IsSupported returns false for not-fully-implemented ISA classes #15514

fiigii · 2017-12-13T23:36:00Z

Resolve #14930

fiigii · 2017-12-13T23:38:53Z

@jkotas @CarolEidt @BruceForstall My current solution is to add a new environment variable COMPlus_EnableIncompleteISAClass to test partially implemented classes.

CarolEidt

I think the new config variable is a good approach, but would like to see it enabled only under DEBUG (i.e. Debug or Checked builds).

CarolEidt · 2017-12-14T01:55:44Z

src/jit/hwintrinsicxarch.cpp

+        case InstructionSet_FMA:
+        case InstructionSet_PCLMULQDQ:
+            // DONE - InstructionSet_LZCNT:
+            // DONE - InstructionSet_POPCNT:


I would prefer to see explicit cases for the ones that are done, and then the default should be unreached().

CarolEidt · 2017-12-14T02:09:34Z

src/jit/hwintrinsicxarch.cpp

-    if ((!compSupports(isa) || (!featureSIMD && isa != InstructionSet_BMI1 && isa != InstructionSet_BMI2 &&
-                                isa != InstructionSet_LZCNT && isa != InstructionSet_POPCNT)) &&
+    // - the ISA class is not yet fully implemented and EnableIncompleteISAClass=false
+    if ((!compSupports(isa) || (!JitConfig.EnableIncompleteISAClass() && !isFullyImplmentedISAClass(isa)) ||


I think it would be cleaner to have this check look something like this:

if ((!compSupports(isa) || !compSupportsIntrinsics(isa)) && !isIntrinsicAnIsSupportedPropertyGetter(intrinsic))

Then, compSupportsIntrinsics(isa) can incorporate both the check for those that are fully supported, as well as, under DEBUG only, checking JitConfig.EnableIncompleteISAClass().

That would be more elegant. Will change.

CarolEidt · 2017-12-14T02:10:38Z

src/jit/hwintrinsicxarch.cpp

@@ -186,6 +186,33 @@ bool Compiler::isIntrinsicAnIsSupportedPropertyGetter(NamedIntrinsic intrinsic)
    }
 }

+// TODO - remove the fully implemented ISA


I think that all these TODO comments should be omitted. I expect that it will take some time to complete the current set of ISAs, and there are likely to be more in future.

fiigii · 2017-12-14T04:51:20Z

I think the new config variable is a good approach, but would like to see it enabled only under DEBUG

@CarolEidt I remember we talked this issue in that email thread. @jkotas mentioned that inconsistent behavior between release and debug build will make some test problems.
If the inconsistency is allowed, I think this new envirnoment variable would be unnecessary.
My original plan was to disable incomplete classes in release and debug both, then add some new CI jobs with COMPlus_EnableIncompleteISAClass.

fiigii · 2017-12-14T21:58:08Z

@jkotas @BruceForstall comments?

jkotas · 2017-12-14T22:01:38Z

Without any environment variable, the debug and release need to behave the same.

It is fine for the environment variable to be respected in debug build only. Debug and release can behave differently with environment variable set.

fiigii · 2017-12-14T22:04:06Z

It is fine for the environment variable to be respected in debug build only. Debug and release can behave differently with environment variable set.

Got it, thank you!

fiigii · 2017-12-17T18:40:23Z

Moved COMPlus_EnableIncompleteISAClass under DEBUG and addressed feedback. @jkotas @CarolEidt PTAL.

CarolEidt · 2017-12-18T18:09:21Z

src/jit/compiler.h

@@ -3044,6 +3044,7 @@ class Compiler
    GenTree* impLZCNTIntrinsic(NamedIntrinsic intrinsic, CORINFO_METHOD_HANDLE method, CORINFO_SIG_INFO* sig);
    GenTree* impPCLMULQDQIntrinsic(NamedIntrinsic intrinsic, CORINFO_METHOD_HANDLE method, CORINFO_SIG_INFO* sig);
    GenTree* impPOPCNTIntrinsic(NamedIntrinsic intrinsic, CORINFO_METHOD_HANDLE method, CORINFO_SIG_INFO* sig);
+    bool compSupportHWIntrinsic(InstructionSet isa);


This should be compSupportsHWIntrinsic

CarolEidt · 2017-12-18T18:12:17Z

src/jit/hwintrinsicxarch.cpp

+            // TODO - remove "JitConfig.EnableIncompleteISAClass()" after fully implement hardware intrinsics of this
+            // class
+            retNode =
+                gtNewIconNode(JitConfig.EnableIncompleteISAClass() && featureSIMD && compSupports(InstructionSet_SSE));


Isn't this test redundant with the one above? The check for JitConfig.EnableIncompleteISAClass() can only be done under DEBUG, but I believe that this check is redundant anyway, unless there's something I've missed.
Also, I still think these "TODO" comments are unnecessary, as I believe that we will need that facility for some time as new ISA feature sets are added and implemented over time.

Isn't this test redundant with the one above?

No, this check lets IsSupported return false for incomplete classes, but the above one lets intrinsics throw PNSE when the call site is not guarded by IsSupported.

The check for JitConfig.EnableIncompleteISAClass() can only be done under DEBUG,

Oops, my mistake. Will fix.

CarolEidt · 2017-12-18T18:58:07Z

src/jit/hwintrinsicxarch.cpp

-                                isa != InstructionSet_LZCNT && isa != InstructionSet_POPCNT)) &&
-        !isIntrinsicAnIsSupportedPropertyGetter(intrinsic))
+    // - JIT does not support this hardware intrinsics (compSupportHWIntrinsic return false)
+    if ((!compSupports(isa) || !compSupportHWIntrinsic(isa)) && !isIntrinsicAnIsSupportedPropertyGetter(intrinsic))


Could you extract the check for the specific ISA intrinsics at the beginning, e.g.

bool supported = (featureSIMD && compSupports(isa) && compSupportsHWIntrinsics(isa));

Then, the above check can be:

if (!supported && !isIntrinsicAnIsSupportedPropertyGetter(intrinsic))

and the checks below can be:

retNode = gtNewIconNode(supported);

Good point, will do.

fiigii · 2017-12-19T20:22:19Z

Streamlined IsSupported code. @CarolEidt @jkotas PTAL

BruceForstall · 2017-12-20T00:48:10Z

src/jit/hwintrinsicxarch.cpp

@@ -186,6 +186,49 @@ bool Compiler::isIntrinsicAnIsSupportedPropertyGetter(NamedIntrinsic intrinsic)
    }
 }

+bool isFullyImplmentedISAClass(InstructionSet isa)


This function needs a standard header comment

Also, please make this a (static) member of the Compiler class.

BruceForstall · 2017-12-20T00:48:31Z

src/jit/hwintrinsicxarch.cpp

+    }
+}
+
+// return true if


Please use a standard header comment format

BruceForstall · 2017-12-20T00:52:23Z

src/jit/hwintrinsicxarch.cpp

+// - isa is fully implemented or EnableIncompleteISAClass=true
+bool Compiler::compSupportsHWIntrinsic(InstructionSet isa)
+{
+    return (featureSIMD || isa == InstructionSet_BMI1 || isa == InstructionSet_BMI2 || isa == InstructionSet_LZCNT ||


Looks like you should add a:

bool Compiler::compIsScalarISA(InstructionSet isa) { return (isa == InstructionSet_BMI1) || (isa == InstructionSet_BMI2) || (isa == InstructionSet_LZCNT) || (isa == InstructionSet_POPCNT); }

Thanks, will change.

fiigii · 2017-12-20T01:51:58Z

@BruceForstall addressed feedback. PTAL.

danmoseley · 2017-12-20T03:16:23Z

FYI @eerhardt

fiigii · 2017-12-20T03:25:54Z

@dotnet-bot test Windows_NT x86 Checked Innerloop Build and Test

fiigii · 2017-12-21T00:30:10Z

@CarolEidt @BruceForstall ping?

fiigii force-pushed the 14930 branch from fd4ef4e to c41a62c Compare December 13, 2017 23:55

jkotas approved these changes Dec 14, 2017

View reviewed changes

CarolEidt suggested changes Dec 14, 2017

View reviewed changes

fiigii force-pushed the 14930 branch from c41a62c to b02a655 Compare December 15, 2017 22:08

CarolEidt suggested changes Dec 18, 2017

View reviewed changes

fiigii mentioned this pull request Dec 18, 2017

SIMD hardware intrinsics throw PNSE #15560

Merged

CarolEidt reviewed Dec 18, 2017

View reviewed changes

fiigii force-pushed the 14930 branch from b02a655 to 2639bb3 Compare December 19, 2017 19:22

fiigii mentioned this pull request Dec 20, 2017

Enable Vector256<T> on AVX CPUs #15528

Merged

BruceForstall suggested changes Dec 20, 2017

View reviewed changes

fiigii force-pushed the 14930 branch 2 times, most recently from f94a3c9 to 58ba214 Compare December 20, 2017 01:35

Add COMPlus_EnableIncompleteISAClass and streamline IsSupported

95c6e74

fiigii force-pushed the 14930 branch from 58ba214 to 95c6e74 Compare December 20, 2017 22:14

BruceForstall approved these changes Dec 21, 2017

View reviewed changes

BruceForstall merged commit 9891c8b into dotnet:master Dec 21, 2017

fiigii deleted the 14930 branch December 21, 2017 04:13

fiigii mentioned this pull request Jan 31, 2020

Need new CI modes for Intel hardware intrinsics dotnet/runtime#9464

Closed

4creators mentioned this pull request Jan 31, 2020

Enable AVX2, SSE42 HW intrinsics: remove unimplemented API surface and mark ISAs as fully implemented dotnet/runtime#25581

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

IsSupported returns false for not-fully-implemented ISA classes #15514

IsSupported returns false for not-fully-implemented ISA classes #15514

fiigii commented Dec 13, 2017

fiigii commented Dec 13, 2017

CarolEidt left a comment

CarolEidt Dec 14, 2017

CarolEidt Dec 14, 2017

fiigii Dec 14, 2017

CarolEidt Dec 14, 2017

fiigii commented Dec 14, 2017 •

edited

Loading

fiigii commented Dec 14, 2017

jkotas commented Dec 14, 2017

fiigii commented Dec 14, 2017

fiigii commented Dec 17, 2017

CarolEidt Dec 18, 2017

CarolEidt Dec 18, 2017

fiigii Dec 18, 2017

CarolEidt Dec 18, 2017

fiigii Dec 18, 2017

fiigii commented Dec 19, 2017

BruceForstall Dec 20, 2017

BruceForstall Dec 20, 2017

BruceForstall Dec 20, 2017

BruceForstall Dec 20, 2017

fiigii Dec 20, 2017

fiigii commented Dec 20, 2017

danmoseley commented Dec 20, 2017

fiigii commented Dec 20, 2017

fiigii commented Dec 21, 2017

IsSupported returns false for not-fully-implemented ISA classes #15514

IsSupported returns false for not-fully-implemented ISA classes #15514

Conversation

fiigii commented Dec 13, 2017

fiigii commented Dec 13, 2017

CarolEidt 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

fiigii commented Dec 14, 2017 • edited Loading

fiigii commented Dec 14, 2017

jkotas commented Dec 14, 2017

fiigii commented Dec 14, 2017

fiigii commented Dec 17, 2017

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

fiigii commented Dec 19, 2017

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

fiigii commented Dec 20, 2017

danmoseley commented Dec 20, 2017

fiigii commented Dec 20, 2017

fiigii commented Dec 21, 2017

fiigii commented Dec 14, 2017 •

edited

Loading