sbcl 1.0.48 shows large compile time regression compilng FriCAS: earlier versions needed few minutes
of compile time, with sbcl 1.0.48 compilations takes more than 3 hours. The machine is 2.4 GHz Core 2
Quatro running 64-bit Fedora 9 (upgraded to 2.6.32 Linux kernel). In attachement there is a standalone
file which take 0.6s to compile using sbcl-1.0.16 but 416s using sbcl-1.0.48. Actually during FriCAS build
compiling this file took more than 1 hour. I had to add a few macros so that the file compiles independently
of the rest of FriCAS. I do not know what causes the difference between FriCAS build and standalone
compilation, but already standalone version shows huge regression.
Introduced in 1.0.46.33: constraint: Don't substitute REFs when replacement LEAF is not visible.
I looked a bit at randomly-sampled backtraces, and the change looks like it substantially increases the pressure on consets. How does that change interact with inline/local functions? Is let conversion guaranteed to run before cprop?