[NET9] [REGRESSION] Gmail API fail with TieredCompilation disabled

[kirsan31]

Description

Google.Apis.Gmail.v1 call to GmailService.Users.GetProfile("someId") fail with strange exception:

Google.GoogleApiException: Parameter validation failed for "userId" : The parameter value must be a non-empty string

Only on .NET9 + <TieredCompilation>false</TieredCompilation> (.NET8 in any combinations and .NET9 + <TieredCompilation>true</TieredCompilation> works normally).

Reproduction Steps

• Build repro app.
• Run .Net9 binary to get an exception and .Net8 to get normal result.

GmailGetProfile.zip

Expected behavior

Must work without Parameter validation failed exception. Repro app must output "Unauthorized" in console.

Actual behavior

Unhandled exception. The service gmail has thrown an exception.
No HttpStatusCode was specified.
No error details were specified.
Google.GoogleApiException: Parameter validation failed for "userId" : The parameter value must be a non-empty string
   at Google.Apis.Requests.ClientServiceRequest`1.AddParameters(RequestBuilder requestBuilder, ParameterCollection inputParameters)
   at Google.Apis.Requests.ClientServiceRequest`1.CreateBuilder()
   at Google.Apis.Requests.ClientServiceRequest`1.CreateRequest(Nullable`1 overrideGZipEnabled)
   at Google.Apis.Requests.ClientServiceRequest`1.ExecuteUnparsedAsync(CancellationToken cancellationToken)
   at Google.Apis.Requests.ClientServiceRequest`1.ExecuteAsync(CancellationToken cancellationToken)
   at GmailGetProfile.Program.Main(String[] args) in ...GmailGetProfile\Program.cs:line 16
   at GmailGetProfile.Program.<Main>(String[] args)

Regression?

Yes it's works in .NET8.

Known Workarounds

No.

Configuration

SDK: 9.0.100-rc.2.24474.11
Runtime: 9.0.0-rc.2.24474.4
OS: Win11 23H2

Other information

No response

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[EgorBo]

cc @dotnet/jit-contrib (I wasn't able to repro it on macos-arm64/x64, don't have any Windows to test it there).

[AndyAyersMS]

I can look later today. If anyone else wants to dig in before then, just let me know.

[hez2010]

Can reproduce on win-x64.

Didn't investigate too far, to summarize the findings so far:

The issue seems to happen in ParameterCollection.FromDictionary.

https://github.com/googleapis/google-api-dotnet-client/blob/6246c9d65cc5a4c4bbaf5fa9524daa96ef1af650/Src/Support/Google.Apis.Core/Requests/Parameters/ParameterCollection.cs#L143

On .NET 8:

the dictionary is

and ParameterCollection.FromDictionary yields

While on .NET 9 the dictionary is the same but ParameterCollection.FromDictionary yields the following result:

codegen diff (net8.0 - net9.0): https://www.diffchecker.com/JbWCBrJw/

[kunalspathak]

shorter repro that has less dependencies:

using Google;
using Google.Apis.Gmail.v1;
using Google.Apis.Requests.Parameters;
using Google.Apis.Util;
using System.Runtime.CompilerServices;

namespace GmailGetProfile;

internal class Program
{
    static async Task Main(string[] args)
    {
        using var service = new GmailService();
        using var cts = new CancellationTokenSource();
        cts.CancelAfter(5000);
        var getProfile = service.Users.GetProfile("test");
        FromMyDictionary(ParameterUtils.CreateParameterDictionary(getProfile));
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    public static void Consume(string a) { Console.WriteLine(a); }

    public static ParameterCollection FromMyDictionary(IDictionary<string, object> dictionary)
    {
        ParameterCollection parameterCollection = new ParameterCollection();
        foreach (KeyValuePair<string, object> item in dictionary)
        {
            IEnumerable<KeyValuePair<string, object>> enumerable = item.Value as IEnumerable<KeyValuePair<string, object>>;
            if (!(item.Value is string) && enumerable != null)
            {
                Consume(item.Key);
            }
            else
            {
                Consume($"else: {item.Key}");
            }
        }

        return parameterCollection;
    }
}

[AndyAyersMS]

Works with DOTNET_JitEnablePhysicalPromotion=0, so possibly related to promotion. FYI @jakobbotsch

[AndyAyersMS]

@hez2010 looks like you implicated the right method

 255: JIT compiled Google.Apis.Requests.Parameters.ParameterCollection:FromDictionary(System.Collections.Generic.IDictionary`2[System.String,System.Object]):Google.Apis.Requests.Parameters.ParameterCollection [FullOpts, IL size=188, code size=715, hash=0xb412e2dc]

set DOTNET_JitEnablePhysicalPromotionRange=b412e2dc
set DOTNET_TieredCompilation=0
(run)
userId : test
userId :
userId :
userId :

set DOTNET_JitEnablePhysicalPromotionRange=0-b412e2db,b412e2dd-ffffffff
set DOTNET_TieredCompilation=0
(run)
userId : test
$.xgafv :
alt :
prettyPrint :

[AndyAyersMS]

We promote a KVP:

Picking promotions
Accesses for V03
  [000..016) as System.Collections.Generic.KeyValuePair`2[System.String,System.Object]
lvaGrabTemp returning 36 (V36 tmp28) (a long lifetime temp) called for V03.[000..008).
lvaGrabTemp returning 37 (V37 tmp29) (a long lifetime temp) called for V03.[008..016).

Looks like the read back (initialization) of V36 happens in a nested QMARK but the uses are unconditional, so we end up with uninitialized reads later on?

Reading back pending replacements before tree with possible exception side effect inside block in try region
  V03.[000..008) -> V36
New statement:
STMT00010 ( ??? ... ??? )
               [000046] DA-X-------                         *  STORE_LCL_VAR int    V11 tmp3         
               [000045] -A-X-------                         \--*  QMARK     int   
               [000036] -----------    if                      +--*  EQ        int   
               [000035] -----------                            |  +--*  LCL_VAR   ref    V10 tmp2         
               [000034] -----------                            |  \--*  CNS_INT   ref    null
               [000044] -A-X-------    if                      \--*  COLON     int   
               [000042] -A-X------- else                          +--*  QMARK     int   
               [000038] -A-X-------    if                         |  +--*  NE        int   
               [000318] -A-X-------                               |  |  +--*  COMMA     long  
               [000317] DA---------                               |  |  |  +--*  STORE_LCL_VAR ref    V36 tmp28        
               [000316] -----------                               |  |  |  |  \--*  LCL_FLD   ref   (RB) V03 loc2         [+0]
               [000037] #--X-------                               |  |  |  \--*  IND       long  
               [000032] -----------                               |  |  |     \--*  LCL_VAR   ref    V10 tmp2          (last use)

Seems like a tricky interaction of promotion, QMARK, and EH. Would guess we need to move the readback even earlier, likely in a preceding statement, or else duplicate it on the other conditional paths.

@jakobbotsch I will defer to you on this one.

Note this was with Kunal's first repro.