diff options
Diffstat (limited to 'lib/builtins/hexagon/fastmath2_dlib_asm.S')
| -rw-r--r-- | lib/builtins/hexagon/fastmath2_dlib_asm.S | 491 |
1 files changed, 491 insertions, 0 deletions
diff --git a/lib/builtins/hexagon/fastmath2_dlib_asm.S b/lib/builtins/hexagon/fastmath2_dlib_asm.S new file mode 100644 index 000000000..9286df06c --- /dev/null +++ b/lib/builtins/hexagon/fastmath2_dlib_asm.S @@ -0,0 +1,491 @@ +//===----------------------Hexagon builtin routine ------------------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is dual licensed under the MIT and the University of Illinois Open +// Source Licenses. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +/* ==================================================================== */ +/* FUNCTIONS Optimized double floating point operators */ +/* ==================================================================== */ +/* c = dadd_asm(a, b) */ +/* ==================================================================== * +fast2_QDOUBLE fast2_dadd(fast2_QDOUBLE a,fast2_QDOUBLE b) { + fast2_QDOUBLE c; + lint manta = a & MANTMASK; + int expa = Q6_R_sxth_R(a) ; + lint mantb = b & MANTMASK; + int expb = Q6_R_sxth_R(b) ; + int exp, expdiff, j, k, hi, lo, cn; + lint mant; + + expdiff = (int) Q6_P_vabsdiffh_PP(a, b); + expdiff = Q6_R_sxth_R(expdiff) ; + if (expdiff > 63) { expdiff = 62;} + if (expa > expb) { + exp = expa + 1; + expa = 1; + expb = expdiff + 1; + } else { + exp = expb + 1; + expb = 1; + expa = expdiff + 1; + } + mant = (manta>>expa) + (mantb>>expb); + + hi = (int) (mant>>32); + lo = (int) (mant); + + k = Q6_R_normamt_R(hi); + if(hi == 0 || hi == -1) k = 31+Q6_R_normamt_R(lo); + + mant = (mant << k); + cn = (mant == 0x8000000000000000LL); + exp = exp - k + cn; + + if (mant == 0 || mant == -1) exp = 0x8001; + c = (mant & MANTMASK) | (((lint) exp) & EXP_MASK); + return(c); + } + * ==================================================================== */ + .text + .global fast2_dadd_asm + .type fast2_dadd_asm, @function +fast2_dadd_asm: +#define manta R0 +#define mantexpa R1:0 +#define lmanta R1:0 +#define mantb R2 +#define mantexpb R3:2 +#define lmantb R3:2 +#define expa R4 +#define expb R5 +#define mantexpd R7:6 +#define expd R6 +#define exp R8 +#define c63 R9 +#define lmant R1:0 +#define manth R1 +#define mantl R0 +#define minmin R11:10 // exactly 0x000000000000008001LL +#define minminl R10 +#define k R4 +#define ce P0 + .falign + { + mantexpd = VABSDIFFH(mantexpa, mantexpb) //represented as 0x08001LL + c63 = #62 + expa = SXTH(manta) + expb = SXTH(mantb) + } { + expd = SXTH(expd) + ce = CMP.GT(expa, expb); + if ( ce.new) exp = add(expa, #1) + if (!ce.new) exp = add(expb, #1) + } { + if ( ce) expa = #1 + if (!ce) expb = #1 + manta.L = #0 + expd = MIN(expd, c63) + } { + if (!ce) expa = add(expd, #1) + if ( ce) expb = add(expd, #1) + mantb.L = #0 + minmin = #0 + } { + lmanta = ASR(lmanta, expa) + lmantb = ASR(lmantb, expb) + } { + lmant = add(lmanta, lmantb) + minminl.L = #0x8001 + } { + k = clb(lmant) + c63 = #58 + } { + k = add(k, #-1) + p0 = cmp.gt(k, c63) + } { + mantexpa = ASL(lmant, k) + exp = SUB(exp, k) + if(p0) jump .Ldenorma + } { + manta = insert(exp, #16, #0) + jumpr r31 + } +.Ldenorma: + { + mantexpa = minmin + jumpr r31 + } +/* =================================================================== * + fast2_QDOUBLE fast2_dsub(fast2_QDOUBLE a,fast2_QDOUBLE b) { + fast2_QDOUBLE c; + lint manta = a & MANTMASK; + int expa = Q6_R_sxth_R(a) ; + lint mantb = b & MANTMASK; + int expb = Q6_R_sxth_R(b) ; + int exp, expdiff, j, k; + lint mant; + + expdiff = (int) Q6_P_vabsdiffh_PP(a, b); + expdiff = Q6_R_sxth_R(expdiff) ; + if (expdiff > 63) { expdiff = 62;} + if (expa > expb) { + exp = expa + 1; + expa = 1; + expb = expdiff + 1; + } else { + exp = expb + 1; + expb = 1; + expa = expdiff + 1; + } + mant = (manta>>expa) - (mantb>>expb); + k = Q6_R_clb_P(mant)-1; + mant = (mant << k); + exp = exp - k; + if (mant == 0 || mant == -1) exp = 0x8001; + c = (mant & MANTMASK) | (((lint) exp) & EXP_MASK); + return(c); + } + * ==================================================================== */ + .text + .global fast2_dsub_asm + .type fast2_dsub_asm, @function +fast2_dsub_asm: + +#define manta R0 +#define mantexpa R1:0 +#define lmanta R1:0 +#define mantb R2 +#define mantexpb R3:2 +#define lmantb R3:2 +#define expa R4 +#define expb R5 +#define mantexpd R7:6 +#define expd R6 +#define exp R8 +#define c63 R9 +#define lmant R1:0 +#define manth R1 +#define mantl R0 +#define minmin R11:10 // exactly 0x000000000000008001LL +#define minminl R10 +#define k R4 +#define ce P0 + .falign + { + mantexpd = VABSDIFFH(mantexpa, mantexpb) //represented as 0x08001LL + c63 = #62 + expa = SXTH(manta) + expb = SXTH(mantb) + } { + expd = SXTH(expd) + ce = CMP.GT(expa, expb); + if ( ce.new) exp = add(expa, #1) + if (!ce.new) exp = add(expb, #1) + } { + if ( ce) expa = #1 + if (!ce) expb = #1 + manta.L = #0 + expd = MIN(expd, c63) + } { + if (!ce) expa = add(expd, #1) + if ( ce) expb = add(expd, #1) + mantb.L = #0 + minmin = #0 + } { + lmanta = ASR(lmanta, expa) + lmantb = ASR(lmantb, expb) + } { + lmant = sub(lmanta, lmantb) + minminl.L = #0x8001 + } { + k = clb(lmant) + c63 = #58 + } { + k = add(k, #-1) + p0 = cmp.gt(k, c63) + } { + mantexpa = ASL(lmant, k) + exp = SUB(exp, k) + if(p0) jump .Ldenorm + } { + manta = insert(exp, #16, #0) + jumpr r31 + } +.Ldenorm: + { + mantexpa = minmin + jumpr r31 + } +/* ==================================================================== * + fast2_QDOUBLE fast2_dmpy(fast2_QDOUBLE a,fast2_QDOUBLE b) { + fast2_QDOUBLE c; + lint manta = a & MANTMASK; + int expa = Q6_R_sxth_R(a) ; + lint mantb = b & MANTMASK; + int expb = Q6_R_sxth_R(b) ; + int exp, k; + lint mant; + int hia, hib, hi, lo; + unsigned int loa, lob; + + hia = (int)(a >> 32); + loa = Q6_R_extractu_RII((int)manta, 31, 1); + hib = (int)(b >> 32); + lob = Q6_R_extractu_RII((int)mantb, 31, 1); + + mant = Q6_P_mpy_RR(hia, lob); + mant = Q6_P_mpyacc_RR(mant,hib, loa); + mant = (mant >> 30) + (Q6_P_mpy_RR(hia, hib)<<1); + + hi = (int) (mant>>32); + + k = Q6_R_normamt_R(hi); + mant = mant << k; + exp = expa + expb - k; + if (mant == 0 || mant == -1) exp = 0x8001; + c = (mant & MANTMASK) | (((lint) exp) & EXP_MASK); + return(c); + } + * ==================================================================== */ + .text + .global fast2_dmpy_asm + .type fast2_dmpy_asm, @function +fast2_dmpy_asm: + +#define mantal R0 +#define mantah R1 +#define mantexpa R1:0 +#define mantbl R2 +#define mantbh R3 +#define mantexpb R3:2 +#define expa R4 +#define expb R5 +#define c8001 R12 +#define mantexpd R7:6 +#define mantdh R7 +#define exp R8 +#define lmantc R11:10 +#define kb R9 +#define guard R11 +#define mantal_ R12 +#define mantbl_ R13 +#define min R15:14 +#define minh R15 + + .falign + { + mantbl_= lsr(mantbl, #16) + expb = sxth(mantbl) + expa = sxth(mantal) + mantal_= lsr(mantal, #16) + } + { + lmantc = mpy(mantah, mantbh) + mantexpd = mpy(mantah, mantbl_) + mantal.L = #0x0 + min = #0 + } + { + lmantc = add(lmantc, lmantc) + mantexpd+= mpy(mantbh, mantal_) + mantbl.L = #0x0 + minh.H = #0x8000 + } + { + mantexpd = asr(mantexpd, #15) + c8001.L = #0x8001 + p1 = cmp.eq(mantexpa, mantexpb) + } + { + mantexpd = add(mantexpd, lmantc) + exp = add(expa, expb) + p2 = cmp.eq(mantexpa, min) + } + { + kb = clb(mantexpd) + mantexpb = abs(mantexpd) + guard = #58 + } + { + p1 = and(p1, p2) + exp = sub(exp, kb) + kb = add(kb, #-1) + p0 = cmp.gt(kb, guard) + } + { + exp = add(exp, #1) + mantexpa = asl(mantexpd, kb) + if(p1) jump .Lsat //rarely happens + } + { + mantal = insert(exp,#16, #0) + if(!p0) jumpr r31 + } + { + mantal = insert(c8001,#16, #0) + jumpr r31 + } +.Lsat: + { + mantexpa = #-1 + } + { + mantexpa = lsr(mantexpa, #1) + } + { + mantal = insert(exp,#16, #0) + jumpr r31 + } + +/* ==================================================================== * + int fast2_qd2f(fast2_QDOUBLE a) { + int exp; + long long int manta; + int ic, rnd, mantb; + + manta = a>>32; + exp = Q6_R_sxth_R(a) ; + ic = 0x80000000 & manta; + manta = Q6_R_abs_R_sat(manta); + mantb = (manta + rnd)>>7; + rnd = 0x40 + exp = (exp + 126); + if((manta & 0xff) == rnd) rnd = 0x00; + if((manta & 0x7fffffc0) == 0x7fffffc0) { + manta = 0x0; exp++; + } else { + manta= mantb & 0x007fffff; + } + exp = (exp << 23) & 0x7fffffc0; + ic = Q6_R_addacc_RR(ic, exp, manta); + return (ic); + } + * ==================================================================== */ + + .text + .global fast2_qd2f_asm + .type fast2_qd2f_asm, @function +fast2_qd2f_asm: +#define mantah R1 +#define mantal R0 +#define cff R0 +#define mant R3 +#define expo R4 +#define rnd R5 +#define mask R6 +#define c07f R7 +#define c80 R0 +#define mantb R2 +#define ic R0 + + .falign + { + mant = abs(mantah):sat + expo = sxth(mantal) + rnd = #0x40 + mask.L = #0xffc0 + } + { + cff = extractu(mant, #8, #0) + p2 = cmp.gt(expo, #126) + p3 = cmp.ge(expo, #-126) + mask.H = #0x7fff + } + { + p1 = cmp.eq(cff,#0x40) + if(p1.new) rnd = #0 + expo = add(expo, #126) + if(!p3) jump .Lmin + } + { + p0 = bitsset(mant, mask) + c80.L = #0x0000 + mantb = add(mant, rnd) + c07f = lsr(mask, #8) + } + { + if(p0) expo = add(expo, #1) + if(p0) mant = #0 + mantb = lsr(mantb, #7) + c80.H = #0x8000 + } + { + ic = and(c80, mantah) + mask &= asl(expo, #23) + if(!p0) mant = and(mantb, c07f) + if(p2) jump .Lmax + } + { + ic += add(mask, mant) + jumpr r31 + } +.Lmax: + { + ic.L = #0xffff; + } + { + ic.H = #0x7f7f; + jumpr r31 + } +.Lmin: + { + ic = #0x0 + jumpr r31 + } + +/* ==================================================================== * +fast2_QDOUBLE fast2_f2qd(int ia) { + lint exp; + lint mant; + fast2_QDOUBLE c; + + mant = ((ia << 7) | 0x40000000)&0x7fffff80 ; + if (ia & 0x80000000) mant = -mant; + exp = ((ia >> 23) & 0xFFLL) - 126; + c = (mant<<32) | Q6_R_zxth_R(exp);; + return(c); +} + * ==================================================================== */ + .text + .global fast2_f2qd_asm + .type fast2_f2qd_asm, @function +fast2_f2qd_asm: +#define ia R0 +#define mag R3 +#define mantr R1 +#define expr R0 +#define zero R2 +#define maxneg R5:4 +#define maxnegl R4 + .falign + { + mantr = asl(ia, #7) + p0 = tstbit(ia, #31) + maxneg = #0 + mag = add(ia,ia) + } + { + mantr = setbit(mantr, #30) + expr= extractu(ia,#8,#23) + maxnegl.L = #0x8001 + p1 = cmp.eq(mag, #0) + } + { + mantr= extractu(mantr, #31, #0) + expr= add(expr, #-126) + zero = #0 + if(p1) jump .Lminqd + } + { + expr = zxth(expr) + if(p0) mantr= sub(zero, mantr) + jumpr r31 + } +.Lminqd: + { + R1:0 = maxneg + jumpr r31 + } |