CPU Toolchain Launch Postmortem. Greg Bedwell

Transcription

1 CPU Toolchain Launch Postmortem Greg Bedwell

2 x86-64 AMD Jaguar 8-core CPU 1.84 TFLOPS AMD Radeon based GPU 8GB GDDR5 RAM

3

4 postmortem noun an analysis or discussion of an event after it is over Now that we have successfully launched PlayStation 4 it is a good time to look back on our initial period of development up to that point

5 First, some history

6 SN Systems Ltd. was founded in 1988 to provide development tools for the games industry 1990 Psy-Q 16-bit home systems

7 Psy-Q included a version of GCC that was highly customized for the needs of game developers 1994 Psy-Q PlayStation

8 Continued to provide GCC but started researching a proprietary compiler technology SNC 2000 ProDG PlayStation 2

9 Provided SNC as part of the ProDG suite of tools although GCC was also available 2004 ProDG PSP (PlayStation Portable)

10 Sony Computer Entertainment Inc. acquired SN Systems in 2005 Provided SNC as part of the ProDG suite of tools although GCC was also available 2006 ProDG PlayStation 3

11 CPU Compiler is SNC 2011 PlayStation Vita

12 CPU Compiler is Clang 2013 PlayStation 4

13 The CPU compiler project is a global effort

14 Builds and build systems (and test systems)

15 We wanted to make use of all of our pre-existing test suites and test systems, which are shared across all targets

16

17 Optimized Assertions Debug Info Debug MinSizeRel Release RelWithDebInfo

18 Optimized Assertions Debug Info Debug MinSizeRel Release RelWithDebInfo

19 seconds Clang configuration effect on game build time Too slow for continuous integration testing! 2h 41m 24s Useful for continuous integration testing but still too much overhead over the pure release configuration of clang 12m 56s 7m 44s 0 Debug Release+Asserts Release

20 Optimized Assertions Debug Info Debug Checking The same set of build configurations as we use for SNC Debug for debugging Release for releasing Release Checking (Release+Asserts) for continuous integration testing

21 Optimized Assertions Debug Info Debug O Checking Release

22 clang.exe size (MB) No executable size overhead for enabling debug data Default /Zi (Produce PDB) /Zi (Produce PDB) and /OPT:REF /OPT:REF

23 Optimized Assertions Debug Info Debug O Checking Release

24 Optimized Assertions Debug Info Debug Checking Release Suppress Windows crash dialog box for Checking and Release builds

25 Improving test coverage

26 Things we know Game code is usually BIG For a new platform, the amount of code that exists is small

27 Things we know Users value CORRECTNESS over all else

28 Why write tests when I can write a test generator?

29 Why write tests when I can write _ a test generator? already have

30 Generate random C++ tests using SIMD language extensions and intrinsics to increase test coverage Random input Result

31 Random input Result

32 Random input Result

33 Random input Result

34 Random input Result

35 Random input Result

36 Random input Result Undefined behaviour makes runtimebehaviour random testing hard!

37 Random input Result Solution: Use a safe wrapper to make all undefined behaviour defined for the purpose of the test

38 Reducing optimization bugs

39 bugpoint Windows Clang integration

40 An alternative approach

41 SNC s max_opts *Not representative of actual visuals

42

43 SNC s max_opts

44 SNC s max_opts

45 SNC s max_opts

46 SNC s max_opts

47 SNC s max_opts

48 SNC s max_opts SNC optimizer SSA Form Rule Based Every transformation is guarded by a specific check

49 SNC keeps an internal counter of the number of transformations it performs Allows the user to specify a limit on the command line after which no further transformations can be performed

50 SNC s max_opts Autochop harness -O2

51 SNC s max_opts Autochop harness -O0

52 SNC s Autochop harness max_opts -O0 -O2

53 SNC s max_opts Autochop harness -O0 -O2

54 SNC s max_opts Autochop harness -O0 -O2

55 SNC s max_opts Autochop harness We ve found the game source file with the bad transformation

56 SNC s max_opts Autochop harness -Xmax_opts=2048

57 SNC s max_opts Autochop harness -Xmax_opts=1024

58 SNC s max_opts Autochop harness -Xmax_opts=512

59 SNC s max_opts Autochop harness -Xmax_opts=988

60 SNC s max_opts Autochop harness -Xmax_opts=989 Now we ve found the specific transformation causing the miscompile

61 A compiler trace shows us the culprit optimization

62 Narrowed down to a single difference in IR Just a single line change in the IR

63 SNC s max_opts A question for the community: Would this work in LLVM/Clang? (even if just at pass level)

64 The release process

65 End-user documentation is lacking Docs Release notes aimed at Clang/LLVM developers, not users

66 Docs We plan to contribute our documentation improvements to the community

67 Docs We plan to contribute our documentation improvements to the community

68 Forward compatibility

69 ABI v3.x v3.x System libraries

70 ABI v3.x v4.x System libraries

71 ABI v3.x v5.x System libraries

72 ABI v3.x v6.x System libraries

73 ABI v3.x v13.x System libraries

74 Maintaining a stable ABI is a ABI MUST (including maintaining existing ABI bugs)

75 We have created a full IA64 ABI test suite

76 ABI TEST SUITE We hope to contribute our test suite to the community (some logistics still to be worked out)

77 Developer reaction

78

79 But

80 #pragma optimize Most requested feature by an order of magnitude and already supported by all the other major compilers

81 This is the most common use-case: #pragma optimize Game runs too slowly at O0, but is very hard to debug at O2

82 #pragma optimize Solution: use a pragma to selectively disable optimization on a small set of functions to be debugged We proposed this on the mailing lists, but it is a major change and we got a limited response

83 #pragma optimize Short term solution: function level attribute to disable optimization Not user-friendly for more than a very small number of functions at a time!

84 #pragma optimize Many of our users abstract this away behind a compilerindependent interface. Function attribute does not fit this model! We still need a rangebased solution

85 In summary

86 Our initial experience with Clang and LLVM has been very positive Thanks to all of you who helped make Clang and LLVM great!

87 There are still improvements that can be made We will be working alongside you to make them