s390x/tcg: Implement Vector enhancements facility and switch to z14
cover s390x/tcg: Implement Vector enhancements facility and switch to z14
This series adds support for the "Vector enhancements facility" and bumps the qemu CPU model tp to a stripped-down z14.
I yet have to find a way to get more test coverage - looks like some of the functions aren't used anywhere yet (e.g., VECTOR FP MAXIMUM), writing unit tests to cover all functions and cases is just nasty. But I might be wrong - I'm planning to at least test basic functionality of all new added instructions.
I have to make excessive use of c macros again to cover different element sizes (32/64/128bit). Any advise to clean things up are welcome.
This is based on:
"[PATCH v2 0/9] s390x/tcg: Implement some z14 facilities"
"[PATCH v2 00/10] softfloat: Implement float128_muladd"Based-on: [email protected]
Based-on: [email protected]
David Hildenbrand (20):
softfloat: Implement
float128_(min|minnum|minnummag|max|maxnum|maxnummag)
s390x/tcg: Implement VECTOR BIT PERMUTE
s390x/tcg: Implement VECTOR MULTIPLY SUM LOGICAL s390x/tcg: Implement 32/128 bit for VECTOR FP ADD s390x/tcg: Implement 32/128 bit for VECTOR FP DIVIDE s390x/tcg: Implement 32/128 bit for VECTOR FP MULTIPLY s390x/tcg: Implement 32/128 bit for VECTOR FP SUBTRACT s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE (AND SIGNAL)
SCALAR
s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE * s390x/tcg: Implement 32/128 bit for VECTOR LOAD FP INTEGER s390x/tcg: Implement 64 bit for VECTOR FP LOAD LENGTHENED s390x/tcg: Implement 128 bit for VECTOR FP LOAD ROUNDED s390x/tcg: Implement 32/128 bit for VECTOR FP PERFORM SIGN OPERATION s390x/tcg: Implement 32/128 bit for VECTOR FP SQUARE ROOT s390x/tcg: Implement 32/128 bit for VECTOR FP TEST DATA CLASS
IMMEDIATE
s390x/tcg: Implement 32/128bit for VECTOR FP MULTIPLY AND
(ADD|SUBTRACT)
s390x/tcg: Implement VECTOR FP NEGATIVE MULTIPLY AND (ADD|SUBTRACT) s390x/tcg: Implement VECTOR FP (MAXIMUM|MINIMUM) s390x/tcg: We support Vector enhancements facility s390x/cpumodel: Bump up QEMU model to a stripped-down IBM z14 GA2
fpu/softfloat.c | 100 +++ hw/s390x/s390-virtio-ccw.c | 2 + include/fpu/softfloat.h | 6 + target/s390x/cpu_models.c | 4 +- target/s390x/gen-features.c | 14 +- target/s390x/helper.h | 72 ++ target/s390x/insn-data.def | 12 + target/s390x/translate_vx.c.inc | 625 ++++++++++++--- target/s390x/vec_fpu_helper.c | 1302 ++++++++++++++++++++++--------- target/s390x/vec_helper.c | 22 + 10 files changed, 1681 insertions(+), 478 deletions(-)
Signature · 1 line
2.26.2
PATCH 1/20 softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)
Implementation inspired by minmax_floats(). Unfortuantely, we don't have any tests we can simply adjust/unlock.
--- fpu/softfloat.c | 100 ++++++++++++++++++++++++++++++++++++++++ include/fpu/softfloat.h | 6 +++ 2 files changed, 106 insertions(+)
fpu/softfloat.c +100 -0
@@ -621,6 +621,8 @@ static inline FloatParts float64_unpack_raw(float64 f)return unpack_raw (float64_params, f); } +static void float128_unpack (FloatParts128 *p,float128 a,float_status *status); +/* Pack a float from parts, but do not canonicalize. */ static inline uint64_t pack_raw (FloatFmt fmt,FloatParts p) {
@@ -3180,6 +3182,89 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, } } +static float128 float128_minmax (float128 a,float128 b,bool ismin,bool ieee, +bool ismag,float_status *s) +{ +FloatParts128 pa, pb; +int a_exp, b_exp; +bool a_less; + +float128_unpack (&pa, a, s); +float128_unpack (&pb, b, s); + +if (unlikely (is_nan (pa.cls) ||is_nan (pb.cls))) { +/* See comment in minmax_floats() */ +if (ieee && !is_snan (pa.cls) && !is_snan (pb.cls)) { +if (is_nan (pa.cls) && !is_nan (pb.cls)) { +return b; + }else if (is_nan (pb.cls) && !is_nan (pa.cls)) { +return a; + } + } + +/* Similar logic to pick_nan(), avoiding re-packing. */ +if (is_snan (pa.cls) ||is_snan (pb.cls)) { + s->float_exception_flags |= float_flag_invalid; + } +if (s->default_nan_mode) { +return float128_default_nan (s); + } +if (pickNaN (pa.cls, pb.cls, + pa.frac0 > pb.frac0 || + (pa.frac0 == pb.frac0 && pa.frac1 > pb.frac1) || + (pa.frac0 == pb.frac0 && pa.frac1 == pb.frac1 && + pa.sign < pb.sign), s)) { +return is_snan (pb.cls) ?float128_silence_nan (b, s) : b; + } +return is_snan (pa.cls) ?float128_silence_nan (a, s) : a; + } + +switch (pa.cls) { +case float_class_normal: + a_exp = pa.exp; +break ; +case float_class_inf: + a_exp =INT_MAX ; +break ; +case float_class_zero: + a_exp =INT_MIN ; +break ; +default : +g_assert_not_reached (); +break ; + } +switch (pb.cls) { +case float_class_normal: + b_exp = pb.exp; +break ; +case float_class_inf: + b_exp =INT_MAX ; +break ; +case float_class_zero: + b_exp =INT_MIN ; +break ; +default : +g_assert_not_reached (); +break ; + } + + a_less = a_exp < b_exp; +if (a_exp == b_exp) { + a_less = pa.frac0 < pb.frac0; +if (pa.frac0 == pb.frac0) { + a_less = pa.frac1 < pb.frac1; + } + } + +if (ismag && + (a_exp != b_exp || pa.frac0 != pb.frac0 || pa.frac1 != pb.frac1)) { +return a_less ^ ismin ? b : a; + }else if (pa.sign == pb.sign) { +return pa.sign ^ a_less ^ ismin ? b : a; + } +return pa.sign ^ ismin ? b : a; +} +#define MINMAX (sz, name, ismin, isiee, ismag) \ float ## sz float ## sz ## _ ## name(float ## sz a, float ## sz b, \ float_status *s) \
@@ -3214,6 +3299,21 @@ MINMAX(64, maxnummag, false, true, true)#undef MINMAX +#define F128_MINMAX (name, ismin, isiee, ismag) \ +float128 float128_ ## name(float128 a, float128 b, float_status *s) \ +{ \ + return float128_minmax(a, b, ismin, isiee, ismag, s); \ +} + +F128_MINMAX(min, true, false, false) +F128_MINMAX(minnum, true, true, false) +F128_MINMAX(minnummag, true, true, true) +F128_MINMAX(max, false, false, false) +F128_MINMAX(maxnum, false, true, false) +F128_MINMAX(maxnummag, false, true, true) + +#undef F128_MINMAX +#define BF16_MINMAX (name, ismin, isiee, ismag) \ bfloat16 bfloat16_ ## name(bfloat16 a, bfloat16 b, float_status *s) \ { \
include/fpu/softfloat.h +6 -0
@@ -1201,6 +1201,12 @@ float128 float128_muladd(float128, float128, float128, int,float128 float128_sqrt (float128 ,float_status *status);FloatRelation float128_compare (float128 ,float128 ,float_status *status);FloatRelation float128_compare_quiet (float128 ,float128 ,float_status *status); +float128 float128_min (float128 ,float128 ,float_status *status); +float128 float128_max (float128 ,float128 ,float_status *status); +float128 float128_minnum (float128 ,float128 ,float_status *status); +float128 float128_maxnum (float128 ,float128 ,float_status *status); +float128 float128_minnummag (float128 ,float128 ,float_status *status); +float128 float128_maxnummag (float128 ,float128 ,float_status *status);bool float128_is_quiet_nan (float128 ,float_status *status);bool float128_is_signaling_nan (float128 ,float_status *status);float128 float128_silence_nan (float128 ,float_status *status);
Signature · 1 line
2.26.2
Re: [PATCH v1 01/20] softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)
Quoted 26 lines · David Hildenbrand
Implementation inspired by
minmax_floats(). Unfortuantely, we don't have any tests we can simply adjust/unlock.Cc Aurelien Jarno <[email protected]>Cc Peter Maydell <[email protected]>Cc "Alex Bennée" <[email protected]>Signed-off-by David Hildenbrand <[email protected]>--- fpu/softfloat.c | 100 ++++++++++++++++++++++++++++++++++++++++ include/fpu/softfloat.h | 6 +++ 2 files changed, 106 insertions(+)fpu/softfloat.c +2
-0@@ -621,6 +621,8 @@ static inline FloatParts float64_unpack_raw(float64 f)return unpack_raw (float64_params, f); } +static void float128_unpack (FloatParts128 *p,float128 a,float_status *status); +/* Pack a float from parts, but do not canonicalize. */ static inline uint64_t pack_raw (FloatFmt fmt,FloatParts p) {
fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3183
Quoted 2 lines
fpu/softfloat.c +0
-0@@ -3180,6 +3182,89 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, } }
It would be desirable to share as much logic for this as possible with the existing minmax_floats code. I appreciate at some point we end up having to deal with fractions and we haven't found a good way to efficiently handle dealing with FloatParts and FloatParts128 in the same unrolled code, however:
fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3194
Quoted 11 lines
fpu/softfloat.c +10
-0+static float128 float128_minmax (float128 a,float128 b,bool ismin,bool ieee, +bool ismag,float_status *s) +{ +FloatParts128 pa, pb; +int a_exp, b_exp; +bool a_less; + +float128_unpack (&pa, a, s); +float128_unpack (&pb, b, s); +
From here:
Quoted 7 lines
fpu/softfloat.c +7
-0+if (unlikely (is_nan (pa.cls) ||is_nan (pb.cls))) { +/* See comment in minmax_floats() */ +if (ieee && !is_snan (pa.cls) && !is_snan (pb.cls)) { +if (is_nan (pa.cls) && !is_nan (pb.cls)) { +return b; + }else if (is_nan (pb.cls) && !is_nan (pa.cls)) { +return a;
fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3211
Quoted 10 lines
fpu/softfloat.c +10
-0+ } + } + +/* Similar logic to pick_nan(), avoiding re-packing. */ +if (is_snan (pa.cls) ||is_snan (pb.cls)) { + s->float_exception_flags |= float_flag_invalid; + } +if (s->default_nan_mode) { +return float128_default_nan (s); + }
to here is common logic - is there anyway we could share it?
Quoted 14 lines
fpu/softfloat.c +14
-0+if (pickNaN (pa.cls, pb.cls, + pa.frac0 > pb.frac0 || + (pa.frac0 == pb.frac0 && pa.frac1 > pb.frac1) || + (pa.frac0 == pb.frac0 && pa.frac1 == pb.frac1 && + pa.sign < pb.sign), s)) { +return is_snan (pb.cls) ?float128_silence_nan (b, s) : b; + } +return is_snan (pa.cls) ?float128_silence_nan (a, s) : a; + } + +switch (pa.cls) { +case float_class_normal: + a_exp = pa.exp; +break ;
fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3235
Quoted 10 lines
fpu/softfloat.c +10
-0+case float_class_inf: + a_exp =INT_MAX ; +break ; +case float_class_zero: + a_exp =INT_MIN ; +break ; +default : +g_assert_not_reached (); +break ; + }
Likewise I wonder if there is scope for a float_minmax_exp helper that could be shared here?
Quoted 21 lines
fpu/softfloat.c +21
-0+switch (pb.cls) { +case float_class_normal: + b_exp = pb.exp; +break ; +case float_class_inf: + b_exp =INT_MAX ; +break ; +case float_class_zero: + b_exp =INT_MIN ; +break ; +default : +g_assert_not_reached (); +break ; + } + + a_less = a_exp < b_exp; +if (a_exp == b_exp) { + a_less = pa.frac0 < pb.frac0; +if (pa.frac0 == pb.frac0) { + a_less = pa.frac1 < pb.frac1; + }
fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3266
Quoted 10 lines
fpu/softfloat.c +10
-0+ } + +if (ismag && + (a_exp != b_exp || pa.frac0 != pb.frac0 || pa.frac1 != pb.frac1)) { +return a_less ^ ismin ? b : a; + }else if (pa.sign == pb.sign) { +return pa.sign ^ a_less ^ ismin ? b : a; + } +return pa.sign ^ ismin ? b : a; +}
Otherwise it seems sane to me.
Signature · 1 line
Alex Bennée
Re: [PATCH v1 01/20] softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)
Quoted 69 lines · Alex Bennée
Quoted 26 lines · David Hildenbrand
Implementation inspired by
minmax_floats(). Unfortuantely, we don't have any tests we can simply adjust/unlock.Cc Aurelien Jarno <[email protected]>Cc Peter Maydell <[email protected]>Cc "Alex Bennée" <[email protected]>Signed-off-by David Hildenbrand <[email protected]>--- fpu/softfloat.c | 100 ++++++++++++++++++++++++++++++++++++++++ include/fpu/softfloat.h | 6 +++ 2 files changed, 106 insertions(+)fpu/softfloat.c +2
-0@@ -621,6 +621,8 @@ static inline FloatParts float64_unpack_raw(float64 f)return unpack_raw (float64_params, f); } +static void float128_unpack (FloatParts128 *p,float128 a,float_status *status); +/* Pack a float from parts, but do not canonicalize. */ static inline uint64_t pack_raw (FloatFmt fmt,FloatParts p) {fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3183
Quoted 2 lines
fpu/softfloat.c +0
-0@@ -3180,6 +3182,89 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, } }It would be desirable to share as much logic for this as possible with the existing
minmax_floatscode. I appreciate at some point we end up having to deal with fractions and we haven't found a good way to efficiently handle dealing with FloatParts and FloatParts128 in the same unrolled code, however:fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3194
Quoted 11 lines
fpu/softfloat.c +10
-0+static float128 float128_minmax (float128 a,float128 b,bool ismin,bool ieee, +bool ismag,float_status *s) +{ +FloatParts128 pa, pb; +int a_exp, b_exp; +bool a_less; + +float128_unpack (&pa, a, s); +float128_unpack (&pb, b, s); +From here:
Quoted 7 lines
fpu/softfloat.c +7
-0+if (unlikely (is_nan (pa.cls) ||is_nan (pb.cls))) { +/* See comment in minmax_floats() */ +if (ieee && !is_snan (pa.cls) && !is_snan (pb.cls)) { +if (is_nan (pa.cls) && !is_nan (pb.cls)) { +return b; + }else if (is_nan (pb.cls) && !is_nan (pa.cls)) { +return a;fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3211
Quoted 10 lines
fpu/softfloat.c +10
-0+ } + } + +/* Similar logic to pick_nan(), avoiding re-packing. */ +if (is_snan (pa.cls) ||is_snan (pb.cls)) { + s->float_exception_flags |= float_flag_invalid; + } +if (s->default_nan_mode) { +return float128_default_nan (s); + }
Quoted 1 line
to here is common logic - is there anyway we could share it?
I can try to factor it out, similar to pickNaN() - passing weird boolean flags and such. It most certainly won't win in a beauty contest, that's for sure.
Quoted 26 lines
Quoted 15 lines
fpu/softfloat.c +14
-0+if (pickNaN (pa.cls, pb.cls, + pa.frac0 > pb.frac0 || + (pa.frac0 == pb.frac0 && pa.frac1 > pb.frac1) || + (pa.frac0 == pb.frac0 && pa.frac1 == pb.frac1 && + pa.sign < pb.sign), s)) { +return is_snan (pb.cls) ?float128_silence_nan (b, s) : b; + } +return is_snan (pa.cls) ?float128_silence_nan (a, s) : a; + } + +switch (pa.cls) { +case float_class_normal: + a_exp = pa.exp; +break ;fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3235
Quoted 10 lines
fpu/softfloat.c +10
-0+case float_class_inf: + a_exp =INT_MAX ; +break ; +case float_class_zero: + a_exp =INT_MIN ; +break ; +default : +g_assert_not_reached (); +break ; + }
Quoted 2 lines
Likewise I wonder if there is scope for a
float_minmax_exphelper that could be shared here?
I'll try, but I have the feeling that it might make the code harder to read than actually help. Will give it a try.
Thanks!
Signature · 3 lines
Thanks, David / dhildenb
Re: [PATCH v1 01/20] softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)
Quoted 76 lines · David Hildenbrand
Quoted 69 lines · Alex Bennée
Quoted 26 lines · David Hildenbrand
Implementation inspired by
minmax_floats(). Unfortuantely, we don't have any tests we can simply adjust/unlock.Cc Aurelien Jarno <[email protected]>Cc Peter Maydell <[email protected]>Cc "Alex Bennée" <[email protected]>Signed-off-by David Hildenbrand <[email protected]>--- fpu/softfloat.c | 100 ++++++++++++++++++++++++++++++++++++++++ include/fpu/softfloat.h | 6 +++ 2 files changed, 106 insertions(+)fpu/softfloat.c +2
-0@@ -621,6 +621,8 @@ static inline FloatParts float64_unpack_raw(float64 f)return unpack_raw (float64_params, f); } +static void float128_unpack (FloatParts128 *p,float128 a,float_status *status); +/* Pack a float from parts, but do not canonicalize. */ static inline uint64_t pack_raw (FloatFmt fmt,FloatParts p) {fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3183
Quoted 2 lines
fpu/softfloat.c +0
-0@@ -3180,6 +3182,89 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, } }It would be desirable to share as much logic for this as possible with the existing
minmax_floatscode. I appreciate at some point we end up having to deal with fractions and we haven't found a good way to efficiently handle dealing with FloatParts and FloatParts128 in the same unrolled code, however:fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3194
Quoted 11 lines
fpu/softfloat.c +10
-0+static float128 float128_minmax (float128 a,float128 b,bool ismin,bool ieee, +bool ismag,float_status *s) +{ +FloatParts128 pa, pb; +int a_exp, b_exp; +bool a_less; + +float128_unpack (&pa, a, s); +float128_unpack (&pb, b, s); +From here:
Quoted 7 lines
fpu/softfloat.c +7
-0+if (unlikely (is_nan (pa.cls) ||is_nan (pb.cls))) { +/* See comment in minmax_floats() */ +if (ieee && !is_snan (pa.cls) && !is_snan (pb.cls)) { +if (is_nan (pa.cls) && !is_nan (pb.cls)) { +return b; + }else if (is_nan (pb.cls) && !is_nan (pa.cls)) { +return a;fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3211
Quoted 10 lines
fpu/softfloat.c +10
-0+ } + } + +/* Similar logic to pick_nan(), avoiding re-packing. */ +if (is_snan (pa.cls) ||is_snan (pb.cls)) { + s->float_exception_flags |= float_flag_invalid; + } +if (s->default_nan_mode) { +return float128_default_nan (s); + }Quoted 1 line
to here is common logic - is there anyway we could share it?
I can try to factor it out, similar to
pickNaN()- passing weird boolean flags and such. It most certainly won't win in a beauty contest, that's for sure.Quoted 1 line
Quoted 5 lines
Quoted 2 lines
Likewise I wonder if there is scope for a
float_minmax_exphelper that could be shared here?I'll try, but I have the feeling that it might make the code harder to read than actually help. Will give it a try.
Give it a try - if it really does become harder to follow then we'll stick with the duplication however if we can have common code you'll know at least the nan handling and minmax behaviour for float128 will be partially tested by the 16/32/64 float code.
Quoted 2 lines
Thanks!
Signature · 1 line
Alex Bennée
Re: [PATCH v1 01/20] softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)
Quoted 82 lines · Alex Bennée
Quoted 79 lines · David Hildenbrand
Quoted 69 lines · Alex Bennée
Quoted 26 lines · David Hildenbrand
Implementation inspired by
minmax_floats(). Unfortuantely, we don't have any tests we can simply adjust/unlock.Cc Aurelien Jarno <[email protected]>Cc Peter Maydell <[email protected]>Cc "Alex Bennée" <[email protected]>Signed-off-by David Hildenbrand <[email protected]>--- fpu/softfloat.c | 100 ++++++++++++++++++++++++++++++++++++++++ include/fpu/softfloat.h | 6 +++ 2 files changed, 106 insertions(+)fpu/softfloat.c +2
-0@@ -621,6 +621,8 @@ static inline FloatParts float64_unpack_raw(float64 f)return unpack_raw (float64_params, f); } +static void float128_unpack (FloatParts128 *p,float128 a,float_status *status); +/* Pack a float from parts, but do not canonicalize. */ static inline uint64_t pack_raw (FloatFmt fmt,FloatParts p) {fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3183
Quoted 2 lines
fpu/softfloat.c +0
-0@@ -3180,6 +3182,89 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, } }It would be desirable to share as much logic for this as possible with the existing
minmax_floatscode. I appreciate at some point we end up having to deal with fractions and we haven't found a good way to efficiently handle dealing with FloatParts and FloatParts128 in the same unrolled code, however:fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3193
Quoted 11 lines
fpu/softfloat.c +10
-0+static float128 float128_minmax (float128 a,float128 b,bool ismin,bool ieee, +bool ismag,float_status *s) +{ +FloatParts128 pa, pb; +int a_exp, b_exp; +bool a_less; + +float128_unpack (&pa, a, s); +float128_unpack (&pb, b, s); +From here:
Quoted 7 lines
fpu/softfloat.c +7
-0+if (unlikely (is_nan (pa.cls) ||is_nan (pb.cls))) { +/* See comment in minmax_floats() */ +if (ieee && !is_snan (pa.cls) && !is_snan (pb.cls)) { +if (is_nan (pa.cls) && !is_nan (pb.cls)) { +return b; + }else if (is_nan (pb.cls) && !is_nan (pa.cls)) { +return a;fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3210
Quoted 10 lines
fpu/softfloat.c +10
-0+ } + } + +/* Similar logic to pick_nan(), avoiding re-packing. */ +if (is_snan (pa.cls) ||is_snan (pb.cls)) { + s->float_exception_flags |= float_flag_invalid; + } +if (s->default_nan_mode) { +return float128_default_nan (s); + }Quoted 1 line
to here is common logic - is there anyway we could share it?
I can try to factor it out, similar to
pickNaN()- passing weird boolean flags and such. It most certainly won't win in a beauty contest, that's for sure.Quoted 3 lines
Likewise I wonder if there is scope for a
float_minmax_exphelper that could be shared here?
Quoted 7 lines
Quoted 2 lines
I'll try, but I have the feeling that it might make the code harder to read than actually help. Will give it a try.
Give it a try - if it really does become harder to follow then we'll stick with the duplication however if we can have common code you'll know at least the nan handling and minmax behaviour for float128 will be partially tested by the 16/32/64 float code.
(finally had time to look into this)
What about something like this:
From dd6cf176c840fc34a095cb2a158032a994aca5ef Mon Sep 17 00:00:00 2001
From: David Hildenbrand <[email protected]>
Date: Tue, 29 Sep 2020 14:36:26 +0200
Subject: [PATCH] softfloat: Implement
float128_(min|minnum|minnummag|max|maxnum|maxnummag)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Implementation inspired by minmax_floats(). Unfortuantely, we don't have any tests we can simply adjust/unlock.
--- fpu/softfloat.c | 141 ++++++++++++++++++++++++++++++++-------- include/fpu/softfloat.h | 6 ++ 2 files changed, 120 insertions(+), 27 deletions(-)
fpu/softfloat.c +114 -27
@@ -525,6 +525,18 @@ typedef struct {bool sign; } FloatParts128; +static inline FloatParts128 floatparts64_to_128 (FloatParts a) +{ +FloatParts128 res = { + .frac0 = a.frac, + .exp = a.exp, + .cls = a.cls, + .sign = a.sign, + }; + +return res; +} +#define DECOMPOSED_BINARY_POINT (64 - 2)#define DECOMPOSED_IMPLICIT_BIT (1ull << DECOMPOSED_BINARY_POINT)#define DECOMPOSED_OVERFLOW_BIT (DECOMPOSED_IMPLICIT_BIT << 1)
@@ -621,6 +633,8 @@ static inline FloatParts float64_unpack_raw(float64 f)return unpack_raw (float64_params, f); } +static void float128_unpack (FloatParts128 *p,float128 a,float_status *status); +/* Pack a float from parts, but do not canonicalize. */ static inline uint64_t pack_raw (FloatFmt fmt,FloatParts p) {
@@ -3093,6 +3107,14 @@ bfloat16 uint16_to_bfloat16(uint16_t a, float_status *status)return uint64_to_bfloat16_scalbn (a,0 , status); } +typedef enum MinMaxRes { +MINMAX_RES_A , +MINMAX_RES_B , +MINMAX_RES_SILENCE_A , +MINMAX_RES_SILENCE_B , +MINMAX_RES_DEFAULT_NAN , +}MinMaxRes ; +/* Float Min/Max */ /* min() and max() functions. These can't be implemented as * 'compare and pick one input' because that would mishandle
@@ -3109,27 +3131,36 @@ bfloat16 uint16_to_bfloat16(uint16_t a, float_status *status)* minnummag() and maxnummag() functions correspond to minNumMag() * and minNumMag() from the IEEE-754 2008. */ -static FloatParts minmax_floats (FloatParts a,FloatParts b,bool ismin, -bool ieee,bool ismag,float_status *s) +static MinMaxRes minmax_floats128 (FloatParts128 a,FloatParts128 b,bool ismin, +bool ieee,bool ismag,float_status *s) {if (unlikely (is_nan (a.cls) ||is_nan (b.cls))) { -if (ieee) { -/* Takes two floating-point values `a' and `b', one of -* which is a NaN, and returns the appropriate NaN -* result. If either `a' or `b' is a signaling NaN, -* the invalid exception is raised. -*/ -if (is_snan (a.cls) ||is_snan (b.cls)) { -return pick_nan (a, b, s); - }else if (is_nan (a.cls) && !is_nan (b.cls)) { -return b; +if (ieee && !is_snan (a.cls) && !is_snan (b.cls)) { +if (is_nan (a.cls) && !is_nan (b.cls)) { +return MINMAX_RES_B ; }else if (is_nan (b.cls) && !is_nan (a.cls)) { -return a; +return MINMAX_RES_A ; } } -return pick_nan (a, b, s); + +/* Similar logic to pick_nan(), avoiding re-packing. */ +if (is_snan (a.cls) ||is_snan (b.cls)) { + s->float_exception_flags |= float_flag_invalid; + } +if (s->default_nan_mode) { +return MINMAX_RES_DEFAULT_NAN ; + } +if (pickNaN (a.cls, b.cls, + a.frac0 > b.frac0 || + (a.frac0 == b.frac0 && a.frac1 > b.frac1) || + (a.frac0 == b.frac0 && a.frac1 == b.frac1 && + a.sign < b.sign), s)) { +return is_snan (b.cls) ?MINMAX_RES_SILENCE_B :MINMAX_RES_B ; + } +return is_snan (a.cls) ?MINMAX_RES_SILENCE_A :MINMAX_RES_A ; }else {int a_exp, b_exp; +bool a_less;switch (a.cls) {case float_class_normal:
@@ -3160,23 +3191,44 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin,break ; } -if (ismag && (a_exp != b_exp || a.frac != b.frac)) { -bool a_less = a_exp < b_exp; -if (a_exp == b_exp) { - a_less = a.frac < b.frac; + a_less = a_exp < b_exp; +if (a_exp == b_exp) { + a_less = a.frac0 < b.frac0; +if (a.frac0 == b.frac0) { + a_less = a.frac1 < b.frac1; } -return a_less ^ ismin ? b : a; } -if (a.sign == b.sign) { -bool a_less = a_exp < b_exp; -if (a_exp == b_exp) { - a_less = a.frac < b.frac; - } -return a.sign ^ a_less ^ ismin ? b : a; - }else { -return a.sign ^ ismin ? b : a; +if (ismag && + (a_exp != b_exp || a.frac0 != b.frac0 || a.frac1 != b.frac1)) { +return a_less ^ ismin ?MINMAX_RES_B :MINMAX_RES_A ; + }else if (a.sign == b.sign) { +return a.sign ^ a_less ^ ismin ?MINMAX_RES_B :MINMAX_RES_A ; } +return a.sign ^ ismin ?MINMAX_RES_B :MINMAX_RES_A ; + } +} + +static FloatParts minmax_floats (FloatParts a,FloatParts b,bool ismin, +bool ieee,bool ismag,float_status *s) +{ +FloatParts128 ta =floatparts64_to_128 (a); +FloatParts128 tb =floatparts64_to_128 (b); +MinMaxRes res =minmax_floats128 (ta, tb, ismin, ieee, ismag, s); + +switch (res) { +case MINMAX_RES_A : +return a; +case MINMAX_RES_B : +return b; +case MINMAX_RES_SILENCE_A : +return parts_silence_nan (a, s); +case MINMAX_RES_SILENCE_B : +return parts_silence_nan (b, s); +case MINMAX_RES_DEFAULT_NAN : +return parts_default_nan (s); +default : +g_assert_not_reached (); } }
@@ -3233,6 +3285,41 @@ BF16_MINMAX(maxnummag, false, true, true)#undef BF16_MINMAX +#define F128_MINMAX (name, ismin, isiee, ismag) \ +float128 float128_ ## name(float128 a, float128 b, float_status *s) \ +{ \ + FloatParts128 pa, pb; \ + MinMaxRes res; \ + \ + float128_unpack(&pa, a, s); \ + float128_unpack(&pb, b, s); \ + res = minmax_floats128(pa, pb, ismin, isiee, ismag, s); \ + \ + switch(res) { \ + case MINMAX_RES_A: \ + return a; \ + case MINMAX_RES_B: \ + return b; \ + case MINMAX_RES_SILENCE_A: \ + return float128_silence_nan(a, s); \ + case MINMAX_RES_SILENCE_B: \ + return float128_silence_nan(b, s); \ + case MINMAX_RES_DEFAULT_NAN: \ + return float128_default_nan(s); \ + default: \ + g_assert_not_reached(); \ + } \ +} + +F128_MINMAX(min, true, false, false) +F128_MINMAX(minnum, true, true, false) +F128_MINMAX(minnummag, true, true, true) +F128_MINMAX(max, false, false, false) +F128_MINMAX(maxnum, false, true, false) +F128_MINMAX(maxnummag, false, true, true) + +#undef F128_MINMAX +/* Floating point compare */ static FloatRelation compare_floats (FloatParts a,FloatParts b,bool is_quiet,float_status *s)
include/fpu/softfloat.h +6 -0
@@ -1201,6 +1201,12 @@ float128 float128_muladd(float128, float128, float128, int,float128 float128_sqrt (float128 ,float_status *status);FloatRelation float128_compare (float128 ,float128 ,float_status *status);FloatRelation float128_compare_quiet (float128 ,float128 ,float_status *status); +float128 float128_min (float128 ,float128 ,float_status *status); +float128 float128_max (float128 ,float128 ,float_status *status); +float128 float128_minnum (float128 ,float128 ,float_status *status); +float128 float128_maxnum (float128 ,float128 ,float_status *status); +float128 float128_minnummag (float128 ,float128 ,float_status *status); +float128 float128_maxnummag (float128 ,float128 ,float_status *status);bool float128_is_quiet_nan (float128 ,float_status *status);bool float128_is_signaling_nan (float128 ,float_status *status);float128 float128_silence_nan (float128 ,float_status *status);
Signature · 8 lines
2.30.2 -- Thanks, David / dhildenb
Re: [PATCH v1 01/20] softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)
Quoted 84 lines · David Hildenbrand
Quoted 83 lines · Alex Bennée
Quoted 81 lines · David Hildenbrand
Quoted 69 lines · Alex Bennée
Quoted 26 lines · David Hildenbrand
Implementation inspired by
minmax_floats(). Unfortuantely, we don't have any tests we can simply adjust/unlock.Cc Aurelien Jarno <[email protected]>Cc Peter Maydell <[email protected]>Cc "Alex Bennée" <[email protected]>Signed-off-by David Hildenbrand <[email protected]>--- fpu/softfloat.c | 100 ++++++++++++++++++++++++++++++++++++++++ include/fpu/softfloat.h | 6 +++ 2 files changed, 106 insertions(+)fpu/softfloat.c +1
-0@@ -621,6 +621,8 @@ static inline FloatParts float64_unpack_raw(float64 f)return unpack_raw (float64_params, f); } +static void float128_unpack(FloatParts128 *p, float128 a,float_status *status); +/* Pack a float from parts, but do not canonicalize. */ static inline uint64_t pack_raw (FloatFmt fmt,FloatParts p) {fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3183
Quoted 2 lines
fpu/softfloat.c +0
-0@@ -3180,6 +3182,89 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, } }It would be desirable to share as much logic for this as possible with the existing
minmax_floatscode. I appreciate at some point we end up having to deal with fractions and we haven't found a good way to efficiently handle dealing with FloatParts and FloatParts128 in the same unrolled code, however:fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3194
Quoted 11 lines
fpu/softfloat.c +9
-0+static float128 float128_minmax(float128 a, float128 b, bool ismin,bool ieee, +bool ismag,float_status *s) +{ +FloatParts128 pa, pb; +int a_exp, b_exp; +bool a_less; + +float128_unpack (&pa, a, s); +float128_unpack (&pb, b, s); +From here:
Quoted 7 lines
fpu/softfloat.c +7
-0+if (unlikely (is_nan (pa.cls) ||is_nan (pb.cls))) { +/* See comment in minmax_floats() */ +if (ieee && !is_snan (pa.cls) && !is_snan (pb.cls)) { +if (is_nan (pa.cls) && !is_nan (pb.cls)) { +return b; + }else if (is_nan (pb.cls) && !is_nan (pa.cls)) { +return a;fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3211
Quoted 10 lines
fpu/softfloat.c +10
-0+ } + } + +/* Similar logic to pick_nan(), avoiding re-packing. */ +if (is_snan (pa.cls) ||is_snan (pb.cls)) { + s->float_exception_flags |= float_flag_invalid; + } +if (s->default_nan_mode) { +return float128_default_nan (s); + }Quoted 1 line
to here is common logic - is there anyway we could share it?
I can try to factor it out, similar to
pickNaN()- passing weird boolean flags and such. It most certainly won't win in a beauty contest, that's for sure.Quoted 3 lines
Likewise I wonder if there is scope for a
float_minmax_exphelper that could be shared here?I'll try, but I have the feeling that it might make the code harder to read than actually help. Will give it a try.
Quoted 9 lines
Quoted 4 lines
Give it a try - if it really does become harder to follow then we'll stick with the duplication however if we can have common code you'll know at least the nan handling and minmax behaviour for float128 will be partially tested by the 16/32/64 float code.
(finally had time to look into this)
What about something like this:
I was just about to look this morning but I see Richard has posted his mega series:
From: Richard Henderson <[email protected]>
Subject: [PATCH 00/72] Convert floatx80 and float128 to FloatParts
Date: Fri, 7 May 2021 18:46:50 -0700
Message-Id: <[email protected]>
which I think also includes the float128_(min|minnum|minnummag|max|maxnum|maxnummag) functions. I'm going to have a look at that first if that's ok.
Quoted 234 lines
From
dd6cf176c840fc34a095cb2a158032a994aca5efMon Sep 17 00:00:00 2001
From: David Hildenbrand <[email protected]>
Date: Tue, 29 Sep 2020 14:36:26 +0200
Subject: [PATCH] softfloat: Implement
float128_(min|minnum|minnummag|max|maxnum|maxnummag)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bitImplementation inspired by
minmax_floats(). Unfortuantely, we don't have any tests we can simply adjust/unlock.Cc Aurelien Jarno <[email protected]>Cc Peter Maydell <[email protected]>Cc "Alex Bennée" <[email protected]>Cc Richard Henderson <[email protected]>Signed-off-by David Hildenbrand <[email protected]>--- fpu/softfloat.c | 141 ++++++++++++++++++++++++++++++++-------- include/fpu/softfloat.h | 6 ++ 2 files changed, 120 insertions(+), 27 deletions(-)fpu/softfloat.c +76
-25@@ -525,6 +525,18 @@ typedef struct {bool sign; } FloatParts128; +static inline FloatParts128 floatparts64_to_128(FloatParts a) +{ +FloatParts128 res = { + .frac0 = a.frac, + .exp = a.exp, + .cls = a.cls, + .sign = a.sign, + }; + +return res; +} +#define DECOMPOSED_BINARY_POINT (64 - 2)#define DECOMPOSED_IMPLICIT_BIT (1ull << DECOMPOSED_BINARY_POINT)#define DECOMPOSED_OVERFLOW_BIT (DECOMPOSED_IMPLICIT_BIT << 1)@@ -621,6 +633,8 @@ static inline FloatParts float64_unpack_raw(float64 f)return unpack_raw (float64_params, f); } +static void float128_unpack(FloatParts128 *p, float128 a,float_status *status); +/* Pack a float from parts, but do not canonicalize. */ static inline uint64_t pack_raw (FloatFmt fmt,FloatParts p) {@@ -3093,6 +3107,14 @@ bfloat16 uint16_to_bfloat16(uint16_t a, float_status *status)return uint64_to_bfloat16_scalbn (a,0 , status); } +typedef enum MinMaxRes { +MINMAX_RES_A , +MINMAX_RES_B , +MINMAX_RES_SILENCE_A , +MINMAX_RES_SILENCE_B , +MINMAX_RES_DEFAULT_NAN , +} MinMaxRes; +/* Float Min/Max */ /* min() and max() functions. These can't be implemented as * 'compare and pick one input' because that would mishandle @@ -3109,27 +3131,36 @@ bfloat16 uint16_to_bfloat16(uint16_t a, float_status *status)* minnummag() and maxnummag() functions correspond to minNumMag() * and minNumMag() from the IEEE-754 2008. */ -static FloatParts minmax_floats (FloatParts a,FloatParts b,bool ismin, -bool ieee,bool ismag,float_status *s) +static MinMaxRes minmax_floats128 (FloatParts128 a,FloatParts128 b,bool ismin, +bool ieee,bool ismag,float_status *s) {if (unlikely (is_nan (a.cls) ||is_nan (b.cls))) { -if (ieee) { -/* Takes two floating-point values `a' and `b', one of -* which is a NaN, and returns the appropriate NaN -* result. If either `a' or `b' is a signaling NaN, -* the invalid exception is raised. -*/ -if (is_snan (a.cls) ||is_snan (b.cls)) { -return pick_nan (a, b, s); - }else if (is_nan (a.cls) && !is_nan (b.cls)) { -return b; +if (ieee && !is_snan (a.cls) && !is_snan (b.cls)) { +if (is_nan (a.cls) && !is_nan (b.cls)) { +return MINMAX_RES_B ; }else if (is_nan (b.cls) && !is_nan (a.cls)) { -return a; +return MINMAX_RES_A ; } } -return pick_nan (a, b, s); + +/* Similar logic to pick_nan(), avoiding re-packing. */ +if (is_snan (a.cls) ||is_snan (b.cls)) { + s->float_exception_flags |= float_flag_invalid; + } +if (s->default_nan_mode) { +return MINMAX_RES_DEFAULT_NAN ; + } +if (pickNaN (a.cls, b.cls, + a.frac0 > b.frac0 || + (a.frac0 == b.frac0 && a.frac1 > b.frac1) || + (a.frac0 == b.frac0 && a.frac1 == b.frac1 && + a.sign < b.sign), s)) { +return is_snan (b.cls) ?MINMAX_RES_SILENCE_B :MINMAX_RES_B ; + } +return is_snan (a.cls) ?MINMAX_RES_SILENCE_A :MINMAX_RES_A ; }else {int a_exp, b_exp; +bool a_less;switch (a.cls) {case float_class_normal:@@ -3160,23 +3191,44 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin,break ; } - if (ismag && (a_exp != b_exp || a.frac != b.frac)) { -bool a_less = a_exp < b_exp; -if (a_exp == b_exp) { - a_less = a.frac < b.frac; + a_less = a_exp < b_exp; +if (a_exp == b_exp) { + a_less = a.frac0 < b.frac0; +if (a.frac0 == b.frac0) { + a_less = a.frac1 < b.frac1; } -return a_less ^ ismin ? b : a; } - if (a.sign == b.sign) { -bool a_less = a_exp < b_exp; -if (a_exp == b_exp) { - a_less = a.frac < b.frac; - } -return a.sign ^ a_less ^ ismin ? b : a; - } else { -return a.sign ^ ismin ? b : a; +if (ismag && + (a_exp != b_exp || a.frac0 != b.frac0 || a.frac1 != b.frac1)) { +return a_less ^ ismin ?MINMAX_RES_B :MINMAX_RES_A ; + }else if (a.sign == b.sign) { +return a.sign ^ a_less ^ ismin ?MINMAX_RES_B :MINMAX_RES_A ; } +return a.sign ^ ismin ?MINMAX_RES_B :MINMAX_RES_A ; + } +} + +static FloatParts minmax_floats (FloatParts a,FloatParts b,bool ismin, +bool ieee,bool ismag,float_status *s) +{ +FloatParts128 ta =floatparts64_to_128 (a); +FloatParts128 tb =floatparts64_to_128 (b); +MinMaxRes res =minmax_floats128 (ta, tb, ismin, ieee, ismag, s); + +switch (res) { +case MINMAX_RES_A : +return a; +case MINMAX_RES_B : +return b; +case MINMAX_RES_SILENCE_A : +return parts_silence_nan (a, s); +case MINMAX_RES_SILENCE_B : +return parts_silence_nan (b, s); +case MINMAX_RES_DEFAULT_NAN : +return parts_default_nan (s); +default : +g_assert_not_reached (); } } @@ -3233,6 +3285,41 @@ BF16_MINMAX(maxnummag, false, true, true) #undef BF16_MINMAX +#define F128_MINMAX(name, ismin, isiee, ismag)\
fpu/softfloat.c +34
-0+float128 float128_ ## name(float128 a, float128 b, float_status *s) \ +{ \ + FloatParts128 pa, pb; \ + MinMaxRes res; \ + \ + float128_unpack(&pa, a, s); \ + float128_unpack(&pb, b, s); \ + res = minmax_floats128(pa, pb, ismin, isiee, ismag, s); \ + \ + switch(res) { \ + case MINMAX_RES_A: \ + return a; \ + case MINMAX_RES_B: \ + return b; \ + case MINMAX_RES_SILENCE_A: \ + return float128_silence_nan(a, s); \ + case MINMAX_RES_SILENCE_B: \ + return float128_silence_nan(b, s); \ + case MINMAX_RES_DEFAULT_NAN: \ + return float128_default_nan(s); \ + default: \ + g_assert_not_reached(); \ + } \ +} + +F128_MINMAX(min, true, false, false) +F128_MINMAX(minnum, true, true, false) +F128_MINMAX(minnummag, true, true, true) +F128_MINMAX(max, false, false, false) +F128_MINMAX(maxnum, false, true, false) +F128_MINMAX(maxnummag, false, true, true) + +#undef F128_MINMAX +/* Floating point compare */ static FloatRelation compare_floats (FloatParts a,FloatParts b,bool is_quiet,float_status *s)include/fpu/softfloat.h +6
-0@@ -1201,6 +1201,12 @@ float128 float128_muladd(float128, float128, float128, int,float128 float128_sqrt (float128 ,float_status *status);FloatRelation float128_compare (float128 ,float128 ,float_status *status);FloatRelation float128_compare_quiet (float128 ,float128 ,float_status *status); +float128 float128_min (float128 ,float128 ,float_status *status); +float128 float128_max (float128 ,float128 ,float_status *status); +float128 float128_minnum (float128 ,float128 ,float_status *status); +float128 float128_maxnum (float128 ,float128 ,float_status *status); +float128 float128_minnummag (float128 ,float128 ,float_status *status); +float128 float128_maxnummag (float128 ,float128 ,float_status *status);bool float128_is_quiet_nan (float128 ,float_status *status);bool float128_is_signaling_nan (float128 ,float_status *status);float128 float128_silence_nan (float128 ,float_status *status);Signature · 1 line
2.30.2
Signature · 1 line
Alex Bennée
Re: [PATCH v1 01/20] softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)
Quoted 100 lines · Alex Bennée
Quoted 84 lines · David Hildenbrand
Quoted 83 lines · Alex Bennée
Quoted 81 lines · David Hildenbrand
Quoted 69 lines · Alex Bennée
Quoted 26 lines · David Hildenbrand
Implementation inspired by
minmax_floats(). Unfortuantely, we don't have any tests we can simply adjust/unlock.Cc Aurelien Jarno <[email protected]>Cc Peter Maydell <[email protected]>Cc "Alex Bennée" <[email protected]>Signed-off-by David Hildenbrand <[email protected]>--- fpu/softfloat.c | 100 ++++++++++++++++++++++++++++++++++++++++ include/fpu/softfloat.h | 6 +++ 2 files changed, 106 insertions(+)fpu/softfloat.c +1
-0@@ -621,6 +621,8 @@ static inline FloatParts float64_unpack_raw(float64 f)return unpack_raw (float64_params, f); } +static void float128_unpack(FloatParts128 *p, float128 a,float_status *status); +/* Pack a float from parts, but do not canonicalize. */ static inline uint64_t pack_raw (FloatFmt fmt,FloatParts p) {fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3183
Quoted 2 lines
fpu/softfloat.c +0
-0@@ -3180,6 +3182,89 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, } }It would be desirable to share as much logic for this as possible with the existing
minmax_floatscode. I appreciate at some point we end up having to deal with fractions and we haven't found a good way to efficiently handle dealing with FloatParts and FloatParts128 in the same unrolled code, however:fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3194
Quoted 11 lines
fpu/softfloat.c +9
-0+static float128 float128_minmax(float128 a, float128 b, bool ismin,bool ieee, +bool ismag,float_status *s) +{ +FloatParts128 pa, pb; +int a_exp, b_exp; +bool a_less; + +float128_unpack (&pa, a, s); +float128_unpack (&pb, b, s); +From here:
Quoted 7 lines
fpu/softfloat.c +7
-0+if (unlikely (is_nan (pa.cls) ||is_nan (pb.cls))) { +/* See comment in minmax_floats() */ +if (ieee && !is_snan (pa.cls) && !is_snan (pb.cls)) { +if (is_nan (pa.cls) && !is_nan (pb.cls)) { +return b; + }else if (is_nan (pb.cls) && !is_nan (pa.cls)) { +return a;fpu/softfloat.c static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin, · line 3211
Quoted 10 lines
fpu/softfloat.c +10
-0+ } + } + +/* Similar logic to pick_nan(), avoiding re-packing. */ +if (is_snan (pa.cls) ||is_snan (pb.cls)) { + s->float_exception_flags |= float_flag_invalid; + } +if (s->default_nan_mode) { +return float128_default_nan (s); + }Quoted 1 line
to here is common logic - is there anyway we could share it?
I can try to factor it out, similar to
pickNaN()- passing weird boolean flags and such. It most certainly won't win in a beauty contest, that's for sure.Quoted 3 lines
Likewise I wonder if there is scope for a
float_minmax_exphelper that could be shared here?I'll try, but I have the feeling that it might make the code harder to read than actually help. Will give it a try.
Quoted 9 lines
Quoted 4 lines
Give it a try - if it really does become harder to follow then we'll stick with the duplication however if we can have common code you'll know at least the nan handling and minmax behaviour for float128 will be partially tested by the 16/32/64 float code.
(finally had time to look into this)
What about something like this:
I was just about to look this morning but I see Richard has posted his mega series:
Quoted 8 lines
From: Richard Henderson <[email protected]>
Subject: [PATCH 00/72] Convert floatx80 and float128 to FloatParts
Date: Fri, 7 May 2021 18:46:50 -0700
Message-Id: <[email protected]>which I think also includes the
float128_(min|minnum|minnummag|max|maxnum|maxnummag)functions. I'm going to have a look at that first if that's ok.
Sure, I have the gut feeling that it follows a similar approach :)
Signature · 3 lines
Thanks, David / dhildenb
PATCH 2/20 s390x/tcg: Implement VECTOR BIT PERMUTE
In reply to David Hildenbrand--- target/s390x/helper.h | 1 + target/s390x/insn-data.def | 2 ++ target/s390x/translate_vx.c.inc | 8 ++++++++ target/s390x/vec_helper.c | 22 ++++++++++++++++++++++ 4 files changed, 33 insertions(+)
target/s390x/helper.h +1 -0
@@ -126,6 +126,7 @@ DEF_HELPER_FLAGS_1(stck, TCG_CALL_NO_RWG_SE, i64, env)DEF_HELPER_FLAGS_3 (probe_write_access,TCG_CALL_NO_WG , void, env, i64, i64) /* === Vector Support Instructions === */ +DEF_HELPER_FLAGS_4(gvec_vbperm, TCG_CALL_NO_RWG, void, ptr, cptr, cptr, i32) DEF_HELPER_FLAGS_4(vll, TCG_CALL_NO_WG, void, env, ptr, i64, i64) DEF_HELPER_FLAGS_4(gvec_vpk16, TCG_CALL_NO_RWG, void, ptr, cptr, cptr, i32)DEF_HELPER_FLAGS_4 (gvec_vpk32,TCG_CALL_NO_RWG , void, ptr, cptr, cptr, i32)
target/s390x/insn-data.def +2 -0
@@ -989,6 +989,8 @@ /* === Vector Support Instructions === */ +/* VECTOR BIT PERMUTE */ + E(0xe785, VBPERM, VRR_c, VE, 0, 0, 0, 0, vbperm, 0, 0, IF_VEC) /* VECTOR GATHER ELEMENT */ E(0xe713, VGEF, VRV, V, la2, 0, 0, 0, vge, 0, ES_32, IF_VEC) E(0xe712, VGEG, VRV, V, la2, 0, 0, 0, vge, 0, ES_64, IF_VEC)
target/s390x/translate_vx.c.inc +8 -0
@@ -327,6 +327,14 @@ static void gen_addi2_i64(TCGv_i64 dl, TCGv_i64 dh, TCGv_i64 al, TCGv_i64 ah, tcg_temp_free_i64(bh); } +static DisasJumpType op_vbperm(DisasContext *s, DisasOps *o) +{ + gen_gvec_3_ool(get_field(s, v1), get_field(s, v2), get_field(s, v3), 0, + gen_helper_gvec_vbperm); + + return DISAS_NEXT; +} + static DisasJumpType op_vge(DisasContext *s, DisasOps *o) { const uint8_t es = s->insn->data;
target/s390x/vec_helper.c +22 -0
@@ -19,6 +19,28 @@#include "exec/cpu_ldst.h" #include "exec/exec-all.h" +void HELPER (gvec_vbperm )(void *v1,const void *v2,const void *v3, +uint32_t desc) +{ +S390Vector tmp = {}; +uint16_t result =0 ; +int i; + +for (i =0 ; i <16 ; i++) { +const uint8_t bit_nr =s390_vec_read_element8 (v3, i); +uint16_t bit; + +if (bit_nr >=128 ) { +continue ; + } + bit = !!(s390_vec_read_element8 (v2, bit_nr /8 ) & + (0x80 >> (bit_nr %8 ))); + result |= (bit << (15 - i)); + } +s390_vec_write_element16 (&tmp,3 , result); + *(S390Vector *)v1 = tmp; +} +void HELPER (vll )(CPUS390XState *env,void *v1,uint64_t addr,uint64_t bytes) {if (likely (bytes >=16 )) {
Signature · 1 line
2.26.2
Re: [PATCH v1 02/20] s390x/tcg: Implement VECTOR BIT PERMUTE
target/s390x/vec_helper.c line 37
Quoted 2 lines · David Hildenbrand
target/s390x/vec_helper.c +2
-0+ bit = !!(s390_vec_read_element8 (v2, bit_nr /8 ) & + (0x80 >> (bit_nr %8 )));
I think this would be clearer as
bit = (s390_vec_read_element8(v2, bit_nr / 8)
>> (7 - (bit_nr % 8))) & 1;Otherwise,
r~
Re: [PATCH v1 02/20] s390x/tcg: Implement VECTOR BIT PERMUTE
Quoted 8 lines · Richard Henderson
target/s390x/vec_helper.c line 37
Quoted 2 lines · David Hildenbrand
target/s390x/vec_helper.c +2
-0+ bit = !!(s390_vec_read_element8 (v2, bit_nr /8 ) & + (0x80 >> (bit_nr %8 )));I think this would be clearer as
bit = (s390_vec_read_element8(v2, bit_nr / 8) >> (7 - (bit_nr % 8))) & 1;
Can do, thanks!
Quoted 7 lines
Otherwise,
Reviewed-by Richard Henderson <[email protected]>r~
Signature · 3 lines
Thanks, David / dhildenb
PATCH 3/20 s390x/tcg: Implement VECTOR MULTIPLY SUM LOGICAL
In reply to David HildenbrandFortunately, we only need the Doubleword implementation.
--- target/s390x/insn-data.def | 2 ++ target/s390x/translate_vx.c.inc | 52 +++++++++++++++++++++++++++++++++ 2 files changed, 54 insertions(+)
target/s390x/insn-data.def +2 -0
@@ -1151,6 +1151,8 @@ F(0xe7a7, VMO, VRR_c, V, 0, 0, 0, 0, vm, 0, IF_VEC) /* VECTOR MULTIPLY LOGICAL ODD */ F(0xe7a5, VMLO, VRR_c, V, 0, 0, 0, 0, vm, 0, IF_VEC) +/* VECTOR MULTIPLY SUM LOGICAL */ + F(0xe7b8, VMSL, VRR_d, VE, 0, 0, 0, 0, vmsl, 0, IF_VEC) /* VECTOR NAND */ F(0xe76e, VNN, VRR_c, VE, 0, 0, 0, 0, vnn, 0, IF_VEC) /* VECTOR NOR */
target/s390x/translate_vx.c.inc +52 -0
58 lines not on this page · raw message
Signature · 1 line
2.26.2
Re: [PATCH v1 03/20] s390x/tcg: Implement VECTOR MULTIPLY SUM LOGICAL
target/s390x/translate_vx.c.inc static DisasJumpType op_vm(DisasContext *s, DisasOps *o) · line 1804
Quoted 9 lines · David Hildenbrand
target/s390x/translate_vx.c.inc +9
-0+ /* Multipy both even elements from v2 and v3 */ + read_vec_element_i64(l1, get_field(s, v2), 0, ES_64); + read_vec_element_i64(h1, get_field(s, v3), 0, ES_64); + tcg_gen_mulu2_i64(l1, h1, l1, h1); + /* Shift result left by one bit if requested */ + if (extract32(get_field(s, m6), 3, 1)) { + tcg_gen_extract2_i64(h1, l1, h1, 63); + tcg_gen_shli_i64(l1, l1, 1); + }
Not a bug, but some hosts require 3 insns for extract2 (so 4 total for this sequence).
This doubling can also be had via add2:
tcg_gen_add2_i64(l1, h1, l1, h1, l1, h1);
At which point most hosts will require only 2 insns for this sequence. The two hosts that don't have a carry bit (mips, riscv), will still be able to perform the add in 3 insns.
So add is never more expensive and sometimes half as expensive.
Regardless,
r~
Re: [PATCH v1 03/20] s390x/tcg: Implement VECTOR MULTIPLY SUM LOGICAL
Quoted 11 lines · Richard Henderson
target/s390x/translate_vx.c.inc static DisasJumpType op_vm(DisasContext *s, DisasOps *o) · line 1804
Quoted 9 lines · David Hildenbrand
target/s390x/translate_vx.c.inc +9
-0+ /* Multipy both even elements from v2 and v3 */ + read_vec_element_i64(l1, get_field(s, v2), 0, ES_64); + read_vec_element_i64(h1, get_field(s, v3), 0, ES_64); + tcg_gen_mulu2_i64(l1, h1, l1, h1); + /* Shift result left by one bit if requested */ + if (extract32(get_field(s, m6), 3, 1)) { + tcg_gen_extract2_i64(h1, l1, h1, 63); + tcg_gen_shli_i64(l1, l1, 1); + }
Quoted 6 lines
Not a bug, but some hosts require 3 insns for extract2 (so 4 total for this sequence).
This doubling can also be had via add2:
tcg_gen_add2_i64(l1, h1, l1, h1, l1, h1);
Took me longer than it should to realize this is really just doubling the value ... will use tcg_gen_add2_i64() and add a comment.
Thanks!
Signature · 3 lines
Thanks, David / dhildenb
PATCH 4/20 s390x/tcg: Implement 32/128 bit for VECTOR FP ADD
In reply to David HildenbrandIn case of 128bit, we always have a single element.
Add some helpers that allow to generically read/write 32/64/128 bit floats. Convert the existing implementation of vop64_3 into a macro that deals with float* instead of uint* instead - the other users keep working as
typedef uint32_t float32;
typedef uint64_t float64;Most of them will get converted next.
--- target/s390x/helper.h | 3 + target/s390x/translate_vx.c.inc | 31 +++++++-- target/s390x/vec_fpu_helper.c | 119 +++++++++++++++++++++----------- 3 files changed, 107 insertions(+), 46 deletions(-)
target/s390x/helper.h +3 -0
@@ -247,8 +247,11 @@ DEF_HELPER_6(gvec_vstrc_cc_rt16, void, ptr, cptr, cptr, cptr, env, i32)DEF_HELPER_6 (gvec_vstrc_cc_rt32, void, ptr, cptr, cptr, cptr, env, i32) /* === Vector Floating-Point Instructions */ +DEF_HELPER_FLAGS_5(gvec_vfa32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfa32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfa64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfa64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfa128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_4(gvec_wfc64, void, cptr, cptr, env, i32) DEF_HELPER_4(gvec_wfk64, void, cptr, cptr, env, i32)DEF_HELPER_FLAGS_5 (gvec_vfce64,TCG_CALL_NO_WG , void, ptr, cptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +24 -7
@@ -2504,16 +2504,27 @@ static DisasJumpType op_vfa(DisasContext *s, DisasOps *o) const uint8_t fpf = get_field(s, m4); const uint8_t m5 = get_field(s, m5); const bool se = extract32(m5, 3, 1); - gen_helper_gvec_3_ptr *fn; - - if (fpf != FPF_LONG || extract32(m5, 0, 3)) { - gen_program_exception(s, PGM_SPECIFICATION); - return DISAS_NORETURN; - } + gen_helper_gvec_3_ptr *fn = NULL; switch (s->fields.op2) { case 0xe3: - fn = se ? gen_helper_gvec_vfa64s : gen_helper_gvec_vfa64; + switch (fpf) { + case FPF_SHORT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = se ? gen_helper_gvec_vfa32s : gen_helper_gvec_vfa32; + } + break; + case FPF_LONG: + fn = se ? gen_helper_gvec_vfa64s : gen_helper_gvec_vfa64; + break; + case FPF_EXT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_vfa128; + } + break; + default: + break; + } break; case 0xe5: fn = se ? gen_helper_gvec_vfd64s : gen_helper_gvec_vfd64;
@@ -2527,6 +2538,12 @@ static DisasJumpType op_vfa(DisasContext *s, DisasOps *o) default: g_assert_not_reached(); } + + if (!fn || extract32(m5, 0, 3)) { + gen_program_exception(s, PGM_SPECIFICATION); + return DISAS_NORETURN; + } + gen_gvec_3_ptr(get_field(s, v1), get_field(s, v2), get_field(s, v3), cpu_env, 0, fn); return DISAS_NEXT;
target/s390x/vec_fpu_helper.c +80 -39
131 lines not on this page · raw message
Signature · 1 line
2.26.2
Re: [PATCH v1 04/20] s390x/tcg: Implement 32/128 bit for VECTOR FP ADD
Quoted 22 lines · David Hildenbrand
target/s390x/vec_fpu_helper.c +0
-22-typedef uint64_t (*vop64_3_fn )(uint64_t a,uint64_t b,float_status *s); -static void vop64_3 (S390Vector *v1,const S390Vector *v2,const S390Vector *v3, -CPUS390XState *env,bool s,vop64_3_fn fn, -uintptr_t retaddr) -{ -uint8_t vxc, vec_exc =0 ; -S390Vector tmp = {}; -int i; - -for (i =0 ; i <2 ; i++) { -const uint64_t a =s390_vec_read_element64 (v2, i); -const uint64_t b =s390_vec_read_element64 (v3, i); - -s390_vec_write_element64 (&tmp, i,fn (a, b, &env->fpu_status)); - vxc =check_ieee_exc (env, i, false, &vec_exc); -if (s || vxc) { -break ; - } - } -handle_ieee_exc (env, vxc, vec_exc, retaddr); - *v1 = tmp; -}
...
Quoted 17 lines
target/s390x/vec_fpu_helper.c +17
-0+#define DEF_VOP_3 (BITS ) \ +typedef float##BITS (*vop##BITS##_3_fn)(float##BITS a, float##BITS b, \ + float_status *s); \ +static void vop##BITS##_3(S390Vector *v1, const S390Vector *v2, \ + const S390Vector *v3, CPUS390XState *env, bool s, \ + vop##BITS##_3_fn fn, uintptr_t retaddr) \ +{ \ + uint8_t vxc, vec_exc = 0; \ + S390Vector tmp = {}; \ + int i; \ + \ + for (i = 0; i < (128 / BITS); i++) { \ + const float##BITS a = s390_vec_read_float##BITS(v2, i); \ + const float##BITS b = s390_vec_read_float##BITS(v3, i); \ + \ + s390_vec_write_float##BITS(&tmp, i, fn(a, b, &env->fpu_status)); \ + vxc = check_ieee_exc(env, i, false, &vec_exc); \
target/s390x/vec_fpu_helper.c static void vop64_2(S390Vector *v1, const S390Vector *v2, CPUS390XState *env, · line 165
Quoted 10 lines
target/s390x/vec_fpu_helper.c +10
-0+if (s || vxc) { \ +break ; \ + } \ + } \ +handle_ieee_exc (env, vxc, vec_exc, retaddr); \ + *v1 = tmp; \ +} +DEF_VOP_3(32) +DEF_VOP_3(64) +DEF_VOP_3 (128 )
While this works, you won't be able to step through this function in the debugger anymore, because it now has one source line: at the point of expansion.
We do have plenty of these around the code base, I know. This is small enough that I think it's reasonable to simply have three copies, one for each type.
Quoted 8 lines
target/s390x/vec_fpu_helper.c +8
-0+#define DEF_GVEC_FVA (BITS ) \ +void HELPER(gvec_vfa##BITS)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + vop##BITS##_3(v1, v2, v3, env, false, float##BITS##_add, GETPC()); \ +} +DEF_GVEC_FVA(32) +DEF_GVEC_FVA (64 )
target/s390x/vec_fpu_helper.c static void vop64_2(S390Vector *v1, const S390Vector *v2, CPUS390XState *env, · line 184
Quoted 10 lines
target/s390x/vec_fpu_helper.c +10
-0+DEF_GVEC_FVA(128) + +#define DEF_GVEC_FVA_S (BITS ) \ +void HELPER(gvec_vfa##BITS##s)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + vop##BITS##_3(v1, v2, v3, env, true, float##BITS##_add, GETPC()); \ +} +DEF_GVEC_FVA_S(32) +DEF_GVEC_FVA_S (64 )
I think you're defining these macros with the wrong parameters. Think of how to use the same macros for all of add/sub/etc.
E.g.
#define DEF_FOP3_B(NAME, OP, BITS) \...
void HELPER(gvec_##NAME##BITS)(void *v1, const void *v2,
const void *v3, CPUS390XState *env, uint32_t desc)
{
vop##BITS##_3(v1, v2, v3, env, false,
float##BITS##_##OP, GETPC());
}
void HELPER(gvec_##NAME##BITS##s)(void *v1, const void *v2,
const void *v3, CPUS390XState *env, uint32_t desc)
{
vop##BITS##_3(v1, v2, v3, env, true,
float##BITS##_##OP, GETPC());}
#define DEF_FOP3(NAME, OP) \
DEF_FOP3_B(NAME, OP, 32) \
DEF_FOP3_B(NAME, OP, 64) \
DEF_FOP3_B(NAME, OP, 128)
DEF_FOP3(vfa, add)DEF_FOP3(vfd, div)DEF_FOP3(vfm, mul)
etc.
r~
Re: [PATCH v1 04/20] s390x/tcg: Implement 32/128 bit for VECTOR FP ADD
In reply to David Hildenbrandtarget/s390x/translate_vx.c.inc static DisasJumpType op_vfa(DisasContext *s, DisasOps *o) · line 2519
Quoted 3 lines · David Hildenbrand
target/s390x/translate_vx.c.inc +3
-0+ case FPF_LONG: + fn = se ? gen_helper_gvec_vfa64s : gen_helper_gvec_vfa64; + break;
BTW, any reason not to pass SE as data, like you do later for SQ? Or potentially the entire M field as is?
Just wondering if it would help tidy up here...
r~
Re: [PATCH v1 04/20] s390x/tcg: Implement 32/128 bit for VECTOR FP ADD
Quoted 7 lines · Richard Henderson
target/s390x/translate_vx.c.inc static DisasJumpType op_vfa(DisasContext *s, DisasOps *o) · line 2519
Quoted 3 lines · David Hildenbrand
target/s390x/translate_vx.c.inc +3
-0+ case FPF_LONG: + fn = se ? gen_helper_gvec_vfa64s : gen_helper_gvec_vfa64; + break;BTW, any reason not to pass SE as data, like you do later for SQ? Or potentially the entire M field as is?
Having a separate implementation for "se" is desirable, because the compiler can optimize-out the complete loop. If we simply pass the M field to the helper, I'm not sure how likely it is that the compiler will specialize (would have to double check).
(if we decide to remove all "s" helpers here, we'd better do it for all
helpers)
Quoted 2 lines
Just wondering if it would help tidy up here...
Signature · 3 lines
Thanks, David / dhildenb
PATCH 5/20 s390x/tcg: Implement 32/128 bit for VECTOR FP DIVIDE
In reply to David HildenbrandJust like VECTOR FP ADD.
--- target/s390x/helper.h | 3 +++ target/s390x/translate_vx.c.inc | 18 +++++++++++++++++- target/s390x/vec_fpu_helper.c | 28 +++++++++++++++------------- 3 files changed, 35 insertions(+), 14 deletions(-)
target/s390x/helper.h +3 -0
@@ -274,8 +274,11 @@ DEF_HELPER_FLAGS_4(gvec_vcgd64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vcgd64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vclgd64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vclgd64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfd32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfd32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfd64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfd64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfd128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfi64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfi64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32)DEF_HELPER_FLAGS_4 (gvec_vfll32,TCG_CALL_NO_WG , void, ptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +17 -1
@@ -2527,7 +2527,23 @@ static DisasJumpType op_vfa(DisasContext *s, DisasOps *o) } break; case 0xe5: - fn = se ? gen_helper_gvec_vfd64s : gen_helper_gvec_vfd64; + switch (fpf) { + case FPF_SHORT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = se ? gen_helper_gvec_vfd32s : gen_helper_gvec_vfd32; + } + break; + case FPF_LONG: + fn = se ? gen_helper_gvec_vfd64s : gen_helper_gvec_vfd64; + break; + case FPF_EXT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_vfd128; + } + break; + default: + break; + } break; case 0xe7: fn = se ? gen_helper_gvec_vfm64s : gen_helper_gvec_vfm64;
target/s390x/vec_fpu_helper.c +15 -13
@@ -411,22 +411,24 @@ void HELPER(gvec_vclgd64s)(void *v1, const void *v2, CPUS390XState *env,vop64_2 (v1, v2, env, true, XxC, erm, vclgd64,GETPC ()); } -static uint64_t vfd64 (uint64_t a,uint64_t b,float_status *s) -{ -return float64_div (a, b, s); -} - -void HELPER (gvec_vfd64 )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ -vop64_3 (v1, v2, v3, env, false, vfd64,GETPC ()); +#define DEF_GVEC_FVD (BITS ) \ +void HELPER(gvec_vfd##BITS)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + vop##BITS##_3(v1, v2, v3, env, false, float##BITS##_div, GETPC()); \ } +DEF_GVEC_FVD(32) +DEF_GVEC_FVD(64) +DEF_GVEC_FVD(128) -void HELPER (gvec_vfd64s )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ -vop64_3 (v1, v2, v3, env, true, vfd64,GETPC ()); +#define DEF_GVEC_FVD_S (BITS ) \ +void HELPER(gvec_vfd##BITS##s)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + vop##BITS##_3(v1, v2, v3, env, true, float##BITS##_div, GETPC()); \ } +DEF_GVEC_FVD_S(32) +DEF_GVEC_FVD_S(64)static uint64_t vfi64 (uint64_t a,float_status *s) {
Signature · 1 line
2.26.2
PATCH 6/20 s390x/tcg: Implement 32/128 bit for VECTOR FP MULTIPLY
In reply to David HildenbrandJust like VECTOR FP ADD.
--- target/s390x/helper.h | 3 +++ target/s390x/translate_vx.c.inc | 18 +++++++++++++++++- target/s390x/vec_fpu_helper.c | 28 +++++++++++++++------------- 3 files changed, 35 insertions(+), 14 deletions(-)
target/s390x/helper.h +3 -0
@@ -285,8 +285,11 @@ DEF_HELPER_FLAGS_4(gvec_vfll32, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfll32s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vflr64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vflr64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfm32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfm32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfm128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfma64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfma64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)DEF_HELPER_FLAGS_6 (gvec_vfms64,TCG_CALL_NO_WG , void, ptr, cptr, cptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +17 -1
@@ -2546,7 +2546,23 @@ static DisasJumpType op_vfa(DisasContext *s, DisasOps *o) } break; case 0xe7: - fn = se ? gen_helper_gvec_vfm64s : gen_helper_gvec_vfm64; + switch (fpf) { + case FPF_SHORT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = se ? gen_helper_gvec_vfm32s : gen_helper_gvec_vfm32; + } + break; + case FPF_LONG: + fn = se ? gen_helper_gvec_vfm64s : gen_helper_gvec_vfm64; + break; + case FPF_EXT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_vfm128; + } + break; + default: + break; + } break; case 0xe2: fn = se ? gen_helper_gvec_vfs64s : gen_helper_gvec_vfs64;
target/s390x/vec_fpu_helper.c +15 -13
@@ -531,22 +531,24 @@ void HELPER(gvec_vflr64s)(void *v1, const void *v2, CPUS390XState *env,vflr64 (v1, v2, env, true, XxC, erm,GETPC ()); } -static uint64_t vfm64 (uint64_t a,uint64_t b,float_status *s) -{ -return float64_mul (a, b, s); -} - -void HELPER (gvec_vfm64 )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ -vop64_3 (v1, v2, v3, env, false, vfm64,GETPC ()); +#define DEF_GVEC_FVM (BITS ) \ +void HELPER(gvec_vfm##BITS)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + vop##BITS##_3(v1, v2, v3, env, false, float##BITS##_mul, GETPC()); \ } +DEF_GVEC_FVM(32) +DEF_GVEC_FVM(64) +DEF_GVEC_FVM(128) -void HELPER (gvec_vfm64s )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ -vop64_3 (v1, v2, v3, env, true, vfm64,GETPC ()); +#define DEF_GVEC_FVM_S (BITS ) \ +void HELPER(gvec_vfm##BITS##s)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + vop##BITS##_3(v1, v2, v3, env, true, float##BITS##_mul, GETPC()); \ } +DEF_GVEC_FVM_S(32) +DEF_GVEC_FVM_S(64) static void vfma64(S390Vector *v1, const S390Vector *v2, const S390Vector *v3, const S390Vector *v4, CPUS390XState *env, bool s, int flags,
Signature · 1 line
2.26.2
PATCH 7/20 s390x/tcg: Implement 32/128 bit for VECTOR FP SUBTRACT
In reply to David HildenbrandJust like VECTOR FP ADD.
--- target/s390x/helper.h | 3 +++ target/s390x/translate_vx.c.inc | 18 +++++++++++++++++- target/s390x/vec_fpu_helper.c | 28 +++++++++++++++------------- 3 files changed, 35 insertions(+), 14 deletions(-)
target/s390x/helper.h +3 -0
@@ -296,8 +296,11 @@ DEF_HELPER_FLAGS_6(gvec_vfms64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env DEF_HELPER_FLAGS_6(gvec_vfms64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfsq64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfsq64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfs32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfs32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfs128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_4(gvec_vftci64, void, ptr, cptr, env, i32)DEF_HELPER_4 (gvec_vftci64s, void, ptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +17 -1
@@ -2565,7 +2565,23 @@ static DisasJumpType op_vfa(DisasContext *s, DisasOps *o) } break; case 0xe2: - fn = se ? gen_helper_gvec_vfs64s : gen_helper_gvec_vfs64; + switch (fpf) { + case FPF_SHORT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = se ? gen_helper_gvec_vfs32s : gen_helper_gvec_vfs32; + } + break; + case FPF_LONG: + fn = se ? gen_helper_gvec_vfs64s : gen_helper_gvec_vfs64; + break; + case FPF_EXT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_vfs128; + } + break; + default: + break; + } break; default: g_assert_not_reached();
target/s390x/vec_fpu_helper.c +15 -13
@@ -615,22 +615,24 @@ void HELPER(gvec_vfsq64s)(void *v1, const void *v2, CPUS390XState *env,vop64_2 (v1, v2, env, true, false,0 , vfsq64,GETPC ()); } -static uint64_t vfs64 (uint64_t a,uint64_t b,float_status *s) -{ -return float64_sub (a, b, s); -} - -void HELPER (gvec_vfs64 )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ -vop64_3 (v1, v2, v3, env, false, vfs64,GETPC ()); +#define DEF_GVEC_FVS (BITS ) \ +void HELPER(gvec_vfs##BITS)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + vop##BITS##_3(v1, v2, v3, env, false, float##BITS##_sub, GETPC()); \ } +DEF_GVEC_FVS(32) +DEF_GVEC_FVS(64) +DEF_GVEC_FVS(128) -void HELPER (gvec_vfs64s )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ -vop64_3 (v1, v2, v3, env, true, vfs64,GETPC ()); +#define DEF_GVEC_FVS_S (BITS ) \ +void HELPER(gvec_vfs##BITS##s)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + vop##BITS##_3(v1, v2, v3, env, true, float##BITS##_sub, GETPC()); \ } +DEF_GVEC_FVS_S(32) +DEF_GVEC_FVS_S(64)static int vftci64 (S390Vector *v1,const S390Vector *v2,CPUS390XState *env,bool s,uint16_t i3)
Signature · 1 line
2.26.2
PATCH 8/20 s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE (AND SIGNAL) SCALAR
In reply to David Hildenbrand--- target/s390x/helper.h | 4 +++ target/s390x/translate_vx.c.inc | 38 +++++++++++++++----- target/s390x/vec_fpu_helper.c | 64 +++++++++++++++++++-------------- 3 files changed, 72 insertions(+), 34 deletions(-)
target/s390x/helper.h +4 -0
@@ -252,8 +252,12 @@ DEF_HELPER_FLAGS_5(gvec_vfa32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfa64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfa64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfa128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_4(gvec_wfc32, void, cptr, cptr, env, i32) +DEF_HELPER_4(gvec_wfk32, void, cptr, cptr, env, i32) DEF_HELPER_4(gvec_wfc64, void, cptr, cptr, env, i32) DEF_HELPER_4(gvec_wfk64, void, cptr, cptr, env, i32) +DEF_HELPER_4(gvec_wfc128, void, cptr, cptr, env, i32) +DEF_HELPER_4(gvec_wfk128, void, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfce64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfce64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)DEF_HELPER_5 (gvec_vfce64_cc, void, ptr, cptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +30 -8
@@ -2601,19 +2601,41 @@ static DisasJumpType op_wfc(DisasContext *s, DisasOps *o) { const uint8_t fpf = get_field(s, m3); const uint8_t m4 = get_field(s, m4); + gen_helper_gvec_2_ptr *fn = NULL; - if (fpf != FPF_LONG || m4) { + switch (fpf) { + case FPF_SHORT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_wfk32; + if (s->fields.op2 == 0xcb) { + fn = gen_helper_gvec_wfc32; + } + } + break; + case FPF_LONG: + fn = gen_helper_gvec_wfk64; + if (s->fields.op2 == 0xcb) { + fn = gen_helper_gvec_wfc64; + } + break; + case FPF_EXT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_wfk128; + if (s->fields.op2 == 0xcb) { + fn = gen_helper_gvec_wfc128; + } + } + break; + default: + break; + }; + + if (!fn || m4) { gen_program_exception(s, PGM_SPECIFICATION); return DISAS_NORETURN; } - if (s->fields.op2 == 0xcb) { - gen_gvec_2_ptr(get_field(s, v1), get_field(s, v2), - cpu_env, 0, gen_helper_gvec_wfc64); - } else { - gen_gvec_2_ptr(get_field(s, v1), get_field(s, v2), - cpu_env, 0, gen_helper_gvec_wfk64); - } + gen_gvec_2_ptr(get_field(s, v1), get_field(s, v2), cpu_env, 0, fn); set_cc_static(s); return DISAS_NEXT; }
target/s390x/vec_fpu_helper.c +38 -26
75 lines not on this page · raw message
Signature · 1 line
2.26.2
Re: [PATCH v1 08/20] s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE (AND SIGNAL) SCALAR
target/s390x/translate_vx.c.inc static DisasJumpType op_wfc(DisasContext *s, DisasOps *o) · line 2610
Quoted 12 lines · David Hildenbrand
target/s390x/translate_vx.c.inc +6
-1@@ -2601,19 +2601,41 @@ static DisasJumpType op_wfc(DisasContext *s, DisasOps *o) { const uint8_t fpf = get_field(s, m3); const uint8_t m4 = get_field(s, m4); + gen_helper_gvec_2_ptr *fn = NULL; - if (fpf != FPF_LONG || m4) { + switch (fpf) { + case FPF_SHORT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_wfk32; + if (s->fields.op2 == 0xcb) {
Hoist and name this comparision (e.g. bool signal = ...).
Quoted 31 lines
target/s390x/vec_fpu_helper.c +13
-18-static int wfc64 (const S390Vector *v1,const S390Vector *v2, -CPUS390XState *env,bool signal,uintptr_t retaddr) -{ -/* only the zero-indexed elements are compared */ -const float64 a =s390_vec_read_element64 (v1,0 ); -const float64 b =s390_vec_read_element64 (v2,0 ); -uint8_t vxc, vec_exc =0 ; -int cmp; - -if (signal) { - cmp =float64_compare (a, b, &env->fpu_status); - }else { - cmp =float64_compare_quiet (a, b, &env->fpu_status); - } - vxc =check_ieee_exc (env,0 , false, &vec_exc); -handle_ieee_exc (env, vxc, vec_exc, retaddr); - -return float_comp_to_cc (env, cmp); +#define DEF_WFC (BITS ) \ +static int wfc##BITS(const S390Vector *v1, const S390Vector *v2, \ + CPUS390XState *env, bool signal, uintptr_t retaddr) \ +{ \ +/* only the zero-indexed elements are compared */ \ + const float##BITS a = s390_vec_read_float##BITS(v1, 0); \ + const float##BITS b = s390_vec_read_float##BITS(v2, 0); \ + uint8_t vxc, vec_exc = 0; \ + int cmp; \ + \ + if (signal) { \ + cmp = float##BITS##_compare(a, b, &env->fpu_status); \ + } else { \
target/s390x/vec_fpu_helper.c void HELPER(gvec_vfa##BITS##s)(void *v1, const void *v2, const void *v3, \ · line 208
Quoted 10 lines
target/s390x/vec_fpu_helper.c +9
-0+ cmp = float##BITS##_compare_quiet(a, b, &env->fpu_status); \ + } \ + vxc = check_ieee_exc(env, 0, false, &vec_exc); \ + handle_ieee_exc(env, vxc, vec_exc, retaddr); \ + \ + return float_comp_to_cc(env, cmp); \ } +DEF_WFC(32) +DEF_WFC(64) +DEF_WFC (128 )
So, same issue here vs debugging.
If you keep this macroized, I don't see the value in two levels of macros...
Quoted 9 lines
target/s390x/vec_fpu_helper.c +8
-0+#define DEF_GVEC_WFC (BITS ) \ +void HELPER(gvec_wfc##BITS)(const void *v1, const void *v2, CPUS390XState *env,\ + uint32_t desc) \ +{ \ + env->cc_op = wfc##BITS(v1, v2, env, false, GETPC()); \ } +DEF_GVEC_WFC(32) +DEF_GVEC_WFC(64) +DEF_GVEC_WFC (128 )
target/s390x/vec_fpu_helper.c void HELPER(gvec_vfa##BITS##s)(void *v1, const void *v2, const void *v3, \ · line 228
Quoted 10 lines
target/s390x/vec_fpu_helper.c +8
-0+#define DEF_GVEC_WFK (BITS ) \ +void HELPER(gvec_wfk##BITS)(const void *v1, const void *v2, CPUS390XState *env,\ + uint32_t desc) \ +{ \ + env->cc_op = wfc##BITS(v1, v2, env, true, GETPC()); \ } +DEF_GVEC_WFK(32) +DEF_GVEC_WFK(64) +DEF_GVEC_WFK (128 )
These could be folded in to the first macro via parameters.
r~
PATCH 9/20 s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE *
In reply to David HildenbrandIn addition to 32/128bit variants, we also have to support the "Signal-on-QNaN (SQ)" bit ... let's pass it as a simple flag, I don't feel like duplicating all helpers and coming up with names like "...s_cc_sq".
--- target/s390x/helper.h | 18 +++ target/s390x/translate_vx.c.inc | 91 +++++++++--- target/s390x/vec_fpu_helper.c | 250 ++++++++++++++++++++++---------- 3 files changed, 258 insertions(+), 101 deletions(-)
target/s390x/helper.h +18 -0
@@ -258,18 +258,36 @@ DEF_HELPER_4(gvec_wfc64, void, cptr, cptr, env, i32) DEF_HELPER_4(gvec_wfk64, void, cptr, cptr, env, i32) DEF_HELPER_4(gvec_wfc128, void, cptr, cptr, env, i32) DEF_HELPER_4(gvec_wfk128, void, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfce32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfce32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfce64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfce64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfce128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_5(gvec_vfce32_cc, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_5(gvec_vfce32s_cc, void, ptr, cptr, cptr, env, i32) DEF_HELPER_5(gvec_vfce64_cc, void, ptr, cptr, cptr, env, i32) DEF_HELPER_5(gvec_vfce64s_cc, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_5(gvec_vfce128_cc, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfch32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfch32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfch64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfch64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfch128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_5(gvec_vfch32_cc, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_5(gvec_vfch32s_cc, void, ptr, cptr, cptr, env, i32) DEF_HELPER_5(gvec_vfch64_cc, void, ptr, cptr, cptr, env, i32) DEF_HELPER_5(gvec_vfch64s_cc, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_5(gvec_vfch128_cc, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfche32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfche32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfche64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfche64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfche128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_5(gvec_vfche32_cc, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_5(gvec_vfche32s_cc, void, ptr, cptr, cptr, env, i32) DEF_HELPER_5(gvec_vfche64_cc, void, ptr, cptr, cptr, env, i32) DEF_HELPER_5(gvec_vfche64s_cc, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_5(gvec_vfche128_cc, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vcdg64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vcdg64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32)DEF_HELPER_FLAGS_4 (gvec_vcdlg64,TCG_CALL_NO_WG , void, ptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +68 -23
113 lines not on this page · raw message
target/s390x/vec_fpu_helper.c +172 -78
287 lines not on this page · raw message
Signature · 1 line
2.26.2
Re: [PATCH v1 09/20] s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE *
Quoted 57 lines · David Hildenbrand
target/s390x/vec_fpu_helper.c +35
-16-void HELPER (gvec_vfce64 )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ -vfc64 (v1, v2, v3, env, false, float64_eq_quiet,GETPC ()); +#define DEF_GVEC_VFCE (BITS ) \ +void HELPER(gvec_vfce##BITS)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + const bool sq = simd_data(desc); \ + \ + vfc##BITS(v1, v2, v3, env, false, \ + sq ? float##BITS##_eq : float##BITS##_eq_quiet, GETPC()); \ } +DEF_GVEC_VFCE(32) +DEF_GVEC_VFCE(64) +DEF_GVEC_VFCE(128) -void HELPER (gvec_vfce64s )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ -vfc64 (v1, v2, v3, env, true, float64_eq_quiet,GETPC ()); +#define DEF_GVEC_VFCE_S (BITS ) \ +void HELPER(gvec_vfce##BITS##s)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + const bool sq = simd_data(desc); \ + \ + vfc##BITS(v1, v2, v3, env, true, \ + sq ? float##BITS##_eq : float##BITS##_eq_quiet, GETPC()); \ } +DEF_GVEC_VFCE_S(32) +DEF_GVEC_VFCE_S(64) -void HELPER (gvec_vfce64_cc )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ - env->cc_op =vfc64 (v1, v2, v3, env, false, float64_eq_quiet,GETPC ()); +#define DEF_GVEC_VFCE_CC (BITS ) \ +void HELPER(gvec_vfce##BITS##_cc)(void *v1, const void *v2, const void *v3, \ + CPUS390XState *env, uint32_t desc) \ +{ \ + const bool sq = simd_data(desc); \ + \ + env->cc_op = vfc##BITS(v1, v2, v3, env, false, \ + sq ? float##BITS##_eq : float##BITS##_eq_quiet, \ + GETPC()); \ } +DEF_GVEC_VFCE_CC(32) +DEF_GVEC_VFCE_CC(64) +DEF_GVEC_VFCE_CC(128) -void HELPER (gvec_vfce64s_cc )(void *v1,const void *v2,const void *v3, -CPUS390XState *env,uint32_t desc) -{ - env->cc_op =vfc64 (v1, v2, v3, env, true, float64_eq_quiet,GETPC ()); +#define DEF_GVEC_VFCE_S_CC (BITS ) \ +void HELPER(gvec_vfce##BITS##s_cc)(void *v1, const void *v2, const void *v3, \
target/s390x/vec_fpu_helper.c DEF_GVEC_WFK(32) · line 322
Quoted 10 lines
target/s390x/vec_fpu_helper.c +9
-0+ CPUS390XState *env, uint32_t desc) \ +{ \ +const bool sq =simd_data (desc); \ + \ + env->cc_op = vfc##BITS(v1, v2, v3, env, true, \ + sq ? float##BITS##_eq : float##BITS##_eq_quiet, \ + GETPC()); \ } +DEF_GVEC_VFCE_S_CC(32) +DEF_GVEC_VFCE_S_CC (64 )
These macros are at the wrong level. You shouldn't need separate macros for EQ vs LT, etc.
r~
PATCH 10/20 s390x/tcg: Implement 32/128 bit for VECTOR LOAD FP INTEGER
In reply to David HildenbrandConvert vop64_2 into a macro, similar to already done with vop64_3.
--- target/s390x/helper.h | 3 ++ target/s390x/translate_vx.c.inc | 51 ++++++++++++++----- target/s390x/vec_fpu_helper.c | 90 ++++++++++++++++++--------------- 3 files changed, 91 insertions(+), 53 deletions(-)
target/s390x/helper.h +3 -0
@@ -301,8 +301,11 @@ DEF_HELPER_FLAGS_5(gvec_vfd32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfd64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfd64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfd128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_4(gvec_vfi32, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) +DEF_HELPER_FLAGS_4(gvec_vfi32s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfi64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfi64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) +DEF_HELPER_FLAGS_4(gvec_vfi128, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfll32, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfll32s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32)DEF_HELPER_FLAGS_4 (gvec_vflr64,TCG_CALL_NO_WG , void, ptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +39 -12
74 lines not on this page · raw message
target/s390x/vec_fpu_helper.c +49 -41
106 lines not on this page · raw message
Signature · 1 line
2.26.2
PATCH 11/20 s390x/tcg: Implement 64 bit for VECTOR FP LOAD LENGTHENED
In reply to David Hildenbrand64 bit -> 128 bit, there is only a single final element.
--- target/s390x/helper.h | 1 + target/s390x/translate_vx.c.inc | 21 ++++++++++++++++----- target/s390x/vec_fpu_helper.c | 13 +++++++++++++ 3 files changed, 30 insertions(+), 5 deletions(-)
target/s390x/helper.h +1 -0
@@ -308,6 +308,7 @@ DEF_HELPER_FLAGS_4(gvec_vfi64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfi128, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfll32, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfll32s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) +DEF_HELPER_FLAGS_4(gvec_vfll64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vflr64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vflr64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32)DEF_HELPER_FLAGS_5 (gvec_vfm32,TCG_CALL_NO_WG , void, ptr, cptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +16 -5
@@ -2807,16 +2807,27 @@ static DisasJumpType op_vfll(DisasContext *s, DisasOps *o) { const uint8_t fpf = get_field(s, m3); const uint8_t m4 = get_field(s, m4); - gen_helper_gvec_2_ptr *fn = gen_helper_gvec_vfll32; + const bool se = extract32(m4, 3, 1); + gen_helper_gvec_2_ptr *fn = NULL; - if (fpf != FPF_SHORT || extract32(m4, 0, 3)) { + switch (fpf) { + case FPF_SHORT: + fn = se ? gen_helper_gvec_vfll32s : gen_helper_gvec_vfll32; + break; + case FPF_LONG: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_vfll64; + } + break; + default: + break; + } + + if (!fn || extract32(m4, 0, 3)) { gen_program_exception(s, PGM_SPECIFICATION); return DISAS_NORETURN; } - if (extract32(m4, 3, 1)) { - fn = gen_helper_gvec_vfll32s; - } gen_gvec_2_ptr(get_field(s, v1), get_field(s, v2), cpu_env, 0, fn); return DISAS_NEXT;
target/s390x/vec_fpu_helper.c +13 -0
@@ -602,6 +602,19 @@ void HELPER(gvec_vfll32s)(void *v1, const void *v2, CPUS390XState *env,vfll32 (v1, v2, env, true,GETPC ()); } +void HELPER (gvec_vfll64 )(void *v1,const void *v2,CPUS390XState *env, +uint32_t desc) +{ +/* load from even element */ +float128 ret =float64_to_float128 (s390_vec_read_float64 (v2,0 ), + &env->fpu_status); +uint8_t vxc, vec_exc =0 ; + + vxc =check_ieee_exc (env,0 , false, &vec_exc); +handle_ieee_exc (env, vxc, vec_exc,GETPC ()); +s390_vec_write_float128 (v1,0 , ret); +} +static void vflr64 (S390Vector *v1,const S390Vector *v2,CPUS390XState *env,bool s,bool XxC,uint8_t erm,uintptr_t retaddr) {
Signature · 1 line
2.26.2
Re: [PATCH v1 11/20] s390x/tcg: Implement 64 bit for VECTOR FP LOAD LENGTHENED
Quoted 8 lines · David Hildenbrand
64 bit -> 128 bit, there is only a single final element.
Signed-off-by David Hildenbrand <[email protected]>--- target/s390x/helper.h | 1 + target/s390x/translate_vx.c.inc | 21 ++++++++++++++++----- target/s390x/vec_fpu_helper.c | 13 +++++++++++++ 3 files changed, 30 insertions(+), 5 deletions(-)
r~
PATCH 12/20 s390x/tcg: Implement 128 bit for VECTOR FP LOAD ROUNDED
In reply to David Hildenbrand128 bit -> 64 bit, there is only a single element to process.
--- target/s390x/helper.h | 1 + target/s390x/translate_vx.c.inc | 11 ++++++++++- target/s390x/vec_fpu_helper.c | 19 +++++++++++++++++++ 3 files changed, 30 insertions(+), 1 deletion(-)
target/s390x/helper.h +1 -0
@@ -311,6 +311,7 @@ DEF_HELPER_FLAGS_4(gvec_vfll32s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfll64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vflr64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vflr64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) +DEF_HELPER_FLAGS_4(gvec_vflr128, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)DEF_HELPER_FLAGS_5 (gvec_vfm64,TCG_CALL_NO_WG , void, ptr, cptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +10 -1
@@ -2785,8 +2785,17 @@ static DisasJumpType op_vcdg(DisasContext *s, DisasOps *o) } break; case 0xc5: - if (fpf == FPF_LONG) { + switch (fpf) { + case FPF_LONG: fn = se ? gen_helper_gvec_vflr64s : gen_helper_gvec_vflr64; + break; + case FPF_EXT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_vflr128; + } + break; + default: + break; } break; default:
target/s390x/vec_fpu_helper.c +19 -0
@@ -658,6 +658,25 @@ void HELPER(gvec_vflr64s)(void *v1, const void *v2, CPUS390XState *env,vflr64 (v1, v2, env, true, XxC, erm,GETPC ()); } +void HELPER (gvec_vflr128 )(void *v1,const void *v2,CPUS390XState *env, +uint32_t desc) +{ +const uint8_t erm =extract32 (simd_data (desc),4 ,4 ); +const bool XxC =extract32 (simd_data (desc),2 ,1 ); +uint8_t vxc, vec_exc =0 ; +int old_mode; +float64 ret; + + old_mode =s390_swap_bfp_rounding_mode (env, erm); + ret =float128_to_float64 (s390_vec_read_float128 (v2,0 ), &env->fpu_status); + vxc =check_ieee_exc (env,0 , XxC, &vec_exc); +s390_restore_bfp_rounding_mode (env, old_mode); +handle_ieee_exc (env, vxc, vec_exc,GETPC ()); + +/* place at even element, odd element is unpredictable */ +s390_vec_write_float64 (v1,0 , ret); +} +#define DEF_GVEC_FVM (BITS ) \ void HELPER(gvec_vfm##BITS)(void *v1, const void *v2, const void *v3, \ CPUS390XState *env, uint32_t desc) \
Signature · 1 line
2.26.2
Re: [PATCH v1 12/20] s390x/tcg: Implement 128 bit for VECTOR FP LOAD ROUNDED
Quoted 8 lines · David Hildenbrand
128 bit -> 64 bit, there is only a single element to process.
Signed-off-by David Hildenbrand <[email protected]>--- target/s390x/helper.h | 1 + target/s390x/translate_vx.c.inc | 11 ++++++++++- target/s390x/vec_fpu_helper.c | 19 +++++++++++++++++++ 3 files changed, 30 insertions(+), 1 deletion(-)
r~
PATCH 13/20 s390x/tcg: Implement 32/128 bit for VECTOR FP PERFORM SIGN OPERATION
In reply to David Hildenbrand--- target/s390x/translate_vx.c.inc | 100 +++++++++++++++++++++----------- 1 file changed, 67 insertions(+), 33 deletions(-)
target/s390x/translate_vx.c.inc +67 -33
@@ -2872,48 +2872,82 @@ static DisasJumpType op_vfpso(DisasContext *s, DisasOps *o) const uint8_t fpf = get_field(s, m3); const uint8_t m4 = get_field(s, m4); const uint8_t m5 = get_field(s, m5); + const bool se = extract32(m4, 3, 1); TCGv_i64 tmp; - if (fpf != FPF_LONG || extract32(m4, 0, 3) || m5 > 2) { + if ((fpf != FPF_LONG && !s390_has_feat(S390_FEAT_VECTOR_ENH)) || + extract32(m4, 0, 3) || m5 > 2) { gen_program_exception(s, PGM_SPECIFICATION); return DISAS_NORETURN; } - if (extract32(m4, 3, 1)) { - tmp = tcg_temp_new_i64(); - read_vec_element_i64(tmp, v2, 0, ES_64); - switch (m5) { - case 0: - /* sign bit is inverted (complement) */ - tcg_gen_xori_i64(tmp, tmp, 1ull << 63); - break; - case 1: - /* sign bit is set to one (negative) */ - tcg_gen_ori_i64(tmp, tmp, 1ull << 63); - break; - case 2: - /* sign bit is set to zero (positive) */ - tcg_gen_andi_i64(tmp, tmp, (1ull << 63) - 1); - break; + switch (fpf) { + case FPF_SHORT: + if (!se) { + switch (m5) { + case 0: + /* sign bit is inverted (complement) */ + gen_gvec_fn_2i(xori, ES_32, v1, v2, 1ull << 31); + break; + case 1: + /* sign bit is set to one (negative) */ + gen_gvec_fn_2i(ori, ES_32, v1, v2, 1ull << 31); + break; + case 2: + /* sign bit is set to zero (positive) */ + gen_gvec_fn_2i(andi, ES_32, v1, v2, (1ull << 31) - 1); + break; + } + return DISAS_NEXT; } - write_vec_element_i64(tmp, v1, 0, ES_64); - tcg_temp_free_i64(tmp); - } else { - switch (m5) { - case 0: - /* sign bit is inverted (complement) */ - gen_gvec_fn_2i(xori, ES_64, v1, v2, 1ull << 63); - break; - case 1: - /* sign bit is set to one (negative) */ - gen_gvec_fn_2i(ori, ES_64, v1, v2, 1ull << 63); - break; - case 2: - /* sign bit is set to zero (positive) */ - gen_gvec_fn_2i(andi, ES_64, v1, v2, (1ull << 63) - 1); - break; + break; + case FPF_LONG: + if (!se) { + switch (m5) { + case 0: + /* sign bit is inverted (complement) */ + gen_gvec_fn_2i(xori, ES_64, v1, v2, 1ull << 63); + break; + case 1: + /* sign bit is set to one (negative) */ + gen_gvec_fn_2i(ori, ES_64, v1, v2, 1ull << 63); + break; + case 2: + /* sign bit is set to zero (positive) */ + gen_gvec_fn_2i(andi, ES_64, v1, v2, (1ull << 63) - 1); + break; + } + return DISAS_NEXT; } + break; + case FPF_EXT: + /* Only a single element. */ + break; + default: + gen_program_exception(s, PGM_SPECIFICATION); + return DISAS_NORETURN; + } + + /* With a single element, we are only interested in bit 0. */ + tmp = tcg_temp_new_i64(); + read_vec_element_i64(tmp, v2, 0, ES_64); + switch (m5) { + case 0: + /* sign bit is inverted (complement) */ + tcg_gen_xori_i64(tmp, tmp, 1ull << 63); + break; + case 1: + /* sign bit is set to one (negative) */ + tcg_gen_ori_i64(tmp, tmp, 1ull << 63); + break; + case 2: + /* sign bit is set to zero (positive) */ + tcg_gen_andi_i64(tmp, tmp, (1ull << 63) - 1); + break; } + write_vec_element_i64(tmp, v1, 0, ES_64); + tcg_temp_free_i64(tmp); + return DISAS_NEXT; }
Signature · 1 line
2.26.2
Re: [PATCH v1 13/20] s390x/tcg: Implement 32/128 bit for VECTOR FP PERFORM SIGN OPERATION
Quoted 4 lines · David Hildenbrand
Signed-off-by David Hildenbrand <[email protected]>--- target/s390x/translate_vx.c.inc | 100 +++++++++++++++++++++----------- 1 file changed, 67 insertions(+), 33 deletions(-)
r~
PATCH 14/20 s390x/tcg: Implement 32/128 bit for VECTOR FP SQUARE ROOT
In reply to David Hildenbrand--- target/s390x/helper.h | 3 +++ target/s390x/translate_vx.c.inc | 26 +++++++++++++++++++++----- target/s390x/vec_fpu_helper.c | 28 +++++++++++++++------------- 3 files changed, 39 insertions(+), 18 deletions(-)
target/s390x/helper.h +3 -0
@@ -321,8 +321,11 @@ DEF_HELPER_FLAGS_6(gvec_vfma64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env DEF_HELPER_FLAGS_6(gvec_vfma64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfms64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfms64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_4(gvec_vfsq32, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) +DEF_HELPER_FLAGS_4(gvec_vfsq32s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfsq64, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfsq64s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) +DEF_HELPER_FLAGS_4(gvec_vfsq128, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)DEF_HELPER_FLAGS_5 (gvec_vfs64,TCG_CALL_NO_WG , void, ptr, cptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +21 -5
@@ -2955,16 +2955,32 @@ static DisasJumpType op_vfsq(DisasContext *s, DisasOps *o) { const uint8_t fpf = get_field(s, m3); const uint8_t m4 = get_field(s, m4); - gen_helper_gvec_2_ptr *fn = gen_helper_gvec_vfsq64; + const bool se = extract32(m4, 3, 1); + gen_helper_gvec_2_ptr *fn = NULL; - if (fpf != FPF_LONG || extract32(m4, 0, 3)) { + switch (fpf) { + case FPF_SHORT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = se ? gen_helper_gvec_vfsq32s : gen_helper_gvec_vfsq32; + } + break; + case FPF_LONG: + fn = se ? gen_helper_gvec_vfsq64s : gen_helper_gvec_vfsq64; + break; + case FPF_EXT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_vfsq128; + } + break; + default: + break; + } + + if (!fn || extract32(m4, 0, 3)) { gen_program_exception(s, PGM_SPECIFICATION); return DISAS_NORETURN; } - if (extract32(m4, 3, 1)) { - fn = gen_helper_gvec_vfsq64s; - } gen_gvec_2_ptr(get_field(s, v1), get_field(s, v2), cpu_env, 0, fn); return DISAS_NEXT;
target/s390x/vec_fpu_helper.c +15 -13
@@ -744,22 +744,24 @@ void HELPER(gvec_vfms64s)(void *v1, const void *v2, const void *v3,vfma64 (v1, v2, v3, v4, env, true, float_muladd_negate_c,GETPC ()); } -static uint64_t vfsq64 (uint64_t a,float_status *s) -{ -return float64_sqrt (a, s); -} - -void HELPER (gvec_vfsq64 )(void *v1,const void *v2,CPUS390XState *env, -uint32_t desc) -{ -vop64_2 (v1, v2, env, false, false,0 , vfsq64,GETPC ()); +#define DEF_GVEC_VFSQ (BITS ) \ +void HELPER(gvec_vfsq##BITS)(void *v1, const void *v2, CPUS390XState *env, \ + uint32_t desc) \ +{ \ + vop##BITS##_2(v1, v2, env, false, false, 0, float##BITS##_sqrt, GETPC()); \ } +DEF_GVEC_VFSQ(32) +DEF_GVEC_VFSQ(64) +DEF_GVEC_VFSQ(128) -void HELPER (gvec_vfsq64s )(void *v1,const void *v2,CPUS390XState *env, -uint32_t desc) -{ -vop64_2 (v1, v2, env, true, false,0 , vfsq64,GETPC ()); +#define DEF_GVEC_VFSQ_S (BITS ) \ +void HELPER(gvec_vfsq##BITS##s)(void *v1, const void *v2, CPUS390XState *env, \ + uint32_t desc) \ +{ \ + vop##BITS##_2(v1, v2, env, true, false, 0, float##BITS##_sqrt, GETPC()); \ } +DEF_GVEC_VFSQ_S(32) +DEF_GVEC_VFSQ_S(64)#define DEF_GVEC_FVS (BITS ) \ void HELPER(gvec_vfs##BITS)(void *v1, const void *v2, const void *v3, \
Signature · 1 line
2.26.2
PATCH 15/20 s390x/tcg: Implement 32/128 bit for VECTOR FP TEST DATA CLASS IMMEDIATE
In reply to David Hildenbrand--- target/s390x/helper.h | 3 ++ target/s390x/translate_vx.c.inc | 26 ++++++++--- target/s390x/vec_fpu_helper.c | 76 +++++++++++++++++++-------------- 3 files changed, 69 insertions(+), 36 deletions(-)
target/s390x/helper.h +3 -0
@@ -331,8 +331,11 @@ DEF_HELPER_FLAGS_5(gvec_vfs32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_4(gvec_vftci32, void, ptr, cptr, env, i32) +DEF_HELPER_4(gvec_vftci32s, void, ptr, cptr, env, i32) DEF_HELPER_4(gvec_vftci64, void, ptr, cptr, env, i32) DEF_HELPER_4(gvec_vftci64s, void, ptr, cptr, env, i32) +DEF_HELPER_4(gvec_vftci128, void, ptr, cptr, env, i32)#ifndef CONFIG_USER_ONLY DEF_HELPER_3 (servc, i32, env, i64, i64)
target/s390x/translate_vx.c.inc +21 -5
@@ -2991,16 +2991,32 @@ static DisasJumpType op_vftci(DisasContext *s, DisasOps *o) const uint16_t i3 = get_field(s, i3); const uint8_t fpf = get_field(s, m4); const uint8_t m5 = get_field(s, m5); - gen_helper_gvec_2_ptr *fn = gen_helper_gvec_vftci64; + const bool se = extract32(m5, 3, 1); + gen_helper_gvec_2_ptr *fn = NULL; - if (fpf != FPF_LONG || extract32(m5, 0, 3)) { + switch (fpf) { + case FPF_SHORT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = se ? gen_helper_gvec_vftci32s : gen_helper_gvec_vftci32; + } + break; + case FPF_LONG: + fn = se ? gen_helper_gvec_vftci64s : gen_helper_gvec_vftci64; + break; + case FPF_EXT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_vftci128; + } + break; + default: + break; + } + + if (!fn || extract32(m5, 0, 3)) { gen_program_exception(s, PGM_SPECIFICATION); return DISAS_NORETURN; } - if (extract32(m5, 3, 1)) { - fn = gen_helper_gvec_vftci64s; - } gen_gvec_2_ptr(get_field(s, v1), get_field(s, v2), cpu_env, i3, fn); set_cc_static(s); return DISAS_NEXT;
target/s390x/vec_fpu_helper.c +45 -31
90 lines not on this page · raw message
Signature · 1 line
2.26.2
Re: [PATCH v1 15/20] s390x/tcg: Implement 32/128 bit for VECTOR FP TEST DATA CLASS IMMEDIATE
Quoted 70 lines · David Hildenbrand
Signed-off-by David Hildenbrand <[email protected]>--- target/s390x/helper.h | 3 ++ target/s390x/translate_vx.c.inc | 26 ++++++++--- target/s390x/vec_fpu_helper.c | 76 +++++++++++++++++++-------------- 3 files changed, 69 insertions(+), 36 deletions(-)target/s390x/helper.h +3
-0@@ -331,8 +331,11 @@ DEF_HELPER_FLAGS_5(gvec_vfs32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfs128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_4(gvec_vftci32, void, ptr, cptr, env, i32) +DEF_HELPER_4(gvec_vftci32s, void, ptr, cptr, env, i32) DEF_HELPER_4(gvec_vftci64, void, ptr, cptr, env, i32) DEF_HELPER_4(gvec_vftci64s, void, ptr, cptr, env, i32) +DEF_HELPER_4(gvec_vftci128, void, ptr, cptr, env, i32)#ifndef CONFIG_USER_ONLY DEF_HELPER_3 (servc, i32, env, i64, i64)target/s390x/translate_vx.c.inc +21
-5@@ -2991,16 +2991,32 @@ static DisasJumpType op_vftci(DisasContext *s, DisasOps *o) const uint16_t i3 = get_field(s, i3); const uint8_t fpf = get_field(s, m4); const uint8_t m5 = get_field(s, m5); - gen_helper_gvec_2_ptr *fn = gen_helper_gvec_vftci64; + const bool se = extract32(m5, 3, 1); + gen_helper_gvec_2_ptr *fn = NULL; - if (fpf != FPF_LONG || extract32(m5, 0, 3)) { + switch (fpf) { + case FPF_SHORT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = se ? gen_helper_gvec_vftci32s : gen_helper_gvec_vftci32; + } + break; + case FPF_LONG: + fn = se ? gen_helper_gvec_vftci64s : gen_helper_gvec_vftci64; + break; + case FPF_EXT: + if (s390_has_feat(S390_FEAT_VECTOR_ENH)) { + fn = gen_helper_gvec_vftci128; + } + break; + default: + break; + } + + if (!fn || extract32(m5, 0, 3)) { gen_program_exception(s, PGM_SPECIFICATION); return DISAS_NORETURN; } - if (extract32(m5, 3, 1)) { - fn = gen_helper_gvec_vftci64s; - } gen_gvec_2_ptr(get_field(s, v1), get_field(s, v2), cpu_env, i3, fn); set_cc_static(s); return DISAS_NEXT;target/s390x/vec_fpu_helper.c +0
-0
target/s390x/vec_fpu_helper.c line 28
Quoted 6 lines
target/s390x/vec_fpu_helper.c +3
-0@@ -23,6 +23,9 @@const float32 float32_ones =make_float32 (-1u );const float64 float64_ones =make_float64 (-1ull );const float128 float128_ones =make_float128 (-1ull ,-1ull ); +const float32 float32_zeroes =make_float32 (0 ); +const float64 float64_zeroes =make_float64 (0 ); +const float128 float128_zeroes =make_float128 (0 ,0 );
These already exist as "zero" not "zeroes".
Otherwise looks ok, modulo the same comments as for all the other macros in this series.
r~
PATCH 16/20 s390x/tcg: Implement 32/128bit for VECTOR FP MULTIPLY AND (ADD|SUBTRACT)
In reply to David Hildenbrand--- target/s390x/helper.h | 6 ++ target/s390x/translate_vx.c.inc | 48 +++++++++++++--- target/s390x/vec_fpu_helper.c | 98 ++++++++++++++++++++------------- 3 files changed, 107 insertions(+), 45 deletions(-)
target/s390x/helper.h +6 -0
@@ -317,10 +317,16 @@ DEF_HELPER_FLAGS_5(gvec_vfm32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfma32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfma32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfma64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfma64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfma128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfms32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfms32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfms64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfms64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfms128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfsq32, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfsq32s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32)DEF_HELPER_FLAGS_4 (gvec_vfsq64,TCG_CALL_NO_WG , void, ptr, cptr, env, i32)
target/s390x/translate_vx.c.inc +41 -7
59 lines not on this page · raw message
target/s390x/vec_fpu_helper.c +60 -38
113 lines not on this page · raw message
Signature · 1 line
2.26.2
PATCH 17/20 s390x/tcg: Implement VECTOR FP NEGATIVE MULTIPLY AND (ADD|SUBTRACT)
In reply to David Hildenbrand--- target/s390x/helper.h | 10 ++++++++ target/s390x/insn-data.def | 4 +++ target/s390x/translate_vx.c.inc | 26 +++++++++++++++++++ target/s390x/vec_fpu_helper.c | 45 +++++++++++++++++++++++++++++++++ 4 files changed, 85 insertions(+)
target/s390x/helper.h +10 -0
@@ -327,6 +327,16 @@ DEF_HELPER_FLAGS_6(gvec_vfms32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, en DEF_HELPER_FLAGS_6(gvec_vfms64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfms64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfms128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnma32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnma32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnma64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnma64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnma128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnms32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnms32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnms64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnms64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_6(gvec_vfnms128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfsq32, TCG_CALL_NO_WG, void, ptr, cptr, env, i32) DEF_HELPER_FLAGS_4(gvec_vfsq32s, TCG_CALL_NO_WG, void, ptr, cptr, env, i32)DEF_HELPER_FLAGS_4 (gvec_vfsq64,TCG_CALL_NO_WG , void, ptr, cptr, env, i32)
target/s390x/insn-data.def +4 -0
@@ -1259,6 +1259,10 @@ F(0xe78f, VFMA, VRR_e, V, 0, 0, 0, 0, vfma, 0, IF_VEC) /* VECTOR FP MULTIPLY AND SUBTRACT */ F(0xe78e, VFMS, VRR_e, V, 0, 0, 0, 0, vfma, 0, IF_VEC) +/* VECTOR FP NEGATIVE MULTIPLY AND ADD */ + F(0xe79f, VFNMA, VRR_e, VE, 0, 0, 0, 0, vfma, 0, IF_VEC) +/* VECTOR FP NEGATIVE MULTIPLY AND SUBTRACT */ + F(0xe79e, VFNMS, VRR_e, VE, 0, 0, 0, 0, vfma, 0, IF_VEC) /* VECTOR FP PERFORM SIGN OPERATION */ F(0xe7cc, VFPSO, VRR_a, V, 0, 0, 0, 0, vfpso, 0, IF_VEC) /* VECTOR FP SQUARE ROOT */
target/s390x/translate_vx.c.inc +26 -0
@@ -2884,6 +2884,32 @@ static DisasJumpType op_vfma(DisasContext *s, DisasOps *o) break; } break; + case 0x9f: + switch (fpf) { + case FPF_SHORT: + fn = se ? gen_helper_gvec_vfnma32s : gen_helper_gvec_vfnma32; + break; + case FPF_LONG: + fn = se ? gen_helper_gvec_vfnma64s : gen_helper_gvec_vfnma64; + break; + default: + fn = gen_helper_gvec_vfnma128; + break; + } + break; + case 0x9e: + switch (fpf) { + case FPF_SHORT: + fn = se ? gen_helper_gvec_vfnms32s : gen_helper_gvec_vfnms32; + break; + case FPF_LONG: + fn = se ? gen_helper_gvec_vfnms64s : gen_helper_gvec_vfnms64; + break; + default: + fn = gen_helper_gvec_vfnms128; + break; + } + break; default: g_assert_not_reached(); }
target/s390x/vec_fpu_helper.c +45 -0
51 lines not on this page · raw message
Signature · 1 line
2.26.2
PATCH 18/20 s390x/tcg: Implement VECTOR FP (MAXIMUM|MINIMUM)
In reply to David HildenbrandFor IEEE functions, we can reuse the softfloat implementations. For the other functions, implement it generically for 32bit/64bit/128bit - carefully taking care of all weird special cases according to the tables defined in the PoP.
While we could add plenty of helpers to do the function selection at translation time, I don't feel like adding 20*(3+2) helpers for that.
--- target/s390x/helper.h | 10 ++ target/s390x/insn-data.def | 4 + target/s390x/translate_vx.c.inc | 44 +++++ target/s390x/vec_fpu_helper.c | 300 ++++++++++++++++++++++++++++++++ 4 files changed, 358 insertions(+)
target/s390x/helper.h +10 -0
@@ -317,6 +317,16 @@ DEF_HELPER_FLAGS_5(gvec_vfm32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_5(gvec_vfm128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmax32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmax32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmax64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmax64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmax128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmin32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmin32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmin64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmin64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) +DEF_HELPER_FLAGS_5(gvec_vfmin128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfma32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32) DEF_HELPER_FLAGS_6(gvec_vfma32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)DEF_HELPER_FLAGS_6 (gvec_vfma64,TCG_CALL_NO_WG , void, ptr, cptr, cptr, cptr, env, i32)
target/s390x/insn-data.def +4 -0
@@ -1253,6 +1253,10 @@ F(0xe7c4, VFLL, VRR_a, V, 0, 0, 0, 0, vfll, 0, IF_VEC) /* VECTOR LOAD ROUNDED */ F(0xe7c5, VFLR, VRR_a, V, 0, 0, 0, 0, vcdg, 0, IF_VEC) +/* VECTOR FP MAXIMUM */ + F(0xe7ef, VFMAX, VRR_c, VE, 0, 0, 0, 0, vfmax, 0, IF_VEC) +/* VECTOR FP MINIMUM */ + F(0xe7ee, VFMIN, VRR_c, VE, 0, 0, 0, 0, vfmax, 0, IF_VEC) /* VECTOR FP MULTIPLY */ F(0xe7e7, VFM, VRR_c, V, 0, 0, 0, 0, vfa, 0, IF_VEC) /* VECTOR FP MULTIPLY AND ADD */
target/s390x/translate_vx.c.inc +44 -0
50 lines not on this page · raw message
target/s390x/vec_fpu_helper.c +300 -0
303 lines not on this page · raw message
Signature · 1 line
2.26.2
Re: [PATCH v1 18/20] s390x/tcg: Implement VECTOR FP (MAXIMUM|MINIMUM)
target/s390x/vec_fpu_helper.c void HELPER(gvec_vftci##BITS##s)(void *v1, const void *v2, CPUS390XState *env, \ · line 912
Quoted 10 lines · David Hildenbrand
target/s390x/vec_fpu_helper.c +10
-0+typedef enum S390MinMaxType { + s390_minmax_java_math_min, + s390_minmax_java_math_max, + s390_minmax_c_macro_min, + s390_minmax_c_macro_max, + s390_minmax_fmin, + s390_minmax_fmax, + s390_minmax_cpp_alg_min, + s390_minmax_cpp_alg_max, +}S390MinMaxType ;
I think you'd do just as well making this enum match M6, so that no translation is necessary.
target/s390x/vec_fpu_helper.c void HELPER(gvec_vftci##BITS##s)(void *v1, const void *v2, CPUS390XState *env, \ · line 922
Quoted 10 lines
target/s390x/vec_fpu_helper.c +10
-0+ +#define S390_MINMAX (BITS ,TYPE ) \ +static float##BITS TYPE##BITS(float##BITS a, float##BITS b, float_status *s) \ +{ \ + const bool zero_a = float##BITS##_is_infinity(a); \ + const bool zero_b = float##BITS##_is_infinity(b); \ + const bool inf_a = float##BITS##_is_infinity(a); \ + const bool inf_b = float##BITS##_is_infinity(b); \ + const bool nan_a = float##BITS##_is_infinity(a); \ + const bool nan_b = float##BITS##_is_infinity(b); \
Wrong lookups.
Note that you've already got float*_dcmask which you could use to help out here. You just need some named constants to help with the 12 different bits.
target/s390x/vec_fpu_helper.c void HELPER(gvec_vftci##BITS##s)(void *v1, const void *v2, CPUS390XState *env, \ · line 935
Quoted 3 lines
target/s390x/vec_fpu_helper.c +3
-0+switch (TYPE ) { \ +case s390_minmax_java_math_min: \ +case s390_minmax_java_math_max: \
I think you should make TYPE a function parameter, and not replicate this function N times, and so you also don't need
Quoted 28 lines
target/s390x/vec_fpu_helper.c +28
-0+static vop32_3_fn const vfmax_fns32[16 ] = { + [0 ] = float32_maxnum, + [1 ] = s390_minmax_java_math_max32, + [2 ] = s390_minmax_c_macro_max32, + [3 ] = s390_minmax_cpp_alg_max32, + [4 ] = s390_minmax_fmax32, + [8 ] = float32_maxnummag, + [9 ] = s390_minmax_java_math_max_abs32, + [10 ] = s390_minmax_c_macro_max_abs32, + [11 ] = s390_minmax_cpp_alg_max_abs32, + [12 ] = s390_minmax_fmax_abs32, +}; + +static vop64_3_fn const vfmax_fns64[16 ] = { + [0 ] = float64_maxnum, + [1 ] = s390_minmax_java_math_max64, + [2 ] = s390_minmax_c_macro_max64, + [3 ] = s390_minmax_cpp_alg_max64, + [4 ] = s390_minmax_fmax64, + [8 ] = float64_maxnummag, + [9 ] = s390_minmax_java_math_max_abs64, + [10 ] = s390_minmax_c_macro_max_abs64, + [11 ] = s390_minmax_cpp_alg_max_abs64, + [12 ] = s390_minmax_fmax_abs64, +}; + +static vop128_3_fn const vfmax_fns128[16 ] = { + [0 ] = float128_maxnum,
target/s390x/vec_fpu_helper.c void HELPER(gvec_vftci##BITS##s)(void *v1, const void *v2, CPUS390XState *env, \ · line 1112
Quoted 10 lines
target/s390x/vec_fpu_helper.c +10
-0+ [1] = s390_minmax_java_math_max128, + [2] = s390_minmax_c_macro_max128, + [3] = s390_minmax_cpp_alg_max128, + [4] = s390_minmax_fmax128, + [8] = float128_maxnummag, + [9] = s390_minmax_java_math_max_abs128, + [10] = s390_minmax_c_macro_max_abs128, + [11] = s390_minmax_cpp_alg_max_abs128, + [12] = s390_minmax_fmax_abs128, +};
these tables.
I think that the bulk of your minmax could be done exclusively with dcmask, so there could be a common non-macroized function that returns an indication of whether A or B (possibly silenced) should be the result, or if we should use one of your two compare cases.
BTW, for your two "simple comparison" cases, have we eliminated all of the special cases? Could we in fact be calling floatN_min/max instead of floatN_le_quiet?
r~
PATCH 19/20 s390x/tcg: We support Vector enhancements facility
In reply to David Hildenbrand--- target/s390x/gen-features.c | 1 + 1 file changed, 1 insertion(+)
target/s390x/gen-features.c +1 -0
@@ -718,6 +718,7 @@ static uint16_t qemu_MAX[] = {S390_FEAT_INSTRUCTION_EXEC_PROT ,S390_FEAT_MISC_INSTRUCTION_EXT2 ,S390_FEAT_MSA_EXT_8 , + S390_FEAT_VECTOR_ENH, };/****** END FEATURE DEFS ******/
Signature · 1 line
2.26.2
Re: [PATCH v1 19/20] s390x/tcg: We support Vector enhancements facility
Quoted 4 lines · David Hildenbrand
Signed-off-by David Hildenbrand <[email protected]>--- target/s390x/gen-features.c | 1 + 1 file changed, 1 insertion(+)
r~
PATCH 20/20 s390x/cpumodel: Bump up QEMU model to a stripped-down IBM z14 GA2
In reply to David HildenbrandTCG implements everything we need to run basic z14 OS+software.
--- hw/s390x/s390-virtio-ccw.c | 2 ++ target/s390x/cpu_models.c | 4 ++-- target/s390x/gen-features.c | 15 +++++++++------ 3 files changed, 13 insertions(+), 8 deletions(-)
hw/s390x/s390-virtio-ccw.c +2 -0
@@ -812,7 +812,9 @@ DEFINE_CCW_MACHINE(5_2, "5.2", true);static void ccw_machine_5_1_instance_options (MachineState *machine) { +static const S390FeatInit qemu_cpu_feat = {S390_FEAT_LIST_QEMU_V5_1 };ccw_machine_5_2_instance_options (machine); +s390_set_qemu_cpu_model (0x2964 ,13 ,2 , qemu_cpu_feat); }static void ccw_machine_5_1_class_options (MachineClass *mc)
target/s390x/cpu_models.c +2 -2
@@ -88,8 +88,8 @@ static S390CPUDef s390_cpu_defs[] = { CPUDEF_INIT(0x8562, 15, 1, 47, 0x08000000U, "gen15b", "IBM 8562 GA1"), }; -#define QEMU_MAX_CPU_TYPE 0x2964 -#define QEMU_MAX_CPU_GEN 13 +#define QEMU_MAX_CPU_TYPE 0x3906 +#define QEMU_MAX_CPU_GEN 14#define QEMU_MAX_CPU_EC_GA 2static const S390FeatInit qemu_max_cpu_feat_init = {S390_FEAT_LIST_QEMU_MAX };static S390FeatBitmap qemu_max_cpu_feat;
target/s390x/gen-features.c +9 -6
@@ -704,23 +704,25 @@ static uint16_t qemu_V4_1[] = { S390_FEAT_VECTOR, }; -static uint16_t qemu_LATEST[] = { +static uint16_t qemu_V5_1[] = {S390_FEAT_ACCESS_EXCEPTION_FS_INDICATION ,S390_FEAT_SIDE_EFFECT_ACCESS_ESOP2 ,S390_FEAT_ESOP , }; -/* add all new definitions before this point */ -static uint16_t qemu_MAX[] = { -/* generates a dependency warning, leave it out for now */ -S390_FEAT_MSA_EXT_5 , -/* features introduced after the z13 */ +static uint16_t qemu_LATEST[] = {S390_FEAT_INSTRUCTION_EXEC_PROT ,S390_FEAT_MISC_INSTRUCTION_EXT2 ,S390_FEAT_MSA_EXT_8 ,S390_FEAT_VECTOR_ENH , }; +/* add all new definitions before this point */ +static uint16_t qemu_MAX[] = { +/* generates a dependency warning, leave it out for now */ +S390_FEAT_MSA_EXT_5 , +}; +/****** END FEATURE DEFS ******/ #define _YEARS "2016"
@@ -837,6 +839,7 @@ static FeatGroupDefSpec QemuFeatDef[] = {QEMU_FEAT_INITIALIZER (V3_1 ),QEMU_FEAT_INITIALIZER (V4_0 ),QEMU_FEAT_INITIALIZER (V4_1 ), +QEMU_FEAT_INITIALIZER (V5_1 ),QEMU_FEAT_INITIALIZER (LATEST ), QEMU_FEAT_INITIALIZER(MAX), };
Signature · 1 line
2.26.2
Re: [PATCH v1 20/20] s390x/cpumodel: Bump up QEMU model to a stripped-down IBM z14 GA2
Quoted 9 lines · David Hildenbrand
TCG implements everything we need to run basic z14 OS+software.
Cc Christian Borntraeger <[email protected]>Signed-off-by David Hildenbrand <[email protected]>--- hw/s390x/s390-virtio-ccw.c | 2 ++ target/s390x/cpu_models.c | 4 ++-- target/s390x/gen-features.c | 15 +++++++++------ 3 files changed, 13 insertions(+), 8 deletions(-)
r~
Patchew URL: https://patchew.org/QEMU/[email protected]/
Patchew URL: https://patchew.org/QEMU/[email protected]/
Hi,
This series seems to have some coding style problems. See output below for more information:
Type: series
Message-id: [email protected]
Subject: [PATCH v1 00/20] s390x/tcg: Implement Vector enhancements facility and switch to z14
=== TEST SCRIPT BEGIN ===
#!/bin/bash
git rev-parse base > /dev/null || exit 0
git config --local diff.renamelimit 0
git config --local diff.renames True
git config --local diff.algorithm histogram./scripts/checkpatch.pl --mailback base..
=== TEST SCRIPT END ===
Switched to a new branch 'test'
855b893 s390x/cpumodel: Bump up QEMU model to a stripped-down IBM z14 GA2 9e2c5a3 s390x/tcg: We support Vector enhancements facility
91355af s390x/tcg: Implement VECTOR FP (MAXIMUM|MINIMUM)
22ae891 s390x/tcg: Implement VECTOR FP NEGATIVE MULTIPLY AND (ADD|SUBTRACT) 599e1d8 s390x/tcg: Implement 32/128bit for VECTOR FP MULTIPLY AND (ADD|SUBTRACT) 2d46f63 s390x/tcg: Implement 32/128 bit for VECTOR FP TEST DATA CLASS IMMEDIATE 7f46bf1 s390x/tcg: Implement 32/128 bit for VECTOR FP SQUARE ROOT 0870741 s390x/tcg: Implement 32/128 bit for VECTOR FP PERFORM SIGN OPERATION 743a4a3 s390x/tcg: Implement 128 bit for VECTOR FP LOAD ROUNDED 793d28c s390x/tcg: Implement 64 bit for VECTOR FP LOAD LENGTHENED ff85384 s390x/tcg: Implement 32/128 bit for VECTOR LOAD FP INTEGER de00280 s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE * ba7aec7 s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE (AND SIGNAL) SCALAR 74af73a s390x/tcg: Implement 32/128 bit for VECTOR FP SUBTRACT 59a6f7a s390x/tcg: Implement 32/128 bit for VECTOR FP MULTIPLY c0903b4 s390x/tcg: Implement 32/128 bit for VECTOR FP DIVIDE
8af7fca s390x/tcg: Implement 32/128 bit for VECTOR FP ADD
e103776 s390x/tcg: Implement VECTOR MULTIPLY SUM LOGICAL
092690a s390x/tcg: Implement VECTOR BIT PERMUTE
46d60ed softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)
=== OUTPUT BEGIN ===
1/20 Checking commit 46d60ede64da (softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)) 2/20 Checking commit 092690a01e4b (s390x/tcg: Implement VECTOR BIT PERMUTE)
3/20 Checking commit e1037765c06f (s390x/tcg: Implement VECTOR MULTIPLY SUM LOGICAL)
4/20 Checking commit 8af7fcaee5ef (s390x/tcg: Implement 32/128 bit for VECTOR FP ADD)
ERROR: space prohibited between function name and open parenthesis '('
#162: FILE: target/s390x/vec_fpu_helper.c:140:
+typedef float##BITS (*vop##BITS##_3_fn)(float##BITS a, float##BITS b, \
total: 1 errors, 0 warnings, 183 lines checked
Patch 4/20 has style problems, please review. If any of these errors are false positives report them to the maintainer, see CHECKPATCH in MAINTAINERS.
5/20 Checking commit c0903b4e139a (s390x/tcg: Implement 32/128 bit for VECTOR FP DIVIDE) 6/20 Checking commit 59a6f7a3d4ca (s390x/tcg: Implement 32/128 bit for VECTOR FP MULTIPLY) 7/20 Checking commit 74af73a36417 (s390x/tcg: Implement 32/128 bit for VECTOR FP SUBTRACT) 8/20 Checking commit ba7aec76b6d7 (s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE (AND SIGNAL) SCALAR)
WARNING: Block comments use a leading /* on a separate line
#113: FILE: target/s390x/vec_fpu_helper.c:190:
+ /* only the zero-indexed elements are compared */ \
total: 0 errors, 1 warnings, 136 lines checked
Patch 8/20 has style problems, please review. If any of these errors
are false positives report them to the maintainer, see
CHECKPATCH in MAINTAINERS.
9/20 Checking commit de00280993e0 (s390x/tcg: Implement 32/128 bit for VECTOR FP COMPARE *)
WARNING: line over 80 characters
#26: FILE: target/s390x/helper.h:262:
+DEF_HELPER_FLAGS_5(gvec_vfce32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#29: FILE: target/s390x/helper.h:265:
+DEF_HELPER_FLAGS_5(gvec_vfce128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#36: FILE: target/s390x/helper.h:272:
+DEF_HELPER_FLAGS_5(gvec_vfch32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#39: FILE: target/s390x/helper.h:275:
+DEF_HELPER_FLAGS_5(gvec_vfch128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#45: FILE: target/s390x/helper.h:281:
+DEF_HELPER_FLAGS_5(gvec_vfche32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#46: FILE: target/s390x/helper.h:282:
+DEF_HELPER_FLAGS_5(gvec_vfche32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#49: FILE: target/s390x/helper.h:285:
+DEF_HELPER_FLAGS_5(gvec_vfche128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: Block comments use a leading /* on a separate line #250: FILE: target/s390x/vec_fpu_helper.c:250:
+ /* swap the parameters, so we can use existing functions */ \
total: 0 errors, 8 warnings, 445 lines checked
Patch 9/20 has style problems, please review. If any of these errors
are false positives report them to the maintainer, see
CHECKPATCH in MAINTAINERS.
10/20 Checking commit ff85384a58d4 (s390x/tcg: Implement 32/128 bit for VECTOR LOAD FP INTEGER)
ERROR: space prohibited between function name and open parenthesis '('
#140: FILE: target/s390x/vec_fpu_helper.c:120:
+typedef float##BITS (*vop##BITS##_2_fn)(float##BITS a, float_status *s); \
total: 1 errors, 0 warnings, 191 lines checked
Patch 10/20 has style problems, please review. If any of these errors are false positives report them to the maintainer, see CHECKPATCH in MAINTAINERS.
11/20 Checking commit 793d28cefc57 (s390x/tcg: Implement 64 bit for VECTOR FP LOAD LENGTHENED) 12/20 Checking commit 743a4a3474e5 (s390x/tcg: Implement 128 bit for VECTOR FP LOAD ROUNDED) 13/20 Checking commit 08707417ca0e (s390x/tcg: Implement 32/128 bit for VECTOR FP PERFORM SIGN OPERATION) 14/20 Checking commit 7f46bf12ce3e (s390x/tcg: Implement 32/128 bit for VECTOR FP SQUARE ROOT) 15/20 Checking commit 2d46f639d6dd (s390x/tcg: Implement 32/128 bit for VECTOR FP TEST DATA CLASS IMMEDIATE) 16/20 Checking commit 599e1d8a08ff (s390x/tcg: Implement 32/128bit for VECTOR FP MULTIPLY AND (ADD|SUBTRACT))
WARNING: line over 80 characters
#20: FILE: target/s390x/helper.h:320:
+DEF_HELPER_FLAGS_6(gvec_vfma32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#21: FILE: target/s390x/helper.h:321:
+DEF_HELPER_FLAGS_6(gvec_vfma32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#24: FILE: target/s390x/helper.h:324:
+DEF_HELPER_FLAGS_6(gvec_vfma128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#25: FILE: target/s390x/helper.h:325:
+DEF_HELPER_FLAGS_6(gvec_vfms32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#26: FILE: target/s390x/helper.h:326:
+DEF_HELPER_FLAGS_6(gvec_vfms32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#29: FILE: target/s390x/helper.h:329:
+DEF_HELPER_FLAGS_6(gvec_vfms128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
total: 0 errors, 6 warnings, 188 lines checked
Patch 16/20 has style problems, please review. If any of these errors
are false positives report them to the maintainer, see
CHECKPATCH in MAINTAINERS.
17/20 Checking commit 22ae891fa9a5 (s390x/tcg: Implement VECTOR FP NEGATIVE MULTIPLY AND (ADD|SUBTRACT))
WARNING: line over 80 characters
#20: FILE: target/s390x/helper.h:330:
+DEF_HELPER_FLAGS_6(gvec_vfnma32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#21: FILE: target/s390x/helper.h:331:
+DEF_HELPER_FLAGS_6(gvec_vfnma32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#22: FILE: target/s390x/helper.h:332:
+DEF_HELPER_FLAGS_6(gvec_vfnma64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#23: FILE: target/s390x/helper.h:333:
+DEF_HELPER_FLAGS_6(gvec_vfnma64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#24: FILE: target/s390x/helper.h:334:
+DEF_HELPER_FLAGS_6(gvec_vfnma128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#25: FILE: target/s390x/helper.h:335:
+DEF_HELPER_FLAGS_6(gvec_vfnms32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#26: FILE: target/s390x/helper.h:336:
+DEF_HELPER_FLAGS_6(gvec_vfnms32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#27: FILE: target/s390x/helper.h:337:
+DEF_HELPER_FLAGS_6(gvec_vfnms64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#28: FILE: target/s390x/helper.h:338:
+DEF_HELPER_FLAGS_6(gvec_vfnms64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#29: FILE: target/s390x/helper.h:339:
+DEF_HELPER_FLAGS_6(gvec_vfnms128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, cptr, env, i32)
total: 0 errors, 10 warnings, 109 lines checked
Patch 17/20 has style problems, please review. If any of these errors are false positives report them to the maintainer, see
CHECKPATCH in MAINTAINERS.
18/20 Checking commit 91355af73073 (s390x/tcg: Implement VECTOR FP (MAXIMUM|MINIMUM))
WARNING: line over 80 characters
#28: FILE: target/s390x/helper.h:320:
+DEF_HELPER_FLAGS_5(gvec_vfmax32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#29: FILE: target/s390x/helper.h:321:
+DEF_HELPER_FLAGS_5(gvec_vfmax32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#30: FILE: target/s390x/helper.h:322:
+DEF_HELPER_FLAGS_5(gvec_vfmax64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#31: FILE: target/s390x/helper.h:323:
+DEF_HELPER_FLAGS_5(gvec_vfmax64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#32: FILE: target/s390x/helper.h:324:
+DEF_HELPER_FLAGS_5(gvec_vfmax128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#33: FILE: target/s390x/helper.h:325:
+DEF_HELPER_FLAGS_5(gvec_vfmin32s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#34: FILE: target/s390x/helper.h:326:
+DEF_HELPER_FLAGS_5(gvec_vfmin32, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#35: FILE: target/s390x/helper.h:327:
+DEF_HELPER_FLAGS_5(gvec_vfmin64s, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#36: FILE: target/s390x/helper.h:328:
+DEF_HELPER_FLAGS_5(gvec_vfmin64, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: line over 80 characters
#37: FILE: target/s390x/helper.h:329:
+DEF_HELPER_FLAGS_5(gvec_vfmin128, TCG_CALL_NO_WG, void, ptr, cptr, cptr, env, i32)
WARNING: Block comments use a leading /* on a separate line #158: FILE: target/s390x/vec_fpu_helper.c:941:
+ /* fall through */ \
WARNING: Block comments use a leading /* on a separate line #211: FILE: target/s390x/vec_fpu_helper.c:994:
+ /* We can process all remaining cases using simple comparison. */ \
total: 0 errors, 12 warnings, 379 lines checked
Patch 18/20 has style problems, please review. If any of these errors
are false positives report them to the maintainer, see
CHECKPATCH in MAINTAINERS.
19/20 Checking commit 9e2c5a3530bd (s390x/tcg: We support Vector enhancements facility) 20/20 Checking commit 855b893d0a4d (s390x/cpumodel: Bump up QEMU model to a stripped-down IBM z14 GA2) === OUTPUT END ===
Test command exited with code: 1
The full log is available at
http://patchew.org/logs/[email protected]/testing.checkpatch/?type=message.
---
Email generated automatically by Patchew [https://patchew.org/].
Please send your feedback to [email protected]
Quoted 8 lines · David Hildenbrand
This series adds support for the "Vector enhancements facility" and bumps the qemu CPU model tp to a stripped-down z14.
I yet have to find a way to get more test coverage - looks like some of the functions aren't used anywhere yet (e.g., VECTOR FP MAXIMUM), writing unit tests to cover all functions and cases is just nasty. But I might be wrong - I'm planning to at least test basic functionality of all new added instructions.
This is where RISU can be helpful. Auto-generate 100k random variations, record known good results on hardware, verify that replay on qemu produces the same results.
r~
Quoted 10 lines · Richard Henderson
Quoted 8 lines · David Hildenbrand
This series adds support for the "Vector enhancements facility" and bumps the qemu CPU model tp to a stripped-down z14.
I yet have to find a way to get more test coverage - looks like some of the functions aren't used anywhere yet (e.g., VECTOR FP MAXIMUM), writing unit tests to cover all functions and cases is just nasty. But I might be wrong - I'm planning to at least test basic functionality of all new added instructions.
Quoted 3 lines
This is where RISU can be helpful. Auto-generate 100k random variations, record known good results on hardware, verify that replay on qemu produces the same results.
Makes sense, however, some corner cases (e.g., MAX(+0, -O)) still might have to be handled manually.
It might take me a while to address the feedback (I'm fairly busy as usual ...). Thanks!
Signature · 3 lines
Thanks, David / dhildenb
Quoted 12 lines · David Hildenbrand
This series adds support for the "Vector enhancements facility" and bumps the qemu CPU model tp to a stripped-down z14.
I yet have to find a way to get more test coverage - looks like some of the functions aren't used anywhere yet (e.g., VECTOR FP MAXIMUM), writing unit tests to cover all functions and cases is just nasty. But I might be wrong - I'm planning to at least test basic functionality of all new added instructions.
I have to make excessive use of c macros again to cover different element sizes (32/64/128bit). Any advise to clean things up are welcome.
Quoted 6 lines
This is based on:
"[PATCH v2 0/9] s390x/tcg: Implement some z14 facilities" "[PATCH v2 00/10] softfloat: Implement float128_muladd"Based-on: [email protected]
Based-on: [email protected]
Hi Richard,
I'll have a new series ready soonish. I did what you suggested and started generating random (valid) Vector FP instructions with (valid) random parameters, executed on randomly generated vectors. It's looking pretty good by now.
I'll still have to see to which degree I can address feedback on "softfloat: Implement float128_(min|minnum|minnummag|max|maxnum|maxnummag)"
Anyhow, I'd need your "softfloat: Implement float128_muladd" series -- any idea when you might get to continue working on that? Thanks!
Signature · 3 lines
Thanks, David / dhildenb