Re: [dpdk-stable] AVX512 bug on SkyLake

[Thomas Monjalon <thomas@monjalon.net>]

11/11/2018 15:15, Ananyev, Konstantin:
> Hi Thomas,
> 
> > Below is my conclusion for this bug.
> > An expert of x86 is required to follow-up.
> > 
> > Summary:
> > 	- CPU: Intel Skylake
> > 	- Linux environment: Ubuntu 18.04
> > 	- Compiler: GCC 7 or 8
> > 	- Scenario: testpmd crashes when it starts forwarding
> > 	- Behaviour: AVX2 version of rte_memcpy() fails if optimized for AVX512
> > 	- Context: inline rte_memcpy() is called from
> > 			inline rte_mempool_put_bulk(), called from
> > 			mlx5_tx_complete() (inline or not)
> > 	- Analysis: AVX512 optimization changes vmovdqu to vmovdqu8
> > 
> > Latest status can be found in Bugzilla:
> > 	https://bugs.dpdk.org/show_bug.cgi?id=97#c35
> 
> 
> Looking at dissamled output from the bug report, it seems that the
> problem is not in vmovdqu8 instruction itself, but in the wrong offsets
> generated by the compiler:
> 
>    vmovdqu8 xmm0,XMMWORD PTR [rax*8+0x2]
>    vinserti128 ymm0,ymm0,XMMWORD PTR [rax*8+0x30],0x1
>     vmovups XMMWORD PTR [rsi+0x20],xmm0
>     vextracti128 XMMWORD PTR [rsi+0x30],ymm0,0x1
>     vmovdqu8 xmm0,XMMWORD PTR [rax*8+0x4]
>     vinserti128 ymm0,ymm0,XMMWORD PTR [rax*8+0x50],0x1
>     vmovups XMMWORD PTR [rsi+0x40],xmm0
>     vextracti128 XMMWORD PTR [rsi+0x50],ymm0,0x1
>     vmovdqu8 xmm0,XMMWORD PTR [rax*8+0x6]
> 
> Should be:
> vmovdqu8 xmm0,XMMWORD PTR [rax*8+0x20]
> I think.
> 
> Same for next two offsets: 0x4 and 0x6 respectively should be 0x40 and 0x60.

Yes, you're right, I missed it, thank you!

The full diff is below:

--- bad-avx512-enabled
+++ good-avx512-disabled
-    vmovdqu8 xmm0,XMMWORD PTR [rax*8+0x0]
+    vmovdqu xmm0,XMMWORD PTR [rax*8+0x0]
     vinserti128 ymm0,ymm0,XMMWORD PTR [rax*8+0x10],0x1
     vmovups XMMWORD PTR [rsi],xmm0
     vextracti128 XMMWORD PTR [rsi+0x10],ymm0,0x1
-    vmovdqu8 xmm0,XMMWORD PTR [rax*8+0x2]
+    vmovdqu xmm0,XMMWORD PTR [rax*8+0x20]
     vinserti128 ymm0,ymm0,XMMWORD PTR [rax*8+0x30],0x1
     vmovups XMMWORD PTR [rsi+0x20],xmm0
     vextracti128 XMMWORD PTR [rsi+0x30],ymm0,0x1
-    vmovdqu8 xmm0,XMMWORD PTR [rax*8+0x4]
+    vmovdqu xmm0,XMMWORD PTR [rax*8+0x40]
     vinserti128 ymm0,ymm0,XMMWORD PTR [rax*8+0x50],0x1
     vmovups XMMWORD PTR [rsi+0x40],xmm0
     vextracti128 XMMWORD PTR [rsi+0x50],ymm0,0x1
-    vmovdqu8 xmm0,XMMWORD PTR [rax*8+0x6]
+    vmovdqu xmm0,XMMWORD PTR [rax*8+0x60]
     vinserti128 ymm0,ymm0,XMMWORD PTR [rax*8+0x70],0x1
     vmovups XMMWORD PTR [rsi+0x60],xmm0
     vextracti128 XMMWORD PTR [rsi+0x70],ymm0,0x1

> Not sure what causing compiler behaves that way.
> BTW, looking though testpmd objdump output - it seems that only mlx5 driver
> exhibits such problem (I didn't enable mlx4 actually, probably same problem here).
> Which looks a bit weird to me.

Yes it's weird. I don't see how the mlx5 code could influence
the compiler to generate this bad code in AVX512 mode.