Assist JIT in eliminating bounds checks #81036
Conversation
|
I couldn't figure out the best area label to add to this PR. If you have write-permissions please help me learn by adding exactly one area label. |
ghost
added
the
community-contribution
| @@ -61,7 +61,7 @@ private static bool TryParseNumber(ReadOnlySpan<byte> source, ref Number.NumberB | |||
| int maxDigitCount = digits.Length - 1; | |||
|
|
|||
| // Throw away any leading zeroes | |||
| while (srcIndex != source.Length) | |||
| while (srcIndex < source.Length) | |||
There was a problem hiding this comment.
Do these need uint-casts to eliminate the bound checks?
There was a problem hiding this comment.
In this case casts appear to regress bounds check elimination.
Perhaps this is unnecessary because srcIndex is a local variable that is zero-initialized?
There was a problem hiding this comment.
In this case casts appear to regress bounds check elimination.
Can you elaborate? I wouldn't expect this to remove the bounds check on source[srcIndex] without casting srcIndex to uint here.
There was a problem hiding this comment.
We already do checks on source.Length before the loop. Adding uint casts seems to cause JIT to "forget" about these.
There was a problem hiding this comment.
I still don't understand what you mean. Can you share a sharplab.io example?
There was a problem hiding this comment.
It's worth noting that SharpLab is .NET 7
@tannergooding For future reference, does the codegen from Compiler Explorer have the same optimizations as a local disassembly using .NET 8 nightly build.
Currently, .e.g. https://csharp.godbolt.org/z/cKhTfjvqd, seems to be using:
// crossgen2 8.0.0-preview.3.23152.99+8d2ddc45e8f6f2a726c2b91c1ba2b9f7208b1f3b
There was a problem hiding this comment.
@xtqqczze for godbolt you need to pass -O or better --Ot argument to compiler to enable optimizations. godbolt uses crossgen (R2R) which is, on average, worse than what JIT emits (sharplab). Although, even sharplab's codegen is a bit synthetic since in the real world our JIT is tiered and Tier1 codegen might be different from what you see in sharplab.
There was a problem hiding this comment.
for godbolt you need to pass -O or better --Ot
@EgorBo Not seeing any differeences in codegen for this example: https://csharp.godbolt.org/z/cKhTfjvqd, is this expected?
There was a problem hiding this comment.
Yes. However, its showing the crossgen code and so isn't strictly the same as what sharplab shows or what a user will see for Tier 1 codegen.
You can customize the instruction set targeted and I'd recommend using --instruction-set x86-x64-v3 as that targets AVX2, BMI1, BMI2, LZCNT, MOVBE, and FMA
You can customize the OS: --targetos windows (defaults to linux)
and the architecture: --targetarch arm64 (defaults to x64)
and in general pass in other options supported by crossgen (such as --Ot to optimize for speed)
There was a problem hiding this comment.
@tannergooding Great help with your tips, I knew about the parameters but couldn't find a useful reference with a search engine.
.../System.Linq.Parallel/src/System/Linq/Parallel/Merging/AsynchronousChannelMergeEnumerator.cs
Show resolved
Hide resolved
...s/System.Linq.Parallel/src/System/Linq/Parallel/Merging/SynchronousChannelMergeEnumerator.cs
Show resolved
Hide resolved
src/libraries/System.Net.Http/src/System/Net/Http/Headers/AltSvcHeaderParser.cs
Show resolved
Hide resolved
xtqqczze
requested review from
marek-safar and
MichalStrehovsky
as code owners
xtqqczze
changed the title
|
Tagging subscribers to this area: @dotnet/area-meta Issue DetailsFollow-up: #81001. Change condition in loops from
cc: @stephentoub
|
|
Should we create an issue to add support to JIT to eliminate range checks for these? |
Issue opened: #84697. |
ghost
locked as resolved and limited conversation to collaborators
Follow-up: #81001.
Change condition in loops from
index != T.Lengthtoindex < T.Lengthwhere T:U[]stringSpan<U>ReadOnlySpan<U>cc: @stephentoub