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

4 files changed, 140 insertions, 103 deletions

diff --git a/coroutine/ppc/Context.S b/coroutine/ppc/Context.S
index fe28390df0..cdda93e179 100644
--- a/coroutine/ppc/Context.S
+++ b/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
-	addi r1,r1,-80
+	; Make space on the stack for caller registers
+	; (Should we rather use red zone? See libphobos example.)
+	subi r1,r1,80
+
+	; Get LR
+	mflr r0
 
-	# 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)
+	; 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 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)
+	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
-	mflr r0
+	; 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
-	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)
+	; 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 r23,40(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)
+	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)
 
-	# Load return address
-	lwz r0,76(r1)
+	; Set LR
 	mtlr r0
 
-	# Pop stack frame
+	; Pop stack frame
 	addi r1,r1,80
 
-	# Jump to return address
+	; Jump to return address
 	blr
-
diff --git a/coroutine/ppc/Context.h b/coroutine/ppc/Context.h
index 9f69390388..1fce112579 100644
--- a/coroutine/ppc/Context.h
+++ b/coroutine/ppc/Context.h
@@ -9,6 +9,7 @@
 #include <string.h>
 
 #define COROUTINE __attribute__((noreturn)) void
+#define COROUTINE_LIMITED_ADDRESS_SPACE
 
 enum {
   COROUTINE_REGISTERS =
diff --git a/coroutine/ppc64/Context.S b/coroutine/ppc64/Context.S
index 1bd9268f93..f8561e0e7d 100644
--- a/coroutine/ppc64/Context.S
+++ b/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)
-PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer):
-	# Make space on the stack for caller registers
-	addi r1,r1,-152
+.align 2
 
-	# 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)
+PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer):
+	; Make space on the stack for caller registers
+	; (Should we rather use red zone? See libphobos example.)
+	subi r1,r1,160
 
-	# Save return address
+	; Get LR
 	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
-	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 r0,152(r1)
+
+	; Restore caller registers
+	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)
 
-	# Load return address
-	ld r0,144(r1)
+	; Set LR
 	mtlr r0
 
-	# Pop stack frame
-	addi r1,r1,152
+	; Pop stack frame
+	addi r1,r1,160
 
-	# Jump to return address
+	; Jump to return address
 	blr
diff --git a/coroutine/ppc64/Context.h b/coroutine/ppc64/Context.h
index 5b47511b9c..3e6f77f55a 100644
--- a/coroutine/ppc64/Context.h
+++ b/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);