Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | #ifndef __ASM_ALPHA_FPU_H |
| 2 | #define __ASM_ALPHA_FPU_H |
| 3 | |
| 4 | /* |
| 5 | * Alpha floating-point control register defines: |
| 6 | */ |
| 7 | #define FPCR_DNOD (1UL<<47) /* denorm INV trap disable */ |
| 8 | #define FPCR_DNZ (1UL<<48) /* denorms to zero */ |
| 9 | #define FPCR_INVD (1UL<<49) /* invalid op disable (opt.) */ |
| 10 | #define FPCR_DZED (1UL<<50) /* division by zero disable (opt.) */ |
| 11 | #define FPCR_OVFD (1UL<<51) /* overflow disable (optional) */ |
| 12 | #define FPCR_INV (1UL<<52) /* invalid operation */ |
| 13 | #define FPCR_DZE (1UL<<53) /* division by zero */ |
| 14 | #define FPCR_OVF (1UL<<54) /* overflow */ |
| 15 | #define FPCR_UNF (1UL<<55) /* underflow */ |
| 16 | #define FPCR_INE (1UL<<56) /* inexact */ |
| 17 | #define FPCR_IOV (1UL<<57) /* integer overflow */ |
| 18 | #define FPCR_UNDZ (1UL<<60) /* underflow to zero (opt.) */ |
| 19 | #define FPCR_UNFD (1UL<<61) /* underflow disable (opt.) */ |
| 20 | #define FPCR_INED (1UL<<62) /* inexact disable (opt.) */ |
| 21 | #define FPCR_SUM (1UL<<63) /* summary bit */ |
| 22 | |
| 23 | #define FPCR_DYN_SHIFT 58 /* first dynamic rounding mode bit */ |
| 24 | #define FPCR_DYN_CHOPPED (0x0UL << FPCR_DYN_SHIFT) /* towards 0 */ |
| 25 | #define FPCR_DYN_MINUS (0x1UL << FPCR_DYN_SHIFT) /* towards -INF */ |
| 26 | #define FPCR_DYN_NORMAL (0x2UL << FPCR_DYN_SHIFT) /* towards nearest */ |
| 27 | #define FPCR_DYN_PLUS (0x3UL << FPCR_DYN_SHIFT) /* towards +INF */ |
| 28 | #define FPCR_DYN_MASK (0x3UL << FPCR_DYN_SHIFT) |
| 29 | |
| 30 | #define FPCR_MASK 0xffff800000000000L |
| 31 | |
| 32 | /* |
| 33 | * IEEE trap enables are implemented in software. These per-thread |
| 34 | * bits are stored in the "ieee_state" field of "struct thread_info". |
| 35 | * Thus, the bits are defined so as not to conflict with the |
| 36 | * floating-point enable bit (which is architected). On top of that, |
| 37 | * we want to make these bits compatible with OSF/1 so |
| 38 | * ieee_set_fp_control() etc. can be implemented easily and |
| 39 | * compatibly. The corresponding definitions are in |
| 40 | * /usr/include/machine/fpu.h under OSF/1. |
| 41 | */ |
| 42 | #define IEEE_TRAP_ENABLE_INV (1UL<<1) /* invalid op */ |
| 43 | #define IEEE_TRAP_ENABLE_DZE (1UL<<2) /* division by zero */ |
| 44 | #define IEEE_TRAP_ENABLE_OVF (1UL<<3) /* overflow */ |
| 45 | #define IEEE_TRAP_ENABLE_UNF (1UL<<4) /* underflow */ |
| 46 | #define IEEE_TRAP_ENABLE_INE (1UL<<5) /* inexact */ |
| 47 | #define IEEE_TRAP_ENABLE_DNO (1UL<<6) /* denorm */ |
| 48 | #define IEEE_TRAP_ENABLE_MASK (IEEE_TRAP_ENABLE_INV | IEEE_TRAP_ENABLE_DZE |\ |
| 49 | IEEE_TRAP_ENABLE_OVF | IEEE_TRAP_ENABLE_UNF |\ |
| 50 | IEEE_TRAP_ENABLE_INE | IEEE_TRAP_ENABLE_DNO) |
| 51 | |
| 52 | /* Denorm and Underflow flushing */ |
| 53 | #define IEEE_MAP_DMZ (1UL<<12) /* Map denorm inputs to zero */ |
| 54 | #define IEEE_MAP_UMZ (1UL<<13) /* Map underflowed outputs to zero */ |
| 55 | |
| 56 | #define IEEE_MAP_MASK (IEEE_MAP_DMZ | IEEE_MAP_UMZ) |
| 57 | |
| 58 | /* status bits coming from fpcr: */ |
| 59 | #define IEEE_STATUS_INV (1UL<<17) |
| 60 | #define IEEE_STATUS_DZE (1UL<<18) |
| 61 | #define IEEE_STATUS_OVF (1UL<<19) |
| 62 | #define IEEE_STATUS_UNF (1UL<<20) |
| 63 | #define IEEE_STATUS_INE (1UL<<21) |
| 64 | #define IEEE_STATUS_DNO (1UL<<22) |
| 65 | |
| 66 | #define IEEE_STATUS_MASK (IEEE_STATUS_INV | IEEE_STATUS_DZE | \ |
| 67 | IEEE_STATUS_OVF | IEEE_STATUS_UNF | \ |
| 68 | IEEE_STATUS_INE | IEEE_STATUS_DNO) |
| 69 | |
| 70 | #define IEEE_SW_MASK (IEEE_TRAP_ENABLE_MASK | \ |
| 71 | IEEE_STATUS_MASK | IEEE_MAP_MASK) |
| 72 | |
| 73 | #define IEEE_CURRENT_RM_SHIFT 32 |
| 74 | #define IEEE_CURRENT_RM_MASK (3UL<<IEEE_CURRENT_RM_SHIFT) |
| 75 | |
| 76 | #define IEEE_STATUS_TO_EXCSUM_SHIFT 16 |
| 77 | |
| 78 | #define IEEE_INHERIT (1UL<<63) /* inherit on thread create? */ |
| 79 | |
| 80 | /* |
| 81 | * Convert the software IEEE trap enable and status bits into the |
| 82 | * hardware fpcr format. |
| 83 | * |
| 84 | * Digital Unix engineers receive my thanks for not defining the |
| 85 | * software bits identical to the hardware bits. The chip designers |
| 86 | * receive my thanks for making all the not-implemented fpcr bits |
| 87 | * RAZ forcing us to use system calls to read/write this value. |
| 88 | */ |
| 89 | |
| 90 | static inline unsigned long |
| 91 | ieee_swcr_to_fpcr(unsigned long sw) |
| 92 | { |
| 93 | unsigned long fp; |
| 94 | fp = (sw & IEEE_STATUS_MASK) << 35; |
| 95 | fp |= (sw & IEEE_MAP_DMZ) << 36; |
| 96 | fp |= (sw & IEEE_STATUS_MASK ? FPCR_SUM : 0); |
| 97 | fp |= (~sw & (IEEE_TRAP_ENABLE_INV |
| 98 | | IEEE_TRAP_ENABLE_DZE |
| 99 | | IEEE_TRAP_ENABLE_OVF)) << 48; |
| 100 | fp |= (~sw & (IEEE_TRAP_ENABLE_UNF | IEEE_TRAP_ENABLE_INE)) << 57; |
| 101 | fp |= (sw & IEEE_MAP_UMZ ? FPCR_UNDZ | FPCR_UNFD : 0); |
| 102 | fp |= (~sw & IEEE_TRAP_ENABLE_DNO) << 41; |
| 103 | return fp; |
| 104 | } |
| 105 | |
| 106 | static inline unsigned long |
| 107 | ieee_fpcr_to_swcr(unsigned long fp) |
| 108 | { |
| 109 | unsigned long sw; |
| 110 | sw = (fp >> 35) & IEEE_STATUS_MASK; |
| 111 | sw |= (fp >> 36) & IEEE_MAP_DMZ; |
| 112 | sw |= (~fp >> 48) & (IEEE_TRAP_ENABLE_INV |
| 113 | | IEEE_TRAP_ENABLE_DZE |
| 114 | | IEEE_TRAP_ENABLE_OVF); |
| 115 | sw |= (~fp >> 57) & (IEEE_TRAP_ENABLE_UNF | IEEE_TRAP_ENABLE_INE); |
| 116 | sw |= (fp >> 47) & IEEE_MAP_UMZ; |
| 117 | sw |= (~fp >> 41) & IEEE_TRAP_ENABLE_DNO; |
| 118 | return sw; |
| 119 | } |
| 120 | |
| 121 | #ifdef __KERNEL__ |
| 122 | |
| 123 | /* The following two functions don't need trapb/excb instructions |
| 124 | around the mf_fpcr/mt_fpcr instructions because (a) the kernel |
| 125 | never generates arithmetic faults and (b) call_pal instructions |
| 126 | are implied trap barriers. */ |
| 127 | |
| 128 | static inline unsigned long |
| 129 | rdfpcr(void) |
| 130 | { |
| 131 | unsigned long tmp, ret; |
| 132 | |
Akinobu Mita | 4b417d0 | 2006-03-26 01:39:01 -0800 | [diff] [blame] | 133 | #if defined(CONFIG_ALPHA_EV6) || defined(CONFIG_ALPHA_EV67) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 134 | __asm__ __volatile__ ( |
| 135 | "ftoit $f0,%0\n\t" |
| 136 | "mf_fpcr $f0\n\t" |
| 137 | "ftoit $f0,%1\n\t" |
| 138 | "itoft %0,$f0" |
| 139 | : "=r"(tmp), "=r"(ret)); |
| 140 | #else |
| 141 | __asm__ __volatile__ ( |
| 142 | "stt $f0,%0\n\t" |
| 143 | "mf_fpcr $f0\n\t" |
| 144 | "stt $f0,%1\n\t" |
| 145 | "ldt $f0,%0" |
| 146 | : "=m"(tmp), "=m"(ret)); |
| 147 | #endif |
| 148 | |
| 149 | return ret; |
| 150 | } |
| 151 | |
| 152 | static inline void |
| 153 | wrfpcr(unsigned long val) |
| 154 | { |
| 155 | unsigned long tmp; |
| 156 | |
Akinobu Mita | 4b417d0 | 2006-03-26 01:39:01 -0800 | [diff] [blame] | 157 | #if defined(CONFIG_ALPHA_EV6) || defined(CONFIG_ALPHA_EV67) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 158 | __asm__ __volatile__ ( |
| 159 | "ftoit $f0,%0\n\t" |
| 160 | "itoft %1,$f0\n\t" |
| 161 | "mt_fpcr $f0\n\t" |
| 162 | "itoft %0,$f0" |
| 163 | : "=&r"(tmp) : "r"(val)); |
| 164 | #else |
| 165 | __asm__ __volatile__ ( |
| 166 | "stt $f0,%0\n\t" |
| 167 | "ldt $f0,%1\n\t" |
| 168 | "mt_fpcr $f0\n\t" |
| 169 | "ldt $f0,%0" |
| 170 | : "=m"(tmp) : "m"(val)); |
| 171 | #endif |
| 172 | } |
| 173 | |
| 174 | static inline unsigned long |
| 175 | swcr_update_status(unsigned long swcr, unsigned long fpcr) |
| 176 | { |
| 177 | /* EV6 implements most of the bits in hardware. Collect |
| 178 | the acrued exception bits from the real fpcr. */ |
| 179 | if (implver() == IMPLVER_EV6) { |
| 180 | swcr &= ~IEEE_STATUS_MASK; |
| 181 | swcr |= (fpcr >> 35) & IEEE_STATUS_MASK; |
| 182 | } |
| 183 | return swcr; |
| 184 | } |
| 185 | |
| 186 | extern unsigned long alpha_read_fp_reg (unsigned long reg); |
| 187 | extern void alpha_write_fp_reg (unsigned long reg, unsigned long val); |
| 188 | extern unsigned long alpha_read_fp_reg_s (unsigned long reg); |
| 189 | extern void alpha_write_fp_reg_s (unsigned long reg, unsigned long val); |
| 190 | |
| 191 | #endif /* __KERNEL__ */ |
| 192 | |
| 193 | #endif /* __ASM_ALPHA_FPU_H */ |