liblink: generate MRC replacement in liblink, not tls_arm

Fixes #8690.

This CL moves the save of LR around BL runtime.read_tls_fallback to liblink as it is not needed when MRC is not replaced.

LGTM=rsc, minux
R=rsc, khr, minux
CC=golang-codereviews
https://golang.org/cl/147310043

src/liblink/obj5.c

@@ -119,14 +119,30 @@ progedit(Link *ctxt, Prog *p)
 				ctxt->diag("%L: TLS MRC instruction must write to R0 as it might get translated into a BL instruction", p->lineno);
 
 			if(ctxt->goarm < 7) {
-				// Replace it with BL runtime.read_tls_fallback(SB).
+				// Replace it with BL runtime.read_tls_fallback(SB) for ARM CPUs that lack the tls extension.
 				if(tlsfallback == nil)
 					tlsfallback = linklookup(ctxt, "runtime.read_tls_fallback", 0);
+				// MOVW	LR, R11
+				p->as = AMOVW;
+				p->from.type = D_REG;
+				p->from.reg = REGLINK;
+				p->to.type = D_REG;
+				p->to.reg = REGTMP;
+
 				// BL	runtime.read_tls_fallback(SB)
+				p = appendp(ctxt, p);
 				p->as = ABL;
 				p->to.type = D_BRANCH;
 				p->to.sym = tlsfallback;
 				p->to.offset = 0;
+
+				// MOVW	R11, LR
+				p = appendp(ctxt, p);
+				p->as = AMOVW;
+				p->from.type = D_REG;
+				p->from.reg = REGTMP;
+				p->to.type = D_REG;
+				p->to.reg = REGLINK;
 				break;
 			}
 		}

src/runtime/tls_arm.s

@@ -31,11 +31,8 @@ TEXT runtime·save_g(SB),NOSPLIT,$-4
 #endif
 	// If the host does not support MRC the linker will replace it with
 	// a call to runtime.read_tls_fallback which jumps to __kuser_get_tls.
-	// Both functions are written to only disturb R0 so it should be safe to 
-	// use R11 here to temporarily store LR.
-	MOVW	LR, R11
+	// The replacement function saves LR in R11 over the call to read_tls_fallback.
 	MRC	15, 0, R0, C13, C0, 3 // fetch TLS base pointer
-	MOVW	R11, LR
 	// $runtime.tlsg(SB) is a special linker symbol.
 	// It is the offset from the TLS base pointer to our
 	// thread-local storage for g.
@@ -57,10 +54,8 @@ TEXT runtime·load_g(SB),NOSPLIT,$0
 	// nothing to do as nacl/arm does not use TLS at all.
 	RET
 #endif
-	// See comment in save_g.
-	MOVW	LR, R11
+	// See save_g
 	MRC	15, 0, R0, C13, C0, 3 // fetch TLS base pointer
-	MOVW	R11, LR
 	// $runtime.tlsg(SB) is a special linker symbol.
 	// It is the offset from the TLS base pointer to our
 	// thread-local storage for g.