On Tue, Aug 12, 2025 at 10:02 PM H.J. Lu <hjl.to...@gmail.com> wrote:
>
> On Tue, Aug 12, 2025 at 06:47:54AM -0700, H.J. Lu wrote:
> > On Mon, Aug 11, 2025 at 11:13 PM Hongtao Liu <crazy...@gmail.com> wrote:
> > >
> > > On Mon, Aug 4, 2025 at 11:33 PM H.J. Lu <hjl.to...@gmail.com> wrote:
> > > >
> > > > On Mon, Aug 04, 2025 at 02:57:39PM +0800, Hongtao Liu wrote:
> > > > > > > > > + rtx_insn *before = nullptr;
> > > > > > > > > + rtx_insn *after = nullptr;
> > > > > > > > > + if (insn == BB_HEAD (bb))
> > > > > > > > > + before = insn;
> > > > > > > > > + else
> > > > > > > > > + after = insn ? PREV_INSN (insn) : BB_END (bb);
> > > > > > > > > +
> > > > > > > > > + /* TLS_GD and TLS_LD_BASE instructions are normal
> > > > > > > > > functions which
> > > > > > > > > + clobber caller-saved registers. TLSDESC instructions
> > > > > > > > > are special
> > > > > > > > > + functions which only clobber RAX. If any registers
> > > > > > > > > clobbered by
> > > > > > > > > + the TLS instruction are live in this basic block, we
> > > > > > > > > must insert
> > > > > > > > > + the TLS instruction after all live registers clobbered
> > > > > > > > > by the TLS
> > > > > > > > > + instruction are dead. */
> > > > > > > > > +
> > > > > > > > > + auto_bitmap live_caller_saved_regs;
> > > > > > > > > + bitmap in = df_live ? DF_LIVE_IN (bb) : DF_LR_IN (bb);
> > > > > > > > > +
> > > > > > > > > + bool flags_live_p = bitmap_bit_p (in, FLAGS_REG);
> > > > > > > > > +
> > > > > > > > > + unsigned int i;
> > > > > > > > > +
> > > > > > > > > + /* Get all live caller-saved registers. */
> > > > > > > > > + if (kind == X86_CSE_TLSDESC)
> > > > > > > > > + {
> > > > > > > > > + if (bitmap_bit_p (in, AX_REG))
> > > > > > > > > + bitmap_set_bit (live_caller_saved_regs, AX_REG);
> > > > > > > >
> > > > > > > > And we don't need to check for those hard registers here and
> > > > > > > > below?
> > > > > > >
> > > > > > > TLS_GD and TLS_LD_BASE instructions are normal functions which
> > > > > > > clobber caller-saved registers. TLSDESC instructions are special
> > > > > > > functions which only clobber RAX. live_caller_saved_regs captures
> > > > > > > live caller-saved registers for these TLS instructions.
> > > > >
> > > > > I notice those insns are CALL_INSN, and for ABI, rax/rdi/rsi is
> > > > > caller_saved registers, so even we explicitly use (clobber (reg: RAX))
> > > >
> > > > Since a single TLS call will clobber caller-saved registers, it must
> > > > be placed where all caller-saved registers are dead. Otherwise, the
> > > > TLS call will clobber some live registers.
> > > I saw in legitimize_tls_address, it doesn't check the liveness of
> > > RAX/RDI(or call clobber registers) even with explicit use of RDI/RAX.
> > > I'm wondering if we can reuse that code to do similar things.
> >
> > Before RA, hard registers are only used as scratch registers:
> >
> > insn 20 set rax ....
> > insn 21 set r300 rax # rax is dead after this.
> >
> > Now we are inserting another rax usage in the same basic block.
> >
> > set rax ...
> > set r400 rax
> >
> > we must insert them where rax isn't alive. We can't insert them
> > between insn 20 and insn 21. My patch inserts them where all
> > caller-saved registers are dead.
I'll try to rewrite ix86_place_single_tls_call to see if it's possible.
will get back to you later.
>
> gcc.target/i386/pr81501-4a.c is an example. We want to insert
>
> (call_insn/u 81 74 4 2 (parallel [
> (set (reg:DI 0 ax)
> (call:DI (mem:QI (symbol_ref:DI ("__tls_get_addr")) [0 S1
> A8])
> (const_int 0 [0])))
> (unspec:DI [
> (symbol_ref:DI ("foo") [flags 0x10] <var_decl
> 0x7fb4f0dd6e40 foo>)
> (reg/f:DI 7 sp)
> ] UNSPEC_TLS_GD)
> (clobber (reg:DI 5 di))
> ]) "pr81501-4a.c":26:1 1657 {*tls_global_dynamic_64_di}
> (expr_list:REG_EH_REGION (const_int -2147483648 [0xffffffff80000000])
> (nil))
> (nil))
>
> in a basic block with
>
> (note 7 0 73 2 [bb 2] NOTE_INSN_BASIC_BLOCK)
> (insn 73 7 74 2 (set (reg:SI 117 [ n ])
> (reg:SI 5 di [ n ])) "pr81501-4a.c":26:1 100 {*movsi_internal}
> (expr_list:REG_DEAD (reg:SI 5 di [ n ])
> (nil)))
> (insn 74 73 4 2 (set (reg/f:DI 118 [ caller_foop ])
> (reg:DI 4 si [ caller_foop ])) "pr81501-4a.c":26:1 99
> {*movdi_internal}
> (expr_list:REG_DEAD (reg:DI 4 si [ caller_foop ])
> (nil)))
> (note 4 74 9 2 NOTE_INSN_FUNCTION_BEG)
> (insn 9 4 10 2 (set (reg/f:DI 105)
> (symbol_ref/f:DI ("*.LC0") [flags 0x2] <var_decl 0x7fffe99fb5f0
> *.LC0>)) "pr81501-4a.c":30:3 99 {*movdi_internal}
> (nil))
> (insn 10 9 11 2 (set (reg:DI 5 di)
> (reg/f:DI 105)) "pr81501-4a.c":30:3 99 {*movdi_internal}
> (expr_list:REG_DEAD (reg/f:DI 105)
> (expr_list:REG_EQUAL (symbol_ref/f:DI ("*.LC0") [flags 0x2] <var_decl
> 0x7fffe99fb5f0 *.LC0>)
> (nil))))
> (call_insn 11 10 12 2 (call (mem:QI (symbol_ref:DI ("bar3") [flags 0x41]
> <function_decl 0x7fffe99cdb00 bar3>) [0 bar3 S1 A8])
> (const_int 0 [0])) "pr81501-4a.c":30:3 1469 {*call}
> (expr_list:REG_DEAD (reg:DI 5 di)
> (expr_list:REG_CALL_DECL (symbol_ref:DI ("bar3") [flags 0x41]
> <function_decl 0x7fffe99cdb00 bar3>)
> (nil)))
> (expr_list:DI (use (reg:DI 5 di))
> (nil)))
>
> We must place it after both RAX and RDI are dead, which is after
>
> (insn 74 73 4 2 (set (reg/f:DI 118 [ caller_foop ])
> (reg:DI 4 si [ caller_foop ])) "pr81501-4a.c":26:1 99
> {*movdi_internal}
> (expr_list:REG_DEAD (reg:DI 4 si [ caller_foop ])
> (nil)))
>
>
> H.J.
--
BR,
Hongtao
