Skip to content

Commit

Permalink
target/arm: Make VMOV scalar <-> gpreg beatwise for MVE
Browse files Browse the repository at this point in the history
In a CPU with MVE, the VMOV (vector lane to general-purpose register)
and VMOV (general-purpose register to vector lane) insns are not
predicated, but they are subject to beatwise execution if they
are not in an IT block.

Since our implementation always executes all 4 beats in one tick,
this means only that we need to handle PSR.ECI:
 * we must do the usual check for bad ECI state
 * we must advance ECI state if the insn succeeds
 * if ECI says we should not be executing the beat corresponding
   to the lane of the vector register being accessed then we
   should skip performing the move

Note that if PSR.ECI is non-zero then we cannot be in an IT block.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210617121628.20116-45-peter.maydell@linaro.org
  • Loading branch information
pm215 committed Jun 21, 2021
1 parent 3e6d6e8 commit d93a468
Show file tree
Hide file tree
Showing 3 changed files with 75 additions and 8 deletions.
2 changes: 2 additions & 0 deletions target/arm/translate-a32.h
Expand Up @@ -47,6 +47,8 @@ long neon_full_reg_offset(unsigned reg);
long neon_element_offset(int reg, int element, MemOp memop);
void gen_rev16(TCGv_i32 dest, TCGv_i32 var);
void clear_eci_state(DisasContext *s);
bool mve_eci_check(DisasContext *s);
void mve_update_and_store_eci(DisasContext *s);

static inline TCGv_i32 load_cpu_offset(int offset)
{
Expand Down
4 changes: 2 additions & 2 deletions target/arm/translate-mve.c
Expand Up @@ -57,7 +57,7 @@ static bool mve_check_qreg_bank(DisasContext *s, int qmask)
return qmask < 8;
}

static bool mve_eci_check(DisasContext *s)
bool mve_eci_check(DisasContext *s)
{
/*
* This is a beatwise insn: check that ECI is valid (not a
Expand Down Expand Up @@ -91,7 +91,7 @@ static void mve_update_eci(DisasContext *s)
}
}

static void mve_update_and_store_eci(DisasContext *s)
void mve_update_and_store_eci(DisasContext *s)
{
/*
* For insns which don't call a helper function that will call
Expand Down
77 changes: 71 additions & 6 deletions target/arm/translate-vfp.c
Expand Up @@ -581,6 +581,48 @@ static bool trans_VCVT(DisasContext *s, arg_VCVT *a)
return true;
}

static bool mve_skip_vmov(DisasContext *s, int vn, int index, int size)
{
/*
* In a CPU with MVE, the VMOV (vector lane to general-purpose register)
* and VMOV (general-purpose register to vector lane) insns are not
* predicated, but they are subject to beatwise execution if they are
* not in an IT block.
*
* Since our implementation always executes all 4 beats in one tick,
* this means only that if PSR.ECI says we should not be executing
* the beat corresponding to the lane of the vector register being
* accessed then we should skip performing the move, and that we need
* to do the usual check for bad ECI state and advance of ECI state.
*
* Note that if PSR.ECI is non-zero then we cannot be in an IT block.
*
* Return true if this VMOV scalar <-> gpreg should be skipped because
* the MVE PSR.ECI state says we skip the beat where the store happens.
*/

/* Calculate the byte offset into Qn which we're going to access */
int ofs = (index << size) + ((vn & 1) * 8);

if (!dc_isar_feature(aa32_mve, s)) {
return false;
}

switch (s->eci) {
case ECI_NONE:
return false;
case ECI_A0:
return ofs < 4;
case ECI_A0A1:
return ofs < 8;
case ECI_A0A1A2:
case ECI_A0A1A2B0:
return ofs < 12;
default:
g_assert_not_reached();
}
}

static bool trans_VMOV_to_gp(DisasContext *s, arg_VMOV_to_gp *a)
{
/* VMOV scalar to general purpose register */
Expand All @@ -603,14 +645,26 @@ static bool trans_VMOV_to_gp(DisasContext *s, arg_VMOV_to_gp *a)
return false;
}

if (dc_isar_feature(aa32_mve, s)) {
if (!mve_eci_check(s)) {
return true;
}
}

if (!vfp_access_check(s)) {
return true;
}

tmp = tcg_temp_new_i32();
read_neon_element32(tmp, a->vn, a->index, a->size | (a->u ? 0 : MO_SIGN));
store_reg(s, a->rt, tmp);
if (!mve_skip_vmov(s, a->vn, a->index, a->size)) {
tmp = tcg_temp_new_i32();
read_neon_element32(tmp, a->vn, a->index,
a->size | (a->u ? 0 : MO_SIGN));
store_reg(s, a->rt, tmp);
}

if (dc_isar_feature(aa32_mve, s)) {
mve_update_and_store_eci(s);
}
return true;
}

Expand All @@ -636,14 +690,25 @@ static bool trans_VMOV_from_gp(DisasContext *s, arg_VMOV_from_gp *a)
return false;
}

if (dc_isar_feature(aa32_mve, s)) {
if (!mve_eci_check(s)) {
return true;
}
}

if (!vfp_access_check(s)) {
return true;
}

tmp = load_reg(s, a->rt);
write_neon_element32(tmp, a->vn, a->index, a->size);
tcg_temp_free_i32(tmp);
if (!mve_skip_vmov(s, a->vn, a->index, a->size)) {
tmp = load_reg(s, a->rt);
write_neon_element32(tmp, a->vn, a->index, a->size);
tcg_temp_free_i32(tmp);
}

if (dc_isar_feature(aa32_mve, s)) {
mve_update_and_store_eci(s);
}
return true;
}

Expand Down

0 comments on commit d93a468

Please sign in to comment.