This patch adds support for the arm_locally_streaming attribute,
which allows a function to use SME internally without changing
the function's ABI. The attribute is valid but redundant for
arm_streaming functions.
gcc/
* config/aarch64/aarch64.cc (aarch64_attribute_table): Add
arm_locally_streaming.
(aarch64_fndecl_is_locally_streaming): New function.
(aarch64_fndecl_sm_state): Handle arm_locally_streaming functions.
(aarch64_cfun_enables_pstate_sm): New function.
(aarch64_add_offset): Add an argument that specifies whether
the streaming vector length should be used instead of the
prevailing one.
(aarch64_split_add_offset, aarch64_add_sp, aarch64_sub_sp): Likewise.
(aarch64_allocate_and_probe_stack_space): Likewise.
(aarch64_expand_mov_immediate): Update calls accordingly.
(aarch64_need_old_pstate_sm): Return true for locally-streaming
streaming-compatible functions.
(aarch64_layout_frame): Force all call-preserved Z and P registers
to be saved and restored if the function switches PSTATE.SM in the
prologue.
(aarch64_get_separate_components): Disable shrink-wrapping of
such Z and P saves and restores.
(aarch64_use_late_prologue_epilogue): New function.
(aarch64_expand_prologue): Measure SVE lengths in the streaming
vector length for locally-streaming functions, then emit code
to enable streaming mode. Combine separate SMSTART ZA and
SMSTART SM instructions into a single SMSTART where possible.
(aarch64_expand_epilogue): Likewise in reverse.
(TARGET_USE_LATE_PROLOGUE_EPILOGUE): Define.
* config/aarch64/aarch64-sme.md (UNSPEC_SMSTART): New unspec.
(UNSPEC_SMSTOP): Likewise.
(aarch64_smstart, aarch64_smstop): New patterns.
gcc/testsuite/
* gcc.target/aarch64/sme/locally_streaming_1.c: New test.
* gcc.target/aarch64/sme/locally_streaming_2.c: Likewise.
* gcc.target/aarch64/sme/locally_streaming_3.c: Likewise.
---
gcc/config/aarch64/aarch64-sme.md | 82 ++++
gcc/config/aarch64/aarch64.cc | 237 ++++++++--
.../aarch64/sme/locally_streaming_1.c | 433 ++++++++++++++++++
.../aarch64/sme/locally_streaming_2.c | 177 +++++++
.../aarch64/sme/locally_streaming_3.c | 273 +++++++++++
5 files changed, 1164 insertions(+), 38 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_1.c
create mode 100644 gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_2.c
create mode 100644 gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_3.c
diff --git a/gcc/config/aarch64/aarch64-sme.md
b/gcc/config/aarch64/aarch64-sme.md
index 7b3ccea2e11..70be7adba28 100644
--- a/gcc/config/aarch64/aarch64-sme.md
+++ b/gcc/config/aarch64/aarch64-sme.md
@@ -281,6 +281,88 @@ (define_insn_and_split "aarch64_restore_za"
DONE;
}
)
+
+;; -------------------------------------------------------------------------
+;; ---- Combined PSTATE.SM and PSTATE.ZA management
+;; -------------------------------------------------------------------------
+;; Includes
+;; - SMSTART
+;; - SMSTOP
+;; -------------------------------------------------------------------------
+
+(define_c_enum "unspec" [
+ UNSPEC_SMSTART
+ UNSPEC_SMSTOP
+])
+
+;; Enable SM and ZA, starting with fresh ZA contents. This is only valid when
+;; SME is present, but the pattern does not depend on TARGET_SME since it can
+;; be used conditionally.
+(define_insn "aarch64_smstart"
+ [(unspec_volatile [(const_int 0)] UNSPEC_SMSTART)
+ (clobber (reg:V4x16QI V0_REGNUM))
+ (clobber (reg:V4x16QI V4_REGNUM))
+ (clobber (reg:V4x16QI V8_REGNUM))
+ (clobber (reg:V4x16QI V12_REGNUM))
+ (clobber (reg:V4x16QI V16_REGNUM))
+ (clobber (reg:V4x16QI V20_REGNUM))
+ (clobber (reg:V4x16QI V24_REGNUM))
+ (clobber (reg:V4x16QI V28_REGNUM))
+ (clobber (reg:VNx16BI P0_REGNUM))
+ (clobber (reg:VNx16BI P1_REGNUM))
+ (clobber (reg:VNx16BI P2_REGNUM))
+ (clobber (reg:VNx16BI P3_REGNUM))
+ (clobber (reg:VNx16BI P4_REGNUM))
+ (clobber (reg:VNx16BI P5_REGNUM))
+ (clobber (reg:VNx16BI P6_REGNUM))
+ (clobber (reg:VNx16BI P7_REGNUM))
+ (clobber (reg:VNx16BI P8_REGNUM))
+ (clobber (reg:VNx16BI P9_REGNUM))
+ (clobber (reg:VNx16BI P10_REGNUM))
+ (clobber (reg:VNx16BI P11_REGNUM))
+ (clobber (reg:VNx16BI P12_REGNUM))
+ (clobber (reg:VNx16BI P13_REGNUM))
+ (clobber (reg:VNx16BI P14_REGNUM))
+ (clobber (reg:VNx16BI P15_REGNUM))
+ (clobber (reg:VNx16QI ZA_REGNUM))]
+ ""
+ "smstart"
+)
+
+;; Disable SM and ZA, and discard its current contents. This is only valid
+;; when SME is present, but the pattern does not depend on TARGET_SME since
+;; it can be used conditionally.
+(define_insn "aarch64_smstop"
+ [(unspec_volatile [(reg:VNx16QI OLD_ZA_REGNUM)] UNSPEC_SMSTOP)
+ (clobber (reg:V4x16QI V0_REGNUM))
+ (clobber (reg:V4x16QI V4_REGNUM))
+ (clobber (reg:V4x16QI V8_REGNUM))
+ (clobber (reg:V4x16QI V12_REGNUM))
+ (clobber (reg:V4x16QI V16_REGNUM))
+ (clobber (reg:V4x16QI V20_REGNUM))
+ (clobber (reg:V4x16QI V24_REGNUM))
+ (clobber (reg:V4x16QI V28_REGNUM))
+ (clobber (reg:VNx16BI P0_REGNUM))
+ (clobber (reg:VNx16BI P1_REGNUM))
+ (clobber (reg:VNx16BI P2_REGNUM))
+ (clobber (reg:VNx16BI P3_REGNUM))
+ (clobber (reg:VNx16BI P4_REGNUM))
+ (clobber (reg:VNx16BI P5_REGNUM))
+ (clobber (reg:VNx16BI P6_REGNUM))
+ (clobber (reg:VNx16BI P7_REGNUM))
+ (clobber (reg:VNx16BI P8_REGNUM))
+ (clobber (reg:VNx16BI P9_REGNUM))
+ (clobber (reg:VNx16BI P10_REGNUM))
+ (clobber (reg:VNx16BI P11_REGNUM))
+ (clobber (reg:VNx16BI P12_REGNUM))
+ (clobber (reg:VNx16BI P13_REGNUM))
+ (clobber (reg:VNx16BI P14_REGNUM))
+ (clobber (reg:VNx16BI P15_REGNUM))
+ (clobber (reg:VNx16QI ZA_REGNUM))]
+ ""
+ "smstop"
+)
+
;; =========================================================================
;; == Loads, stores and moves
;; =========================================================================
diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
index 966d13abe4c..48bf2de4b3d 100644
--- a/gcc/config/aarch64/aarch64.cc
+++ b/gcc/config/aarch64/aarch64.cc
@@ -2790,6 +2790,7 @@ static const struct attribute_spec
aarch64_attribute_table[] =
NULL, attr_streaming_exclusions },
{ "arm_streaming_compatible", 0, 0, false, true, true, true,
NULL, attr_streaming_exclusions },
+ { "arm_locally_streaming", 0, 0, true, false, false, false, NULL, NULL },
{ "arm_new_za", 0, 0, true, false, false, false,
handle_arm_new_za_attribute,
attr_arm_new_za_exclusions },
@@ -4162,6 +4163,15 @@ aarch64_fndecl_has_new_za_state (const_tree fndecl)
return lookup_attribute ("arm_new_za", DECL_ATTRIBUTES (fndecl));
}
+/* Return true if FNDECL uses streaming mode internally, as an
+ implementation choice. */
+
+static bool
+aarch64_fndecl_is_locally_streaming (const_tree fndecl)
+{
+ return lookup_attribute ("arm_locally_streaming", DECL_ATTRIBUTES (fndecl));
+}
+
/* Return the state of PSTATE.SM when compiling the body of
function FNDECL. This might be different from the state of
PSTATE.SM on entry. */
@@ -4169,6 +4179,9 @@ aarch64_fndecl_has_new_za_state (const_tree fndecl)
static aarch64_feature_flags
aarch64_fndecl_sm_state (const_tree fndecl)
{
+ if (aarch64_fndecl_is_locally_streaming (fndecl))
+ return AARCH64_FL_SM_ON;
+
return aarch64_fntype_sm_state (TREE_TYPE (fndecl));
}
@@ -4222,6 +4235,16 @@ aarch64_cfun_incoming_za_state ()
return aarch64_fntype_za_state (TREE_TYPE (cfun->decl));
}
+/* Return true if PSTATE.SM is 1 in the body of the current function,
+ but is not guaranteed to be 1 on entry. */
+
+static bool
+aarch64_cfun_enables_pstate_sm ()
+{
+ return (aarch64_fndecl_is_locally_streaming (cfun->decl)
+ && aarch64_cfun_incoming_sm_state () != AARCH64_FL_SM_ON);
+}
+
/* Return true if the current function creates new ZA state (as opposed
to sharing ZA with its callers or ignoring ZA altogether). */
@@ -6432,6 +6455,10 @@ aarch64_add_offset_temporaries (rtx x)
TEMP2, if nonnull, is a second temporary register that doesn't
overlap either DEST or REG.
+ FORCE_ISA_MODE is AARCH64_FL_SM_ON if any variable component of OFFSET
+ is measured relative to the SME vector length instead of the current
+ prevailing vector length. It is 0 otherwise.
+
Since this function may be used to adjust the stack pointer, we must
ensure that it cannot cause transient stack deallocation (for example
by first incrementing SP and then decrementing when adjusting by a
@@ -6440,6 +6467,7 @@ aarch64_add_offset_temporaries (rtx x)
static void
aarch64_add_offset (scalar_int_mode mode, rtx dest, rtx src,
poly_int64 offset, rtx temp1, rtx temp2,
+ aarch64_feature_flags force_isa_mode,
bool frame_related_p, bool emit_move_imm = true)
{
gcc_assert (emit_move_imm || temp1 != NULL_RTX);
@@ -6452,9 +6480,17 @@ aarch64_add_offset (scalar_int_mode mode, rtx dest, rtx
src,
/* Try using ADDVL or ADDPL to add the whole value. */
if (src != const0_rtx && aarch64_sve_addvl_addpl_immediate_p (offset))
{
- rtx offset_rtx = gen_int_mode (offset, mode);
+ rtx offset_rtx;
+ if (force_isa_mode == 0)
+ offset_rtx = gen_int_mode (offset, mode);
+ else
+ offset_rtx = aarch64_sme_vq_immediate (mode, offset.coeffs[0], 0);
rtx_insn *insn = emit_insn (gen_add3_insn (dest, src, offset_rtx));
RTX_FRAME_RELATED_P (insn) = frame_related_p;
+ if (frame_related_p && (force_isa_mode & AARCH64_FL_SM_ON))
+ add_reg_note (insn, REG_CFA_ADJUST_CFA,
+ gen_rtx_SET (dest, plus_constant (Pmode, src,
+ offset)));
return;
}
@@ -6470,11 +6506,19 @@ aarch64_add_offset (scalar_int_mode mode, rtx dest, rtx
src,
if (src != const0_rtx
&& aarch64_sve_addvl_addpl_immediate_p (poly_offset))
{
- rtx offset_rtx = gen_int_mode (poly_offset, mode);
+ rtx offset_rtx;
+ if (force_isa_mode == 0)
+ offset_rtx = gen_int_mode (poly_offset, mode);
+ else
+ offset_rtx = aarch64_sme_vq_immediate (mode, factor, 0);
if (frame_related_p)
{
rtx_insn *insn = emit_insn (gen_add3_insn (dest, src, offset_rtx));
RTX_FRAME_RELATED_P (insn) = true;
+ if (force_isa_mode & AARCH64_FL_SM_ON)
+ add_reg_note (insn, REG_CFA_ADJUST_CFA,
+ gen_rtx_SET (dest, plus_constant (Pmode, src,
+ poly_offset)));
src = dest;
}
else
@@ -6505,8 +6549,18 @@ aarch64_add_offset (scalar_int_mode mode, rtx dest, rtx
src,
rtx val;
if (IN_RANGE (rel_factor, -32, 31))
{
+ if (force_isa_mode & AARCH64_FL_SM_ON)
+ {
+ /* Try to use an unshifted RDSVL, otherwise fall back on
+ a shifted RDSVL #1. */
+ if (aarch64_sve_rdvl_addvl_factor_p (factor))
+ shift = 0;
+ else
+ factor = rel_factor * 16;
+ val = aarch64_sme_vq_immediate (mode, factor, 0);
+ }
/* Try to use an unshifted CNT[BHWD]. */
- if (aarch64_sve_cnt_factor_p (factor))
+ else if (aarch64_sve_cnt_factor_p (factor))
{
val = gen_int_mode (poly_int64 (factor, factor), mode);
shift = 0;
@@ -6542,12 +6596,19 @@ aarch64_add_offset (scalar_int_mode mode, rtx dest, rtx
src,
a shift and add sequence for the multiplication.
If CNTB << SHIFT is out of range, stick with the current
shift factor. */
- if (IN_RANGE (low_bit, 2, 16 * 16))
+ if (force_isa_mode == 0
+ && IN_RANGE (low_bit, 2, 16 * 16))
{
val = gen_int_mode (poly_int64 (low_bit, low_bit), mode);
shift = 0;
}
- else
+ else if ((force_isa_mode & AARCH64_FL_SM_ON)
+ && aarch64_sve_rdvl_addvl_factor_p (low_bit))
+ {
+ val = aarch64_sme_vq_immediate (mode, low_bit, 0);
+ shift = 0;
+ }
+ else
val = gen_int_mode (BYTES_PER_SVE_VECTOR, mode);
val = aarch64_force_temporary (mode, temp1, val);
@@ -6634,30 +6695,34 @@ aarch64_split_add_offset (scalar_int_mode mode, rtx
dest, rtx src,
rtx offset_rtx, rtx temp1, rtx temp2)
{
aarch64_add_offset (mode, dest, src, rtx_to_poly_int64 (offset_rtx),
- temp1, temp2, false);
+ temp1, temp2, 0, false);
}
/* Add DELTA to the stack pointer, marking the instructions frame-related.
- TEMP1 is available as a temporary if nonnull. EMIT_MOVE_IMM is false
- if TEMP1 already contains abs (DELTA). */
+ TEMP1 is available as a temporary if nonnull. FORCE_ISA_MODE is as
+ for aarch64_add_offset. EMIT_MOVE_IMM is false if TEMP1 already
+ contains abs (DELTA). */
static inline void
-aarch64_add_sp (rtx temp1, rtx temp2, poly_int64 delta, bool emit_move_imm)
+aarch64_add_sp (rtx temp1, rtx temp2, poly_int64 delta,
+ aarch64_feature_flags force_isa_mode, bool emit_move_imm)
{
aarch64_add_offset (Pmode, stack_pointer_rtx, stack_pointer_rtx, delta,
- temp1, temp2, true, emit_move_imm);
+ temp1, temp2, force_isa_mode, true, emit_move_imm);
}
/* Subtract DELTA from the stack pointer, marking the instructions
- frame-related if FRAME_RELATED_P. TEMP1 is available as a temporary
- if nonnull. */
+ frame-related if FRAME_RELATED_P. FORCE_ISA_MODE is as for
+ aarch64_add_offset. TEMP1 is available as a temporary if nonnull. */
static inline void
-aarch64_sub_sp (rtx temp1, rtx temp2, poly_int64 delta, bool frame_related_p,
- bool emit_move_imm = true)
+aarch64_sub_sp (rtx temp1, rtx temp2, poly_int64 delta,
+ aarch64_feature_flags force_isa_mode,
+ bool frame_related_p, bool emit_move_imm = true)
{
aarch64_add_offset (Pmode, stack_pointer_rtx, stack_pointer_rtx, -delta,
- temp1, temp2, frame_related_p, emit_move_imm);
+ temp1, temp2, force_isa_mode, frame_related_p,
+ emit_move_imm);
}
/* A streaming-compatible function needs to switch temporarily to the known
@@ -7673,11 +7738,11 @@ aarch64_expand_mov_immediate (rtx dest, rtx imm)
{
base = aarch64_force_temporary (int_mode, dest, base);
aarch64_add_offset (int_mode, dest, base, offset,
- NULL_RTX, NULL_RTX, false);
+ NULL_RTX, NULL_RTX, 0, false);
}
else
aarch64_add_offset (int_mode, dest, base, offset,
- dest, NULL_RTX, false);
+ dest, NULL_RTX, 0, false);
}
return;
}
@@ -7704,7 +7769,7 @@ aarch64_expand_mov_immediate (rtx dest, rtx imm)
gcc_assert (can_create_pseudo_p ());
base = aarch64_force_temporary (int_mode, dest, base);
aarch64_add_offset (int_mode, dest, base, const_offset,
- NULL_RTX, NULL_RTX, false);
+ NULL_RTX, NULL_RTX, 0, false);
return;
}
@@ -7744,7 +7809,7 @@ aarch64_expand_mov_immediate (rtx dest, rtx imm)
gcc_assert(can_create_pseudo_p ());
base = aarch64_force_temporary (int_mode, dest, base);
aarch64_add_offset (int_mode, dest, base, const_offset,
- NULL_RTX, NULL_RTX, false);
+ NULL_RTX, NULL_RTX, 0, false);
return;
}
/* FALLTHRU */
@@ -9212,6 +9277,9 @@ aarch64_need_old_pstate_sm ()
if (aarch64_cfun_incoming_sm_state () != 0)
return false;
+ if (aarch64_cfun_enables_pstate_sm ())
+ return true;
+
if (cfun->machine->call_switches_sm_state)
for (auto insn = get_insns (); insn; insn = NEXT_INSN (insn))
if (auto *call = dyn_cast<rtx_call_insn *> (insn))
@@ -9238,6 +9306,7 @@ aarch64_layout_frame (void)
poly_int64 vector_save_size = GET_MODE_SIZE (vector_save_mode);
bool frame_related_fp_reg_p = false;
aarch64_frame &frame = cfun->machine->frame;
+ bool enables_pstate_sm = aarch64_cfun_enables_pstate_sm ();
frame.emit_frame_chain = aarch64_needs_frame_chain ();
@@ -9277,7 +9346,7 @@ aarch64_layout_frame (void)
frame.reg_offset[regno] = SLOT_REQUIRED;
for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++)
- if (df_regs_ever_live_p (regno)
+ if ((enables_pstate_sm || df_regs_ever_live_p (regno))
&& !fixed_regs[regno]
&& !crtl->abi->clobbers_full_reg_p (regno))
{
@@ -9306,7 +9375,7 @@ aarch64_layout_frame (void)
}
for (regno = P0_REGNUM; regno <= P15_REGNUM; regno++)
- if (df_regs_ever_live_p (regno)
+ if ((enables_pstate_sm || df_regs_ever_live_p (regno))
&& !fixed_regs[regno]
&& !crtl->abi->clobbers_full_reg_p (regno))
frame.reg_offset[regno] = SLOT_REQUIRED;
@@ -10121,9 +10190,14 @@ aarch64_get_separate_components (void)
bitmap_clear (components);
/* The registers we need saved to the frame. */
+ bool enables_pstate_sm = aarch64_cfun_enables_pstate_sm ();
for (unsigned regno = 0; regno <= LAST_SAVED_REGNUM; regno++)
if (aarch64_register_saved_on_entry (regno))
{
+ if (enables_pstate_sm
+ && (FP_REGNUM_P (regno) || PR_REGNUM_P (regno)))
+ continue;
+
/* Punt on saves and restores that use ST1D and LD1D. We could
try to be smarter, but it would involve making sure that the
spare predicate register itself is safe to use at the save
@@ -10438,6 +10512,7 @@ aarch64_stack_clash_protection_alloca_probe_range (void)
static void
aarch64_allocate_and_probe_stack_space (rtx temp1, rtx temp2,
poly_int64 poly_size,
+ aarch64_feature_flags force_isa_mode,
bool frame_related_p,
bool final_adjustment_p)
{
@@ -10498,7 +10573,8 @@ aarch64_allocate_and_probe_stack_space (rtx temp1, rtx
temp2,
if (known_lt (poly_size, min_probe_threshold)
|| !flag_stack_clash_protection)
{
- aarch64_sub_sp (temp1, temp2, poly_size, frame_related_p);
+ aarch64_sub_sp (temp1, temp2, poly_size, force_isa_mode,
+ frame_related_p);
return;
}
@@ -10515,7 +10591,8 @@ aarch64_allocate_and_probe_stack_space (rtx temp1, rtx
temp2,
/* First calculate the amount of bytes we're actually spilling. */
aarch64_add_offset (Pmode, temp1, CONST0_RTX (Pmode),
- poly_size, temp1, temp2, false, true);
+ poly_size, temp1, temp2, force_isa_mode,
+ false, true);
rtx_insn *insn = get_last_insn ();
@@ -10573,7 +10650,7 @@ aarch64_allocate_and_probe_stack_space (rtx temp1, rtx
temp2,
{
for (HOST_WIDE_INT i = 0; i < rounded_size; i += guard_size)
{
- aarch64_sub_sp (NULL, temp2, guard_size, true);
+ aarch64_sub_sp (NULL, temp2, guard_size, force_isa_mode, true);
emit_stack_probe (plus_constant (Pmode, stack_pointer_rtx,
guard_used_by_caller));
emit_insn (gen_blockage ());
@@ -10584,7 +10661,7 @@ aarch64_allocate_and_probe_stack_space (rtx temp1, rtx
temp2,
{
/* Compute the ending address. */
aarch64_add_offset (Pmode, temp1, stack_pointer_rtx, -rounded_size,
- temp1, NULL, false, true);
+ temp1, NULL, force_isa_mode, false, true);
rtx_insn *insn = get_last_insn ();
/* For the initial allocation, we don't have a frame pointer
@@ -10654,7 +10731,7 @@ aarch64_allocate_and_probe_stack_space (rtx temp1, rtx
temp2,
else if (final_adjustment_p && rounded_size == 0)
residual_probe_offset = 0;
- aarch64_sub_sp (temp1, temp2, residual, frame_related_p);
+ aarch64_sub_sp (temp1, temp2, residual, force_isa_mode, frame_related_p);
if (residual >= min_probe_threshold)
{
if (dump_file)
@@ -10670,6 +10747,14 @@ aarch64_allocate_and_probe_stack_space (rtx temp1, rtx
temp2,
}
}
+/* Implement TARGET_USE_LATE_PROLOGUE_EPILOGUE. */
+
+static bool
+aarch64_use_late_prologue_epilogue ()
+{
+ return aarch64_cfun_enables_pstate_sm ();
+}
+
/* Return 1 if the register is used by the epilogue. We need to say the
return register is used, but only after epilogue generation is complete.
Note that in the case of sibcalls, the values "used by the epilogue" are
@@ -10826,6 +10911,9 @@ aarch64_expand_prologue (void)
unsigned reg2 = cfun->machine->frame.wb_push_candidate2;
bool emit_frame_chain = cfun->machine->frame.emit_frame_chain;
rtx_insn *insn;
+ aarch64_feature_flags force_isa_mode = 0;
+ if (aarch64_cfun_enables_pstate_sm ())
+ force_isa_mode = AARCH64_FL_SM_ON;
if (flag_stack_clash_protection && known_eq (callee_adjust, 0))
{
@@ -10887,7 +10975,7 @@ aarch64_expand_prologue (void)
less the amount of the guard reserved for use by the caller's
outgoing args. */
aarch64_allocate_and_probe_stack_space (tmp0_rtx, tmp1_rtx, initial_adjust,
- true, false);
+ force_isa_mode, true, false);
if (callee_adjust != 0)
aarch64_push_regs (reg1, reg2, callee_adjust);
@@ -10913,7 +11001,8 @@ aarch64_expand_prologue (void)
gcc_assert (known_eq (chain_offset, 0));
aarch64_add_offset (Pmode, hard_frame_pointer_rtx,
stack_pointer_rtx, chain_offset,
- tmp1_rtx, tmp0_rtx, frame_pointer_needed);
+ tmp1_rtx, tmp0_rtx, force_isa_mode,
+ frame_pointer_needed);
if (frame_pointer_needed && !frame_size.is_constant ())
{
/* Variable-sized frames need to describe the save slot
@@ -10956,6 +11045,7 @@ aarch64_expand_prologue (void)
|| known_eq (initial_adjust, 0));
aarch64_allocate_and_probe_stack_space (tmp1_rtx, tmp0_rtx,
sve_callee_adjust,
+ force_isa_mode,
!frame_pointer_needed, false);
saved_regs_offset += sve_callee_adjust;
}
@@ -10968,10 +11058,13 @@ aarch64_expand_prologue (void)
/* We may need to probe the final adjustment if it is larger than the guard
that is assumed by the called. */
aarch64_allocate_and_probe_stack_space (tmp1_rtx, tmp0_rtx, final_adjust,
+ force_isa_mode,
!frame_pointer_needed, true);
- /* Save the incoming value of PSTATE.SM, if required. */
- if (known_ge (cfun->machine->frame.old_svcr_offset, 0))
+ /* Save the incoming value of PSTATE.SM, if required. Code further
+ down does this for locally-streaming functions. */
+ if (known_ge (cfun->machine->frame.old_svcr_offset, 0)
+ && !aarch64_cfun_enables_pstate_sm ())
{
rtx mem = aarch64_old_svcr_mem ();
MEM_VOLATILE_P (mem) = 1;
@@ -11022,7 +11115,40 @@ aarch64_expand_prologue (void)
emit_insn (gen_aarch64_tpidr2_save ());
emit_insn (gen_aarch64_clear_tpidr2 ());
emit_label (label);
- emit_insn (gen_aarch64_smstart_za ());
+ if (!aarch64_cfun_enables_pstate_sm ()
+ || known_ge (cfun->machine->frame.old_svcr_offset, 0))
+ emit_insn (gen_aarch64_smstart_za ());
+ }
+
+ /* Enable PSTATE.SM, if required. */
+ if (aarch64_cfun_enables_pstate_sm ())
+ {
+ rtx_insn *guard_label = nullptr;
+ if (known_ge (cfun->machine->frame.old_svcr_offset, 0))
+ {
+ /* The current function is streaming-compatible. Save the
+ original state of PSTATE.SM. */
+ rtx svcr = gen_rtx_REG (DImode, IP0_REGNUM);
+ emit_insn (gen_aarch64_read_svcr (svcr));
+ emit_move_insn (aarch64_old_svcr_mem (), svcr);
+ guard_label = aarch64_guard_switch_pstate_sm (svcr,
+ aarch64_isa_flags);
+ }
+ aarch64_sme_mode_switch_regs args_switch;
+ auto &args = crtl->args.info;
+ for (unsigned int i = 0; i < args.num_sme_mode_switch_args; ++i)
+ {
+ rtx x = args.sme_mode_switch_args[i];
+ args_switch.add_reg (GET_MODE (x), REGNO (x));
+ }
+ args_switch.emit_prologue ();
+ if (cfun->machine->frame.has_new_za_state && !guard_label)
+ emit_insn (gen_aarch64_smstart ());
+ else
+ emit_insn (gen_aarch64_smstart_sm ());
+ args_switch.emit_epilogue ();
+ if (guard_label)
+ emit_label (guard_label);
}
}
@@ -11073,6 +11199,9 @@ aarch64_expand_epilogue (rtx_call_insn *sibcall)
HOST_WIDE_INT guard_size
= 1 << param_stack_clash_protection_guard_size;
HOST_WIDE_INT guard_used_by_caller = STACK_CLASH_CALLER_GUARD;
+ aarch64_feature_flags force_isa_mode = 0;
+ if (aarch64_cfun_enables_pstate_sm ())
+ force_isa_mode = AARCH64_FL_SM_ON;
/* We can re-use the registers when:
@@ -11097,7 +11226,33 @@ aarch64_expand_epilogue (rtx_call_insn *sibcall)
= maybe_ne (get_frame_size ()
+ cfun->machine->frame.saved_varargs_size, 0);
- if (cfun->machine->frame.has_new_za_state)
+ /* Reset PSTATE.SM, if required. Fold an unconditional SMSTOP SM
+ and SMSTOP ZA into a single SMSTOP. */
+ bool pending_smstop_za = cfun->machine->frame.has_new_za_state;
+ if (aarch64_cfun_enables_pstate_sm ())
+ {
+ rtx_insn *guard_label = nullptr;
+ if (known_ge (cfun->machine->frame.old_svcr_offset, 0))
+ guard_label = aarch64_guard_switch_pstate_sm (IP0_REGNUM,
+ aarch64_isa_flags);
+ aarch64_sme_mode_switch_regs args_switch;
+ if (crtl->return_rtx && REG_P (crtl->return_rtx))
+ args_switch.add_reg (GET_MODE (crtl->return_rtx),
+ REGNO (crtl->return_rtx));
+ args_switch.emit_prologue ();
+ if (pending_smstop_za && !guard_label)
+ {
+ emit_insn (gen_aarch64_smstop ());
+ pending_smstop_za = false;
+ }
+ else
+ emit_insn (gen_aarch64_smstop_sm ());
+ args_switch.emit_epilogue ();
+ if (guard_label)
+ emit_label (guard_label);
+ }
+
+ if (pending_smstop_za)
/* Turn ZA off before returning. TPIDR2_EL0 is already null at
this point. */
emit_insn (gen_aarch64_smstop_za ());
@@ -11122,12 +11277,13 @@ aarch64_expand_epilogue (rtx_call_insn *sibcall)
aarch64_add_offset (Pmode, stack_pointer_rtx,
hard_frame_pointer_rtx,
-callee_offset - below_hard_fp_saved_regs_size,
- tmp1_rtx, tmp0_rtx, callee_adjust == 0);
+ tmp1_rtx, tmp0_rtx, force_isa_mode,
+ callee_adjust == 0);
else
/* The case where we need to re-use the register here is very rare, so
avoid the complicated condition and just always emit a move if the
immediate doesn't fit. */
- aarch64_add_sp (tmp1_rtx, tmp0_rtx, final_adjust, true);
+ aarch64_add_sp (tmp1_rtx, tmp0_rtx, final_adjust, force_isa_mode, true);
/* Restore the vector registers before the predicate registers,
so that we can use P4 as a temporary for big-endian SVE frames. */
@@ -11136,7 +11292,8 @@ aarch64_expand_epilogue (rtx_call_insn *sibcall)
aarch64_restore_callee_saves (callee_offset, P0_REGNUM, P15_REGNUM,
false, &cfi_ops);
if (maybe_ne (sve_callee_adjust, 0))
- aarch64_add_sp (NULL_RTX, NULL_RTX, sve_callee_adjust, true);
+ aarch64_add_sp (NULL_RTX, NULL_RTX, sve_callee_adjust,
+ force_isa_mode, true);
/* When shadow call stack is enabled, the scs_pop in the epilogue will
restore x30, we don't need to restore x30 again in the traditional
@@ -11167,7 +11324,7 @@ aarch64_expand_epilogue (rtx_call_insn *sibcall)
/* Liveness of EP0_REGNUM can not be trusted across function calls either, so
add restriction on emit_move optimization to leaf functions. */
- aarch64_add_sp (tmp0_rtx, tmp1_rtx, initial_adjust,
+ aarch64_add_sp (tmp0_rtx, tmp1_rtx, initial_adjust, force_isa_mode,
(!can_inherit_p || !crtl->is_leaf
|| df_regs_ever_live_p (EP0_REGNUM)));
@@ -11300,7 +11457,8 @@ aarch64_output_mi_thunk (FILE *file, tree thunk
ATTRIBUTE_UNUSED,
temp1 = gen_rtx_REG (Pmode, EP1_REGNUM);
if (vcall_offset == 0)
- aarch64_add_offset (Pmode, this_rtx, this_rtx, delta, temp1, temp0, false);
+ aarch64_add_offset (Pmode, this_rtx, this_rtx, delta, temp1, temp0,
+ 0, false);
else
{
gcc_assert ((vcall_offset & (POINTER_BYTES - 1)) == 0);
@@ -11313,7 +11471,7 @@ aarch64_output_mi_thunk (FILE *file, tree thunk
ATTRIBUTE_UNUSED,
plus_constant (Pmode, this_rtx, delta));
else
aarch64_add_offset (Pmode, this_rtx, this_rtx, delta,
- temp1, temp0, false);
+ temp1, temp0, 0, false);
}
if (Pmode == ptr_mode)
@@ -29469,6 +29627,9 @@ aarch64_libgcc_floating_mode_supported_p
#undef TARGET_HAVE_SHADOW_CALL_STACK
#define TARGET_HAVE_SHADOW_CALL_STACK true
+#undef TARGET_USE_LATE_PROLOGUE_EPILOGUE
+#define TARGET_USE_LATE_PROLOGUE_EPILOGUE aarch64_use_late_prologue_epilogue
+
#undef TARGET_EMIT_EPILOGUE_FOR_SIBCALL
#define TARGET_EMIT_EPILOGUE_FOR_SIBCALL aarch64_expand_epilogue
diff --git a/gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_1.c
b/gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_1.c
new file mode 100644
index 00000000000..ab9c8cd6bac
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_1.c
@@ -0,0 +1,433 @@
+// { dg-options "-O -fomit-frame-pointer -fno-optimize-sibling-calls" }
+// { dg-final { check-function-bodies "**" "" } }
+
+__attribute__((arm_streaming, arm_shared_za)) void consume_za ();
+
+/*
+** n_ls:
+** stp d8, d9, \[sp, #?-64\]!
+** stp d10, d11, \[sp, #?16\]
+** stp d12, d13, \[sp, #?32\]
+** stp d14, d15, \[sp, #?48\]
+** smstart sm
+** smstop sm
+** ldp d10, d11, \[sp, #?16\]
+** ldp d12, d13, \[sp, #?32\]
+** ldp d14, d15, \[sp, #?48\]
+** ldp d8, d9, \[sp\], #?64
+** ret
+*/
+void __attribute__((arm_locally_streaming))
+n_ls ()
+{
+ asm ("");
+}
+
+/*
+** s_ls:
+** ret
+*/
+void __attribute__((arm_streaming, arm_locally_streaming))
+s_ls ()
+{
+ asm ("");
+}
+
+/*
+** sc_ls:
+** stp x29, x30, \[sp, #?-96\]!
+** mov x29, sp
+** stp d8, d9, \[sp, #?32\]
+** stp d10, d11, \[sp, #?48\]
+** stp d12, d13, \[sp, #?64\]
+** stp d14, d15, \[sp, #?80\]
+** mrs x16, svcr
+** str x16, \[x29, #?16\]
+** tbnz x16, 0, .*
+** smstart sm
+** ldr x16, \[x29, #?16\]
+** tbnz x16, 0, .*
+** smstop sm
+** ldp d8, d9, \[sp, #?32\]
+** ldp d10, d11, \[sp, #?48\]
+** ldp d12, d13, \[sp, #?64\]
+** ldp d14, d15, \[sp, #?80\]
+** ldp x29, x30, \[sp\], #?96
+** ret
+*/
+void __attribute__((arm_streaming_compatible, arm_locally_streaming))
+sc_ls ()
+{
+ asm ("");
+}
+
+/*
+** n_ls_new_za:
+** str x30, \[sp, #?-80\]!
+** stp d8, d9, \[sp, #?16\]
+** stp d10, d11, \[sp, #?32\]
+** stp d12, d13, \[sp, #?48\]
+** stp d14, d15, \[sp, #?64\]
+** mrs x11, tpidr2_el0
+** cbz x11, .*
+** bl __arm_tpidr2_save
+** msr tpidr2_el0, xzr
+** smstart
+** bl consume_za
+** smstop
+** ldp d8, d9, \[sp, #?16\]
+** ldp d10, d11, \[sp, #?32\]
+** ldp d12, d13, \[sp, #?48\]
+** ldp d14, d15, \[sp, #?64\]
+** ldr x30, \[sp\], #?80
+** ret
+*/
+void __attribute__((arm_locally_streaming, arm_new_za))
+n_ls_new_za ()
+{
+ consume_za ();
+}
+
+/*
+** s_ls_new_za:
+** str x30, \[sp, #?-16\]!
+** mrs x11, tpidr2_el0
+** cbz x11, .*
+** bl __arm_tpidr2_save
+** msr tpidr2_el0, xzr
+** smstart za
+** bl consume_za
+** smstop za
+** ldr x30, \[sp\], #?16
+** ret
+*/
+void __attribute__((arm_locally_streaming, arm_streaming, arm_new_za))
+s_ls_new_za ()
+{
+ consume_za ();
+}
+
+/*
+** sc_ls_new_za:
+** stp x29, x30, \[sp, #?-96\]!
+** mov x29, sp
+** stp d8, d9, \[sp, #?32\]
+** stp d10, d11, \[sp, #?48\]
+** stp d12, d13, \[sp, #?64\]
+** stp d14, d15, \[sp, #?80\]
+** mrs x11, tpidr2_el0
+** cbz x11, .*
+** bl __arm_tpidr2_save
+** msr tpidr2_el0, xzr
+** smstart za
+** mrs x16, svcr
+** str x16, \[x29, #?16\]
+** tbnz x16, 0, .*
+** smstart sm
+** bl consume_za
+** ldr x16, \[x29, #?16\]
+** tbnz x16, 0, .*
+** smstop sm
+** smstop za
+** ldp d8, d9, \[sp, #?32\]
+** ldp d10, d11, \[sp, #?48\]
+** ldp d12, d13, \[sp, #?64\]
+** ldp d14, d15, \[sp, #?80\]
+** ldp x29, x30, \[sp\], #?96
+** ret
+*/
+void __attribute__((arm_streaming_compatible, arm_locally_streaming,
arm_new_za))
+sc_ls_new_za ()
+{
+ consume_za ();
+}
+
+/*
+** n_ls_shared_za:
+** str x30, \[sp, #?-80\]!
+** stp d8, d9, \[sp, #?16\]
+** stp d10, d11, \[sp, #?32\]
+** stp d12, d13, \[sp, #?48\]
+** stp d14, d15, \[sp, #?64\]
+** smstart sm
+** bl consume_za
+** smstop sm
+** ldp d8, d9, \[sp, #?16\]
+** ldp d10, d11, \[sp, #?32\]
+** ldp d12, d13, \[sp, #?48\]
+** ldp d14, d15, \[sp, #?64\]
+** ldr x30, \[sp\], #?80
+** ret
+*/
+void __attribute__((arm_locally_streaming, arm_shared_za))
+n_ls_shared_za ()
+{
+ consume_za ();
+}
+
+/*
+** s_ls_shared_za:
+** str x30, \[sp, #?-16\]!
+** bl consume_za
+** ldr x30, \[sp\], #?16
+** ret
+*/
+void __attribute__((arm_streaming, arm_locally_streaming, arm_shared_za))
+s_ls_shared_za ()
+{
+ consume_za ();
+}
+
+/*
+** sc_ls_shared_za:
+** stp x29, x30, \[sp, #?-96\]!
+** mov x29, sp
+** stp d8, d9, \[sp, #?32\]
+** stp d10, d11, \[sp, #?48\]
+** stp d12, d13, \[sp, #?64\]
+** stp d14, d15, \[sp, #?80\]
+** mrs x16, svcr
+** str x16, \[x29, #?16\]
+** tbnz x16, 0, .*
+** smstart sm
+** bl consume_za
+** ldr x16, \[x29, #?16\]
+** tbnz x16, 0, .*
+** smstop sm
+** ldp d8, d9, \[sp, #?32\]
+** ldp d10, d11, \[sp, #?48\]
+** ldp d12, d13, \[sp, #?64\]
+** ldp d14, d15, \[sp, #?80\]
+** ldp x29, x30, \[sp\], #?96
+** ret
+*/
+void __attribute__((arm_streaming_compatible, arm_locally_streaming,
arm_shared_za))
+sc_ls_shared_za ()
+{
+ consume_za ();
+}
+
+/*
+** n_ls_vector_pcs:
+** stp q8, q9, \[sp, #?-256\]!
+** stp q10, q11, \[sp, #?32\]
+** stp q12, q13, \[sp, #?64\]
+** stp q14, q15, \[sp, #?96\]
+** stp q16, q17, \[sp, #?128\]
+** stp q18, q19, \[sp, #?160\]
+** stp q20, q21, \[sp, #?192\]
+** stp q22, q23, \[sp, #?224\]
+** smstart sm
+** smstop sm
+** ldp q10, q11, \[sp, #?32\]
+** ldp q12, q13, \[sp, #?64\]
+** ldp q14, q15, \[sp, #?96\]
+** ldp q16, q17, \[sp, #?128\]
+** ldp q18, q19, \[sp, #?160\]
+** ldp q20, q21, \[sp, #?192\]
+** ldp q22, q23, \[sp, #?224\]
+** ldp q8, q9, \[sp\], #?256
+** ret
+*/
+void __attribute__((arm_locally_streaming, aarch64_vector_pcs))
+n_ls_vector_pcs ()
+{
+ asm ("");
+}
+
+/*
+** n_ls_sve_pcs:
+** addsvl sp, sp, #-18
+** str p4, \[sp\]
+** str p5, \[sp, #1, mul vl\]
+** str p6, \[sp, #2, mul vl\]
+** str p7, \[sp, #3, mul vl\]
+** str p8, \[sp, #4, mul vl\]
+** str p9, \[sp, #5, mul vl\]
+** str p10, \[sp, #6, mul vl\]
+** str p11, \[sp, #7, mul vl\]
+** str p12, \[sp, #8, mul vl\]
+** str p13, \[sp, #9, mul vl\]
+** str p14, \[sp, #10, mul vl\]
+** str p15, \[sp, #11, mul vl\]
+** str z8, \[sp, #2, mul vl\]
+** str z9, \[sp, #3, mul vl\]
+** str z10, \[sp, #4, mul vl\]
+** str z11, \[sp, #5, mul vl\]
+** str z12, \[sp, #6, mul vl\]
+** str z13, \[sp, #7, mul vl\]
+** str z14, \[sp, #8, mul vl\]
+** str z15, \[sp, #9, mul vl\]
+** str z16, \[sp, #10, mul vl\]
+** str z17, \[sp, #11, mul vl\]
+** str z18, \[sp, #12, mul vl\]
+** str z19, \[sp, #13, mul vl\]
+** str z20, \[sp, #14, mul vl\]
+** str z21, \[sp, #15, mul vl\]
+** str z22, \[sp, #16, mul vl\]
+** str z23, \[sp, #17, mul vl\]
+** addvl sp, sp, #-1
+** str p0, \[sp\]
+** smstart sm
+** ldr p0, \[sp\]
+** addvl sp, sp, #1
+** smstop sm
+** ldr z8, \[sp, #2, mul vl\]
+** ldr z9, \[sp, #3, mul vl\]
+** ldr z10, \[sp, #4, mul vl\]
+** ldr z11, \[sp, #5, mul vl\]
+** ldr z12, \[sp, #6, mul vl\]
+** ldr z13, \[sp, #7, mul vl\]
+** ldr z14, \[sp, #8, mul vl\]
+** ldr z15, \[sp, #9, mul vl\]
+** ldr z16, \[sp, #10, mul vl\]
+** ldr z17, \[sp, #11, mul vl\]
+** ldr z18, \[sp, #12, mul vl\]
+** ldr z19, \[sp, #13, mul vl\]
+** ldr z20, \[sp, #14, mul vl\]
+** ldr z21, \[sp, #15, mul vl\]
+** ldr z22, \[sp, #16, mul vl\]
+** ldr z23, \[sp, #17, mul vl\]
+** ldr p4, \[sp\]
+** ldr p5, \[sp, #1, mul vl\]
+** ldr p6, \[sp, #2, mul vl\]
+** ldr p7, \[sp, #3, mul vl\]
+** ldr p8, \[sp, #4, mul vl\]
+** ldr p9, \[sp, #5, mul vl\]
+** ldr p10, \[sp, #6, mul vl\]
+** ldr p11, \[sp, #7, mul vl\]
+** ldr p12, \[sp, #8, mul vl\]
+** ldr p13, \[sp, #9, mul vl\]
+** ldr p14, \[sp, #10, mul vl\]
+** ldr p15, \[sp, #11, mul vl\]
+** addsvl sp, sp, #18
+** ret
+*/
+void __attribute__((arm_locally_streaming))
+n_ls_sve_pcs (__SVBool_t x)
+{
+ asm ("");
+}
+
+/*
+** n_ls_v0:
+** addsvl sp, sp, #-1
+** ...
+** smstart sm
+** add x[0-9]+, .*
+** smstop sm
+** ...
+** addsvl sp, sp, #1
+** ...
+*/
+#define TEST(VN) __SVInt32_t VN; asm ("" :: "r" (&VN));
+void __attribute__((arm_locally_streaming))
+n_ls_v0 ()
+{
+ TEST (v0);
+}
+
+/*
+** n_ls_v32:
+** addsvl sp, sp, #-32
+** ...
+** smstart sm
+** ...
+** smstop sm
+** ...
+** rdsvl (x[0-9]+), #1
+** lsl (x[0-9]+), \1, #?5
+** add sp, sp, \2
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+n_ls_v32 ()
+{
+ TEST (v0);
+ TEST (v1);
+ TEST (v2);
+ TEST (v3);
+ TEST (v4);
+ TEST (v5);
+ TEST (v6);
+ TEST (v7);
+ TEST (v8);
+ TEST (v9);
+ TEST (v10);
+ TEST (v11);
+ TEST (v12);
+ TEST (v13);
+ TEST (v14);
+ TEST (v15);
+ TEST (v16);
+ TEST (v17);
+ TEST (v18);
+ TEST (v19);
+ TEST (v20);
+ TEST (v21);
+ TEST (v22);
+ TEST (v23);
+ TEST (v24);
+ TEST (v25);
+ TEST (v26);
+ TEST (v27);
+ TEST (v28);
+ TEST (v29);
+ TEST (v30);
+ TEST (v31);
+}
+
+/*
+** n_ls_v33:
+** rdsvl (x[0-9]+), #1
+** mov (x[0-9]+), #?33
+** mul (x[0-9]+), (?:\1, \2|\2, \1)
+** sub sp, sp, \3
+** ...
+** smstart sm
+** ...
+** smstop sm
+** ...
+** rdsvl (x[0-9]+), #1
+** mov (x[0-9]+), #?33
+** mul (x[0-9]+), (?:\4, \5|\5, \4)
+** add sp, sp, \6
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+n_ls_v33 ()
+{
+ TEST (v0);
+ TEST (v1);
+ TEST (v2);
+ TEST (v3);
+ TEST (v4);
+ TEST (v5);
+ TEST (v6);
+ TEST (v7);
+ TEST (v8);
+ TEST (v9);
+ TEST (v10);
+ TEST (v11);
+ TEST (v12);
+ TEST (v13);
+ TEST (v14);
+ TEST (v15);
+ TEST (v16);
+ TEST (v17);
+ TEST (v18);
+ TEST (v19);
+ TEST (v20);
+ TEST (v21);
+ TEST (v22);
+ TEST (v23);
+ TEST (v24);
+ TEST (v25);
+ TEST (v26);
+ TEST (v27);
+ TEST (v28);
+ TEST (v29);
+ TEST (v30);
+ TEST (v31);
+ TEST (v32);
+}
diff --git a/gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_2.c
b/gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_2.c
new file mode 100644
index 00000000000..4c9caf5d078
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_2.c
@@ -0,0 +1,177 @@
+// { dg-options "-O -fomit-frame-pointer -fno-optimize-sibling-calls" }
+// { dg-final { check-function-bodies "**" "" } }
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+
+/*
+** test_d0:
+** ...
+** smstart sm
+** .*
+** fmov x10, d0
+** smstop sm
+** fmov d0, x10
+** ...
+*/
+double __attribute__((arm_locally_streaming))
+test_d0 ()
+{
+ asm ("");
+ return 1.0f;
+}
+
+/*
+** test_d0_vec:
+** ...
+** smstart sm
+** .*
+** (
+** fmov x10, d0
+** |
+** umov x10, v0.d\[0\]
+** )
+** smstop sm
+** fmov d0, x10
+** ...
+*/
+int8x8_t __attribute__((arm_locally_streaming))
+test_d0_vec ()
+{
+ asm volatile ("");
+ return (int8x8_t) {};
+}
+
+/*
+** test_q0:
+** ...
+** smstart sm
+** .*
+** str q0, \[sp, #?-16\]!
+** smstop sm
+** ldr q0, \[sp\], #?16
+** ...
+*/
+int8x16_t __attribute__((arm_locally_streaming))
+test_q0 ()
+{
+ asm volatile ("");
+ return (int8x16_t) {};
+}
+
+/*
+** test_q1:
+** ...
+** smstart sm
+** ...
+** stp q0, q1, \[sp, #?-32\]!
+** smstop sm
+** ldp q0, q1, \[sp\], #?32
+** ...
+*/
+int8x16x2_t __attribute__((arm_locally_streaming))
+test_q1 ()
+{
+ asm volatile ("");
+ return (int8x16x2_t) {};
+}
+
+/*
+** test_q2:
+** ...
+** smstart sm
+** ...
+** stp q0, q1, \[sp, #?-48\]!
+** str q2, \[sp, #?32\]
+** smstop sm
+** ldr q2, \[sp, #?32\]
+** ldp q0, q1, \[sp\], #?48
+** ...
+*/
+int8x16x3_t __attribute__((arm_locally_streaming))
+test_q2 ()
+{
+ asm volatile ("");
+ return (int8x16x3_t) {};
+}
+
+/*
+** test_q3:
+** ...
+** smstart sm
+** ...
+** stp q0, q1, \[sp, #?-64\]!
+** stp q2, q3, \[sp, #?32\]
+** smstop sm
+** ldp q2, q3, \[sp, #?32\]
+** ldp q0, q1, \[sp\], #?64
+** ...
+*/
+int8x16x4_t __attribute__((arm_locally_streaming))
+test_q3 ()
+{
+ asm volatile ("");
+ return (int8x16x4_t) {};
+}
+
+/*
+** test_z0:
+** ...
+** smstart sm
+** mov z0\.b, #0
+** addvl sp, sp, #-1
+** str z0, \[sp\]
+** smstop sm
+** ldr z0, \[sp\]
+** addvl sp, sp, #1
+** ...
+*/
+svint8_t __attribute__((arm_locally_streaming))
+test_z0 ()
+{
+ asm volatile ("");
+ return (svint8_t) {};
+}
+
+/*
+** test_z3:
+** ...
+** smstart sm
+** ...
+** addvl sp, sp, #-4
+** str z0, \[sp\]
+** str z1, \[sp, #1, mul vl\]
+** str z2, \[sp, #2, mul vl\]
+** str z3, \[sp, #3, mul vl\]
+** smstop sm
+** ldr z0, \[sp\]
+** ldr z1, \[sp, #1, mul vl\]
+** ldr z2, \[sp, #2, mul vl\]
+** ldr z3, \[sp, #3, mul vl\]
+** ...
+*/
+svint8x4_t __attribute__((arm_locally_streaming))
+test_z3 ()
+{
+ asm volatile ("");
+ return (svint8x4_t) {};
+}
+
+/*
+** test_p0:
+** ...
+** smstart sm
+** pfalse p0\.b
+** addvl sp, sp, #-1
+** str p0, \[sp\]
+** smstop sm
+** ldr p0, \[sp\]
+** addvl sp, sp, #1
+** ...
+*/
+svbool_t __attribute__((arm_locally_streaming))
+test_p0 ()
+{
+ asm volatile ("");
+ return (svbool_t) {};
+}
diff --git a/gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_3.c
b/gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_3.c
new file mode 100644
index 00000000000..e6cbd9d176d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sme/locally_streaming_3.c
@@ -0,0 +1,273 @@
+// { dg-options "-O -fomit-frame-pointer -fno-optimize-sibling-calls" }
+// { dg-final { check-function-bodies "**" "" } }
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+
+/*
+** test_d0:
+** ...
+** fmov x10, d0
+** smstart sm
+** fmov d0, x10
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_d0 (double d0)
+{
+ asm ("");
+}
+
+/*
+** test_d7:
+** ...
+** fmov x10, d0
+** fmov x11, d1
+** fmov x12, d2
+** fmov x13, d3
+** fmov x14, d4
+** fmov x15, d5
+** fmov x16, d6
+** fmov x17, d7
+** smstart sm
+** fmov d0, x10
+** fmov d1, x11
+** fmov d2, x12
+** fmov d3, x13
+** fmov d4, x14
+** fmov d5, x15
+** fmov d6, x16
+** fmov d7, x17
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_d7 (double d0, double d1, double d2, double d3,
+ double d4, double d5, double d6, double d7)
+{
+ asm volatile ("");
+}
+
+/*
+** test_d0_vec:
+** ...
+** (
+** fmov x10, d0
+** |
+** umov x10, v0.d\[0\]
+** )
+** smstart sm
+** fmov d0, x10
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_d0_vec (int8x8_t d0)
+{
+ asm volatile ("");
+}
+
+/*
+** test_d7_vec:
+** ...
+** (
+** fmov x10, d0
+** fmov x11, d1
+** fmov x12, d2
+** fmov x13, d3
+** fmov x14, d4
+** fmov x15, d5
+** fmov x16, d6
+** fmov x17, d7
+** |
+** umov x10, v0.d\[0\]
+** umov x11, v1.d\[0\]
+** umov x12, v2.d\[0\]
+** umov x13, v3.d\[0\]
+** umov x14, v4.d\[0\]
+** umov x15, v5.d\[0\]
+** umov x16, v6.d\[0\]
+** umov x17, v7.d\[0\]
+** )
+** smstart sm
+** fmov d0, x10
+** fmov d1, x11
+** fmov d2, x12
+** fmov d3, x13
+** fmov d4, x14
+** fmov d5, x15
+** fmov d6, x16
+** fmov d7, x17
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_d7_vec (int8x8_t d0, int8x8_t d1, int8x8_t d2, int8x8_t d3,
+ int8x8_t d4, int8x8_t d5, int8x8_t d6, int8x8_t d7)
+{
+ asm volatile ("");
+}
+
+/*
+** test_q0:
+** ...
+** str q0, \[sp, #?-16\]!
+** smstart sm
+** ldr q0, \[sp\], #?16
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_q0 (int8x16_t q0)
+{
+ asm volatile ("");
+}
+
+/*
+** test_q7:
+** ...
+** stp q0, q1, \[sp, #?-128\]!
+** stp q2, q3, \[sp, #?32\]
+** stp q4, q5, \[sp, #?64\]
+** stp q6, q7, \[sp, #?96\]
+** smstart sm
+** ldp q2, q3, \[sp, #?32\]
+** ldp q4, q5, \[sp, #?64\]
+** ldp q6, q7, \[sp, #?96\]
+** ldp q0, q1, \[sp\], #?128
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_q7 (int8x16x4_t q0, int8x16x4_t q4)
+{
+ asm volatile ("");
+}
+
+/*
+** test_z0:
+** ...
+** addvl sp, sp, #-1
+** str z0, \[sp\]
+** smstart sm
+** ldr z0, \[sp\]
+** addvl sp, sp, #1
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_z0 (svint8_t z0)
+{
+ asm volatile ("");
+}
+
+/*
+** test_z7:
+** ...
+** addvl sp, sp, #-8
+** str z0, \[sp\]
+** str z1, \[sp, #1, mul vl\]
+** str z2, \[sp, #2, mul vl\]
+** str z3, \[sp, #3, mul vl\]
+** str z4, \[sp, #4, mul vl\]
+** str z5, \[sp, #5, mul vl\]
+** str z6, \[sp, #6, mul vl\]
+** str z7, \[sp, #7, mul vl\]
+** smstart sm
+** ldr z0, \[sp\]
+** ldr z1, \[sp, #1, mul vl\]
+** ldr z2, \[sp, #2, mul vl\]
+** ldr z3, \[sp, #3, mul vl\]
+** ldr z4, \[sp, #4, mul vl\]
+** ldr z5, \[sp, #5, mul vl\]
+** ldr z6, \[sp, #6, mul vl\]
+** ldr z7, \[sp, #7, mul vl\]
+** addvl sp, sp, #8
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_z7 (svint8x4_t z0, svint8x4_t z4)
+{
+ asm volatile ("");
+}
+
+/*
+** test_p0:
+** ...
+** addvl sp, sp, #-1
+** str p0, \[sp\]
+** smstart sm
+** ldr p0, \[sp\]
+** addvl sp, sp, #1
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_p0 (svbool_t p0)
+{
+ asm volatile ("");
+}
+
+/*
+** test_p3:
+** ...
+** addvl sp, sp, #-1
+** str p0, \[sp\]
+** str p1, \[sp, #1, mul vl\]
+** str p2, \[sp, #2, mul vl\]
+** str p3, \[sp, #3, mul vl\]
+** smstart sm
+** ldr p0, \[sp\]
+** ldr p1, \[sp, #1, mul vl\]
+** ldr p2, \[sp, #2, mul vl\]
+** ldr p3, \[sp, #3, mul vl\]
+** addvl sp, sp, #1
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_p3 (svbool_t p0, svbool_t p1, svbool_t p2, svbool_t p3)
+{
+ asm volatile ("");
+}
+
+/*
+** test_mixed:
+** ...
+** addvl sp, sp, #-3
+** str p0, \[sp\]
+** str p1, \[sp, #1, mul vl\]
+** str p2, \[sp, #2, mul vl\]
+** str p3, \[sp, #3, mul vl\]
+** str z3, \[sp, #1, mul vl\]
+** str z7, \[sp, #2, mul vl\]
+** stp q2, q6, \[sp, #?-32\]!
+** fmov w10, s0
+** fmov x11, d1
+** fmov w12, s4
+** fmov x13, d5
+** smstart sm
+** fmov s0, w10
+** fmov d1, x11
+** fmov s4, w12
+** fmov d5, x13
+** ldp q2, q6, \[sp\], #?32
+** ldr p0, \[sp\]
+** ldr p1, \[sp, #1, mul vl\]
+** ldr p2, \[sp, #2, mul vl\]
+** ldr p3, \[sp, #3, mul vl\]
+** ldr z3, \[sp, #1, mul vl\]
+** ldr z7, \[sp, #2, mul vl\]
+** addvl sp, sp, #3
+** smstop sm
+** ...
+*/
+void __attribute__((arm_locally_streaming))
+test_mixed (float s0, double d1, float32x4_t q2, svfloat32_t z3,
+ float s4, double d5, float64x2_t q6, svfloat64_t z7,
+ svbool_t p0, svbool_t p1, svbool_t p2, svbool_t p3)
+{
+ asm volatile ("");
+}
--
2.25.1
