gdb/riscv: Split ISA and ABI features

Message ID 20181216003021.2563-1-andrew.burgess@embecosm.com
State New, archived
Headers

Commit Message

Andrew Burgess Dec. 16, 2018, 12:30 a.m. UTC
  The goal of this commit is to allow RV64 binaries compiled for the 'F'
extension to run on a target that supports both the 'F' and 'D'
extensions.

The 'D' extension is a depends of the 'F' extension and chapter 9 of
the RISC-V ISA manual indicates that running a program compiled for
'F' on a 'D' target should be fine.

To support this the gdbarch now holds two feature sets, one represents
the features that are present on the target, and one represents the
features requested in the ELF flags.

The existing error checks are relaxed slightly to allow binaries
compiled for 32-bit 'F' extension to run on targets with the 64-bit
'D' extension.

A new set of functions called riscv_abi_{xlen,flen} are added to
compliment the existing riscv_isa_{xlen,flen}, and some callers to the
isa functions now call the abi functions when that is appropriate.

In riscv_call_arg_struct two asserts are removed, these asserts no
longer make sense.  The asserts were both like this:

    gdb_assert (TYPE_LENGTH (ainfo->type)
                <= (cinfo->flen + cinfo->xlen));

And were made in two cases, when passing structures like these:

   struct {
     integer field1;
     float   field2;
   };

or,

   struct {
     float   field1;
     integer field2;
   };

When running on an RV64 target which only has 32-bit float then the
integer field could be 64-bits, while if the float field is 32-bits
the overall size of the structure can be 128-bits (with 32-bits of
padding).  In this case the assertion would fail, however, the code
isn't incorrect, so its safe to just remove the assertion.

This was tested by running on an RV64IMFDC target using a compiler
configured for RV64IMFC, and comparing the results with those obtained
when using a compiler configured for RV64IMFDC.  The only regressions
I see (now) are in gdb.base/store.exp and are related too different
code generation choices GCC makes between the two targets.

gdb/ChangeLog:

	* arch/riscv.h (struct riscv_gdbarch_features) <hw_float_abi>:
	Delete.
	<operator==>: Update with for removed field.
	<hash>: Likewise.
	* riscv-tdep.h (struct gdbarch_tdep) <features>: Renamed to...
	<isa_features>: ...this.
	<abi_features>: New field.
	(riscv_isa_flen): Update comment.
	(riscv_abi_flen): New declaration.x
	* riscv-tdep.c (riscv_isa_xlen): Update to get answer from
	isa_features.
	(riscv_abi_xlen): New static function.
	(riscv_isa_flen): Update to get answer from isa_features.
	(riscv_abi_flen): New function.
	(riscv_has_fp_abi): Update to get answer from abi_features.
	(riscv_call_info::riscv_call_info): Use abi xlen and flen, not isa
	xlen and flen.
	(riscv_call_info) <xlen, flen>: Update comment.
	(riscv_call_arg_struct): Remove invalid assertions
	(riscv_features_from_gdbarch_info): Update now hw_float_abi field
	is removed.
	(riscv_gdbarch_init): Gather isa features and abi features
	separately, ensure both match on the gdbarch when reusing an old
	gdbarch.  Relax an error check to allow 32-bit abi float to run on
	a target with 64-bit float hardware.
---
 gdb/ChangeLog    | 28 +++++++++++++++++++
 gdb/arch/riscv.h | 15 ++--------
 gdb/riscv-tdep.c | 83 +++++++++++++++++++++++++++++++-------------------------
 gdb/riscv-tdep.h | 29 +++++++++++++++-----
 4 files changed, 98 insertions(+), 57 deletions(-)
  

Comments

John Baldwin Dec. 17, 2018, 6:30 p.m. UTC | #1
On 12/15/18 4:30 PM, Andrew Burgess wrote:
> When running on an RV64 target which only has 32-bit float then the
> integer field could be 64-bits, while if the float field is 32-bits
> the overall size of the structure can be 128-bits (with 32-bits of
> padding).  In this case the assertion would fail, however, the code
> isn't incorrect, so its safe to just remove the assertion.
> 
> This was tested by running on an RV64IMFDC target using a compiler
> configured for RV64IMFC, and comparing the results with those obtained
> when using a compiler configured for RV64IMFDC.  The only regressions
> I see (now) are in gdb.base/store.exp and are related too different
> code generation choices GCC makes between the two targets.
> 
> gdb/ChangeLog:
> 
> 	* arch/riscv.h (struct riscv_gdbarch_features) <hw_float_abi>:
> 	Delete.
> 	<operator==>: Update with for removed field.
> 	<hash>: Likewise.
> 	* riscv-tdep.h (struct gdbarch_tdep) <features>: Renamed to...
> 	<isa_features>: ...this.
> 	<abi_features>: New field.
> 	(riscv_isa_flen): Update comment.
> 	(riscv_abi_flen): New declaration.x

Extra 'x'?

> diff --git a/gdb/riscv-tdep.c b/gdb/riscv-tdep.c
> index 88b79af866f..aeb9f2e5627 100644
> --- a/gdb/riscv-tdep.c
> +++ b/gdb/riscv-tdep.c
> @@ -361,7 +361,18 @@ static unsigned int riscv_debug_gdbarch = 0;
>  int
>  riscv_isa_xlen (struct gdbarch *gdbarch)
>  {
> -  return gdbarch_tdep (gdbarch)->features.xlen;
> +  return gdbarch_tdep (gdbarch)->isa_features.xlen;
> +}
> +
> +/* Return the x-register length that is being used for the abi.  This
> +   exists only so we have similar interfaces for flen and xlen, however,
> +   this always calls RISCV_ISA_XLEN, as running a binary compiled for RV32
> +   on an RV64 target doesn't make sense.  */

I think in theory you could do this just fine?  It would be akin to running
an i386 binary on x86_64.  I thought the XLEN field in one of the status
registers was allowed to be writable so that you could switch modes at
least for U?  That said, I'm not sure if that will actually get used.  I'm
not sure if Linux is planning on supporting RV32.  FreeBSD will likely only
ever support RV64.  I think having riscv_abi_xlen is fine in its current
form.

Note that I think the current patch will always report the floating point
registers as the target width?  I think for other architectures in GDB we
store the target registers as "raw" registers, but would present the
abi registers as "cooked".  I think this is how we treat an i386 binary on
x86_64 for example where 'info registers' shows 32-bit registers instead of
64-bit.  MIPS does some similar things to handle 32-bit MIPS binaries on
64-bit targets.  Perhaps this is less important for FP registers, but I'm
curious if GDB will DTRT if you have a 'float' variable that DWARF indicates
is stored in $f0 if $f0 is actually a 64-bit double register?
  
Jim Wilson Dec. 18, 2018, 9:04 p.m. UTC | #2
On Mon, Dec 17, 2018 at 10:30 AM John Baldwin <jhb@freebsd.org> wrote:
> I think in theory you could do this just fine?  It would be akin to running
> an i386 binary on x86_64.  I thought the XLEN field in one of the status
> registers was allowed to be writable so that you could switch modes at
> least for U?  That said, I'm not sure if that will actually get used.  I'm
> not sure if Linux is planning on supporting RV32.  FreeBSD will likely only
> ever support RV64.  I think having riscv_abi_xlen is fine in its current
> form.

riscv32-linux will be supported because there are people that want
this for embedded sytems.  The desktop/server distros will probably
never support riscv32-linux.  The riscv32-linux support is not
complete upstream yet though.  There is a pending glibc patch that
won't go until after the next glibc release, because they are waiting
for 64-bit time_t support (or something like that), to avoid a future
ABI change.

While you can in theory run 32-bit code on a riscv64-linux, there is
currently no linux kernel support for this, and no plans to add it.  A
n32/x32 style ABI would be simpler to implement, but there are no
current plans for this either.

> Note that I think the current patch will always report the floating point
> registers as the target width?  I think for other architectures in GDB we
> store the target registers as "raw" registers, but would present the
> abi registers as "cooked".  I think this is how we treat an i386 binary on
> x86_64 for example where 'info registers' shows 32-bit registers instead of
> 64-bit.  MIPS does some similar things to handle 32-bit MIPS binaries on
> 64-bit targets.  Perhaps this is less important for FP registers, but I'm
> curious if GDB will DTRT if you have a 'float' variable that DWARF indicates
> is stored in $f0 if $f0 is actually a 64-bit double register?

When gdb knows that it is a float, it prints a float, so this works
OK.  The problem occurs when you use a command like "info registers
float" and gdb doesn't know if it is a float or a double value.  This
is why I added support to treat 64-bit FP registers as a union of
float and double, so that gdb would print both.  I also added support
to nan-box float values for inferior calls.  I haven't seen any
serious FP register related issues when I run the gdb testsuite
natively on my Fedora Core system.  I haven't tested Andrew's latest
patch yet, but if I see issues I will report them.

Jim
  
Andrew Burgess Dec. 18, 2018, 11:29 p.m. UTC | #3
Thanks for the feedback from you both.

* Jim Wilson <jimw@sifive.com> [2018-12-18 13:04:13 -0800]:

> On Mon, Dec 17, 2018 at 10:30 AM John Baldwin <jhb@freebsd.org> wrote:
> > I think in theory you could do this just fine?  It would be akin to running
> > an i386 binary on x86_64.  I thought the XLEN field in one of the status
> > registers was allowed to be writable so that you could switch modes at
> > least for U?  That said, I'm not sure if that will actually get used.  I'm
> > not sure if Linux is planning on supporting RV32.  FreeBSD will likely only
> > ever support RV64.  I think having riscv_abi_xlen is fine in its current
> > form.
> 
> riscv32-linux will be supported because there are people that want
> this for embedded sytems.  The desktop/server distros will probably
> never support riscv32-linux.  The riscv32-linux support is not
> complete upstream yet though.  There is a pending glibc patch that
> won't go until after the next glibc release, because they are waiting
> for 64-bit time_t support (or something like that), to avoid a future
> ABI change.
> 
> While you can in theory run 32-bit code on a riscv64-linux, there is
> currently no linux kernel support for this, and no plans to add it.  A
> n32/x32 style ABI would be simpler to implement, but there are no
> current plans for this either.

I think the code as proposed doesn't prevent supporting rv32i on rv64i
in the future, I'll update the wording in the comments and commit
messages to remove the "this will not work" to "this isn't currently
supported, but could be in future".

> 
> > Note that I think the current patch will always report the floating point
> > registers as the target width?  I think for other architectures in GDB we
> > store the target registers as "raw" registers, but would present the
> > abi registers as "cooked".  I think this is how we treat an i386 binary on
> > x86_64 for example where 'info registers' shows 32-bit registers instead of
> > 64-bit.  MIPS does some similar things to handle 32-bit MIPS binaries on
> > 64-bit targets.  Perhaps this is less important for FP registers, but I'm
> > curious if GDB will DTRT if you have a 'float' variable that DWARF indicates
> > is stored in $f0 if $f0 is actually a 64-bit double register?
> 
> When gdb knows that it is a float, it prints a float, so this works
> OK.  The problem occurs when you use a command like "info registers
> float" and gdb doesn't know if it is a float or a double value.  This
> is why I added support to treat 64-bit FP registers as a union of
> float and double, so that gdb would print both.  I also added support
> to nan-box float values for inferior calls.  I haven't seen any
> serious FP register related issues when I run the gdb testsuite
> natively on my Fedora Core system.  I haven't tested Andrew's latest
> patch yet, but if I see issues I will report them.

I did think about having a cooked view as John described, so if you
load a binary marked as being for 'F' extension then you'd only see
32-bit float registers, even on a target with 64-bit float.  However,
I think such a feature would be orthogonal to that changes I've
presented here.  On my list of things to fix / improve it's a nice to
have, but not a super high priority, especially as we have your (Jim)
union view of 64-bit float registers now.

I'll roll a new version of the patch with the typo John spotted fixed,
and with some improved comments and post this tomorrow.

Thanks,
Andrew
  
John Baldwin Dec. 18, 2018, 11:56 p.m. UTC | #4
On 12/18/18 3:29 PM, Andrew Burgess wrote:
>>> Note that I think the current patch will always report the floating point
>>> registers as the target width?  I think for other architectures in GDB we
>>> store the target registers as "raw" registers, but would present the
>>> abi registers as "cooked".  I think this is how we treat an i386 binary on
>>> x86_64 for example where 'info registers' shows 32-bit registers instead of
>>> 64-bit.  MIPS does some similar things to handle 32-bit MIPS binaries on
>>> 64-bit targets.  Perhaps this is less important for FP registers, but I'm
>>> curious if GDB will DTRT if you have a 'float' variable that DWARF indicates
>>> is stored in $f0 if $f0 is actually a 64-bit double register?
>>
>> When gdb knows that it is a float, it prints a float, so this works
>> OK.  The problem occurs when you use a command like "info registers
>> float" and gdb doesn't know if it is a float or a double value.  This
>> is why I added support to treat 64-bit FP registers as a union of
>> float and double, so that gdb would print both.  I also added support
>> to nan-box float values for inferior calls.  I haven't seen any
>> serious FP register related issues when I run the gdb testsuite
>> natively on my Fedora Core system.  I haven't tested Andrew's latest
>> patch yet, but if I see issues I will report them.
> 
> I did think about having a cooked view as John described, so if you
> load a binary marked as being for 'F' extension then you'd only see
> 32-bit float registers, even on a target with 64-bit float.  However,
> I think such a feature would be orthogonal to that changes I've
> presented here.  On my list of things to fix / improve it's a nice to
> have, but not a super high priority, especially as we have your (Jim)
> union view of 64-bit float registers now.

Please don't take my questions/suggestions as requirements, btw.  I'm not
certain that cooked registers would really be the right thing.  Having
floats as a union the way the vector registers are on x86 is definitely a
plus and probably does give you most of the gain a cooked register set
would bring.
  

Patch

diff --git a/gdb/arch/riscv.h b/gdb/arch/riscv.h
index 79f5ec622d4..68297a8b73e 100644
--- a/gdb/arch/riscv.h
+++ b/gdb/arch/riscv.h
@@ -46,19 +46,10 @@  struct riscv_gdbarch_features
      that there are no f-registers.  No other value is valid.  */
   int flen = 0;
 
-  /* This indicates if hardware floating point abi is in use.  If the FLEN
-     field is 0 then this value _must_ be false.  If the FLEN field is
-     non-zero and this field is false then this indicates the target has
-     floating point registers, but is still using the soft-float abi.  If
-     this field is true then the hardware floating point abi is in use, and
-     values are passed in f-registers matching the size of FLEN.  */
-  bool hw_float_abi = false;
-
   /* Equality operator.  */
   bool operator== (const struct riscv_gdbarch_features &rhs) const
   {
-    return (xlen == rhs.xlen && flen == rhs.flen
-	    && hw_float_abi == rhs.hw_float_abi);
+    return (xlen == rhs.xlen && flen == rhs.flen);
   }
 
   /* Inequality operator.  */
@@ -70,9 +61,7 @@  struct riscv_gdbarch_features
   /* Used by std::unordered_map to hash feature sets.  */
   std::size_t hash () const noexcept
   {
-    std::size_t val = ((xlen & 0x1f) << 6
-                       | (flen & 0x1f) << 1
-                       | (hw_float_abi ? 1 : 0));
+    std::size_t val = ((xlen & 0x1f) << 5 | (flen & 0x1f) << 0);
     return val;
   }
 };
diff --git a/gdb/riscv-tdep.c b/gdb/riscv-tdep.c
index 88b79af866f..aeb9f2e5627 100644
--- a/gdb/riscv-tdep.c
+++ b/gdb/riscv-tdep.c
@@ -361,7 +361,18 @@  static unsigned int riscv_debug_gdbarch = 0;
 int
 riscv_isa_xlen (struct gdbarch *gdbarch)
 {
-  return gdbarch_tdep (gdbarch)->features.xlen;
+  return gdbarch_tdep (gdbarch)->isa_features.xlen;
+}
+
+/* Return the x-register length that is being used for the abi.  This
+   exists only so we have similar interfaces for flen and xlen, however,
+   this always calls RISCV_ISA_XLEN, as running a binary compiled for RV32
+   on an RV64 target doesn't make sense.  */
+
+static int
+riscv_abi_xlen (struct gdbarch *gdbarch)
+{
+  return riscv_isa_xlen (gdbarch);
 }
 
 /* See riscv-tdep.h.  */
@@ -369,7 +380,15 @@  riscv_isa_xlen (struct gdbarch *gdbarch)
 int
 riscv_isa_flen (struct gdbarch *gdbarch)
 {
-  return gdbarch_tdep (gdbarch)->features.flen;
+  return gdbarch_tdep (gdbarch)->isa_features.flen;
+}
+
+/* See riscv-tdep.h.  */
+
+int
+riscv_abi_flen (struct gdbarch *gdbarch)
+{
+  return gdbarch_tdep (gdbarch)->abi_features.flen;
 }
 
 /* Return true if the target for GDBARCH has floating point hardware.  */
@@ -385,7 +404,7 @@  riscv_has_fp_regs (struct gdbarch *gdbarch)
 static bool
 riscv_has_fp_abi (struct gdbarch *gdbarch)
 {
-  return gdbarch_tdep (gdbarch)->features.hw_float_abi;
+  return gdbarch_tdep (gdbarch)->abi_features.flen > 0;
 }
 
 /* Return true if REGNO is a floating pointer register.  */
@@ -1777,8 +1796,8 @@  struct riscv_call_info
     : int_regs (RISCV_A0_REGNUM, RISCV_A0_REGNUM + 7),
       float_regs (RISCV_FA0_REGNUM, RISCV_FA0_REGNUM + 7)
   {
-    xlen = riscv_isa_xlen (gdbarch);
-    flen = riscv_isa_flen (gdbarch);
+    xlen = riscv_abi_xlen (gdbarch);
+    flen = riscv_abi_flen (gdbarch);
 
     /* Disable use of floating point registers if needed.  */
     if (!riscv_has_fp_abi (gdbarch))
@@ -1798,7 +1817,7 @@  struct riscv_call_info
   struct riscv_arg_reg float_regs;
 
   /* The XLEN and FLEN are copied in to this structure for convenience, and
-     are just the results of calling RISCV_ISA_XLEN and RISCV_ISA_FLEN.  */
+     are just the results of calling RISCV_ABI_XLEN and RISCV_ABI_FLEN.  */
   int xlen;
   int flen;
 };
@@ -2139,9 +2158,6 @@  riscv_call_arg_struct (struct riscv_arg_info *ainfo,
 	{
 	  int len0, len1, offset;
 
-	  gdb_assert (TYPE_LENGTH (ainfo->type)
-		      <= (cinfo->flen + cinfo->xlen));
-
 	  len0 = TYPE_LENGTH (sinfo.field_type (0));
 	  if (!riscv_assign_reg_location (&ainfo->argloc[0],
 					  &cinfo->float_regs, len0, 0))
@@ -2165,9 +2181,6 @@  riscv_call_arg_struct (struct riscv_arg_info *ainfo,
 	{
 	  int len0, len1, offset;
 
-	  gdb_assert (TYPE_LENGTH (ainfo->type)
-		      <= (cinfo->flen + cinfo->xlen));
-
 	  len0 = TYPE_LENGTH (sinfo.field_type (0));
 	  len1 = TYPE_LENGTH (sinfo.field_type (1));
 	  offset = align_up (len0, riscv_type_alignment (sinfo.field_type (1)));
@@ -2827,15 +2840,9 @@  riscv_features_from_gdbarch_info (const struct gdbarch_info info)
 			_("unknown ELF header class %d"), eclass);
 
       if (e_flags & EF_RISCV_FLOAT_ABI_DOUBLE)
-	{
-	  features.flen = 8;
-	  features.hw_float_abi = true;
-	}
+	features.flen = 8;
       else if (e_flags & EF_RISCV_FLOAT_ABI_SINGLE)
-	{
-	  features.flen = 4;
-	  features.hw_float_abi = true;
-	}
+	features.flen = 4;
     }
   else
     {
@@ -3039,25 +3046,25 @@  riscv_gdbarch_init (struct gdbarch_info info,
   /* Have a look at what the supplied (if any) bfd object requires of the
      target, then check that this matches with what the target is
      providing.  */
-  struct riscv_gdbarch_features info_features
+  struct riscv_gdbarch_features abi_features
     = riscv_features_from_gdbarch_info (info);
-  if (info_features.xlen != 0 && info_features.xlen != features.xlen)
+  if (abi_features.xlen != 0 && abi_features.xlen != features.xlen)
     error (_("bfd requires xlen %d, but target has xlen %d"),
-           info_features.xlen, features.xlen);
-  if (info_features.flen != 0 && info_features.flen != features.flen)
+            abi_features.xlen, features.xlen);
+  /* Unlike the x-register case, running a binary compiled for the 32-bit
+     float abi should work fine on a target with 64-bit (or more) floating
+     point hardware.  As a result, we only complain if the ABI_FEATURES
+     requires more than the target has to offer.  */
+  if (abi_features.flen > features.flen)
     error (_("bfd requires flen %d, but target has flen %d"),
-           info_features.flen, features.flen);
-
-  /* If the xlen from INFO_FEATURES is 0 then this indicates either there
-     is no bfd object, or nothing useful could be extracted from it, in
-     this case we enable hardware float abi if the target has floating
-     point registers.
+            abi_features.flen, features.flen);
 
-     If the xlen from INFO_FEATURES is not 0, and the flen in
-     INFO_FEATURES is also not 0, then this indicates that the supplied
-     bfd does require hardware floating point abi.  */
-  if (info_features.xlen == 0 || info_features.flen != 0)
-    features.hw_float_abi = (features.flen > 0);
+  /* If the ABI_FEATURES xlen is 0 then this indicates we got no useful abi
+     features from the INFO object.  In this case we assume that the xlen
+     abi matches the hardware type.  If we had gotten anything useful from
+     INFO then this would have been enforced above.  */
+  if (abi_features.xlen == 0)
+    abi_features.xlen = features.xlen;
 
   /* Find a candidate among the list of pre-declared architectures.  */
   for (arches = gdbarch_list_lookup_by_info (arches, &info);
@@ -3069,7 +3076,8 @@  riscv_gdbarch_init (struct gdbarch_info info,
          gdbarch.  */
       struct gdbarch_tdep *other_tdep = gdbarch_tdep (arches->gdbarch);
 
-      if (other_tdep->features != features)
+      if (other_tdep->isa_features != features
+	  || other_tdep->abi_features != abi_features)
         continue;
 
       break;
@@ -3084,7 +3092,8 @@  riscv_gdbarch_init (struct gdbarch_info info,
   /* None found, so create a new architecture from the information provided.  */
   tdep = new (struct gdbarch_tdep);
   gdbarch = gdbarch_alloc (&info, tdep);
-  tdep->features = features;
+  tdep->isa_features = features;
+  tdep->abi_features = abi_features;
 
   /* Target data types.  */
   set_gdbarch_short_bit (gdbarch, 16);
diff --git a/gdb/riscv-tdep.h b/gdb/riscv-tdep.h
index 59ad37df6ce..bfb65c002fa 100644
--- a/gdb/riscv-tdep.h
+++ b/gdb/riscv-tdep.h
@@ -59,9 +59,14 @@  enum
 /* RISC-V specific per-architecture information.  */
 struct gdbarch_tdep
 {
-  /* Features about the target that impact how the gdbarch is configured.
-     Two gdbarch instances are compatible only if this field matches.  */
-  struct riscv_gdbarch_features features;
+  /* Features about the target hardware that impact how the gdbarch is
+     configured.  Two gdbarch instances are compatible only if this field
+     matches.  */
+  struct riscv_gdbarch_features isa_features;
+
+  /* Features about the abi that impact how the gdbarch is configured.  Two
+     gdbarch instances are compatible only if this field matches.  */
+  struct riscv_gdbarch_features abi_features;
 
   /* ISA-specific data types.  */
   struct type *riscv_fpreg_d_type = nullptr;
@@ -73,12 +78,22 @@  struct gdbarch_tdep
    RV128.  */
 extern int riscv_isa_xlen (struct gdbarch *gdbarch);
 
-/* Return the width in bytes of the floating point registers for GDBARCH.
-   If this architecture has no floating point registers, then return 0.
-   Possible values are 4, 8, or 16 for depending on which of single, double
-   or quad floating point support is available.  */
+/* Return the width in bytes of the hardware floating point registers for
+   GDBARCH.  If this architecture has no floating point registers, then
+   return 0.  Possible values are 4, 8, or 16 for depending on which of
+   single, double or quad floating point support is available.  */
 extern int riscv_isa_flen (struct gdbarch *gdbarch);
 
+/* Return the width in bytes of the floating point register abi for
+   GDBARCH.  This reflects how the binary was compiled rather than the
+   hardware that is available.  It is possible that a binary is compiled
+   for single precision floating point, and then run on a target with
+   double precision floating point.  A return value of 0 indicates that no
+   floating point abi is in use (floating point arguments will be passed
+   in integer registers) other possible return value are 4, 8, or 16 as
+   with RISCV_ISA_FLEN.  */
+extern int riscv_abi_flen (struct gdbarch *gdbarch);
+
 /* Single step based on where the current instruction will take us.  */
 extern std::vector<CORE_ADDR> riscv_software_single_step
   (struct regcache *regcache);