The addition of TLS to ARM rewrote the MRC instruction
differently depending on whether we were using internal
or external linking mode. That's clearly not okay, since we
don't know that during compilation, which is when we now
generate the code. Also, because the change did not introduce
a real MRC instruction but instead just macro-expanded it
in the assembler, liblink is rewriting a WORD instruction that
may actually be looking for that specific constant, which would
lead to very unexpected results. It was also using one value
that happened to be 8 where a different value that also
happened to be 8 belonged. So the code was correct for those
values but not correct in general, and very confusing.
Throw it all away.
Replace with the following. There is a linker-provided symbol
runtime.tlsgm with a value (address) set to the offset from the
hardware-provided TLS base register to the g and m storage.
Any reference to that name emits an appropriate TLS relocation
to be resolved by either the internal linker or the external linker,
depending on the link mode. The relocation has exactly the
semantics of the R_ARM_TLS_LE32 relocation, which is what
the external linker provides.
This symbol is only used in two routines, runtime.load_gm and
runtime.save_gm. In both cases it is now used like this:
MRC 15, 0, R0, C13, C0, 3 // fetch TLS base pointer
MOVW $runtime·tlsgm(SB), R2
ADD R2, R0 // now R0 points at thread-local g+m storage
It is likely that this change breaks the generation of shared libraries
on ARM, because the MOVW needs to be rewritten to use the global
offset table and a different relocation type. But let's get the supported
functionality working again before we worry about unsupported
functionality.
LGTM=dave, iant
R=iant, dave
CC=golang-codereviews
https://golang.org/cl/
56120043
diag("unknown reloc %d", r->type);
break;
case D_TLS:
+ if(linkmode == LinkInternal && iself && thechar == '5') {
+ // On ELF ARM, the thread pointer is 8 bytes before
+ // the start of the thread-local data block, so add 8
+ // to the actual TLS offset (r->sym->value).
+ // This 8 seems to be a fundamental constant of
+ // ELF on ARM (or maybe Glibc on ARM); it is not
+ // related to the fact that our own TLS storage happens
+ // to take up 8 bytes.
+ o = 8 + r->sym->value;
+ break;
+ }
r->done = 0;
o = 0;
if(thechar != '6')
}
sect->len = datsize;
} else {
- // References to STLSBSS symbols may be in the binary
- // but should not be used. Give them an invalid address
- // so that any uses will fault. Using 1 instead of 0 so that
- // if used as an offset on ARM it will result in an unaligned
- // address and still cause a fault.
- for(; s != nil && s->type == STLSBSS; s = s->next)
- s->value = 1;
+ // Might be internal linking but still using cgo.
+ // In that case, the only possible STLSBSS symbol is tlsgm.
+ // Give it offset 0, because it's the only thing here.
+ if(s != nil && s->type == STLSBSS && strcmp(s->name, "runtime.tlsgm") == 0) {
+ s->value = 0;
+ s = s->next;
+ }
}
if(s != nil) {
gmsym->type = STLSBSS;
gmsym->size = 2*PtrSize;
gmsym->hide = 1;
- if(linkmode == LinkExternal && iself && HEADTYPE != Hopenbsd)
- gmsym->reachable = 1;
- else
- gmsym->reachable = 0;
+ gmsym->reachable = 1;
// Now that we know the link mode, trim the dynexp list.
x = CgoExportDynamic;
aclass(ctxt, &p->to);
o1 = ctxt->instoffset;
if(p->to.sym != nil) {
+ // This case happens with words generated
+ // in the PC stream as part of the literal pool.
rel = addrel(ctxt->cursym);
rel->off = ctxt->pc;
rel->siz = 4;
rel->sym = p->to.sym;
rel->add = p->to.offset;
+
+ // runtime.tlsgm (aka gmsym) is special.
+ // Its "address" is the offset from the TLS thread pointer
+ // to the thread-local g and m pointers.
+ // Emit a TLS relocation instead of a standard one.
+ // The TLS flag_shared case is not tested and probably now wrong.
if(rel->sym == ctxt->gmsym) {
rel->type = D_TLS;
if(ctxt->flag_shared)
- rel->add += ctxt->pc - p->pcrel->pc - 8 - rel->siz;
+ rel->add += ctxt->pc - p->pcrel->pc - 8 - rel->siz; // TODO: probably wrong
rel->xadd = rel->add;
rel->xsym = rel->sym;
} else if(ctxt->flag_shared) {
{
Prog *p, *pl, *q, *q1, *q2;
int o;
- LSym *tlsfallback;
int32 autosize, autoffset;
autosize = 0;
if(ctxt->symmorestack[0] == nil)
ctxt->symmorestack[0] = linklookup(ctxt, "runtime.morestack", 0);
- tlsfallback = linklookup(ctxt, "runtime.read_tls_fallback", 0);
if(ctxt->gmsym == nil)
ctxt->gmsym = linklookup(ctxt, "runtime.tlsgm", 0);
q = nil;
}
}
break;
- case AWORD:
- // Rewrite TLS register fetch: MRC 15, 0, <reg>, C13, C0, 3
- if((p->to.offset & 0xffff0fff) == 0xee1d0f70) {
- if(ctxt->headtype == Hopenbsd) {
- p->as = ARET;
- break;
- }
- if(ctxt->goarm < 7) {
- // BL runtime.read_tls_fallback(SB)
- p->as = ABL;
- p->to.type = D_BRANCH;
- p->to.sym = tlsfallback;
- p->to.offset = 0;
- cursym->text->mark &= ~LEAF;
- }
- // runtime.tlsgm is relocated with R_ARM_TLS_LE32
- // and $runtime.tlsgm will contain the TLS offset.
- //
- // MOV $runtime.tlsgm+ctxt->tlsoffset(SB), REGTMP
- // ADD REGTMP, <reg>
- //
- // In shared mode, runtime.tlsgm is relocated with
- // R_ARM_TLS_IE32 and runtime.tlsgm(SB) will point
- // to the GOT entry containing the TLS offset.
- //
- // MOV runtime.tlsgm(SB), REGTMP
- // ADD REGTMP, <reg>
- // SUB -ctxt->tlsoffset, <reg>
- //
- // The SUB compensates for ctxt->tlsoffset
- // used in runtime.save_gm and runtime.load_gm.
- q = p;
- p = appendp(ctxt, p);
- p->as = AMOVW;
- p->scond = C_SCOND_NONE;
- p->reg = NREG;
- if(ctxt->flag_shared) {
- p->from.type = D_OREG;
- p->from.offset = 0;
- } else {
- p->from.type = D_CONST;
- p->from.offset = ctxt->tlsoffset;
- }
- p->from.sym = ctxt->gmsym;
- p->from.name = D_EXTERN;
- p->to.type = D_REG;
- p->to.reg = REGTMP;
- p->to.offset = 0;
-
- p = appendp(ctxt, p);
- p->as = AADD;
- p->scond = C_SCOND_NONE;
- p->reg = NREG;
- p->from.type = D_REG;
- p->from.reg = REGTMP;
- p->to.type = D_REG;
- p->to.reg = (q->to.offset & 0xf000) >> 12;
- p->to.offset = 0;
-
- if(ctxt->flag_shared) {
- p = appendp(ctxt, p);
- p->as = ASUB;
- p->scond = C_SCOND_NONE;
- p->reg = NREG;
- p->from.type = D_CONST;
- p->from.offset = -ctxt->tlsoffset;
- p->to.type = D_REG;
- p->to.reg = (q->to.offset & 0xf000) >> 12;
- p->to.offset = 0;
- }
- break;
- }
}
q = p;
}
// Note: all three functions will clobber R0, and the last
// two can be called from 5c ABI code.
+// save_gm saves the g and m registers into pthread-provided
+// thread-local memory, so that we can call externally compiled
+// ARM code that will overwrite those registers.
+// NOTE: runtime.gogo assumes that R1 is preserved by this function.
TEXT runtime·save_gm(SB),NOSPLIT,$0
- // NOTE: Liblink adds some instructions following the MRC
- // to adjust R0 so that 8(R0) and 12(R0) are the TLS copies of
- // the g and m registers. It's a bit too magical for its own good.
- MRC 15, 0, R0, C13, C0, 3 // Fetch TLS register
- MOVW g, 8(R0)
- MOVW m, 12(R0)
+ MRC 15, 0, R0, C13, C0, 3 // fetch TLS base pointer
+ // $runtime.tlsgm(SB) is a special linker symbol.
+ // It is the offset from the TLS base pointer to our
+ // thread-local storage for g and m.
+ MOVW $runtime·tlsgm(SB), R11
+ ADD R11, R0
+ MOVW g, 0(R0)
+ MOVW m, 4(R0)
RET
+// load_gm loads the g and m registers from pthread-provided
+// thread-local memory, for use after calling externally compiled
+// ARM code that overwrote those registers.
TEXT runtime·load_gm(SB),NOSPLIT,$0
- // NOTE: Liblink adds some instructions following the MRC
- // to adjust R0 so that 8(R0) and 12(R0) are the TLS copies of
- // the g and m registers. It's a bit too magical for its own good.
- MRC 15, 0, R0, C13, C0, 3 // Fetch TLS register
- MOVW 8(R0), g
- MOVW 12(R0), m
+ MRC 15, 0, R0, C13, C0, 3 // fetch TLS base pointer
+ // $runtime.tlsgm(SB) is a special linker symbol.
+ // It is the offset from the TLS base pointer to our
+ // thread-local storage for g and m.
+ MOVW $runtime·tlsgm(SB), R11
+ ADD R11, R0
+ MOVW 0(R0), g
+ MOVW 4(R0), m
RET
// void setmg_gcc(M*, G*); set m and g called from gcc.