[lworld] RFR: 8272566: [lworld] [AArch64] G1 write barrier slow path needs to save more registers

[Nick Gasson]

We call MacroAssembler::store_heap_oop() when packing inline type
arguments in C1's scalarised entry point and that in turn uses the
interpreter's barrier set assembler for the write barrier.  With G1GC
this can call into the runtime on the slow path, but it doesn't preserve
the call-clobbered registers, which include the Java argument registers.

This was fixed on x86 in JDK-8242210 and JDK-8251398 but I missed that
when I updated the AArch64 port earlier in the year.  I haven't actually
seen any failures caused by this, but from looking at the generated
assembly it's certainly possible.

This patch saves only the argument registers, and only if Valhalla is
enabled (unlike x86 which saves all registers unconditionally).  There
are 32 caller-saved registers on AArch64 so I wanted to avoid generating
many additional store/load instructions.

-------------

Commit messages:
 - 8272566: [lworld] [AArch64] G1 write barrier slow path needs to save more registers

Changes: https://git.openjdk.java.net/valhalla/pull/534/files
 Webrev: https://webrevs.openjdk.java.net/?repo=valhalla&pr=534&range=00
  Issue: https://bugs.openjdk.java.net/browse/JDK-8272566
  Stats: 44 lines in 2 files changed: 35 ins; 0 del; 9 mod
  Patch: https://git.openjdk.java.net/valhalla/pull/534.diff
  Fetch: git fetch https://git.openjdk.java.net/valhalla pull/534/head:pull/534

PR: https://git.openjdk.java.net/valhalla/pull/534