Ruby31: add support for Darwin ppc/ppc64 (#5927)

* add coroutines for ppc & ppc64

* fix universal coroutine to include ppc & ppc64

* add powerpc*-darwin to configure.ac

* fix thread_pthread for older systems

configure.ac

@@ -1866,7 +1866,7 @@ AS_IF([test $rb_cv_stack_end_address != no], [
 dnl Checks for library functions.
 AC_TYPE_GETGROUPS
 AS_CASE(["${target_cpu}-${target_os}:${target_archs}"],
-[powerpc-darwin*], [
+[powerpc*-darwin*], [
   AC_LIBSOURCES(alloca.c)
   AC_SUBST([ALLOCA], [\${LIBOBJDIR}alloca.${ac_objext}])
   AC_DEFINE(C_ALLOCA)
@@ -2546,6 +2546,12 @@ AS_CASE([$coroutine_type], [yes|''], [
         [arm64-darwin*], [
             coroutine_type=arm64
         ],
+        [powerpc-darwin*], [
+            coroutine_type=ppc
+        ],
+        [powerpc64-darwin*], [
+            coroutine_type=ppc64
+        ],
         [x*64-linux*], [
             AS_CASE(["$ac_cv_sizeof_voidp"],
                 [8], [ coroutine_type=amd64 ],

coroutine/ppc/Context.S

@@ -0,0 +1,73 @@
+#define TOKEN_PASTE(x,y) x##y
+#define PREFIXED_SYMBOL(prefix,name) TOKEN_PASTE(prefix,name)
+
+.text
+.align 2
+
+.globl PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer)
+PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer):
+	# Make space on the stack for caller registers
+	addi r1,r1,-80
+
+	# Save caller registers
+	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
+	stw r0,76(r1)
+
+	# Save stack pointer to first argument
+	stw r1,0(r3)
+
+	# Load stack pointer from second argument
+	lwz r1,0(r4)
+
+	# Restore caller registers
+	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)
+	mtlr r0
+
+	# Pop stack frame
+	addi r1,r1,80
+
+	# Jump to return address
+	blr
+

coroutine/ppc/Context.h

@@ -0,0 +1,57 @@
+#ifndef COROUTINE_PPC_CONTEXT_H
+#define COROUTINE_PPC_CONTEXT_H 1
+
+#pragma once
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+#define COROUTINE __attribute__((noreturn)) void
+
+enum {
+  COROUTINE_REGISTERS =
+  20  /* 19 general purpose registers (r13–r31) and 1 return address */
+  + 4  /* space for fiber_entry() to store the link register */
+};
+
+struct coroutine_context
+{
+    void **stack_pointer;
+    void *argument;
+};
+
+typedef COROUTINE(* coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
+
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->stack_pointer = NULL;
+}
+
+static inline void coroutine_initialize(
+    struct coroutine_context *context,
+    coroutine_start start,
+    void *stack,
+    size_t size
+) {
+    assert(start && stack && size >= 1024);
+
+    // Stack grows down. Force 16-byte alignment.
+    char * top = (char*)stack + size;
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
+
+    context->stack_pointer -= COROUTINE_REGISTERS;
+    memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
+
+    /* Skip a global prologue that sets the TOC register */
+    context->stack_pointer[19] = ((char*)start) + 8;
+}
+
+struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);
+
+static inline void coroutine_destroy(struct coroutine_context * context)
+{
+    context->stack_pointer = NULL;
+}
+
+#endif /* COROUTINE_PPC_CONTEXT_H */

coroutine/ppc64/Context.S

@@ -0,0 +1,70 @@
+#define TOKEN_PASTE(x,y) x##y
+#define PREFIXED_SYMBOL(prefix,name) TOKEN_PASTE(prefix,name)
+
+.text
+.align 3
+
+.globl PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer)
+PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer):
+	# Make space on the stack for caller registers
+	addi r1,r1,-152
+
+	# Save caller registers
+	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
+	std r1,0(r3)
+
+	# Load stack pointer from second argument
+	ld r1,0(r4)
+
+	# Restore caller registers
+	ld r14,0(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
+	ld r0,144(r1)
+	mtlr r0
+
+	# Pop stack frame
+	addi r1,r1,152
+
+	# Jump to return address
+	blr

coroutine/ppc64/Context.h

@@ -0,0 +1,57 @@
+#ifndef COROUTINE_PPC64_CONTEXT_H
+#define COROUTINE_PPC64_CONTEXT_H 1
+
+#pragma once
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+#define COROUTINE __attribute__((noreturn)) void
+
+enum {
+  COROUTINE_REGISTERS =
+  19  /* 18 general purpose registers (r14–r31) and 1 return address */
+  + 4  /* space for fiber_entry() to store the link register */
+};
+
+struct coroutine_context
+{
+    void **stack_pointer;
+    void *argument;
+};
+
+typedef COROUTINE(* coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
+
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->stack_pointer = NULL;
+}
+
+static inline void coroutine_initialize(
+    struct coroutine_context *context,
+    coroutine_start start,
+    void *stack,
+    size_t size
+) {
+    assert(start && stack && size >= 1024);
+
+    // Stack grows down. Force 16-byte alignment.
+    char * top = (char*)stack + size;
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
+
+    context->stack_pointer -= COROUTINE_REGISTERS;
+    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;
+}
+
+struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);
+
+static inline void coroutine_destroy(struct coroutine_context * context)
+{
+    context->stack_pointer = NULL;
+}
+
+#endif /* COROUTINE_PPC64_CONTEXT_H */

coroutine/universal/Context.S

@@ -3,7 +3,11 @@
 # include "coroutine/amd64/Context.S"
 #elif defined __i386__
 # include "coroutine/x86/Context.S"
-#elif defined __ppc64__
+#elif defined __ppc__
+# include "coroutine/ppc/Context.S"
+#elif defined __ppc64__ && defined(WORDS_BIGENDIAN)
+# include "coroutine/ppc64/Context.S"
+#elif defined __ppc64__ && !defined(WORDS_BIGENDIAN)
 # include "coroutine/ppc64le/Context.S"
 #elif defined __arm64__
 # include "coroutine/arm64/Context.S"

coroutine/universal/Context.h

@@ -6,7 +6,11 @@
 # include "coroutine/amd64/Context.h"
 #elif defined __i386__
 # include "coroutine/x86/Context.h"
-#elif defined __ppc64__
+#elif defined __ppc__
+# include "coroutine/ppc/Context.h"
+#elif defined __ppc64__ && defined(WORDS_BIGENDIAN)
+# include "coroutine/ppc64/Context.h"
+#elif defined __ppc64__ && !defined(WORDS_BIGENDIAN)
 # include "coroutine/ppc64le/Context.h"
 #elif defined __arm64__
 # include "coroutine/arm64/Context.h"

thread_pthread.c

@@ -1765,10 +1765,23 @@ native_thread_native_thread_id(rb_thread_t *target_th)
     if (tid == 0) return Qnil;
     return INT2FIX(tid);
 #elif defined(__APPLE__)
-    uint64_t tid;
-    int e = pthread_threadid_np(target_th->nt->thread_id, &tid);
-    if (e != 0) rb_syserr_fail(e, "pthread_threadid_np");
-    return ULL2NUM((unsigned long long)tid);
+    #if MAC_OS_X_VERSION_MAX_ALLOWED < 1060
+        uint64_t tid;
+        tid = pthread_mach_thread_np(pthread_self());
+    #elif MAC_OS_X_VERSION_MIN_REQUIRED < 1060
+        uint64_t tid;
+        if (&pthread_threadid_np) {
+            int error = pthread_threadid_np(target_th->thread_id, &tid);
+            if (error != 0) rb_syserr_fail(error, "pthread_threadid_np");
+        } else {
+            uint64_t tid;
+            tid = pthread_mach_thread_np(pthread_self());
+        }
+    #else
+        int error = pthread_threadid_np(target_th->thread_id, &tid);
+        if (error != 0) rb_syserr_fail(error, "pthread_threadid_np");
+        return ULL2NUM((unsigned long long)tid);
+    #endif
 #endif
 }
 # define USE_NATIVE_THREAD_NATIVE_THREAD_ID 1