GitLab

Thanks to Peter Wortmann for pointing out this bug.

Previously these were caught by removeDeadAssignments, but we aren't
doing that now.

The new sinker is at least as powerful, and the mini-inliner sometimes
even makes things worse.

We now track stack regions, so that we can float a stack load past a
stack store if they don't conflict.  Also, we now use the CmmType to
more accurately identify heap addresses.

Also lots of refactoring and tidyup

See Note [shortcut call returns]

Instead of relying on common-block-elimination to share return
continuations in the common case (case-alternative heap checks) we do
it explicitly.  This isn't hard to do, is more robust, and saves some
compilation time.  Full commentary in Note [sharing continuations].

There was a long-standing ToDo here that I just did: if a real
register is clobbered by the current instruction, then we should move
it to another free register rather than spilling it to memory.  This
case crops up more often now that the register allocator can allocate
into the fixed Rn registers.

We do this in lots of other places (blockConcat, CmmLayoutStack, etc.)

Proc-point splitting is only required by backends that do not support
having proc-points within a code block (that is, everything except the
native backend, i.e. LLVM and C).

Not doing proc-point splitting saves some compilation time, and might
produce slightly better code in some cases.

We didn't trip over this one until I started generating code without
proc-point splitting, which has more join points and loops.