Stream: git-wasmtime

Topic: wasmtime / issue #4923 Cranelift: Quadratic compile time ...

Wasmtime GitHub notifications bot (Sep 19 2022 at 15:01):

adambratschikaye labeled issue #4923:

It seems that a wasm source with a chain of blocks each containing a local.get instruction can have quadratic compile time because can_optimize_var_lookup will traverse the entire chain when handling each local.get instruction. As an example, the following wat takes 2.4 seconds to compile on my machine:

(module
  (func (param i32)
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    ... repeated 15,000 times
  )
)

and a flamegraph:  shows that almost the entire time is spent in can_optimize_var_lookup. Note that the wasm file is 147 KB.

Steps to Reproduce

• Download the attached blocks.wasm.gz and unzip it (or take the above wat and convert it to wasm).
• Compile it with time target/release/wasmtime compile blocks.wasm.

Expected Results

I'd expect the compilation to be on the order of 100ms as it is for other modules with a single function of similar size on my machine.

Actual Results

The compilation takes over 2 seconds.

Versions and Environment

Cranelift version or commit: commit 27435ae398bcce74bf990b9683e5c47c3f6c5d51

Operating system: Ubuntu 20.04, Linux kernel 5.4.0-122-generic

Architecture:

Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   43 bits physical, 48 bits virtual
CPU(s):                          64
On-line CPU(s) list:             0-63
Thread(s) per core:              2
Core(s) per socket:              16
Socket(s):                       2
NUMA node(s):                    1
Vendor ID:                       AuthenticAMD
CPU family:                      23
Model:                           49
Model name:                      AMD EPYC 7302 16-Core Processor

Extra Info

I'm working on a PR which should help with this.

Wasmtime GitHub notifications bot (Sep 19 2022 at 15:01):

adambratschikaye opened issue #4923:

It seems that a wasm source with a chain of blocks each containing a local.get instruction can have quadratic compile time because can_optimize_var_lookup will traverse the entire chain when handling each local.get instruction. As an example, the following wat takes 2.4 seconds to compile on my machine:

(module
  (func (param i32)
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    ... repeated 15,000 times
  )
)

and a flamegraph:  shows that almost the entire time is spent in can_optimize_var_lookup. Note that the wasm file is 147 KB.

Steps to Reproduce

• Download the attached blocks.wasm.gz and unzip it (or take the above wat and convert it to wasm).
• Compile it with time target/release/wasmtime compile blocks.wasm.

Expected Results

I'd expect the compilation to be on the order of 100ms as it is for other modules with a single function of similar size on my machine.

Actual Results

The compilation takes over 2 seconds.

Versions and Environment

Cranelift version or commit: commit 27435ae398bcce74bf990b9683e5c47c3f6c5d51

Operating system: Ubuntu 20.04, Linux kernel 5.4.0-122-generic

Architecture:

Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   43 bits physical, 48 bits virtual
CPU(s):                          64
On-line CPU(s) list:             0-63
Thread(s) per core:              2
Core(s) per socket:              16
Socket(s):                       2
NUMA node(s):                    1
Vendor ID:                       AuthenticAMD
CPU family:                      23
Model:                           49
Model name:                      AMD EPYC 7302 16-Core Processor

Extra Info

I'm working on a PR which should help with this.

Wasmtime GitHub notifications bot (Sep 19 2022 at 15:01):

adambratschikaye labeled issue #4923:

It seems that a wasm source with a chain of blocks each containing a local.get instruction can have quadratic compile time because can_optimize_var_lookup will traverse the entire chain when handling each local.get instruction. As an example, the following wat takes 2.4 seconds to compile on my machine:

(module
  (func (param i32)
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    ... repeated 15,000 times
  )
)

and a flamegraph:  shows that almost the entire time is spent in can_optimize_var_lookup. Note that the wasm file is 147 KB.

Steps to Reproduce

• Download the attached blocks.wasm.gz and unzip it (or take the above wat and convert it to wasm).
• Compile it with time target/release/wasmtime compile blocks.wasm.

Expected Results

I'd expect the compilation to be on the order of 100ms as it is for other modules with a single function of similar size on my machine.

Actual Results

The compilation takes over 2 seconds.

Versions and Environment

Cranelift version or commit: commit 27435ae398bcce74bf990b9683e5c47c3f6c5d51

Operating system: Ubuntu 20.04, Linux kernel 5.4.0-122-generic

Architecture:

Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   43 bits physical, 48 bits virtual
CPU(s):                          64
On-line CPU(s) list:             0-63
Thread(s) per core:              2
Core(s) per socket:              16
Socket(s):                       2
NUMA node(s):                    1
Vendor ID:                       AuthenticAMD
CPU family:                      23
Model:                           49
Model name:                      AMD EPYC 7302 16-Core Processor

Extra Info

I'm working on a PR which should help with this.

Wasmtime GitHub notifications bot (Sep 20 2022 at 07:48):

adambratschikaye closed issue #4923:

It seems that a wasm source with a chain of blocks each containing a local.get instruction can have quadratic compile time because can_optimize_var_lookup will traverse the entire chain when handling each local.get instruction. As an example, the following wat takes 2.4 seconds to compile on my machine:

(module
  (func (param i32)
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    ... repeated 15,000 times
  )
)

and a flamegraph:  shows that almost the entire time is spent in can_optimize_var_lookup. Note that the wasm file is 147 KB.

Steps to Reproduce

• Download the attached blocks.wasm.gz and unzip it (or take the above wat and convert it to wasm).
• Compile it with time target/release/wasmtime compile blocks.wasm.

Expected Results

I'd expect the compilation to be on the order of 100ms as it is for other modules with a single function of similar size on my machine.

Actual Results

The compilation takes over 2 seconds.

Versions and Environment

Cranelift version or commit: commit 27435ae398bcce74bf990b9683e5c47c3f6c5d51

Operating system: Ubuntu 20.04, Linux kernel 5.4.0-122-generic

Architecture:

Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   43 bits physical, 48 bits virtual
CPU(s):                          64
On-line CPU(s) list:             0-63
Thread(s) per core:              2
Core(s) per socket:              16
Socket(s):                       2
NUMA node(s):                    1
Vendor ID:                       AuthenticAMD
CPU family:                      23
Model:                           49
Model name:                      AMD EPYC 7302 16-Core Processor

Extra Info

I'm working on a PR which should help with this.

Wasmtime GitHub notifications bot (Sep 21 2022 at 18:16):

jameysharp commented on issue #4923:

My draft fix for this is in #4939.

Wasmtime GitHub notifications bot (Sep 21 2022 at 18:16):

jameysharp reopened issue #4923:

It seems that a wasm source with a chain of blocks each containing a local.get instruction can have quadratic compile time because can_optimize_var_lookup will traverse the entire chain when handling each local.get instruction. As an example, the following wat takes 2.4 seconds to compile on my machine:

(module
  (func (param i32)
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    ... repeated 15,000 times
  )
)

and a flamegraph:  shows that almost the entire time is spent in can_optimize_var_lookup. Note that the wasm file is 147 KB.

Steps to Reproduce

• Download the attached blocks.wasm.gz and unzip it (or take the above wat and convert it to wasm).
• Compile it with time target/release/wasmtime compile blocks.wasm.

Expected Results

I'd expect the compilation to be on the order of 100ms as it is for other modules with a single function of similar size on my machine.

Actual Results

The compilation takes over 2 seconds.

Versions and Environment

Cranelift version or commit: commit 27435ae398bcce74bf990b9683e5c47c3f6c5d51

Operating system: Ubuntu 20.04, Linux kernel 5.4.0-122-generic

Architecture:

Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   43 bits physical, 48 bits virtual
CPU(s):                          64
On-line CPU(s) list:             0-63
Thread(s) per core:              2
Core(s) per socket:              16
Socket(s):                       2
NUMA node(s):                    1
Vendor ID:                       AuthenticAMD
CPU family:                      23
Model:                           49
Model name:                      AMD EPYC 7302 16-Core Processor

Extra Info

I'm working on a PR which should help with this.

Wasmtime GitHub notifications bot (Sep 23 2022 at 16:41):

jameysharp closed issue #4923:

It seems that a wasm source with a chain of blocks each containing a local.get instruction can have quadratic compile time because can_optimize_var_lookup will traverse the entire chain when handling each local.get instruction. As an example, the following wat takes 2.4 seconds to compile on my machine:

(module
  (func (param i32)
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    (block (local.set 0 (i32.add (i32.const 1) (local.get 0))))
    ... repeated 15,000 times
  )
)

and a flamegraph:  shows that almost the entire time is spent in can_optimize_var_lookup. Note that the wasm file is 147 KB.

Steps to Reproduce

• Download the attached blocks.wasm.gz and unzip it (or take the above wat and convert it to wasm).
• Compile it with time target/release/wasmtime compile blocks.wasm.

Expected Results

I'd expect the compilation to be on the order of 100ms as it is for other modules with a single function of similar size on my machine.

Actual Results

The compilation takes over 2 seconds.

Versions and Environment

Cranelift version or commit: commit 27435ae398bcce74bf990b9683e5c47c3f6c5d51

Operating system: Ubuntu 20.04, Linux kernel 5.4.0-122-generic

Architecture:

Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   43 bits physical, 48 bits virtual
CPU(s):                          64
On-line CPU(s) list:             0-63
Thread(s) per core:              2
Core(s) per socket:              16
Socket(s):                       2
NUMA node(s):                    1
Vendor ID:                       AuthenticAMD
CPU family:                      23
Model:                           49
Model name:                      AMD EPYC 7302 16-Core Processor

Extra Info

I'm working on a PR which should help with this.

Last updated: Dec 23 2024 at 12:05 UTC