From cdc7a4110282d30caa25666855928848fc87031cb137e454b3605ec8890bb384 Mon Sep 17 00:00:00 2001
From: Aaron Puchert <aaronpuchert@alice-dsl.net>
Date: Thu, 19 Jan 2023 20:24:57 +0000
Subject: [PATCH] - Add llvm-workaround-superfluous-branches.patch: hints LLVM
 to   eliminate branches until gh#llvm/llvm-project#28804 is solved.

OBS-URL: https://build.opensuse.org/package/show/devel:tools:compiler/llvm15?expand=0&rev=18
---
 llvm-workaround-superfluous-branches.patch | 14 ++++++++++++++
 llvm15.changes                             | 11 ++++++-----
 llvm15.spec                                | 15 +++++++++------
 3 files changed, 29 insertions(+), 11 deletions(-)
 create mode 100644 llvm-workaround-superfluous-branches.patch

diff --git a/llvm-workaround-superfluous-branches.patch b/llvm-workaround-superfluous-branches.patch
new file mode 100644
index 0000000..8f0c9dc
--- /dev/null
+++ b/llvm-workaround-superfluous-branches.patch
@@ -0,0 +1,14 @@
+diff --git a/llvm/include/llvm/Support/Casting.h b/llvm/include/llvm/Support/Casting.h
+index b6bbff8..1b68640 100644
+--- a/llvm/include/llvm/Support/Casting.h
++++ b/llvm/include/llvm/Support/Casting.h
+@@ -605,6 +605,9 @@ LLVM_NODISCARD inline decltype(auto) dyn_cast(From &Val) {
+ 
+ template <typename To, typename From>
+ LLVM_NODISCARD inline decltype(auto) dyn_cast(From *Val) {
++#if defined(__clang__) && defined(NDEBUG)
++  __builtin_assume(Val);
++#endif
+   return CastInfo<To, From *>::doCastIfPossible(Val);
+ }
+ 
diff --git a/llvm15.changes b/llvm15.changes
index adf010c..a152056 100644
--- a/llvm15.changes
+++ b/llvm15.changes
@@ -5,11 +5,12 @@ Sat Jan 14 14:06:38 UTC 2023 - Aaron Puchert <aaronpuchert@alice-dsl.net>
   * This release contains bug-fixes for the LLVM 15.0.0 release.
     This release is API and ABI compatible with 15.0.0.
 - Rebase llvm-do-not-install-static-libraries.patch.
-- Build stage 2 with -fno-plt if we're using LTO: since building
-  with -Wl,-z,now the PLT stubs are basically dead code, and
-  eliminating the indirection should improve code locality and
-  reduce BTB pressure for the quite frequent cross-DSO calls.
-  With LTO we should not need linker relaxation.
+- Build stage 2 with -fno-plt on x86_64: since building with
+  -Wl,-z,now the PLT stubs are basically dead code, so eliminating
+  the indirection reduces the number of branches and improves code
+  locality for the quite frequent cross-DSO calls.
+- Add llvm-workaround-superfluous-branches.patch: hints LLVM to
+  eliminate branches until gh#llvm/llvm-project#28804 is solved.
 
 -------------------------------------------------------------------
 Sun Dec  4 21:43:38 UTC 2022 - Aaron Puchert <aaronpuchert@alice-dsl.net>
diff --git a/llvm15.spec b/llvm15.spec
index 5c20842..262346a 100644
--- a/llvm15.spec
+++ b/llvm15.spec
@@ -367,6 +367,8 @@ Patch13:        llvm-normally-versioned-libllvm.patch
 Patch14:        llvm-do-not-install-static-libraries.patch
 # PATCH-FIX-OPENSUSE (or -UPSTREAM?): we disable RPATHs, but the test driver drops LD_LIBRARY_PATH.
 Patch15:        libcxx-test-library-path.patch
+# PATCH-FIX-UPSTREAM (?): Work around gh#llvm/llvm-project#28804 by hinting with __builtin_assume.
+Patch16:        llvm-workaround-superfluous-branches.patch
 Patch20:        llvm_build_tablegen_component_as_shared_library.patch
 Patch21:        tests-use-python3.patch
 Patch22:        llvm-better-detect-64bit-atomics-support.patch
@@ -807,6 +809,7 @@ This package contains the development files for Polly.
 %patch5 -p1
 %patch13 -p1
 %patch14 -p1
+%patch16 -p2
 %patch20 -p1
 %patch21 -p1
 %patch22 -p1
@@ -979,12 +982,12 @@ if ! ./stage1/bin/clang -c -xc -Werror -fstack-clash-protection -o /dev/null /de
 then
     flags=$(echo -n %flags | sed 's/-fstack-clash-protection//');
 fi
-# 4) Add -fno-plt: With -Wl,-z,now we don't need the PLT anymore, allowing us to
-#    reduce the number of branches for the quite frequent cross-DSO calls. This
-#    is good for code locality and reduces the pressure on the BTB.
-#    However, do this only when we're using LTO, since otherwise indirect
-#    branches have to be relaxed by the linker, which might cause regressions.
-%if %{with thin_lto}
+# 4) Add -fno-plt: With -Wl,-z,now the PLT is basically dead code, so we can
+#    now go the direct route for quite frequent cross-DSO calls. This reduces
+#    branches in a typical execution by ~5 percent, instructions/cycles
+#    by ~4 percent, and reduces pressure on the instruction cache. We do this
+#    only on x86_64 where it doesn't increase the code size significantly.
+%ifarch x86_64
 flags="$flags -fno-plt"
 %endif