARYEN_API BinaryenIndex BinaryenGetMaxCombinedBinarySize(void);
+
+// Sets the limit for the combined size of the code after inlining.
+// Applies to all modules, globally.
+BINARYEN_API void BinaryenSetMaxCombinedBinarySize(BinaryenIndex size);
+
 // Gets the function size which we inline when there is only one caller.
 // Applies to all modules, globally.
 BINARYEN_API BinaryenIndex BinaryenGetOneCallerInlineMaxSize(void);

src/pass.h

@@ -89,6 +89,17 @@ struct InliningOptions {
   // This is checked after alwaysInlineMaxSize and oneCallerInlineMaxSize, but
   // the order normally won't matter.
   Index flexibleInlineMaxSize = 20;
+  // The limit for the combined size of the code after inlining.
+  // We have an absolute limit in order to avoid extremely-large sizes after
+  // inlining, as they may hit limits in VMs and/or slow down startup
+  // (measurements there indicate something like ~1 second to optimize a 100K
+  // function). See e.g.
+  // https://github.com/WebAssembly/binaryen/pull/3730#issuecomment-867939138
+  // https://github.com/emscripten-core/emscripten/issues/13899#issuecomment-825073344
+  // The limit is arbitrary, but based on the links above. It is a very high
+  // value that should appear very rarely in practice (for example, it does
+  // not occur on the Emscripten benchmark suite of real-world codebases).
+  Index maxCombinedBinarySize = 400 * 1024;
   // Loops usually mean the function does heavy work, so the call overhead
   // is not significant and we do not inline such functions by default.
   bool allowFunctionsWithLoops = false;

src/passes/Inlining.cpp

@@ -1398,7 +1398,7 @@ struct Inlining : public Pass {
           continue;
         }
         Name inlinedName = inlinedFunction->name;
-        if (!isUnderSizeLimit(func->name, inlinedName)) {
+        if (!isUnderSizeLimit(func->name, inlinedName, getPassRunner()->options)) {
           continue;
         }
 
@@ -1501,21 +1501,12 @@ struct Inlining : public Pass {
   }
 
   // Checks if the combined size of the code after inlining is under the
-  // absolute size limit. We have an absolute limit in order to avoid
-  // extremely-large sizes after inlining, as they may hit limits in VMs and/or
-  // slow down startup (measurements there indicate something like ~1 second to
-  // optimize a 100K function). See e.g.
-  // https://github.com/WebAssembly/binaryen/pull/3730#issuecomment-867939138
-  // https://github.com/emscripten-core/emscripten/issues/13899#issuecomment-825073344
-  bool isUnderSizeLimit(Name target, Name source) {
+  // absolute size limit.
+  bool isUnderSizeLimit(Name target, Name source, PassOptions& options) {
     // Estimate the combined binary size from the number of instructions.
     auto combinedSize = infos[target].size + infos[source].size;
     auto estimatedBinarySize = Measurer::BytesPerExpr * combinedSize;
-    // The limit is arbitrary, but based on the links above. It is a very high
-    // value that should appear very rarely in practice (for example, it does
-    // not occur on the Emscripten benchmark suite of real-world codebases).
-    const Index MaxCombinedBinarySize = 400 * 1024;
-    return estimatedBinarySize < MaxCombinedBinarySize;
+    return estimatedBinarySize < options.inlining.maxCombinedBinarySize;
   }
 };
 

src/tools/optimization-options.h

@@ -246,6 +246,17 @@ struct OptimizationOptions : public ToolOptions {
              passOptions.inlining.oneCallerInlineMaxSize =
                static_cast<Index>(atoi(argument.c_str()));
            })
+      .add("--inline-max-combined-binary-size",
+           "-imcbs",
+           "Max size of combined functions after inlining. "
+           "Default: " +
+             std::to_string(InliningOptions().maxCombinedBinarySize),
+           OptimizationOptionsCategory,
+           Options::Arguments::One,
+           [this](Options* o, const std::string& argument) {
+             passOptions.inlining.maxCombinedBinarySize =
+               static_cast<Index>(atoi(argument.c_str()));
+           })
       .add("--inline-functions-with-loops",
            "-ifwl",
            "Allow inlining functions with loops",