Skip to contentSkip to conversation

clang: Space optimize clang recipes

PATCH In review Khem Raj 2 participants 3 messages May 7, 2025 – May 9, 2025

lore
Share & tools
lore

Khem Raj

author
#0 lore raw

PATCH v3 1/2 clang: Space optimize clang recipes

They can consume quite a bit of build space, we don't debug them as often

---
v2: No change
v3: No change

 meta/conf/distro/include/yocto-space-optimize.inc | 4 +++-
 meta/recipes-devtools/clang/clang_git.bb          | 2 --
 2 files changed, 3 insertions(+), 3 deletions(-)
meta/conf/distro/include/yocto-space-optimize.inc +3 -1
index 407e8937f08..b83ba426e78 100644
@@ -3,6 +3,8 @@
 # and build speed.
 
 # With -g, llvm workdir is 40GB, llvm-staticdev is 1.3GB, llvm-dbg is 1GB
-DEBUG_LEVELFLAG:pn-llvm = ""
+DEBUG_LEVELFLAG:pn-clang = ""
+DEBUG_LEVELFLAG:pn-clang-native = ""
+DEBUG_LEVELFLAG:pn-nativesdk-clang = ""
 # With -g, qemu workdir is 10GB, qemu-dbg is 363MB
 DEBUG_LEVELFLAG:pn-qemu = ""
meta/recipes-devtools/clang/clang_git.bb +0 -2
index ba037e59785..c93554622f1 100644
@@ -153,8 +153,6 @@ EXTRA_OECMAKE += "-DLLVM_ENABLE_ASSERTIONS=OFF \
                   -DLLVM_BUILD_EXTERNAL_COMPILER_RT=ON \
                   -DCMAKE_SYSTEM_NAME=Linux \
                   -DCMAKE_BUILD_TYPE=Release \
-                  -DCMAKE_CXX_FLAGS_RELEASE='${CXXFLAGS} -DNDEBUG -g0' \
-                  -DCMAKE_C_FLAGS_RELEASE='${CFLAGS} -DNDEBUG -g0' \
                   -DLLVM_ENABLE_PROJECTS='${LLVM_PROJECTS}' \
                   -DLLVM_BINUTILS_INCDIR=${STAGING_INCDIR} \
                   -DLLVM_VERSION_SUFFIX='${VER_SUFFIX}' \
author
#1 lore raw

PATCH v3 2/2 clang: Build limited targets for target clang recipe

On target its sufficient to build what mesa needs + one host target It should speed up the build.

Fix bug in function mapping target architecture to LLVM backend

Reduces size of libLLVM

Before - 131M May 6 20:42 ./clang-libllvm/usr/lib/libLLVM.so.20.1 After - 93M May 6 18:47 ./clang-libllvm/usr/lib/libLLVM.so.20.1

---
v2: Fix bug in function mapping target architecture to LLVM backend
v3: Use HOST_ARCH instead of get_clang_arch()

 meta/recipes-devtools/clang/clang_git.bb | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)
meta/recipes-devtools/clang/clang_git.bb +9 -5
index c93554622f1..b41b700292c 100644
@@ -35,10 +35,10 @@ def get_clang_arch(bb, d, arch_var):
     elif re.match('aarch64$', a):                      return 'AArch64'
     elif re.match('aarch64_be$', a):                   return 'AArch64'
     elif re.match('mips(isa|)(32|64|)(r6|)(el|)$', a): return 'Mips'
-    elif re.match('riscv32$', a):                      return 'riscv32'
-    elif re.match('riscv64$', a):                      return 'riscv64'
+    elif re.match('riscv32$', a):                      return 'RISCV'
+    elif re.match('riscv64$', a):                      return 'RISCV'
     elif re.match('p(pc|owerpc)(|64)', a):             return 'PowerPC'
-    elif re.match('loongarch64$', a):                  return 'loongarch64'
+    elif re.match('loongarch64$', a):                  return 'LoongArch'
     else:
         bb.note("'%s' is not a primary llvm architecture" % a)
     return ""
@@ -115,8 +115,12 @@ LLVM_BUILD_TOOLS;LLVM_USE_HOST_TOOLS;LLVM_CONFIG_PATH;LLVM_EXTERNAL_SPIRV_HEADER
 # Default to build all OE-Core supported target arches (user overridable).
 # Gennerally setting LLVM_TARGETS_TO_BUILD = "" in local.conf is ok in most simple situations
 # where only one target architecture is needed along with just one build arch (usually X86)
-#
+# Core tier targets:
+# AArch64;AMDGPU;ARM;AVR;BPF;Hexagon;Lanai;LoongArch;Mips;MSP430;NVPTX;PowerPC;RISCV;Sparc;SPIRV;SystemZ;VE;WebAssembly;X86;XCore
+# Known experimental targets: ARC;CSKY;DirectX;M68k;Xtensa
+
 LLVM_TARGETS_TO_BUILD ?= "AMDGPU;AArch64;ARM;BPF;Mips;PowerPC;RISCV;X86;LoongArch;NVPTX;SPIRV"
+LLVM_TARGETS_TO_BUILD:class-target ?= "${@get_clang_host_arch(bb, d)};AMDGPU;BPF;NVPTX;SPIRV"
 
 LLVM_EXPERIMENTAL_TARGETS_TO_BUILD ?= ""
 
@@ -189,7 +193,7 @@ EXTRA_OECMAKE:append:class-target = "\
                   -DCMAKE_AR=${STAGING_BINDIR_TOOLCHAIN}/${TARGET_PREFIX}llvm-ar \
                   -DCMAKE_NM=${STAGING_BINDIR_TOOLCHAIN}/${TARGET_PREFIX}llvm-nm \
                   -DCMAKE_STRIP=${STAGING_BINDIR_TOOLCHAIN}/${TARGET_PREFIX}llvm-strip \
-                  -DLLVM_TARGET_ARCH=${@get_clang_target_arch(bb, d)} \
+                  -DLLVM_TARGET_ARCH=${HOST_ARCH} \
                   -DLLVM_DEFAULT_TARGET_TRIPLE=${TARGET_SYS}${HF} \
                   -DLLVM_HOST_TRIPLE=${TARGET_SYS}${HF} \
                   -DPYTHON_LIBRARY=${STAGING_LIBDIR}/lib${PYTHON_DIR}${PYTHON_ABI}.so \
#2 lore raw
In reply to Khem Raj
Quoted 1 line · Khem Raj via lists.openembedded.org

They can consume quite a bit of build space, we don't debug them as often

I did get some measurements before and after these two were applied:

https://valkyrie.yocto.io/pub/non-release/20250509-59/testresults/buildperf-debian11/perf-debian12-vk_master_20250509082553_2859d5fb9d.html
https://valkyrie.yocto.io/pub/non-release/20250509-58/testresults/buildperf-alma8/perf-alma8-vk_master_20250509082530_2859d5fb9d.html

Sadly the improvements aren't that large, particularly considering the original impact of clang.

Cheers,

Richard