Convert Math{F}.CoreCLR methods from FCall to QCall #39474

am11 · 2020-07-16T20:59:49Z

TODOs:

decide which c-runtime functions to call directly
- CRT functions are directly registered to QCalls, without the wrappers.
how to handle /fp:fast vs. /fp:precise when p/invoking crt functions on Windows
- remove the wrappers and /fp:fast logic. we will use the default settings of the project (/fp:fast is always on).

src/coreclr/src/System.Private.CoreLib/src/System/Math.CoreCLR.cs

jkotas · 2020-07-17T13:51:55Z

Convert math intrinsics to named intrinsics

This looks like a signifiant change on its own. Would you mind submitting it in a separate PR to make it easier to review, test and get in?

src/coreclr/src/jit/codegenarmarch.cpp

src/coreclr/src/jit/gentree.h

src/coreclr/src/tools/Common/JitInterface/CorInfoImpl.cs

am11 · 2020-07-20T08:56:56Z

@jkotas, @tannergooding, in CLR ~~checked~~ debug x64 builds (all platforms), Single API MathF.Abs(-3.3f) is returning -3.3 instead of 3.3. Other build flavors do not have this issue. Also the double API Math.Abs(-3.3) does not has this issue in ~~checked~~ debug build. I couldn't figure out the reason. Is there any special handling for ~~Checked~~ Debug build which I should look into?

Edit: locally, I can only reproduce this issue with CLR's Debug build, but CI encounters in CLR Checked build.

# repros `MathF.Abs(-3.3f)` issue
./build.sh -c Debug

# do not produce the issue
./build.sh -rc Checked
./build.sh -c Release

src/coreclr/src/jit/importer.cpp

src/coreclr/src/zap/zapinfo.cpp

src/coreclr/src/tools/Common/JitInterface/CorInfoBase.cs

src/coreclr/src/System.Private.CoreLib/src/System/Math.CoreCLR.cs

jkotas · 2020-07-20T13:42:20Z

CLR Checked build.

The difference between checked and debug builds are optimizations settings. Debug is asserts on + optimizations off. Checked is assert on + optimizations on.

src/coreclr/src/vm/ecalllist.h

am11 · 2020-07-21T18:30:55Z

Remaining failures are related to 'Insert GC Polls', which seems to have recent changes (#39111).

ARM64 checked tests are failing this assertion:

runtime/src/coreclr/src/jit/flowgraph.cpp

Lines 2341 to 2342 in c5b39f3

    
           // After the DFS reverse postorder is completed, we must have visited all the basic blocks. 
        
           noway_assert(postIndex == fgBBcount + 1);

and x86 checked test is failing this one:

runtime/src/coreclr/src/jit/flowgraph.cpp

Lines 4142 to 4144 in c5b39f3

    
           // We are allowed to split loops and we need to keep a few other flags... 
        
           // 
        
           noway_assert((originalFlags & (BBF_SPLIT_NONEXIST & ~(BBF_LOOP_HEAD | BBF_LOOP_CALL0 | BBF_LOOP_CALL1))) == 0);

jkotas · 2020-07-21T18:36:08Z

Thanks! I have opened #39726 on this. This change will be blocked until it is fixed.

jkotas · 2020-07-21T18:39:10Z

Convert math intrinsics to named intrinsics

This looks like a signifiant change on its own. Would you mind submitting it in a separate PR to make it easier to review, test and get in?

Doing this in the meantime may be useful.

am11 · 2020-07-21T19:07:44Z

Submitted #39730. First and third commits are for this PR. Will rebase after JIT changes are merged.

erozenfeld · 2020-07-21T19:40:45Z

Which tests are failing with the JIT asserts?

am11 · 2020-07-21T19:47:43Z

@erozenfeld,

WinNT x86: https://helix.dot.net/api/2019-06-17/jobs/b855baea-b8a7-4cf9-b532-0e653190ec7c/workitems/JIT/console
WinNT arm64: https://helix.dot.net/api/2019-06-17/jobs/9639e468-8d03-4f7e-b173-5382b9cad8cf/workitems/JIT.HardwareIntrinsics/console
Linux arm64: https://helix.dot.net/api/2019-06-17/jobs/ac231456-0184-4011-8db7-b037ec59e37c/workitems/JIT.HardwareIntrinsics/console

It is the checked build per the CI categorization, but locally it might only repro with debug (./build.sh instead of ./build.sh -rc Checked). YMMV. :)

`fgComputeDoms` has an assumption that the flow graph has no unreachable blocks. It's checked with this assert: https://github.com/dotnet/runtime/blob/ad325b014124b1adb9306abf95fdac85e3f7f2f4/src/coreclr/src/jit/flowgraph.cpp#L2342 This assert fired when testing dotnet#39474 (`Convert Math{F}.CoreCLR methods from FCall to QCall`) when we are updating the flow graph after inserting GC polls. This change switches `fgUpdateChangedFlowGraph` to call `fgComputeReachability`, which both removes unreachable blocks and calls `fgComputeDoms`. pin-icount reported a 0.0043% throughput improvement, which is within noise level.

* Don't inline GC polls in cold basic blocks. * Allow GC poll inlining in basic blocks with `BBF_LOOP_PREHEADER` or `BBF_RETLESS_CALL` set. This fixes one of the assert seen in dotnet#39474 (comment) Contributes to resolving dotnet#39726.

erozenfeld · 2020-07-24T01:32:02Z

I opened #39878 and #39881, which will address the asserts in #39474 (comment)

jkotas · 2020-07-30T12:46:02Z

You should be able to build a benchmark using the same pattern as the GC latency benchmark for sorting from https://devblogs.microsoft.com/dotnet/performance-improvements-in-net-5/ that shows that algorithms that call Math operations a lot can stall GC for a long time.

The overhead of the GC poll should be very small once everything else works correctly. Also, this change should remove extra layer of indirection (the QCall wrapper) that will help to pay for this a bit of extra overhead.

I have looked what the JIT generates with the latest iteration of the change. I have used the same test as in #39474 (comment). It still does not seem to be working correctly. The PInvoke call is not getting inlined and there is an extra method frame that adds significant overhead. The problem is that the JIT recognizes the PInvoke as intrinsic and does not bother inlining it. This needs to be fixed to make this perform well.

jkotas · 2020-07-30T12:46:41Z

I will create a separate PR for mono and cleanup #if CORECLR from src/libraries.

You can get this refactoring PR going in parallel if you would like.

am11 · 2020-07-30T14:55:00Z

Thank you for sharing that link. I will try to compare the GC polls before and after these changes.

This non-QCall DllImport version is ready for use: am11@e8c4c31. It was the fastest among regressions per the benchmark. I can update the PR with it. However, it sill incurs the overhead due to P/Invoke+Intrinsic not getting inlined. Is there an issue tracking it? I can also try to investigate over the weekend.

The mono change was made yesterday in a separate branch, based off of this PR branch: https://github.com/am11/runtime/commits/feature/mathf-abs (top commit: Intrinsify MathF.Abs(float) instead of Math.Abs). I was thinking about creating that PR after these changes go in.

jkotas · 2020-07-30T16:39:18Z

It was the fastest among regressions per the benchmark.

Do you understand the reason why was the fastest? I suspect that it was just a measurement noise. I have stepped through the code on Windows and Linux (I do not have an easy access to mac at the moment) and the code was identical to the version with QCall. Also, the change is not right on Windows. msvcrt.dll is a legacy Windows C-runtime library that we do not want to use.

it still incurs the overhead due to P/Invoke+Intrinsic not getting inlined. Is there an issue tracking it?

There is no issue tracking it. This issue is exposed by this PR, there are no intrinsics that are DllImports today. If you can look into it, it would be great.

tannergooding · 2020-07-30T16:45:19Z

@AndyAyersMS might also have an idea if there is something that might block this in inlining today.

am11 · 2020-07-31T08:45:03Z

With the current state of this PR vs. master, i have found another difference in intrinsic recognition of Abs() flavors:

Method	Branch	Named Intrinsic {0}: Recognized
Math.Abs(double)	master	Yes, {0} -> `System.Math.Abs`
Math.Abs(double)	pr	Yes, {0} -> `System.Math.Abs`
Math.Abs(float)	master	Yes, {0} -> `System.Math.Abs`
Math.Abs(float)	pr	No
MathF.Abs(float)	master	Yes, {0} -> `System.Math.Abs`
MathF.Abs(float)	pr	Yes, {0} -> `System.MathF.Abs`

Perhaps the agressive-inlining on Math.Abs(float), which now points to [Intrinsic] MathF.Abs(float) is behaving unexpectedly.

jkotas · 2020-07-31T12:06:01Z

It may be best to do the Math.Abs refactoring in separate PR where it will be easier to validate that it is not regressing anything.

am11 · 2020-08-01T05:44:28Z

I took JIT dumps from master branch using COMPlus_JitDump=* COMPlus_JitPrintInlinedMethods=1, in order to set the expectation. With FCalls even, there is no math in the context of inlin* found in the dumps, although there are many other inline/inliner/inlining related infos found:

...
----------- Statements (and blocks) added due to the inlining of call [000293] -----------
Inlinee method body:fgInlineAppendStatements: no gc ref inline locals.
Successfully inlined Internal.Runtime.CompilerServices.Unsafe:As(byref):byref (2 IL bytes) (depth 2) [aggressive inline attribute]
INLINER: during 'fgInline' result 'success' reason 'aggressive inline attribute' for 'System.SpanHelpers:Contains(byref,ushort,int):bool' calling 'Internal.Runtime.CompilerServices.Unsafe:As(byref):byref'
INLINER: during 'fgInline' result 'success' reason 'aggressive inline attribute'
Inlines into 060013C6 System.SpanHelpers:Contains(byref,ushort,int):bool
  [1 IL=0007 TR=000005 06005898] [profitable inline] System.Diagnostics.Debug:Assert(bool)
...
INLINER: during 'fgInline' result 'failed this call site' reason 'unprofitable inline' for 'ILStubClass:IL_STUB_PInvoke(System.Runtime.InteropServices.SafeHandle,long,int):int' calling 'System.StubHelpers.StubHelpers:SafeHandleAddRef(System.Runtime.InteropServices.SafeHandle,byref):long'
INLINER: during 'fgInline' result 'failed this call site' reason 'unprofitable inline'
INLINER: inlineInfo.tokenLookupContextHandle for System.StubHelpers.StubHelpers:SafeHandleRelease(System.Runtime.InteropServices.SafeHandle) set to 0x0000000115FEFCE1:
Invoking compiler for the inlinee method System.StubHelpers.StubHelpers:SafeHandleRelease(System.Runtime.InteropServices.SafeHandle) :
INLINER impTokenLookupContextHandle for System.StubHelpers.StubHelpers:SafeHandleRelease(System.Runtime.InteropServices.SafeHandle) is 0x0000000115FEFCE1.
Inline candidate has an arg that feeds a constant test.  Multiplier increased to 2.
Native estimate for function size exceeds threshold for inlining 30.8 > 28 (multiplier = 3.3)
Inline expansion aborted, inline not profitable
INLINER: during 'fgInline' result 'failed this call site' reason 'unprofitable inline' for 'ILStubClass:IL_STUB_PInvoke(System.Runtime.InteropServices.SafeHandle,long,int):int' calling 'System.StubHelpers.StubHelpers:SafeHandleRelease(System.Runtime.InteropServices.SafeHandle)'
INLINER: during 'fgInline' result 'failed this call site' reason 'unprofitable inline'
Inlines into 06000000 ILStubClass:IL_STUB_PInvoke(System.Runtime.InteropServices.SafeHandle,long,int):int
...

am11 · 2020-08-07T14:22:54Z

Also, this change should remove extra layer of indirection (the QCall wrapper) that will help to pay for this a bit of extra overhead.

@jkotas, I have ran CI job with direct P/Invoke 3957e2d, but found strange issues during the test runs. For example,

this C program:

// ilogb-test.c
#include<math.h>
#include<stdio.h>

int main()
{
    printf("ilogb(NAN): %d\n", ilogb(NAN));
}

when compiled with clang -lm ilogb-test.c on Linux x64, results in ilogb(NAN): -2147483648.
when compiled with cl ilogb-test.c on Windows x64, results in ilogb(NAN): 2147483647.

Whereas, this .NET program in master (FCall) and QCall:

using System;

class Program
{
    static void Main(string[] args) => Console.WriteLine("Math.ILogB(Double.NaN): {0}", Math.ILogB(Double.NaN));
}

produces Math.ILogB(Double.NaN): 2147483647 on both Linux and Windows.
With direct P/Invoke, we get Math.ILogB(Double.NaN): -2147483648.

I am not sure how QCall/FCall mitigate this disparity. Perhaps some compiler flag?

(keeping QCall implementation for now)

am11 · 2020-08-07T14:39:14Z

Changes:

Revert Math.Abs refactoring.
Ignore SuppressGCTransitionAttribute for ref.
- Some recent change in master is forcing it (see APICompat failures on Linux #40416 e.g.). After the rebase with master, it started giving errors like:
  
  .packages/microsoft.dotnet.apicompat/5.0.0-beta.20381.6/build/Microsoft.DotNet.ApiCompat.targets(145,5): error : (NETCORE_ENGINEERING_TELEMETRY=Build) CannotRemoveAttribute : Attribute 'System.Runtime.InteropServices.SuppressGCTransitionAttribute' exists on 'System.Math.Abs(System.Double)' in the implementation but not the reference.

After manually comparing some of the pr vs. master control flows in JIT, I have found that importer needs the following patch to match code reachability of master in Compiler::impIntrinsic function:

--- a/src/coreclr/src/jit/importer.cpp
+++ b/src/coreclr/src/jit/importer.cpp
@@ -3608,7 +3608,7 @@ GenTree* Compiler::impIntrinsic(GenTree*                newobjThis,
     // If that call is an intrinsic and is expanded, codegen for NextCallReturnAddress will fail.
     // To avoid that we conservatively expand only required intrinsics in methods that call
     // the NextCallReturnAddress intrinsic.
-    if (!mustExpand && (opts.OptimizationDisabled() || info.compHasNextCallRetAddr))
+    if (!mustExpand && (opts.OptimizationDisabled() || info.compHasNextCallRetAddr) && (ni < NI_System_Math_FusedMultiplyAdd || ni > NI_System_Math_Floor))

I am not sure about the side-effects (just that it compiles and works on my machine 😅).

However, this still does not solve our lcl frame size regression problem:

-; rsp based frame
+; rbp based frame
 ; partially interruptible
 ; Final local variable assignments
 ;
-;  V00 arg0         [V00    ] (  5,  5   )  double  ->  [rsp+0x00]   do-not-enreg[X] addr-exposed ld-addr-op
+;  V00 arg0         [V00    ] (  5,  5   )  double  ->  [rbp-0x08]   do-not-enreg[X] addr-exposed ld-addr-op
 ;# V01 OutArgs      [V01    ] (  1,  1   )  lclBlk ( 0) [rsp+0x00]   "OutgoingArgSpace"
 ;
-; Lcl frame size = 8
+; Lcl frame size = 16

@jkotas, at this point, I think this will require help from someone more familiar with JIT.

src/coreclr/src/System.Private.CoreLib/src/System/Math.CoreCLR.cs

jkotas · 2020-08-10T02:35:26Z

I am not sure how QCall/FCall mitigate this disparity. Perhaps some compiler flag?

It is mitigated here:

runtime/src/coreclr/src/pal/src/cruntime/math.cpp

Line 308 in 6072e4d

PALIMPORT int __cdecl PAL_ilogb(double x)

jkotas · 2020-08-10T02:40:49Z

After manually comparing some of the pr vs. master control flows in JIT, I have found that importer needs the following patch to match code reachability of master

What is the problem that this is attempting to fix?

* Don't inline GC polls in cold basic blocks. * Allow GC poll inlining in basic blocks with `BBF_LOOP_PREHEADER` or `BBF_RETLESS_CALL` set. This fixes one of the assert seen in dotnet#39474 (comment) Contributes to resolving dotnet#39726.

`fgComputeDoms` has an assumption that the flow graph has no unreachable blocks. It's checked with this assert: https://github.com/dotnet/runtime/blob/ad325b014124b1adb9306abf95fdac85e3f7f2f4/src/coreclr/src/jit/flowgraph.cpp#L2342 This assert fired when testing dotnet#39474 (`Convert Math{F}.CoreCLR methods from FCall to QCall`) when we are updating the flow graph after inserting GC polls. This change switches `fgUpdateChangedFlowGraph` to call `fgComputeReachability`, which both removes unreachable blocks and calls `fgComputeDoms`. pin-icount reported a 0.0043% throughput improvement, which is within noise level.

am11 · 2020-08-10T07:34:06Z

What is the problem that this is attempting to fix?

I was tracing the code path due to math function calls in JIT with FCall vs. QCall. With QCall, it bails out earlier from

runtime/src/coreclr/src/jit/importer.cpp

Lines 3617 to 3620 in 4ec3a25

    
           if (!mustExpand && (opts.OptimizationDisabled() || info.compHasNextCallRetAddr)) 
        
           { 
        
               *pIntrinsicID = CORINFO_INTRINSIC_Illegal; 
        
               return retNode;

with FCall, it goes all the way to

runtime/src/coreclr/src/jit/importer.cpp

Lines 4285 to 4286 in 4ec3a25

    
           retNode = impMathIntrinsic(method, sig, callType, ni, tailCall); 
        
           break;

jkotas · 2020-08-11T00:13:15Z

I was tracing the code path due to math function calls in JIT with FCall vs. QCall. With QCall, it bails out earlier from

This sounds like you have been looking at Tier0 code or optimizations off. It is fine to inline this with optimizations off.

jkotas · 2020-08-11T00:31:00Z

an extra method frame that adds significant overhead.

There are several ways this can be fixed. I think that the best way to fix this is:

Add a new flag to GenTreeIntrinsic that says that the call will need gcpoll if the intrinsic is converted to one later in RewriteIntrinsicAsUserCall
The importer needs to create GenTreeIntrinsic with the actual unmanaged target for these math intrinsics
Compiler::fgInsertGCPolls needs to recognize GenTreeIntrinsic with the gcpoll flag set in addition to the regular call

Would you like to give it a shot?

jkotas · 2020-08-11T07:28:10Z

is there really an extra frame due to P/Invoke not getting inlined

Yes, there is an extra frame not getting inlined. Run this program and set a breakpoint at the cos method in libc / ucrtbase:

using System;

class Program
{
    static void Main() { for (;;) Work(); }

    static double Work()
    {
        double arg = 0;
        double sum = 0;
        for (int i = 0; i < 100000000; i++)
        {
            sum += Math.Cos(arg); arg += 1e-8;
        }
        return sum;
    }
}

The stacktrace when the cos method is called should look like this:

ucrtbase!cos 
System.Math.Cos(Double)+0x1b <- this is the extra frame that we need to get rid off
test!Program.Work()+0x27

am11 · 2020-08-11T15:13:43Z

Thanks. With libSOS, I can see that the call to Math.Cos is not present in case of FCall:

--- master	2020-08-11 17:30:15.000000000 +0300
+++ pr		2020-08-11 17:30:43.000000000 +0300
@@ -1,7 +1,7 @@
-(lldb) dumpstack
-OS Thread Id: 0x121b2fc (1)
+OS Thread Id: 0x11f94c5 (1)
 TEB information is not available so a stack size of 0xFFFF is assumed
 Current frame: libsystem_m.dylib!cos
 Child-SP         RetAddr          Caller, Callee
-00007FFEEFBFE6F0 000000012069595c (MethodDesc 0000000120731f60 + 0x5c Foo.Program.Work()), calling 0000000101936e40 (stub for System.Math.Cos(Double))
-00007FFEEFBFE6F8 0000000101a6e027 libcoreclr.dylib!MetaSig::Init(unsigned char const*, unsigned int, Module*, SigTypeContext const*, MetaSig::MetaSigKind) + 0x297, calling libcoreclr.dylib!SigParser::SkipExactlyOne()
+00007FFEEFBFE6D0 00000001210bd3fb (MethodDesc 00000001217a89c8 + 0x1b System.Math.Cos(Double))
+00007FFEEFBFE6F0 000000012169669c (MethodDesc 0000000121736168 + 0x5c Foo.Program.Work()), calling 0000000121696220 (stub for System.Math.Cos(Double))
+00007FFEEFBFE6F8 0000000102935bf7 libcoreclr.dylib!MetaSig::Init(unsigned char const*, unsigned int, Module*, SigTypeContext const*, MetaSig::MetaSigKind) + 0x297, calling libcoreclr.dylib!SigParser::SkipExactlyOne()

with this test, we never reach Compiler::impMathIntrinsic in importer, where GentTreeIntrinsic nodes are created (with both master and pr; tested with release and checked builds).

am11 · 2020-08-20T18:56:58Z

@jkotas, I tried implementing this approach #39474 (comment), but could not make it elide that libm call and the regression persists. I will defer to the JIT experts for the inlining part.

am11 · 2020-10-05T16:21:13Z

using System;

class Program
{
    static void Main() { for (;;) Work(); }

    static double Work()
    {
        double arg = 0;
        double sum = 0;
        for (int i = 0; i < 100000000; i++)
        {
            sum += Math.Cos(arg); arg += 1e-8;
        }
        return sum;
    }
}

The stacktrace when the cos method is called should look like this:

ucrtbase!cos 
System.Math.Cos(Double)+0x1b <- this is the extra frame that we need to get rid off
test!Program.Work()+0x27

With today's master branch (56950ad), I built and ran this program on macOS, that also breaks on frame #0: 0x00007fff789e0c1c libsystem_m.dylib`cos, and jit-diff only reports textual PMI diffs with master and rebased of this PR. If ucrtbase's cos is also hitting on Windows with master branch, then looks like perf. has regressed since?

Pref. results: https://gist.github.com/am11/65e492e856d81090530b03f0b8662f34 (vs. old results from July https://gist.github.com/am11/8dcbea6e3ca864585f14f888d10953e3)

Dotnet-GitSync-Bot added the area-Infrastructure-coreclr label Jul 16, 2020

jkotas reviewed Jul 16, 2020

View reviewed changes

src/coreclr/src/System.Private.CoreLib/src/System/Math.CoreCLR.cs Outdated Show resolved Hide resolved

jkotas added area-VM-coreclr and removed area-Infrastructure-coreclr labels Jul 16, 2020

jaredpar mentioned this pull request Jul 17, 2020

OSX machines are de-provisioned during CI / PR runs leading to failures #34472

Closed

tannergooding reviewed Jul 17, 2020

View reviewed changes

src/coreclr/src/jit/codegenarmarch.cpp Outdated Show resolved Hide resolved

tannergooding reviewed Jul 17, 2020

View reviewed changes

src/coreclr/src/jit/gentree.h Outdated Show resolved Hide resolved

am11 commented Jul 17, 2020

View reviewed changes

src/coreclr/src/tools/Common/JitInterface/CorInfoImpl.cs Outdated Show resolved Hide resolved

jkotas reviewed Jul 20, 2020

View reviewed changes

src/coreclr/src/jit/importer.cpp Outdated Show resolved Hide resolved

jkotas reviewed Jul 20, 2020

View reviewed changes

src/coreclr/src/zap/zapinfo.cpp Outdated Show resolved Hide resolved

jkotas reviewed Jul 20, 2020

View reviewed changes

src/coreclr/src/tools/Common/JitInterface/CorInfoBase.cs Outdated Show resolved Hide resolved

jkotas reviewed Jul 20, 2020

View reviewed changes

src/coreclr/src/System.Private.CoreLib/src/System/Math.CoreCLR.cs Show resolved Hide resolved

jkotas reviewed Jul 21, 2020

View reviewed changes

src/coreclr/src/vm/ecalllist.h Outdated Show resolved Hide resolved

jkotas mentioned this pull request Jul 21, 2020

SuppressGCTransitionAttribute hitting asserts in the JIT #39726

Closed

jkotas added the blocked Issue/PR is blocked on something - see comments label Jul 21, 2020

am11 mentioned this pull request Jul 21, 2020

Convert math intrinsics to named intrinsics #39730

Merged

erozenfeld mentioned this pull request Jul 24, 2020

Fix fgUpdateChangedFlowGraph. #39878

Merged

erozenfeld mentioned this pull request Jul 24, 2020

Fix GC Poll inlining. #39881

Merged

danmoseley mentioned this pull request Aug 7, 2020

Improving P95+ latency #37534

Closed

am11 commented Aug 7, 2020

View reviewed changes

src/coreclr/src/System.Private.CoreLib/src/System/Math.CoreCLR.cs Show resolved Hide resolved

am11 marked this pull request as draft August 20, 2020 18:01

Convert Math{F}.CoreCLR methods from FCall to QCall

c773c75

am11 mentioned this pull request Nov 26, 2020

Convert several FCalls to QCalls, plus misc cleanup #45258

Merged

am11 closed this Nov 27, 2020

am11 deleted the feature/fcalls-to-qcalls branch November 27, 2020 01:02

jkotas mentioned this pull request Nov 27, 2020

Calling System.Math floating point operations in a loop causes long GC pause times #13820

Open

ghost locked as resolved and limited conversation to collaborators Dec 27, 2020

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Convert Math{F}.CoreCLR methods from FCall to QCall #39474

Convert Math{F}.CoreCLR methods from FCall to QCall #39474

am11 commented Jul 16, 2020 •

edited

Loading

jkotas commented Jul 17, 2020

am11 commented Jul 20, 2020 •

edited

Loading

jkotas commented Jul 20, 2020

am11 commented Jul 21, 2020

jkotas commented Jul 21, 2020

jkotas commented Jul 21, 2020

am11 commented Jul 21, 2020

erozenfeld commented Jul 21, 2020

am11 commented Jul 21, 2020

erozenfeld commented Jul 24, 2020

jkotas commented Jul 30, 2020

jkotas commented Jul 30, 2020

am11 commented Jul 30, 2020

jkotas commented Jul 30, 2020

tannergooding commented Jul 30, 2020

am11 commented Jul 31, 2020

jkotas commented Jul 31, 2020

am11 commented Aug 1, 2020

am11 commented Aug 7, 2020 •

edited

Loading

am11 commented Aug 7, 2020

jkotas commented Aug 10, 2020

jkotas commented Aug 10, 2020

am11 commented Aug 10, 2020

jkotas commented Aug 11, 2020

jkotas commented Aug 11, 2020

jkotas commented Aug 11, 2020

am11 commented Aug 11, 2020

am11 commented Aug 20, 2020

am11 commented Oct 5, 2020

Convert Math{F}.CoreCLR methods from FCall to QCall #39474

Convert Math{F}.CoreCLR methods from FCall to QCall #39474

Conversation

am11 commented Jul 16, 2020 • edited Loading

jkotas commented Jul 17, 2020

am11 commented Jul 20, 2020 • edited Loading

jkotas commented Jul 20, 2020

am11 commented Jul 21, 2020

jkotas commented Jul 21, 2020

jkotas commented Jul 21, 2020

am11 commented Jul 21, 2020

erozenfeld commented Jul 21, 2020

am11 commented Jul 21, 2020

erozenfeld commented Jul 24, 2020

jkotas commented Jul 30, 2020

jkotas commented Jul 30, 2020

am11 commented Jul 30, 2020

jkotas commented Jul 30, 2020

tannergooding commented Jul 30, 2020

am11 commented Jul 31, 2020

jkotas commented Jul 31, 2020

am11 commented Aug 1, 2020

am11 commented Aug 7, 2020 • edited Loading

am11 commented Aug 7, 2020

jkotas commented Aug 10, 2020

jkotas commented Aug 10, 2020

am11 commented Aug 10, 2020

jkotas commented Aug 11, 2020

jkotas commented Aug 11, 2020

jkotas commented Aug 11, 2020

am11 commented Aug 11, 2020

am11 commented Aug 20, 2020

am11 commented Oct 5, 2020

am11 commented Jul 16, 2020 •

edited

Loading

am11 commented Jul 20, 2020 •

edited

Loading

am11 commented Aug 7, 2020 •

edited

Loading