[JSC] Work around GCC miscompilation in testmasm

https://bugs.webkit.org/show_bug.cgi?id=307056

Reviewed by Justin Michaud.

It seems that GCC miscompiles the switch statement in
testMoveConditionallyFloatingPointSameArg when -funswitch-loops is on
(that's the case for -O3). The symptom is that we don't handle
MacroAssembler::DoubleEqualAndOrdered (at least) and end up hitting the
following RELEASE_ASSERT_NOT_REACHED();.

Globally turn off this specific transformation when compiling with older
versions of GCC, to avoid running into further hard-to-diagnose bugs.

Canonical link: https://commits.webkit.org/307067@main

Author: aoikonomopoulos

Source/cmake/WebKitCompilerFlags.cmake

@@ -190,6 +190,16 @@ if (COMPILER_IS_GCC_OR_CLANG)
         WEBKIT_PREPEND_GLOBAL_COMPILER_FLAGS(-Wno-uninitialized)
     endif ()
 
+    # Older GCC versions sometimes miscompile switches with that flag on.
+    # Observed in testMoveConditionallyFloatingPointSameArg (testmasm), turn it
+    # off throughout to avoid hard-to-diagnose bugs.
+    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU")
+      # This and later versions don't seem to exhibit the issue.
+      if (${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS "14.0.1")
+        WEBKIT_PREPEND_GLOBAL_COMPILER_FLAGS(-fno-unswitch-loops)
+      endif ()
+    endif ()
+
     WEBKIT_PREPEND_GLOBAL_CXX_FLAGS(-Wno-noexcept-type)
 
     # These GCC warnings produce too many false positives to be useful. We'll