X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=arch%2Farm%2Fvfp%2Fvfpsingle.c;h=0221ba3bc799bcef222032a5545013bb91a286fe;hb=refs%2Fheads%2Fvserver;hp=14dd696ddeb1e98177cbc3b9a9883d3218ca592f;hpb=76828883507a47dae78837ab5dec5a5b4513c667;p=linux-2.6.git diff --git a/arch/arm/vfp/vfpsingle.c b/arch/arm/vfp/vfpsingle.c index 14dd696dd..0221ba3bc 100644 --- a/arch/arm/vfp/vfpsingle.c +++ b/arch/arm/vfp/vfpsingle.c @@ -198,12 +198,14 @@ u32 vfp_single_normaliseround(int sd, struct vfp_single *vs, u32 fpscr, u32 exce vfp_single_dump("pack: final", vs); { s32 d = vfp_single_pack(vs); +#ifdef DEBUG pr_debug("VFP: %s: d(s%d)=%08x exceptions=%08x\n", func, sd, d, exceptions); - vfp_put_float(sd, d); +#endif + vfp_put_float(d, sd); } - return exceptions & ~VFP_NAN_FLAG; + return exceptions; } /* @@ -257,19 +259,19 @@ vfp_propagate_nan(struct vfp_single *vsd, struct vfp_single *vsn, */ static u32 vfp_single_fabs(int sd, int unused, s32 m, u32 fpscr) { - vfp_put_float(sd, vfp_single_packed_abs(m)); + vfp_put_float(vfp_single_packed_abs(m), sd); return 0; } static u32 vfp_single_fcpy(int sd, int unused, s32 m, u32 fpscr) { - vfp_put_float(sd, m); + vfp_put_float(m, sd); return 0; } static u32 vfp_single_fneg(int sd, int unused, s32 m, u32 fpscr) { - vfp_put_float(sd, vfp_single_packed_negate(m)); + vfp_put_float(vfp_single_packed_negate(m), sd); return 0; } @@ -333,7 +335,7 @@ static u32 vfp_single_fsqrt(int sd, int unused, s32 m, u32 fpscr) vsp = &vfp_single_default_qnan; ret = FPSCR_IOC; } - vfp_put_float(sd, vfp_single_pack(vsp)); + vfp_put_float(vfp_single_pack(vsp), sd); return ret; } @@ -506,7 +508,7 @@ static u32 vfp_single_fcvtd(int dd, int unused, s32 m, u32 fpscr) */ if (tm & (VFP_INFINITY|VFP_NAN)) { vdd.exponent = 2047; - if (tm & VFP_NAN) + if (tm == VFP_QNAN) vdd.significand |= VFP_DOUBLE_SIGNIFICAND_QNAN; goto pack_nan; } else if (tm & VFP_ZERO) @@ -514,14 +516,10 @@ static u32 vfp_single_fcvtd(int dd, int unused, s32 m, u32 fpscr) else vdd.exponent = vsm.exponent + (1023 - 127); - /* - * Technically, if bit 0 of dd is set, this is an invalid - * instruction. However, we ignore this for efficiency. - */ return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fcvtd"); pack_nan: - vfp_put_double(dd, vfp_double_pack(&vdd)); + vfp_put_double(vfp_double_pack(&vdd), dd); return exceptions; } @@ -617,7 +615,7 @@ static u32 vfp_single_ftoui(int sd, int unused, s32 m, u32 fpscr) pr_debug("VFP: ftoui: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions); - vfp_put_float(sd, d); + vfp_put_float(d, sd); return exceptions; } @@ -632,6 +630,7 @@ static u32 vfp_single_ftosi(int sd, int unused, s32 m, u32 fpscr) struct vfp_single vsm; u32 d, exceptions = 0; int rmode = fpscr & FPSCR_RMODE_MASK; + int tm; vfp_single_unpack(&vsm, m); vfp_single_dump("VSM", &vsm); @@ -639,10 +638,14 @@ static u32 vfp_single_ftosi(int sd, int unused, s32 m, u32 fpscr) /* * Do we have a denormalised number? */ + tm = vfp_single_type(&vsm); if (vfp_single_type(&vsm) & VFP_DENORMAL) exceptions |= FPSCR_IDC; - if (vsm.exponent >= 127 + 32) { + if (tm & VFP_NAN) { + d = 0; + exceptions |= FPSCR_IOC; + } else if (vsm.exponent >= 127 + 32) { /* * m >= 2^31-2^7: invalid */ @@ -691,7 +694,7 @@ static u32 vfp_single_ftosi(int sd, int unused, s32 m, u32 fpscr) pr_debug("VFP: ftosi: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions); - vfp_put_float(sd, (s32)d); + vfp_put_float((s32)d, sd); return exceptions; } @@ -701,22 +704,22 @@ static u32 vfp_single_ftosiz(int sd, int unused, s32 m, u32 fpscr) return vfp_single_ftosi(sd, unused, m, FPSCR_ROUND_TOZERO); } -static u32 (* const fop_extfns[32])(int sd, int unused, s32 m, u32 fpscr) = { - [FEXT_TO_IDX(FEXT_FCPY)] = vfp_single_fcpy, - [FEXT_TO_IDX(FEXT_FABS)] = vfp_single_fabs, - [FEXT_TO_IDX(FEXT_FNEG)] = vfp_single_fneg, - [FEXT_TO_IDX(FEXT_FSQRT)] = vfp_single_fsqrt, - [FEXT_TO_IDX(FEXT_FCMP)] = vfp_single_fcmp, - [FEXT_TO_IDX(FEXT_FCMPE)] = vfp_single_fcmpe, - [FEXT_TO_IDX(FEXT_FCMPZ)] = vfp_single_fcmpz, - [FEXT_TO_IDX(FEXT_FCMPEZ)] = vfp_single_fcmpez, - [FEXT_TO_IDX(FEXT_FCVT)] = vfp_single_fcvtd, - [FEXT_TO_IDX(FEXT_FUITO)] = vfp_single_fuito, - [FEXT_TO_IDX(FEXT_FSITO)] = vfp_single_fsito, - [FEXT_TO_IDX(FEXT_FTOUI)] = vfp_single_ftoui, - [FEXT_TO_IDX(FEXT_FTOUIZ)] = vfp_single_ftouiz, - [FEXT_TO_IDX(FEXT_FTOSI)] = vfp_single_ftosi, - [FEXT_TO_IDX(FEXT_FTOSIZ)] = vfp_single_ftosiz, +static struct op fops_ext[32] = { + [FEXT_TO_IDX(FEXT_FCPY)] = { vfp_single_fcpy, 0 }, + [FEXT_TO_IDX(FEXT_FABS)] = { vfp_single_fabs, 0 }, + [FEXT_TO_IDX(FEXT_FNEG)] = { vfp_single_fneg, 0 }, + [FEXT_TO_IDX(FEXT_FSQRT)] = { vfp_single_fsqrt, 0 }, + [FEXT_TO_IDX(FEXT_FCMP)] = { vfp_single_fcmp, OP_SCALAR }, + [FEXT_TO_IDX(FEXT_FCMPE)] = { vfp_single_fcmpe, OP_SCALAR }, + [FEXT_TO_IDX(FEXT_FCMPZ)] = { vfp_single_fcmpz, OP_SCALAR }, + [FEXT_TO_IDX(FEXT_FCMPEZ)] = { vfp_single_fcmpez, OP_SCALAR }, + [FEXT_TO_IDX(FEXT_FCVT)] = { vfp_single_fcvtd, OP_SCALAR|OP_DD }, + [FEXT_TO_IDX(FEXT_FUITO)] = { vfp_single_fuito, OP_SCALAR }, + [FEXT_TO_IDX(FEXT_FSITO)] = { vfp_single_fsito, OP_SCALAR }, + [FEXT_TO_IDX(FEXT_FTOUI)] = { vfp_single_ftoui, OP_SCALAR }, + [FEXT_TO_IDX(FEXT_FTOUIZ)] = { vfp_single_ftouiz, OP_SCALAR }, + [FEXT_TO_IDX(FEXT_FTOSI)] = { vfp_single_ftosi, OP_SCALAR }, + [FEXT_TO_IDX(FEXT_FTOSIZ)] = { vfp_single_ftosiz, OP_SCALAR }, }; @@ -1126,7 +1129,7 @@ static u32 vfp_single_fdiv(int sd, int sn, s32 m, u32 fpscr) vsn_nan: exceptions = vfp_propagate_nan(&vsd, &vsn, &vsm, fpscr); pack: - vfp_put_float(sd, vfp_single_pack(&vsd)); + vfp_put_float(vfp_single_pack(&vsd), sd); return exceptions; vsm_nan: @@ -1146,20 +1149,20 @@ static u32 vfp_single_fdiv(int sd, int sn, s32 m, u32 fpscr) goto pack; invalid: - vfp_put_float(sd, vfp_single_pack(&vfp_single_default_qnan)); + vfp_put_float(vfp_single_pack(&vfp_single_default_qnan), sd); return FPSCR_IOC; } -static u32 (* const fop_fns[16])(int sd, int sn, s32 m, u32 fpscr) = { - [FOP_TO_IDX(FOP_FMAC)] = vfp_single_fmac, - [FOP_TO_IDX(FOP_FNMAC)] = vfp_single_fnmac, - [FOP_TO_IDX(FOP_FMSC)] = vfp_single_fmsc, - [FOP_TO_IDX(FOP_FNMSC)] = vfp_single_fnmsc, - [FOP_TO_IDX(FOP_FMUL)] = vfp_single_fmul, - [FOP_TO_IDX(FOP_FNMUL)] = vfp_single_fnmul, - [FOP_TO_IDX(FOP_FADD)] = vfp_single_fadd, - [FOP_TO_IDX(FOP_FSUB)] = vfp_single_fsub, - [FOP_TO_IDX(FOP_FDIV)] = vfp_single_fdiv, +static struct op fops[16] = { + [FOP_TO_IDX(FOP_FMAC)] = { vfp_single_fmac, 0 }, + [FOP_TO_IDX(FOP_FNMAC)] = { vfp_single_fnmac, 0 }, + [FOP_TO_IDX(FOP_FMSC)] = { vfp_single_fmsc, 0 }, + [FOP_TO_IDX(FOP_FNMSC)] = { vfp_single_fnmsc, 0 }, + [FOP_TO_IDX(FOP_FMUL)] = { vfp_single_fmul, 0 }, + [FOP_TO_IDX(FOP_FNMUL)] = { vfp_single_fnmul, 0 }, + [FOP_TO_IDX(FOP_FADD)] = { vfp_single_fadd, 0 }, + [FOP_TO_IDX(FOP_FSUB)] = { vfp_single_fsub, 0 }, + [FOP_TO_IDX(FOP_FDIV)] = { vfp_single_fdiv, 0 }, }; #define FREG_BANK(x) ((x) & 0x18) @@ -1169,60 +1172,68 @@ u32 vfp_single_cpdo(u32 inst, u32 fpscr) { u32 op = inst & FOP_MASK; u32 exceptions = 0; - unsigned int sd = vfp_get_sd(inst); + unsigned int dest; unsigned int sn = vfp_get_sn(inst); unsigned int sm = vfp_get_sm(inst); unsigned int vecitr, veclen, vecstride; - u32 (*fop)(int, int, s32, u32); + struct op *fop; - veclen = fpscr & FPSCR_LENGTH_MASK; vecstride = 1 + ((fpscr & FPSCR_STRIDE_MASK) == FPSCR_STRIDE_MASK); + fop = (op == FOP_EXT) ? &fops_ext[FEXT_TO_IDX(inst)] : &fops[FOP_TO_IDX(op)]; + + /* + * fcvtsd takes a dN register number as destination, not sN. + * Technically, if bit 0 of dd is set, this is an invalid + * instruction. However, we ignore this for efficiency. + * It also only operates on scalars. + */ + if (fop->flags & OP_DD) + dest = vfp_get_dd(inst); + else + dest = vfp_get_sd(inst); + /* * If destination bank is zero, vector length is always '1'. * ARM DDI0100F C5.1.3, C5.3.2. */ - if (FREG_BANK(sd) == 0) + if ((fop->flags & OP_SCALAR) || FREG_BANK(dest) == 0) veclen = 0; + else + veclen = fpscr & FPSCR_LENGTH_MASK; pr_debug("VFP: vecstride=%u veclen=%u\n", vecstride, (veclen >> FPSCR_LENGTH_BIT) + 1); - fop = (op == FOP_EXT) ? fop_extfns[sn] : fop_fns[FOP_TO_IDX(op)]; - if (!fop) + if (!fop->fn) goto invalid; for (vecitr = 0; vecitr <= veclen; vecitr += 1 << FPSCR_LENGTH_BIT) { s32 m = vfp_get_float(sm); u32 except; + char type; + type = fop->flags & OP_DD ? 'd' : 's'; if (op == FOP_EXT) - pr_debug("VFP: itr%d (s%u) = op[%u] (s%u=%08x)\n", - vecitr >> FPSCR_LENGTH_BIT, sd, sn, sm, m); + pr_debug("VFP: itr%d (%c%u) = op[%u] (s%u=%08x)\n", + vecitr >> FPSCR_LENGTH_BIT, type, dest, sn, + sm, m); else - pr_debug("VFP: itr%d (s%u) = (s%u) op[%u] (s%u=%08x)\n", - vecitr >> FPSCR_LENGTH_BIT, sd, sn, + pr_debug("VFP: itr%d (%c%u) = (s%u) op[%u] (s%u=%08x)\n", + vecitr >> FPSCR_LENGTH_BIT, type, dest, sn, FOP_TO_IDX(op), sm, m); - except = fop(sd, sn, m, fpscr); + except = fop->fn(dest, sn, m, fpscr); pr_debug("VFP: itr%d: exceptions=%08x\n", vecitr >> FPSCR_LENGTH_BIT, except); exceptions |= except; - /* - * This ensures that comparisons only operate on scalars; - * comparisons always return with one FPSCR status bit set. - */ - if (except & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V)) - break; - /* * CHECK: It appears to be undefined whether we stop when * we encounter an exception. We continue. */ - - sd = FREG_BANK(sd) + ((FREG_IDX(sd) + vecstride) & 7); + dest = FREG_BANK(dest) + ((FREG_IDX(dest) + vecstride) & 7); sn = FREG_BANK(sn) + ((FREG_IDX(sn) + vecstride) & 7); if (FREG_BANK(sm) != 0) sm = FREG_BANK(sm) + ((FREG_IDX(sm) + vecstride) & 7);