[PATCH] D32971: [scudo] CRC32 optimizations

[Kostya Kortchinsky via Phabricator via llvm-commits]

cryptoad updated this revision to Diff 98207.
cryptoad added a comment.

This new solution should fit both Dmitry's and Aleksey's suggestions.

Wrap up the checksum logic in a function that takes as parameters all the
information needed to compute the checksum, and returns the checksum using
the fastest way available.

Regarding the assembly, the full SSE one looks like (no extra `HashAlgorithm`
load and inlined `crc32` calls):

  mov     rax, 0FFFFFFFFFFFFFFh
  lea     rcx, [r13-10h]
  shl     r10, 38h
  and     r14, rax
  mov     rax, cs:_ZN7__scudoL6CookieE ; __scudo::Cookie
  or      r14, r10
  mov     rdx, r14
  crc32   eax, rcx
  xor     dx, dx
  crc32   eax, rdx
  mov     r14w, ax
  mov     rax, r13
  mov     [r13-10h], r14

And the partial SSE one (only one `HashAlgorithm` load):

                  mov     rax, cs:_ZN7__scudoL6CookieE ; __scudo::Cookie
                  mov     dl, cs:_ZN7__scudoL13HashAlgorithmE ; unsigned __int64
                  cmp     dl, 1
                  jnz     short loc_8859
                  and     rbx, 0FFFFFFFFFFFF0000h
                  mov     edi, eax        ; this
                  mov     rsi, rcx        ; unsigned int
                  call    _ZN7__scudo20computeHardwareCRC32Ejm ; __scudo::computeHardwareCRC32(uint,ulong)
                  mov     edi, eax        ; this
                  mov     rsi, rbx        ; unsigned int
                  call    _ZN7__scudo20computeHardwareCRC32Ejm ; __scudo::computeHardwareCRC32(uint,ulong)
                  jmp     loc_8974
  ; ---------------------------------------------------------------------------
  
  loc_8859:                               ; CODE XREF: __scudo::ScudoChunk::computeChecksum(__scudo::UnpackedHeader *)+17j
  
  ... software CRC32 ...

So this seems pretty good.


https://reviews.llvm.org/D32971

Files:
  lib/scudo/scudo_allocator.cpp
  lib/scudo/scudo_crc32.cpp
  lib/scudo/scudo_crc32.h
  lib/scudo/scudo_utils.h

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D32971.98207.patch
Type: text/x-patch
Size: 13635 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170508/4733df43/attachment.bin>