runtime: correct seh installation during callbacks

Every time we enter callback from Windows, it is
possible that go exception handler is not at the top
of per-thread exception handlers chain. So it needs
to be installed again. At this moment this is done
by replacing top SEH frame with SEH frame as at time
of syscall for the time of callback. This is incorrect,
because, if exception strike, we won't be able to call
any exception handlers installed inside syscall,
because they are not in the chain. This changes
procedure to add new SEH frame on top of existing
chain instead.

I also removed m sehframe field, because I don't
think it is needed. We use single global exception
handler everywhere.

R=golang-dev, r
CC=golang-dev, hectorchu
https://golang.org/cl/4832060

src/pkg/runtime/runtime.h

@@ -248,12 +248,9 @@ struct	M
 	uint32	freghi[16];	// D[i] msb and F[i+16]
 	uint32	fflag;		// floating point compare flags
 #ifdef __WINDOWS__
-	void*	sehframe;
-
 #ifdef _64BIT	
 	void*	gostack;
 #endif
-
 #endif
 };
 

src/pkg/runtime/windows/386/sys.s

@@ -14,8 +14,6 @@ TEXT runtime·stdcall_raw(SB),7,$0
 	// Switch to m->g0 if needed.
 	get_tls(DI)
 	MOVL	m(DI), DX
-	MOVL	0(FS), SI
-	MOVL	SI, m_sehframe(DX)
 	MOVL	m_g0(DX), SI
 	CMPL	g(DI), SI
 	MOVL	SP, BX
@@ -116,7 +114,7 @@ TEXT runtime·ctrlhandler(SB),7,$0
 	MOVL	SP, BX
 
 	// setup dummy m, g
-	SUBL	$(m_sehframe+4), SP	// at least space for m_sehframe
+	SUBL	$(m_fflag+4), SP	// at least space for m_fflag
 	LEAL	m_tls(SP), CX
 	MOVL	CX, 0x2c(FS)
 	MOVL	SP, m(CX)
@@ -159,33 +157,39 @@ TEXT runtime·callbackasm+0(SB),7,$0
 	ADDL	$4, DX			// extend argsize by size of return value
 
 	// save registers as required for windows callback
-	PUSHL	0(FS)
 	PUSHL	DI
 	PUSHL	SI
 	PUSHL	BP
 	PUSHL	BX
+
+	// set up SEH frame again
+	PUSHL	$runtime·sigtramp(SB)
+	PUSHL	0(FS)
+	MOVL	SP, 0(FS)
+
+	// callback parameters
 	PUSHL	DX
 	PUSHL	CX
 	PUSHL	AX
 
-	// reinstall our SEH handler
-	get_tls(CX)
-	MOVL	m(CX), CX
-	MOVL	m_sehframe(CX), CX
-	MOVL	CX, 0(FS)
 	CLD
 
 	CALL	runtime·cgocallback(SB)
 
-	// restore registers as required for windows callback
 	POPL	AX
 	POPL	CX
 	POPL	DX
+
+	// pop SEH frame
+	POPL	0(FS)
+	POPL	BX
+
+	// restore registers as required for windows callback
 	POPL	BX
 	POPL	BP
 	POPL	SI
 	POPL	DI
-	POPL	0(FS)
+
 	CLD
 
 	MOVL	-4(CX)(DX*1), AX