diff options
Diffstat (limited to 'lib/builtins/hexagon/fastmath2_ldlib_asm.S')
| -rw-r--r-- | lib/builtins/hexagon/fastmath2_ldlib_asm.S | 345 |
1 files changed, 345 insertions, 0 deletions
diff --git a/lib/builtins/hexagon/fastmath2_ldlib_asm.S b/lib/builtins/hexagon/fastmath2_ldlib_asm.S new file mode 100644 index 000000000..419255535 --- /dev/null +++ b/lib/builtins/hexagon/fastmath2_ldlib_asm.S @@ -0,0 +1,345 @@ +//===----------------------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. +// +//===----------------------------------------------------------------------===// +/* ==================================================================== * + +fast2_QLDOUBLE fast2_ldadd(fast2_QLDOUBLE a,fast2_QLDOUBLE b) { + fast2_QLDOUBLE 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_ldadd_asm + .type fast2_ldadd_asm, @function +fast2_ldadd_asm: +#define manta R1:0 +#define lmanta R1:0 +#define mantb R3:2 +#define lmantb R3:2 +#define expa R4 +#define expb R5 +#define expd R6 +#define exp R8 +#define c63 R9 +#define lmant R1:0 +#define k R4 +#define ce P0 +#define zero R3:2 + .falign + { + expa = memw(r29+#8) + expb = memw(r29+#24) + r7 = r0 + } + { + expd = sub(expa, expb):sat + ce = CMP.GT(expa, expb); + if ( ce.new) exp = add(expa, #1) + if (!ce.new) exp = add(expb, #1) + } { + expd = abs(expd):sat + if ( ce) expa = #1 + if (!ce) expb = #1 + c63 = #62 + } { + expd = MIN(expd, c63) + manta = memd(r29+#0) + mantb = memd(r29+#16) + } { + if (!ce) expa = add(expd, #1) + if ( ce) expb = add(expd, #1) + } { + lmanta = ASR(lmanta, expa) + lmantb = ASR(lmantb, expb) + } { + lmant = add(lmanta, lmantb) + zero = #0 + } { + k = clb(lmant) + c63.L =#0x0001 + } { + exp -= add(k, #-1) //exp = exp - (k-1) + k = add(k, #-1) + p0 = cmp.gt(k, #58) + c63.H =#0x8000 + } { + if(!p0)memw(r7+#8) = exp + lmant = ASL(lmant, k) + if(p0) jump .Ldenorma + } { + memd(r7+#0) = lmant + jumpr r31 + } +.Ldenorma: + memd(r7+#0) = zero + { + memw(r7+#8) = c63 + jumpr r31 + } +/* =================================================================== * + fast2_QLDOUBLE fast2_ldsub(fast2_QLDOUBLE a,fast2_QLDOUBLE b) { + fast2_QLDOUBLE 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_ldsub_asm + .type fast2_ldsub_asm, @function +fast2_ldsub_asm: +#define manta R1:0 +#define lmanta R1:0 +#define mantb R3:2 +#define lmantb R3:2 +#define expa R4 +#define expb R5 +#define expd R6 +#define exp R8 +#define c63 R9 +#define lmant R1:0 +#define k R4 +#define ce P0 +#define zero R3:2 + .falign + { + expa = memw(r29+#8) + expb = memw(r29+#24) + r7 = r0 + } + { + expd = sub(expa, expb):sat + ce = CMP.GT(expa, expb); + if ( ce.new) exp = add(expa, #1) + if (!ce.new) exp = add(expb, #1) + } { + expd = abs(expd):sat + if ( ce) expa = #1 + if (!ce) expb = #1 + c63 = #62 + } { + expd = min(expd, c63) + manta = memd(r29+#0) + mantb = memd(r29+#16) + } { + if (!ce) expa = add(expd, #1) + if ( ce) expb = add(expd, #1) + } { + lmanta = ASR(lmanta, expa) + lmantb = ASR(lmantb, expb) + } { + lmant = sub(lmanta, lmantb) + zero = #0 + } { + k = clb(lmant) + c63.L =#0x0001 + } { + exp -= add(k, #-1) //exp = exp - (k+1) + k = add(k, #-1) + p0 = cmp.gt(k, #58) + c63.H =#0x8000 + } { + if(!p0)memw(r7+#8) = exp + lmant = asl(lmant, k) + if(p0) jump .Ldenorma_s + } { + memd(r7+#0) = lmant + jumpr r31 + } +.Ldenorma_s: + memd(r7+#0) = zero + { + memw(r7+#8) = c63 + jumpr r31 + } + +/* ==================================================================== * + fast2_QLDOUBLE fast2_ldmpy(fast2_QLDOUBLE a,fast2_QLDOUBLE b) { + fast2_QLDOUBLE 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_ldmpy_asm + .type fast2_ldmpy_asm, @function +fast2_ldmpy_asm: + +#define mantxl_ R9 +#define mantxl R14 +#define mantxh R15 +#define mantx R15:14 +#define mantbl R2 +#define mantbl_ R8 +#define mantbh R3 +#define mantb R3:2 +#define expa R4 +#define expb R5 +#define c8001 R8 +#define mantd R7:6 +#define lmantc R11:10 +#define kp R9 +#define min R13:12 +#define minh R13 +#define max R13:12 +#define maxh R13 +#define ret R0 + + .falign + { + mantx = memd(r29+#0) + mantb = memd(r29+#16) + min = #0 + } + { + mantbl_= extractu(mantbl, #31, #1) + mantxl_= extractu(mantxl, #31, #1) + minh.H = #0x8000 + } + { + lmantc = mpy(mantxh, mantbh) + mantd = mpy(mantxh, mantbl_) + expa = memw(r29+#8) + expb = memw(r29+#24) + } + { + lmantc = add(lmantc, lmantc) + mantd += mpy(mantbh, mantxl_) + } + { + mantd = asr(mantd, #30) + c8001.L = #0x0001 + p1 = cmp.eq(mantx, mantb) + } + { + mantd = add(mantd, lmantc) + expa= add(expa, expb) + p2 = cmp.eq(mantb, min) + } + { + kp = clb(mantd) + c8001.H = #0x8000 + p1 = and(p1, p2) + } + { + expa-= add(kp, #-1) + kp = add(kp, #-1) + if(p1) jump .Lsat + } + { + mantd = asl(mantd, kp) + memw(ret+#8) = expa + p0 = cmp.gt(kp, #58) + if(p0.new) jump:NT .Ldenorm //rarely happens + } + { + memd(ret+#0) = mantd + jumpr r31 + } +.Lsat: + { + max = #0 + expa+= add(kp, #1) + } + { + maxh.H = #0x4000 + memw(ret+#8) = expa + } + { + memd(ret+#0) = max + jumpr r31 + } +.Ldenorm: + { + memw(ret+#8) = c8001 + mantx = #0 + } + { + memd(ret+#0) = mantx + jumpr r31 + } |