[libcxx-commits] [PATCH] D73291: Inline basic_string::assign(s [, n]) for compile time known short strings

[Martijn Vels via Phabricator via libcxx-commits]

mvels created this revision.
mvels added a reviewer: EricWF.
Herald added subscribers: libcxx-commits, christof.
Herald added a project: libc++.

This change uses __has_builtin_constant_p if supported to check for compile time known short inputs, and directly inlines the assignment. The compiler will elide the outlined or inlined code as needed.

Outlined calls have been placed in __assign_external(...) calls. Combined with ABI changes for unstable (inline assign, outline assign_external), this results in considerable speed ups.

Before

  -----------------------------------------------------------------------------------
  Benchmark                                        Time             CPU   Iterations
  -----------------------------------------------------------------------------------
  BM_StringAssignAsciiz_Empty_Opaque            6.83 ns         6.83 ns    100196352
  BM_StringAssignAsciiz_Empty_Transparent       6.57 ns         6.58 ns    105873408
  BM_StringAssignAsciiz_Small_Opaque            8.47 ns         8.47 ns     82526208
  BM_StringAssignAsciiz_Small_Transparent       8.25 ns         8.25 ns     84996096
  BM_StringAssignAsciiz_Large_Opaque            19.0 ns         19.0 ns     36720640
  BM_StringAssignAsciiz_Large_Transparent       18.5 ns         18.5 ns     37986304
  BM_StringAssignAsciiz_Huge_Opaque             1683 ns         1683 ns       401408
  BM_StringAssignAsciiz_Huge_Transparent        1680 ns         1679 ns       397312
  BM_StringAssignAsciizMix_Opaque               11.0 ns         11.0 ns     63692800
  BM_StringAssignAsciizMix_Transparent          11.2 ns         11.1 ns     62808064

After with ABI change

  -----------------------------------------------------------------------------------
  Benchmark                                        Time             CPU   Iterations
  -----------------------------------------------------------------------------------
  BM_StringAssignAsciiz_Empty_Opaque            6.74 ns         6.75 ns    100999168
  BM_StringAssignAsciiz_Empty_Transparent      0.863 ns        0.864 ns    809365504
  BM_StringAssignAsciiz_Small_Opaque            8.53 ns         8.54 ns     81907712
  BM_StringAssignAsciiz_Small_Transparent       1.15 ns         1.15 ns    605679616
  BM_StringAssignAsciiz_Large_Opaque            18.9 ns         18.9 ns     37072896
  BM_StringAssignAsciiz_Large_Transparent       18.6 ns         18.6 ns     37535744
  BM_StringAssignAsciiz_Huge_Opaque             1687 ns         1687 ns       405504
  BM_StringAssignAsciiz_Huge_Transparent        1690 ns         1690 ns       409600
  BM_StringAssignAsciizMix_Opaque               10.9 ns         10.9 ns     64319488
  BM_StringAssignAsciizMix_Transparent          5.31 ns         5.31 ns    131194880

Inlined code for assign empty string

  void AssignEmpty(std::string*p) {
      p->assign("");
  }
  
  AssignEmpty:
          test    byte ptr [rdi], 1
          jne     .LBB1_2
          mov     byte ptr [rdi], 0
          add     rdi, 1
          mov     byte ptr [rdi], 0
          ret
  .LBB1_2:
          mov     rax, qword ptr [rdi + 16]
          mov     qword ptr [rdi + 8], 0
          mov     byte ptr [rax], 0
          ret

Inlined code for assign short string

  void Assign(std::string*p) {
      p->assign("Hello world");
  }
  
  Assign:
          test    byte ptr [rdi], 1
          jne     .LBB0_2
          mov     byte ptr [rdi], 22
          add     rdi, 1
          jmp     .LBB0_3
  .LBB0_2:
          mov     rax, qword ptr [rdi + 16]
          mov     qword ptr [rdi + 8], 11
          mov     rdi, rax
  .LBB0_3:
          movabs  rax, 8031924123371070792
          mov     qword ptr [rdi], rax
          mov     dword ptr [rdi + 7], 1684828783
          mov     byte ptr [rdi + 11], 0
          ret


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D73291

Files:
  libcxx/include/__config
  libcxx/include/string

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D73291.239994.patch
Type: text/x-patch
Size: 4653 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/libcxx-commits/attachments/20200123/ed8dc4ec/attachment.bin>