tracepoint:random:*_read { ... }
kprobe:sys_read, uprobe:/bin/bash:readline { ... }

Previously the auto-generated key type was the same as the value type,
which made maps containing string values use far more stack memory than
they should when accessing them.

Commit 3dd703a5 didn't fix it as intended

For count() maps, sort by value
For others, sort by key

BPF programs can't span multiple functions - add always inliner pass
so multiple calls to strcmp will work.

LLVM version-independent workaround from BCC

e.g. @x = "a"; @y = @x;

Fixes case where two printf statements might have the same format
string but still require their outputs to be displayed differently.
This can happen with the new func builtin, which is printed as a
string, but passes through an address to be symbolicated.
e.g.
kprobe:sys_open { printf("hi %s\n", func); printf("hi %s\n", comm) }

Improves accuracy and efficiency

Issue #5

e.g.: kprobe:sys_read,sys_write { @[func] = count() }

Issue #2

Issue #1

Behave like regular variables from other languages
- scoped within each probe block
- can only be used after being assigned
- not saved to BPF maps

Codegen: Use unsigned versions of division/modulo operations (LLVM BPF backend doesn't support signed)