target/arm: Add TB flag for "MVE insns not predicated"

Our current codegen for MVE always calls out to helper functions,
because some byte lanes might be predicated.  The common case is that
in fact there is no predication active and all lanes should be
updated together, so we can produce better code by detecting that and
using the TCG generic vector infrastructure.

Add a TB flag that is set when we can guarantee that there is no
active MVE predication, and a bool in the DisasContext.  Subsequent
patches will use this flag to generate improved code for some
instructions.

In most cases when the predication state changes we simply end the TB
after that instruction.  For the code called from vfp_access_check()
that handles lazy state preservation and creating a new FP context,
we can usually avoid having to try to end the TB because luckily the
new value of the flag following the register changes in those
sequences doesn't depend on any runtime decisions.  We do have to end
the TB if the guest has enabled lazy FP state preservation but not
automatic state preservation, but this is an odd corner case that is
not going to be common in real-world code.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210913095440.13462-4-peter.maydell@linaro.org
This commit is contained in:
Peter Maydell 2021-09-13 10:54:31 +01:00
parent 85e7d1e9ff
commit 2670221397
7 changed files with 92 additions and 9 deletions

View File

@ -3441,7 +3441,7 @@ typedef ARMCPU ArchCPU;
* | TBFLAG_AM32 | +-----+----------+
* | | |TBFLAG_M32|
* +-------------+----------------+----------+
* 31 23 5 4 0
* 31 23 6 5 0
*
* Unless otherwise noted, these bits are cached in env->hflags.
*/
@ -3499,6 +3499,8 @@ FIELD(TBFLAG_M32, LSPACT, 2, 1) /* Not cached. */
FIELD(TBFLAG_M32, NEW_FP_CTXT_NEEDED, 3, 1) /* Not cached. */
/* Set if FPCCR.S does not match current security state */
FIELD(TBFLAG_M32, FPCCR_S_WRONG, 4, 1) /* Not cached. */
/* Set if MVE insns are definitely not predicated by VPR or LTPSIZE */
FIELD(TBFLAG_M32, MVE_NO_PRED, 5, 1) /* Not cached. */
/*
* Bit usage when in AArch64 state

View File

@ -13637,6 +13637,35 @@ static inline void assert_hflags_rebuild_correctly(CPUARMState *env)
#endif
}
static bool mve_no_pred(CPUARMState *env)
{
/*
* Return true if there is definitely no predication of MVE
* instructions by VPR or LTPSIZE. (Returning false even if there
* isn't any predication is OK; generated code will just be
* a little worse.)
* If the CPU does not implement MVE then this TB flag is always 0.
*
* NOTE: if you change this logic, the "recalculate s->mve_no_pred"
* logic in gen_update_fp_context() needs to be updated to match.
*
* We do not include the effect of the ECI bits here -- they are
* tracked in other TB flags. This simplifies the logic for
* "when did we emit code that changes the MVE_NO_PRED TB flag
* and thus need to end the TB?".
*/
if (cpu_isar_feature(aa32_mve, env_archcpu(env))) {
return false;
}
if (env->v7m.vpr) {
return false;
}
if (env->v7m.ltpsize < 4) {
return false;
}
return true;
}
void cpu_get_tb_cpu_state(CPUARMState *env, target_ulong *pc,
target_ulong *cs_base, uint32_t *pflags)
{
@ -13676,6 +13705,10 @@ void cpu_get_tb_cpu_state(CPUARMState *env, target_ulong *pc,
if (env->v7m.fpccr[is_secure] & R_V7M_FPCCR_LSPACT_MASK) {
DP_TBFLAG_M32(flags, LSPACT, 1);
}
if (mve_no_pred(env)) {
DP_TBFLAG_M32(flags, MVE_NO_PRED, 1);
}
} else {
/*
* Note that XSCALE_CPAR shares bits with VECSTRIDE.

View File

@ -95,7 +95,10 @@ static bool trans_VLLDM_VLSTM(DisasContext *s, arg_VLLDM_VLSTM *a)
clear_eci_state(s);
/* End the TB, because we have updated FP control bits */
/*
* End the TB, because we have updated FP control bits,
* and possibly VPR or LTPSIZE.
*/
s->base.is_jmp = DISAS_UPDATE_EXIT;
return true;
}
@ -397,6 +400,7 @@ static bool gen_M_fp_sysreg_write(DisasContext *s, int regno,
store_cpu_field(control, v7m.control[M_REG_S]);
tcg_gen_andi_i32(tmp, tmp, ~FPCR_NZCV_MASK);
gen_helper_vfp_set_fpscr(cpu_env, tmp);
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
tcg_temp_free_i32(tmp);
tcg_temp_free_i32(sfpa);
break;
@ -409,6 +413,7 @@ static bool gen_M_fp_sysreg_write(DisasContext *s, int regno,
}
tmp = loadfn(s, opaque, true);
store_cpu_field(tmp, v7m.vpr);
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
break;
case ARM_VFP_P0:
{
@ -418,6 +423,7 @@ static bool gen_M_fp_sysreg_write(DisasContext *s, int regno,
tcg_gen_deposit_i32(vpr, vpr, tmp,
R_V7M_VPR_P0_SHIFT, R_V7M_VPR_P0_LENGTH);
store_cpu_field(vpr, v7m.vpr);
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
tcg_temp_free_i32(tmp);
break;
}

View File

@ -810,7 +810,12 @@ DO_LOGIC(VORR, gen_helper_mve_vorr)
DO_LOGIC(VORN, gen_helper_mve_vorn)
DO_LOGIC(VEOR, gen_helper_mve_veor)
DO_LOGIC(VPSEL, gen_helper_mve_vpsel)
static bool trans_VPSEL(DisasContext *s, arg_2op *a)
{
/* This insn updates predication bits */
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
return do_2op(s, a, gen_helper_mve_vpsel);
}
#define DO_2OP(INSN, FN) \
static bool trans_##INSN(DisasContext *s, arg_2op *a) \
@ -1366,6 +1371,8 @@ static bool trans_VPNOT(DisasContext *s, arg_VPNOT *a)
}
gen_helper_mve_vpnot(cpu_env);
/* This insn updates predication bits */
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
mve_update_eci(s);
return true;
}
@ -1852,6 +1859,8 @@ static bool do_vcmp(DisasContext *s, arg_vcmp *a, MVEGenCmpFn *fn)
/* VPT */
gen_vpst(s, a->mask);
}
/* This insn updates predication bits */
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
mve_update_eci(s);
return true;
}
@ -1883,6 +1892,8 @@ static bool do_vcmp_scalar(DisasContext *s, arg_vcmp_scalar *a,
/* VPT */
gen_vpst(s, a->mask);
}
/* This insn updates predication bits */
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
mve_update_eci(s);
return true;
}

View File

@ -109,7 +109,7 @@ static inline long vfp_f16_offset(unsigned reg, bool top)
* Generate code for M-profile lazy FP state preservation if needed;
* this corresponds to the pseudocode PreserveFPState() function.
*/
static void gen_preserve_fp_state(DisasContext *s)
static void gen_preserve_fp_state(DisasContext *s, bool skip_context_update)
{
if (s->v7m_lspact) {
/*
@ -128,6 +128,20 @@ static void gen_preserve_fp_state(DisasContext *s)
* any further FP insns in this TB.
*/
s->v7m_lspact = false;
/*
* The helper might have zeroed VPR, so we do not know the
* correct value for the MVE_NO_PRED TB flag any more.
* If we're about to create a new fp context then that
* will precisely determine the MVE_NO_PRED value (see
* gen_update_fp_context()). Otherwise, we must:
* - set s->mve_no_pred to false, so this instruction
* is generated to use helper functions
* - end the TB now, without chaining to the next TB
*/
if (skip_context_update || !s->v7m_new_fp_ctxt_needed) {
s->mve_no_pred = false;
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
}
}
}
@ -169,12 +183,19 @@ static void gen_update_fp_context(DisasContext *s)
TCGv_i32 z32 = tcg_const_i32(0);
store_cpu_field(z32, v7m.vpr);
}
/*
* We don't need to arrange to end the TB, because the only
* parts of FPSCR which we cache in the TB flags are the VECLEN
* and VECSTRIDE, and those don't exist for M-profile.
* We just updated the FPSCR and VPR. Some of this state is cached
* in the MVE_NO_PRED TB flag. We want to avoid having to end the
* TB here, which means we need the new value of the MVE_NO_PRED
* flag to be exactly known here and the same for all executions.
* Luckily FPDSCR.LTPSIZE is always constant 4 and the VPR is
* always set to 0, so the new MVE_NO_PRED flag is always 1
* if and only if we have MVE.
*
* (The other FPSCR state cached in TB flags is VECLEN and VECSTRIDE,
* but those do not exist for M-profile, so are not relevant here.)
*/
s->mve_no_pred = dc_isar_feature(aa32_mve, s);
if (s->v8m_secure) {
bits |= R_V7M_CONTROL_SFPA_MASK;
@ -238,7 +259,7 @@ bool vfp_access_check_m(DisasContext *s, bool skip_context_update)
/* Handle M-profile lazy FP state mechanics */
/* Trigger lazy-state preservation if necessary */
gen_preserve_fp_state(s);
gen_preserve_fp_state(s, skip_context_update);
if (!skip_context_update) {
/* Update ownership of FP context and create new FP context if needed */

View File

@ -8496,6 +8496,7 @@ static bool trans_DLS(DisasContext *s, arg_DLS *a)
/* DLSTP: set FPSCR.LTPSIZE */
tmp = tcg_const_i32(a->size);
store_cpu_field(tmp, v7m.ltpsize);
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
}
return true;
}
@ -8561,6 +8562,10 @@ static bool trans_WLS(DisasContext *s, arg_WLS *a)
assert(ok);
tmp = tcg_const_i32(a->size);
store_cpu_field(tmp, v7m.ltpsize);
/*
* LTPSIZE updated, but MVE_NO_PRED will always be the same thing (0)
* when we take this upcoming exit from this TB, so gen_jmp_tb() is OK.
*/
}
gen_jmp_tb(s, s->base.pc_next, 1);
@ -8743,6 +8748,8 @@ static bool trans_VCTP(DisasContext *s, arg_VCTP *a)
gen_helper_mve_vctp(cpu_env, masklen);
tcg_temp_free_i32(masklen);
tcg_temp_free_i32(rn_shifted);
/* This insn updates predication bits */
s->base.is_jmp = DISAS_UPDATE_NOCHAIN;
mve_update_eci(s);
return true;
}
@ -9413,6 +9420,7 @@ static void arm_tr_init_disas_context(DisasContextBase *dcbase, CPUState *cs)
dc->v7m_new_fp_ctxt_needed =
EX_TBFLAG_M32(tb_flags, NEW_FP_CTXT_NEEDED);
dc->v7m_lspact = EX_TBFLAG_M32(tb_flags, LSPACT);
dc->mve_no_pred = EX_TBFLAG_M32(tb_flags, MVE_NO_PRED);
} else {
dc->debug_target_el = EX_TBFLAG_ANY(tb_flags, DEBUG_TARGET_EL);
dc->sctlr_b = EX_TBFLAG_A32(tb_flags, SCTLR__B);

View File

@ -100,6 +100,8 @@ typedef struct DisasContext {
bool align_mem;
/* True if PSTATE.IL is set */
bool pstate_il;
/* True if MVE insns are definitely not predicated by VPR or LTPSIZE */
bool mve_no_pred;
/*
* >= 0, a copy of PSTATE.BTYPE, which will be 0 without v8.5-BTI.
* < 0, set by the current instruction.