PowerPC {32, 64}-bit Block Trampolines (#272)

* Implement PowerPC block trampoline

* Adjust pagesize on ppc64

* Skip UnexpectedException test for PowerPC

* Move PAGE_SIZE to asmconstants.h

* Use PAGE_SIZE and PAGE_SHIFT macros for PowerPC

* Add ppc64el and powerpc qemu-crossbuild targets

* Add NO_SAFE_CACHING definition and guards

* Do not export objc_method_cache_version on ppc32

---------

Co-authored-by: David Chisnall <davidchisnall@users.noreply.github.com>

.github/workflows/main.yml

@@ -88,6 +88,10 @@ jobs:
             system-processor: riscv64
             triple: riscv64-linux-gnu
             rtld: ld-linux-riscv64-lp64d.so.1
+          - name: ppc64el
+            system-processor: powerpc64le
+            triple: powerpc64le-linux-gnu
+            rtld: ld64.so.2
         # lld versions prior to 15 do not support R_RISCV_ALIGN relocations
         exclude:
           - llvm-version: 13
@@ -108,7 +112,7 @@ jobs:
           sudo apt install libstdc++-9-dev-${{ matrix.arch.name }}-cross qemu-user ninja-build
       - name: Configure CMake
         run: |
-          export LDFLAGS="-L/usr/lib/llvm-${{ matrix.llvm-version }}/lib/ -fuse-ld=lld -Wl,--dynamic-linker=/usr/${{ matrix.arch.triple }}/lib/${{ matrix.arch.rtld }},-rpath,/usr/${{ matrix.arch.triple }}/lib"
+          export LDFLAGS="-L/usr/lib/llvm-${{ matrix.llvm-version }}/lib/ -fuse-ld=lld-${{ matrix.llvm-version}} -Wl,--dynamic-linker=/usr/${{ matrix.arch.triple }}/lib/${{ matrix.arch.rtld }},-rpath,/usr/${{ matrix.arch.triple }}/lib"
           cmake -B ${{github.workspace}}/build \
             -DCMAKE_SYSTEM_NAME=Linux \
             -DCMAKE_SYSTEM_PROCESSOR=${{ matrix.arch.system-processor }} \

CMake/detect_arch.c

@@ -7,6 +7,10 @@
 #error i386
 #elif defined(__x86_64__)
 #error x86_64
+#elif defined(__powerpc64__)
+#error powerpc64
+#elif defined(__powerpc__)
+#error powerpc
 #else
 #error unknown
 #endif

CMakeLists.txt

@@ -51,7 +51,7 @@ try_compile(
 )
 
 if(NOT COMPILE_SUCCESS)
-    string(REGEX MATCH "(aarch64|arm|i386|x86_64|unknown)" ARCHITECTURE ${COMPILE_OUTPUT})
+    string(REGEX MATCH "(aarch64|arm|i386|x86_64|powerpc64|powerpc|unknown)" ARCHITECTURE ${COMPILE_OUTPUT})
 endif()
 
 set(ARCHITECTURE ${ARCHITECTURE} CACHE STRING "Architecture Type")
@@ -187,6 +187,14 @@ set(INCLUDE_DIRECTORY "objc" CACHE STRING
 
 add_compile_options($<$<STREQUAL:${CMAKE_SYSTEM_PROCESSOR},i686>:-march=i586>)
 
+# PowerPC 32-bit does not support native 64-bit atomic operations,
+# which is used in safe caching.
+# You must also update the guard in objc/runtime.h, when updating
+# this macro.
+if (ARCHITECTURE STREQUAL "powerpc")
+	add_definitions(-DNO_SAFE_CACHING)
+endif()
+
 set(INSTALL_TARGETS objc)
 
 if(WIN32)

Test/UnexpectedException.m

@@ -31,7 +31,7 @@ LONG WINAPI _UnhandledExceptionFilter(struct _EXCEPTION_POINTERS* exceptionInfo)
 
 int main(void)
 {
-#if !(defined(__arm__) || defined(__ARM_ARCH_ISA_A64))
+#if !(defined(__arm__) || defined(__ARM_ARCH_ISA_A64)) && !defined(__powerpc__)
 #if defined(_WIN32) && !defined(__MINGW32__)
 	// also verify that an existing handler still gets called after we set ours
 	SetUnhandledExceptionFilter(&_UnhandledExceptionFilter);

asmconstants.h

@@ -19,3 +19,11 @@
 #define SLOT_OFFSET    0
 #endif
 #define SMALLOBJ_MASK  ((1<<SMALLOBJ_BITS) - 1)
+
+// Page size configuration
+#if defined(__powerpc64__)
+#	define PAGE_SHIFT 16
+#else
+#	define PAGE_SHIFT 12
+#endif
+#define PAGE_SIZE (1<<PAGE_SHIFT)

block_to_imp.c

@@ -20,6 +20,7 @@
 #include "blocks_runtime.h"
 #include "lock.h"
 #include "visibility.h"
+#include "asmconstants.h" // For PAGE_SIZE
 
 #ifndef __has_builtin
 #define __has_builtin(x) 0
@@ -95,22 +96,26 @@ static int mprotect(void *buffer, size_t len, int prot)
 #	endif
 #endif
 
-#define PAGE_SIZE 4096
-
 struct block_header
 {
 	void *block;
 	void(*fnptr)(void);
 	/**
 	 * On 64-bit platforms, we have 16 bytes for instructions, which ought to
-	 * be enough without padding.  On MIPS, we need 
+	 * be enough without padding.
 	 * Note: If we add too much padding, then we waste space but have no other
 	 * ill effects.  If we get this too small, then the assert in
 	 * `init_trampolines` will fire on library load.
+	 *
+	 * PowerPC: We need INSTR_CNT * INSTR_LEN = 7*4 = 28 bytes
+	 * for instruction. sizeof(block_header) must be a divisor of
+	 * PAGE_SIZE, so we need to pad block_header to 32 bytes.
+	 * On PowerPC 64-bit where sizeof(void *) = 8 bytes, we
+	 * add 16 bytes of padding.
 	 */
-#if defined(__i386__) || (defined(__mips__) && !defined(__mips_n64))
+#if defined(__i386__) || (defined(__mips__) && !defined(__mips_n64)) || (defined(__powerpc__) && !defined(__powerpc64__))
 	uint64_t padding[3];
-#elif defined(__mips__)
+#elif defined(__mips__) || defined(__powerpc64__)
 	uint64_t padding[2];
 #elif defined(__arm__)
 	uint64_t padding;

block_trampolines.S

@@ -1,4 +1,5 @@
 #include "common.S"
+#include "asmconstants.h"
 
 #
 # This file defines some trampolines for calling blocks.  A block function
@@ -98,6 +99,45 @@
 #define ARG1 $a1
 #define ARG2 $a2
 
+#elif defined(__powerpc__)
+////////////////////////////////////////////////////////////////////////////////
+// PowerPC trampoline
+////////////////////////////////////////////////////////////////////////////////
+
+#if defined(__powerpc64__)
+#define LOAD ld
+#define OFFSET 8
+#else
+#define LOAD lwz
+#define OFFSET 4
+#endif
+
+.macro trampoline arg0, arg1
+	mfctr %r12                   # The block trampoline is always called
+                                 # via a function pointer. We can thus
+                                 # assume that ctr contains the trampline
+                                 # entry point address from the previous
+                                 # branch to this trampoline (bctrl).
+
+	#if PAGE_SHIFT < 16
+	addi  %r12, %r12, -PAGE_SIZE # Substract page size from entry point
+	#else
+	addis %r12, %r12, (-0x1 << (PAGE_SHIFT - 16))
+	#endif
+
+	mr    \arg1, \arg0
+	LOAD  \arg0, 0(%r12)
+	LOAD  %r12, OFFSET(%r12)
+	mtctr %r12                   # Move block function pointer into ctr
+	bctr                         # Branch to block function
+.endm
+
+#define ARG0 %r3
+#define ARG1 %r4
+#define ARG2 %r5
+#define SARG0 ARG1
+#define SARG1 ARG2
+
 #elif defined(__riscv) && (__riscv_xlen == 64)
 ////////////////////////////////////////////////////////////////////////////////
 // RISC-V trampoline

dtable.c

@@ -43,7 +43,9 @@ PRIVATE mutex_t initialize_lock;
  * 2^x in increments of 8. */
 static uint32_t dtable_depth = 8;
 
+#ifndef NO_SAFE_CACHING
 _Atomic(uint64_t) objc_method_cache_version;
+#endif
 
 /**
  * Starting at `cls`, finds the class that provides the implementation of the
@@ -404,7 +406,9 @@ static BOOL installMethodInDtable(Class class,
 	// Invalidate the old slot, if there is one.
 	if (NULL != oldMethod)
 	{
+#ifndef NO_SAFE_CACHING
 		objc_method_cache_version++;
+#endif
 	}
 	return YES;
 }
@@ -520,7 +524,9 @@ PRIVATE void objc_update_dtable_for_new_superclass(Class cls, Class newSuper)
 	LOCK_RUNTIME_FOR_SCOPE();
 	rebaseDtableRecursive(cls, newSuper);
 	// Invalidate all caches after this operation.
-	objc_method_cache_version++;
+#ifndef NO_SAFE_CACHING
+		objc_method_cache_version++;
+#endif
 
 	return;
 }

objc/slot.h

@@ -27,8 +27,14 @@ struct objc_slot2
  * A counter that is incremented whenever one or more cached slots become
  * invalid, for example if a subclass loads a category containing methods that
  * were inherited from the superclass.
+ *
+ * Caching is disabled on targets without native 64-bit atomics support such
+ * as PowerPC 32-bit.
  */
+#if defined(__powerpc__) && !defined(__powerpc64__)
+#else
 OBJC_PUBLIC extern _Atomic(uint64_t) objc_method_cache_version;
+#endif
 
 /**
  * Legacy cache structure.  This is no longer maintained in the runtime and is

sendmsg2.c

@@ -93,7 +93,12 @@ struct objc_slot2 *objc_msg_lookup_internal(id *receiver, SEL selector, uint64_t
 {
 	if (version)
 	{
+#ifdef NO_SAFE_CACHING
+		// Always write 0 to version, marking the slot as uncacheable.
+		*version = 0;
+#else
 		*version = objc_method_cache_version;
+#endif
 	}
 	Class class = classForObject((*receiver));
 retry:;
@@ -118,10 +123,12 @@ retry:;
 		{
 			if ((result = objc_dtable_lookup(dtable, get_untyped_idx(selector))))
 			{
+#ifndef NO_SAFE_CACHING
 				if (version)
 				{
 					*version = 0;
 				}
+#endif
 				uncacheable_slot.imp = call_mismatch_hook(class, selector, result);
 				result = (struct objc_slot2*)&uncacheable_slot;
 			}
@@ -135,10 +142,12 @@ retry:;
 			}
 			if (0 == result)
 			{
+#ifndef NO_SAFE_CACHING
 				if (version)
 				{
 					*version = 0;
 				}
+#endif
 				uncacheable_slot.imp = __objc_msg_forward2(*receiver, selector);
 				result = (struct objc_slot2*)&uncacheable_slot;
 			}
@@ -236,10 +245,12 @@ struct objc_slot2 *objc_slot_lookup_version(id *receiver, SEL selector, uint64_t
 	// inlined trivially.
 	if (UNLIKELY(*receiver == nil))
 	{
+#ifndef NO_SAFE_CACHING
 		if (version)
 		{
 			*version = 0;
 		}
+#endif
 		// Return the correct kind of zero, depending on the type encoding.
 		if (selector->types)
 		{
@@ -350,10 +361,12 @@ struct objc_slot *objc_slot_lookup_super(struct objc_super *super, SEL selector)
  */
 struct objc_slot2 *objc_get_slot2(Class cls, SEL selector, uint64_t *version)
 {
+#ifndef NO_SAFE_CACHING
 	if (version)
 	{
 		*version = objc_method_cache_version;
 	}
+#endif
 	struct objc_slot2 * result = objc_dtable_lookup(cls->dtable, selector->index);
 	if (0 == result)
 	{
@@ -374,10 +387,12 @@ struct objc_slot2 *objc_get_slot2(Class cls, SEL selector, uint64_t *version)
 		{
 			if ((result = objc_dtable_lookup(dtable, get_untyped_idx(selector))))
 			{
+#ifndef NO_SAFE_CACHING
 				if (version)
 				{
 					*version = 0;
 				}
+#endif
 				uncacheable_slot.imp = call_mismatch_hook(cls, selector, result);
 				result = (struct objc_slot2*)&uncacheable_slot;
 			}