[v5] RISCV: Inline subword atomic ops

Message ID 20230418142858.2424851-1-patrick@rivosinc.com
State Accepted
Headers
Series [v5] RISCV: Inline subword atomic ops |

Checks

Context Check Description
snail/gcc-patch-check success Github commit url

Commit Message

Patrick O'Neill April 18, 2023, 2:28 p.m. UTC
  RISC-V has no support for subword atomic operations; code currently
generates libatomic library calls.

This patch changes the default behavior to inline subword atomic calls 
(using the same logic as the existing library call).
Behavior can be specified using the -minline-atomics and
-mno-inline-atomics command line flags.

gcc/libgcc/config/riscv/atomic.c has the same logic implemented in asm.
This will need to stay for backwards compatibility and the
-mno-inline-atomics flag.

2023-04-18 Patrick O'Neill <patrick@rivosinc.com>

	PR target/104338
	* riscv-protos.h: Add helper function stubs.
	* riscv.cc: Add helper functions for subword masking.
	* riscv.opt: Add command-line flag.
	* sync.md: Add masking logic and inline asm for fetch_and_op,
	fetch_and_nand, CAS, and exchange ops.
	* invoke.texi: Add blurb regarding command-line flag.
	* inline-atomics-1.c: New test.
	* inline-atomics-2.c: Likewise.
	* inline-atomics-3.c: Likewise.
	* inline-atomics-4.c: Likewise.
	* inline-atomics-5.c: Likewise.
	* inline-atomics-6.c: Likewise.
	* inline-atomics-7.c: Likewise.
	* inline-atomics-8.c: Likewise.
	* atomic.c: Add reference to duplicate logic.

Signed-off-by: Patrick O'Neill <patrick@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
---
v4: https://inbox.sourceware.org/gcc-patches/20220821215823.18207-1-palmer@rivosinc.com/

Rebased v4 and addressed Kito Cheng's comments.
No new failures on trunk.
---
The mapping implemented here matches Libatomic. That mapping changes if
"Implement ISA Manual Table A.6 Mappings" is merged. Depending on which
patch is merged first, I will update the other to make sure the
correct mapping is emitted.
  https://gcc.gnu.org/pipermail/gcc-patches/2023-April/615748.html
---
 gcc/config/riscv/riscv-protos.h               |   2 +
 gcc/config/riscv/riscv.cc                     |  50 ++
 gcc/config/riscv/riscv.opt                    |   4 +
 gcc/config/riscv/sync.md                      | 314 ++++++++++
 gcc/doc/invoke.texi                           |   8 +
 .../gcc.target/riscv/inline-atomics-1.c       |  18 +
 .../gcc.target/riscv/inline-atomics-2.c       |  19 +
 .../gcc.target/riscv/inline-atomics-3.c       | 569 ++++++++++++++++++
 .../gcc.target/riscv/inline-atomics-4.c       | 566 +++++++++++++++++
 .../gcc.target/riscv/inline-atomics-5.c       |  87 +++
 .../gcc.target/riscv/inline-atomics-6.c       |  87 +++
 .../gcc.target/riscv/inline-atomics-7.c       |  69 +++
 .../gcc.target/riscv/inline-atomics-8.c       |  69 +++
 libgcc/config/riscv/atomic.c                  |   2 +
 14 files changed, 1864 insertions(+)
 create mode 100644 gcc/testsuite/gcc.target/riscv/inline-atomics-1.c
 create mode 100644 gcc/testsuite/gcc.target/riscv/inline-atomics-2.c
 create mode 100644 gcc/testsuite/gcc.target/riscv/inline-atomics-3.c
 create mode 100644 gcc/testsuite/gcc.target/riscv/inline-atomics-4.c
 create mode 100644 gcc/testsuite/gcc.target/riscv/inline-atomics-5.c
 create mode 100644 gcc/testsuite/gcc.target/riscv/inline-atomics-6.c
 create mode 100644 gcc/testsuite/gcc.target/riscv/inline-atomics-7.c
 create mode 100644 gcc/testsuite/gcc.target/riscv/inline-atomics-8.c
  

Comments

Andreas Schwab April 18, 2023, 3:06 p.m. UTC | #1
On Apr 18 2023, Patrick O'Neill wrote:

> diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
> index a38547f53e5..9c3e91d2fee 100644
> --- a/gcc/doc/invoke.texi
> +++ b/gcc/doc/invoke.texi
> @@ -1227,6 +1227,7 @@ See RS/6000 and PowerPC Options.
>  -mstack-protector-guard=@var{guard}  -mstack-protector-guard-reg=@var{reg}
>  -mstack-protector-guard-offset=@var{offset}
>  -mcsr-check -mno-csr-check}
> +-minline-atomics  -mno-inline-atomics

The options need to be inside @gccoptlist.

> @@ -29006,6 +29007,13 @@ Do or don't use smaller but slower prologue and epilogue code that uses
>  library function calls.  The default is to use fast inline prologues and
>  epilogues.
>  
> +@item -minline-atomics
> +@itemx -mno-inline-atomics
> +@opindex minline-atomics

@opindex should precede @item.
  
Jeff Law April 18, 2023, 4:59 p.m. UTC | #2
On 4/18/23 08:28, Patrick O'Neill wrote:
> RISC-V has no support for subword atomic operations; code currently
> generates libatomic library calls.
> 
> This patch changes the default behavior to inline subword atomic calls
> (using the same logic as the existing library call).
> Behavior can be specified using the -minline-atomics and
> -mno-inline-atomics command line flags.
> 
> gcc/libgcc/config/riscv/atomic.c has the same logic implemented in asm.
> This will need to stay for backwards compatibility and the
> -mno-inline-atomics flag.
> 
> 2023-04-18 Patrick O'Neill <patrick@rivosinc.com>
> 
> 	PR target/104338
> 	* riscv-protos.h: Add helper function stubs.
> 	* riscv.cc: Add helper functions for subword masking.
> 	* riscv.opt: Add command-line flag.
> 	* sync.md: Add masking logic and inline asm for fetch_and_op,
> 	fetch_and_nand, CAS, and exchange ops.
> 	* invoke.texi: Add blurb regarding command-line flag.
> 	* inline-atomics-1.c: New test.
> 	* inline-atomics-2.c: Likewise.
> 	* inline-atomics-3.c: Likewise.
> 	* inline-atomics-4.c: Likewise.
> 	* inline-atomics-5.c: Likewise.
> 	* inline-atomics-6.c: Likewise.
> 	* inline-atomics-7.c: Likewise.
> 	* inline-atomics-8.c: Likewise.
> 	* atomic.c: Add reference to duplicate logic.
So for others who may be interested.  The motivation here is that for a 
sub-word atomic we currently have to explicitly link in libatomic or we 
get undefined symbols.

This is particularly problematical for the distros because we're one of 
the few (only?) architectures supported by the distros that require 
linking in libatomic for these cases.  THe distros don't want to adjust 
each affected packages and be stuck carrying that change forward or 
negotiating with all the relevant upstreams.  The distros might tackle 
this problem by porting this patch into their compiler tree which has 
its own set of problems with long term maintenance.

The net is from a usability standpoint it's best if we get this problem 
addressed and backported to our gcc-13 RISC-V coordination branch.

We had held this up pending resolution of some other issues in the 
atomics space.  In retrospect that might have been a mistake.

So with that background...  Here we go...

>   
> +/* Helper function for extracting a subword from memory.  */
> +
> +void
> +riscv_subword_address (rtx mem, rtx *aligned_mem, rtx *shift, rtx *mask,
> +		       rtx *not_mask)
So I'd expand on that comment.  The idea is we would like someone 
working in the backend to be able to read the function comment and have 
a reasonable sense of what the function does as well as the inputs and 
return value.  So perhaps something like this:

/* Given memory reference MEM, expand code to compute the aligned
    memory address, shift and mask values and store them into
    *ALIGNED_MEM, *SHIFT, *MASK and *NOT_MASK.  */

Or something like that.


> +{
> +  /* Align the memory addess to a word.  */
s/addess/address/


> +  rtx addr = force_reg (Pmode, XEXP (mem, 0));
> +
> +  rtx aligned_addr = gen_reg_rtx (Pmode);
> +  emit_move_insn (aligned_addr,  gen_rtx_AND (Pmode, addr,
> +					      gen_int_mode (-4, Pmode)));
So rather than -4 as a magic number, GET_MODE_MASK would be better. 
That may result in needing to rewrap this code.  I'd bring the 
gen_rtx_AND down on a new line, aligned with aligned_addr.

Presumably using SImode is intentional here rather than wanting to use 
word_mode which would be SImode for rv32 and DImode for rv64?  I'm going 
to work based on that assumption, but if it isn't there's more work to 
do to generalize this code.



> +
> +  /* Calculate the shift amount.  */
> +  emit_move_insn (*shift, gen_rtx_AND (SImode, gen_lowpart (SImode, addr),
> +				       gen_int_mode (3, SImode)));
> +  emit_move_insn (*shift, gen_rtx_ASHIFT (SImode, *shift,
> +					  gen_int_mode(3, SImode)));


Formatting nit.  space after before open paren in the gen_int_mode call 
on that last line above.  This minor goof shows up in various places, 
please review the patch as a whole looking for similar nits.



> +
> +  /* Calculate the mask.  */
> +  int unshifted_mask;
> +  if (GET_MODE (mem) == QImode)
> +    unshifted_mask = 0xFF;
> +  else
> +    unshifted_mask = 0xFFFF;
Can you just use GET_MASK_MODE here which should simplify this to 
something like

unshifted_mask = GET_MODE_MASK (GET_MODE (mem));








> +
> +(define_expand "atomic_fetch_nand<mode>"
> +  [(set (match_operand:SHORT 0 "register_operand" "=&r")
> +	(match_operand:SHORT 1 "memory_operand" "+A"))
> +   (set (match_dup 1)
> +	(unspec_volatile:SHORT
> +	  [(not:SHORT (and:SHORT (match_dup 1)
> +				 (match_operand:SHORT 2 "reg_or_0_operand" "rJ")))
> +	   (match_operand:SI 3 "const_int_operand")] ;; model
> +	 UNSPEC_SYNC_OLD_OP_SUBWORD))]
> +  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
Just a note, constraints aren't necessary for a define_expand.  They 
don't hurt anything though.  They do document expectations, but then you 
have to maintain them over time.  I'm OK leaving them, mostly wanted to 
make sure you're aware they aren't strictly necessary for a define_expand.






> +
> +(define_expand "atomic_fetch_<atomic_optab><mode>"
> +  [(set (match_operand:SHORT 0 "register_operand" "=&r")	      ;; old value at mem
> +	(match_operand:SHORT 1 "memory_operand" "+A"))		      ;; mem location
> +   (set (match_dup 1)
> +	(unspec_volatile:SHORT
> +	  [(any_atomic:SHORT (match_dup 1)
> +		     (match_operand:SHORT 2 "reg_or_0_operand" "rJ")) ;; value for op
> +	   (match_operand:SI 3 "const_int_operand")]		      ;; model
> +	 UNSPEC_SYNC_OLD_OP_SUBWORD))]
> +  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
> +{
> +  /* We have no QImode/HImode atomics, so form a mask, then use
> +     subword_atomic_fetch_strong_<mode> to implement a LR/SC version of the
> +     operation. */
> +
> +  /* Logic duplicated in gcc/libgcc/config/riscv/atomic.c for use when inlining
> +     is disabled */
Makes me wonder if we should expose builtins that could be used by 
atomic.c rather than having the logic open coded in two places.  That 
could be a follow-up in my opinion rather than a blocker for this patch.


>   
> +(define_expand "atomic_compare_and_swap<mode>"
> +  [(match_operand:SI 0 "register_operand" "")    ;; bool output
> +   (match_operand:SHORT 1 "register_operand" "") ;; val output
> +   (match_operand:SHORT 2 "memory_operand" "")   ;; memory
> +   (match_operand:SHORT 3 "reg_or_0_operand" "") ;; expected value
> +   (match_operand:SHORT 4 "reg_or_0_operand" "") ;; desired value
> +   (match_operand:SI 5 "const_int_operand" "")   ;; is_weak
> +   (match_operand:SI 6 "const_int_operand" "")   ;; mod_s
> +   (match_operand:SI 7 "const_int_operand" "")]  ;; mod_f
> +  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
I nearly suggested you use this form for the earlier expanders where the 
RTL in the expander is ultimately thrown away because we generate all 
the RTL we need in the C fragment and finish with a "DONE" tag.


> diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-1.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-1.c
> new file mode 100644
> index 00000000000..5c5623d9b2f
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-1.c
> @@ -0,0 +1,18 @@
> +/* { dg-do compile } */
> +/* { dg-options "-mno-inline-atomics" } */
> +/* { dg-message "note: '__sync_fetch_and_nand' changed semantics in GCC 4.4" "fetch_and_nand" { target *-*-* } 0 } */
> +/* { dg-final { scan-assembler "\tcall\t__sync_fetch_and_add_1" } } */
> +/* { dg-final { scan-assembler "\tcall\t__sync_fetch_and_nand_1" } } */
> +/* { dg-final { scan-assembler "\tcall\t__sync_bool_compare_and_swap_1" } } */
> +
> +char foo;
> +char bar;
> +char baz;
> +
> +int
> +main ()
> +{
> +  __sync_fetch_and_add(&foo, 1);
> +  __sync_fetch_and_nand(&bar, 1);
> +  __sync_bool_compare_and_swap (&baz, 1, 2);
> +}
> diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-2.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-2.c
> new file mode 100644
> index 00000000000..fdce7a5d71f
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-2.c
> @@ -0,0 +1,19 @@
> +/* { dg-do compile } */
> +/* Verify that subword atomics do not generate calls.  */
> +/* { dg-options "-minline-atomics" } */
> +/* { dg-message "note: '__sync_fetch_and_nand' changed semantics in GCC 4.4" "fetch_and_nand" { target *-*-* } 0 } */
> +/* { dg-final { scan-assembler-not "\tcall\t__sync_fetch_and_add_1" } } */
> +/* { dg-final { scan-assembler-not "\tcall\t__sync_fetch_and_nand_1" } } */
> +/* { dg-final { scan-assembler-not "\tcall\t__sync_bool_compare_and_swap_1" } } */
Note that you can #include another test file.  When you do that the 
dg-directives in the #included file are ignored.   Meaning that you can 
share test code, but use different dg-directives.  So for inline-2.c, 
you can have the dg-* directives above, followed by:

#include "inline-atomics-1.c"

Overall it looks pretty close to ready.

Jeff
  
Patrick O'Neill April 18, 2023, 8:48 p.m. UTC | #3
On 4/18/23 09:59, Jeff Law wrote:
> On 4/18/23 08:28, Patrick O'Neill wrote:
> ...
>> +  rtx addr = force_reg (Pmode, XEXP (mem, 0));
>> +
>> +  rtx aligned_addr = gen_reg_rtx (Pmode);
>> +  emit_move_insn (aligned_addr,  gen_rtx_AND (Pmode, addr,
>> +                          gen_int_mode (-4, Pmode)));
> So rather than -4 as a magic number, GET_MODE_MASK would be better. 
> That may result in needing to rewrap this code.  I'd bring the 
> gen_rtx_AND down on a new line, aligned with aligned_addr.
IIUC GET_MODE_MASK generates masks like 0xFF for QI (for example). It 
doesn't have the granularity to generate 0x3 (which we can NOT to get 
-4). I searched the GCC internals docs but couldn't find a function that 
does address alignment masks.
> Presumably using SImode is intentional here rather than wanting to use 
> word_mode which would be SImode for rv32 and DImode for rv64?  I'm 
> going to work based on that assumption, but if it isn't there's more 
> work to do to generalize this code.
It's been a year but IIRC it was just simpler to implement (and to me it 
didn't make sense to use 64 bits for a subword op).
Is there a benefit in using 64 bit instructions when computing subwords?
>> +
>> +(define_expand "atomic_fetch_nand<mode>"
>> +  [(set (match_operand:SHORT 0 "register_operand" "=&r")
>> +    (match_operand:SHORT 1 "memory_operand" "+A"))
>> +   (set (match_dup 1)
>> +    (unspec_volatile:SHORT
>> +      [(not:SHORT (and:SHORT (match_dup 1)
>> +                 (match_operand:SHORT 2 "reg_or_0_operand" "rJ")))
>> +       (match_operand:SI 3 "const_int_operand")] ;; model
>> +     UNSPEC_SYNC_OLD_OP_SUBWORD))]
>> +  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
> Just a note, constraints aren't necessary for a define_expand. They 
> don't hurt anything though.  They do document expectations, but then 
> you have to maintain them over time.  I'm OK leaving them, mostly 
> wanted to make sure you're aware they aren't strictly necessary for a 
> define_expand.
I wasn't aware, thanks for pointing it out! - you're referring to the 
"TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC", (not the register 
constraints) right?
> ...
Thanks for reviewing!
Patrick
  
Jeff Law April 18, 2023, 9:04 p.m. UTC | #4
On 4/18/23 14:48, Patrick O'Neill wrote:
> On 4/18/23 09:59, Jeff Law wrote:
>> On 4/18/23 08:28, Patrick O'Neill wrote:
>> ...
>>> +  rtx addr = force_reg (Pmode, XEXP (mem, 0));
>>> +
>>> +  rtx aligned_addr = gen_reg_rtx (Pmode);
>>> +  emit_move_insn (aligned_addr,  gen_rtx_AND (Pmode, addr,
>>> +                          gen_int_mode (-4, Pmode)));
>> So rather than -4 as a magic number, GET_MODE_MASK would be better. 
>> That may result in needing to rewrap this code.  I'd bring the 
>> gen_rtx_AND down on a new line, aligned with aligned_addr.
> IIUC GET_MODE_MASK generates masks like 0xFF for QI (for example). It 
> doesn't have the granularity to generate 0x3 (which we can NOT to get 
> -4). I searched the GCC internals docs but couldn't find a function that 
> does address alignment masks.
Yea, yea.  Big "duh" on my side.

>> Presumably using SImode is intentional here rather than wanting to use 
>> word_mode which would be SImode for rv32 and DImode for rv64?  I'm 
>> going to work based on that assumption, but if it isn't there's more 
>> work to do to generalize this code.
> It's been a year but IIRC it was just simpler to implement (and to me it 
> didn't make sense to use 64 bits for a subword op).
> Is there a benefit in using 64 bit instructions when computing subwords?
Given that rv64 should have 32bit load/stores, I don't offhand see any 
advantage.


>>> +
>>> +(define_expand "atomic_fetch_nand<mode>"
>>> +  [(set (match_operand:SHORT 0 "register_operand" "=&r")
>>> +    (match_operand:SHORT 1 "memory_operand" "+A"))
>>> +   (set (match_dup 1)
>>> +    (unspec_volatile:SHORT
>>> +      [(not:SHORT (and:SHORT (match_dup 1)
>>> +                 (match_operand:SHORT 2 "reg_or_0_operand" "rJ")))
>>> +       (match_operand:SI 3 "const_int_operand")] ;; model
>>> +     UNSPEC_SYNC_OLD_OP_SUBWORD))]
>>> +  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
>> Just a note, constraints aren't necessary for a define_expand. They 
>> don't hurt anything though.  They do document expectations, but then 
>> you have to maintain them over time.  I'm OK leaving them, mostly 
>> wanted to make sure you're aware they aren't strictly necessary for a 
>> define_expand.
> I wasn't aware, thanks for pointing it out! - you're referring to the 
> "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC", (not the register 
> constraints) right?
I was referring to the register constraints like "=&r".  They're ignored 
on define_expand constructors.  A define_expand generates RTL that will 
be matched later by a define_insn.

The "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC" is usually referred 
to as the insn condition.


> Thanks for reviewing!
NP.  Looking forward to V6 which I expect will be ready for inclusion.

jeff
  

Patch

diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h
index 5244e8dcbf0..02b33e02020 100644
--- a/gcc/config/riscv/riscv-protos.h
+++ b/gcc/config/riscv/riscv-protos.h
@@ -79,6 +79,8 @@  extern void riscv_reinit (void);
 extern poly_uint64 riscv_regmode_natural_size (machine_mode);
 extern bool riscv_v_ext_vector_mode_p (machine_mode);
 extern bool riscv_shamt_matches_mask_p (int, HOST_WIDE_INT);
+extern void riscv_subword_address (rtx, rtx *, rtx *, rtx *, rtx *);
+extern void riscv_lshift_subword (machine_mode, rtx, rtx, rtx *);
 
 /* Routines implemented in riscv-c.cc.  */
 void riscv_cpu_cpp_builtins (cpp_reader *);
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index e4937d1af25..fa0247be22f 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -7143,6 +7143,56 @@  riscv_zero_call_used_regs (HARD_REG_SET need_zeroed_hardregs)
 							& ~zeroed_hardregs);
 }
 
+/* Helper function for extracting a subword from memory.  */
+
+void
+riscv_subword_address (rtx mem, rtx *aligned_mem, rtx *shift, rtx *mask,
+		       rtx *not_mask)
+{
+  /* Align the memory addess to a word.  */
+  rtx addr = force_reg (Pmode, XEXP (mem, 0));
+
+  rtx aligned_addr = gen_reg_rtx (Pmode);
+  emit_move_insn (aligned_addr,  gen_rtx_AND (Pmode, addr,
+					      gen_int_mode (-4, Pmode)));
+
+  *aligned_mem = change_address (mem, SImode, aligned_addr);
+
+  /* Calculate the shift amount.  */
+  emit_move_insn (*shift, gen_rtx_AND (SImode, gen_lowpart (SImode, addr),
+				       gen_int_mode (3, SImode)));
+  emit_move_insn (*shift, gen_rtx_ASHIFT (SImode, *shift,
+					  gen_int_mode(3, SImode)));
+
+  /* Calculate the mask.  */
+  int unshifted_mask;
+  if (GET_MODE (mem) == QImode)
+    unshifted_mask = 0xFF;
+  else
+    unshifted_mask = 0xFFFF;
+
+  emit_move_insn (*mask, gen_int_mode(unshifted_mask, SImode));
+
+  emit_move_insn (*mask, gen_rtx_ASHIFT(SImode, *mask,
+					gen_lowpart (QImode, *shift)));
+
+  emit_move_insn (*not_mask, gen_rtx_NOT(SImode, *mask));
+}
+
+/* Leftshift a subword within an SImode register.  */
+
+void
+riscv_lshift_subword (machine_mode mode, rtx value, rtx shift,
+		      rtx *shifted_value)
+{
+  rtx value_reg = gen_reg_rtx (SImode);
+  emit_move_insn (value_reg, simplify_gen_subreg (SImode, value,
+						  mode, 0));
+
+  emit_move_insn(*shifted_value, gen_rtx_ASHIFT(SImode, value_reg,
+						gen_lowpart (QImode, shift)));
+}
+
 /* Initialize the GCC target structure.  */
 #undef TARGET_ASM_ALIGNED_HI_OP
 #define TARGET_ASM_ALIGNED_HI_OP "\t.half\t"
diff --git a/gcc/config/riscv/riscv.opt b/gcc/config/riscv/riscv.opt
index ff1dd4ddd4f..bc5e63ab3e6 100644
--- a/gcc/config/riscv/riscv.opt
+++ b/gcc/config/riscv/riscv.opt
@@ -254,3 +254,7 @@  Enum(isa_spec_class) String(20191213) Value(ISA_SPEC_CLASS_20191213)
 misa-spec=
 Target RejectNegative Joined Enum(isa_spec_class) Var(riscv_isa_spec) Init(TARGET_DEFAULT_ISA_SPEC)
 Set the version of RISC-V ISA spec.
+
+minline-atomics
+Target Var(TARGET_INLINE_SUBWORD_ATOMIC) Init(1)
+Always inline subword atomic operations.
diff --git a/gcc/config/riscv/sync.md b/gcc/config/riscv/sync.md
index c932ef87b9d..5adbe08e2d6 100644
--- a/gcc/config/riscv/sync.md
+++ b/gcc/config/riscv/sync.md
@@ -21,8 +21,11 @@ 
 
 (define_c_enum "unspec" [
   UNSPEC_COMPARE_AND_SWAP
+  UNSPEC_COMPARE_AND_SWAP_SUBWORD
   UNSPEC_SYNC_OLD_OP
+  UNSPEC_SYNC_OLD_OP_SUBWORD
   UNSPEC_SYNC_EXCHANGE
+  UNSPEC_SYNC_EXCHANGE_SUBWORD
   UNSPEC_ATOMIC_STORE
   UNSPEC_MEMORY_BARRIER
 ])
@@ -91,6 +94,143 @@ 
   [(set_attr "type" "atomic")
    (set (attr "length") (const_int 8))])
 
+(define_insn "subword_atomic_fetch_strong_<atomic_optab>"
+  [(set (match_operand:SI 0 "register_operand" "=&r")		   ;; old value at mem
+	(match_operand:SI 1 "memory_operand" "+A"))		   ;; mem location
+   (set (match_dup 1)
+	(unspec_volatile:SI
+	  [(any_atomic:SI (match_dup 1)
+		     (match_operand:SI 2 "register_operand" "rI")) ;; value for op
+	   (match_operand:SI 3 "register_operand" "rI")]	   ;; mask
+	 UNSPEC_SYNC_OLD_OP_SUBWORD))
+    (match_operand:SI 4 "register_operand" "rI")		   ;; not_mask
+    (clobber (match_scratch:SI 5 "=&r"))			   ;; tmp_1
+    (clobber (match_scratch:SI 6 "=&r"))]			   ;; tmp_2
+  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
+  {
+    return "1:\;"
+	   "lr.w.aq\t%0, %1\;"
+	   "<insn>\t%5, %0, %2\;"
+	   "and\t%5, %5, %3\;"
+	   "and\t%6, %0, %4\;"
+	   "or\t%6, %6, %5\;"
+	   "sc.w.rl\t%5, %6, %1\;"
+	   "bnez\t%5, 1b";
+  }
+  [(set (attr "length") (const_int 28))])
+
+(define_expand "atomic_fetch_nand<mode>"
+  [(set (match_operand:SHORT 0 "register_operand" "=&r")
+	(match_operand:SHORT 1 "memory_operand" "+A"))
+   (set (match_dup 1)
+	(unspec_volatile:SHORT
+	  [(not:SHORT (and:SHORT (match_dup 1)
+				 (match_operand:SHORT 2 "reg_or_0_operand" "rJ")))
+	   (match_operand:SI 3 "const_int_operand")] ;; model
+	 UNSPEC_SYNC_OLD_OP_SUBWORD))]
+  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
+{
+  /* We have no QImode/HImode atomics, so form a mask, then use
+     subword_atomic_fetch_strong_nand to implement a LR/SC version of the
+     operation. */
+
+  /* Logic duplicated in gcc/libgcc/config/riscv/atomic.c for use when inlining
+     is disabled */
+
+  rtx old = gen_reg_rtx (SImode);
+  rtx mem = operands[1];
+  rtx value = operands[2];
+  rtx aligned_mem = gen_reg_rtx (SImode);
+  rtx shift = gen_reg_rtx (SImode);
+  rtx mask = gen_reg_rtx (SImode);
+  rtx not_mask = gen_reg_rtx (SImode);
+
+  riscv_subword_address (mem, &aligned_mem, &shift, &mask, &not_mask);
+
+  rtx shifted_value = gen_reg_rtx (SImode);
+  riscv_lshift_subword (<MODE>mode, value, shift, &shifted_value);
+
+  emit_insn (gen_subword_atomic_fetch_strong_nand (old, aligned_mem,
+						   shifted_value,
+						   mask, not_mask));
+
+  emit_move_insn (old, gen_rtx_ASHIFTRT (SImode, old,
+					 gen_lowpart (QImode, shift)));
+
+  emit_move_insn (operands[0], gen_lowpart (<MODE>mode, old));
+
+  DONE;
+})
+
+(define_insn "subword_atomic_fetch_strong_nand"
+  [(set (match_operand:SI 0 "register_operand" "=&r")			  ;; old value at mem
+	(match_operand:SI 1 "memory_operand" "+A"))			  ;; mem location
+   (set (match_dup 1)
+	(unspec_volatile:SI
+	  [(not:SI (and:SI (match_dup 1)
+			   (match_operand:SI 2 "register_operand" "rI"))) ;; value for op
+	   (match_operand:SI 3 "register_operand" "rI")]		  ;; mask
+	 UNSPEC_SYNC_OLD_OP_SUBWORD))
+    (match_operand:SI 4 "register_operand" "rI")			  ;; not_mask
+    (clobber (match_scratch:SI 5 "=&r"))				  ;; tmp_1
+    (clobber (match_scratch:SI 6 "=&r"))]				  ;; tmp_2
+  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
+  {
+    return "1:\;"
+	   "lr.w.aq\t%0, %1\;"
+	   "and\t%5, %0, %2\;"
+	   "not\t%5, %5\;"
+	   "and\t%5, %5, %3\;"
+	   "and\t%6, %0, %4\;"
+	   "or\t%6, %6, %5\;"
+	   "sc.w.rl\t%5, %6, %1\;"
+	   "bnez\t%5, 1b";
+  }
+  [(set (attr "length") (const_int 32))])
+
+(define_expand "atomic_fetch_<atomic_optab><mode>"
+  [(set (match_operand:SHORT 0 "register_operand" "=&r")	      ;; old value at mem
+	(match_operand:SHORT 1 "memory_operand" "+A"))		      ;; mem location
+   (set (match_dup 1)
+	(unspec_volatile:SHORT
+	  [(any_atomic:SHORT (match_dup 1)
+		     (match_operand:SHORT 2 "reg_or_0_operand" "rJ")) ;; value for op
+	   (match_operand:SI 3 "const_int_operand")]		      ;; model
+	 UNSPEC_SYNC_OLD_OP_SUBWORD))]
+  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
+{
+  /* We have no QImode/HImode atomics, so form a mask, then use
+     subword_atomic_fetch_strong_<mode> to implement a LR/SC version of the
+     operation. */
+
+  /* Logic duplicated in gcc/libgcc/config/riscv/atomic.c for use when inlining
+     is disabled */
+
+  rtx old = gen_reg_rtx (SImode);
+  rtx mem = operands[1];
+  rtx value = operands[2];
+  rtx aligned_mem = gen_reg_rtx (SImode);
+  rtx shift = gen_reg_rtx (SImode);
+  rtx mask = gen_reg_rtx (SImode);
+  rtx not_mask = gen_reg_rtx (SImode);
+
+  riscv_subword_address (mem, &aligned_mem, &shift, &mask, &not_mask);
+
+  rtx shifted_value = gen_reg_rtx (SImode);
+  riscv_lshift_subword (<MODE>mode, value, shift, &shifted_value);
+
+  emit_insn (gen_subword_atomic_fetch_strong_<atomic_optab> (old, aligned_mem,
+							     shifted_value,
+							     mask, not_mask));
+
+  emit_move_insn (old, gen_rtx_ASHIFTRT (SImode, old,
+					 gen_lowpart (QImode, shift)));
+
+  emit_move_insn (operands[0], gen_lowpart (<MODE>mode, old));
+
+  DONE;
+})
+
 (define_insn "atomic_exchange<mode>"
   [(set (match_operand:GPR 0 "register_operand" "=&r")
 	(unspec_volatile:GPR
@@ -104,6 +244,59 @@ 
   [(set_attr "type" "atomic")
    (set (attr "length") (const_int 8))])
 
+(define_expand "atomic_exchange<mode>"
+  [(set (match_operand:SHORT 0 "register_operand" "=&r")
+	(unspec_volatile:SHORT
+	  [(match_operand:SHORT 1 "memory_operand" "+A")
+	   (match_operand:SI 3 "const_int_operand")] ;; model
+	  UNSPEC_SYNC_EXCHANGE_SUBWORD))
+   (set (match_dup 1)
+	(match_operand:SHORT 2 "register_operand" "0"))]
+  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
+{
+  rtx old = gen_reg_rtx (SImode);
+  rtx mem = operands[1];
+  rtx value = operands[2];
+  rtx aligned_mem = gen_reg_rtx (SImode);
+  rtx shift = gen_reg_rtx (SImode);
+  rtx mask = gen_reg_rtx (SImode);
+  rtx not_mask = gen_reg_rtx (SImode);
+
+  riscv_subword_address (mem, &aligned_mem, &shift, &mask, &not_mask);
+
+  rtx shifted_value = gen_reg_rtx (SImode);
+  riscv_lshift_subword (<MODE>mode, value, shift, &shifted_value);
+
+  emit_insn (gen_subword_atomic_exchange_strong (old, aligned_mem,
+						 shifted_value, not_mask));
+
+  emit_move_insn (old, gen_rtx_ASHIFTRT (SImode, old,
+					 gen_lowpart (QImode, shift)));
+
+  emit_move_insn (operands[0], gen_lowpart (<MODE>mode, old));
+  DONE;
+})
+
+(define_insn "subword_atomic_exchange_strong"
+  [(set (match_operand:SI 0 "register_operand" "=&r")	 ;; old value at mem
+	(match_operand:SI 1 "memory_operand" "+A"))	 ;; mem location
+   (set (match_dup 1)
+	(unspec_volatile:SI
+	  [(match_operand:SI 2 "reg_or_0_operand" "rI")  ;; value
+	   (match_operand:SI 3 "reg_or_0_operand" "rI")] ;; not_mask
+      UNSPEC_SYNC_EXCHANGE_SUBWORD))
+    (clobber (match_scratch:SI 4 "=&r"))]		 ;; tmp_1
+  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
+  {
+    return "1:\;"
+	   "lr.w.aq\t%0, %1\;"
+	   "and\t%4, %0, %3\;"
+	   "or\t%4, %4, %2\;"
+	   "sc.w.rl\t%4, %4, %1\;"
+	   "bnez\t%4, 1b";
+  }
+  [(set (attr "length") (const_int 20))])
+
 (define_insn "atomic_cas_value_strong<mode>"
   [(set (match_operand:GPR 0 "register_operand" "=&r")
 	(match_operand:GPR 1 "memory_operand" "+A"))
@@ -153,6 +346,127 @@ 
   DONE;
 })
 
+(define_expand "atomic_compare_and_swap<mode>"
+  [(match_operand:SI 0 "register_operand" "")    ;; bool output
+   (match_operand:SHORT 1 "register_operand" "") ;; val output
+   (match_operand:SHORT 2 "memory_operand" "")   ;; memory
+   (match_operand:SHORT 3 "reg_or_0_operand" "") ;; expected value
+   (match_operand:SHORT 4 "reg_or_0_operand" "") ;; desired value
+   (match_operand:SI 5 "const_int_operand" "")   ;; is_weak
+   (match_operand:SI 6 "const_int_operand" "")   ;; mod_s
+   (match_operand:SI 7 "const_int_operand" "")]  ;; mod_f
+  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
+{
+  emit_insn (gen_atomic_cas_value_strong<mode> (operands[1], operands[2],
+						operands[3], operands[4],
+						operands[6], operands[7]));
+
+  rtx val = gen_reg_rtx (SImode);
+  if (operands[1] != const0_rtx)
+    emit_move_insn (val, gen_rtx_SIGN_EXTEND (SImode, operands[1]));
+  else
+    emit_move_insn (val, const0_rtx);
+
+  rtx exp = gen_reg_rtx (SImode);
+  if (operands[3] != const0_rtx)
+    emit_move_insn (exp, gen_rtx_SIGN_EXTEND (SImode, operands[3]));
+  else
+    emit_move_insn (exp, const0_rtx);
+
+  rtx compare = val;
+  if (exp != const0_rtx)
+    {
+      rtx difference = gen_rtx_MINUS (SImode, val, exp);
+      compare = gen_reg_rtx (SImode);
+      emit_move_insn  (compare, difference);
+    }
+
+  if (word_mode != SImode)
+    {
+      rtx reg = gen_reg_rtx (word_mode);
+      emit_move_insn (reg, gen_rtx_SIGN_EXTEND (word_mode, compare));
+      compare = reg;
+    }
+
+  emit_move_insn (operands[0], gen_rtx_EQ (SImode, compare, const0_rtx));
+  DONE;
+})
+
+(define_expand "atomic_cas_value_strong<mode>"
+  [(set (match_operand:SHORT 0 "register_operand" "=&r")			;; val output
+	(match_operand:SHORT 1 "memory_operand" "+A"))				;; memory
+   (set (match_dup 1)
+	(unspec_volatile:SHORT [(match_operand:SHORT 2 "reg_or_0_operand" "rJ")	;; expected val
+				(match_operand:SHORT 3 "reg_or_0_operand" "rJ")	;; desired val
+				(match_operand:SI 4 "const_int_operand")	;; mod_s
+				(match_operand:SI 5 "const_int_operand")]	;; mod_f
+	 UNSPEC_COMPARE_AND_SWAP_SUBWORD))
+   (clobber (match_scratch:SHORT 6 "=&r"))]
+  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
+{
+  /* We have no QImode/HImode atomics, so form a mask, then use
+     subword_atomic_cas_strong<mode> to implement a LR/SC version of the
+     operation. */
+
+  /* Logic duplicated in gcc/libgcc/config/riscv/atomic.c for use when inlining
+     is disabled */
+
+  rtx old = gen_reg_rtx (SImode);
+  rtx mem = operands[1];
+  rtx aligned_mem = gen_reg_rtx (SImode);
+  rtx shift = gen_reg_rtx (SImode);
+  rtx mask = gen_reg_rtx (SImode);
+  rtx not_mask = gen_reg_rtx (SImode);
+
+  riscv_subword_address (mem, &aligned_mem, &shift, &mask, &not_mask);
+
+  rtx o = operands[2];
+  rtx n = operands[3];
+  rtx shifted_o = gen_reg_rtx (SImode);
+  rtx shifted_n = gen_reg_rtx (SImode);
+
+  riscv_lshift_subword (<MODE>mode, o, shift, &shifted_o);
+  riscv_lshift_subword (<MODE>mode, n, shift, &shifted_n);
+
+  emit_move_insn (shifted_o, gen_rtx_AND (SImode, shifted_o, mask));
+  emit_move_insn (shifted_n, gen_rtx_AND (SImode, shifted_n, mask));
+
+  emit_insn (gen_subword_atomic_cas_strong (old, aligned_mem,
+					    shifted_o, shifted_n,
+					    mask, not_mask));
+
+  emit_move_insn (old, gen_rtx_ASHIFTRT (SImode, old,
+					 gen_lowpart (QImode, shift)));
+
+  emit_move_insn (operands[0], gen_lowpart(<MODE>mode, old));
+
+  DONE;
+})
+
+(define_insn "subword_atomic_cas_strong"
+  [(set (match_operand:SI 0 "register_operand" "=&r")			   ;; old value at mem
+	(match_operand:SI 1 "memory_operand" "+A"))			   ;; mem location
+   (set (match_dup 1)
+	(unspec_volatile:SI [(match_operand:SI 2 "reg_or_0_operand" "rJ")  ;; o
+			     (match_operand:SI 3 "reg_or_0_operand" "rJ")] ;; n
+	 UNSPEC_COMPARE_AND_SWAP_SUBWORD))
+	(match_operand:SI 4 "register_operand" "rI")			   ;; mask
+	(match_operand:SI 5 "register_operand" "rI")			   ;; not_mask
+	(clobber (match_scratch:SI 6 "=&r"))]				   ;; tmp_1
+  "TARGET_ATOMIC && TARGET_INLINE_SUBWORD_ATOMIC"
+  {
+    return "1:\;"
+	   "lr.w.aq\t%0, %1\;"
+	   "and\t%6, %0, %4\;"
+	   "bne\t%6, %z2, 1f\;"
+	   "and\t%6, %0, %5\;"
+	   "or\t%6, %6, %3\;"
+	   "sc.w.rl\t%6, %6, %1\;"
+	   "bnez\t%6, 1b\;"
+	   "1:";
+  }
+  [(set (attr "length") (const_int 28))])
+
 (define_expand "atomic_test_and_set"
   [(match_operand:QI 0 "register_operand" "")     ;; bool output
    (match_operand:QI 1 "memory_operand" "+A")    ;; memory
diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
index a38547f53e5..9c3e91d2fee 100644
--- a/gcc/doc/invoke.texi
+++ b/gcc/doc/invoke.texi
@@ -1227,6 +1227,7 @@  See RS/6000 and PowerPC Options.
 -mstack-protector-guard=@var{guard}  -mstack-protector-guard-reg=@var{reg}
 -mstack-protector-guard-offset=@var{offset}
 -mcsr-check -mno-csr-check}
+-minline-atomics  -mno-inline-atomics
 
 @emph{RL78 Options}
 @gccoptlist{-msim  -mmul=none  -mmul=g13  -mmul=g14  -mallregs
@@ -29006,6 +29007,13 @@  Do or don't use smaller but slower prologue and epilogue code that uses
 library function calls.  The default is to use fast inline prologues and
 epilogues.
 
+@item -minline-atomics
+@itemx -mno-inline-atomics
+@opindex minline-atomics
+Do or don't use smaller but slower subword atomic emulation code that uses
+libatomic function calls.  The default is to use fast inline subword atomics
+that do not require libatomic.
+
 @opindex mshorten-memrefs
 @item -mshorten-memrefs
 @itemx -mno-shorten-memrefs
diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-1.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-1.c
new file mode 100644
index 00000000000..5c5623d9b2f
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-1.c
@@ -0,0 +1,18 @@ 
+/* { dg-do compile } */
+/* { dg-options "-mno-inline-atomics" } */
+/* { dg-message "note: '__sync_fetch_and_nand' changed semantics in GCC 4.4" "fetch_and_nand" { target *-*-* } 0 } */
+/* { dg-final { scan-assembler "\tcall\t__sync_fetch_and_add_1" } } */
+/* { dg-final { scan-assembler "\tcall\t__sync_fetch_and_nand_1" } } */
+/* { dg-final { scan-assembler "\tcall\t__sync_bool_compare_and_swap_1" } } */
+
+char foo;
+char bar;
+char baz;
+
+int
+main ()
+{
+  __sync_fetch_and_add(&foo, 1);
+  __sync_fetch_and_nand(&bar, 1);
+  __sync_bool_compare_and_swap (&baz, 1, 2);
+}
diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-2.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-2.c
new file mode 100644
index 00000000000..fdce7a5d71f
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-2.c
@@ -0,0 +1,19 @@ 
+/* { dg-do compile } */
+/* Verify that subword atomics do not generate calls.  */
+/* { dg-options "-minline-atomics" } */
+/* { dg-message "note: '__sync_fetch_and_nand' changed semantics in GCC 4.4" "fetch_and_nand" { target *-*-* } 0 } */
+/* { dg-final { scan-assembler-not "\tcall\t__sync_fetch_and_add_1" } } */
+/* { dg-final { scan-assembler-not "\tcall\t__sync_fetch_and_nand_1" } } */
+/* { dg-final { scan-assembler-not "\tcall\t__sync_bool_compare_and_swap_1" } } */
+
+char foo;
+char bar;
+char baz;
+
+int
+main ()
+{
+  __sync_fetch_and_add(&foo, 1);
+  __sync_fetch_and_nand(&bar, 1);
+  __sync_bool_compare_and_swap (&baz, 1, 2);
+}
diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-3.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-3.c
new file mode 100644
index 00000000000..709f3734377
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-3.c
@@ -0,0 +1,569 @@ 
+/* Check all char alignments.  */
+/* Duplicate logic as libatomic/testsuite/libatomic.c/atomic-op-1.c */
+/* Test __atomic routines for existence and proper execution on 1 byte
+   values with each valid memory model.  */
+/* { dg-do run } */
+/* { dg-options "-minline-atomics -Wno-address-of-packed-member" } */
+
+/* Test the execution of the __atomic_*OP builtin routines for a char.  */
+
+extern void abort(void);
+
+char count, res;
+const char init = ~0;
+
+struct A
+{
+   char a;
+   char b;
+   char c;
+   char d;
+} __attribute__ ((packed)) A;
+
+/* The fetch_op routines return the original value before the operation.  */
+
+void
+test_fetch_add (char* v)
+{
+  *v = 0;
+  count = 1;
+
+  if (__atomic_fetch_add (v, count, __ATOMIC_RELAXED) != 0)
+    abort ();
+
+  if (__atomic_fetch_add (v, 1, __ATOMIC_CONSUME) != 1)
+    abort ();
+
+  if (__atomic_fetch_add (v, count, __ATOMIC_ACQUIRE) != 2)
+    abort ();
+
+  if (__atomic_fetch_add (v, 1, __ATOMIC_RELEASE) != 3)
+    abort ();
+
+  if (__atomic_fetch_add (v, count, __ATOMIC_ACQ_REL) != 4)
+    abort ();
+
+  if (__atomic_fetch_add (v, 1, __ATOMIC_SEQ_CST) != 5)
+    abort ();
+}
+
+
+void
+test_fetch_sub (char* v)
+{
+  *v = res = 20;
+  count = 0;
+
+  if (__atomic_fetch_sub (v, count + 1, __ATOMIC_RELAXED) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, 1, __ATOMIC_CONSUME) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, count + 1, __ATOMIC_ACQUIRE) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, 1, __ATOMIC_RELEASE) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, count + 1, __ATOMIC_ACQ_REL) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, 1, __ATOMIC_SEQ_CST) !=  res--)
+    abort ();
+}
+
+void
+test_fetch_and (char* v)
+{
+  *v = init;
+
+  if (__atomic_fetch_and (v, 0, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_fetch_and (v, init, __ATOMIC_CONSUME) !=  0)
+    abort ();
+
+  if (__atomic_fetch_and (v, 0, __ATOMIC_ACQUIRE) !=  0)
+    abort ();
+
+  *v = ~*v;
+  if (__atomic_fetch_and (v, init, __ATOMIC_RELEASE) !=  init)
+    abort ();
+
+  if (__atomic_fetch_and (v, 0, __ATOMIC_ACQ_REL) !=  init)
+    abort ();
+
+  if (__atomic_fetch_and (v, 0, __ATOMIC_SEQ_CST) !=  0)
+    abort ();
+}
+
+void
+test_fetch_nand (char* v)
+{
+  *v = init;
+
+  if (__atomic_fetch_nand (v, 0, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_fetch_nand (v, init, __ATOMIC_CONSUME) !=  init)
+    abort ();
+
+  if (__atomic_fetch_nand (v, 0, __ATOMIC_ACQUIRE) !=  0 )
+    abort ();
+
+  if (__atomic_fetch_nand (v, init, __ATOMIC_RELEASE) !=  init)
+    abort ();
+
+  if (__atomic_fetch_nand (v, init, __ATOMIC_ACQ_REL) !=  0)
+    abort ();
+
+  if (__atomic_fetch_nand (v, 0, __ATOMIC_SEQ_CST) !=  init)
+    abort ();
+}
+
+void
+test_fetch_xor (char* v)
+{
+  *v = init;
+  count = 0;
+
+  if (__atomic_fetch_xor (v, count, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_fetch_xor (v, ~count, __ATOMIC_CONSUME) !=  init)
+    abort ();
+
+  if (__atomic_fetch_xor (v, 0, __ATOMIC_ACQUIRE) !=  0)
+    abort ();
+
+  if (__atomic_fetch_xor (v, ~count, __ATOMIC_RELEASE) !=  0)
+    abort ();
+
+  if (__atomic_fetch_xor (v, 0, __ATOMIC_ACQ_REL) !=  init)
+    abort ();
+
+  if (__atomic_fetch_xor (v, ~count, __ATOMIC_SEQ_CST) !=  init)
+    abort ();
+}
+
+void
+test_fetch_or (char* v)
+{
+  *v = 0;
+  count = 1;
+
+  if (__atomic_fetch_or (v, count, __ATOMIC_RELAXED) !=  0)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, 2, __ATOMIC_CONSUME) !=  1)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, count, __ATOMIC_ACQUIRE) !=  3)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, 8, __ATOMIC_RELEASE) !=  7)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, count, __ATOMIC_ACQ_REL) !=  15)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, count, __ATOMIC_SEQ_CST) !=  31)
+    abort ();
+}
+
+/* The OP_fetch routines return the new value after the operation.  */
+
+void
+test_add_fetch (char* v)
+{
+  *v = 0;
+  count = 1;
+
+  if (__atomic_add_fetch (v, count, __ATOMIC_RELAXED) != 1)
+    abort ();
+
+  if (__atomic_add_fetch (v, 1, __ATOMIC_CONSUME) != 2)
+    abort ();
+
+  if (__atomic_add_fetch (v, count, __ATOMIC_ACQUIRE) != 3)
+    abort ();
+
+  if (__atomic_add_fetch (v, 1, __ATOMIC_RELEASE) != 4)
+    abort ();
+
+  if (__atomic_add_fetch (v, count, __ATOMIC_ACQ_REL) != 5)
+    abort ();
+
+  if (__atomic_add_fetch (v, count, __ATOMIC_SEQ_CST) != 6)
+    abort ();
+}
+
+
+void
+test_sub_fetch (char* v)
+{
+  *v = res = 20;
+  count = 0;
+
+  if (__atomic_sub_fetch (v, count + 1, __ATOMIC_RELAXED) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, 1, __ATOMIC_CONSUME) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, count + 1, __ATOMIC_ACQUIRE) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, 1, __ATOMIC_RELEASE) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, count + 1, __ATOMIC_ACQ_REL) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, count + 1, __ATOMIC_SEQ_CST) !=  --res)
+    abort ();
+}
+
+void
+test_and_fetch (char* v)
+{
+  *v = init;
+
+  if (__atomic_and_fetch (v, 0, __ATOMIC_RELAXED) !=  0)
+    abort ();
+
+  *v = init;
+  if (__atomic_and_fetch (v, init, __ATOMIC_CONSUME) !=  init)
+    abort ();
+
+  if (__atomic_and_fetch (v, 0, __ATOMIC_ACQUIRE) !=  0)
+    abort ();
+
+  *v = ~*v;
+  if (__atomic_and_fetch (v, init, __ATOMIC_RELEASE) !=  init)
+    abort ();
+
+  if (__atomic_and_fetch (v, 0, __ATOMIC_ACQ_REL) !=  0)
+    abort ();
+
+  *v = ~*v;
+  if (__atomic_and_fetch (v, 0, __ATOMIC_SEQ_CST) !=  0)
+    abort ();
+}
+
+void
+test_nand_fetch (char* v)
+{
+  *v = init;
+
+  if (__atomic_nand_fetch (v, 0, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_nand_fetch (v, init, __ATOMIC_CONSUME) !=  0)
+    abort ();
+
+  if (__atomic_nand_fetch (v, 0, __ATOMIC_ACQUIRE) !=  init)
+    abort ();
+
+  if (__atomic_nand_fetch (v, init, __ATOMIC_RELEASE) !=  0)
+    abort ();
+
+  if (__atomic_nand_fetch (v, init, __ATOMIC_ACQ_REL) !=  init)
+    abort ();
+
+  if (__atomic_nand_fetch (v, 0, __ATOMIC_SEQ_CST) !=  init)
+    abort ();
+}
+
+
+
+void
+test_xor_fetch (char* v)
+{
+  *v = init;
+  count = 0;
+
+  if (__atomic_xor_fetch (v, count, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_xor_fetch (v, ~count, __ATOMIC_CONSUME) !=  0)
+    abort ();
+
+  if (__atomic_xor_fetch (v, 0, __ATOMIC_ACQUIRE) !=  0)
+    abort ();
+
+  if (__atomic_xor_fetch (v, ~count, __ATOMIC_RELEASE) !=  init)
+    abort ();
+
+  if (__atomic_xor_fetch (v, 0, __ATOMIC_ACQ_REL) !=  init)
+    abort ();
+
+  if (__atomic_xor_fetch (v, ~count, __ATOMIC_SEQ_CST) !=  0)
+    abort ();
+}
+
+void
+test_or_fetch (char* v)
+{
+  *v = 0;
+  count = 1;
+
+  if (__atomic_or_fetch (v, count, __ATOMIC_RELAXED) !=  1)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, 2, __ATOMIC_CONSUME) !=  3)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, count, __ATOMIC_ACQUIRE) !=  7)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, 8, __ATOMIC_RELEASE) !=  15)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, count, __ATOMIC_ACQ_REL) !=  31)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, count, __ATOMIC_SEQ_CST) !=  63)
+    abort ();
+}
+
+
+/* Test the OP routines with a result which isn't used. Use both variations
+   within each function.  */
+
+void
+test_add (char* v)
+{
+  *v = 0;
+  count = 1;
+
+  __atomic_add_fetch (v, count, __ATOMIC_RELAXED);
+  if (*v != 1)
+    abort ();
+
+  __atomic_fetch_add (v, count, __ATOMIC_CONSUME);
+  if (*v != 2)
+    abort ();
+
+  __atomic_add_fetch (v, 1 , __ATOMIC_ACQUIRE);
+  if (*v != 3)
+    abort ();
+
+  __atomic_fetch_add (v, 1, __ATOMIC_RELEASE);
+  if (*v != 4)
+    abort ();
+
+  __atomic_add_fetch (v, count, __ATOMIC_ACQ_REL);
+  if (*v != 5)
+    abort ();
+
+  __atomic_fetch_add (v, count, __ATOMIC_SEQ_CST);
+  if (*v != 6)
+    abort ();
+}
+
+
+void
+test_sub (char* v)
+{
+  *v = res = 20;
+  count = 0;
+
+  __atomic_sub_fetch (v, count + 1, __ATOMIC_RELAXED);
+  if (*v != --res)
+    abort ();
+
+  __atomic_fetch_sub (v, count + 1, __ATOMIC_CONSUME);
+  if (*v != --res)
+    abort ();
+
+  __atomic_sub_fetch (v, 1, __ATOMIC_ACQUIRE);
+  if (*v != --res)
+    abort ();
+
+  __atomic_fetch_sub (v, 1, __ATOMIC_RELEASE);
+  if (*v != --res)
+    abort ();
+
+  __atomic_sub_fetch (v, count + 1, __ATOMIC_ACQ_REL);
+  if (*v != --res)
+    abort ();
+
+  __atomic_fetch_sub (v, count + 1, __ATOMIC_SEQ_CST);
+  if (*v != --res)
+    abort ();
+}
+
+void
+test_and (char* v)
+{
+  *v = init;
+
+  __atomic_and_fetch (v, 0, __ATOMIC_RELAXED);
+  if (*v != 0)
+    abort ();
+
+  *v = init;
+  __atomic_fetch_and (v, init, __ATOMIC_CONSUME);
+  if (*v != init)
+    abort ();
+
+  __atomic_and_fetch (v, 0, __ATOMIC_ACQUIRE);
+  if (*v != 0)
+    abort ();
+
+  *v = ~*v;
+  __atomic_fetch_and (v, init, __ATOMIC_RELEASE);
+  if (*v != init)
+    abort ();
+
+  __atomic_and_fetch (v, 0, __ATOMIC_ACQ_REL);
+  if (*v != 0)
+    abort ();
+
+  *v = ~*v;
+  __atomic_fetch_and (v, 0, __ATOMIC_SEQ_CST);
+  if (*v != 0)
+    abort ();
+}
+
+void
+test_nand (char* v)
+{
+  *v = init;
+
+  __atomic_fetch_nand (v, 0, __ATOMIC_RELAXED);
+  if (*v != init)
+    abort ();
+
+  __atomic_fetch_nand (v, init, __ATOMIC_CONSUME);
+  if (*v != 0)
+    abort ();
+
+  __atomic_nand_fetch (v, 0, __ATOMIC_ACQUIRE);
+  if (*v != init)
+    abort ();
+
+  __atomic_nand_fetch (v, init, __ATOMIC_RELEASE);
+  if (*v != 0)
+    abort ();
+
+  __atomic_fetch_nand (v, init, __ATOMIC_ACQ_REL);
+  if (*v != init)
+    abort ();
+
+  __atomic_nand_fetch (v, 0, __ATOMIC_SEQ_CST);
+  if (*v != init)
+    abort ();
+}
+
+
+
+void
+test_xor (char* v)
+{
+  *v = init;
+  count = 0;
+
+  __atomic_xor_fetch (v, count, __ATOMIC_RELAXED);
+  if (*v != init)
+    abort ();
+
+  __atomic_fetch_xor (v, ~count, __ATOMIC_CONSUME);
+  if (*v != 0)
+    abort ();
+
+  __atomic_xor_fetch (v, 0, __ATOMIC_ACQUIRE);
+  if (*v != 0)
+    abort ();
+
+  __atomic_fetch_xor (v, ~count, __ATOMIC_RELEASE);
+  if (*v != init)
+    abort ();
+
+  __atomic_fetch_xor (v, 0, __ATOMIC_ACQ_REL);
+  if (*v != init)
+    abort ();
+
+  __atomic_xor_fetch (v, ~count, __ATOMIC_SEQ_CST);
+  if (*v != 0)
+    abort ();
+}
+
+void
+test_or (char* v)
+{
+  *v = 0;
+  count = 1;
+
+  __atomic_or_fetch (v, count, __ATOMIC_RELAXED);
+  if (*v != 1)
+    abort ();
+
+  count *= 2;
+  __atomic_fetch_or (v, count, __ATOMIC_CONSUME);
+  if (*v != 3)
+    abort ();
+
+  count *= 2;
+  __atomic_or_fetch (v, 4, __ATOMIC_ACQUIRE);
+  if (*v != 7)
+    abort ();
+
+  count *= 2;
+  __atomic_fetch_or (v, 8, __ATOMIC_RELEASE);
+  if (*v != 15)
+    abort ();
+
+  count *= 2;
+  __atomic_or_fetch (v, count, __ATOMIC_ACQ_REL);
+  if (*v != 31)
+    abort ();
+
+  count *= 2;
+  __atomic_fetch_or (v, count, __ATOMIC_SEQ_CST);
+  if (*v != 63)
+    abort ();
+}
+
+int
+main ()
+{
+  char* V[] = {&A.a, &A.b, &A.c, &A.d};
+
+  for (int i = 0; i < 4; i++) {
+    test_fetch_add (V[i]);
+    test_fetch_sub (V[i]);
+    test_fetch_and (V[i]);
+    test_fetch_nand (V[i]);
+    test_fetch_xor (V[i]);
+    test_fetch_or (V[i]);
+
+    test_add_fetch (V[i]);
+    test_sub_fetch (V[i]);
+    test_and_fetch (V[i]);
+    test_nand_fetch (V[i]);
+    test_xor_fetch (V[i]);
+    test_or_fetch (V[i]);
+
+    test_add (V[i]);
+    test_sub (V[i]);
+    test_and (V[i]);
+    test_nand (V[i]);
+    test_xor (V[i]);
+    test_or (V[i]);
+  }
+
+  return 0;
+}
diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-4.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-4.c
new file mode 100644
index 00000000000..eecfaae5cc6
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-4.c
@@ -0,0 +1,566 @@ 
+/* Check all short alignments.  */
+/* Duplicate logic as libatomic/testsuite/libatomic.c/atomic-op-2.c */
+/* Test __atomic routines for existence and proper execution on 2 byte
+   values with each valid memory model.  */
+/* { dg-do run } */
+/* { dg-options "-minline-atomics -Wno-address-of-packed-member" } */
+
+/* Test the execution of the __atomic_*OP builtin routines for a short.  */
+
+extern void abort(void);
+
+short count, res;
+const short init = ~0;
+
+struct A
+{
+   short a;
+   short b;
+} __attribute__ ((packed)) A;
+
+/* The fetch_op routines return the original value before the operation.  */
+
+void
+test_fetch_add (short* v)
+{
+  *v = 0;
+  count = 1;
+
+  if (__atomic_fetch_add (v, count, __ATOMIC_RELAXED) != 0)
+    abort ();
+
+  if (__atomic_fetch_add (v, 1, __ATOMIC_CONSUME) != 1)
+    abort ();
+
+  if (__atomic_fetch_add (v, count, __ATOMIC_ACQUIRE) != 2)
+    abort ();
+
+  if (__atomic_fetch_add (v, 1, __ATOMIC_RELEASE) != 3)
+    abort ();
+
+  if (__atomic_fetch_add (v, count, __ATOMIC_ACQ_REL) != 4)
+    abort ();
+
+  if (__atomic_fetch_add (v, 1, __ATOMIC_SEQ_CST) != 5)
+    abort ();
+}
+
+
+void
+test_fetch_sub (short* v)
+{
+  *v = res = 20;
+  count = 0;
+
+  if (__atomic_fetch_sub (v, count + 1, __ATOMIC_RELAXED) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, 1, __ATOMIC_CONSUME) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, count + 1, __ATOMIC_ACQUIRE) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, 1, __ATOMIC_RELEASE) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, count + 1, __ATOMIC_ACQ_REL) !=  res--)
+    abort ();
+
+  if (__atomic_fetch_sub (v, 1, __ATOMIC_SEQ_CST) !=  res--)
+    abort ();
+}
+
+void
+test_fetch_and (short* v)
+{
+  *v = init;
+
+  if (__atomic_fetch_and (v, 0, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_fetch_and (v, init, __ATOMIC_CONSUME) !=  0)
+    abort ();
+
+  if (__atomic_fetch_and (v, 0, __ATOMIC_ACQUIRE) !=  0)
+    abort ();
+
+  *v = ~*v;
+  if (__atomic_fetch_and (v, init, __ATOMIC_RELEASE) !=  init)
+    abort ();
+
+  if (__atomic_fetch_and (v, 0, __ATOMIC_ACQ_REL) !=  init)
+    abort ();
+
+  if (__atomic_fetch_and (v, 0, __ATOMIC_SEQ_CST) !=  0)
+    abort ();
+}
+
+void
+test_fetch_nand (short* v)
+{
+  *v = init;
+
+  if (__atomic_fetch_nand (v, 0, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_fetch_nand (v, init, __ATOMIC_CONSUME) !=  init)
+    abort ();
+
+  if (__atomic_fetch_nand (v, 0, __ATOMIC_ACQUIRE) !=  0 )
+    abort ();
+
+  if (__atomic_fetch_nand (v, init, __ATOMIC_RELEASE) !=  init)
+    abort ();
+
+  if (__atomic_fetch_nand (v, init, __ATOMIC_ACQ_REL) !=  0)
+    abort ();
+
+  if (__atomic_fetch_nand (v, 0, __ATOMIC_SEQ_CST) !=  init)
+    abort ();
+}
+
+void
+test_fetch_xor (short* v)
+{
+  *v = init;
+  count = 0;
+
+  if (__atomic_fetch_xor (v, count, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_fetch_xor (v, ~count, __ATOMIC_CONSUME) !=  init)
+    abort ();
+
+  if (__atomic_fetch_xor (v, 0, __ATOMIC_ACQUIRE) !=  0)
+    abort ();
+
+  if (__atomic_fetch_xor (v, ~count, __ATOMIC_RELEASE) !=  0)
+    abort ();
+
+  if (__atomic_fetch_xor (v, 0, __ATOMIC_ACQ_REL) !=  init)
+    abort ();
+
+  if (__atomic_fetch_xor (v, ~count, __ATOMIC_SEQ_CST) !=  init)
+    abort ();
+}
+
+void
+test_fetch_or (short* v)
+{
+  *v = 0;
+  count = 1;
+
+  if (__atomic_fetch_or (v, count, __ATOMIC_RELAXED) !=  0)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, 2, __ATOMIC_CONSUME) !=  1)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, count, __ATOMIC_ACQUIRE) !=  3)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, 8, __ATOMIC_RELEASE) !=  7)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, count, __ATOMIC_ACQ_REL) !=  15)
+    abort ();
+
+  count *= 2;
+  if (__atomic_fetch_or (v, count, __ATOMIC_SEQ_CST) !=  31)
+    abort ();
+}
+
+/* The OP_fetch routines return the new value after the operation.  */
+
+void
+test_add_fetch (short* v)
+{
+  *v = 0;
+  count = 1;
+
+  if (__atomic_add_fetch (v, count, __ATOMIC_RELAXED) != 1)
+    abort ();
+
+  if (__atomic_add_fetch (v, 1, __ATOMIC_CONSUME) != 2)
+    abort ();
+
+  if (__atomic_add_fetch (v, count, __ATOMIC_ACQUIRE) != 3)
+    abort ();
+
+  if (__atomic_add_fetch (v, 1, __ATOMIC_RELEASE) != 4)
+    abort ();
+
+  if (__atomic_add_fetch (v, count, __ATOMIC_ACQ_REL) != 5)
+    abort ();
+
+  if (__atomic_add_fetch (v, count, __ATOMIC_SEQ_CST) != 6)
+    abort ();
+}
+
+
+void
+test_sub_fetch (short* v)
+{
+  *v = res = 20;
+  count = 0;
+
+  if (__atomic_sub_fetch (v, count + 1, __ATOMIC_RELAXED) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, 1, __ATOMIC_CONSUME) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, count + 1, __ATOMIC_ACQUIRE) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, 1, __ATOMIC_RELEASE) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, count + 1, __ATOMIC_ACQ_REL) !=  --res)
+    abort ();
+
+  if (__atomic_sub_fetch (v, count + 1, __ATOMIC_SEQ_CST) !=  --res)
+    abort ();
+}
+
+void
+test_and_fetch (short* v)
+{
+  *v = init;
+
+  if (__atomic_and_fetch (v, 0, __ATOMIC_RELAXED) !=  0)
+    abort ();
+
+  *v = init;
+  if (__atomic_and_fetch (v, init, __ATOMIC_CONSUME) !=  init)
+    abort ();
+
+  if (__atomic_and_fetch (v, 0, __ATOMIC_ACQUIRE) !=  0)
+    abort ();
+
+  *v = ~*v;
+  if (__atomic_and_fetch (v, init, __ATOMIC_RELEASE) !=  init)
+    abort ();
+
+  if (__atomic_and_fetch (v, 0, __ATOMIC_ACQ_REL) !=  0)
+    abort ();
+
+  *v = ~*v;
+  if (__atomic_and_fetch (v, 0, __ATOMIC_SEQ_CST) !=  0)
+    abort ();
+}
+
+void
+test_nand_fetch (short* v)
+{
+  *v = init;
+
+  if (__atomic_nand_fetch (v, 0, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_nand_fetch (v, init, __ATOMIC_CONSUME) !=  0)
+    abort ();
+
+  if (__atomic_nand_fetch (v, 0, __ATOMIC_ACQUIRE) !=  init)
+    abort ();
+
+  if (__atomic_nand_fetch (v, init, __ATOMIC_RELEASE) !=  0)
+    abort ();
+
+  if (__atomic_nand_fetch (v, init, __ATOMIC_ACQ_REL) !=  init)
+    abort ();
+
+  if (__atomic_nand_fetch (v, 0, __ATOMIC_SEQ_CST) !=  init)
+    abort ();
+}
+
+
+
+void
+test_xor_fetch (short* v)
+{
+  *v = init;
+  count = 0;
+
+  if (__atomic_xor_fetch (v, count, __ATOMIC_RELAXED) !=  init)
+    abort ();
+
+  if (__atomic_xor_fetch (v, ~count, __ATOMIC_CONSUME) !=  0)
+    abort ();
+
+  if (__atomic_xor_fetch (v, 0, __ATOMIC_ACQUIRE) !=  0)
+    abort ();
+
+  if (__atomic_xor_fetch (v, ~count, __ATOMIC_RELEASE) !=  init)
+    abort ();
+
+  if (__atomic_xor_fetch (v, 0, __ATOMIC_ACQ_REL) !=  init)
+    abort ();
+
+  if (__atomic_xor_fetch (v, ~count, __ATOMIC_SEQ_CST) !=  0)
+    abort ();
+}
+
+void
+test_or_fetch (short* v)
+{
+  *v = 0;
+  count = 1;
+
+  if (__atomic_or_fetch (v, count, __ATOMIC_RELAXED) !=  1)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, 2, __ATOMIC_CONSUME) !=  3)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, count, __ATOMIC_ACQUIRE) !=  7)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, 8, __ATOMIC_RELEASE) !=  15)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, count, __ATOMIC_ACQ_REL) !=  31)
+    abort ();
+
+  count *= 2;
+  if (__atomic_or_fetch (v, count, __ATOMIC_SEQ_CST) !=  63)
+    abort ();
+}
+
+
+/* Test the OP routines with a result which isn't used. Use both variations
+   within each function.  */
+
+void
+test_add (short* v)
+{
+  *v = 0;
+  count = 1;
+
+  __atomic_add_fetch (v, count, __ATOMIC_RELAXED);
+  if (*v != 1)
+    abort ();
+
+  __atomic_fetch_add (v, count, __ATOMIC_CONSUME);
+  if (*v != 2)
+    abort ();
+
+  __atomic_add_fetch (v, 1 , __ATOMIC_ACQUIRE);
+  if (*v != 3)
+    abort ();
+
+  __atomic_fetch_add (v, 1, __ATOMIC_RELEASE);
+  if (*v != 4)
+    abort ();
+
+  __atomic_add_fetch (v, count, __ATOMIC_ACQ_REL);
+  if (*v != 5)
+    abort ();
+
+  __atomic_fetch_add (v, count, __ATOMIC_SEQ_CST);
+  if (*v != 6)
+    abort ();
+}
+
+
+void
+test_sub (short* v)
+{
+  *v = res = 20;
+  count = 0;
+
+  __atomic_sub_fetch (v, count + 1, __ATOMIC_RELAXED);
+  if (*v != --res)
+    abort ();
+
+  __atomic_fetch_sub (v, count + 1, __ATOMIC_CONSUME);
+  if (*v != --res)
+    abort ();
+
+  __atomic_sub_fetch (v, 1, __ATOMIC_ACQUIRE);
+  if (*v != --res)
+    abort ();
+
+  __atomic_fetch_sub (v, 1, __ATOMIC_RELEASE);
+  if (*v != --res)
+    abort ();
+
+  __atomic_sub_fetch (v, count + 1, __ATOMIC_ACQ_REL);
+  if (*v != --res)
+    abort ();
+
+  __atomic_fetch_sub (v, count + 1, __ATOMIC_SEQ_CST);
+  if (*v != --res)
+    abort ();
+}
+
+void
+test_and (short* v)
+{
+  *v = init;
+
+  __atomic_and_fetch (v, 0, __ATOMIC_RELAXED);
+  if (*v != 0)
+    abort ();
+
+  *v = init;
+  __atomic_fetch_and (v, init, __ATOMIC_CONSUME);
+  if (*v != init)
+    abort ();
+
+  __atomic_and_fetch (v, 0, __ATOMIC_ACQUIRE);
+  if (*v != 0)
+    abort ();
+
+  *v = ~*v;
+  __atomic_fetch_and (v, init, __ATOMIC_RELEASE);
+  if (*v != init)
+    abort ();
+
+  __atomic_and_fetch (v, 0, __ATOMIC_ACQ_REL);
+  if (*v != 0)
+    abort ();
+
+  *v = ~*v;
+  __atomic_fetch_and (v, 0, __ATOMIC_SEQ_CST);
+  if (*v != 0)
+    abort ();
+}
+
+void
+test_nand (short* v)
+{
+  *v = init;
+
+  __atomic_fetch_nand (v, 0, __ATOMIC_RELAXED);
+  if (*v != init)
+    abort ();
+
+  __atomic_fetch_nand (v, init, __ATOMIC_CONSUME);
+  if (*v != 0)
+    abort ();
+
+  __atomic_nand_fetch (v, 0, __ATOMIC_ACQUIRE);
+  if (*v != init)
+    abort ();
+
+  __atomic_nand_fetch (v, init, __ATOMIC_RELEASE);
+  if (*v != 0)
+    abort ();
+
+  __atomic_fetch_nand (v, init, __ATOMIC_ACQ_REL);
+  if (*v != init)
+    abort ();
+
+  __atomic_nand_fetch (v, 0, __ATOMIC_SEQ_CST);
+  if (*v != init)
+    abort ();
+}
+
+
+
+void
+test_xor (short* v)
+{
+  *v = init;
+  count = 0;
+
+  __atomic_xor_fetch (v, count, __ATOMIC_RELAXED);
+  if (*v != init)
+    abort ();
+
+  __atomic_fetch_xor (v, ~count, __ATOMIC_CONSUME);
+  if (*v != 0)
+    abort ();
+
+  __atomic_xor_fetch (v, 0, __ATOMIC_ACQUIRE);
+  if (*v != 0)
+    abort ();
+
+  __atomic_fetch_xor (v, ~count, __ATOMIC_RELEASE);
+  if (*v != init)
+    abort ();
+
+  __atomic_fetch_xor (v, 0, __ATOMIC_ACQ_REL);
+  if (*v != init)
+    abort ();
+
+  __atomic_xor_fetch (v, ~count, __ATOMIC_SEQ_CST);
+  if (*v != 0)
+    abort ();
+}
+
+void
+test_or (short* v)
+{
+  *v = 0;
+  count = 1;
+
+  __atomic_or_fetch (v, count, __ATOMIC_RELAXED);
+  if (*v != 1)
+    abort ();
+
+  count *= 2;
+  __atomic_fetch_or (v, count, __ATOMIC_CONSUME);
+  if (*v != 3)
+    abort ();
+
+  count *= 2;
+  __atomic_or_fetch (v, 4, __ATOMIC_ACQUIRE);
+  if (*v != 7)
+    abort ();
+
+  count *= 2;
+  __atomic_fetch_or (v, 8, __ATOMIC_RELEASE);
+  if (*v != 15)
+    abort ();
+
+  count *= 2;
+  __atomic_or_fetch (v, count, __ATOMIC_ACQ_REL);
+  if (*v != 31)
+    abort ();
+
+  count *= 2;
+  __atomic_fetch_or (v, count, __ATOMIC_SEQ_CST);
+  if (*v != 63)
+    abort ();
+}
+
+int
+main () {
+  short* V[] = {&A.a, &A.b};
+
+  for (int i = 0; i < 2; i++) {
+    test_fetch_add (V[i]);
+    test_fetch_sub (V[i]);
+    test_fetch_and (V[i]);
+    test_fetch_nand (V[i]);
+    test_fetch_xor (V[i]);
+    test_fetch_or (V[i]);
+
+    test_add_fetch (V[i]);
+    test_sub_fetch (V[i]);
+    test_and_fetch (V[i]);
+    test_nand_fetch (V[i]);
+    test_xor_fetch (V[i]);
+    test_or_fetch (V[i]);
+
+    test_add (V[i]);
+    test_sub (V[i]);
+    test_and (V[i]);
+    test_nand (V[i]);
+    test_xor (V[i]);
+    test_or (V[i]);
+  }
+
+  return 0;
+}
diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-5.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-5.c
new file mode 100644
index 00000000000..52093894a79
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-5.c
@@ -0,0 +1,87 @@ 
+/* Test __atomic routines for existence and proper execution on 1 byte
+   values with each valid memory model.  */
+/* Duplicate logic as libatomic/testsuite/libatomic.c/atomic-compare-exchange-1.c */
+/* { dg-do run } */
+/* { dg-options "-minline-atomics" } */
+
+/* Test the execution of the __atomic_compare_exchange_n builtin for a char.  */
+
+extern void abort(void);
+
+char v = 0;
+char expected = 0;
+char max = ~0;
+char desired = ~0;
+char zero = 0;
+
+#define STRONG 0
+#define WEAK 1
+
+int
+main ()
+{
+
+  if (!__atomic_compare_exchange_n (&v, &expected, max, STRONG , __ATOMIC_RELAXED, __ATOMIC_RELAXED))
+    abort ();
+  if (expected != 0)
+    abort ();
+
+  if (__atomic_compare_exchange_n (&v, &expected, 0, STRONG , __ATOMIC_ACQUIRE, __ATOMIC_RELAXED))
+    abort ();
+  if (expected != max)
+    abort ();
+
+  if (!__atomic_compare_exchange_n (&v, &expected, 0, STRONG , __ATOMIC_RELEASE, __ATOMIC_ACQUIRE))
+    abort ();
+  if (expected != max)
+    abort ();
+  if (v != 0)
+    abort ();
+
+  if (__atomic_compare_exchange_n (&v, &expected, desired, WEAK, __ATOMIC_ACQ_REL, __ATOMIC_ACQUIRE))
+    abort ();
+  if (expected != 0)
+    abort ();
+
+  if (!__atomic_compare_exchange_n (&v, &expected, desired, STRONG , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
+    abort ();
+  if (expected != 0)
+    abort ();
+  if (v != max)
+    abort ();
+
+  /* Now test the generic version.  */
+
+  v = 0;
+
+  if (!__atomic_compare_exchange (&v, &expected, &max, STRONG, __ATOMIC_RELAXED, __ATOMIC_RELAXED))
+    abort ();
+  if (expected != 0)
+    abort ();
+
+  if (__atomic_compare_exchange (&v, &expected, &zero, STRONG , __ATOMIC_ACQUIRE, __ATOMIC_RELAXED))
+    abort ();
+  if (expected != max)
+    abort ();
+
+  if (!__atomic_compare_exchange (&v, &expected, &zero, STRONG , __ATOMIC_RELEASE, __ATOMIC_ACQUIRE))
+    abort ();
+  if (expected != max)
+    abort ();
+  if (v != 0)
+    abort ();
+
+  if (__atomic_compare_exchange (&v, &expected, &desired, WEAK, __ATOMIC_ACQ_REL, __ATOMIC_ACQUIRE))
+    abort ();
+  if (expected != 0)
+    abort ();
+
+  if (!__atomic_compare_exchange (&v, &expected, &desired, STRONG , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
+    abort ();
+  if (expected != 0)
+    abort ();
+  if (v != max)
+    abort ();
+
+  return 0;
+}
diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-6.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-6.c
new file mode 100644
index 00000000000..8fee8c44811
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-6.c
@@ -0,0 +1,87 @@ 
+/* Test __atomic routines for existence and proper execution on 2 byte
+   values with each valid memory model.  */
+/* Duplicate logic as libatomic/testsuite/libatomic.c/atomic-compare-exchange-2.c */
+/* { dg-do run } */
+/* { dg-options "-minline-atomics" } */
+
+/* Test the execution of the __atomic_compare_exchange_n builtin for a short.  */
+
+extern void abort(void);
+
+short v = 0;
+short expected = 0;
+short max = ~0;
+short desired = ~0;
+short zero = 0;
+
+#define STRONG 0
+#define WEAK 1
+
+int
+main ()
+{
+
+  if (!__atomic_compare_exchange_n (&v, &expected, max, STRONG , __ATOMIC_RELAXED, __ATOMIC_RELAXED))
+    abort ();
+  if (expected != 0)
+    abort ();
+
+  if (__atomic_compare_exchange_n (&v, &expected, 0, STRONG , __ATOMIC_ACQUIRE, __ATOMIC_RELAXED))
+    abort ();
+  if (expected != max)
+    abort ();
+
+  if (!__atomic_compare_exchange_n (&v, &expected, 0, STRONG , __ATOMIC_RELEASE, __ATOMIC_ACQUIRE))
+    abort ();
+  if (expected != max)
+    abort ();
+  if (v != 0)
+    abort ();
+
+  if (__atomic_compare_exchange_n (&v, &expected, desired, WEAK, __ATOMIC_ACQ_REL, __ATOMIC_ACQUIRE))
+    abort ();
+  if (expected != 0)
+    abort ();
+
+  if (!__atomic_compare_exchange_n (&v, &expected, desired, STRONG , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
+    abort ();
+  if (expected != 0)
+    abort ();
+  if (v != max)
+    abort ();
+
+  /* Now test the generic version.  */
+
+  v = 0;
+
+  if (!__atomic_compare_exchange (&v, &expected, &max, STRONG, __ATOMIC_RELAXED, __ATOMIC_RELAXED))
+    abort ();
+  if (expected != 0)
+    abort ();
+
+  if (__atomic_compare_exchange (&v, &expected, &zero, STRONG , __ATOMIC_ACQUIRE, __ATOMIC_RELAXED))
+    abort ();
+  if (expected != max)
+    abort ();
+
+  if (!__atomic_compare_exchange (&v, &expected, &zero, STRONG , __ATOMIC_RELEASE, __ATOMIC_ACQUIRE))
+    abort ();
+  if (expected != max)
+    abort ();
+  if (v != 0)
+    abort ();
+
+  if (__atomic_compare_exchange (&v, &expected, &desired, WEAK, __ATOMIC_ACQ_REL, __ATOMIC_ACQUIRE))
+    abort ();
+  if (expected != 0)
+    abort ();
+
+  if (!__atomic_compare_exchange (&v, &expected, &desired, STRONG , __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
+    abort ();
+  if (expected != 0)
+    abort ();
+  if (v != max)
+    abort ();
+
+  return 0;
+}
diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-7.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-7.c
new file mode 100644
index 00000000000..24c344c0ce3
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-7.c
@@ -0,0 +1,69 @@ 
+/* Test __atomic routines for existence and proper execution on 1 byte
+   values with each valid memory model.  */
+/* Duplicate logic as libatomic/testsuite/libatomic.c/atomic-exchange-1.c */
+/* { dg-do run } */
+/* { dg-options "-minline-atomics" } */
+
+/* Test the execution of the __atomic_exchange_n builtin for a char.  */
+
+extern void abort(void);
+
+char v, count, ret;
+
+int
+main ()
+{
+  v = 0;
+  count = 0;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_RELAXED) != count)
+    abort ();
+  count++;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_ACQUIRE) != count)
+    abort ();
+  count++;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_RELEASE) != count)
+    abort ();
+  count++;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_ACQ_REL) != count)
+    abort ();
+  count++;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_SEQ_CST) != count)
+    abort ();
+  count++;
+
+  /* Now test the generic version.  */
+
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_RELAXED);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_ACQUIRE);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_RELEASE);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_ACQ_REL);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_SEQ_CST);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  return 0;
+}
diff --git a/gcc/testsuite/gcc.target/riscv/inline-atomics-8.c b/gcc/testsuite/gcc.target/riscv/inline-atomics-8.c
new file mode 100644
index 00000000000..edc212df04e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/inline-atomics-8.c
@@ -0,0 +1,69 @@ 
+/* Test __atomic routines for existence and proper execution on 2 byte
+   values with each valid memory model.  */
+/* Duplicate logic as libatomic/testsuite/libatomic.c/atomic-exchange-2.c */
+/* { dg-do run } */
+/* { dg-options "-minline-atomics" } */
+
+/* Test the execution of the __atomic_X builtin for a short.  */
+
+extern void abort(void);
+
+short v, count, ret;
+
+int
+main ()
+{
+  v = 0;
+  count = 0;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_RELAXED) != count)
+    abort ();
+  count++;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_ACQUIRE) != count)
+    abort ();
+  count++;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_RELEASE) != count)
+    abort ();
+  count++;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_ACQ_REL) != count)
+    abort ();
+  count++;
+
+  if (__atomic_exchange_n (&v, count + 1, __ATOMIC_SEQ_CST) != count)
+    abort ();
+  count++;
+
+  /* Now test the generic version.  */
+
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_RELAXED);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_ACQUIRE);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_RELEASE);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_ACQ_REL);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  __atomic_exchange (&v, &count, &ret, __ATOMIC_SEQ_CST);
+  if (ret != count - 1 || v != count)
+    abort ();
+  count++;
+
+  return 0;
+}
diff --git a/libgcc/config/riscv/atomic.c b/libgcc/config/riscv/atomic.c
index 69f53623509..573d163ea04 100644
--- a/libgcc/config/riscv/atomic.c
+++ b/libgcc/config/riscv/atomic.c
@@ -30,6 +30,8 @@  see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
 #define INVERT		"not %[tmp1], %[tmp1]\n\t"
 #define DONT_INVERT	""
 
+/* Logic duplicated in gcc/gcc/config/riscv/sync.md for use when inlining is enabled */
+
 #define GENERATE_FETCH_AND_OP(type, size, opname, insn, invert, cop)	\
   type __sync_fetch_and_ ## opname ## _ ## size (type *p, type v)	\
   {									\