[Bug target/97715] New: [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[Bug target/97715] New: [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

            Bug ID: 97715
           Summary: [11 Regression] ICE in insn_default_length, at
                    config/i386/i386.md:15325 since
                    r11-4578-gd10f3e900b0377b4
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Keywords: ice-on-valid-code
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: marxin at gcc dot gnu.org
                CC: hjl at gcc dot gnu.org, qinzhao at gcc dot gnu.org
  Target Milestone: ---

The following fails:

$ gcc
/home/marxin/Programming/gcc/gcc/testsuite/c-c++-common/zero-scratch-regs-11.c
-mno-80387
In file included from
/home/marxin/Programming/gcc/gcc/testsuite/c-c++-common/zero-scratch-regs-11.c:4:
/home/marxin/Programming/gcc/gcc/testsuite/c-c++-common/zero-scratch-regs-10.c:
In function ‘foo9’:
/home/marxin/Programming/gcc/gcc/testsuite/c-c++-common/zero-scratch-regs-10.c:77:1:
error: unrecognizable insn:
   77 | }
      | ^
(insn 27 67 28 2 (set (reg:XF 8 st)
        (const_double:XF 0.0 [0x0.0p+0]))
"/home/marxin/Programming/gcc/gcc/testsuite/c-c++-common/zero-scratch-regs-10.c":77:1
-1
     (nil))
during RTL pass: shorten
/home/marxin/Programming/gcc/gcc/testsuite/c-c++-common/zero-scratch-regs-10.c:77:1:
internal compiler error: in insn_default_length, at config/i386/i386.md:15325
0x6c85cb _fatal_insn(char const*, rtx_def const*, char const*, int, char
const*)
        /home/marxin/Programming/gcc/gcc/rtl-error.c:108
0x6c85e7 _fatal_insn_not_found(rtx_def const*, char const*, int, char const*)
        /home/marxin/Programming/gcc/gcc/rtl-error.c:116
0x7afca5 insn_default_length(rtx_insn*)
        /home/marxin/Programming/gcc/gcc/config/i386/i386.md:15325
0xa90559 shorten_branches(rtx_insn*)
        /home/marxin/Programming/gcc/gcc/final.c:1118
0xa905bf rest_of_handle_shorten_branches
        /home/marxin/Programming/gcc/gcc/final.c:4753
0xa905bf execute
        /home/marxin/Programming/gcc/gcc/final.c:4782
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

Martin Liška <marxin at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to fail|                            |11.0
   Last reconfirmed|                            |2020-11-04
      Known to work|                            |10.2.0
   Target Milestone|---                         |11.0
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qinzhao at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #1 from qinzhao at gcc dot gnu.org ---
for -fzero-call-used-regs=all, when zeroing st/mm registers under x87 exit
mode, However, command line option force no 80387 mode, the following insn
generated to zero st registers is not recognized:

(insn 27 67 28 2 (set (reg:XF 8 st)
        (const_double:XF 0.0 [0x0.0p+0])) "zero-scratch-regs-10.c":8:1 -1
     (nil))

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|hjl at gcc dot gnu.org             |hjl.tools at gmail dot com

--- Comment #2 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to qinzhao from comment #1)
> for -fzero-call-used-regs=all, when zeroing st/mm registers under x87 exit
> mode, However, command line option force no 80387 mode, the following insn
> generated to zero st registers is not recognized:
> 
> (insn 27 67 28 2 (set (reg:XF 8 st)
>         (const_double:XF 0.0 [0x0.0p+0])) "zero-scratch-regs-10.c":8:1 -1
>      (nil))

You should avoid zeroing fixed registers.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
;; Floating-point register constraints.
(define_register_constraint "f"
 "TARGET_80387 || TARGET_FLOAT_RETURNS_IN_80387 ? FLOAT_REGS : NO_REGS"
 "Any 80387 floating-point (stack) register.")

So, zero_all_st_registers really should just
  if (!TARGET_80387 && !TARGET_FLOAT_RETURNS_IN_80387)
    return false;
somewhere early.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Another thing to check is how does this option handle e.g. TUs that have some
registers fixed and use global register vars.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qinzhao at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #5 from qinzhao at gcc dot gnu.org ---
(In reply to H.J. Lu from comment #2)
> (In reply to qinzhao from comment #1)
> > for -fzero-call-used-regs=all, when zeroing st/mm registers under x87 exit
> > mode, However, command line option force no 80387 mode, the following insn
> > generated to zero st registers is not recognized:
> > 
> > (insn 27 67 28 2 (set (reg:XF 8 st)
> >         (const_double:XF 0.0 [0x0.0p+0])) "zero-scratch-regs-10.c":8:1 -1
> >      (nil))
> 
> You should avoid zeroing fixed registers.

fixed registers should already be excluded from zeroing. 
are ST registers considered FIXED registers when -mno-80387 is specified?

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qinzhao at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #6 from qinzhao at gcc dot gnu.org ---
(In reply to Jakub Jelinek from comment #3)
> ;; Floating-point register constraints.
> (define_register_constraint "f"
>  "TARGET_80387 || TARGET_FLOAT_RETURNS_IN_80387 ? FLOAT_REGS : NO_REGS"
>  "Any 80387 floating-point (stack) register.")
> 
> So, zero_all_st_registers really should just
>   if (!TARGET_80387 && !TARGET_FLOAT_RETURNS_IN_80387)
>     return false;
> somewhere early.

I can do this to avoid the ICE. however, the st/mm register set will NOT be
zeroed under such situation when user requests to zero them with
-fzero-call-used-regs=all.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #7 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to qinzhao from comment #5)
> (In reply to H.J. Lu from comment #2)
> > (In reply to qinzhao from comment #1)
> > > for -fzero-call-used-regs=all, when zeroing st/mm registers under x87 exit
> > > mode, However, command line option force no 80387 mode, the following insn
> > > generated to zero st registers is not recognized:
> > > 
> > > (insn 27 67 28 2 (set (reg:XF 8 st)
> > >         (const_double:XF 0.0 [0x0.0p+0])) "zero-scratch-regs-10.c":8:1 -1
> > >      (nil))
> > 
> > You should avoid zeroing fixed registers.
> 
> fixed registers should already be excluded from zeroing. 
> are ST registers considered FIXED registers when -mno-80387 is specified?

They used to be, but now they are cleared from accessible_reg_set.

  /* If the FPU is disabled, disable the registers.  */
  if (! (TARGET_80387 || TARGET_FLOAT_RETURNS_IN_80387))
    accessible_reg_set &= ~reg_class_contents[FLOAT_REGS];

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
It is not fixed, but it is removed from accessible_reg_set.
The same way how e.g. MMX regs are removed from it if !TARGET_MMX or SSE
registers if !TARGET_SSE.
You shouldn't try to zero unaccessible registers either.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to qinzhao from comment #6)
> (In reply to Jakub Jelinek from comment #3)
> > ;; Floating-point register constraints.
> > (define_register_constraint "f"
> >  "TARGET_80387 || TARGET_FLOAT_RETURNS_IN_80387 ? FLOAT_REGS : NO_REGS"
> >  "Any 80387 floating-point (stack) register.")
> > 
> > So, zero_all_st_registers really should just
> >   if (!TARGET_80387 && !TARGET_FLOAT_RETURNS_IN_80387)
> >     return false;
> > somewhere early.
> 
> I can do this to avoid the ICE. however, the st/mm register set will NOT be
> zeroed under such situation when user requests to zero them with
> -fzero-call-used-regs=all.

True, but we can't do anything about that.  Function has to exit in x87 mode,
even when there are no x87 registers.  That means emms has to be used at the
end of a function, and no %mm can be touched afterwards.

The above proposed solution is correct.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qing.zhao at oracle dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #10 from Qing Zhao <qing.zhao at oracle dot com> ---
> On Nov 4, 2020, at 9:45 AM, ubizjak at gmail dot com <gcc-bugzilla@gcc.gnu.org> wrote:
>> fixed registers should already be excluded from zeroing. 
>> are ST registers considered FIXED registers when -mno-80387 is specified?
> 
> They used to be,

I used the following in middle end to exclude fixed registers from being
zeroed:
      if (fixed_regs[regno])
        continue;
So, looks like that ST registers are not included in fixed_regs when
!TARGET_80387. 
Is this a bug in gcc?

> but now they are cleared from accessible_reg_set.
> 
>  /* If the FPU is disabled, disable the registers.  */
>  if (! (TARGET_80387 || TARGET_FLOAT_RETURNS_IN_80387))
>    accessible_reg_set &= ~reg_class_contents[FLOAT_REGS];

The above means that ST registers are excluded from the accessible_reg_set,
i.e, they cannot be used in the function body anymore? Then, no need to zero
them?

Is this information available to middle end?
Or I have to delete them from the register set in i386 backend?

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I think you should do:
--- gcc/function.c      2020-10-31 17:41:19.756740009 +0100
+++ gcc/function.c      2020-11-04 17:02:51.199298173 +0100
@@ -5871,6 +5871,8 @@ gen_call_used_regs_seq (rtx_insn *ret, u
        continue;
       if (fixed_regs[regno])
        continue;
+      if (!TEST_HARD_REG_BIT (accessible_reg_set, regno))
+       continue;
       if (REGNO_REG_SET_P (live_out, regno))
        continue;
       if (only_gpr

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #12 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Qing Zhao from comment #10)
> > On Nov 4, 2020, at 9:45 AM, ubizjak at gmail dot com <gcc-bugzilla@gcc.gnu.org> wrote:
> >> fixed registers should already be excluded from zeroing. 
> >> are ST registers considered FIXED registers when -mno-80387 is specified?
> > 
> > They used to be,
> 
> I used the following in middle end to exclude fixed registers from being
> zeroed:
>       if (fixed_regs[regno])
>         continue;
> So, looks like that ST registers are not included in fixed_regs when
> !TARGET_80387. 
> Is this a bug in gcc?

fixed_regs members are only set with -ffixed-REG, in addition to members,
initialized by the target-dependent initializer. It is not a bug.

> > but now they are cleared from accessible_reg_set.
> > 
> >  /* If the FPU is disabled, disable the registers.  */
> >  if (! (TARGET_80387 || TARGET_FLOAT_RETURNS_IN_80387))
> >    accessible_reg_set &= ~reg_class_contents[FLOAT_REGS];
> 
> The above means that ST registers are excluded from the accessible_reg_set,
> i.e, they cannot be used in the function body anymore? Then, no need to zero
> them?

Yes. You should exit early from zero_all_st_registers as advised by Jakub in
Comment #3.

> Is this information available to middle end?
> Or I have to delete them from the register set in i386 backend?

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qing.zhao at oracle dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #13 from Qing Zhao <qing.zhao at oracle dot com> ---
> On Nov 4, 2020, at 10:04 AM, jakub at gcc dot gnu.org <gcc-bugzilla@gcc.gnu.org> wrote:
> --- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> I think you should do:
> --- gcc/function.c      2020-10-31 17:41:19.756740009 +0100
> +++ gcc/function.c      2020-11-04 17:02:51.199298173 +0100
> @@ -5871,6 +5871,8 @@ gen_call_used_regs_seq (rtx_insn *ret, u
>        continue;
>       if (fixed_regs[regno])
>        continue;
> +      if (!TEST_HARD_REG_BIT (accessible_reg_set, regno))
> +       continue;
>       if (REGNO_REG_SET_P (live_out, regno))\

Yes, this looks like a better solution. I will add this fix.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #14 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Jakub Jelinek from comment #11)
> I think you should do:
> --- gcc/function.c	2020-10-31 17:41:19.756740009 +0100
> +++ gcc/function.c	2020-11-04 17:02:51.199298173 +0100
> @@ -5871,6 +5871,8 @@ gen_call_used_regs_seq (rtx_insn *ret, u
>  	continue;
>        if (fixed_regs[regno])
>  	continue;
> +      if (!TEST_HARD_REG_BIT (accessible_reg_set, regno))
> +	continue;
>        if (REGNO_REG_SET_P (live_out, regno))
>  	continue;
>        if (only_gpr

These should never be live, otherwise reg allocator is faulty.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
They aren't live.  But that loop checks that only if only_used is true, when
one uses =all, it marks all regs that aren't fixed, aren't live at the end of
the function and a few other checks.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #16 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Jakub Jelinek from comment #15)
> They aren't live.  But that loop checks that only if only_used is true, when
> one uses =all, it marks all regs that aren't fixed, aren't live at the end
> of the function and a few other checks.

Ah, yes. The logic is inverted.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qing.zhao at oracle dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #17 from Qing Zhao <qing.zhao at oracle dot com> ---
> On Nov 4, 2020, at 10:08 AM, ubizjak at gmail dot com <gcc-bugzilla@gcc.gnu.org> wrote:
>> 
>> I used the following in middle end to exclude fixed registers from being
>> zeroed:
>>      if (fixed_regs[regno])
>>        continue;
>> So, looks like that ST registers are not included in fixed_regs when
>> !TARGET_80387. 
>> Is this a bug in gcc?
> 
> fixed_regs members are only set with -ffixed-REG, in addition to members,
> initialized by the target-dependent initializer. It is not a bug.
Yes.

Then, we need to check whether the register is in “accessible_reg_set” to
exclude it from being
Zeroed as Jacub suggested in  Comment #14.  That looks like a better fix.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #18 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
--- gcc/function.c.jj   2020-10-31 17:41:19.756740009 +0100
+++ gcc/function.c      2020-11-04 17:56:53.790403571 +0100
@@ -5871,6 +5871,8 @@ gen_call_used_regs_seq (rtx_insn *ret, u
        continue;
       if (fixed_regs[regno])
        continue;
+      if (!TEST_HARD_REG_BIT (accessible_reg_set, regno))
+       continue;
       if (REGNO_REG_SET_P (live_out, regno))
        continue;
       if (only_gpr
--- gcc/testsuite/gcc.target/i386/zero-scratch-regs-32.c.jj     2020-11-04
17:59:19.063795531 +0100
+++ gcc/testsuite/gcc.target/i386/zero-scratch-regs-32.c        2020-11-04
17:59:01.956984879 +0100
@@ -0,0 +1,5 @@
+/* PR target/97715 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fzero-call-used-regs=all -mno-80387" } */
+
+#include "../../c-c++-common/zero-scratch-regs-10.c"

isn't enough though, because while the st to st(7) registers are inaccessible
with -mno-80387, the MMX registers are accessible, so I think the
zero_all_st_registers change is needed too.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #19 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
And, actually the function.c change is probably unnecessary, because the
      if (!crtl->abi->clobbers_full_reg_p (regno))
        continue;
      if (!fixed_regs[i])
        continue;
should already rule them out.
  operand_reg_set &= accessible_reg_set;
...
      /* If a register is too limited to be treated as a register operand,
         then it should never be allocated to a pseudo.  */
      if (!TEST_HARD_REG_BIT (operand_reg_set, i))
        fixed_regs[i] = 1;

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qing.zhao at oracle dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #20 from Qing Zhao <qing.zhao at oracle dot com> ---
> On Nov 4, 2020, at 11:48 AM, jakub at gcc dot gnu.org <gcc-bugzilla@gcc.gnu.org> wrote:
> 
> --- Comment #18 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> --- gcc/function.c.jj   2020-10-31 17:41:19.756740009 +0100
> +++ gcc/function.c      2020-11-04 17:56:53.790403571 +0100
> @@ -5871,6 +5871,8 @@ gen_call_used_regs_seq (rtx_insn *ret, u
>        continue;
>       if (fixed_regs[regno])
>        continue;
> +      if (!TEST_HARD_REG_BIT (accessible_reg_set, regno))
> +       continue;
>       if (REGNO_REG_SET_P (live_out, regno))
>        continue;
>       if (only_gpr
> --- gcc/testsuite/gcc.target/i386/zero-scratch-regs-32.c.jj     2020-11-04
> 17:59:19.063795531 +0100
> +++ gcc/testsuite/gcc.target/i386/zero-scratch-regs-32.c        2020-11-04
> 17:59:01.956984879 +0100
> @@ -0,0 +1,5 @@
> +/* PR target/97715 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fzero-call-used-regs=all -mno-80387" } */
> +
> +#include "../../c-c++-common/zero-scratch-regs-10.c"
> 
> isn't enough though, because while the st to st(7) registers are inaccessible
> with -mno-80387, the MMX registers are accessible, so I think the
> zero_all_st_registers change is needed too.

Yeh, I noticed the same issue too.
I will add the fix you proposed previously.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qing.zhao at oracle dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #21 from Qing Zhao <qing.zhao at oracle dot com> ---
> --- Comment #19 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> And, actually the function.c change is probably unnecessary, because the
>      if (!crtl->abi->clobbers_full_reg_p (regno))
>        continue;
>      if (!fixed_regs[i])
>        continue;
> should already rule them out.
>  operand_reg_set &= accessible_reg_set;
> ...
>      /* If a register is too limited to be treated as a register operand,
>         then it should never be allocated to a pseudo.  */
>      if (!TEST_HARD_REG_BIT (operand_reg_set, i))
>        fixed_regs[i] = 1;

So, fixed_regs should already include the information of accessible_reg_set, no
need to check accessible_reg_set anymore.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qinzhao at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #22 from qinzhao at gcc dot gnu.org ---
proposed patch:

This change fixes a bug in the i386 backend when adding
-fzero-call-used-regs=all on a target that has no x87
registers.

When there is no x87 registers available, we should not
zero stack registers.

gcc/Changelog:

        PR target/97715
        * config/i386/i386.c (zero_all_st_registers): Return
        earlier when the FPU is disabled.

gcc/testsuite/ChnageLog:

        PR target/97715
        * gcc.target/i386/zero-scratch-regs-32.c: New test.
---
 gcc/config/i386/i386.c                               |  5 +++++
 gcc/testsuite/gcc.target/i386/zero-scratch-regs-32.c | 11 +++++++++++
 2 files changed, 16 insertions(+)
 create mode 100644 gcc/testsuite/gcc.target/i386/zero-scratch-regs-32.c

diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 6fc6228a26e..789ef727cf8 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -3640,6 +3640,11 @@ zero_all_vector_registers (HARD_REG_SET
need_zeroed_hardregs)
 static bool
 zero_all_st_registers (HARD_REG_SET need_zeroed_hardregs)
 {
+
+  /* If the FPU is disabled, no need to zero all st registers.  */
+  if (! (TARGET_80387 || TARGET_FLOAT_RETURNS_IN_80387))
+    return false;
+
   unsigned int num_of_st = 0;
   for (unsigned int regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
     if ((STACK_REGNO_P (regno) || MMX_REGNO_P (regno))
diff --git a/gcc/testsuite/gcc.target/i386/zero-scratch-regs-32.c
b/gcc/testsuite/gcc.target/i386/zero-scratch-regs-32.c
new file mode 100644
index 00000000000..ca3261fe5ea
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/zero-scratch-regs-32.c
@@ -0,0 +1,11 @@
+/* { dg-do compile { target *-*-linux* } } */
+/* { dg-options "-O2 -fzero-call-used-regs=all -mno-80387" } */
+
+int
+foo (int x)
+{
+  return (x + 1);
+}
+
+/* { dg-final { scan-assembler-not "fldz" } } */
+
--

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #23 from Martin Liška <marxin at gcc dot gnu.org> ---
*** Bug 97724 has been marked as a duplicate of this bug. ***

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

--- Comment #24 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Qing Zhao <qinzhao@gcc.gnu.org>:

https://gcc.gnu.org/g:cc32e81cdbb7696cd571bdb5ffe52f228f125df5

commit r11-4741-gcc32e81cdbb7696cd571bdb5ffe52f228f125df5
Author: qing zhao <qinzhao@gcc.gnu.org>
Date:   Thu Nov 5 15:57:46 2020 +0100

    i386: Fix PR97715

    This change fixes a bug in the i386 backend when adding
    -fzero-call-used-regs=all on a target that has no x87
    registers.

    When there is no x87 registers available, we should not
    zero stack registers.

    gcc/ChangeLog:

            PR target/97715
            * config/i386/i386.c (zero_all_st_registers): Return
            earlier when the FPU is disabled.

    gcc/testsuite/ChangeLog:

            PR target/97715
            * gcc.target/i386/zero-scratch-regs-32.c: New test.

[Bug target/97715] [11 Regression] ICE in insn_default_length, at config/i386/i386.md:15325 since r11-4578-gd10f3e900b0377b4

[qinzhao at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97715

qinzhao at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED

--- Comment #25 from qinzhao at gcc dot gnu.org ---
fixed in gcc11