Still not super-confident I understand why the newer version is slower. Inner loops look pretty comparable with one exception. I think this is the clone version of the first deeply nested loop:

                        for (k = 0; k < i; k++)
                            sum -= a[i][k] * a[k][j];

In the after version the loop body is disconnected and there is a spill restore from [ebp-38H]. So this may be why the test is now running slower.

;;; Before

G_M19375_IG39:
       8B5D88       mov      ebx, gword ptr [ebp-78H]
       8BFB         mov      edi, ebx
       C4E17B1054D708 vmovsd   xmm2, qword ptr [edi+8*edx+8]
       8B7C9108     mov      edi, gword ptr [ecx+4*edx+8]
       3B7704       cmp      esi, dword ptr [edi+4]
       0F8342080000 jae      G_M19375_IG101
       C4E16B5954F708 vmulsd   xmm2, qword ptr [edi+8*esi+8]
       C4E1735CCA   vsubsd   xmm1, xmm2
       42           inc      edx
       3BD6         cmp      edx, esi
       895D88       mov      gword ptr [ebp-78H], ebx
       7CD3         jl       SHORT G_M19375_IG39

;;; AFTER

G_M19375_IG37:
       8B5D90       mov      ebx, gword ptr [ebp-70H]
       8BCB         mov      ecx, ebx
       C4E17B1054D108 vmovsd   xmm2, qword ptr [ecx+8*edx+8]
       8B4DC8       mov      ecx, gword ptr [ebp-38H]
       8B449108     mov      eax, gword ptr [ecx+4*edx+8]
       3B7004       cmp      esi, dword ptr [eax+4]
       0F8346080000 jae      G_M19375_IG102
       C4E16B5954F008 vmulsd   xmm2, qword ptr [eax+8*esi+8]
       C4E1735CCA   vsubsd   xmm1, xmm2
       42           inc      edx
       3BD6         cmp      edx, esi
       895D90       mov      gword ptr [ebp-70H], ebx
       7C0B         jl       SHORT G_M19375_IG38

       8B5DEC       mov      ebx, dword ptr [ebp-14H]
       897DD0       mov      dword ptr [ebp-30H], edi
       8B45CC       mov      eax, dword ptr [ebp-34H]
       EB7B         jmp      SHORT G_M19375_IG48

G_M19375_IG38:
       894DC8       mov      gword ptr [ebp-38H], ecx
       EBC0         jmp      SHORT G_M19375_IG37

This method is probably a good stress test for spill placement as a whole, as there are large runs of blocks to spill and reload. So linking in dotnet/coreclr#16857. Likely cloning plays a big role in creating lots of pressure and backegdes.

G_M19375_IG40:
       8B5D88       mov      ebx, gword ptr [ebp-78H]
       EB2F         jmp      SHORT G_M19375_IG48

G_M19375_IG41:
       8B5D88       mov      ebx, gword ptr [ebp-78H]
       EB2A         jmp      SHORT G_M19375_IG48

G_M19375_IG42:
       895D88       mov      gword ptr [ebp-78H], ebx
       EB5E         jmp      SHORT G_M19375_IG49

G_M19375_IG43:
       895D88       mov      gword ptr [ebp-78H], ebx
       EB59         jmp      SHORT G_M19375_IG49

G_M19375_IG44:
       8B55F0       mov      edx, dword ptr [ebp-10H]
       E972010000   jmp      G_M19375_IG56

G_M19375_IG45:
       8B5DC0       mov      ebx, gword ptr [ebp-40H]
       E9C0FEFFFF   jmp      G_M19375_IG35

G_M19375_IG46:
       895DC0       mov      gword ptr [ebp-40H], ebx
       E9F5FEFFFF   jmp      G_M19375_IG36

G_M19375_IG47:
       895DC0       mov      gword ptr [ebp-40H], ebx
       E9EDFEFFFF   jmp      G_M19375_IG36

So for the immediate 2.1 vs 2.0 issue I don't see anything fixable in the 2.1 timeframe.

I am still going to try and understand the larger regression that both these branches saw a while back...

[Perf][Windows_NT][x86] Investigate ByteMark/BenchLUDecomp regression. #9833

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions