runtime: use reflect.call during panic instead of newstackcall

newstackcall creates a new stack segment, and we want to
be able to throw away all that code.

LGTM=khr
R=khr, iant
CC=dvyukov, golang-codereviews, r
https://codereview.appspot.com/139270043

17 files changed, 187 insertions, 284 deletions

diff --git a/misc/cgo/test/callback.go b/misc/cgo/test/callback.go
index 98f653ef7..281e79494 100644
--- a/misc/cgo/test/callback.go
+++ b/misc/cgo/test/callback.go
@@ -153,6 +153,7 @@ func testCallbackCallers(t *testing.T) {
 	n := 0
 	name := []string{
 		"test.goCallback",
+		"runtime.call16",
 		"runtime.cgocallbackg1",
 		"runtime.cgocallbackg",
 		"runtime.cgocallback_gofunc",
diff --git a/misc/cgo/test/issue7695_test.go b/misc/cgo/test/issue7695_test.go
index 4bd6f8e73..de2fc03d4 100644
--- a/misc/cgo/test/issue7695_test.go
+++ b/misc/cgo/test/issue7695_test.go
@@ -2,6 +2,9 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
+// +build ignore
+// This test depends on running C code on Go stacks. Not allowed anymore.
+
 // Demo of deferred C function with untrue prototype
 // breaking stack copying. See golang.org/issue/7695.
 
diff --git a/src/cmd/dist/buildruntime.c b/src/cmd/dist/buildruntime.c
index 3a274e05c..246bd2710 100644
--- a/src/cmd/dist/buildruntime.c
+++ b/src/cmd/dist/buildruntime.c
@@ -244,6 +244,8 @@ ok:
 				aggr = "seh";
 			else if(streq(fields.p[1], "Alg"))
 				aggr = "alg";
+			else if(streq(fields.p[1], "Panic"))
+				aggr = "panic";
 		}
 		if(hasprefix(lines.p[i], "}"))
 			aggr = nil;
diff --git a/src/pkg/reflect/value.go b/src/pkg/reflect/value.go
index 76086c561..b02b8ea0c 100644
--- a/src/pkg/reflect/value.go
+++ b/src/pkg/reflect/value.go
@@ -563,7 +563,7 @@ func (v Value) call(op string, in []Value) []Value {
 	}
 
 	// Call.
-	call(fn, args, uint32(frametype.size), uint32(retOffset))
+	call(fn, args, uint32(frametype.size), uint32(retOffset), nil)
 
 	// For testing; see TestCallMethodJump.
 	if callGC {
@@ -761,7 +761,7 @@ func callMethod(ctxt *methodValue, frame unsafe.Pointer) {
 	memmove(unsafe.Pointer(uintptr(args)+ptrSize), frame, argSize-ptrSize)
 
 	// Call.
-	call(fn, args, uint32(frametype.size), uint32(retOffset))
+	call(fn, args, uint32(frametype.size), uint32(retOffset), nil)
 
 	// Copy return values. On amd64p32, the beginning of return values
 	// is 64-bit aligned, so the caller's frame layout (which doesn't have
@@ -2700,7 +2700,9 @@ func mapiterkey(it unsafe.Pointer) (key unsafe.Pointer)
 func mapiternext(it unsafe.Pointer)
 func maplen(m unsafe.Pointer) int
 
-func call(fn, arg unsafe.Pointer, n uint32, retoffset uint32)
+// panicpos is for use by runtime and should be nil in all calls in this package
+func call(fn, arg unsafe.Pointer, n uint32, retoffset uint32, panicpos unsafe.Pointer)
+
 func ifaceE2I(t *rtype, src interface{}, dst unsafe.Pointer)
 
 // Dummy annotation marking that the value x escapes,
diff --git a/src/pkg/runtime/asm_386.s b/src/pkg/runtime/asm_386.s
index 25911a5b9..0b5ded683 100644
--- a/src/pkg/runtime/asm_386.s
+++ b/src/pkg/runtime/asm_386.s
@@ -283,6 +283,12 @@ TEXT runtime·morestack(SB),NOSPLIT,$0-0
 	JNE	2(PC)
 	INT	$3
 
+	// Cannot grow signal stack.
+	MOVL	m_gsignal(BX), SI
+	CMPL	g(CX), SI
+	JNE	2(PC)
+	INT	$3
+
 	// frame size in DI
 	// arg size in AX
 	// Save in m.
@@ -322,56 +328,8 @@ TEXT runtime·morestack_noctxt(SB),NOSPLIT,$0-0
 	MOVL	$0, DX
 	JMP runtime·morestack(SB)
 
-// Called from panic.  Mimics morestack,
-// reuses stack growth code to create a frame
-// with the desired args running the desired function.
-//
-// func call(fn *byte, arg *byte, argsize uint32).
-TEXT runtime·newstackcall(SB), NOSPLIT, $0-12
-	get_tls(CX)
-	MOVL	g(CX), BX
-	MOVL	g_m(BX), BX
-
-	// Save our caller's state as the PC and SP to
-	// restore when returning from f.
-	MOVL	0(SP), AX	// our caller's PC
-	MOVL	AX, (m_morebuf+gobuf_pc)(BX)
-	LEAL	fv+0(FP), AX	// our caller's SP
-	MOVL	AX, (m_morebuf+gobuf_sp)(BX)
-	MOVL	g(CX), AX
-	MOVL	AX, (m_morebuf+gobuf_g)(BX)
-
-	// Save our own state as the PC and SP to restore
-	// if this goroutine needs to be restarted.
-	MOVL	$runtime·newstackcall(SB), (g_sched+gobuf_pc)(AX)
-	MOVL	SP, (g_sched+gobuf_sp)(AX)
-
-	// Set up morestack arguments to call f on a new stack.
-	// We set f's frame size to 1, as a hint to newstack
-	// that this is a call from runtime·newstackcall.
-	// If it turns out that f needs a larger frame than
-	// the default stack, f's usual stack growth prolog will
-	// allocate a new segment (and recopy the arguments).
-	MOVL	fv+0(FP), AX	// fn
-	MOVL	addr+4(FP), DX	// arg frame
-	MOVL	size+8(FP), CX	// arg size
-
-	MOVL	AX, m_cret(BX)	// f's PC
-	MOVL	DX, m_moreargp(BX)	// f's argument pointer
-	MOVL	CX, m_moreargsize(BX)	// f's argument size
-	MOVL	$1, m_moreframesize(BX)	// f's frame size
-
-	// Call newstack on m->g0's stack.
-	MOVL	m_g0(BX), BP
-	get_tls(CX)
-	MOVL	BP, g(CX)
-	MOVL	(g_sched+gobuf_sp)(BP), SP
-	CALL	runtime·newstack(SB)
-	MOVL	$0, 0x1103	// crash if newstack returns
-	RET
-
 // reflect·call: call a function with the given argument list
-// func call(f *FuncVal, arg *byte, argsize uint32).
+// func call(f *FuncVal, arg *byte, argsize, retoffset uint32, p *Panic).
 // we don't have variable-sized frames, so we use a small number
 // of constant-sized-frame functions to encode a few bits of size in the pc.
 // Caution: ugly multiline assembly macros in your future!
@@ -379,11 +337,11 @@ TEXT runtime·newstackcall(SB), NOSPLIT, $0-12
 #define DISPATCH(NAME,MAXSIZE)		\
 	CMPL	CX, $MAXSIZE;		\
 	JA	3(PC);			\
-	MOVL	$NAME(SB), AX;	\
+	MOVL	$NAME(SB), AX;		\
 	JMP	AX
 // Note: can't just "JMP NAME(SB)" - bad inlining results.
 
-TEXT reflect·call(SB), NOSPLIT, $0-16
+TEXT reflect·call(SB), NOSPLIT, $0-20
 	MOVL	argsize+8(FP), CX
 	DISPATCH(runtime·call16, 16)
 	DISPATCH(runtime·call32, 32)
@@ -415,11 +373,11 @@ TEXT reflect·call(SB), NOSPLIT, $0-16
 	MOVL	$runtime·badreflectcall(SB), AX
 	JMP	AX
 
-// Argument map for the callXX frames.  Each has one
-// stack map (for the single call) with 3 arguments.
+// Argument map for the callXX frames.  Each has one stack map.
 DATA gcargs_reflectcall<>+0x00(SB)/4, $1  // 1 stackmap
-DATA gcargs_reflectcall<>+0x04(SB)/4, $6  // 3 args
-DATA gcargs_reflectcall<>+0x08(SB)/4, $(const_BitsPointer+(const_BitsPointer<<2)+(const_BitsScalar<<4))
+DATA gcargs_reflectcall<>+0x04(SB)/4, $10  // 5 words
+DATA gcargs_reflectcall<>+0x08(SB)/1, $(const_BitsPointer+(const_BitsPointer<<2)+(const_BitsScalar<<4)+(const_BitsScalar<<6))
+DATA gcargs_reflectcall<>+0x09(SB)/1, $(const_BitsPointer)
 GLOBL gcargs_reflectcall<>(SB),RODATA,$12
 
 // callXX frames have no locals
@@ -428,7 +386,7 @@ DATA gclocals_reflectcall<>+0x04(SB)/4, $0  // 0 locals
 GLOBL gclocals_reflectcall<>(SB),RODATA,$8
 
 #define CALLFN(NAME,MAXSIZE)			\
-TEXT NAME(SB), WRAPPER, $MAXSIZE-16;	\
+TEXT NAME(SB), WRAPPER, $MAXSIZE-20;	\
 	FUNCDATA $FUNCDATA_ArgsPointerMaps,gcargs_reflectcall<>(SB);	\
 	FUNCDATA $FUNCDATA_LocalsPointerMaps,gclocals_reflectcall<>(SB);\
 	/* copy arguments to stack */		\
@@ -436,11 +394,22 @@ TEXT NAME(SB), WRAPPER, $MAXSIZE-16;	\
 	MOVL	argsize+8(FP), CX;		\
 	MOVL	SP, DI;				\
 	REP;MOVSB;				\
+	/* initialize panic argp */		\
+	MOVL	panic+16(FP), CX;		\
+	CMPL	CX, $0;				\
+	JEQ	3(PC);				\
+	LEAL	(MAXSIZE+4)(SP), BX;		\
+	MOVL	BX, panic_argp(CX);		\
 	/* call function */			\
 	MOVL	f+0(FP), DX;			\
 	MOVL	(DX), AX; 			\
 	PCDATA  $PCDATA_StackMapIndex, $0;	\
 	CALL	AX;				\
+	/* clear panic argp */			\
+	MOVL	panic+16(FP), CX;		\
+	CMPL	CX, $0;				\
+	JEQ	2(PC);				\
+	MOVL	$0, panic_argp(CX);		\
 	/* copy return values back */		\
 	MOVL	argptr+4(FP), DI;		\
 	MOVL	argsize+8(FP), CX;		\
diff --git a/src/pkg/runtime/asm_amd64.s b/src/pkg/runtime/asm_amd64.s
index 962074582..587fcf480 100644
--- a/src/pkg/runtime/asm_amd64.s
+++ b/src/pkg/runtime/asm_amd64.s
@@ -274,6 +274,12 @@ TEXT runtime·morestack(SB),NOSPLIT,$0-0
 	JNE	2(PC)
 	INT	$3
 
+	// Cannot grow signal stack (m->gsignal).
+	MOVQ	m_gsignal(BX), SI
+	CMPQ	g(CX), SI
+	JNE	2(PC)
+	INT	$3
+
 	// Called from f.
 	// Set m->morebuf to f's caller.
 	MOVQ	8(SP), AX	// f's caller's PC
@@ -301,57 +307,8 @@ TEXT runtime·morestack(SB),NOSPLIT,$0-0
 	MOVQ	$0, 0x1003	// crash if newstack returns
 	RET
 
-// Called from panic.  Mimics morestack,
-// reuses stack growth code to create a frame
-// with the desired args running the desired function.
-//
-// func call(fn *byte, arg *byte, argsize uint32).
-TEXT runtime·newstackcall(SB), NOSPLIT, $0-20
-	get_tls(CX)
-	MOVQ	g(CX), BX
-	MOVQ	g_m(BX), BX
-
-	// Save our caller's state as the PC and SP to
-	// restore when returning from f.
-	MOVQ	0(SP), AX	// our caller's PC
-	MOVQ	AX, (m_morebuf+gobuf_pc)(BX)
-	LEAQ	fv+0(FP), AX	// our caller's SP
-	MOVQ	AX, (m_morebuf+gobuf_sp)(BX)
-	MOVQ	g(CX), AX
-	MOVQ	AX, (m_morebuf+gobuf_g)(BX)
-	
-	// Save our own state as the PC and SP to restore
-	// if this goroutine needs to be restarted.
-	MOVQ	$runtime·newstackcall(SB), BP
-	MOVQ	BP, (g_sched+gobuf_pc)(AX)
-	MOVQ	SP, (g_sched+gobuf_sp)(AX)
-
-	// Set up morestack arguments to call f on a new stack.
-	// We set f's frame size to 1, as a hint to newstack
-	// that this is a call from runtime·newstackcall.
-	// If it turns out that f needs a larger frame than
-	// the default stack, f's usual stack growth prolog will
-	// allocate a new segment (and recopy the arguments).
-	MOVQ	fv+0(FP), AX	// fn
-	MOVQ	addr+8(FP), DX	// arg frame
-	MOVL	size+16(FP), CX	// arg size
-
-	MOVQ	AX, m_cret(BX)	// f's PC
-	MOVQ	DX, m_moreargp(BX)	// argument frame pointer
-	MOVL	CX, m_moreargsize(BX)	// f's argument size
-	MOVL	$1, m_moreframesize(BX)	// f's frame size
-
-	// Call newstack on m->g0's stack.
-	MOVQ	m_g0(BX), BP
-	get_tls(CX)
-	MOVQ	BP, g(CX)
-	MOVQ	(g_sched+gobuf_sp)(BP), SP
-	CALL	runtime·newstack(SB)
-	MOVQ	$0, 0x1103	// crash if newstack returns
-	RET
-
 // reflect·call: call a function with the given argument list
-// func call(f *FuncVal, arg *byte, argsize uint32).
+// func call(f *FuncVal, arg *byte, argsize, retoffset uint32, p *Panic).
 // we don't have variable-sized frames, so we use a small number
 // of constant-sized-frame functions to encode a few bits of size in the pc.
 // Caution: ugly multiline assembly macros in your future!
@@ -363,7 +320,7 @@ TEXT runtime·newstackcall(SB), NOSPLIT, $0-20
 	JMP	AX
 // Note: can't just "JMP NAME(SB)" - bad inlining results.
 
-TEXT reflect·call(SB), NOSPLIT, $0-24
+TEXT reflect·call(SB), NOSPLIT, $0-32
 	MOVLQZX argsize+16(FP), CX
 	DISPATCH(runtime·call16, 16)
 	DISPATCH(runtime·call32, 32)
@@ -395,11 +352,10 @@ TEXT reflect·call(SB), NOSPLIT, $0-24
 	MOVQ	$runtime·badreflectcall(SB), AX
 	JMP	AX
 
-// Argument map for the callXX frames.  Each has one
-// stack map (for the single call) with 3 arguments.
+// Argument map for the callXX frames.  Each has one stack map.
 DATA gcargs_reflectcall<>+0x00(SB)/4, $1  // 1 stackmap
-DATA gcargs_reflectcall<>+0x04(SB)/4, $6  // 3 args
-DATA gcargs_reflectcall<>+0x08(SB)/4, $(const_BitsPointer+(const_BitsPointer<<2)+(const_BitsScalar<<4))
+DATA gcargs_reflectcall<>+0x04(SB)/4, $8  // 4 words
+DATA gcargs_reflectcall<>+0x08(SB)/1, $(const_BitsPointer+(const_BitsPointer<<2)+(const_BitsScalar<<4)+(const_BitsPointer<<6))
 GLOBL gcargs_reflectcall<>(SB),RODATA,$12
 
 // callXX frames have no locals
@@ -407,8 +363,16 @@ DATA gclocals_reflectcall<>+0x00(SB)/4, $1  // 1 stackmap
 DATA gclocals_reflectcall<>+0x04(SB)/4, $0  // 0 locals
 GLOBL gclocals_reflectcall<>(SB),RODATA,$8
 
+// CALLFN is marked as a WRAPPER so that a deferred reflect.call func will
+// see the right answer for recover. However, CALLFN is also how we start
+// the panic in the first place. We record the panic argp if this is the start of
+// a panic. Since the wrapper adjustment has already happened, though
+// (in the implicit prologue), we have to write not SP but MAXSIZE+8+SP into
+// p.argp. The MAXSIZE+8 will counter the MAXSIZE+8 the wrapper prologue
+// added to g->panicwrap.
+
 #define CALLFN(NAME,MAXSIZE)			\
-TEXT NAME(SB), WRAPPER, $MAXSIZE-24;	\
+TEXT NAME(SB), WRAPPER, $MAXSIZE-32;		\
 	FUNCDATA $FUNCDATA_ArgsPointerMaps,gcargs_reflectcall<>(SB);	\
 	FUNCDATA $FUNCDATA_LocalsPointerMaps,gclocals_reflectcall<>(SB);\
 	/* copy arguments to stack */		\
@@ -416,10 +380,21 @@ TEXT NAME(SB), WRAPPER, $MAXSIZE-24;	\
 	MOVLQZX argsize+16(FP), CX;		\
 	MOVQ	SP, DI;				\
 	REP;MOVSB;				\
+	/* initialize panic argp */		\
+	MOVQ	panic+24(FP), CX;		\
+	CMPQ	CX, $0;				\
+	JEQ	3(PC);				\
+	LEAQ	(MAXSIZE+8)(SP), BX;		\
+	MOVQ	BX, panic_argp(CX);		\
 	/* call function */			\
 	MOVQ	f+0(FP), DX;			\
 	PCDATA  $PCDATA_StackMapIndex, $0;	\
 	CALL	(DX);				\
+	/* clear panic argp */			\
+	MOVQ	panic+24(FP), CX;		\
+	CMPQ	CX, $0;				\
+	JEQ	2(PC);				\
+	MOVQ	$0, panic_argp(CX);		\
 	/* copy return values back */		\
 	MOVQ	argptr+8(FP), DI;		\
 	MOVLQZX	argsize+16(FP), CX;		\
diff --git a/src/pkg/runtime/asm_amd64p32.s b/src/pkg/runtime/asm_amd64p32.s
index 64a436495..5647d7762 100644
--- a/src/pkg/runtime/asm_amd64p32.s
+++ b/src/pkg/runtime/asm_amd64p32.s
@@ -247,6 +247,12 @@ TEXT runtime·morestack(SB),NOSPLIT,$0-0
 	JNE	2(PC)
 	MOVL	0, AX
 
+	// Cannot grow signal stack (m->gsignal).
+	MOVL	m_gsignal(BX), SI
+	CMPL	g(CX), SI
+	JNE	2(PC)
+	MOVL	0, AX
+
 	// Called from f.
 	// Set m->morebuf to f's caller.
 	MOVL	8(SP), AX	// f's caller's PC
@@ -274,57 +280,8 @@ TEXT runtime·morestack(SB),NOSPLIT,$0-0
 	MOVL	$0, 0x1003	// crash if newstack returns
 	RET
 
-// Called from panic.  Mimics morestack,
-// reuses stack growth code to create a frame
-// with the desired args running the desired function.
-//
-// func call(fn *byte, arg *byte, argsize uint32).
-TEXT runtime·newstackcall(SB), NOSPLIT, $0-12
-	get_tls(CX)
-	MOVL	g(CX), BX
-	MOVL	g_m(BX), BX
-
-	// Save our caller's state as the PC and SP to
-	// restore when returning from f.
-	MOVL	0(SP), AX	// our caller's PC
-	MOVL	AX, (m_morebuf+gobuf_pc)(BX)
-	LEAL	fv+0(FP), AX	// our caller's SP
-	MOVL	AX, (m_morebuf+gobuf_sp)(BX)
-	MOVL	g(CX), AX
-	MOVL	AX, (m_morebuf+gobuf_g)(BX)
-	
-	// Save our own state as the PC and SP to restore
-	// if this goroutine needs to be restarted.
-	MOVL	$runtime·newstackcall(SB), DI
-	MOVL	DI, (g_sched+gobuf_pc)(AX)
-	MOVL	SP, (g_sched+gobuf_sp)(AX)
-
-	// Set up morestack arguments to call f on a new stack.
-	// We set f's frame size to 1, as a hint to newstack
-	// that this is a call from runtime·newstackcall.
-	// If it turns out that f needs a larger frame than
-	// the default stack, f's usual stack growth prolog will
-	// allocate a new segment (and recopy the arguments).
-	MOVL	fv+0(FP), AX	// fn
-	MOVL	addr+4(FP), DX	// arg frame
-	MOVL	size+8(FP), CX	// arg size
-
-	MOVQ	AX, m_cret(BX)	// f's PC
-	MOVL	DX, m_moreargp(BX)	// argument frame pointer
-	MOVL	CX, m_moreargsize(BX)	// f's argument size
-	MOVL	$1, m_moreframesize(BX)	// f's frame size
-
-	// Call newstack on m->g0's stack.
-	MOVL	m_g0(BX), BX
-	get_tls(CX)
-	MOVL	BX, g(CX)
-	MOVL	(g_sched+gobuf_sp)(BX), SP
-	CALL	runtime·newstack(SB)
-	MOVL	$0, 0x1103	// crash if newstack returns
-	RET
-
 // reflect·call: call a function with the given argument list
-// func call(f *FuncVal, arg *byte, argsize uint32).
+// func call(f *FuncVal, arg *byte, argsize, retoffset uint32, p *Panic).
 // we don't have variable-sized frames, so we use a small number
 // of constant-sized-frame functions to encode a few bits of size in the pc.
 // Caution: ugly multiline assembly macros in your future!
@@ -368,17 +325,42 @@ TEXT reflect·call(SB), NOSPLIT, $0-20
 	MOVL	$runtime·badreflectcall(SB), AX
 	JMP	AX
 
+// Argument map for the callXX frames.  Each has one stack map.
+DATA gcargs_reflectcall<>+0x00(SB)/4, $1  // 1 stackmap
+DATA gcargs_reflectcall<>+0x04(SB)/4, $10  // 5 words
+DATA gcargs_reflectcall<>+0x08(SB)/1, $(const_BitsPointer+(const_BitsPointer<<2)+(const_BitsScalar<<4)+(const_BitsScalar<<6))
+DATA gcargs_reflectcall<>+0x09(SB)/1, $(const_BitsPointer)
+GLOBL gcargs_reflectcall<>(SB),RODATA,$12
+
+// callXX frames have no locals
+DATA gclocals_reflectcall<>+0x00(SB)/4, $1  // 1 stackmap
+DATA gclocals_reflectcall<>+0x04(SB)/4, $0  // 0 locals
+GLOBL gclocals_reflectcall<>(SB),RODATA,$8
+
 #define CALLFN(NAME,MAXSIZE)			\
-TEXT NAME(SB), WRAPPER, $MAXSIZE-16;		\
+TEXT NAME(SB), WRAPPER, $MAXSIZE-20;		\
+	FUNCDATA $FUNCDATA_ArgsPointerMaps,gcargs_reflectcall<>(SB);	\
+	FUNCDATA $FUNCDATA_LocalsPointerMaps,gclocals_reflectcall<>(SB);\
 	/* copy arguments to stack */		\
 	MOVL	argptr+4(FP), SI;		\
 	MOVL	argsize+8(FP), CX;		\
 	MOVL	SP, DI;				\
 	REP;MOVSB;				\
+	/* initialize panic argp */		\
+	MOVL	panic+16(FP), CX;		\
+	CMPL	CX, $0;				\
+	JEQ	3(PC);				\
+	LEAL	(MAXSIZE+8)(SP), BX;		\
+	MOVL	BX, panic_argp(CX);		\
 	/* call function */			\
 	MOVL	f+0(FP), DX;			\
 	MOVL	(DX), AX;				\
 	CALL	AX; \
+	/* clear panic argp */			\
+	MOVL	panic+16(FP), CX;		\
+	CMPL	CX, $0;				\
+	JEQ	2(PC);				\
+	MOVL	$0, panic_argp(CX);		\
 	/* copy return values back */		\
 	MOVL	argptr+4(FP), DI;		\
 	MOVL	argsize+8(FP), CX;		\
diff --git a/src/pkg/runtime/asm_arm.s b/src/pkg/runtime/asm_arm.s
index 9e68e099e..96a21ceac 100644
--- a/src/pkg/runtime/asm_arm.s
+++ b/src/pkg/runtime/asm_arm.s
@@ -269,6 +269,11 @@ TEXT runtime·morestack(SB),NOSPLIT,$-4-0
 	CMP	g, R4
 	BL.EQ	runtime·abort(SB)
 
+	// Cannot grow signal stack (m->gsignal).
+	MOVW	m_gsignal(R8), R4
+	CMP	g, R4
+	BL.EQ	runtime·abort(SB)
+
 	MOVW	R1, m_moreframesize(R8)
 	MOVW	R2, m_moreargsize(R8)
 
@@ -300,49 +305,8 @@ TEXT runtime·morestack_noctxt(SB),NOSPLIT,$-4-0
 	MOVW	$0, R7
 	B runtime·morestack(SB)
 
-// Called from panic.  Mimics morestack,
-// reuses stack growth code to create a frame
-// with the desired args running the desired function.
-//
-// func call(fn *byte, arg *byte, argsize uint32).
-TEXT runtime·newstackcall(SB),NOSPLIT,$-4-12
-	// Save our caller's state as the PC and SP to
-	// restore when returning from f.
-	MOVW	g_m(g), R8
-	MOVW	LR, (m_morebuf+gobuf_pc)(R8)	// our caller's PC
-	MOVW	SP, (m_morebuf+gobuf_sp)(R8)	// our caller's SP
-	MOVW	g,  (m_morebuf+gobuf_g)(R8)
-
-	// Save our own state as the PC and SP to restore
-	// if this goroutine needs to be restarted.
-	MOVW	$runtime·newstackcall(SB), R11
-	MOVW	R11, (g_sched+gobuf_pc)(g)
-	MOVW	LR, (g_sched+gobuf_lr)(g)
-	MOVW	SP, (g_sched+gobuf_sp)(g)
-
-	// Set up morestack arguments to call f on a new stack.
-	// We set f's frame size to 1, as a hint to newstack
-	// that this is a call from runtime·newstackcall.
-	// If it turns out that f needs a larger frame than
-	// the default stack, f's usual stack growth prolog will
-	// allocate a new segment (and recopy the arguments).
-	MOVW	4(SP), R0			// fn
-	MOVW	8(SP), R1			// arg frame
-	MOVW	12(SP), R2			// arg size
-
-	MOVW	R0, m_cret(R8)			// f's PC
-	MOVW	R1, m_moreargp(R8)		// f's argument pointer
-	MOVW	R2, m_moreargsize(R8)		// f's argument size
-	MOVW	$1, R3
-	MOVW	R3, m_moreframesize(R8)		// f's frame size
-
-	// Call newstack on m->g0's stack.
-	MOVW	m_g0(R8), g
-	MOVW	(g_sched+gobuf_sp)(g), SP
-	B	runtime·newstack(SB)
-
 // reflect·call: call a function with the given argument list
-// func call(f *FuncVal, arg *byte, argsize uint32).
+// func call(f *FuncVal, arg *byte, argsize, retoffset uint32, p *Panic).
 // we don't have variable-sized frames, so we use a small number
 // of constant-sized-frame functions to encode a few bits of size in the pc.
 // Caution: ugly multiline assembly macros in your future!
@@ -350,10 +314,10 @@ TEXT runtime·newstackcall(SB),NOSPLIT,$-4-12
 #define DISPATCH(NAME,MAXSIZE)		\
 	CMP	$MAXSIZE, R0;		\
 	B.HI	3(PC);			\
-	MOVW	$NAME(SB), R1;	\
+	MOVW	$NAME(SB), R1;		\
 	B	(R1)
 
-TEXT reflect·call(SB),NOSPLIT,$-4-16
+TEXT reflect·call(SB),NOSPLIT,$-4-20
 	MOVW	argsize+8(FP), R0
 	DISPATCH(runtime·call16, 16)
 	DISPATCH(runtime·call32, 32)
@@ -385,11 +349,11 @@ TEXT reflect·call(SB),NOSPLIT,$-4-16
 	MOVW	$runtime·badreflectcall(SB), R1
 	B	(R1)
 
-// Argument map for the callXX frames.  Each has one
-// stack map (for the single call) with 3 arguments.
+// Argument map for the callXX frames.  Each has one stack map.
 DATA gcargs_reflectcall<>+0x00(SB)/4, $1  // 1 stackmap
-DATA gcargs_reflectcall<>+0x04(SB)/4, $6  // 3 args
-DATA gcargs_reflectcall<>+0x08(SB)/4, $(const_BitsPointer+(const_BitsPointer<<2)+(const_BitsScalar<<4))
+DATA gcargs_reflectcall<>+0x04(SB)/4, $10  // 5 words
+DATA gcargs_reflectcall<>+0x08(SB)/1, $(const_BitsPointer+(const_BitsPointer<<2)+(const_BitsScalar<<4)+(const_BitsScalar<<6))
+DATA gcargs_reflectcall<>+0x09(SB)/1, $(const_BitsPointer)
 GLOBL gcargs_reflectcall<>(SB),RODATA,$12
 
 // callXX frames have no locals
@@ -398,7 +362,7 @@ DATA gclocals_reflectcall<>+0x04(SB)/4, $0  // 0 locals
 GLOBL gclocals_reflectcall<>(SB),RODATA,$8
 
 #define CALLFN(NAME,MAXSIZE)			\
-TEXT NAME(SB), WRAPPER, $MAXSIZE-16;	\
+TEXT NAME(SB), WRAPPER, $MAXSIZE-20;		\
 	FUNCDATA $FUNCDATA_ArgsPointerMaps,gcargs_reflectcall<>(SB);	\
 	FUNCDATA $FUNCDATA_LocalsPointerMaps,gclocals_reflectcall<>(SB);\
 	/* copy arguments to stack */		\
@@ -411,11 +375,23 @@ TEXT NAME(SB), WRAPPER, $MAXSIZE-16;	\
 	MOVBU.P R5, 1(R1);			\
 	SUB	$1, R2, R2;			\
 	B	-5(PC);				\
+	/* initialize panic argp */		\
+	MOVW	panic+16(FP), R4;		\
+	CMP	$0, R4;				\
+	B.EQ	3(PC);				\
+	ADD	$(4+MAXSIZE+4), R13, R5;	\
+	MOVW	R5, panic_argp(R4);		\
 	/* call function */			\
 	MOVW	f+0(FP), R7;			\
 	MOVW	(R7), R0;			\
 	PCDATA  $PCDATA_StackMapIndex, $0;	\
 	BL	(R0);				\
+	/* clear panic argp */			\
+	MOVW	panic+16(FP), R4;		\
+	CMP	$0, R4;				\
+	B.EQ	3(PC);				\
+	MOVW	$0, R5;				\
+	MOVW	R5, panic_argp(R4);		\
 	/* copy return values back */		\
 	MOVW	argptr+4(FP), R0;		\
 	MOVW	argsize+8(FP), R2;		\
diff --git a/src/pkg/runtime/cgocall.go b/src/pkg/runtime/cgocall.go
index 1037c5dc2..4040fee8e 100644
--- a/src/pkg/runtime/cgocall.go
+++ b/src/pkg/runtime/cgocall.go
@@ -225,7 +225,7 @@ func cgocallbackg1() {
 	}
 
 	// Invoke callback.
-	newstackcall(cb.fn, cb.arg, uint32(cb.argsize))
+	reflectcall(unsafe.Pointer(cb.fn), unsafe.Pointer(cb.arg), uint32(cb.argsize), 0, nil)
 
 	if raceenabled {
 		racereleasemerge(unsafe.Pointer(&racecgosync))
diff --git a/src/pkg/runtime/malloc.go b/src/pkg/runtime/malloc.go
index 883ca0cef..664b03b15 100644
--- a/src/pkg/runtime/malloc.go
+++ b/src/pkg/runtime/malloc.go
@@ -743,7 +743,7 @@ func runfinq() {
 				default:
 					gothrow("bad kind in runfinq")
 				}
-				reflectcall(unsafe.Pointer(f.fn), frame, uint32(framesz), uint32(framesz))
+				reflectcall(unsafe.Pointer(f.fn), frame, uint32(framesz), uint32(framesz), nil)
 
 				// drop finalizer queue references to finalized object
 				f.fn = nil
diff --git a/src/pkg/runtime/panic.go b/src/pkg/runtime/panic.go
index 1e35561d1..b8fa213f6 100644
--- a/src/pkg/runtime/panic.go
+++ b/src/pkg/runtime/panic.go
@@ -208,7 +208,7 @@ func Goexit() {
 	for gp._defer != nil {
 		d := gp._defer
 		gp._defer = d.link
-		reflectcall(unsafe.Pointer(d.fn), unsafe.Pointer(&d.args), uint32(d.siz), uint32(d.siz))
+		reflectcall(unsafe.Pointer(d.fn), unsafe.Pointer(&d.args), uint32(d.siz), uint32(d.siz), nil)
 		freedefer(d)
 		// Note: we ignore recovers here because Goexit isn't a panic
 	}
diff --git a/src/pkg/runtime/panic1.go b/src/pkg/runtime/panic1.go
index 6d939703d..7bdfb4b2c 100644
--- a/src/pkg/runtime/panic1.go
+++ b/src/pkg/runtime/panic1.go
@@ -56,10 +56,17 @@ func gopanic(e interface{}) {
 		dabort.link = gp._defer
 		gp._defer = (*_defer)(noescape(unsafe.Pointer(&dabort)))
 		p._defer = d
+		p.outerwrap = gp.panicwrap
 
-		newstackcall(d.fn, unsafe.Pointer(&d.args), uint32(d.siz))
+		// TODO(rsc): I am pretty sure the panicwrap manipulation here is not correct.
+		// It is close enough to pass all the tests we have, but I think it needs to be
+		// restored during recovery too. I will write better tests and fix it in a separate CL.
 
-		// Newstackcall did not panic. Remove dabort.
+		gp.panicwrap = 0
+		reflectcall(unsafe.Pointer(d.fn), unsafe.Pointer(&d.args), uint32(d.siz), uint32(d.siz), (*_panic)(noescape(unsafe.Pointer(&p))))
+		gp.panicwrap = p.outerwrap
+
+		// reflectcall did not panic. Remove dabort.
 		if gp._defer != &dabort {
 			gothrow("bad defer entry in panic")
 		}
@@ -114,9 +121,11 @@ func gorecover(argp uintptr) interface{} {
 	// while they are active on the stack. The linker emits adjustments of
 	// g.panicwrap in the prologue and epilogue of functions marked as wrappers.
 	gp := getg()
-	top := (*stktop)(unsafe.Pointer(gp.stackbase))
 	p := gp._panic
-	if p != nil && !p.recovered && top._panic && argp == gp.stackbase-uintptr(top.argsize+gp.panicwrap) {
+	//	if p != nil {
+	//		println("recover?", p, p.recovered, hex(argp), hex(p.argp), uintptr(gp.panicwrap), p != nil && !p.recovered && argp == p.argp-uintptr(gp.panicwrap))
+	//	}
+	if p != nil && !p.recovered && argp == p.argp-uintptr(gp.panicwrap) {
 		p.recovered = true
 		return p.arg
 	}
diff --git a/src/pkg/runtime/runtime.h b/src/pkg/runtime/runtime.h
index b3d1a9422..4f279db3f 100644
--- a/src/pkg/runtime/runtime.h
+++ b/src/pkg/runtime/runtime.h
@@ -452,7 +452,6 @@ struct	Stktop
 	uint32	panicwrap;
 
 	uint8*	argp;	// pointer to arguments in old frame
-	bool	panic;	// is this frame the top of a panic?
 };
 struct	SigTab
 {
@@ -658,6 +657,8 @@ struct Panic
 	Eface	arg;		// argument to panic
 	Panic*	link;		// link to earlier panic
 	Defer*	defer;		// current executing defer
+	uintptr	argp;		// pointer to arguments of deferred call, for recover
+	uint32	outerwrap;	// outer gp->panicwrap
 	bool	recovered;	// whether this panic is over
 	bool	aborted;	// the panic was aborted
 };
@@ -1015,8 +1016,6 @@ void	runtime·printcomplex(Complex128);
 /*
  * runtime go-called
  */
-void	runtime·newstackcall(FuncVal*, byte*, uint32);
-void	reflect·call(FuncVal*, byte*, uint32, uint32);
 void	runtime·gopanic(Eface);
 void	runtime·panicindex(void);
 void	runtime·panicslice(void);
diff --git a/src/pkg/runtime/stack.c b/src/pkg/runtime/stack.c
index f0861e408..facf0c5e8 100644
--- a/src/pkg/runtime/stack.c
+++ b/src/pkg/runtime/stack.c
@@ -25,6 +25,8 @@ enum
 	StackFaultOnFree = 0,	// old stacks are mapped noaccess to detect use after free
 
 	StackCache = 1,
+	
+	StackCopyAlways = 1,	// expect to be able to copy stacks 100% of the time
 };
 
 // Global pool of spans that have free stacks.
@@ -324,6 +326,9 @@ runtime·oldstack(void)
 	int64 goid;
 	int32 oldstatus;
 
+	if(StackCopyAlways)
+		runtime·throw("unexpected call to oldstack");
+
 	gp = g->m->curg;
 	top = (Stktop*)gp->stackbase;
 	if(top == nil)
@@ -442,14 +447,14 @@ checkframecopy(Stkframe *frame, void *arg)
 		stackmap = runtime·funcdata(f, FUNCDATA_LocalsPointerMaps);
 		if(stackmap == nil) {
 			cinfo->frames = -1;
-			if(StackDebug >= 1)
-				runtime·printf("copystack: no locals info for %s\n", runtime·funcname(f));
+			if(StackDebug >= 1 || StackCopyAlways)
+				runtime·printf("runtime: copystack: no locals info for %s\n", runtime·funcname(f));
 			return false;
 		}
 		if(stackmap->n <= 0) {
 			cinfo->frames = -1;
-			if(StackDebug >= 1)
-				runtime·printf("copystack: locals size info only for %s\n", runtime·funcname(f));
+			if(StackDebug >= 1 || StackCopyAlways)
+				runtime·printf("runtime: copystack: locals size info only for %s\n", runtime·funcname(f));
 			return false;
 		}
 	}
@@ -457,8 +462,8 @@ checkframecopy(Stkframe *frame, void *arg)
 		stackmap = runtime·funcdata(f, FUNCDATA_ArgsPointerMaps);
 		if(stackmap == nil) {
 			cinfo->frames = -1;
-			if(StackDebug >= 1)
-				runtime·printf("copystack: no arg info for %s\n", runtime·funcname(f));
+			if(StackDebug >= 1 || StackCopyAlways)
+				runtime·printf("runtime: copystack: no arg info for %s\n", runtime·funcname(f));
 			return false;
 		}
 	}
@@ -510,8 +515,8 @@ copyabletopsegment(G *gp)
 			continue;
 		f = runtime·findfunc((uintptr)fn->fn);
 		if(f == nil) {
-			if(StackDebug >= 1)
-				runtime·printf("copystack: no func for deferred pc %p\n", fn->fn);
+			if(StackDebug >= 1 || StackCopyAlways)
+				runtime·printf("runtime: copystack: no func for deferred pc %p\n", fn->fn);
 			return -1;
 		}
 
@@ -522,14 +527,14 @@ copyabletopsegment(G *gp)
 		// C (particularly, cgo) lies to us.  See issue 7695.
 		stackmap = runtime·funcdata(f, FUNCDATA_ArgsPointerMaps);
 		if(stackmap == nil || stackmap->n <= 0) {
-			if(StackDebug >= 1)
-				runtime·printf("copystack: no arg info for deferred %s\n", runtime·funcname(f));
+			if(StackDebug >= 1 || StackCopyAlways)
+				runtime·printf("runtime: copystack: no arg info for deferred %s\n", runtime·funcname(f));
 			return -1;
 		}
 		stackmap = runtime·funcdata(f, FUNCDATA_LocalsPointerMaps);
 		if(stackmap == nil || stackmap->n <= 0) {
-			if(StackDebug >= 1)
-				runtime·printf("copystack: no local info for deferred %s\n", runtime·funcname(f));
+			if(StackDebug >= 1 || StackCopyAlways)
+				runtime·printf("runtime: copystack: no local info for deferred %s\n", runtime·funcname(f));
 			return -1;
 		}
 
@@ -755,20 +760,17 @@ adjustdefers(G *gp, AdjustInfo *adjinfo)
 static void
 adjustpanics(G *gp, AdjustInfo *adjinfo)
 {
-	Panic *p;
+	Panic *p, **l;
 
 	// only the topmost panic is on the current stack
-	p = gp->panic;
-	if(p == nil)
-		return;
-	if(p->link != nil) {
-		// only the topmost panic can be on the current stack
-		// (because panic runs defers on a new stack)
-		if(adjinfo->oldstk <= (byte*)p->link && (byte*)p->link < adjinfo->oldbase)
-			runtime·throw("two panics on one stack");
+	for(l = &gp->panic; (p = *l) != nil; ) {
+		if(adjinfo->oldstk <= (byte*)p && (byte*)p < adjinfo->oldbase)
+			*l = (Panic*)((byte*)p + adjinfo->delta);
+		l = &p->link;
+		
+		if(adjinfo->oldstk <= (byte*)p->argp && (byte*)p->argp < adjinfo->oldbase)
+			p->argp += adjinfo->delta;
 	}
-	if(adjinfo->oldstk <= (byte*)p && (byte*)p < adjinfo->oldbase)
-		gp->panic = (Panic*)((byte*)p + adjinfo->delta);
 }
 
 static void
@@ -867,11 +869,10 @@ runtime·round2(int32 x)
 	return 1 << s;
 }
 
-// Called from runtime·newstackcall or from runtime·morestack when a new
-// stack segment is needed.  Allocate a new stack big enough for
-// m->moreframesize bytes, copy m->moreargsize bytes to the new frame,
-// and then act as though runtime·lessstack called the function at
-// m->morepc.
+// Called from runtime·morestack when a new stack segment is needed.
+// Allocate a new stack big enough for m->moreframesize bytes,
+// copy m->moreargsize bytes to the new frame,
+// and then act as though runtime·lessstack called the function at m->morepc.
 //
 // g->atomicstatus will be Grunning, Gsyscall or Gscanrunning, Gscansyscall upon entry. 
 // If the GC is trying to stop this g then it will set preemptscan to true.
@@ -879,14 +880,13 @@ void
 runtime·newstack(void)
 {
 	int32 framesize, argsize, oldstatus, oldsize, newsize, nframes;
-	Stktop *top, *oldtop;
+	Stktop *top;
 	byte *stk, *oldstk, *oldbase;
 	uintptr sp;
 	uintptr *src, *dst, *dstend;
 	G *gp;
 	Gobuf label, morebuf;
 	void *moreargp;
-	bool newstackcall;
 
 	if(g->m->forkstackguard)
 		runtime·throw("split stack after fork");
@@ -894,6 +894,8 @@ runtime·newstack(void)
 		runtime·printf("runtime: newstack called from g=%p\n"
 			"\tm=%p m->curg=%p m->g0=%p m->gsignal=%p\n",
 			g->m->morebuf.g, g->m, g->m->curg, g->m->g0, g->m->gsignal);
+		morebuf = g->m->morebuf;
+		runtime·traceback(morebuf.pc, morebuf.sp, morebuf.lr, morebuf.g);
 		runtime·throw("runtime: wrong goroutine in newstack");
 	}
 	if(g->throwsplit)
@@ -917,13 +919,8 @@ runtime·newstack(void)
 
 	runtime·casgstatus(gp, oldstatus, Gwaiting); // oldstatus is not in a Gscan status
 	gp->waitreason = runtime·gostringnocopy((byte*)"stack growth");
-	newstackcall = framesize==1;
-	if(newstackcall)
-		framesize = 0;
 
-	// For newstackcall the context already points to beginning of runtime·newstackcall.
-	if(!newstackcall)
-		runtime·rewindmorestack(&gp->sched);
+	runtime·rewindmorestack(&gp->sched);
 
 	if(gp->stackbase == 0)
 		runtime·throw("nil stackbase");
@@ -1008,6 +1005,9 @@ runtime·newstack(void)
 		// end of C code to trigger a copy as soon as C code exits.  That way, we'll
 		// have stack available if we get this deep again.
 	}
+	
+	if(StackCopyAlways)
+		runtime·throw("split stack not allowed");
 
 	// allocate new segment.
 	framesize += argsize;
@@ -1033,17 +1033,6 @@ runtime·newstack(void)
 	top->argp = moreargp;
 	top->argsize = argsize;
 
-	// copy flag from panic
-	top->panic = gp->ispanic;
-	gp->ispanic = false;
-	
-	// if this isn't a panic, maybe we're splitting the stack for a panic.
-	// if we're splitting in the top frame, propagate the panic flag
-	// forward so that recover will know we're in a panic.
-	oldtop = (Stktop*)top->stackbase;
-	if(oldtop != nil && oldtop->panic && top->argp == (byte*)oldtop - oldtop->argsize - gp->panicwrap)
-		top->panic = true;
-
 	top->panicwrap = gp->panicwrap;
 	gp->panicwrap = 0;
 
@@ -1060,6 +1049,7 @@ runtime·newstack(void)
 		while(dst < dstend)
 			*dst++ = *src++;
 	}
+	
 	if(thechar == '5') {
 		// caller would have saved its LR below args.
 		sp -= sizeof(void*);
@@ -1072,12 +1062,8 @@ runtime·newstack(void)
 	label.sp = sp;
 	label.pc = (uintptr)runtime·lessstack;
 	label.g = g->m->curg;
-	if(newstackcall)
-		runtime·gostartcallfn(&label, (FuncVal*)g->m->cret);
-	else {
-		runtime·gostartcall(&label, (void(*)(void))gp->sched.pc, gp->sched.ctxt);
-		gp->sched.ctxt = nil;
-	}
+	runtime·gostartcall(&label, (void(*)(void))gp->sched.pc, gp->sched.ctxt);
+	gp->sched.ctxt = nil;
 	runtime·casgstatus(gp, Gwaiting, oldstatus); // oldstatus is Grunning or Gsyscall
 	runtime·gogo(&label);
 
diff --git a/src/pkg/runtime/stubs.go b/src/pkg/runtime/stubs.go
index 7ed4aaa55..c97831a7c 100644
--- a/src/pkg/runtime/stubs.go
+++ b/src/pkg/runtime/stubs.go
@@ -171,8 +171,7 @@ func cputicks() int64
 func mmap(addr unsafe.Pointer, n uintptr, prot, flags, fd int32, off uint32) unsafe.Pointer
 func munmap(addr unsafe.Pointer, n uintptr)
 func madvise(addr unsafe.Pointer, n uintptr, flags int32)
-func newstackcall(fv *funcval, addr unsafe.Pointer, size uint32)
-func reflectcall(fn, arg unsafe.Pointer, n uint32, retoffset uint32)
+func reflectcall(fn, arg unsafe.Pointer, n uint32, retoffset uint32, p *_panic)
 func osyield()
 func procyield(cycles uint32)
 func cgocallback_gofunc(fv *funcval, frame unsafe.Pointer, framesize uintptr)
diff --git a/test/fixedbugs/issue4388.go b/test/fixedbugs/issue4388.go
index 2e052e138..b18c98bac 100644
--- a/test/fixedbugs/issue4388.go
+++ b/test/fixedbugs/issue4388.go
@@ -17,18 +17,18 @@ type T struct {
 }
 
 func f1() {
-	// The 4 here and below depends on the number of internal runtime frames
+	// The 5 here and below depends on the number of internal runtime frames
 	// that sit between a deferred function called during panic and
 	// the original frame. If that changes, this test will start failing and
 	// the number here will need to be updated.
-	defer checkLine(4)
+	defer checkLine(5)
 	var t *T
 	var c io.Closer = t
 	c.Close()
 }
 
 func f2() {
-	defer checkLine(4)
+	defer checkLine(5)
 	var t T
 	var c io.Closer = t
 	c.Close()
diff --git a/test/fixedbugs/issue5856.go b/test/fixedbugs/issue5856.go
index 35cadf8c9..78ca3b9f6 100644
--- a/test/fixedbugs/issue5856.go
+++ b/test/fixedbugs/issue5856.go
@@ -29,7 +29,7 @@ func f() {
 }
 
 func g() {
-	_, file, line, _ := runtime.Caller(2)
+	_, file, line, _ := runtime.Caller(3)
 	if !strings.HasSuffix(file, "issue5856.go") || line != 28 {
 		fmt.Printf("BUG: defer called from %s:%d, want issue5856.go:28\n", file, line)
 		os.Exit(1)