Add support for compartmentalizing on ARM

package/crossfiles/aarch64-linux-clang.meson

@@ -1,5 +1,5 @@
 [binaries]
-c = 'clang'
+c = '/home/amunoz/ia2/llvm-project/build/bin/clang'
 cpp = 'clang++'
 ar = 'aarch64-linux-gnu-ar'
 strip = 'aarch64-linux-gnu-strip'

rewrite.py

@@ -195,12 +195,14 @@ def main(
             rt_libs_build_dir = custom_llvm_project / "build-rtlibs"
             cmake_cflags += [
                 "--rtlib=compiler-rt",
+                "--unwindlib=libunwind",
                 "--stdlib=libc++",
                 f"-I{str(rt_libs_build_dir / "include/c++/v1")}",
             ]
             cmake_link_flags += [
                 "--rtlib=compiler-rt",
                 "--unwindlib=libunwind",
+                "--stdlib=libc++",
                 f"-B{str(rt_libs_build_dir / "compiler-rt/lib/linux")}",
                 f"-L{str(rt_libs_build_dir /"lib")}",
             ]
@@ -239,7 +241,7 @@ def main(
             f"-DCMAKE_SHARED_LINKER_FLAGS={cmake_shared_linker_flags}",
         ]
 
-        cross_file = original_dir / "package" / "crossfiles" / f"{cross_target}.meson"
+        cross_file = original_dir / "package" / "crossfiles" / "aarch64-linux-clang.meson"
         cross = parse_machine_files(
             filenames=[str(cross_file)], sourcedir=str(original_dir)
         )
@@ -252,8 +254,8 @@ def main(
             ia2_dir,
             "-G",
             "Ninja",
-            f"-DClang_DIR={str(llvm_cmake_dir / ".." / "clang")}",
-            f"-DLLVM_DIR={str(llvm_cmake_dir)}",
+            f"-DClang_DIR=/usr/lib/llvm-14/lib/cmake/clang",
+            f"-DLLVM_DIR=/usr/lib/llvm-14/lib/cmake/llvm",
             f"-DLLVM_EXTERNAL_LIT={str(lit.executable)}",
             f"-DCMAKE_C_COMPILER={str(clang.executable)}",
             f"-DCMAKE_CXX_COMPILER={str(clang_cpp.executable)}",
@@ -292,7 +294,7 @@ def main(
         if stashed:
             git["stash", "pop"]()
 
-    clang_include_dir = find_clang_include_dir(llvm_config)
+    clang_include_dir = "/usr/lib/llvm-14/lib/clang/14.0.0/include"
 
     cc_text = cc_db.read_text()
     cmds = json.loads(cc_text)
@@ -312,13 +314,11 @@ def main(
         original_dir,
         "--output-directory",
         rewritten_dir,
-        f"--enable-dav1d_get_picture-post-condition={enable_dav1d_get_picture_post_condition}",
         "-p",
         cc_db.parent,
         *extra_args(
             "-isystem",
-            "include-fixed",
-            "-isystem",
+            "--extra-arg ",
             clang_include_dir,
         ),
         *srcs_to_rewrite,
@@ -389,10 +389,10 @@ def main(
                 "validate_input_or_ret(IA2_ADDR(s->allocator.release_picture_callback) != NULL,",
             ),
             (
-                Path("callgate_wrapper.h"),
+                src / "../../dav1d-ia2/callgate_wrapper.h",
                 {
                     TargetArch.X86_64: "struct __va_list_tag *",
-                    TargetArch.AArch64: "struct __va_list",
+                    TargetArch.AArch64: "struct std::__va_list",
                 }[target_arch],
                 "va_list",
             ),
@@ -421,25 +421,31 @@ def main(
             f"-Dia2_permissive_mode={permissive_mode}",
             f"--buildtype={dav1d_meson_build_type.value}",
         ]()
-        retcode, stdout, stderr = ninja["tools/dav1d"].run(
+        retcode, stdout, stderr = ninja["tools/dav1d", "-vvv"].run(
             # retcode=None,
             stdout=sys.stdout,
             stderr=sys.stderr,
         )
         # Path("ninja.out").write_text(stdout)
         # Path("ninja.err").write_text(stderr)
         assert retcode == 0
+        retcode, stdout, stderr = ninja["src/libdav1d.so"].run(
+            # retcode=None,
+            stdout=sys.stdout,
+            stderr=sys.stderr,
+        )
+        assert retcode == 0
         canonicalize_compile_command_paths()
 
     dav1d = rewritten_build_dir / "tools/dav1d"
     pad_tls[dav1d]()
 
-    for ldd in parse_ldd(ldd[dav1d]()):
-        padded = rpath / ldd.name
-        if padded.exists() and ldd.path.samefile(padded):
-            continue
-        shutil.copy(ldd.path, padded)
-        pad_tls["--allow-no-tls", padded]()
+    #for ldd in parse_ldd(ldd[dav1d]()):
+    #    padded = rpath / ldd.name
+    #    if padded.exists() and ldd.path.samefile(padded):
+    #        continue
+    #    shutil.copy(ldd.path, padded)
+    #    pad_tls["--allow-no-tls", padded]()
 
 
 if __name__ == "__main__":

src/arm/64/mc.S

@@ -30,6 +30,10 @@
 #include "util.S"
 
 .macro avg dst, t0, t1, t2, t3
+        extr x2, x2, x18, #56
+        extr x2, x2, x2, #8
+        extr x3, x3, x18, #56
+        extr x3, x3, x3, #8
         ld1             {\t0\().8h,\t1\().8h},   [x2],  32
         ld1             {\t2\().8h,\t3\().8h},   [x3],  32
         add             \t0\().8h,   \t0\().8h,   \t2\().8h
@@ -39,6 +43,10 @@
 .endm
 
 .macro w_avg dst, t0, t1, t2, t3
+        extr x2, x2, x18, #56
+        extr x2, x2, x2, #8
+        extr x3, x3, x18, #56
+        extr x3, x3, x3, #8
         ld1             {\t0\().8h,\t1\().8h},   [x2],  32
         ld1             {\t2\().8h,\t3\().8h},   [x3],  32
         sub             \t0\().8h,   \t2\().8h,   \t0\().8h
@@ -52,6 +60,12 @@
 .endm
 
 .macro mask dst, t0, t1, t2, t3
+        extr x6, x6, x18, #56
+        extr x6, x6, x6, #8
+        extr x2, x2, x18, #56
+        extr x2, x2, x2, #8
+        extr x3, x3, x18, #56
+        extr x3, x3, x3, #8
         ld1             {v30.16b}, [x6],  16
         ld1             {\t0\().8h,\t1\().8h},   [x2],  32
         mul             v30.16b, v30.16b, v31.16b
@@ -436,6 +450,8 @@ w_mask_fn 420
 
 
 function blend_8bpc_neon, export=1
+        extr x5, x5, x18, #56
+        extr x5, x5, x5, #8
         movrel          x6,  blend_tbl
         clz             w3,  w3
         sub             w3,  w3,  #26

src/arm/64/mc16.S

@@ -118,6 +118,8 @@
 
 .macro bidir_fn type, bdmax
 function \type\()_16bpc_neon, export=1
+        extr x6, x6, x18, #56
+        extr x6, x6, x6, #8
         clz             w4,  w4
 .ifnc \type, avg
         dup             v31.8h,  \bdmax // bitdepth_max
@@ -565,6 +567,8 @@ w_mask_fn 420
 
 
 function blend_16bpc_neon, export=1
+        extr x5, x5, x18, #56
+        extr x5, x5, x5, #8
         movrel          x6,  blend_tbl
         clz             w3,  w3
         sub             w3,  w3,  #26

src/arm/asm.S

@@ -31,8 +31,6 @@
 #include "config.h"
 
 #if ARCH_AARCH64
-#define x18 do_not_use_x18
-#define w18 do_not_use_w18
 
 #if HAVE_AS_ARCH_DIRECTIVE
         .arch AS_ARCH_LEVEL

src/arm/cpu.c

@@ -42,7 +42,7 @@
 #define HWCAP2_AARCH64_SVE2   (1 << 1)
 #define HWCAP2_AARCH64_I8MM   (1 << 13)
 
-COLD unsigned dav1d_get_cpu_flags_arm(void) {
+DAV1D_API COLD unsigned dav1d_get_cpu_flags_arm(void) {
 #if HAVE_GETAUXVAL
     unsigned long hw_cap = getauxval(AT_HWCAP);
     unsigned long hw_cap2 = getauxval(AT_HWCAP2);
@@ -68,7 +68,7 @@ COLD unsigned dav1d_get_cpu_flags_arm(void) {
 #define HWCAP_ARM_ASIMDDP (1 << 24)
 #define HWCAP_ARM_I8MM    (1 << 27)
 
-COLD unsigned dav1d_get_cpu_flags_arm(void) {
+DAV1D_API COLD unsigned dav1d_get_cpu_flags_arm(void) {
 #if HAVE_GETAUXVAL
     unsigned long hw_cap = getauxval(AT_HWCAP);
 #else
@@ -96,7 +96,7 @@ static int have_feature(const char *feature) {
     return supported;
 }
 
-COLD unsigned dav1d_get_cpu_flags_arm(void) {
+DAV1D_API COLD unsigned dav1d_get_cpu_flags_arm(void) {
     unsigned flags = dav1d_get_default_cpu_flags();
     if (have_feature("hw.optional.arm.FEAT_DotProd"))
         flags |= DAV1D_ARM_CPU_FLAG_DOTPROD;
@@ -112,7 +112,7 @@ COLD unsigned dav1d_get_cpu_flags_arm(void) {
 #include <sys/types.h>
 #include <sys/sysctl.h>
 
-COLD unsigned dav1d_get_cpu_flags_arm(void) {
+DAV1D_API COLD unsigned dav1d_get_cpu_flags_arm(void) {
      unsigned flags = dav1d_get_default_cpu_flags();
 
 #ifdef CPU_ID_AA64ISAR0
@@ -146,7 +146,7 @@ COLD unsigned dav1d_get_cpu_flags_arm(void) {
 #elif defined(_WIN32)
 #include <windows.h>
 
-COLD unsigned dav1d_get_cpu_flags_arm(void) {
+DAV1D_API COLD unsigned dav1d_get_cpu_flags_arm(void) {
     unsigned flags = dav1d_get_default_cpu_flags();
 #ifdef PF_ARM_V82_DP_INSTRUCTIONS_AVAILABLE
     if (IsProcessorFeaturePresent(PF_ARM_V82_DP_INSTRUCTIONS_AVAILABLE))
@@ -211,7 +211,7 @@ static unsigned parse_proc_cpuinfo(const char *flag) {
     return 0;
 }
 
-COLD unsigned dav1d_get_cpu_flags_arm(void) {
+DAV1D_API COLD unsigned dav1d_get_cpu_flags_arm(void) {
     unsigned flags = dav1d_get_default_cpu_flags();
     flags |= parse_proc_cpuinfo("neon") ? DAV1D_ARM_CPU_FLAG_NEON : 0;
     flags |= parse_proc_cpuinfo("asimd") ? DAV1D_ARM_CPU_FLAG_NEON : 0;
@@ -226,7 +226,7 @@ COLD unsigned dav1d_get_cpu_flags_arm(void) {
 
 #else  /* Unsupported OS */
 
-COLD unsigned dav1d_get_cpu_flags_arm(void) {
+DAV1D_API COLD unsigned dav1d_get_cpu_flags_arm(void) {
     return dav1d_get_default_cpu_flags();
 }
 

src/meson.build

@@ -320,10 +320,17 @@ link_args = []
 
 if ia2_path != ''
     libdav1d_flags += [
+        '-O3',
+        '-ffixed-x18',
+        '-march=armv8.5-a+memtag',
+        '-B/home/amunoz/ia2/llvm-project/build-rtlibs/compiler-rt/lib/linux',
+        '-L/home/amunoz/ia2/llvm-project/build-rtlibs/lib',
+        '--rtlib=compiler-rt',
         f'-DPKEY=@pkey@',
     ]
     if ia2_enable
         link_args += [
+            '--rtlib=compiler-rt',
             '-Wl,@' + join_paths(dav1d_src_root, f'callgate_wrapper_@pkey@.ld'),
         ]
     endif

tools/meson.build

@@ -46,11 +46,21 @@ link_args = []
 
 if ia2_path != ''
     c_args += [
+        '-O3',
+        '-g',
+        '-ffixed-x18',
+        '-march=armv8.5-a+memtag',
+        '-B/home/amunoz/ia2/llvm-project/build-rtlibs/compiler-rt/lib/linux',
+        '-L/home/amunoz/ia2/llvm-project/build-rtlibs/lib',
+        '--rtlib=compiler-rt',
         f'-DPKEY=@pkey@',
     ]
     if ia2_enable
         link_args += [
             '-Wl,@' + join_paths(dav1d_src_root, f'callgate_wrapper_@pkey@.ld'),
+            '-Wl,--dynamic-linker=/tmp/build/aarch64/tools/ld-linux-aarch64.so.1',
+            '-Wl,--rpath=/tmp/build/aarch64/tools/',
+            '--rtlib=compiler-rt',
             '-Wl,--wrap=main',
             '-Wl,--dynamic-list=' + join_paths(ia2_path, 'runtime/libia2/dynsym.syms'),
             '-Wl,--export-dynamic',
@@ -77,9 +87,6 @@ if ia2_path != ''
             # <stdio.h> functions
             '-Wl,-wrap,asprintf',
             '-Wl,-wrap,vasprintf',
-
-            # for `__tls_get_addr@@GLIBC_2.3`, which didn't use to error for being undefined, but now does
-            '/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2',
         ]
     endif
 endif