[TySan][Clang] Add clang flag to use tysan outlined instrumentation a… #166170
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters. Learn more about bidirectional Unicode characters
llvmbot added clang 
| @llvm/pr-subscribers-clang-driver Author: Matthew Nagy (gbMattN) Changes…nd update docs Full diff: https://github.com/llvm/llvm-project/pull/166170.diff 6 Files Affected:
diff --git a/clang/docs/TypeSanitizer.rst b/clang/docs/TypeSanitizer.rst index 3c683a6c24bb4..6420d9e1d0ebd 100644 --- a/clang/docs/TypeSanitizer.rst +++ b/clang/docs/TypeSanitizer.rst @@ -20,9 +20,13 @@ code is build with ``-fno-strict-aliasing``, sacrificing performance. TypeSanitizer is built to catch when these strict aliasing rules have been violated, helping users find where such bugs originate in their code despite the code looking valid at first glance. -As TypeSanitizer is still experimental, it can currently have a large impact on runtime speed, -memory use, and code size. It also has a large compile-time overhead. Work is being done to -reduce these impacts. +Typical memory overhead introduced by TypeSanitizer is about **8x**. Runtime slowdown varies greatly +depending on how often the instrumented code relies on type aliasing. In the best case slowdown is +**2x-3x**. + +The compiler instrumentation also has an impact on code size and compilation overhead. There is an +experimental :ref:`instrumentation outlining option<outlining_flag>` which can greatly reduce this +but this may decrease runtime performance. The TypeSanitizer Algorithm =========================== @@ -128,6 +132,23 @@ references to LLVM IR specific terms. Sanitizer features ================== +.. _outlining_flag: + +Instrumentation code outlining +------------------------------ + +By default TypeSanitizer inlines the instrumentation code. This leads to increased +binary size and compilation time. Using the clang flag +``-fsanitize-type-outline-instrumentation`` (default: ``false``) +forces all code instrumentation to be outlined. This reduces the size of the +generated code and reduces compile-time overhead, but it also reduces runtime +performance. + +This outlined instrumentation is new. If you wish to verify that the outlined instrumentation +is behaving in the same way as the inline instrumentation, you can force TypeSanitizer +to use both types of instrumentation. You can use the clang flag +``-fsanitize-type-verify-outlined-instrumentation`` (default: ``false``) to do this. + ``__has_feature(type_sanitizer)`` ------------------------------------ diff --git a/clang/docs/UsersManual.rst b/clang/docs/UsersManual.rst index fb22ad3c90af4..04da152bd0aff 100644 --- a/clang/docs/UsersManual.rst +++ b/clang/docs/UsersManual.rst @@ -2277,6 +2277,15 @@ are listed below. See :doc: `AddressSanitizer` for more details. +.. option:: -f[no-]sanitize-type-outline-instrumentation + + Controls how type sanitizer code is generated. If enabled will always use + a function call instead of inlining the code. Turning this option on may + reduce the binary size and compilation overhead, but might result in a worse + run-time performance. + + See :doc: `TypeSanitizer` for more details. + .. option:: -f[no-]sanitize-stats Enable simple statistics gathering for the enabled sanitizers. diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td index 6e1c9425d8d75..adef9a592b55e 100644 --- a/clang/include/clang/Driver/Options.td +++ b/clang/include/clang/Driver/Options.td @@ -2454,6 +2454,18 @@ def fsanitize_address_outline_instrumentation : Flag<["-"], "fsanitize-address-o def fno_sanitize_address_outline_instrumentation : Flag<["-"], "fno-sanitize-address-outline-instrumentation">, Group<f_clang_Group>, HelpText<"Use default code inlining logic for the address sanitizer">; +def fsanitize_type_outline_instrumentation : Flag<["-"], "fsanitize-type-outline-instrumentation">, + Group<f_clang_Group>, + HelpText<"Always generate function calls for type sanitizer instrumentation">; +def fno_sanitize_type_outline_instrumentation : Flag<["-"], "fno-sanitize-type-outline-instrumentation">, + Group<f_clang_Group>, + HelpText<"Use default code inlining logic for the type sanitizer">; +def fsanitize_type_verify_outlined_instrumentation : Flag<["-"], "fsanitize_type_verify_outlined_instrumentation">, + Group<f_clang_Group>, + HelpText<"Use both inlined and outlined instrumentation for type sanitizer to verify equivilence">; +def fno_sanitize_type_verify_outlined_instrumentation : Flag<["-"], "fno_sanitize_type_verify_outlined_instrumentation">, + Group<f_clang_Group>, + HelpText<"Don't use both inlined and outlined instrumentation for type sanitizer to verify equivilence">; defm sanitize_stable_abi : OptInCC1FFlag<"sanitize-stable-abi", "Stable ", "Conventional ", "ABI instrumentation for sanitizer runtime. Default: Conventional">; diff --git a/clang/include/clang/Driver/SanitizerArgs.h b/clang/include/clang/Driver/SanitizerArgs.h index eea7897e96afd..29569bcf7fa25 100644 --- a/clang/include/clang/Driver/SanitizerArgs.h +++ b/clang/include/clang/Driver/SanitizerArgs.h @@ -67,6 +67,8 @@ class SanitizerArgs { bool TsanFuncEntryExit = true; bool TsanAtomics = true; bool MinimalRuntime = false; + bool TysanOutlineInstrumentation = false; + bool TysanVerifyOutlinedInstrumentation = false; // True if cross-dso CFI support if provided by the system (i.e. Android). bool ImplicitCfiRuntime = false; bool NeedsMemProfRt = false; diff --git a/clang/lib/Driver/SanitizerArgs.cpp b/clang/lib/Driver/SanitizerArgs.cpp index 5dd48f53b9069..c106dd63c2805 100644 --- a/clang/lib/Driver/SanitizerArgs.cpp +++ b/clang/lib/Driver/SanitizerArgs.cpp @@ -1176,6 +1176,17 @@ SanitizerArgs::SanitizerArgs(const ToolChain &TC, options::OPT_fno_sanitize_alloc_token_extended, AllocTokenExtended); } + if (AllAddedKinds & SanitizerKind::Type) { + TysanOutlineInstrumentation = + Args.hasFlag(options::OPT_fsanitize_type_outline_instrumentation, + options::OPT_fno_sanitize_type_outline_instrumentation, + TysanOutlineInstrumentation); + TysanVerifyOutlinedInstrumentation = Args.hasFlag( + options::OPT_fsanitize_type_verify_outlined_instrumentation, + options::OPT_fno_sanitize_type_verify_outlined_instrumentation, + TysanVerifyOutlinedInstrumentation); + } + LinkRuntimes = Args.hasFlag(options::OPT_fsanitize_link_runtime, options::OPT_fno_sanitize_link_runtime, !Args.hasArg(options::OPT_r)); @@ -1500,6 +1511,15 @@ void SanitizerArgs::addArgs(const ToolChain &TC, const llvm::opt::ArgList &Args, CmdArgs.push_back("-asan-instrumentation-with-call-threshold=0"); } + if (TysanOutlineInstrumentation || TysanVerifyOutlinedInstrumentation) { + CmdArgs.push_back("-mllvm"); + CmdArgs.push_back("-tysan-outline-instrumentation"); + } + if (TysanVerifyOutlinedInstrumentation) { + CmdArgs.push_back("-mllvm"); + CmdArgs.push_back("-tysan-verify-outlined-instrumentation"); + } + // When emitting Stable ABI instrumentation, force outlining calls and avoid // inlining shadow memory poisoning. While this is a big performance burden // for now it allows full abstraction from implementation details. diff --git a/clang/test/CodeGen/sanitize-type-outlined.cpp b/clang/test/CodeGen/sanitize-type-outlined.cpp new file mode 100644 index 0000000000000..3c22f15c93d94 --- /dev/null +++ b/clang/test/CodeGen/sanitize-type-outlined.cpp @@ -0,0 +1,23 @@ +// RUN: %clang -S -fsanitize=type -emit-llvm -o - -fsanitize=type %s \ +// RUN: | FileCheck %s --check-prefixes=CHECK-NO-OUTLINE +// RUN: %clang -S -fsanitize=type -emit-llvm -o - -fsanitize=type %s \ +// RUN: -fsanitize-type-outline-instrumentation \ +// RUN: | FileCheck %s --check-prefixes=CHECK-OUTLINE + +// RUN: %clang -S -fsanitize=type -emit-llvm -o - -fsanitize=type %s \ +// RUN: -fsanitize-type-outline-instrumentation \ +// RUN: -fsanitize-type-verify-outlined-instrumentation \ +// RUN: | FileCheck %s --check-prefixes=CHECK-OUTLINE-VERIFY +// RUN: %clang -S -fsanitize=type -emit-llvm -o - -fsanitize=type %s \ +// RUN: -fsanitize-type-verify-outlined-instrumentation \ +// RUN: | FileCheck %s --check-prefixes=CHECK-VERIFY + +// CHECK-NO-OUTLINE-NOT: call{{.*}}@__tysan_instrument_mem_inst +// CHECK-OUTLINE: call{{.*}}@__tysan_instrument_mem_inst +// CHECK-OUTLINE-VERIFY: call{{.*}}@__tysan_instrument_mem_inst +// CHECK-VERIFY: call{{.*}}@__tysan_instrument_mem_inst + +float alias(int *ptr){ + float *aliasedPtr = (float *)ptr; + return *aliasedPtr; +} |
| @llvm/pr-subscribers-clang Author: Matthew Nagy (gbMattN) Changes…nd update docs Full diff: https://github.com/llvm/llvm-project/pull/166170.diff 6 Files Affected:
diff --git a/clang/docs/TypeSanitizer.rst b/clang/docs/TypeSanitizer.rst index 3c683a6c24bb4..6420d9e1d0ebd 100644 --- a/clang/docs/TypeSanitizer.rst +++ b/clang/docs/TypeSanitizer.rst @@ -20,9 +20,13 @@ code is build with ``-fno-strict-aliasing``, sacrificing performance. TypeSanitizer is built to catch when these strict aliasing rules have been violated, helping users find where such bugs originate in their code despite the code looking valid at first glance. -As TypeSanitizer is still experimental, it can currently have a large impact on runtime speed, -memory use, and code size. It also has a large compile-time overhead. Work is being done to -reduce these impacts. +Typical memory overhead introduced by TypeSanitizer is about **8x**. Runtime slowdown varies greatly +depending on how often the instrumented code relies on type aliasing. In the best case slowdown is +**2x-3x**. + +The compiler instrumentation also has an impact on code size and compilation overhead. There is an +experimental :ref:`instrumentation outlining option<outlining_flag>` which can greatly reduce this +but this may decrease runtime performance. The TypeSanitizer Algorithm =========================== @@ -128,6 +132,23 @@ references to LLVM IR specific terms. Sanitizer features ================== +.. _outlining_flag: + +Instrumentation code outlining +------------------------------ + +By default TypeSanitizer inlines the instrumentation code. This leads to increased +binary size and compilation time. Using the clang flag +``-fsanitize-type-outline-instrumentation`` (default: ``false``) +forces all code instrumentation to be outlined. This reduces the size of the +generated code and reduces compile-time overhead, but it also reduces runtime +performance. + +This outlined instrumentation is new. If you wish to verify that the outlined instrumentation +is behaving in the same way as the inline instrumentation, you can force TypeSanitizer +to use both types of instrumentation. You can use the clang flag +``-fsanitize-type-verify-outlined-instrumentation`` (default: ``false``) to do this. + ``__has_feature(type_sanitizer)`` ------------------------------------ diff --git a/clang/docs/UsersManual.rst b/clang/docs/UsersManual.rst index fb22ad3c90af4..04da152bd0aff 100644 --- a/clang/docs/UsersManual.rst +++ b/clang/docs/UsersManual.rst @@ -2277,6 +2277,15 @@ are listed below. See :doc: `AddressSanitizer` for more details. +.. option:: -f[no-]sanitize-type-outline-instrumentation + + Controls how type sanitizer code is generated. If enabled will always use + a function call instead of inlining the code. Turning this option on may + reduce the binary size and compilation overhead, but might result in a worse + run-time performance. + + See :doc: `TypeSanitizer` for more details. + .. option:: -f[no-]sanitize-stats Enable simple statistics gathering for the enabled sanitizers. diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td index 6e1c9425d8d75..adef9a592b55e 100644 --- a/clang/include/clang/Driver/Options.td +++ b/clang/include/clang/Driver/Options.td @@ -2454,6 +2454,18 @@ def fsanitize_address_outline_instrumentation : Flag<["-"], "fsanitize-address-o def fno_sanitize_address_outline_instrumentation : Flag<["-"], "fno-sanitize-address-outline-instrumentation">, Group<f_clang_Group>, HelpText<"Use default code inlining logic for the address sanitizer">; +def fsanitize_type_outline_instrumentation : Flag<["-"], "fsanitize-type-outline-instrumentation">, + Group<f_clang_Group>, + HelpText<"Always generate function calls for type sanitizer instrumentation">; +def fno_sanitize_type_outline_instrumentation : Flag<["-"], "fno-sanitize-type-outline-instrumentation">, + Group<f_clang_Group>, + HelpText<"Use default code inlining logic for the type sanitizer">; +def fsanitize_type_verify_outlined_instrumentation : Flag<["-"], "fsanitize_type_verify_outlined_instrumentation">, + Group<f_clang_Group>, + HelpText<"Use both inlined and outlined instrumentation for type sanitizer to verify equivilence">; +def fno_sanitize_type_verify_outlined_instrumentation : Flag<["-"], "fno_sanitize_type_verify_outlined_instrumentation">, + Group<f_clang_Group>, + HelpText<"Don't use both inlined and outlined instrumentation for type sanitizer to verify equivilence">; defm sanitize_stable_abi : OptInCC1FFlag<"sanitize-stable-abi", "Stable ", "Conventional ", "ABI instrumentation for sanitizer runtime. Default: Conventional">; diff --git a/clang/include/clang/Driver/SanitizerArgs.h b/clang/include/clang/Driver/SanitizerArgs.h index eea7897e96afd..29569bcf7fa25 100644 --- a/clang/include/clang/Driver/SanitizerArgs.h +++ b/clang/include/clang/Driver/SanitizerArgs.h @@ -67,6 +67,8 @@ class SanitizerArgs { bool TsanFuncEntryExit = true; bool TsanAtomics = true; bool MinimalRuntime = false; + bool TysanOutlineInstrumentation = false; + bool TysanVerifyOutlinedInstrumentation = false; // True if cross-dso CFI support if provided by the system (i.e. Android). bool ImplicitCfiRuntime = false; bool NeedsMemProfRt = false; diff --git a/clang/lib/Driver/SanitizerArgs.cpp b/clang/lib/Driver/SanitizerArgs.cpp index 5dd48f53b9069..c106dd63c2805 100644 --- a/clang/lib/Driver/SanitizerArgs.cpp +++ b/clang/lib/Driver/SanitizerArgs.cpp @@ -1176,6 +1176,17 @@ SanitizerArgs::SanitizerArgs(const ToolChain &TC, options::OPT_fno_sanitize_alloc_token_extended, AllocTokenExtended); } + if (AllAddedKinds & SanitizerKind::Type) { + TysanOutlineInstrumentation = + Args.hasFlag(options::OPT_fsanitize_type_outline_instrumentation, + options::OPT_fno_sanitize_type_outline_instrumentation, + TysanOutlineInstrumentation); + TysanVerifyOutlinedInstrumentation = Args.hasFlag( + options::OPT_fsanitize_type_verify_outlined_instrumentation, + options::OPT_fno_sanitize_type_verify_outlined_instrumentation, + TysanVerifyOutlinedInstrumentation); + } + LinkRuntimes = Args.hasFlag(options::OPT_fsanitize_link_runtime, options::OPT_fno_sanitize_link_runtime, !Args.hasArg(options::OPT_r)); @@ -1500,6 +1511,15 @@ void SanitizerArgs::addArgs(const ToolChain &TC, const llvm::opt::ArgList &Args, CmdArgs.push_back("-asan-instrumentation-with-call-threshold=0"); } + if (TysanOutlineInstrumentation || TysanVerifyOutlinedInstrumentation) { + CmdArgs.push_back("-mllvm"); + CmdArgs.push_back("-tysan-outline-instrumentation"); + } + if (TysanVerifyOutlinedInstrumentation) { + CmdArgs.push_back("-mllvm"); + CmdArgs.push_back("-tysan-verify-outlined-instrumentation"); + } + // When emitting Stable ABI instrumentation, force outlining calls and avoid // inlining shadow memory poisoning. While this is a big performance burden // for now it allows full abstraction from implementation details. diff --git a/clang/test/CodeGen/sanitize-type-outlined.cpp b/clang/test/CodeGen/sanitize-type-outlined.cpp new file mode 100644 index 0000000000000..3c22f15c93d94 --- /dev/null +++ b/clang/test/CodeGen/sanitize-type-outlined.cpp @@ -0,0 +1,23 @@ +// RUN: %clang -S -fsanitize=type -emit-llvm -o - -fsanitize=type %s \ +// RUN: | FileCheck %s --check-prefixes=CHECK-NO-OUTLINE +// RUN: %clang -S -fsanitize=type -emit-llvm -o - -fsanitize=type %s \ +// RUN: -fsanitize-type-outline-instrumentation \ +// RUN: | FileCheck %s --check-prefixes=CHECK-OUTLINE + +// RUN: %clang -S -fsanitize=type -emit-llvm -o - -fsanitize=type %s \ +// RUN: -fsanitize-type-outline-instrumentation \ +// RUN: -fsanitize-type-verify-outlined-instrumentation \ +// RUN: | FileCheck %s --check-prefixes=CHECK-OUTLINE-VERIFY +// RUN: %clang -S -fsanitize=type -emit-llvm -o - -fsanitize=type %s \ +// RUN: -fsanitize-type-verify-outlined-instrumentation \ +// RUN: | FileCheck %s --check-prefixes=CHECK-VERIFY + +// CHECK-NO-OUTLINE-NOT: call{{.*}}@__tysan_instrument_mem_inst +// CHECK-OUTLINE: call{{.*}}@__tysan_instrument_mem_inst +// CHECK-OUTLINE-VERIFY: call{{.*}}@__tysan_instrument_mem_inst +// CHECK-VERIFY: call{{.*}}@__tysan_instrument_mem_inst + +float alias(int *ptr){ + float *aliasedPtr = (float *)ptr; + return *aliasedPtr; +} |
melver reviewed Nov 3, 2025
| HelpText<"Use default code inlining logic for the type sanitizer">; | ||
| def fsanitize_type_verify_outlined_instrumentation : Flag<["-"], "fsanitize-type-verify-outlined-instrumentation">, | ||
| Group<f_clang_Group>, | ||
| HelpText<"Use both inlined and outlined instrumentation for type sanitizer to verify equivilence">; |
| HelpText<"Use both inlined and outlined instrumentation for type sanitizer to verify equivilence">; | ||
| def fno_sanitize_type_verify_outlined_instrumentation : Flag<["-"], "fno_sanitize-type-verify-outlined-instrumentation">, | ||
| Group<f_clang_Group>, | ||
| HelpText<"Don't use both inlined and outlined instrumentation for type sanitizer to verify equivilence">; |
| // CHECK-OUTLINE-VERIFY: call{{.*}}@__tysan_instrument_mem_inst | ||
| // CHECK-VERIFY: call{{.*}}@__tysan_instrument_mem_inst | ||
| | ||
| float alias(int *ptr){ |
There was a problem hiding this comment.
Can you use CHECK-LABEL to avoid accidentally passing the test for some reason?
| // RUN: -fsanitize-type-verify-outlined-instrumentation \ | ||
| // RUN: | FileCheck %s --check-prefixes=CHECK-VERIFY | ||
| | ||
| // CHECK-NO-OUTLINE-NOT: call{{.*}}@__tysan_instrument_mem_inst |
There was a problem hiding this comment.
The existing IR test for outline instrumentation (llvm/test/Instrumentation/TypeSanitizer/basic_outlined.ll) checks __tysan_instrument_with_shadow_update.
Why does this test pass? Are there any memintrinsics here?
There was a problem hiding this comment.
Both __tysan_instrument_mem_inst and __tysan_instrument_with_shadow_update are functions which can be inserted when outlining is enabled, so either (perhaps better if both) can be used to check that the outlining of instrumentation is occurring. I made this test mimicking the minimal one ASan has but on second thought maybe I should make it a bit more verbose to properly check for all possible outlined functions
There was a problem hiding this comment.
Right. Note the ASan tests were written ages ago, so maybe we know better now :-)
fhahn reviewed Nov 4, 2025
There was a problem hiding this comment.
@gbMattN did you already get a chance to verify the outlined instrumentation matches via -fsanitize-type-verify-outlined-instrumentation on a larger set of workloads?
Are you aware of any cases where the outlined insturmentation performs worse than the inline instrumentation?
If not, I think it would be better to due some due diligence testing using -fsanitize-type-verify-outlined-instrumentation and flip the default, instead of adding the plumbing to make it convenient to use
| I'm not aware of any cases where the outlined instrumentation gives you worse diagnostics than the inlined one, but I think before being comfortable changing the default I will want to do a little more testing. What I've done is a little more "ad-hoc" than exhaustive |
Sounds good, I think it would probably be worth doing that and then flip the default, instead of adding more plumbing, unless the results are not that clear cut |
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit. This suggestion is invalid because no changes were made to the code. Suggestions cannot be applied while the pull request is closed. Suggestions cannot be applied while viewing a subset of changes. Only one suggestion per line can be applied in a batch. Add this suggestion to a batch that can be applied as a single commit. Applying suggestions on deleted lines is not supported. You must change the existing code in this line in order to create a valid suggestion. Outdated suggestions cannot be applied. This suggestion has been applied or marked resolved. Suggestions cannot be applied from pending reviews. Suggestions cannot be applied on multi-line comments. Suggestions cannot be applied while the pull request is queued to merge. Suggestion cannot be applied right now. Please check back later.
…nd update docs