JIT: extend VN to understand method table fetch from newobj #52476
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
If we see a GT_IND whose base is JitNew, the value number of the IND is the same as the value number of the class handle for the new. Partial fix for dotnet#52370.
ghost
added
the
area-CodeGen-coreclr
Member
Author
|
@EgorBo PTAL Results on the #52370 case are similar to Egor's fix. Would be nice to get rid of that frame. Also more changes in this mode are possible, eg new feeding a type test helper, or loading a field of a newly allocated object. IN0002: 000000 sub rsp, 40
G_M4678_IG02: ; offs=000004H, size=0005H, bbWeight=1 PerfScore 0.25, gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref
IN0001: 000004 mov eax, 3
G_M4678_IG03: ; offs=000009H, size=0005H, bbWeight=1 PerfScore 1.25, epilog, nogc, extend
IN0003: 000009 add rsp, 40
IN0004: 00000D ret Quite a few diffs in SPMI testing (which is a bit ragged still with missing info). Will summarize shortly. |
Member
|
Nice, but will it work without JitNew? e.g.: using System.Runtime.CompilerServices;
using System.Threading;
public class Program
{
public static void Main()
{
// Promote Test to tier1 and collect some profile data
for (int i = 0; i < 100; i++)
{
Test();
Thread.Sleep(16);
}
}
[MethodImpl(MethodImplOptions.NoInlining)]
static B GetB() => new B(); // not inlineable, but returns sealed type
static A GetA() => GetB();
[MethodImpl(MethodImplOptions.NoInlining)]
static int Test()
{
return GetA().GetValue();
}
}
public class A
{
public virtual int GetValue() => 1;
}
public sealed class B : A
{
public override int GetValue() => 3;
} |
EgorBo
approved these changes
May 8, 2021
Member
Author
|
SPMI diffs Some large improvements in places, typically where the optimization leads to folding a branch. A few large regressions in test cases that do a lot of type testing, as we are now propagating a large literal method table value to many uses. We should consider either not propagating such a large constant, or CSEing it back later. asm.aspnet.run.windows.x64.checked Detail diffsasm.benchmarks.run.windows.x64.checked Detail diffsasm.libraries.crossgen.windows.x64.checked Detail diffsasm.libraries.pmi.windows.x64.checked Detail diffsasm.tests.pmi.windows.x64.checked Detail diffsasm.tests_libraries.pmi.windows.x64.checked Detail diffs |
Member
Author
No. Which is why you should go ahead with your proposed fix too. VN generally doesn't handle conditional facts. So say we learn the exact type of some ssa value because of a type test introduced by GDV. VN has no way of expressing this unless. Some compilers I've seen will introduce "fake" ssa defs in order to extend the power of analyses like VN, eg if (x == 3) S1; else S2;would become if (x == 3) { x = 3; S1; } else S2;Assertion prop can also operate on conditional facts like these; not sure if it handles this case. Worth investigating. |
Member
Author
|
Failure is a known issue (#52464). |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
If we see a GT_IND whose base is JitNew, the value number of the IND
is the same as the value number of the class handle for the new.
Partial fix for #52370.