From 7dde13f87544c59807436e2eefe0f1b65fad841c Mon Sep 17 00:00:00 2001 From: Jannik Vogel Date: Mon, 16 May 2016 10:55:00 +0200 Subject: [PATCH] Fix exception propagation for VFP double precision --- src/core/arm/skyeye_common/vfp/vfp_helper.h | 8 ++- src/core/arm/skyeye_common/vfp/vfpdouble.cpp | 65 ++++++++++---------- 2 files changed, 39 insertions(+), 34 deletions(-) diff --git a/src/core/arm/skyeye_common/vfp/vfp_helper.h b/src/core/arm/skyeye_common/vfp/vfp_helper.h index 210972917d..c1cc4fd7e0 100644 --- a/src/core/arm/skyeye_common/vfp/vfp_helper.h +++ b/src/core/arm/skyeye_common/vfp/vfp_helper.h @@ -357,8 +357,9 @@ inline int vfp_double_type(const vfp_double* s) // Unpack a double-precision float. Note that this returns the magnitude // of the double-precision float mantissa with the 1. if necessary, // aligned to bit 62. -inline void vfp_double_unpack(vfp_double* s, s64 val, u32* fpscr) +inline u32 vfp_double_unpack(vfp_double* s, s64 val, u32 fpscr) { + u32 exceptions = 0; s->sign = vfp_double_packed_sign(val) >> 48; s->exponent = vfp_double_packed_exponent(val); @@ -369,12 +370,13 @@ inline void vfp_double_unpack(vfp_double* s, s64 val, u32* fpscr) // If flush-to-zero mode is enabled, turn the denormal into zero. // On a VFPv2 architecture, the sign of the zero is always positive. - if ((*fpscr & FPSCR_FLUSH_TO_ZERO) != 0 && (vfp_double_type(s) & VFP_DENORMAL) != 0) { + if ((fpscr & FPSCR_FLUSH_TO_ZERO) != 0 && (vfp_double_type(s) & VFP_DENORMAL) != 0) { s->sign = 0; s->exponent = 0; s->significand = 0; - *fpscr |= FPSCR_IDC; + exceptions |= FPSCR_IDC; } + return exceptions; } // Re-pack a double-precision float. This assumes that the float is diff --git a/src/core/arm/skyeye_common/vfp/vfpdouble.cpp b/src/core/arm/skyeye_common/vfp/vfpdouble.cpp index 45914d4792..0182b90b81 100644 --- a/src/core/arm/skyeye_common/vfp/vfpdouble.cpp +++ b/src/core/arm/skyeye_common/vfp/vfpdouble.cpp @@ -291,8 +291,9 @@ static u32 vfp_double_fsqrt(ARMul_State* state, int dd, int unused, int dm, u32 LOG_TRACE(Core_ARM11, "In %s", __FUNCTION__); vfp_double vdm, vdd, *vdp; int ret, tm; + u32 exceptions = 0; - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); tm = vfp_double_type(&vdm); if (tm & (VFP_NAN|VFP_INFINITY)) { @@ -369,7 +370,9 @@ sqrt_invalid: } vdd.significand = vfp_shiftright64jamming(vdd.significand, 1); - return vfp_double_normaliseround(state, dd, &vdd, fpscr, 0, "fsqrt"); + exceptions |= vfp_double_normaliseround(state, dd, &vdd, fpscr, 0, "fsqrt"); + + return exceptions; } /* @@ -475,7 +478,7 @@ static u32 vfp_double_fcvts(ARMul_State* state, int sd, int unused, int dm, u32 u32 exceptions = 0; LOG_TRACE(Core_ARM11, "In %s", __FUNCTION__); - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); tm = vfp_double_type(&vdm); @@ -545,7 +548,7 @@ static u32 vfp_double_ftoui(ARMul_State* state, int sd, int unused, int dm, u32 int tm; LOG_TRACE(Core_ARM11, "In %s", __FUNCTION__); - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); /* * Do we have a denormalised number? @@ -626,7 +629,7 @@ static u32 vfp_double_ftosi(ARMul_State* state, int sd, int unused, int dm, u32 int tm; LOG_TRACE(Core_ARM11, "In %s", __FUNCTION__); - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); vfp_double_dump("VDM", &vdm); /* @@ -892,21 +895,21 @@ static u32 vfp_double_multiply_accumulate(ARMul_State* state, int dd, int dn, int dm, u32 fpscr, u32 negate, const char *func) { struct vfp_double vdd, vdp, vdn, vdm; - u32 exceptions; + u32 exceptions = 0; - vfp_double_unpack(&vdn, vfp_get_double(state, dn), &fpscr); + exceptions |= vfp_double_unpack(&vdn, vfp_get_double(state, dn), fpscr); if (vdn.exponent == 0 && vdn.significand) vfp_double_normalise_denormal(&vdn); - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); if (vdm.exponent == 0 && vdm.significand) vfp_double_normalise_denormal(&vdm); - exceptions = vfp_double_multiply(&vdp, &vdn, &vdm, fpscr); + exceptions |= vfp_double_multiply(&vdp, &vdn, &vdm, fpscr); if (negate & NEG_MULTIPLY) vdp.sign = vfp_sign_negate(vdp.sign); - vfp_double_unpack(&vdn, vfp_get_double(state, dd), &fpscr); + exceptions |= vfp_double_unpack(&vdn, vfp_get_double(state, dd), fpscr); if (vdn.exponent == 0 && vdn.significand != 0) vfp_double_normalise_denormal(&vdn); @@ -964,18 +967,18 @@ static u32 vfp_double_fnmsc(ARMul_State* state, int dd, int dn, int dm, u32 fpsc static u32 vfp_double_fmul(ARMul_State* state, int dd, int dn, int dm, u32 fpscr) { struct vfp_double vdd, vdn, vdm; - u32 exceptions; + u32 exceptions = 0; LOG_TRACE(Core_ARM11, "In %s", __FUNCTION__); - vfp_double_unpack(&vdn, vfp_get_double(state, dn), &fpscr); + exceptions |= vfp_double_unpack(&vdn, vfp_get_double(state, dn), fpscr); if (vdn.exponent == 0 && vdn.significand) vfp_double_normalise_denormal(&vdn); - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); if (vdm.exponent == 0 && vdm.significand) vfp_double_normalise_denormal(&vdm); - exceptions = vfp_double_multiply(&vdd, &vdn, &vdm, fpscr); + exceptions |= vfp_double_multiply(&vdd, &vdn, &vdm, fpscr); return vfp_double_normaliseround(state, dd, &vdd, fpscr, exceptions, "fmul"); } @@ -985,18 +988,18 @@ static u32 vfp_double_fmul(ARMul_State* state, int dd, int dn, int dm, u32 fpscr static u32 vfp_double_fnmul(ARMul_State* state, int dd, int dn, int dm, u32 fpscr) { struct vfp_double vdd, vdn, vdm; - u32 exceptions; + u32 exceptions = 0; LOG_TRACE(Core_ARM11, "In %s", __FUNCTION__); - vfp_double_unpack(&vdn, vfp_get_double(state, dn), &fpscr); + exceptions |= vfp_double_unpack(&vdn, vfp_get_double(state, dn), fpscr); if (vdn.exponent == 0 && vdn.significand) vfp_double_normalise_denormal(&vdn); - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); if (vdm.exponent == 0 && vdm.significand) vfp_double_normalise_denormal(&vdm); - exceptions = vfp_double_multiply(&vdd, &vdn, &vdm, fpscr); + exceptions |= vfp_double_multiply(&vdd, &vdn, &vdm, fpscr); vdd.sign = vfp_sign_negate(vdd.sign); return vfp_double_normaliseround(state, dd, &vdd, fpscr, exceptions, "fnmul"); @@ -1008,18 +1011,18 @@ static u32 vfp_double_fnmul(ARMul_State* state, int dd, int dn, int dm, u32 fpsc static u32 vfp_double_fadd(ARMul_State* state, int dd, int dn, int dm, u32 fpscr) { struct vfp_double vdd, vdn, vdm; - u32 exceptions; + u32 exceptions = 0; LOG_TRACE(Core_ARM11, "In %s", __FUNCTION__); - vfp_double_unpack(&vdn, vfp_get_double(state, dn), &fpscr); + exceptions |= vfp_double_unpack(&vdn, vfp_get_double(state, dn), fpscr); if (vdn.exponent == 0 && vdn.significand) vfp_double_normalise_denormal(&vdn); - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); if (vdm.exponent == 0 && vdm.significand) vfp_double_normalise_denormal(&vdm); - exceptions = vfp_double_add(&vdd, &vdn, &vdm, fpscr); + exceptions |= vfp_double_add(&vdd, &vdn, &vdm, fpscr); return vfp_double_normaliseround(state, dd, &vdd, fpscr, exceptions, "fadd"); } @@ -1030,14 +1033,14 @@ static u32 vfp_double_fadd(ARMul_State* state, int dd, int dn, int dm, u32 fpscr static u32 vfp_double_fsub(ARMul_State* state, int dd, int dn, int dm, u32 fpscr) { struct vfp_double vdd, vdn, vdm; - u32 exceptions; + u32 exceptions = 0; LOG_TRACE(Core_ARM11, "In %s", __FUNCTION__); - vfp_double_unpack(&vdn, vfp_get_double(state, dn), &fpscr); + exceptions |= vfp_double_unpack(&vdn, vfp_get_double(state, dn), fpscr); if (vdn.exponent == 0 && vdn.significand) vfp_double_normalise_denormal(&vdn); - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); if (vdm.exponent == 0 && vdm.significand) vfp_double_normalise_denormal(&vdm); @@ -1046,7 +1049,7 @@ static u32 vfp_double_fsub(ARMul_State* state, int dd, int dn, int dm, u32 fpscr */ vdm.sign = vfp_sign_negate(vdm.sign); - exceptions = vfp_double_add(&vdd, &vdn, &vdm, fpscr); + exceptions |= vfp_double_add(&vdd, &vdn, &vdm, fpscr); return vfp_double_normaliseround(state, dd, &vdd, fpscr, exceptions, "fsub"); } @@ -1061,8 +1064,8 @@ static u32 vfp_double_fdiv(ARMul_State* state, int dd, int dn, int dm, u32 fpscr int tm, tn; LOG_TRACE(Core_ARM11, "In %s", __FUNCTION__); - vfp_double_unpack(&vdn, vfp_get_double(state, dn), &fpscr); - vfp_double_unpack(&vdm, vfp_get_double(state, dm), &fpscr); + exceptions |= vfp_double_unpack(&vdn, vfp_get_double(state, dn), fpscr); + exceptions |= vfp_double_unpack(&vdm, vfp_get_double(state, dm), fpscr); vdd.sign = vdn.sign ^ vdm.sign; @@ -1134,13 +1137,13 @@ static u32 vfp_double_fdiv(ARMul_State* state, int dd, int dn, int dm, u32 fpscr return vfp_double_normaliseround(state, dd, &vdd, fpscr, 0, "fdiv"); vdn_nan: - exceptions = vfp_propagate_nan(&vdd, &vdn, &vdm, fpscr); + exceptions |= vfp_propagate_nan(&vdd, &vdn, &vdm, fpscr); pack: vfp_put_double(state, vfp_double_pack(&vdd), dd); return exceptions; vdm_nan: - exceptions = vfp_propagate_nan(&vdd, &vdm, &vdn, fpscr); + exceptions |= vfp_propagate_nan(&vdd, &vdm, &vdn, fpscr); goto pack; zero: @@ -1149,7 +1152,7 @@ zero: goto pack; divzero: - exceptions = FPSCR_DZC; + exceptions |= FPSCR_DZC; infinity: vdd.exponent = 2047; vdd.significand = 0;