Use -ffloat-store option when building lmathlib.c with mingw

This fixes the result of math.log with some powers of 2 not generating a whole number

lua/premake5.lua

@@ -19,15 +19,24 @@ project "lua"
 					"src/lutf8lib.c",
 					"src/onelua.c"
 				}
-	
+
 	includedirs { "./" }
 	forceincludes { "luaconf-customize.h" }
 	if _OPTIONS["lua-apicheck"] then
 		defines "LUA_EPRO_APICHECK"
 	end
-	
+
+	-- Workaround for MinGW i386 targetting crtdll.dll, on that C runtime functions returning double
+	-- don't round the values to 64-bit IEEE floats but instead leaves them in the 80-bit float
+	-- registers since they use x87 float instructions. This causes an erroneous results in Lua's
+	-- math.log as the operation log(x)/log(base) is done on 80-bit doubles instead of 64, specifically,
+	-- it affects the log of power of 2s.
+	-- Force GCC to emit code that rounds double results to 64-bits
+	filter { "action:not vs*", "system:windows", "files:src/lmathlib.c"  }
+                buildoptions { "-ffloat-store" }
+
 	filter "configurations:Release"
-		optimize "Speed"	
+		optimize "Speed"
 	filter "configurations:Debug"
 		optimize "Off"