Implement z/OS XPLINK system linkage #3779

fjeremic · 2019-04-24T21:31:00Z

To get JitBuilder working on z/OS we need a proper z/OS system linkage in place, that means both frame shape and callout support. At the moment we have some half-baked solution in OMR which does not cut it, since we cannot even run the simplest JitBuilder test cases (see #3693).

This PR introduces minimal, but proper XPLINK support to OMR and lays out the groundwork for pushing Java-isms out of OMR and consolidating much of the system linkage support across Linux and z/OS. There are many changes in the PR, each of which deserves it's own commit. Each commit comment outlines the specific change made. The work is not done yet, but to avoid feature creep as this is a deep rabbit hole I've chosen this as the point where the code is good enough to be reviewed and contributed to the project while I continue to work on it in the background.

The following is a summary of some of the goals of this PR:

Implement proper XPLINK z/OS system linkage to allow for JitBuilder test cases in Migrate z/OS to a CMake build system #3693 to run
Create necessary snippets for XPLINK
Partition Linux and z/OS system linkages so they're as identical as possible to make the code simple to understand
Sink OpenJ9 code into OMR and vice-versa
Document existing and newly added APIs

fjeremic · 2019-04-24T21:34:14Z

@mstoodle FYI with this PR and #3693 we are able to from start to finish build JitBuilder on z/OS and run all the tests successfully.

fjeremic

FYI note to reviewers:

This is not the final linkage change I will be delivering. It is merely a good synchronization point for review and initial delivery. I have left many TODOs for myself which I will continue to work on in the background. The system linkage rabbit hole is very deep, and to avoid a super-bloated PR I'm trying to get this portion delivered first as I continue to work on it.

I'll continue adding self reviews to further explain some of the decisions made.

fjeremic · 2019-04-24T21:38:49Z

compiler/z/codegen/OMRCodeGenerator.cpp

+   TR::Instruction* cursor = data.cursorInstruction;
+
+   // TODO: We should be caching the PROC instruction as it's used in several places and is pretty important
+   while (cursor && cursor->getOpCodeValue() != TR::InstOpCode::PROC)


I also want to consolidate this so it's the same on all platforms. I'm working on making such a change at the common level.

fjeremic · 2019-04-24T21:40:57Z

compiler/z/codegen/OMRLinkage.hpp

 #define AggregatesPassedOnParmStack   0x400
 #define AggregatesPassedInParmRegs    0x800
 #define AggregatesReturnedInRegs      0x1000
 // Available                          0x2000
 #define SmallIntParmsAlignedRight     0x4000  ///< < gprSize int parms aligned into the parmword
-#define ParmBlockRegister             0x8000  ///< Has a parameter block register: OS Linkage (non-Java)


This is only for z/OS OS linkage which doesn't exist in OMR, so deprecating it. We only support XPLINK, and currently have no plans to support other system linkages.

fjeremic · 2019-04-24T21:42:19Z

compiler/z/codegen/OMRInstOpCodeEnum.hpp

@@ -60,7 +60,6 @@
   XHHR,                // Exclusive OR High (High <- High)
   XHLR,                // Exclusive OR High (High <- Low)
   XLHR,                // Exclusive OR High (Low <- High)
-   XPCALLDESC,          // zOS-31 LE Call Descriptor.


This used to be used for an "inline" XPLINK call descriptor. It is now replaced but a much more robust and documented XPLINKCallDescriptorSnippet

fjeremic · 2019-04-24T21:43:30Z

compiler/z/codegen/SystemLinkageLinux.cpp

+   setIntegerArgumentRegister(0, TR::RealRegister::GPR2);
+   setIntegerArgumentRegister(1, TR::RealRegister::GPR3);
+
+   if (!comp()->getCurrentMethod()->isJ9())


@0xdaryl I didn't get to this here but I will be moving this out in my subsequent PR.

fjeremic · 2019-04-24T21:45:40Z

compiler/z/codegen/snippet/XPLINKCallDescriptorSnippet.cpp

+   uint8_t* cursor = cg()->getBinaryBufferCursor();
+
+   // TODO: We should not have to do this here. This should be done by the caller.
+   getSnippetLabel()->setCodeLocation(cursor);


I'll be fixing this as well at the common level.

jitbuilder/control/Jit.cpp

0xdaryl · 2019-04-25T00:34:50Z

FYI... Please see the Travis CI build failure.

fjeremic · 2019-04-25T01:35:50Z

FYI... Please see the Travis CI build failure.

Ah, yes. I had to resolve a conflict with one of your recent commits in that file. I must have merged it incorrectly. I'll fix this first thing in the morning.

fjeremic · 2019-04-25T20:04:26Z

@genie-omr build all

fjeremic · 2019-04-26T16:53:50Z

z/OS build failed due to /tmp not being mounted properly. I've contacted IT and go it fixed.

@genie-omr build zos

mstoodle

For the one tiny little bit I have knowledge about, I approve :) .

Maybe @joransiu could provide a review on the XPLINK aspects, please?

jitbuilder/control/Jit.cpp

0xdaryl · 2019-05-10T19:56:17Z

I'll approve once rebased.

Deprecate `TRzOSSystemLinkageBase` and push it's APIs into `S390zOSSystemLinkage` since we only support a single z/OS system linkage. Signed-off-by: Filip Jeremic <[email protected]>

Signed-off-by: Filip Jeremic <[email protected]>

The `generateBinaryEncodingPrologue` has generic code for the following: 1. Creating the recompilation preprologue 2. Creating the method preprologue 3. Creating the method prologue On Z there is nothing specific to OpenJ9 about the above steps. In fact the steps are missing from OMR and hence no prologue/preprologue is generated of system linkages at the OMR level during binary encoding. This is a problem and the fix is to simply migrate this code over. Signed-off-by: Filip Jeremic <[email protected]>

The entry point marker is a 4-word struct right before the entry point of the method body which contains method metadata. [1] https://www.ibm.com/support/knowledgecenter/en/SSLTBW_2.1.0/com.ibm.zos.v2r1.ceev100/rtnly6.htm Signed-off-by: Filip Jeremic <[email protected]>

These labels mark the start and end of the entire method. They are convenience labels such that we always have an instruction at which to insert other instructions, thus effectively prepending instructions to the start of the instruction stream. Signed-off-by: Filip Jeremic <[email protected]>

This hash table is never referenced yet compilation time is wasted adding branches into the table. Remove it to improve compilation time and footprint. Signed-off-by: Filip Jeremic <[email protected]>

We have redundant loops which estimate the binary encoding length of an instruction. Because of the way we generate prologues and recompilation pre-prologues we may miss to estimate some instructions. Avoid this by only having one estimation loop which traverses all instructions and estimates them. Signed-off-by: Filip Jeremic <[email protected]>

Signed-off-by: Filip Jeremic <[email protected]>

Now that we have proper implementations of PPA1 and PPA2 we can generate both snippets as part of prologue creation and add the appropriate relocations in the Entry Point Marker in the function preprologue to correctly update the offset to PPA1. With this commit we have the bare minimum we need for this JIT generated method body to be XPLINK conformant. Signed-off-by: Filip Jeremic <[email protected]>

There is no reason to assert because of an unknown snippet type when tracing. Instead if we end up calling the base class virtual implementation we will print a sane default value indicating to the user that there is indeed a snippet there, however no custom printing logic has been implemented yet. Signed-off-by: Filip Jeremic <[email protected]>

Signed-off-by: Filip Jeremic <[email protected]>

The Linux and z/OS system linkage classes are large and complicated enough that they deserve to live within their own compilation units. This commit partitions up the two classes into: - SystemLinkageLinux.cpp - SystemLinkagezOS.cpp and updates the corresponding include files in the rest of the codebase. We take this opportunity to completely fold `S390SystemLinkage` as it really serves no purpose since the `SystemLinkage` class already exists in the `TR` namespace. Signed-off-by: Filip Jeremic <[email protected]>

The frame types always default to `standardFrame` and we currently do not handle other frame types, such as leaf frames for simplicity. This commit cleans up all related APIs which serve no purpose at the moment. Signed-off-by: Filip Jeremic <[email protected]>

This linkage strategy is no better than what we do by default. In fact it looks almost identical. We deprecate it here to simplify the implementation of the system linkage as this option provides us no benefit. Signed-off-by: Filip Jeremic <[email protected]>

Some APIs such as the creation of a prologue and epilogue are system linkage specific, depending on the calling convention of the system. Such APIs belong in the system linkage classes. Signed-off-by: Filip Jeremic <[email protected]>

With the introduction of four new APIs: - `fillGPRsInEpilogue` - `fillFPRsInEpilogue` - `spillGPRsInPrologue` - `spillFPRsInPrologue` We can migrate code from the the prologue and epilogue creation into the aforementioned APIs to simplify the implementation. There can be more work done within these functions, for example taking advantage of the GPR/FPR masks which will be handled in subsequent commits. Signed-off-by: Filip Jeremic <[email protected]>

Signed-off-by: Filip Jeremic <[email protected]>

The `preProcInstruction` is the instruction that was before the JIT to JIT entry point before the arguments are loaded from the stack. This is the instruction we want to use as the entry point into the method. Signed-off-by: Filip Jeremic <[email protected]>

Currently the z/OS XPLINK only supports 64-bit targets. We currently do not intend to support 31-bit XPLINK targets at the OMR level. There was some pre-existing support for floating point parameter descriptors which are required only for 31-bit to describe floating point arguments. We not longer need this support as the PPA1 structure does not support generating float parameter descriptors. Signed-off-by: Filip Jeremic <[email protected]>

We prepare the code to the be partitioned into the following functions: - `fillGPRsInEpilogue` - `fillFPRsInEpilogue` - `spillGPRsInPrologue` - `spillFPRsInPrologue` similarly to what we did on Linux. This will make the code much easier to understand across the two system linkages. Signed-off-by: Filip Jeremic <[email protected]>

- Create the new snippet - Migrate Call Descriptor generation function to the snippet - Document all the new APIs - Make use of the new snippet when generating the call outs - Clean up generation of Call Descriptor relocation - Remove XPLINKDESC, now unused pseudo-instruction - Enhance internal relocation class to be more generic Signed-off-by: Filip Jeremic <[email protected]>

These relocations can be used by other codegens as well. Document and expose them at the common level. Signed-off-by: Filip Jeremic <[email protected]>

Both linkages should look as identical as possible in terms of API definitions so as to make it easier for a reader to understand. We partition mainly the z/OS system linkage here so it looks more or less like the Linux system linkage. During this exercise we identify that quite some code related to prologue and epilogue creation is actually common, as is the filling and spilling of the registers. This opens up the doors for more commoning to happen at the common system linkage level, a change which will follow shortly. Signed-off-by: Filip Jeremic <[email protected]>

We may be dispatching a call to an XPLINK function from a non-XPLINK function, in which case we may not have an Entry Point Marker for the non-XPLINK body. We still however need the call descriptor to be present as the XPLINK body we are calling will return past the NOP descriptor. Signed-off-by: Filip Jeremic <[email protected]>

Signed-off-by: Filip Jeremic <[email protected]>

fjeremic · 2019-05-10T21:31:15Z

@genie-omr build all

fjeremic added comp:compiler comp:jitbuilder os:linux os:zos arch:z labels Apr 24, 2019

fjeremic requested a review from 0xdaryl April 24, 2019 21:31

fjeremic requested review from charliegracie, Leonardo2718 and mstoodle as code owners April 24, 2019 21:31

fjeremic mentioned this pull request Apr 24, 2019

Implement z/OS XPLINK system linkage eclipse-openj9/openj9#5609

Merged

fjeremic commented Apr 24, 2019

View reviewed changes

fjeremic mentioned this pull request Apr 24, 2019

Migrate z/OS to a CMake build system #3693

Merged

fjeremic force-pushed the xplink-zos-pr branch 2 times, most recently from d5ce5c9 to 0ddae4d Compare April 25, 2019 17:07

mstoodle approved these changes Apr 29, 2019

View reviewed changes

jitbuilder/control/Jit.cpp Show resolved Hide resolved

dchopra001 mentioned this pull request May 7, 2019

Reoarganize fieldwatch-codegen implementation and add support for Z eclipse-openj9/openj9#4743

Merged

fjeremic added 8 commits May 10, 2019 15:56

Deprecate TRzOSSystemLinkageBase

50e6a2c

Deprecate `TRzOSSystemLinkageBase` and push it's APIs into `S390zOSSystemLinkage` since we only support a single z/OS system linkage. Signed-off-by: Filip Jeremic <[email protected]>

Reinstate XPLINK linkage for z/OS

9a5ae83

Signed-off-by: Filip Jeremic <[email protected]>

Add hack to compileMethodBuilder to get z/OS working

3e9e53d

Signed-off-by: Filip Jeremic <[email protected]>

Deprecate branchHashTable

7f5c8da

This hash table is never referenced yet compilation time is wasted adding branches into the table. Remove it to improve compilation time and footprint. Signed-off-by: Filip Jeremic <[email protected]>

fjeremic added 18 commits May 10, 2019 16:05

Create PPA1 snippet

2251d85

Signed-off-by: Filip Jeremic <[email protected]>

Create PPA2 snippet

0d38d30

Signed-off-by: Filip Jeremic <[email protected]>

Add snippets to the back of the snippet list by default

c9ad96f

Signed-off-by: Filip Jeremic <[email protected]>

Remove unused z/OS system linkage APIs

5a9e757

Signed-off-by: Filip Jeremic <[email protected]>

Migrate the instruction to label relocations to the common codegen

78bc1eb

These relocations can be used by other codegens as well. Document and expose them at the common level. Signed-off-by: Filip Jeremic <[email protected]>

fjeremic force-pushed the xplink-zos-pr branch from 0ddae4d to 56d5dde Compare May 10, 2019 20:15

Reference Linkage_inlines.hpp in newly added files

0c66434

Signed-off-by: Filip Jeremic <[email protected]>

0xdaryl self-assigned this May 11, 2019

0xdaryl approved these changes May 11, 2019

View reviewed changes

0xdaryl merged commit 3befa3f into eclipse-omr:master May 11, 2019

fjeremic mentioned this pull request May 24, 2019

Implement proper RX-b instruction support on Z #3900

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement z/OS XPLINK system linkage #3779

Implement z/OS XPLINK system linkage #3779

fjeremic commented Apr 24, 2019

fjeremic commented Apr 24, 2019 •

edited

Loading

fjeremic left a comment

fjeremic Apr 24, 2019

fjeremic Apr 24, 2019

fjeremic Apr 24, 2019

fjeremic Apr 24, 2019

fjeremic Apr 24, 2019

0xdaryl commented Apr 25, 2019

fjeremic commented Apr 25, 2019

fjeremic commented Apr 25, 2019

fjeremic commented Apr 26, 2019

mstoodle left a comment

0xdaryl commented May 10, 2019

fjeremic commented May 10, 2019

Implement z/OS XPLINK system linkage #3779

Implement z/OS XPLINK system linkage #3779

Conversation

fjeremic commented Apr 24, 2019

fjeremic commented Apr 24, 2019 • edited Loading

fjeremic left a comment

Choose a reason for hiding this comment

fjeremic Apr 24, 2019

Choose a reason for hiding this comment

fjeremic Apr 24, 2019

Choose a reason for hiding this comment

fjeremic Apr 24, 2019

Choose a reason for hiding this comment

fjeremic Apr 24, 2019

Choose a reason for hiding this comment

fjeremic Apr 24, 2019

Choose a reason for hiding this comment

0xdaryl commented Apr 25, 2019

fjeremic commented Apr 25, 2019

fjeremic commented Apr 25, 2019

fjeremic commented Apr 26, 2019

mstoodle left a comment

Choose a reason for hiding this comment

0xdaryl commented May 10, 2019

fjeremic commented May 10, 2019

fjeremic commented Apr 24, 2019 •

edited

Loading