I have not built or tested this on Windows, only written the code to hopefully do roughly the right thing! 😀

Looks good, after a preliminary review. I did not get a chance yet to test this properly, hopefully by the end of next week.

The red appveyor builds puzzle me!

The red appveyor builds puzzle me!

Indeed. It seems that ../src/curl.exe --version segfaults... :-(

I presume then that something in this new curlrc code of mine is totally borked...

I can't make it crash with --version but I don't know how to build it the same exact way as appveyor. VS in debug mode doesn't detect anything. I chose rebuild pr I'm curious if it's arbitrary or the results will be the same.

@jay I can if I do:

set APPDATA=
set HOME=
curl.exe --version

The WinDbg call-stack points the finger at:

ucrtbased!strcmp(unsigned char * str1 = 0x00000000 "", unsigned char * str2 = 0x005bf900 "???")+0x10 [minkernel\crts\ucrt\src\appcrt\string\i386\strcmp.asm @ 82] 
curl!parseconfig(char * filename = 0x00000000 "", struct GlobalConfig * global = 0x005bf900)+0x1b1 [F:\MingW32\src\inet\curl\src\tool_parsecfg.c @ 133] 

Unusual I know to have neither of these. Something that should be considered?

Ah, that remark made me realize the new logic only checks for curlrc in the excutable's directory if it figures out a home directory, which seems wrong! I'll fix.

edited, squashed and rebased

Oops, torture test failures. This is leaking memory in some exit paths. I'll fix...

Considering the message in url.c:

      infof(data, "Couldn't find host %s in the "
            DOT_CHAR "netrc file; using defaults\n",
            conn->host.name);

Shouldn't curl_setup.h be patched too?

--- a/curl_setup.h 2019-08-04 11:16:15
+++ b/curl_setup.h 2019-08-20 10:48:34
@@ -486,7 +486,7 @@
 #ifdef WIN32

 #  define DIR_CHAR      "\\"
-#  define DOT_CHAR      "_"
+#  define DOT_CHAR      "."

 #else /* WIN32 */