build: Refactor get_concurrent_links.py

build/toolchain/concurrent_links.gni

Index: build/toolchain/concurrent_links.gni
diff --git a/build/toolchain/concurrent_links.gni b/build/toolchain/concurrent_links.gni
index 9705255b4d48a30725b489bada9666c6caa8a8b5..75d4932d5c524d07e8713a9db22dc4a2f965d5bd 100644
--- a/build/toolchain/concurrent_links.gni
+++ b/build/toolchain/concurrent_links.gni
@@ -20,7 +20,14 @@ declare_args() {
 
 if (concurrent_links == -1) {
   if (allow_posix_link_time_opt || is_cfi) {
-    _args = [ "--lto" ]
+    _args = [
+      "--mem_per_link_gb=24",
+      "--reserve_mem_gb=10",
+    ]
+  } else if (is_win) {
+    _args = [ "--mem_per_link_gb=5" ]

[brucedawson, 2016/08/22 19:31:32]

The Windows memory used per linker is highly dependent on the binary being
linked and the build type. The worst-case memory consumption amounts (recorded a
few months ago) are:

64-bit official: 27 GB
32-bit official: 14 GB
64-bit/32-bit static release: 11 GB
64-bit/32-bit component release: 6 GB

These are worst-case numbers, but towards the end of the build process there are
often quite a few worst-cast (test) binaries trying to link simultaneously.

I don't have component debug numbers handy, but I think they are similar to the
component release numbers.

Starting with 5 for compatibility with the existing values is fine, but it might
be worth tweaking after that.

[boliu, 2016/08/22 20:49:03]

This was just intended to be a refactor, without changing behavior for any
platform.

I don't personally build for windows, but hopefully your proposed changes will
get easier after this lands.

+  } else if (is_mac) {
+    _args = [ "--mem_per_link_gb=4" ]
   } else {
     _args = []
   }