Fix and improve coroutines for Darwin (macOS) ppc/ppc64. (#5975)

Merged-By: ioquatix <samuel@codeotaku.com>

addr2line.c

@@ -869,8 +869,11 @@ typedef struct {
     int type;
 } DebugInfoValue;
 
-/* TODO: Big Endian */
+#if defined(WORDS_BIGENDIAN)
+#define MERGE_2INTS(a,b,sz) (((uint64_t)(a)<<sz)|(b))
+#else
 #define MERGE_2INTS(a,b,sz) (((uint64_t)(b)<<sz)|(a))
+#endif
 
 static uint16_t
 get_uint16(const uint8_t *p)

configure.ac

@@ -1902,8 +1902,8 @@ AS_CASE(["${target_cpu}-${target_os}:${target_archs}"],
 [universal-darwin*:*ppc*], [
   AC_LIBSOURCES(alloca.c)
   AC_SUBST([ALLOCA], [\${LIBOBJDIR}alloca.${ac_objext}])
-  RUBY_DEFINE_IF([defined __powerpc__], C_ALLOCA, 1)
+  RUBY_DEFINE_IF([defined __POWERPC__], C_ALLOCA, 1) # Darwin defines __POWERPC__ for ppc and ppc64 both
-  RUBY_DEFINE_IF([defined __powerpc__], alloca, alloca)
+  RUBY_DEFINE_IF([defined __POWERPC__], alloca, alloca)
   ],
 [
   AC_FUNC_ALLOCA
@@ -2573,10 +2573,13 @@ AS_CASE([$coroutine_type], [yes|''], [
         [arm64-darwin*], [
             coroutine_type=arm64
         ],
-        [powerpc-darwin*], [
+        # Correct target name is powerpc*-, but Ruby seems to prefer ppc*-.
+        # Notice that Darwin PPC ABI differs from AIX and ELF.
+        # Adding PPC targets for AIX, *BSD and *Linux will require separate implementations.
+        [powerpc-darwin*|ppc-darwin*], [
             coroutine_type=ppc
         ],
-        [powerpc64-darwin*], [
+        [powerpc64-darwin*|ppc64-darwin*], [
             coroutine_type=ppc64
         ],
         [x*64-linux*], [

coroutine/ppc/Context.S

@@ -1,73 +1,90 @@
+; Based on the code by Samuel Williams. Created by Sergey Fedorov on 04/06/2022.
+; Credits to Samuel Williams, Rei Odaira and Iain Sandoe. Errors, if any, are mine.
+; Some relevant examples: https://github.com/gcc-mirror/gcc/blob/master/libphobos/libdruntime/config/powerpc/switchcontext.S
+; https://github.com/gcc-mirror/gcc/blob/master/libgcc/config/rs6000/darwin-gpsave.S
+; https://www.ibm.com/docs/en/aix/7.2?topic=epilogs-saving-gprs-only
+; ppc32 version may be re-written compactly with stmw/lwm, but the code wonʼt be faster, see: https://github.com/ruby/ruby/pull/5927#issuecomment-1139730541
+
+; Notice that this code is only for Darwin (macOS). Darwin ABI differs from AIX and ELF.
+; To add support for AIX, *BSD or *Linux, please make separate implementations.
+
 #define TOKEN_PASTE(x,y) x##y
 #define PREFIXED_SYMBOL(prefix,name) TOKEN_PASTE(prefix,name)
 
+.machine ppc7400 ; = G4, Rosetta
 .text
-.align 2
 
 .globl PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer)
+.align 2
+
 PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer):
-	# Make space on the stack for caller registers
+	; Make space on the stack for caller registers
-	addi r1,r1,-80
+	; (Should we rather use red zone? See libphobos example.)
+	subi r1,r1,80
 
-	# Save caller registers
+	; Get LR
-	stw r13,0(r1)
-	stw r14,4(r1)
-	stw r15,8(r1)
-	stw r16,12(r1)
-	stw r17,16(r1)
-	stw r18,20(r1)
-	stw r19,24(r1)
-	stw r20,28(r1)
-	stw r21,32(r1)
-	stw r22,36(r1)
-	stw r23,40(r1)
-	stw r24,44(r1)
-	stw r25,48(r1)
-	stw r26,52(r1)
-	stw r27,56(r1)
-	stw r28,60(r1)
-	stw r29,64(r1)
-	stw r30,68(r1)
-	stw r31,72(r1)
-
-	# Save return address
 	mflr r0
+
+	; Save caller registers
+	stw r31,0(r1)
+	stw r30,4(r1)
+	stw r29,8(r1)
+	stw r28,12(r1)
+	stw r27,16(r1)
+	stw r26,20(r1)
+	stw r25,24(r1)
+	stw r24,28(r1)
+	stw r23,32(r1)
+	stw r22,36(r1)
+	stw r21,40(r1)
+	stw r20,44(r1)
+	stw r19,48(r1)
+	stw r18,52(r1)
+	stw r17,56(r1)
+	stw r16,60(r1)
+	stw r15,64(r1)
+	stw r14,68(r1)
+	stw r13,72(r1)
+
+	; Save return address
+	; Possibly should rather be saved into linkage area, see libphobos and IBM docs
 	stw r0,76(r1)
 
-	# Save stack pointer to first argument
+	; Save stack pointer to first argument
 	stw r1,0(r3)
 
-	# Load stack pointer from second argument
+	; Load stack pointer from second argument
 	lwz r1,0(r4)
 
-	# Restore caller registers
+	; Load return address
-	lwz r13,0(r1)
-	lwz r14,4(r1)
-	lwz r15,8(r1)
-	lwz r16,12(r1)
-	lwz r17,16(r1)
-	lwz r18,20(r1)
-	lwz r19,24(r1)
-	lwz r20,28(r1)
-	lwz r21,32(r1)
-	lwz r22,36(r1)
-	lwz r23,40(r1)
-	lwz r24,44(r1)
-	lwz r25,48(r1)
-	lwz r26,52(r1)
-	lwz r27,56(r1)
-	lwz r28,60(r1)
-	lwz r29,64(r1)
-	lwz r30,68(r1)
-	lwz r31,72(r1)
-
-	# Load return address
 	lwz r0,76(r1)
+
+	; Restore caller registers
+	lwz r13,72(r1)
+	lwz r14,68(r1)
+	lwz r15,64(r1)
+	lwz r16,60(r1)
+	lwz r17,56(r1)
+	lwz r18,52(r1)
+	lwz r19,48(r1)
+	lwz r20,44(r1)
+	lwz r21,40(r1)
+	lwz r22,36(r1)
+	lwz r23,32(r1)
+	lwz r24,28(r1)
+	lwz r25,24(r1)
+	lwz r26,20(r1)
+	lwz r27,16(r1)
+	lwz r28,12(r1)
+	lwz r29,8(r1)
+	lwz r30,4(r1)
+	lwz r31,0(r1)
+
+	; Set LR
 	mtlr r0
 
-	# Pop stack frame
+	; Pop stack frame
 	addi r1,r1,80
 
-	# Jump to return address
+	; Jump to return address
 	blr
-

coroutine/ppc/Context.h

@@ -9,6 +9,7 @@
 #include <string.h>
 
 #define COROUTINE __attribute__((noreturn)) void
+#define COROUTINE_LIMITED_ADDRESS_SPACE
 
 enum {
   COROUTINE_REGISTERS =

coroutine/ppc64/Context.S

@@ -1,70 +1,89 @@
+; Based on the code by Samuel Williams. Created by Sergey Fedorov on 04/06/2022.
+; Credits to Samuel Williams, Rei Odaira and Iain Sandoe. Errors, if any, are mine.
+; Some relevant examples: https://github.com/gcc-mirror/gcc/blob/master/libphobos/libdruntime/config/powerpc/switchcontext.S
+; https://github.com/gcc-mirror/gcc/blob/master/libgcc/config/rs6000/darwin-gpsave.S
+; https://www.ibm.com/docs/en/aix/7.2?topic=epilogs-saving-gprs-only
+
+; Notice that this code is only for Darwin (macOS). Darwin ABI differs from AIX and ELF.
+; To add support for AIX, *BSD or *Linux, please make separate implementations.
+
 #define TOKEN_PASTE(x,y) x##y
 #define PREFIXED_SYMBOL(prefix,name) TOKEN_PASTE(prefix,name)
 
+.machine ppc64 ; = G5
 .text
-.align 3
 
 .globl PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer)
+.align 2
+
 PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer):
-	# Make space on the stack for caller registers
+	; Make space on the stack for caller registers
-	addi r1,r1,-152
+	; (Should we rather use red zone? See libphobos example.)
+	subi r1,r1,160
 
-	# Save caller registers
+	; Get LR
-	std r14,0(r1)
-	std r15,8(r1)
-	std r16,16(r1)
-	std r17,24(r1)
-	std r18,32(r1)
-	std r19,40(r1)
-	std r20,48(r1)
-	std r21,56(r1)
-	std r22,64(r1)
-	std r23,72(r1)
-	std r24,80(r1)
-	std r25,88(r1)
-	std r26,96(r1)
-	std r27,104(r1)
-	std r28,112(r1)
-	std r29,120(r1)
-	std r30,128(r1)
-	std r31,136(r1)
-
-	# Save return address
 	mflr r0
-	std r0,144(r1)
 
-	# Save stack pointer to first argument
+	; Save caller registers
+	std r31,0(r1)
+	std r30,8(r1)
+	std r29,16(r1)
+	std r28,24(r1)
+	std r27,32(r1)
+	std r26,40(r1)
+	std r25,48(r1)
+	std r24,56(r1)
+	std r23,64(r1)
+	std r22,72(r1)
+	std r21,80(r1)
+	std r20,88(r1)
+	std r19,96(r1)
+	std r18,104(r1)
+	std r17,112(r1)
+	std r16,120(r1)
+	std r15,128(r1)
+	std r14,136(r1)
+	std r13,144(r1)
+
+	; Save return address
+	; Possibly should rather be saved into linkage area, see libphobos and IBM docs
+	std r0,152(r1)
+
+	; Save stack pointer to first argument
 	std r1,0(r3)
 
-	# Load stack pointer from second argument
+	; Load stack pointer from second argument
 	ld r1,0(r4)
 
-	# Restore caller registers
+	; Load return address
-	ld r14,0(r1)
+	ld r0,152(r1)
-	ld r15,8(r1)
-	ld r16,16(r1)
-	ld r17,24(r1)
-	ld r18,32(r1)
-	ld r19,40(r1)
-	ld r20,48(r1)
-	ld r21,56(r1)
-	ld r22,64(r1)
-	ld r23,72(r1)
-	ld r24,80(r1)
-	ld r25,88(r1)
-	ld r26,96(r1)
-	ld r27,104(r1)
-	ld r28,112(r1)
-	ld r29,120(r1)
-	ld r30,128(r1)
-	ld r31,136(r1)
 
-	# Load return address
+	; Restore caller registers
-	ld r0,144(r1)
+	ld r13,144(r1)
+	ld r14,136(r1)
+	ld r15,128(r1)
+	ld r16,120(r1)
+	ld r17,112(r1)
+	ld r18,104(r1)
+	ld r19,96(r1)
+	ld r20,88(r1)
+	ld r21,80(r1)
+	ld r22,72(r1)
+	ld r23,64(r1)
+	ld r24,56(r1)
+	ld r25,48(r1)
+	ld r26,40(r1)
+	ld r27,32(r1)
+	ld r28,24(r1)
+	ld r29,16(r1)
+	ld r30,8(r1)
+	ld r31,0(r1)
+
+	; Set LR
 	mtlr r0
 
-	# Pop stack frame
+	; Pop stack frame
-	addi r1,r1,152
+	addi r1,r1,160
 
-	# Jump to return address
+	; Jump to return address
 	blr

coroutine/ppc64/Context.h

@@ -12,7 +12,7 @@
 
 enum {
   COROUTINE_REGISTERS =
-  19  /* 18 general purpose registers (r14–r31) and 1 return address */
+  20  /* 19 general purpose registers (r13–r31) and 1 return address */
   + 4  /* space for fiber_entry() to store the link register */
 };
 
@@ -44,7 +44,7 @@ static inline void coroutine_initialize(
     memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
 
     /* Skip a global prologue that sets the TOC register */
-    context->stack_pointer[18] = ((char*)start) + 8;
+    context->stack_pointer[19] = ((char*)start) + 8;
 }
 
 struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);

dln.c

@@ -41,6 +41,10 @@ static void dln_loaderror(const char *format, ...);
 # include <strings.h>
 #endif
 
+#if defined __APPLE__
+# include <AvailabilityMacros.h>
+#endif
+
 #ifndef xmalloc
 void *xmalloc();
 void *xcalloc();
@@ -58,7 +62,7 @@ void *xrealloc();
 #include <sys/stat.h>
 
 #ifndef S_ISDIR
-#   define S_ISDIR(m) (((m) & S_IFMT) == S_IFDIR)
+# define S_ISDIR(m) (((m) & S_IFMT) == S_IFDIR)
 #endif
 
 #ifdef HAVE_SYS_PARAM_H

error.c

@@ -668,7 +668,7 @@ bug_important_message(FILE *out, const char *const msg, size_t len)
 
 #undef CRASH_REPORTER_MAY_BE_CREATED
 #if defined(__APPLE__) && \
-    (!defined(MAC_OS_X_VERSION_10_6) || MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_X_VERSION_10_6)
+    (!defined(MAC_OS_X_VERSION_10_6) || MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_X_VERSION_10_6 || defined(__POWERPC__)) /* 10.6 PPC case */
 # define CRASH_REPORTER_MAY_BE_CREATED
 #endif
 static void

gc.c

@@ -1359,6 +1359,27 @@ tick(void)
     return val;
 }
 
+/* Implementation for macOS PPC by @nobu
+ * See: https://github.com/ruby/ruby/pull/5975#discussion_r890045558
+ */
+#elif defined(__POWERPC__) && defined(__APPLE__)
+typedef unsigned long long tick_t;
+#define PRItick "llu"
+
+static __inline__ tick_t
+tick(void)
+{
+    unsigned long int upper, lower, tmp;
+    # define mftbu(r) __asm__ volatile("mftbu   %0" : "=r"(r))
+    # define mftb(r)  __asm__ volatile("mftb    %0" : "=r"(r))
+        do {
+            mftbu(upper);
+            mftb(lower);
+            mftbu(tmp);
+        } while (tmp != upper);
+    return ((tick_t)upper << 32) | lower;
+}
+
 #elif defined(__aarch64__) &&  defined(__GNUC__)
 typedef unsigned long tick_t;
 #define PRItick "lu"

gc.h

@@ -6,10 +6,12 @@
 #define SET_MACHINE_STACK_END(p) __asm__ __volatile__ ("movq\t%%rsp, %0" : "=r" (*(p)))
 #elif defined(__i386) && defined(__GNUC__)
 #define SET_MACHINE_STACK_END(p) __asm__ __volatile__ ("movl\t%%esp, %0" : "=r" (*(p)))
-#elif (defined(__powerpc__) || defined(__powerpc64__)) && defined(__GNUC__) && !defined(_AIX)
+#elif (defined(__powerpc__) || defined(__powerpc64__)) && defined(__GNUC__) && !defined(_AIX) && !defined(__APPLE__) // Not Apple is NEEDED to unbreak ppc64 build on Darwin. Don't ask.
 #define SET_MACHINE_STACK_END(p) __asm__ __volatile__ ("mr\t%0, %%r1" : "=r" (*(p)))
 #elif (defined(__powerpc__) || defined(__powerpc64__)) && defined(__GNUC__) && defined(_AIX)
 #define SET_MACHINE_STACK_END(p) __asm__ __volatile__ ("mr %0,1" : "=r" (*(p)))
+#elif defined(__POWERPC__) && defined(__APPLE__) // Darwin ppc and ppc64
+#define SET_MACHINE_STACK_END(p) __asm__ volatile("mr %0, r1" : "=r" (*(p)))
 #elif defined(__aarch64__) && defined(__GNUC__)
 #define SET_MACHINE_STACK_END(p) __asm__ __volatile__ ("mov\t%0, sp" : "=r" (*(p)))
 #else

include/ruby/internal/config.h

@@ -113,6 +113,8 @@
 # define UNALIGNED_WORD_ACCESS 1
 #elif defined(__powerpc64__)
 # define UNALIGNED_WORD_ACCESS 1
+#elif defined(__POWERPC__) // __POWERPC__ is defined for ppc and ppc64 on Darwin
+# define UNALIGNED_WORD_ACCESS 1
 #elif defined(__aarch64__)
 # define UNALIGNED_WORD_ACCESS 1
 #elif defined(__mc68020__)

regint.h

@@ -49,10 +49,11 @@
 # endif
 #endif
 
+/* __POWERPC__ added to accommodate Darwin case. */
 #ifndef UNALIGNED_WORD_ACCESS
 # if defined(__i386) || defined(__i386__) || defined(_M_IX86) || \
      defined(__x86_64) || defined(__x86_64__) || defined(_M_AMD64) || \
-     defined(__powerpc64__) || defined(__aarch64__) || \
+     defined(__powerpc64__) || defined(__POWERPC__) || defined(__aarch64__) || \
      defined(__mc68020__)
 #  define UNALIGNED_WORD_ACCESS 1
 # else

siphash.c

@@ -34,10 +34,11 @@
   #error "Only strictly little or big endian supported"
 #endif
 
+/* __POWERPC__ added to accommodate Darwin case. */
 #ifndef UNALIGNED_WORD_ACCESS
 # if defined(__i386) || defined(__i386__) || defined(_M_IX86) || \
      defined(__x86_64) || defined(__x86_64__) || defined(_M_AMD64) || \
-     defined(__powerpc64__) || defined(__aarch64__) || \
+     defined(__powerpc64__) || defined(__POWERPC__) || defined(__aarch64__) || \
      defined(__mc68020__)
 #   define UNALIGNED_WORD_ACCESS 1
 # endif

st.c

@@ -1671,10 +1671,11 @@ st_values_check(st_table *tab, st_data_t *values, st_index_t size,
  */
 #define FNV_32_PRIME 0x01000193
 
+/* __POWERPC__ added to accommodate Darwin case. */
 #ifndef UNALIGNED_WORD_ACCESS
 # if defined(__i386) || defined(__i386__) || defined(_M_IX86) || \
      defined(__x86_64) || defined(__x86_64__) || defined(_M_AMD64) || \
-     defined(__powerpc64__) || defined(__aarch64__) || \
+     defined(__powerpc64__) || defined(__POWERPC__) || defined(__aarch64__) || \
      defined(__mc68020__)
 #   define UNALIGNED_WORD_ACCESS 1
 # endif

tool/m4/ruby_default_arch.m4

@@ -5,6 +5,7 @@ AS_CASE([$1],
 	[arm64],      [],
 	[*64],        [ARCH_FLAG=-m64],
 	[[i[3-6]86]], [ARCH_FLAG=-m32],
+	[ppc],        [ARCH_FLAG=-m32],
 	[AC_MSG_ERROR(unknown target architecture: $target_archs)]
 	)
 AC_MSG_RESULT([$ARCH_FLAG])

vm_core.h

@@ -780,8 +780,8 @@ typedef struct rb_vm_struct {
 #define RUBY_VM_FIBER_VM_STACK_SIZE           (  16 * 1024 * sizeof(VALUE)) /*   64 KB or  128 KB */
 #define RUBY_VM_FIBER_VM_STACK_SIZE_MIN       (   2 * 1024 * sizeof(VALUE)) /*    8 KB or   16 KB */
 #define RUBY_VM_FIBER_MACHINE_STACK_SIZE      (  64 * 1024 * sizeof(VALUE)) /*  256 KB or  512 KB */
-#if defined(__powerpc64__)
+#if defined(__powerpc64__) || defined(__ppc64__) // macOS has __ppc64__
-#define RUBY_VM_FIBER_MACHINE_STACK_SIZE_MIN  (  32 * 1024 * sizeof(VALUE)) /*   128 KB or  256 KB */
+#define RUBY_VM_FIBER_MACHINE_STACK_SIZE_MIN  (  32 * 1024 * sizeof(VALUE)) /*  128 KB or  256 KB */
 #else
 #define RUBY_VM_FIBER_MACHINE_STACK_SIZE_MIN  (  16 * 1024 * sizeof(VALUE)) /*   64 KB or  128 KB */
 #endif

vm_exec.c

@@ -55,7 +55,7 @@ static void vm_insns_counter_count_insn(int insn) {}
 #elif defined(__GNUC__) && defined(__i386__)
 #define DECL_SC_REG(type, r, reg) register type reg_##r __asm__("e" reg)
 
-#elif defined(__GNUC__) && defined(__powerpc64__)
+#elif defined(__GNUC__) && (defined(__powerpc64__) || defined(__POWERPC__))
 #define DECL_SC_REG(type, r, reg) register type reg_##r __asm__("r" reg)
 
 #elif defined(__GNUC__) && defined(__aarch64__)
@@ -92,7 +92,7 @@ vm_exec_core(rb_execution_context_t *ec, VALUE initial)
     DECL_SC_REG(rb_control_frame_t *, cfp, "15");
 #define USE_MACHINE_REGS 1
 
-#elif defined(__GNUC__) && defined(__powerpc64__)
+#elif defined(__GNUC__) && (defined(__powerpc64__) || defined(__POWERPC__))
     DECL_SC_REG(const VALUE *, pc, "14");
     DECL_SC_REG(rb_control_frame_t *, cfp, "15");
 #define USE_MACHINE_REGS 1