Reduce UInt256 in EVMOperations by shemnon · Pull Request #5331 · besu-eth/besu

shemnon · 2023-04-11T16:21:54Z

PR description

Improve the performance of the EVM by reducing the usage of UInt256 in almost all operations and moving the main switch to a Java 17 expression switch

Fixed Issue(s)

Reduce the use of UInt256 in operations, replacing them with Bytes as appropriate. Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

github-actions · 2023-04-11T16:22:16Z

I thought about documentation and added the doc-change-required label to this PR if updates are required.
I have considered running ./gradlew acceptanceTestNonMainnet locally if my PR affects non-mainnet modules.
I thought about the changelog and included a changelog update if required.

…6-in-operations Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

shemnon · 2023-04-11T23:54:42Z

Benchmark results - https://docs.google.com/spreadsheets/d/1qbvlAveoLCiWbhbxouHgcpFeDVjws7UD0dgZghJZ6oU/edit?usp=sharing

Macbook Pro M1 MAX

Top line result - 10-15% "all in"
Target opcodes - 100-500% improvement. Most notably in JumpDest& SDIV

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

macfarla

a few comments on parts I don't understand
overall I think it's fine
performance improvement sounds fab
think a second opinion would be good :)

macfarla · 2023-04-13T21:54:33Z

-        .thenReturn(UInt256.fromBytes(Bytes32.fromHexStringLenient(shift)))
-        .thenReturn(UInt256.fromHexString(number));
+        .thenReturn(Bytes32.fromHexStringLenient(shift))
+        .thenReturn(Bytes.fromHexString(number));


why not Bytes32 everywhere?

That's something worth benchmarking, but the number of times we need to take a 256 bit number and consider it as a 8 or 32 bit number are very high (memory access, shifts, etc). So my first pass was trimmed bytes, where we can assess size much quicker.

I'll look into Bytes32 or possibly ByteBuffer integrations in a future pass. It did cross my mind that consistently sized words may improve garbage collection implications and (by extension) speed.

macfarla · 2023-04-13T22:21:53Z

+  private static Bytes getByte(final Bytes seq, final Bytes offset) {
+    Bytes trimmedOffset = offset.trimLeadingZeros();
+    if (trimmedOffset.size() > 1) {
+      return Bytes.EMPTY;


in the cases where return UInt256.ZERO has changed to Bytes.EMPTY - I'm not an EVM expert - there is already test coverage for this? no side effects?

Semantically they are the same on the stack, it's quicker to get an int or byte zero out of a Bytes.EMPTY than a UInt256.ZERO (UInt256 would have to check 7 ints as zero before returning the last one, and possibly bit checking for size).

The tests that detected the Tuweni 3.2 constant values bug (where Bytes32.ZERO couldn't be trimmed to 20 bytes in length) would trip this if it was a problem.

macfarla · 2023-04-13T23:18:37Z

@shemnon should this get a changelog entry?

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

…6-in-operations Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

garyschulte

the slicing is a bit scary tbh, but happy to see evm perf improvements. This PR would be a good candidate for fuzzing on a private testnet. What kind of testing did you do already?

garyschulte · 2023-04-14T00:10:37Z

      frame.incrementRemainingGas(gasAvailableForChildCall(frame) + cost);
      frame.popStackItems(getStackItemsConsumed());
-      frame.pushStackItem(UInt256.ZERO);
+      frame.pushStackItem(Bytes.EMPTY);


FAILURE_STACK_ITEM here?

garyschulte · 2023-04-14T00:20:14Z

  public Operation.OperationResult executeFixedCostOperation(
      final MessageFrame frame, final EVM evm) {
-    final UInt256 blockArg = UInt256.fromBytes(frame.popStackItem());
+    final Bytes blockArg = frame.popStackItem().trimLeadingZeros();


might this lead to problems if we are not aligned on a word boundary?

It won't. The call to blockArg.toLong() on line 54 handles any byte length of 8 or less. That's the only use of this variable.

garyschulte · 2023-04-14T00:34:11Z

 public class JumpDestOperation extends AbstractFixedCostOperation {

+  /** constant for a successful jumpdest * */
+  public static final OperationResult jumpdestSuccess = new OperationResult(1L, null);


suggest all caps on public static final

garyschulte · 2023-04-14T00:43:19Z

+    final BigInteger value1 = new BigInteger(1, frame.popStackItem().toArrayUnsafe());

-    final UInt256 result = value0.subtract(value1);
+    final BigInteger result = value0.subtract(value1);


surprising that BigInteger is more performant than UInt256

I expect there are some JVM optimizations targeting that stretch of code. It doesn't look like it's directly an @intrinsic function but it ought to be. Also, the zero check in UInt256 is kind of expensive.

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

shemnon · 2023-04-14T16:52:18Z

Before I press Merge I was going to give it a day or so on Marius's fuzzer. It found the other corner cases fairly quickly.

non-fungible-nelson · 2023-04-18T15:32:42Z

How did it go? @shemnon

Use a primitive to bytes function that is more direct than the Bytes methods. Signed-off-by: Danno Ferrin <danno.ferrin@gmail.com>

Signed-off-by: Danno Ferrin <danno.ferrin@gmail.com>

shemnon · 2023-04-19T21:23:40Z

Fuzz testing was useful.

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

Performance improvements - Reduce the use of UInt256 in operations, replacing them with Bytes as appropriate. - Use Java17 expression switch Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

shemnon added 2 commits April 7, 2023 23:47

Reduce use of UInt256 in operations

6a9766f

Reduce the use of UInt256 in operations, replacing them with Bytes as appropriate. Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

Use Java17 switch

dfb909a

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

shemnon added 6 commits April 11, 2023 10:22

Merge branch 'main' of github.com:hyperledger/besu into reduce-uint25…

8b603e9

…6-in-operations Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

exp, jump, and jumpi

b245536

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

javadoc

5ca0482

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

unit tests are still tied deeply to internal types

b636f15

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

spotless

8fea8a6

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

spotless

9d1d176

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

merge

9b4390c

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

macfarla reviewed Apr 13, 2023

View reviewed changes

shemnon added 3 commits April 13, 2023 17:53

remove dead code

7666d5f

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

changelog

9ba0008

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

Merge branch 'main' of github.com:hyperledger/besu into reduce-uint25…

a814aa8

…6-in-operations Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

macfarla approved these changes Apr 14, 2023

View reviewed changes

garyschulte approved these changes Apr 14, 2023

View reviewed changes

merge and review fixes

1fe6e74

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

shemnon added 2 commits April 19, 2023 15:20

Change primitive to bytes function

f7cca55

Use a primitive to bytes function that is more direct than the Bytes methods. Signed-off-by: Danno Ferrin <danno.ferrin@gmail.com>

merge

2421bfa

Signed-off-by: Danno Ferrin <danno.ferrin@gmail.com>

Merge branch 'main' into reduce-uint256-in-operations

7ab6f63

shemnon enabled auto-merge (squash) April 20, 2023 03:22

javadoc

1b343e1

Signed-off-by: Danno Ferrin <danno.ferrin@swirldslabs.com>

shemnon force-pushed the reduce-uint256-in-operations branch from b8ab092 to 1b343e1 Compare April 20, 2023 04:50

shemnon merged commit b3d3f54 into besu-eth:main Apr 20, 2023

siladu mentioned this pull request Feb 18, 2026

Improve shift opcodes performance (SAR, SHR and SHL) #9796

Merged

10 tasks

Conversation

shemnon commented Apr 11, 2023

PR description

Fixed Issue(s)

Uh oh!

github-actions Bot commented Apr 11, 2023 • edited by shemnon Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

shemnon commented Apr 11, 2023

Uh oh!

macfarla left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

macfarla commented Apr 13, 2023

Uh oh!

garyschulte left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

shemnon commented Apr 14, 2023

Uh oh!

non-fungible-nelson commented Apr 18, 2023

Uh oh!

shemnon commented Apr 19, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

github-actions Bot commented Apr 11, 2023 •

edited by shemnon

Loading