| Message ID | 20220222143646.1268606-39-matheus.ferst@eldorado.org.br |
|---|---|
| State | New |
| Headers | show |
| Series | target/ppc: PowerISA Vector/VSX instruction batch | expand |
On 2/22/22 04:36, matheus.ferst@eldorado.org.br wrote: > From: Víctor Colombo <victor.colombo@eldorado.org.br> > > Refactor VSX_SCALAR_CMP_DP, changing its name to VSX_SCALAR_CMP and > prepare the helper to be used for quadword comparisons. > > Signed-off-by: Víctor Colombo <victor.colombo@eldorado.org.br> > Signed-off-by: Matheus Ferst <matheus.ferst@eldorado.org.br> > --- > target/ppc/fpu_helper.c | 31 ++++++++++++++----------------- > 1 file changed, 14 insertions(+), 17 deletions(-) > > diff --git a/target/ppc/fpu_helper.c b/target/ppc/fpu_helper.c > index 9b034d1fe4..5ebbcfe3b7 100644 > --- a/target/ppc/fpu_helper.c > +++ b/target/ppc/fpu_helper.c > @@ -2265,28 +2265,30 @@ VSX_MADDQ(XSNMSUBQP, NMSUB_FLGS, 0) > VSX_MADDQ(XSNMSUBQPO, NMSUB_FLGS, 0) > > /* > - * VSX_SCALAR_CMP_DP - VSX scalar floating point compare double precision > + * VSX_SCALAR_CMP - VSX scalar floating point compare > * op - instruction mnemonic > + * tp - type > * cmp - comparison operation > * exp - expected result of comparison > + * fld - vsr_t field > * svxvc - set VXVC bit > */ > -#define VSX_SCALAR_CMP_DP(op, cmp, exp, svxvc) \ > +#define VSX_SCALAR_CMP(op, tp, cmp, fld, exp, svxvc) \ > void helper_##op(CPUPPCState *env, ppc_vsr_t *xt, \ > ppc_vsr_t *xa, ppc_vsr_t *xb) \ > { \ > - ppc_vsr_t t = *xt; \ > + ppc_vsr_t t = { }; \ > bool vxsnan_flag = false, vxvc_flag = false, vex_flag = false; \ > \ > - if (float64_is_signaling_nan(xa->VsrD(0), &env->fp_status) || \ > - float64_is_signaling_nan(xb->VsrD(0), &env->fp_status)) { \ > + if (tp##_is_signaling_nan(xa->fld, &env->fp_status) || \ > + tp##_is_signaling_nan(xb->fld, &env->fp_status)) { \ > vxsnan_flag = true; \ > if (fpscr_ve == 0 && svxvc) { \ > vxvc_flag = true; \ > } \ > } else if (svxvc) { \ > - vxvc_flag = float64_is_quiet_nan(xa->VsrD(0), &env->fp_status) || \ > - float64_is_quiet_nan(xb->VsrD(0), &env->fp_status); \ > + vxvc_flag = tp##_is_quiet_nan(xa->fld, &env->fp_status) || \ > + tp##_is_quiet_nan(xb->fld, &env->fp_status); \ > } Note that this can be simplified further, using the full FloatRelation result and float_flag_invalid_snan. Note that do_scalar_cmp gets half-way there, only checking for NaNs once we have float_relation_unordered as a comparision result. But it could go further and check float_flag_invalid_snan and drop all of the other checks vs snan and qnan. r~
On 22/02/2022 21:20, Richard Henderson wrote:> On 2/22/22 04:36, matheus.ferst@eldorado.org.br wrote: >> From: Víctor Colombo <victor.colombo@eldorado.org.br> >> >> Refactor VSX_SCALAR_CMP_DP, changing its name to VSX_SCALAR_CMP and >> prepare the helper to be used for quadword comparisons. >> >> Signed-off-by: Víctor Colombo <victor.colombo@eldorado.org.br> >> Signed-off-by: Matheus Ferst <matheus.ferst@eldorado.org.br> >> --- >> target/ppc/fpu_helper.c | 31 ++++++++++++++----------------- >> 1 file changed, 14 insertions(+), 17 deletions(-) >> >> diff --git a/target/ppc/fpu_helper.c b/target/ppc/fpu_helper.c >> index 9b034d1fe4..5ebbcfe3b7 100644 >> --- a/target/ppc/fpu_helper.c >> +++ b/target/ppc/fpu_helper.c >> @@ -2265,28 +2265,30 @@ VSX_MADDQ(XSNMSUBQP, NMSUB_FLGS, 0) >> VSX_MADDQ(XSNMSUBQPO, NMSUB_FLGS, 0) >> >> /* >> - * VSX_SCALAR_CMP_DP - VSX scalar floating point compare double >> precision >> + * VSX_SCALAR_CMP - VSX scalar floating point compare >> * op - instruction mnemonic >> + * tp - type >> * cmp - comparison operation >> * exp - expected result of comparison >> + * fld - vsr_t field >> * svxvc - set VXVC bit >> */ >> -#define VSX_SCALAR_CMP_DP(op, cmp, exp, >> svxvc) \ >> +#define VSX_SCALAR_CMP(op, tp, cmp, fld, exp, >> svxvc) \ >> void helper_##op(CPUPPCState *env, ppc_vsr_t >> *xt, \ >> ppc_vsr_t *xa, ppc_vsr_t >> *xb) \ >> >> { >> \ >> - ppc_vsr_t t = >> *xt; \ >> + ppc_vsr_t t = { >> }; \ >> bool vxsnan_flag = false, vxvc_flag = false, vex_flag = >> false; \ >> >> \ >> - if (float64_is_signaling_nan(xa->VsrD(0), &env->fp_status) >> || \ >> - float64_is_signaling_nan(xb->VsrD(0), &env->fp_status)) >> { \ >> + if (tp##_is_signaling_nan(xa->fld, &env->fp_status) >> || \ >> + tp##_is_signaling_nan(xb->fld, &env->fp_status)) >> { \ >> vxsnan_flag = >> true; \ >> if (fpscr_ve == 0 && svxvc) >> { \ >> vxvc_flag = >> true; \ >> >> } \ >> } else if (svxvc) >> { \ >> - vxvc_flag = float64_is_quiet_nan(xa->VsrD(0), >> &env->fp_status) || \ >> - float64_is_quiet_nan(xb->VsrD(0), >> &env->fp_status); \ >> + vxvc_flag = tp##_is_quiet_nan(xa->fld, &env->fp_status) >> || \ >> + tp##_is_quiet_nan(xb->fld, >> &env->fp_status); \ >> } > > Note that this can be simplified further, using the full FloatRelation > result and > float_flag_invalid_snan. > > Note that do_scalar_cmp gets half-way there, only checking for NaNs once > we have > float_relation_unordered as a comparision result. But it could go > further and check > float_flag_invalid_snan and drop all of the other checks vs snan and qnan. > > > r~ Hello Richard! Thanks for your review Could you please elaborate more on how do you think using float*_compare and its FloatRelation result would work here? I noticed do_scalar_cmp modifies CR and sets FPCC flag, which is not what VSX_SCALAR_CMP do. Using that function would require a rework. An option I though would be to bring into VSX_SCALAR_CMP the important necessary parts, something like this: #define VSX_SCALAR_CMP(op, tp, cmp, fld, svxvc, expr) ... r = tp##_compare(xa->fld, xb->fld, &env->fp_status); \ if (expr) { \ memset(&t.fld, 0xFF, sizeof(t.fld)); \ } else if (r == float_relation_unordered) { \ if (env->fp_status.float_exception_flags & float_flag_invalid_snan) { \ float_invalid_op_vxsnan(env, GETPC()); \ if (fpscr_ve == 0 && svxvc) { \ float_invalid_op_vxvc(env, 0, GETPC()); \ } \ } else if (svxvc) { \ if (tp##_is_quiet_nan(xa->fld, &env->fp_status) || \ tp##_is_quiet_nan(xb->fld, &env->fp_status)) { \ float_invalid_op_vxvc(env, 0, GETPC()); \ } \ } \ } \ ... VSX_SCALAR_CMP(XSCMPEQDP, float64, eq, VsrD(0), 0, r == float_relation_equal) VSX_SCALAR_CMP(XSCMPGEDP, float64, le, VsrD(0), 1, \ r == float_relation_equal || r == float_relation_greater) VSX_SCALAR_CMP(XSCMPGTDP, float64, lt, VsrD(0), 1, r == float_relation_greater) But this still looks convoluted. Another option I came with would be: ppc_vsr_t t = { }; \ \ helper_reset_fpstatus(env); \ \ if (tp##_##cmp##_quiet(xb->fld, xa->fld, &env->fp_status)) { \ memset(&t.fld, 0xFF, sizeof(t.fld)); \ } \ \ if (env->fp_status.float_exception_flags & float_flag_invalid_snan) { \ float_invalid_op_vxsnan(env, GETPC()); \ if (fpscr_ve == 0 && svxvc) { \ float_invalid_op_vxvc(env, 0, GETPC()); \ } \ } else if (svxvc) { \ if (tp##_is_quiet_nan(xa->fld, &env->fp_status) || \ tp##_is_quiet_nan(xb->fld, &env->fp_status)) { \ float_invalid_op_vxvc(env, 0, GETPC()); \ } \ } \ Is this close to what you were thinking? Thank you very much! -- Víctor
On 2/24/22 09:16, Víctor Colombo wrote: > Could you please elaborate more on how do you think using > float*_compare and its FloatRelation result would work here? > I noticed do_scalar_cmp modifies CR and sets FPCC flag, which > is not what VSX_SCALAR_CMP do. Using that function would require a > rework. > > An option I though would be to bring into VSX_SCALAR_CMP the > important necessary parts, something like this: > > #define VSX_SCALAR_CMP(op, tp, cmp, fld, svxvc, expr) ... > r = tp##_compare(xa->fld, xb->fld, &env->fp_status); \ > if (expr) { \ > memset(&t.fld, 0xFF, sizeof(t.fld)); \ > } else if (r == float_relation_unordered) { \ > if (env->fp_status.float_exception_flags & float_flag_invalid_snan) { \ > float_invalid_op_vxsnan(env, GETPC()); \ > if (fpscr_ve == 0 && svxvc) { \ > float_invalid_op_vxvc(env, 0, GETPC()); \ > } \ > } else if (svxvc) { \ > if (tp##_is_quiet_nan(xa->fld, &env->fp_status) || \ > tp##_is_quiet_nan(xb->fld, &env->fp_status)) { \ > float_invalid_op_vxvc(env, 0, GETPC()); \ > } \ > } \ > } \ > ... > VSX_SCALAR_CMP(XSCMPEQDP, float64, eq, VsrD(0), 0, r == float_relation_equal) > VSX_SCALAR_CMP(XSCMPGEDP, float64, le, VsrD(0), 1, \ > r == float_relation_equal || r == float_relation_greater) > VSX_SCALAR_CMP(XSCMPGTDP, float64, lt, VsrD(0), 1, r == float_relation_greater) I was thinking along the lines of: bool r; int flags; helper_reset_fpstatus(env); if (svxvc) { r = tp##cmp(...); } else { r = tp##cmp##_quiet(...); } flags = get_float_exception_flags(&env->fp_status); if (unlikely(flags & float_flag_invalid)) { bool vxvc = svxvc; if (flags & float_flag_invalid_snan)) { float_invalid_op_vxsnan(...); vxvc &= fpscr_ve == 0; } if (vxvc) { float_invalid_op_vxvc(...); } } memset(xt, 0, sizeof(*xt)); memset(&xt->fld, -r, sizeof(xt->fld)); do_float_check_status(...); r~
diff --git a/target/ppc/fpu_helper.c b/target/ppc/fpu_helper.c index 9b034d1fe4..5ebbcfe3b7 100644 --- a/target/ppc/fpu_helper.c +++ b/target/ppc/fpu_helper.c @@ -2265,28 +2265,30 @@ VSX_MADDQ(XSNMSUBQP, NMSUB_FLGS, 0) VSX_MADDQ(XSNMSUBQPO, NMSUB_FLGS, 0) /* - * VSX_SCALAR_CMP_DP - VSX scalar floating point compare double precision + * VSX_SCALAR_CMP - VSX scalar floating point compare * op - instruction mnemonic + * tp - type * cmp - comparison operation * exp - expected result of comparison + * fld - vsr_t field * svxvc - set VXVC bit */ -#define VSX_SCALAR_CMP_DP(op, cmp, exp, svxvc) \ +#define VSX_SCALAR_CMP(op, tp, cmp, fld, exp, svxvc) \ void helper_##op(CPUPPCState *env, ppc_vsr_t *xt, \ ppc_vsr_t *xa, ppc_vsr_t *xb) \ { \ - ppc_vsr_t t = *xt; \ + ppc_vsr_t t = { }; \ bool vxsnan_flag = false, vxvc_flag = false, vex_flag = false; \ \ - if (float64_is_signaling_nan(xa->VsrD(0), &env->fp_status) || \ - float64_is_signaling_nan(xb->VsrD(0), &env->fp_status)) { \ + if (tp##_is_signaling_nan(xa->fld, &env->fp_status) || \ + tp##_is_signaling_nan(xb->fld, &env->fp_status)) { \ vxsnan_flag = true; \ if (fpscr_ve == 0 && svxvc) { \ vxvc_flag = true; \ } \ } else if (svxvc) { \ - vxvc_flag = float64_is_quiet_nan(xa->VsrD(0), &env->fp_status) || \ - float64_is_quiet_nan(xb->VsrD(0), &env->fp_status); \ + vxvc_flag = tp##_is_quiet_nan(xa->fld, &env->fp_status) || \ + tp##_is_quiet_nan(xb->fld, &env->fp_status); \ } \ if (vxsnan_flag) { \ float_invalid_op_vxsnan(env, GETPC()); \ @@ -2297,22 +2299,17 @@ void helper_##op(CPUPPCState *env, ppc_vsr_t *xt, \ vex_flag = fpscr_ve && (vxvc_flag || vxsnan_flag); \ \ if (!vex_flag) { \ - if (float64_##cmp(xb->VsrD(0), xa->VsrD(0), \ - &env->fp_status) == exp) { \ - t.VsrD(0) = -1; \ - t.VsrD(1) = 0; \ - } else { \ - t.VsrD(0) = 0; \ - t.VsrD(1) = 0; \ + if (tp##_##cmp(xb->fld, xa->fld, &env->fp_status) == exp) { \ + memset(&t.fld, 0xFF, sizeof(t.fld)); \ } \ } \ *xt = t; \ do_float_check_status(env, GETPC()); \ } -VSX_SCALAR_CMP_DP(xscmpeqdp, eq, 1, 0) -VSX_SCALAR_CMP_DP(xscmpgedp, le, 1, 1) -VSX_SCALAR_CMP_DP(xscmpgtdp, lt, 1, 1) +VSX_SCALAR_CMP(xscmpeqdp, float64, eq, VsrD(0), 1, 0) +VSX_SCALAR_CMP(xscmpgedp, float64, le, VsrD(0), 1, 1) +VSX_SCALAR_CMP(xscmpgtdp, float64, lt, VsrD(0), 1, 1) void helper_xscmpexpdp(CPUPPCState *env, uint32_t opcode, ppc_vsr_t *xa, ppc_vsr_t *xb)