llvm-project: link against jemalloc

[nickdesaulniers]

The following symbols from musl's malloc-ng implementation were
appearing in a perf profile of a kernel build within the top 20 hottest
symbols by cycle count.

malloc-ng:
Samples: 589K of event 'cycles:ppu', Event count (approx.): 11053523397362
  Overhead  Shared Object              Symbol
+    1.57%  clang                     [.] __libc_malloc_impl
     0.69%  clang                     [.] __libc_free
+    0.65%  clang                     [.] alloc_slot
+    0.58%  clang                     [.] get_meta

$ /usr/bin/time -v make CC=clang LD=ld.lld -j72 -s
Command being timed: "make CC=clang LD=ld.lld -j72 -s"
  User time (seconds): 2801.78
  System time (seconds): 395.48
  Percent of CPU this job got: 5212%
  Elapsed (wall clock) time (h:mm:ss or m:ss): 1:01.33

jemalloc:
Samples: 569K of event 'cycles:ppu', Event count (approx.): 10681965442287
  Overhead  Shared Object              Symbol
+    0.55%  clang                     [.] free
+    0.41%  clang                     [.] malloc

$ /usr/bin/time -v make CC=clang LD=ld.lld -j72 -s
Command being timed: "make CC=clang LD=ld.lld -j72 -s"
User time (seconds): 2626.97
System time (seconds): 248.37
Percent of CPU this job got: 4774%
Elapsed (wall clock) time (h:mm:ss or m:ss): 1:00.22

And these symbols are no longer in the top 20 hottest symbols by cycle
count.

Replace musl's malloc with jemalloc. Nothing major to write home about,
but gives a ~2.5% speedup by cycle count (N=1).

Fixes #36

[compnerd]

This seems fine procedurally. I do have to ask - which malloc in musl was being used? There is oldmalloc and malloc-ng IIRC, and I'm curious if switching between those has any impact on performance. It would be nice to have that sort of information available to better explain the reasoning for the switch (and perhaps expand on that a bit more in the commit message).

[nickdesaulniers]

Based on the disassembly, it looked like malloc-ng IIRC.

Let me confirm, then collect some more number, then update the commit message.

[nickdesaulniers]

jemalloc:

$ PATH=/android2/containers/llvm-project:$PATH /usr/bin/time -v make CC=clang LD=ld.lld -j72 -s        
	Command being timed: "make CC=clang LD=ld.lld -j72 -s"
	User time (seconds): 2626.97
	System time (seconds): 248.37
	Percent of CPU this job got: 4774%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 1:00.22

 0.55%  clang                     [.] free
 0.41%  clang                     [.] malloc

so malloc and free are still hot BUT they fall out of the top 20 hottest functions like we have with malloc-ng, and they take less time.  Maybe `-Xclang -disable-free` for another slight improvement...

malloc-ng:
```sh
$ PATH=/android2/containers/llvm-project:$PATH /usr/bin/time -v make CC=clang LD=ld.lld -j72 -s
	Command being timed: "make CC=clang LD=ld.lld -j72 -s"
	User time (seconds): 2801.78
	System time (seconds): 395.48
	Percent of CPU this job got: 5212%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 1:01.33

In the container, I can see that ${BUILD_DIR}/obj/src/malloc/ contains mallocng/ but not oldmalloc/.