Patch addresses in LLILF/IIHF pairs on class unload and HCR

[Spencer-Comin]

With eclipse-omr/omr#6596, LLILF/IIHF pairs can be used to hold 64 bit address constants in two 32 bit immediate fields. This commit adds logic to HCR and class unloading compensate functions to patch the split constants.

Also, in the case of 32 bit compressed class pointers stored in LLILF (without a following IIHF) the opcode is changed to LGFI when patched on class unload, so that the -1 patched over the address immediate is sign-extended rather than zero-extended.

[Spencer-Comin]

A bit of an explanation/proof of concept for patching. For class unload, the class address is patched over with a -1 value. In the nocompressedrefs case, the 64 class pointer is placed in a LLILF/IIHF pair, and patching is as simple as patching the 32 bit immediate of both instructions with -1 (FFFF FFFF).

class unload nocompressedrefs
BEFORE:
0x3fb77146a46 {UnitTest.virtualCallTest} +80                c00f883ea700 llilf     %r0, 0x883ea700 
0x3fb77146a4c {UnitTest.virtualCallTest} +86                c008000003ff iihf      %r0, 0x3ff
AFTER:
0x3fb77146a46 {UnitTest.virtualCallTest} +80                c00fffffffff llilf     %r0, 0xffffffff 
0x3fb77146a4c {UnitTest.virtualCallTest} +86                c008ffffffff iihf      %r0, 0xffffffff

In the default compressed refs case, the class pointer is guaranteed to fit into 32 bits, so only an LLILF is needed. However, if the immediate is simply patched over with -1, the FFFF FFFF will get zero-extended to 0000 0000 FFFF FFFF rather than sign-extended to keep the -1 value. To fix this, the LLILF needs to be changed to LGFI.

class unload default
BEFORE:
0x3ff7388aa32 {UnitTest.virtualCallTest} +60                c00f00227500 llilf     %r0, 0x227500 MemBlk
AFTER:
0x3ff7388aa32 {UnitTest.virtualCallTest} +60                c001ffffffff lgfi      %r0, -1

For HCR, I could not come up with a unit test where the class address changes on reload. However, when stepping through the compensate function in gdb, the immediate location is patched to the same class address.

HCR default
BEFORE:
0x3ff80113d28 {redef/ClassRedefUnitTest.unitTest} +44                c00f0023ec00 llilf     %r0, 0x23ec00 Ptr
AFTER:
0x3ff80113d28 {redef/ClassRedefUnitTest.unitTest} +44                c00f0023ec00 llilf     %r0, 0x23ec00 Ptr

HCR nocompressedrefs
BEFORE:
0x3fb8f674d28 {redef/ClassRedefUnitTest.unitTest} +44                c00fa0404400 llilf     %r0, 0xa0404400 
0x3fb8f674d2e {redef/ClassRedefUnitTest.unitTest} +50                c008000003ff iihf      %r0, 0x3ff
AFTER:
0x3fb8f674d28 {redef/ClassRedefUnitTest.unitTest} +44                c00fa0404400 llilf     %r0, 0xa0404400 
0x3fb8f674d2e {redef/ClassRedefUnitTest.unitTest} +50                c008000003ff iihf      %r0, 0x3ff

As a sidenote, there might be room for improvement in HCR by skipping patching altogether when the class address doesn't change.

[r30shah]

@Spencer-Comin Is this ready for review or there are some changes needed to be added on ? Also I see that on x86 and Z, we check if the Size is 4 and check the instructions if it is what you generated, can you verify if eclipse-omr/omr#6596 (comment) is actually an issue?

[Spencer-Comin]

I've marked this as draft while I investigate to see how we can avoid the bug @r30shah brought up in eclipse-omr/omr#6596 (comment)

[Spencer-Comin]

FYI @joransiu @r30shah

[r30shah]

@Spencer-Comin can you also run the sanity test on P and X (Reason being it touches the routines that are used by other code-gens as well.) Just for sanity, we should run the test on other platforms. Overall change looks good to me, besides that odd comment regarding to tab.

[Spencer-Comin]

Passed sanity test on P, X, and Z

[r30shah]

@joransiu Can I request you to review and merge this PR?

[joransiu]

jenkins test sanity xlinux,zlinux JDK17

[joransiu]

The x86 sanity.functional cmdLineTester_criu_jitPostRestore failure is not related to this PR: #17367