On Tue, Jan 21, 2020 at 8:16 PM H.J. Lu <hjl.to...@gmail.com> wrote: > > On Tue, Jan 21, 2020 at 2:29 AM Uros Bizjak <ubiz...@gmail.com> wrote: > > > > On Tue, Jan 21, 2020 at 9:47 AM Uros Bizjak <ubiz...@gmail.com> wrote: > > > > > > On Mon, Jan 20, 2020 at 10:46 PM H.J. Lu <hjl.to...@gmail.com> wrote: > > > > > > > > > OK. Let's go with this version, but please investigate if we need to > > > > > > calculate TLS address in ptr_mode instead of Pmode. Due to quite > > > > > > some > > > > > > zero-extension from ptr_mode to Pmode hacks in this area, it looks > > > > > > to > > > > > > me that the whole calculation should be performed in ptr_mode > > > > > > (SImode > > > > > > in case of x32), and the result zero-extended to Pmode in case when > > > > > > Pmode = DImode. > > > > > > > > > > > > > > > > I checked it in. I will investigate if we can use ptr_mode for TLS. > > > > > > > > Here is a patch to perform GNU2 TLS address computation in ptr_mode > > > > and zero-extend result to Pmode. > > > > > > case TLS_MODEL_GLOBAL_DYNAMIC: > > > - dest = gen_reg_rtx (Pmode); > > > + dest = gen_reg_rtx (TARGET_GNU2_TLS ? ptr_mode : Pmode); > > > > > > Please put these in their respective arms of "if (TARGET_GNU2_TLS). > > > > > > case TLS_MODEL_LOCAL_DYNAMIC: > > > - base = gen_reg_rtx (Pmode); > > > + base = gen_reg_rtx (TARGET_GNU2_TLS ? ptr_mode : Pmode); > > > > > > Also here. > > > > > > A question: Do we need to emit the following part in Pmode? > > > > To answer my own question: Yes. Linker doesn't like SImode relocs fox > > x86_64 and for > > > > addl $foo@dtpoff, %eax > > > > errors out with: > > > > pr93319-1a.s: Assembler messages: > > pr93319-1a.s:20: Error: relocated field and relocation type differ in > > signedness > > > > So, the part below is OK, except: > > > > - tp = get_thread_pointer (Pmode, true); > > - set_unique_reg_note (get_last_insn (), REG_EQUAL, > > - gen_rtx_MINUS (Pmode, tmp, tp)); > > + tp = get_thread_pointer (ptr_mode, true); > > + tmp = gen_rtx_MINUS (ptr_mode, tmp, tp); > > + if (GET_MODE (tmp) != Pmode) > > + tmp = gen_rtx_ZERO_EXTEND (Pmode, tmp); > > + set_unique_reg_note (get_last_insn (), REG_EQUAL, tmp); > > > > I don't think we should attach this note to the thread pointer > > initialization. I have removed this part from the patch, but please > > review the decision. > > > > and > > > > - dest = gen_rtx_PLUS (Pmode, tp, dest); > > + dest = gen_rtx_PLUS (ptr_mode, tp, dest); > > > > Please leave Pmode here. ptr_mode == Pmode at this point, but Pmode > > better documents the mode selection logic. > > > > Also, the tests fail for me with: > > > > /usr/include/gnu/stubs.h:13:11: fatal error: gnu/stubs-x32.h: No such > > file or directory > > > > so either use __builtin_printf or some other approach that doesn't > > need to #include stdio.h. > > > > A patch that implements above changes is attached to the message. > > > > Here is the updated patch. OK for master?
OK if it passes regtests. Thanks, Uros.
