AArch64 Capstone Update / SME2 support #429

FinnWilkinson · 2024-09-04T10:35:35Z

This PR updates SimEng to work with a newer release of Capstone who's AArch64 engine is based on LLVM18, hence allowing SME2 support.

The reccommended version of LLVM has also been changed to 18.1.8 and the test suite has been updated to work with this version in order to dissassemble new AArch64 instructions.

Closes Issue #349

There have been large changes to Capstone internally for AArch64, some of the main ones which effect our current development / usage include:

A writeback from a pre/post index load/store is now included as an implicit destination. Hence, in instruction_execute the updated base address register is resuts[0] rather than results[1] (for example).
Post index immediate operands are now included as part of the memory operand in mem.disp. Post indexing with a register however is added as an additional operand.
Aliasing can be dissabled for auto-sync architectures, and has been done for AArch64. This means expected registers for the operand are now correct and allows us to remove revertalias() from instruction_metadata completely! The only oddity is the mnemonic and operandStr are still that of the alias. This means exceptions or debugging print statements may be confusing. To indicate aliases better, an isAlias bool has been added to the metadata and exception print out.
Get automatic LLVM download and install working for version 18.1.8
~~Add AArch64 multi-vector operand enum decoding logic to instruction_decode~~
- No longer required as multi-vector operands already give each register individually rather than using the multi-register enums

ABenC377 · 2024-09-20T09:45:25Z

CMakeLists.txt

@@ -149,14 +140,17 @@ if(SIMENG_ENABLE_TESTS)
    find_package(LLVM REQUIRED CONFIG NO_CMAKE_BUILDS_PATH)

      # Check LLVM version
-    if ((${LLVM_PACKAGE_VERSION} VERSION_LESS "8.0") OR (${LLVM_PACKAGE_VERSION} VERSION_GREATER_EQUAL "14.1"))
-      message(FATAL_ERROR "LLVM version must be >= 8.0 and <= 14.0")
+    if ((${LLVM_PACKAGE_VERSION} VERSION_LESS "8.0") OR (${LLVM_PACKAGE_VERSION} VERSION_GREATER_EQUAL "18.2"))


I think the if statement checks for >= 18.2, but the message says can be <= 18.2. Is 18.2 itself okay?

Currently updating how we build LLVM in CMake so will see if this comment is still relevant then

Was there an outcome for this?

These checks are still in place so I will update them inline with Alex's initial comment

Updated message

ABenC377 · 2024-09-20T13:05:19Z

src/include/simeng/arch/aarch64/Instruction.hh

+inline uint8_t getDataSize(cs_aarch64_op op) {
+  // No V-register enum identifiers exist. Instead, depending on whether a full
+  // or half vector is accessed, a Q or D register is used instead.
+  // A `is_vreg` bool in `op` defines if we are using v-vecotr registers.


typo: 'vector'

ABenC377 · 2024-09-20T13:58:56Z

src/lib/arch/aarch64/ExceptionHandler.cc

@@ -658,8 +658,17 @@ bool ExceptionHandler::init() {
    if (metadata.opcode == Opcode::AArch64_MSR) {
      newSVCR = instruction_.getSourceOperands()[0].get<uint64_t>();
    } else if (metadata.opcode == Opcode::AArch64_MSRpstatesvcrImm1) {
+      // Enusre operand metadata is as expected


typo -- 'Ensure'

ABenC377 · 2024-09-20T14:41:04Z

src/lib/arch/aarch64/Instruction_decode.cc

+        sourceRegisters_[sourceRegisterCount_] = csRegToRegister(op.reg);
+        sourceRegisterCount_++;
+        sourceOperandsPending_++;
+
        // TODO checking of the shift type is a temporary fix to help reduce the


Has this todo become actionable by the capstone update?

ABenC377 · 2024-09-23T08:52:37Z

src/lib/config/ModelConfig.cc

@@ -370,8 +370,7 @@ void ModelConfig::setExpectations(bool isDefault) {
    expectations_["Core"].addChild(ExpectationNode::createExpectation<uint64_t>(


If we know that the maximum value for this is 2048, do we need to use unit64_t? Elsewhere we have gone for the smallest suitable datatype. Not sure if there is a technical advantage to using a smaller int type, but if there is then it might apply here as well.

Good point / spot - changing to uint16_t

jj16791

A fair few comments but there's a lot of work done here so expected. Great job on this, a significant amount of effort done!

CMakeLists.txt

jj16791 · 2024-10-26T10:50:36Z

CMakeLists.txt

@@ -149,14 +140,17 @@ if(SIMENG_ENABLE_TESTS)
    find_package(LLVM REQUIRED CONFIG NO_CMAKE_BUILDS_PATH)

      # Check LLVM version
-    if ((${LLVM_PACKAGE_VERSION} VERSION_LESS "8.0") OR (${LLVM_PACKAGE_VERSION} VERSION_GREATER_EQUAL "14.1"))
-      message(FATAL_ERROR "LLVM version must be >= 8.0 and <= 14.0")
+    if ((${LLVM_PACKAGE_VERSION} VERSION_LESS "8.0") OR (${LLVM_PACKAGE_VERSION} VERSION_GREATER_EQUAL "18.2"))


Was there an outcome for this?

docs/sphinx/developer/arch/index.rst

docs/sphinx/developer/arch/supported/aarch64.rst

src/lib/arch/aarch64/Instruction_decode.cc

jj16791 · 2024-10-27T12:17:55Z

src/lib/arch/aarch64/Instruction_decode.cc

+  if (((Opcode::AArch64_FMOVD0 <= metadata_.opcode &&
+        metadata_.opcode <= Opcode::AArch64_FMOVS0) ||
+       (Opcode::AArch64_FMOVDXHighr <= metadata_.opcode &&
+        metadata_.opcode <= Opcode::AArch64_FMOVv8f16_ns)) &&


Aren't the _ns instructions moving immediates into vector registers? (and thus aren't "move to general instructions")

For this I replaced the numbers with Opcodes (and referred to the previous Capstone version to make sure they aligned). If this is the case then we can remove the _ns ones from this catch

Double-checked and we should remove _ns from this catch

Updated catch

src/lib/arch/aarch64/Instruction_decode.cc

dANW34V3R

Not much to add on top of the others' comments. Will approve once theirs have been resolved

dANW34V3R · 2024-10-29T13:15:36Z

docs/sphinx/user/building_simeng.rst

@@ -41,7 +41,7 @@ With this configuration, the build files will be generated in a directory called
                More information about the LLVM_DIR value can be found `here <https://llvm.org/docs/CMake.html#embedding-llvm-in-your-project>`_.

        .. Note::
-                LLVM versions greater than 14 or less than 8 are not supported. We'd recommend using LLVM 14.0.5 where possible as this has been verified by us to work correctly.
+                LLVM versions greater than 18 or less than 8 are not supported. We'd recommend using LLVM 18.1.8 where possible as this has been verified by us to work correctly for the most recent version of SimEng.


Do lower versions of LLVM not support SME? If so that would be worth mentioning here (or at least somewhere in the docs)

Versions from ~14 do support SME. Can add something to this

docs/sphinx/user/configuring_simeng.rst

src/include/simeng/arch/aarch64/helpers/sve.hh

src/lib/arch/aarch64/Architecture.cc

src/lib/arch/riscv/InstructionMetadata.cc

dANW34V3R · 2024-10-29T14:00:25Z

src/lib/arch/aarch64/Instruction_decode.cc

-  // Check from top of the range downwards
+ * `aarch64_reg` enum. Updates to the Capstone library version may cause this to
+ * break.
+ * TODO: Add multi-register enum decoding.


Can this TODO be removed?

…manual DDI0487K issue K.a 20th March 2024.

… it is updated

…h case.

FinnWilkinson added the enhancement New feature or request label Sep 4, 2024

FinnWilkinson self-assigned this Sep 4, 2024

FinnWilkinson force-pushed the sme2-support branch from ce8e138 to 4921682 Compare September 5, 2024 13:31

FinnWilkinson marked this pull request as ready for review September 17, 2024 17:08

FinnWilkinson requested review from dANW34V3R, jj16791, JosephMoore25 and ABenC377 September 17, 2024 17:09

FinnWilkinson added the 0.9.7 Part of SimEng Release 0.9.7 label Sep 17, 2024

FinnWilkinson force-pushed the sme2-support branch from e3f2486 to 8ea3727 Compare September 20, 2024 08:44

ABenC377 requested changes Sep 23, 2024

View reviewed changes

FinnWilkinson changed the title ~~[WIP] AArch64 Capstone Update / SME2 support~~ AArch64 Capstone Update / SME2 support Oct 2, 2024

FinnWilkinson linked an issue Oct 24, 2024 that may be closed by this pull request

Add SME2 support #349

Open

jj16791 requested changes Oct 27, 2024

View reviewed changes

FinnWilkinson mentioned this pull request Oct 28, 2024

Full SME(1) instruction support and STREAMING Groups #415

Open

2 tasks

dANW34V3R reviewed Oct 29, 2024

View reviewed changes

FinnWilkinson mentioned this pull request Nov 4, 2024

[AArch64] NEON, SVE2 and SME2 instruction support with tests #439

Open

FinnWilkinson added 13 commits November 6, 2024 16:45

Temp change of UoB-HPC captone branch to next-update.

cef268e

Began migrating to new Capstone structure.

82ec752

Continued with Capstone update in SimEng.

4c8adeb

Further changes to try get regression tests to all pass.

0e6412c

More fixes after Capstone update.

478a1ac

Finished ensuring regression tests pass.

abe4ac2

Got unit tests to pass.

71a130f

Final fixes to get default program working in emulation mode.

306917e

More changes and fixes.

ed909f8

Updated SVE predicate pattern to use new enum in instruction.

89a9ca3

Began changes to update LLVM version to 18.1.8

87bfdd3

Updated valid SVE vector lengths inline with Arm A profile reference …

ae8b2c5

…manual DDI0487K issue K.a 20th March 2024.

Added table physical register count to model config.

077edd0

FinnWilkinson and others added 17 commits November 6, 2024 16:45

Fixed integration test.

0784017

Added return identification properly.

2068322

Fixed RV64 regression tests / compressed instruction implementation.

1b4a1f4

Updated aarch64 comment.

fd3d013

Removed print out.

81c1f8d

Added new capstone archs to non-build list in cmake.

3130ad2

Fixed microDecode to accomodate new sysop member of aarch64_cs_op.

0614a4c

Updated aarch64 metadata.

8083d88

Final fixes for capstone update.

3a42bbb

Attended PR comments.

f671b8f

Migrate away from add_subdirectory LLVM (#433)

1570b33

Updated LLVM version to 18.1.8 in CMake download.

7001485

Updated docs.

5f2854a

Updated cmake file.

a3e1d14

Added note to CMakeLists about making a new capstone branch each time…

c23f85c

… it is updated

Attended PR comments.

92c2c2d

Reverted to using old-values for SVE VL and fixed AArch64 decode catc…

e7d34e1

…h case.

FinnWilkinson force-pushed the sme2-support branch from ec02455 to e7d34e1 Compare November 6, 2024 16:45

Cmake fix for building LLVM.

b8a2647

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

AArch64 Capstone Update / SME2 support #429

AArch64 Capstone Update / SME2 support #429

FinnWilkinson commented Sep 4, 2024 •

edited

Loading

ABenC377 Sep 20, 2024

FinnWilkinson Oct 2, 2024

jj16791 Oct 26, 2024

FinnWilkinson Oct 28, 2024

FinnWilkinson Oct 31, 2024

ABenC377 Sep 20, 2024

ABenC377 Sep 20, 2024

ABenC377 Sep 20, 2024

ABenC377 Sep 23, 2024

FinnWilkinson Oct 2, 2024

jj16791 left a comment

jj16791 Oct 26, 2024

jj16791 Oct 27, 2024

FinnWilkinson Oct 28, 2024

jj16791 Nov 2, 2024

FinnWilkinson Nov 4, 2024

dANW34V3R left a comment

dANW34V3R Oct 29, 2024

FinnWilkinson Oct 31, 2024

dANW34V3R Oct 29, 2024

		@@ -370,8 +370,7 @@ void ModelConfig::setExpectations(bool isDefault) {
		expectations_["Core"].addChild(ExpectationNode::createExpectation<uint64_t>(

AArch64 Capstone Update / SME2 support #429

Are you sure you want to change the base?

AArch64 Capstone Update / SME2 support #429

Conversation

FinnWilkinson commented Sep 4, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jj16791 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dANW34V3R left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

FinnWilkinson commented Sep 4, 2024 •

edited

Loading