Incorrect value after truncating cast

[TechPizzaDev]

Description

When running a Release build, performing a (ushort)(uint) cast on int and storing the result in a ushort* seems to write out 32 bits instead of 16 bits.

Reproduction Steps

1. Clone repo https://github.com/TechPizzaDev/SoftwareRasterizer
2. Checkout branch net8-upgrade
3. Download Castle folder from https://github.com/rawrunprotected/rasterizer/tree/master/SoftwareRasterizer and put next to built SoftwareRasterizer.exe
4. Run project in Release mode to see broken render (it works in Debug mode)
5. Remove BUG define (https://github.com/TechPizzaDev/SoftwareRasterizer/blob/318adc376bcdaee1b6c863dd9384871aed82897e/SoftwareRasterizer/Avx2Rasterizer.cs#L2)
6. Run project again to see correct render

Expected behavior

Correct raster like in net7:

Actual behavior

Regression?

Regression in net8, worked on net7 and net6.

Known Workarounds

Rewriting the code seems like the only workaround. The project also works in Debug mode.

Configuration

.NET: 8.0.0-rc.1.23419.4
OS: Microsoft Windows 11 Home (10.0.22621 version 22621)
CPU: 11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz
Arch: x64

Env variables

DOTNET_EnableAVX512F=0: Required since net8 seems to cause unexpected AccessViolationExceptions.

Other information

Sse41.MinHorizontal(Vector128<ushort>) returns a Vector128<ushort> where the first ushort element is the minimum value from the vector, and the second ushort element is the index of the min element within the vector.

Problematic code: https://github.com/TechPizzaDev/SoftwareRasterizer/blob/318adc376bcdaee1b6c863dd9384871aed82897e/SoftwareRasterizer/Avx2Rasterizer.cs#L1262-L1264

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

Issue Details

---

Description

Performing a (ushort)(uint) cast on int and storing the result in a ushort* seems to write out 32 bits instead of 16 bits.

Reproduction Steps

1. Clone repo https://github.com/TechPizzaDev/SoftwareRasterizer
2. Checkout branch net8-upgrade
3. Run project to see broken render
4. Remove BUG define (https://github.com/TechPizzaDev/SoftwareRasterizer/blob/318adc376bcdaee1b6c863dd9384871aed82897e/SoftwareRasterizer/Avx2Rasterizer.cs#L2)
5. Run project again to see correct render

Expected behavior

Correct raster like in net7:

Actual behavior

Regression?

Worked on net7.0 and net6.0

Known Workarounds

Rewriting the code seems like the only workaround

Configuration

.NET: 8.0.0-rc.1.23419.4
OS: Microsoft Windows 11 Home (10.0.22621 version 22621)
CPU: 11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz
Arch: x64

Other information

Sse41.MinHorizontal(Vector128<ushort>) returns a Vector128<ushort> where the first ushort element is the minimum value from the vector, and the second ushort element is the index of the min element within the vector.

Author:	TechPizzaDev
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[TechPizzaDev]

Disassembly of suspected code:

   vphminposuw xmm11, xmm11
-  vmovd    r11d, xmm11
-  movzx    rax, r11w
-  mov      word  ptr [r13], ax
+  vmovd    dword ptr [r13], xmm11
   mov      qword ptr [rsp+0x6D0], r10

Seems as if 32 instead of 16 bits are moved, judging by that word ptr to dword ptr diff.

correct jitdisasm.txt
bug jitdisasm.txt

[AndyAyersMS]

DOTNET_EnableAVX512F=0: Required since net8 seems to cause unexpected AccessViolationExceptions.

Are you saying there are two problems?

• fails with AV if AVX-512 enabled
• wrong behavior if AVX-512 disabled

[TechPizzaDev]

I have not gotten to what causes access violations on AVX512, but DOTNET_EnableAVX512=1 with fixed code still AVs, so I am sure these are two different problems.

[AndyAyersMS]

@dotnet/jit-contrib FYI

[JulieLeeMSFT]

@TIHan, PTAL.

[TIHan]

Was actually just looking at it :) should be interesting

[TIHan]

I can confirm that that the suspected disassembly is cause of the issue.

Looking at a dump of a similar example, we remove the CAST int <- ushort <- int:

fgMorphTree BB01, STMT00002 (before)
               [000029] -A-XG------                         *  STOREIND  short 
               [000025] -----------                         +--*  LCL_VAR   long   V00 arg0          (last use)
               [000028] -----------                         \--*  CAST      int <- ushort <- int
               [000027] -----------                            \--*  HWINTRINSIC int    int ConvertToInt32
               [000023] -----------                               \--*  HWINTRINSIC simd16 ushort MinHorizontal
               [000022] -----------                                  \--*  HWINTRINSIC simd16 ushort Min
               [000019] -----------                                     +--*  HWINTRINSIC simd16 ushort GetLower
               [000018] -----------                                     |  \--*  LCL_VAR   simd32<System.Runtime.Intrinsics.Vector256`1> V01 loc0         
               [000021] -----------                                     \--*  HWINTRINSIC simd16 ushort GetUpper
               [000020] -----------                                        \--*  LCL_VAR   simd32<System.Runtime.Intrinsics.Vector256`1> V01 loc0          (last use)

fgMorphTree BB01, STMT00002 (after)
               [000029] -A-XG+-----                         *  STOREIND  short 
               [000025] -----+-----                         +--*  LCL_VAR   long   V00 arg0          (last use)
               [000027] -----+-----                         \--*  HWINTRINSIC int    int ConvertToInt32
               [000023] -----+-----                            \--*  HWINTRINSIC simd16 ushort MinHorizontal
               [000022] -----+-----                               \--*  HWINTRINSIC simd16 ushort Min
               [000019] -----+-----                                  +--*  HWINTRINSIC simd16 ushort GetLower
               [000018] -----+-----                                  |  \--*  LCL_VAR   simd32<System.Runtime.Intrinsics.Vector256`1> V01 loc0         
               [000021] -----+-----                                  \--*  HWINTRINSIC simd16 ushort GetUpper
               [000020] -----+-----                                     \--*  LCL_VAR   simd32<System.Runtime.Intrinsics.Vector256`1> V01 loc0          (last use)

However, it isn't clear if the CAST removal itself is the main issue. We may not be handling something correctly later. Such as, we are not respecting the short of STOREIND in codegen.