[Clang] Enable -fextend-variable-liveness at -Og #118026

SLTozer · 2024-11-28T17:23:54Z

This patch follows two other ongoing reviews, #110000 and #110102, which add a new feature to Clang - -fextend-variable-liveness, which generates fake uses for user variables (and the this pointer) to prevent them from being optimized out, providing a better debug experience at the cost of less efficient generated code. This is useful for debugging code without the need to disable optimizations outright, which is particularly important when debugging applications where some level of optimization is necessary, e.g. game code.

Following the inclusion of the flag, this patch enables it by default when -Og is set. Currently, -Og is equivalent to -O1 - it is effectively just an alias. By enabling -fextend-variable-liveness, this patch changes the code generated by Clang with -Og to have reduced optimization and greater debuggability than -O1, differentiating the two according to their respective purposes. This idea was discussed previously on Discourse, where there was general agreement with the principle of this change.

llvmbot · 2024-11-28T17:24:14Z

@llvm/pr-subscribers-clang-driver

@llvm/pr-subscribers-clang

Author: Stephen Tozer (SLTozer)

Changes

This patch follows two other ongoing reviews, #110000 and #110102, which add a new feature to Clang - -fextend-lifetimes, which generates fake uses for user variables (and the this pointer) to prevent them from being optimized out, providing a better debug experience at the cost of less efficient generated code. This is useful for debugging code without the need to disable optimizations outright, which is particularly important when debugging applications where some level of optimization is necessary, e.g. game code.

Following the inclusion of the flag, this patch enables it by default when -Og is set. Currently, -Og is equivalent to -O1 - it is effectively just an alias. By enabling -fextend-lifetimes, this patch changes the code generated by Clang with -Og to have reduced optimization and greater debuggability than -O1, differentiating the two according to their respective purposes. This idea was discussed previously on Discourse, where there was general agreement with the principle of this change.

Full diff: https://github.com/llvm/llvm-project/pull/118026.diff

4 Files Affected:

(modified) clang/docs/CommandGuide/clang.rst (+5-2)
(modified) clang/docs/ReleaseNotes.rst (+4)
(modified) clang/lib/Frontend/CompilerInvocation.cpp (+8)
(added) clang/test/CodeGen/fake-use-opt-flags.cpp (+30)

diff --git a/clang/docs/CommandGuide/clang.rst b/clang/docs/CommandGuide/clang.rst
index ca8176f854729b..8f3f0de9446b41 100644
--- a/clang/docs/CommandGuide/clang.rst
+++ b/clang/docs/CommandGuide/clang.rst
@@ -442,8 +442,11 @@ Code Generation Options
     :option:`-Oz` Like :option:`-Os` (and thus :option:`-O2`), but reduces code
     size further.
 
-    :option:`-Og` Like :option:`-O1`. In future versions, this option might
-    disable different optimizations in order to improve debuggability.
+    :option:`-Og` Similar to :option:`-O1`, but with slightly reduced
+    optimization and better variable visibility. The same optimizations are run
+    as at :option:`-O1`, but the :option:`-fextend-lifetimes` flag is also
+    set, which tries to prevent optimizations from reducing the lifetime of
+    user variables.
 
     :option:`-O` Equivalent to :option:`-O1`.
 
diff --git a/clang/docs/ReleaseNotes.rst b/clang/docs/ReleaseNotes.rst
index 601a233b81904f..14fec70dfb85f7 100644
--- a/clang/docs/ReleaseNotes.rst
+++ b/clang/docs/ReleaseNotes.rst
@@ -441,6 +441,10 @@ Modified Compiler Flags
   ``memset`` and similar functions for which it is a documented undefined
   behavior. It is implied by ``-Wnontrivial-memaccess``
 
+- The ``-Og`` optimization flag now sets ``-fextend-lifetimes``, a new compiler
+  flag, resulting in slightly reduced optimization compared to ``-O1`` in
+  exchange for improved variable visibility.
+
 Removed Compiler Flags
 -------------------------
 
diff --git a/clang/lib/Frontend/CompilerInvocation.cpp b/clang/lib/Frontend/CompilerInvocation.cpp
index 3dd94c31b2bc7a..09322d0617aaad 100644
--- a/clang/lib/Frontend/CompilerInvocation.cpp
+++ b/clang/lib/Frontend/CompilerInvocation.cpp
@@ -1834,6 +1834,14 @@ bool CompilerInvocation::ParseCodeGenArgs(CodeGenOptions &Opts, ArgList &Args,
     Opts.setInlining(CodeGenOptions::NormalInlining);
   }
 
+  // If we have specified -Og and have not set any explicit -fextend-lifetimes
+  // value, then default to -fextend-lifetimes=all.
+  if (Arg *A = Args.getLastArg(options::OPT_O_Group);
+      A && A->containsValue("g")) {
+    if (!Args.hasArg(options::OPT_fextend_lifetimes))
+      Opts.ExtendLifetimes = true;
+  }
+
   // PIC defaults to -fno-direct-access-external-data while non-PIC defaults to
   // -fdirect-access-external-data.
   Opts.DirectAccessExternalData =
diff --git a/clang/test/CodeGen/fake-use-opt-flags.cpp b/clang/test/CodeGen/fake-use-opt-flags.cpp
new file mode 100644
index 00000000000000..a99ed7a573ab22
--- /dev/null
+++ b/clang/test/CodeGen/fake-use-opt-flags.cpp
@@ -0,0 +1,30 @@
+/// Check that we generate fake uses only when -fextend-lifetimes is set and we
+/// are not setting optnone, or when we have optimizations set to -Og and we have
+/// not passed -fno-extend-lifetimes.
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O0 -fextend-lifetimes %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O0 -disable-O0-optnone -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Og %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Og -fno-extend-lifetimes %s -o - | FileCheck %s
+
+/// Test various optimization flags with -fextend-lifetimes...
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O1 %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O1 -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O2 %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O2 -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O3 %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O3 -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Os %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Os -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Oz %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Oz -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+
+// CHECK-LABEL: define{{.*}} void @_Z3fooi(i32{{.*}} %a)
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    %a.addr = alloca i32
+// CHECK-NEXT:    store i32 %a, ptr %a.addr
+// EXTEND-NEXT:   %fake.use = load i32, ptr %a.addr
+// EXTEND-NEXT:   call void (...) @llvm.fake.use(i32 %fake.use)
+// CHECK-NEXT:    ret void
+// CHECK-NEXT:  }
+
+void foo(int a) {}

llvmbot · 2024-11-28T17:24:14Z

@llvm/pr-subscribers-clang-codegen

Author: Stephen Tozer (SLTozer)

Changes

This patch follows two other ongoing reviews, #110000 and #110102, which add a new feature to Clang - -fextend-lifetimes, which generates fake uses for user variables (and the this pointer) to prevent them from being optimized out, providing a better debug experience at the cost of less efficient generated code. This is useful for debugging code without the need to disable optimizations outright, which is particularly important when debugging applications where some level of optimization is necessary, e.g. game code.

Following the inclusion of the flag, this patch enables it by default when -Og is set. Currently, -Og is equivalent to -O1 - it is effectively just an alias. By enabling -fextend-lifetimes, this patch changes the code generated by Clang with -Og to have reduced optimization and greater debuggability than -O1, differentiating the two according to their respective purposes. This idea was discussed previously on Discourse, where there was general agreement with the principle of this change.

Full diff: https://github.com/llvm/llvm-project/pull/118026.diff

4 Files Affected:

(modified) clang/docs/CommandGuide/clang.rst (+5-2)
(modified) clang/docs/ReleaseNotes.rst (+4)
(modified) clang/lib/Frontend/CompilerInvocation.cpp (+8)
(added) clang/test/CodeGen/fake-use-opt-flags.cpp (+30)

diff --git a/clang/docs/CommandGuide/clang.rst b/clang/docs/CommandGuide/clang.rst
index ca8176f854729b..8f3f0de9446b41 100644
--- a/clang/docs/CommandGuide/clang.rst
+++ b/clang/docs/CommandGuide/clang.rst
@@ -442,8 +442,11 @@ Code Generation Options
     :option:`-Oz` Like :option:`-Os` (and thus :option:`-O2`), but reduces code
     size further.
 
-    :option:`-Og` Like :option:`-O1`. In future versions, this option might
-    disable different optimizations in order to improve debuggability.
+    :option:`-Og` Similar to :option:`-O1`, but with slightly reduced
+    optimization and better variable visibility. The same optimizations are run
+    as at :option:`-O1`, but the :option:`-fextend-lifetimes` flag is also
+    set, which tries to prevent optimizations from reducing the lifetime of
+    user variables.
 
     :option:`-O` Equivalent to :option:`-O1`.
 
diff --git a/clang/docs/ReleaseNotes.rst b/clang/docs/ReleaseNotes.rst
index 601a233b81904f..14fec70dfb85f7 100644
--- a/clang/docs/ReleaseNotes.rst
+++ b/clang/docs/ReleaseNotes.rst
@@ -441,6 +441,10 @@ Modified Compiler Flags
   ``memset`` and similar functions for which it is a documented undefined
   behavior. It is implied by ``-Wnontrivial-memaccess``
 
+- The ``-Og`` optimization flag now sets ``-fextend-lifetimes``, a new compiler
+  flag, resulting in slightly reduced optimization compared to ``-O1`` in
+  exchange for improved variable visibility.
+
 Removed Compiler Flags
 -------------------------
 
diff --git a/clang/lib/Frontend/CompilerInvocation.cpp b/clang/lib/Frontend/CompilerInvocation.cpp
index 3dd94c31b2bc7a..09322d0617aaad 100644
--- a/clang/lib/Frontend/CompilerInvocation.cpp
+++ b/clang/lib/Frontend/CompilerInvocation.cpp
@@ -1834,6 +1834,14 @@ bool CompilerInvocation::ParseCodeGenArgs(CodeGenOptions &Opts, ArgList &Args,
     Opts.setInlining(CodeGenOptions::NormalInlining);
   }
 
+  // If we have specified -Og and have not set any explicit -fextend-lifetimes
+  // value, then default to -fextend-lifetimes=all.
+  if (Arg *A = Args.getLastArg(options::OPT_O_Group);
+      A && A->containsValue("g")) {
+    if (!Args.hasArg(options::OPT_fextend_lifetimes))
+      Opts.ExtendLifetimes = true;
+  }
+
   // PIC defaults to -fno-direct-access-external-data while non-PIC defaults to
   // -fdirect-access-external-data.
   Opts.DirectAccessExternalData =
diff --git a/clang/test/CodeGen/fake-use-opt-flags.cpp b/clang/test/CodeGen/fake-use-opt-flags.cpp
new file mode 100644
index 00000000000000..a99ed7a573ab22
--- /dev/null
+++ b/clang/test/CodeGen/fake-use-opt-flags.cpp
@@ -0,0 +1,30 @@
+/// Check that we generate fake uses only when -fextend-lifetimes is set and we
+/// are not setting optnone, or when we have optimizations set to -Og and we have
+/// not passed -fno-extend-lifetimes.
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O0 -fextend-lifetimes %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O0 -disable-O0-optnone -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Og %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Og -fno-extend-lifetimes %s -o - | FileCheck %s
+
+/// Test various optimization flags with -fextend-lifetimes...
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O1 %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O1 -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O2 %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O2 -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O3 %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O3 -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Os %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Os -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Oz %s -o - | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -Oz -fextend-lifetimes %s -o - | FileCheck %s --check-prefixes=CHECK,EXTEND
+
+// CHECK-LABEL: define{{.*}} void @_Z3fooi(i32{{.*}} %a)
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    %a.addr = alloca i32
+// CHECK-NEXT:    store i32 %a, ptr %a.addr
+// EXTEND-NEXT:   %fake.use = load i32, ptr %a.addr
+// EXTEND-NEXT:   call void (...) @llvm.fake.use(i32 %fake.use)
+// CHECK-NEXT:    ret void
+// CHECK-NEXT:  }
+
+void foo(int a) {}

jmorse

Some minor test comments, otherwise looking good,

jmorse · 2024-11-29T15:28:21Z

clang/lib/Frontend/CompilerInvocation.cpp

  }

+  // If we have specified -Og and have not set any explicit -fextend-lifetimes
+  // value, then default to -fextend-lifetimes=all.


This is going to create confusion by speculating the existence of an ={all,params,this} fextend-lifetimes when it doesn't exist yet, and might not.

Ah, I implemented the = option on a local-only branch, and accidentally left this comment in. My mistake, though I may yet decide to merge the feature into the earlier PRs.

jmorse · 2024-11-29T15:28:56Z

clang/docs/ReleaseNotes.rst

  behavior. It is implied by ``-Wnontrivial-memaccess``

+- The ``-Og`` optimization flag now sets ``-fextend-lifetimes``, a new compiler
+  flag, resulting in slightly reduced optimization compared to ``-O1`` in


Can we just say "trading optimization for improved..." to indicate that this is a conscious decision the developer might make?

jmorse · 2024-11-29T15:36:21Z

clang/test/CodeGen/fake-use-opt-flags.cpp

+/// Check that we generate fake uses only when -fextend-lifetimes is set and we
+/// are not setting optnone, or when we have optimizations set to -Og and we have
+/// not passed -fno-extend-lifetimes.
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-passes -O0 -fextend-lifetimes %s -o - | FileCheck %s


Intention of this RUN line is that no fake uses are generated, but there are no test-lines for ensuring that -> add an implicit-fake-not? And the other check-only FileCheck commands.

Normally I would use implicit-check-not, but in this case we're CHECK-NEXTing from the start of the function to the closing brace, which equally guarantees that no fake uses have been generated.

We just need two RUN lines, one with -fextended-lifetimes and one without. -O group is completely orthogonal and does not need to be tested

MaskRay · 2024-11-30T03:34:07Z

clang/lib/Frontend/CompilerInvocation.cpp

A->getValue() == "g"

Perhaps move the code around ToolChains/Clang.cpp:9198 and let the driver pass -fextend-lifetimes to frontend.

Is there an advantage to having it passed through the driver? I ask mainly because I see -fextend-lifetimes as being a feature of -Og optimization, and if you invoke clang -cc1 -Og directly then you'd expect/hope that the optimization behaviour would be the same as clang -Og - I'm not too familiar with expected Clang behaviour though, so if this is quite normal for driver/frontend behaviour then I'm fine to switch it.

I'm not sure if we have a fundamental principle about how flags should work in the driver V the frontend, but yeah - in some ways the driver acts as a lowering/canonicalizer - representing flag defaults, etc. So I think the expectation that clang -cc1 -Og does the same thing as clang -Og isn't valid/not something we're striving for (& in some sense counter to what we are striving for: that all the defaults/grouping is handled by the driver, and the frontend can just think about whether a flag is enabled or disabled)

I agree, the driver's job is usually to unpack high level flags into semi-orthogonal cc1 flags. So, for example, in clang-cl, /O1 becomes a collection of things like -Os -ffunction-sections -fdata-sections and maybe something else I've forgotten. clang++ -fexceptions becomes -fexceptions -fcxx-exceptions in most contexts, separating destructor cleanups from try/throw language support.

cc1 flags kind of represent features that are useful to toggle on and off for testing and test case reduction, since they're one of the things we reduce in the clang crash reducer script.

I've moved this behaviour to the driver now. Also regarding the first request to use A->getValue() == "g" - if just -O is passed then A->getValue() will crash, as there are no values. An alternative would be to change the check to A->getNumValues() && A->getValue() == "g", but I think containsValue is more concise.

SLTozer · 2024-12-12T16:17:59Z

I've updated this patch, and made it dependent on a previous PR - the logic for enabling the flag is coupled with code from the earlier PR, and also we now slightly extend the test added in that PR instead of creating a new one. To see just the code added in this patch, ignore the first commit.

jmorse

This looks good, but I'd like to reword some of it

jmorse · 2025-01-24T17:27:01Z

clang/docs/ReleaseNotes.rst

+  Clang to generate code that tries to preserve the liveness of source variables
+  through optimizations, meaning that variables will typically be visible in a
+  debugger more often. The flag has two levels: ``-fextend-variable-liveness``,
+  or ``-fextend-variable-liveness=all``, extendes the liveness of all user


Suggested change

or ``-fextend-variable-liveness=all``, extendes the liveness of all user

or ``-fextend-variable-liveness=all``, extends the liveness of all user

jmorse · 2025-01-24T17:28:40Z

clang/docs/ReleaseNotes.rst

+  in reduced performance in generated code; however, this feature will not
+  extend the liveness of some variables in cases where doing so would likely
+  have a severe impact on generated code performance.


IMO, just drop the late sentence for brevity -- it's up to the user to work out whether the trade-off is worth it.

The intention here is just to specify to the user that some variables won't have their liveness extended, so that they're not caught off-guard if they notice some variables getting excluded.

jmorse · 2025-01-24T17:31:05Z

clang/include/clang/Driver/Options.td

+  HelpText<"Extend the liveness of user variables through optimizations to "
+           "prevent stale or optimized-out variable values when debugging. Can "
+           "be applied to all user variables, or just to the C++ 'this' ptr. "
+           "May choose not to extend the liveness of some variables, such as "
+           "non-scalars larger than 4 unsigned ints, or variables in any "
+           "inlined functions.">,


Are there other compiler flags which have such long help-text? IIRC this is a super-short summary that appears on the command line with --help, and thus conciseness is key. Developers looking for more information will go to the full docs.

SLTozer · 2025-01-28T15:56:37Z

With the prior patches having landed, this is now in a state where it could land; I've rebased it onto main, meaning all of the code diff in this PR is now relevant.

rnk

Is this ready to go? I can't recall if we ultimately accepted the RFC.

SLTozer · 2025-04-17T14:07:55Z

Is this ready to go? I can't recall if we ultimately accepted the RFC.

My interpretation of the RFC is that there was a general acceptance of -Og = -O1 + -fextend-variable-liveness, so with the acceptance here I believe this is ready!

gulfemsavrun · 2025-04-23T02:08:26Z

We started seeing the following failure, and I bisected it to this commit.

[1343/43547](319) CXX host_x64/obj/third_party/protobuf/src/google/protobuf/libprotobuf_lite.generated_message_tctable_lite.cc.o
FAILED: [code=1] host_x64/obj/third_party/protobuf/src/google/protobuf/libprotobuf_lite.generated_message_tctable_lite.cc.o 
../../build/rbe/reclient_cxx.sh --working-subdir=out/not-default --exec_strategy=remote_local_fallback --preserve_unchanged_output_mtime --  ../../prebuilt/third_party/clang/linux-x64/bin/clang++ -MD -MF host_x64/obj/third_party/protobuf/src/google/protobuf/libprotobuf_lite.generated_message_tctable_lite.cc.o.d -DTOOLCHAIN_VERSION=xBAWf3R2sN1ViZaJjFVJTCaaAZEuAtpSDTaGs7iC35IC -D_LIBCPP_DISABLE_VISIBILITY_ANNOTATIONS -D_LIBCPP_REMOVE_TRANSITIVE_INCLUDES -DGOOGLE_PROTOBUF_NO_RTTI -DHAVE_PTHREAD -I../.. -Ihost_x64/gen -I../../third_party/protobuf/src -I../../third_party/protobuf/third_party/utf8_range -I../../third_party/abseil-cpp -fcolor-diagnostics -fcrash-diagnostics-dir=clang-crashreports -fcrash-diagnostics=all -ffp-contract=off --sysroot=../../prebuilt/third_party/sysroot/linux --target=x86_64-unknown-linux-gnu -ffile-compilation-dir=. -no-canonical-prefixes -fomit-frame-pointer -fdata-sections -ffunction-sections -Og -Xclang -debug-info-kind=constructor -g3 -grecord-gcc-switches -gdwarf-5 -gz=zstd -Wall -Wextra -Wconversion -Wextra-semi -Wimplicit-fallthrough -Wnewline-eof -Wstrict-prototypes -Wwrite-strings -Wno-sign-conversion -Wno-unused-parameter -Wnonportable-system-include-path -Wno-missing-field-initializers -Wno-extra-qualification -Wno-cast-function-type-strict -Wno-cast-function-type-mismatch -Wno-unknown-warning-option -Wno-missing-template-arg-list-after-template-kw -Wno-deprecated-pragma -Wno-nontrivial-memaccess -fvisibility=hidden -Werror -Wa,--fatal-warnings --sysroot=../../prebuilt/third_party/sysroot/linux --target=x86_64-unknown-linux-gnu -Wno-deprecated-pragma -Wno-enum-enum-conversion -Wno-extra-semi -Wno-float-conversion -Wno-implicit-float-conversion -Wno-implicit-int-conversion -Wno-implicit-int-float-conversion -Wno-invalid-noreturn -Wno-missing-field-initializers -Wno-sign-compare -Wno-unused-function -Wno-unused-private-field -Wno-deprecated-declarations -Wno-deprecated-this-capture -Wno-unnecessary-virtual-specifier -Wno-c++98-compat-extra-semi -Wno-shorten-64-to-32 -Wno-sign-compare -Wno-missing-field-initializers -Wno-conversion -Wno-unknown-warning-option -Wno-implicit-int-float-conversion -Wno-array-parameter -Wno-deprecated-builtins -fvisibility-inlines-hidden -stdlib=libc++ -stdlib=libc++ -std=c++20 -Wno-deprecated-this-capture -fno-exceptions -fno-rtti -ftemplate-backtrace-limit=0 -stdlib=libc++ -c ../../third_party/protobuf/src/google/protobuf/generated_message_tctable_lite.cc -o host_x64/obj/third_party/protobuf/src/google/protobuf/libprotobuf_lite.generated_message_tctable_lite.cc.o
../../third_party/protobuf/src/google/protobuf/generated_message_tctable_lite.cc:68:28: error: cannot compile this tail call skipping over cleanups yet
   68 |   PROTOBUF_MUSTTAIL return GenericFallbackImpl<MessageLite, std::string>(
      |                            ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
   69 |       PROTOBUF_TC_PARAM_PASS);
      |       ~~~~~~~~~~~~~~~~~~~~~~~

I'm uploading a reproducer.
clang-crashreports.zip

Any ideas why we might be hitting this issue? This happens during a built with -Og, and issue does not remove if I remove -Og flag.

SLTozer · 2025-04-23T09:26:29Z

Any ideas why we might be hitting this issue? This happens during a built with -Og, and issue does not remove if I remove -Og flag.

At a glance that looks like it could be related, we generate (no-op) cleanups as part of -fextend-variable-liveness. I'll look into this, but could you clarify what you mean - did you mean that the error occurs with -Og and does not occur when you remove -Og?

Fixes the issue reported following the merge of llvm#118026. When a valid `musttail` call is made, the function it is made from must return immediately after the call; if there are any cleanups left in the function, then an error is triggered. This is not necessary for fake uses however - it is perfectly valid to simply emit the fake use "cleanup" code before the tail call, and indeed LLVM will automatically move any fake uses following a tail call to come before the tail call. Therefore, this patch specifically choose to handle fake use cleanups when a musttail call is present by simply emitting them immediately before the call.

SLTozer · 2025-04-23T13:45:21Z

We started seeing the following failure, and I bisected it to this commit.

I've created a fix for this issue here: #136867

gulfemsavrun · 2025-04-23T16:29:17Z

Any ideas why we might be hitting this issue? This happens during a built with -Og, and issue does not remove if I remove -Og flag.

At a glance that looks like it could be related, we generate (no-op) cleanups as part of -fextend-variable-liveness. I'll look into this, but could you clarify what you mean - did you mean that the error occurs with -Og and does not occur when you remove -Og?

Yes, it does not reproduce if I remove -Og.

mikaelholmen · 2025-04-25T07:24:30Z

I did some downstream testing with "-Og" since this patch and noticed that e.g.

llc -mtriple=avr bbi-106478.ll

crashes with

SoftenFloatOperand Op #1: t2: ch = fake_use t0, ConstantFP:f64<0.000000e+00>

LLVM ERROR: Do not know how to soften this operator's operand!

and

llc -mtriple=systemz bbi-106478.ll

fails with

llc: ../lib/Target/SystemZ/SystemZLongBranch.cpp:226: unsigned int getInstSizeInBytes(const MachineInstr &, const SystemZInstrInfo *): Assertion `(Size || MI.isDebugOrPseudoInstr() || MI.isPosition() || MI.isKill() || MI.isImplicitDef() || MI.getOpcode() == TargetOpcode::MEMBARRIER || MI.getOpcode() == TargetOpcode::INIT_UNDEF || MI.isInlineAsm() || MI.getOpcode() == SystemZ::STACKMAP || MI.getOpcode() == SystemZ::PATCHPOINT || MI.getOpcode() == SystemZ::EH_SjLj_Setup) && "Missing size value for instruction."' failed.

for bbi-106478.ll just being

define double @idd() {
entry:
  notail call void (...) @llvm.fake.use(double 0.000000e+00)
  ret double 0.000000e+00
}

(and if removing the llvm.fake.use call they don't crash)
so I guess there is some testing and fixes around llvm.fake.use left to do.

Fixes the issue reported following the merge of #118026. When a valid `musttail` call is made, the function it is made from must return immediately after the call; if there are any cleanups left in the function, then an error is triggered. This is not necessary for fake uses however - it is perfectly valid to simply emit the fake use "cleanup" code before the tail call, and indeed LLVM will automatically move any fake uses following a tail call to come before the tail call. Therefore, this patch specifically choose to handle fake use cleanups when a musttail call is present by simply emitting them immediately before the call.

SLTozer · 2025-04-25T11:25:58Z

I did some downstream testing with "-Og" since this patch and noticed that e.g.

Thanks for catching these - in terms of legalizer support it looks like everything relevant is implemented except for SoftenFloatOperand, so that should be a straightforward fix. The target-specific error is also trivial (FAKE_USE has a size of 0), but it would be preferable if every target didn't need to account for it - previously there was a discussion of whether FAKE_USE should have isPosition, and this probably settles that since it would allow them to be treated correctly by-default.

I'll revert this for now (after verifying), since it may take a little time to confirm that changing isPosition() doesn't introduce any new bugs; I'll try to expand the fake use tests to cover more targets at the same time.

mikaelholmen · 2025-04-25T11:32:42Z

I'll revert this for now (after verifying), since it may take a little time to confirm that changing isPosition() doesn't introduce any new bugs; I'll try to expand the fake use tests to cover more targets at the same time.

Thanks!

Reverted following several issues appearing related to fake uses, reported on the github PR. This reverts commit a9dff35.

Recently, a new flag -fextend-variable-liveness was added that prevents optimizations from removing the values of source variables in some cases, improving the quality of debugging for optimized builds where it is enabled. Following the inclusion of the flag, this patch enables it by default when `-Og` is set. Currently, `-Og` is equivalent to `-O1` - it is effectively just an alias. By enabling `-fextend-lifetimes`, this patch changes the code generated by Clang with `-Og` to have reduced optimization and greater debuggability than `-O1`, differentiating the two according to their respective purposes. This idea was discussed previously on Discourse where there was general agreement with the principle of this change: https://discourse.llvm.org/t/rfc-redefine-og-o1-and-add-a-new-level-of-og

Fixes the issue reported following the merge of llvm#118026. When a valid `musttail` call is made, the function it is made from must return immediately after the call; if there are any cleanups left in the function, then an error is triggered. This is not necessary for fake uses however - it is perfectly valid to simply emit the fake use "cleanup" code before the tail call, and indeed LLVM will automatically move any fake uses following a tail call to come before the tail call. Therefore, this patch specifically choose to handle fake use cleanups when a musttail call is present by simply emitting them immediately before the call.

Reverted following several issues appearing related to fake uses, reported on the github PR. This reverts commit a9dff35.

MaskRay · 2025-05-11T06:13:57Z

Since the option has been renamed to -fextend-variable-liveness, I think you want to edit this title and the description to reflect the naming change as well. I wanted to check the status the effort and I tried git log --grep fextend-lifetimes but found nothing, then I realized that the option has been renamed :)

Also, make sure that you update https://discourse.llvm.org/t/rfc-redefine-og-o1-and-add-a-new-level-of-og/72850/25 after the option relands after the musttail fix.

Relands this feature after several fixes: * Force fake uses to be emitted before musttail calls (#136867) * Added soften-float legalization for fake uses (#142714) * Treat fake uses as size-less instructions in a SystemZ assert (#144390) If further issues with fake uses are found then this may be reverted again, but all currently-known issues are resolved. This reverts commit 2dc6e98.

sbc100 · 2025-09-22T17:32:02Z

This change seems to have broken some emscripten users downstream: emscripten-core/emscripten#25301

Seems likely a bug in WebAssembly backend, but posting here in case anything looks obvious/strange.

efriedma-quic · 2025-09-22T17:42:32Z

There's a corresponding intrinsic llvm.fake.use (FAKE_USE in MIR), which isn't used by anything else. Possibly there's a bug in the implementation of that intrinsic.

SLTozer · 2025-09-22T18:40:14Z

This change seems to have broken some emscripten users downstream: emscripten-core/emscripten#25301

I'm not familiar with emscripten/wasm/binaryen; do you have the last state of the IR/MIR? Potentially the --print-after-all output for the reproducer would be usable too.

I don't know exactly how to read the binaryen IR, but for some speculation: the "fake use" feature that was added by this change intentionally preserves values that are not used by the program but that are (likely to be) referenced by debug information, to prevent variables from appearing optimized out. At a glance, the error in question seems to be that binaryen expects some values in the block to have been dropped, which could be related to the existence of fake use intrinsics forcing an unused value to be emitted?

SLTozer added clang Clang issues not falling into any other category clang:frontend Language frontend issues, e.g. anything involving "Sema" clang:codegen IR generation bugs: mangling, exceptions, etc. labels Nov 28, 2024

SLTozer requested review from AaronBallman, dwblaikie, jansvoboda11, jmorse, rnk and tru November 28, 2024 17:23

SLTozer self-assigned this Nov 28, 2024

jmorse reviewed Nov 29, 2024

View reviewed changes

MaskRay reviewed Nov 30, 2024

View reviewed changes

SLTozer mentioned this pull request Dec 12, 2024

[Clang] Add "extend lifetime" flags and release note #110000

Closed

SLTozer force-pushed the extend-lifetimes-at-og branch from ec32824 to 2b59ce7 Compare December 12, 2024 16:12

llvmbot added the clang:driver 'clang' and 'clang++' user-facing binaries. Not 'clang-cl' label Dec 12, 2024

jmorse reviewed Jan 24, 2025

View reviewed changes

SLTozer force-pushed the extend-lifetimes-at-og branch from 2b59ce7 to beb5e19 Compare January 28, 2025 15:17

rnk approved these changes Apr 16, 2025

View reviewed changes

MaskRay approved these changes Apr 16, 2025

View reviewed changes

SLTozer force-pushed the extend-lifetimes-at-og branch from beb5e19 to e503747 Compare April 17, 2025 14:15

SLTozer added 2 commits April 17, 2025 15:52

Enable -fextend-lifetimes at -Og

c0369e1

Update docs to fit current compiler/doc state

7126cb5

SLTozer force-pushed the extend-lifetimes-at-og branch from e503747 to 7126cb5 Compare April 17, 2025 14:57

SLTozer merged commit a9dff35 into llvm:main Apr 17, 2025
8 of 12 checks passed

SLTozer mentioned this pull request Apr 23, 2025

[Clang][CodeGen] Emit fake uses before musttail calls #136867

Merged

SLTozer mentioned this pull request Apr 25, 2025

[ARM] Crash with -Og or -fextend-variable-liveness: "Do not know how to soften this operator's operand" #137308

Closed

SLTozer added a commit that referenced this pull request Apr 25, 2025

Revert "[Clang] Enable -fextend-lifetimes at -Og (#118026)"

2dc6e98

Reverted following several issues appearing related to fake uses, reported on the github PR. This reverts commit a9dff35.

IanWood1 pushed a commit to IanWood1/llvm-project that referenced this pull request May 6, 2025

Revert "[Clang] Enable -fextend-lifetimes at -Og (llvm#118026)"

04b0df3

Reverted following several issues appearing related to fake uses, reported on the github PR. This reverts commit a9dff35.

SLTozer changed the title ~~[Clang] Enable -fextend-lifetimes at -Og~~ [Clang] Enable -fextend-variable-liveness at -Og Jun 19, 2025

	or ``-fextend-variable-liveness=all``, extendes the liveness of all user
	or ``-fextend-variable-liveness=all``, extends the liveness of all user

[Clang] Enable -fextend-variable-liveness at -Og #118026

[Clang] Enable -fextend-variable-liveness at -Og #118026

Uh oh!

Conversation

SLTozer commented Nov 28, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Nov 28, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Nov 28, 2024

Uh oh!

jmorse left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SLTozer Nov 29, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SLTozer commented Dec 12, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jmorse left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SLTozer commented Jan 28, 2025

Uh oh!

rnk left a comment

Choose a reason for hiding this comment

Uh oh!

SLTozer commented Apr 17, 2025

Uh oh!

Uh oh!

gulfemsavrun commented Apr 23, 2025

Uh oh!

SLTozer commented Apr 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

SLTozer commented Apr 23, 2025

Uh oh!

gulfemsavrun commented Apr 23, 2025

Uh oh!

mikaelholmen commented Apr 25, 2025

Uh oh!

SLTozer commented Apr 25, 2025

Uh oh!

mikaelholmen commented Apr 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MaskRay commented May 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

SLTozer commented Nov 28, 2024 •

edited

Loading

llvmbot commented Nov 28, 2024 •

edited

Loading

SLTozer Nov 29, 2024 •

edited

Loading

SLTozer commented Dec 12, 2024 •

edited

Loading

SLTozer commented Apr 23, 2025 •

edited

Loading

mikaelholmen commented Apr 25, 2025 •

edited

Loading

MaskRay commented May 11, 2025 •

edited

Loading

SLTozer commented Sep 22, 2025 •

edited

Loading