| Message ID | 20210823150235.35759-1-luis.pires@eldorado.org.br |
|---|---|
| State | New |
| Headers | show |
| Series | target/ppc: fix setting of CR flags in bcdcfsq | expand |
On Mon, Aug 23, 2021 at 12:02:35PM -0300, Luis Pires wrote: > According to the ISA, CR should be set based on the source value, and > not on the packed decimal result. > The way this was implemented would cause GT, LT and EQ to be set > incorrectly when the source value was too large and the 31 least > significant digits of the packed decimal result ended up being all zero. > This would happen for source values of +/-10^31, +/-10^32, etc. > > The new implementation fixes this and also skips the result calculation > altogether in case of src overflow. > > Signed-off-by: Luis Pires <luis.pires@eldorado.org.br> CCing Richard Henderson. In general Richard's probably a better person than me to review actual ppc TCG mechanics. > --- > target/ppc/int_helper.c | 61 ++++++++++++++++++++++++++++++++--------- > 1 file changed, 48 insertions(+), 13 deletions(-) > > diff --git a/target/ppc/int_helper.c b/target/ppc/int_helper.c > index efa833ef64..de56056429 100644 > --- a/target/ppc/int_helper.c > +++ b/target/ppc/int_helper.c > @@ -2498,10 +2498,26 @@ uint32_t helper_bcdctz(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) > return cr; > } > > +/** > + * Compare 2 128-bit unsigned integers, passed in as unsigned 64-bit pairs > + * > + * Returns: > + * > 0 if ahi|alo > bhi|blo, > + * 0 if ahi|alo == bhi|blo, > + * < 0 if ahi|alo < bhi|blo > + */ > +static inline int ucmp128(uint64_t alo, uint64_t ahi, > + uint64_t blo, uint64_t bhi) > +{ > + return (ahi == bhi) ? > + (alo > blo ? 1 : (alo == blo ? 0 : -1)) : > + (ahi > bhi ? 1 : -1); > +} > + > uint32_t helper_bcdcfsq(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) > { > int i; > - int cr = 0; > + int cr; > uint64_t lo_value; > uint64_t hi_value; > ppc_avr_t ret = { .u64 = { 0, 0 } }; > @@ -2510,28 +2526,47 @@ uint32_t helper_bcdcfsq(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) > lo_value = -b->VsrSD(1); > hi_value = ~b->VsrD(0) + !lo_value; > bcd_put_digit(&ret, 0xD, 0); > + > + cr = CRF_LT; > } else { > lo_value = b->VsrD(1); > hi_value = b->VsrD(0); > bcd_put_digit(&ret, bcd_preferred_sgn(0, ps), 0); > - } > > - if (divu128(&lo_value, &hi_value, 1000000000000000ULL) || > - lo_value > 9999999999999999ULL) { > - cr = CRF_SO; > + if (hi_value == 0 && lo_value == 0) { > + cr = CRF_EQ; > + } else { > + cr = CRF_GT; > + } > } > > - for (i = 1; i < 16; hi_value /= 10, i++) { > - bcd_put_digit(&ret, hi_value % 10, i); > - } > + /* > + * Check src limits: abs(src) <= 10^31 - 1 > + * > + * 10^31 - 1 = 0x0000007e37be2022 c0914b267fffffff > + */ > + if (ucmp128(lo_value, hi_value, > + 0xc0914b267fffffffULL, 0x7e37be2022ULL) > 0) { > + cr |= CRF_SO; > > - for (; i < 32; lo_value /= 10, i++) { > - bcd_put_digit(&ret, lo_value % 10, i); > - } > + /* > + * According to the ISA, if src wouldn't fit in the destination > + * register, the result is undefined. > + * In that case, we leave r unchanged. > + */ > + } else { > + divu128(&lo_value, &hi_value, 1000000000000000ULL); > > - cr |= bcd_cmp_zero(&ret); > + for (i = 1; i < 16; hi_value /= 10, i++) { > + bcd_put_digit(&ret, hi_value % 10, i); > + } > > - *r = ret; > + for (; i < 32; lo_value /= 10, i++) { > + bcd_put_digit(&ret, lo_value % 10, i); > + } > + > + *r = ret; > + } > > return cr; > }
On 8/23/21 5:02 PM, Luis Pires wrote: > According to the ISA, CR should be set based on the source value, and > not on the packed decimal result. > The way this was implemented would cause GT, LT and EQ to be set > incorrectly when the source value was too large and the 31 least > significant digits of the packed decimal result ended up being all zero. > This would happen for source values of +/-10^31, +/-10^32, etc. > > The new implementation fixes this and also skips the result calculation > altogether in case of src overflow. > > Signed-off-by: Luis Pires<luis.pires@eldorado.org.br> > --- > target/ppc/int_helper.c | 61 ++++++++++++++++++++++++++++++++--------- > 1 file changed, 48 insertions(+), 13 deletions(-) Reviewed-by: Richard Henderson <richard.henderson@linaro.org> r~
On Mon, Aug 23, 2021 at 12:02:35PM -0300, Luis Pires wrote: > According to the ISA, CR should be set based on the source value, and > not on the packed decimal result. > The way this was implemented would cause GT, LT and EQ to be set > incorrectly when the source value was too large and the 31 least > significant digits of the packed decimal result ended up being all zero. > This would happen for source values of +/-10^31, +/-10^32, etc. > > The new implementation fixes this and also skips the result calculation > altogether in case of src overflow. > > Signed-off-by: Luis Pires <luis.pires@eldorado.org.br> Applied to ppc-for-6.2, thanks. > --- > target/ppc/int_helper.c | 61 ++++++++++++++++++++++++++++++++--------- > 1 file changed, 48 insertions(+), 13 deletions(-) > > diff --git a/target/ppc/int_helper.c b/target/ppc/int_helper.c > index efa833ef64..de56056429 100644 > --- a/target/ppc/int_helper.c > +++ b/target/ppc/int_helper.c > @@ -2498,10 +2498,26 @@ uint32_t helper_bcdctz(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) > return cr; > } > > +/** > + * Compare 2 128-bit unsigned integers, passed in as unsigned 64-bit pairs > + * > + * Returns: > + * > 0 if ahi|alo > bhi|blo, > + * 0 if ahi|alo == bhi|blo, > + * < 0 if ahi|alo < bhi|blo > + */ > +static inline int ucmp128(uint64_t alo, uint64_t ahi, > + uint64_t blo, uint64_t bhi) > +{ > + return (ahi == bhi) ? > + (alo > blo ? 1 : (alo == blo ? 0 : -1)) : > + (ahi > bhi ? 1 : -1); > +} > + > uint32_t helper_bcdcfsq(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) > { > int i; > - int cr = 0; > + int cr; > uint64_t lo_value; > uint64_t hi_value; > ppc_avr_t ret = { .u64 = { 0, 0 } }; > @@ -2510,28 +2526,47 @@ uint32_t helper_bcdcfsq(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) > lo_value = -b->VsrSD(1); > hi_value = ~b->VsrD(0) + !lo_value; > bcd_put_digit(&ret, 0xD, 0); > + > + cr = CRF_LT; > } else { > lo_value = b->VsrD(1); > hi_value = b->VsrD(0); > bcd_put_digit(&ret, bcd_preferred_sgn(0, ps), 0); > - } > > - if (divu128(&lo_value, &hi_value, 1000000000000000ULL) || > - lo_value > 9999999999999999ULL) { > - cr = CRF_SO; > + if (hi_value == 0 && lo_value == 0) { > + cr = CRF_EQ; > + } else { > + cr = CRF_GT; > + } > } > > - for (i = 1; i < 16; hi_value /= 10, i++) { > - bcd_put_digit(&ret, hi_value % 10, i); > - } > + /* > + * Check src limits: abs(src) <= 10^31 - 1 > + * > + * 10^31 - 1 = 0x0000007e37be2022 c0914b267fffffff > + */ > + if (ucmp128(lo_value, hi_value, > + 0xc0914b267fffffffULL, 0x7e37be2022ULL) > 0) { > + cr |= CRF_SO; > > - for (; i < 32; lo_value /= 10, i++) { > - bcd_put_digit(&ret, lo_value % 10, i); > - } > + /* > + * According to the ISA, if src wouldn't fit in the destination > + * register, the result is undefined. > + * In that case, we leave r unchanged. > + */ > + } else { > + divu128(&lo_value, &hi_value, 1000000000000000ULL); > > - cr |= bcd_cmp_zero(&ret); > + for (i = 1; i < 16; hi_value /= 10, i++) { > + bcd_put_digit(&ret, hi_value % 10, i); > + } > > - *r = ret; > + for (; i < 32; lo_value /= 10, i++) { > + bcd_put_digit(&ret, lo_value % 10, i); > + } > + > + *r = ret; > + } > > return cr; > }
diff --git a/target/ppc/int_helper.c b/target/ppc/int_helper.c index efa833ef64..de56056429 100644 --- a/target/ppc/int_helper.c +++ b/target/ppc/int_helper.c @@ -2498,10 +2498,26 @@ uint32_t helper_bcdctz(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) return cr; } +/** + * Compare 2 128-bit unsigned integers, passed in as unsigned 64-bit pairs + * + * Returns: + * > 0 if ahi|alo > bhi|blo, + * 0 if ahi|alo == bhi|blo, + * < 0 if ahi|alo < bhi|blo + */ +static inline int ucmp128(uint64_t alo, uint64_t ahi, + uint64_t blo, uint64_t bhi) +{ + return (ahi == bhi) ? + (alo > blo ? 1 : (alo == blo ? 0 : -1)) : + (ahi > bhi ? 1 : -1); +} + uint32_t helper_bcdcfsq(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) { int i; - int cr = 0; + int cr; uint64_t lo_value; uint64_t hi_value; ppc_avr_t ret = { .u64 = { 0, 0 } }; @@ -2510,28 +2526,47 @@ uint32_t helper_bcdcfsq(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) lo_value = -b->VsrSD(1); hi_value = ~b->VsrD(0) + !lo_value; bcd_put_digit(&ret, 0xD, 0); + + cr = CRF_LT; } else { lo_value = b->VsrD(1); hi_value = b->VsrD(0); bcd_put_digit(&ret, bcd_preferred_sgn(0, ps), 0); - } - if (divu128(&lo_value, &hi_value, 1000000000000000ULL) || - lo_value > 9999999999999999ULL) { - cr = CRF_SO; + if (hi_value == 0 && lo_value == 0) { + cr = CRF_EQ; + } else { + cr = CRF_GT; + } } - for (i = 1; i < 16; hi_value /= 10, i++) { - bcd_put_digit(&ret, hi_value % 10, i); - } + /* + * Check src limits: abs(src) <= 10^31 - 1 + * + * 10^31 - 1 = 0x0000007e37be2022 c0914b267fffffff + */ + if (ucmp128(lo_value, hi_value, + 0xc0914b267fffffffULL, 0x7e37be2022ULL) > 0) { + cr |= CRF_SO; - for (; i < 32; lo_value /= 10, i++) { - bcd_put_digit(&ret, lo_value % 10, i); - } + /* + * According to the ISA, if src wouldn't fit in the destination + * register, the result is undefined. + * In that case, we leave r unchanged. + */ + } else { + divu128(&lo_value, &hi_value, 1000000000000000ULL); - cr |= bcd_cmp_zero(&ret); + for (i = 1; i < 16; hi_value /= 10, i++) { + bcd_put_digit(&ret, hi_value % 10, i); + } - *r = ret; + for (; i < 32; lo_value /= 10, i++) { + bcd_put_digit(&ret, lo_value % 10, i); + } + + *r = ret; + } return cr; }
According to the ISA, CR should be set based on the source value, and not on the packed decimal result. The way this was implemented would cause GT, LT and EQ to be set incorrectly when the source value was too large and the 31 least significant digits of the packed decimal result ended up being all zero. This would happen for source values of +/-10^31, +/-10^32, etc. The new implementation fixes this and also skips the result calculation altogether in case of src overflow. Signed-off-by: Luis Pires <luis.pires@eldorado.org.br> --- target/ppc/int_helper.c | 61 ++++++++++++++++++++++++++++++++--------- 1 file changed, 48 insertions(+), 13 deletions(-)