From aa95fc13f5b02044eadc3af3d9e1c025f2e1edda Mon Sep 17 00:00:00 2001 From: Anton Blanchard Date: Tue, 28 Jun 2016 21:59:40 +1000 Subject: powerpc: Add a POWER8-optimized version of sinf() This uses the implementation of sinf() in sysdeps/x86_64/fpu/s_sinf.S as inspiration. --- sysdeps/powerpc/powerpc64/power8/fpu/s_sinf.S | 519 ++++++++++++++++++++++++++ 1 file changed, 519 insertions(+) create mode 100644 sysdeps/powerpc/powerpc64/power8/fpu/s_sinf.S (limited to 'sysdeps/powerpc/powerpc64/power8') diff --git a/sysdeps/powerpc/powerpc64/power8/fpu/s_sinf.S b/sysdeps/powerpc/powerpc64/power8/fpu/s_sinf.S new file mode 100644 index 0000000000..3b8f5af292 --- /dev/null +++ b/sysdeps/powerpc/powerpc64/power8/fpu/s_sinf.S @@ -0,0 +1,519 @@ +/* Optimized sinf(). PowerPC64/POWER8 version. + Copyright (C) 2016 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + . */ + +#include +#define _ERRNO_H 1 +#include + +#define FRAMESIZE (FRAME_MIN_SIZE+16) + +#define FLOAT_EXPONENT_SHIFT 23 +#define FLOAT_EXPONENT_BIAS 127 +#define INTEGER_BITS 3 + +#define PI_4 0x3f490fdb /* PI/4 */ +#define NINEPI_4 0x40e231d6 /* 9 * PI/4 */ +#define TWO_PN5 0x3d000000 /* 2^-5 */ +#define TWO_PN27 0x32000000 /* 2^-27 */ +#define INFINITY 0x7f800000 +#define TWO_P23 0x4b000000 /* 2^27 */ +#define FX_FRACTION_1_28 0x9249250 /* 0x100000000 / 28 + 1 */ + + /* Implements the function + + float [fp1] sinf (float [fp1] x) */ + + .machine power8 +EALIGN(__sinf, 4, 0) + addis r9,r2,L(anchor)@toc@ha + addi r9,r9,L(anchor)@toc@l + + lis r4,PI_4@h + ori r4,r4,PI_4@l + + xscvdpspn v0,v1 + mfvsrd r8,v0 + rldicl r3,r8,32,33 /* Remove sign bit. */ + + cmpw r3,r4 + bge L(greater_or_equal_pio4) + + lis r4,TWO_PN5@h + ori r4,r4,TWO_PN5@l + + cmpw r3,r4 + blt L(less_2pn5) + + /* Chebyshev polynomial of the form: + * x+x^3*(S0+x^2*(S1+x^2*(S2+x^2*(S3+x^2*S4)))). */ + + lfd fp9,(L(S0)-L(anchor))(r9) + lfd fp10,(L(S1)-L(anchor))(r9) + lfd fp11,(L(S2)-L(anchor))(r9) + lfd fp12,(L(S3)-L(anchor))(r9) + lfd fp13,(L(S4)-L(anchor))(r9) + + fmul fp2,fp1,fp1 /* x^2 */ + fmul fp3,fp2,fp1 /* x^3 */ + + fmadd fp4,fp2,fp13,fp12 /* S3+x^2*S4 */ + fmadd fp4,fp2,fp4,fp11 /* S2+x^2*(S3+x^2*S4) */ + fmadd fp4,fp2,fp4,fp10 /* S1+x^2*(S2+x^2*(S3+x^2*S4)) */ + fmadd fp4,fp2,fp4,fp9 /* S0+x^2*(S1+x^2*(S2+x^2*(S3+x^2*S4))) */ + fmadd fp1,fp3,fp4,fp1 /* x+x^3*(S0+x^2*(S1+x^2*(S2+x^2*(S3+x^2*S4)))) */ + frsp fp1,fp1 /* Round to single precision. */ + + blr + + .balign 16 +L(greater_or_equal_pio4): + lis r4,NINEPI_4@h + ori r4,r4,NINEPI_4@l + cmpw r3,r4 + bge L(greater_or_equal_9pio4) + + /* Calculate quotient of |x|/(PI/4). */ + lfd fp2,(L(invpio4)-L(anchor))(r9) + fabs fp1,fp1 /* |x| */ + fmul fp2,fp1,fp2 /* |x|/(PI/4) */ + fctiduz fp2,fp2 + mfvsrd r3,v2 /* n = |x| mod PI/4 */ + + /* Now use that quotient to find |x| mod (PI/2). */ + addi r7,r3,1 + rldicr r5,r7,2,60 /* ((n+1) >> 1) << 3 */ + addi r6,r9,(L(pio2_table)-L(anchor)) + lfdx fp4,r5,r6 + fsub fp1,fp1,fp4 + + .balign 16 +L(reduced): + /* Now we are in the range -PI/4 to PI/4. */ + + /* Work out if we are in a positive or negative primary interval. */ + rldicl r4,r7,62,63 /* ((n+1) >> 2) & 1 */ + + /* We are operating on |x|, so we need to add back the original + sign. */ + rldicl r8,r8,33,63 /* (x >> 31) & 1, ie the sign bit. */ + xor r4,r4,r8 /* 0 if result should be positive, + 1 if negative. */ + + /* Load a 1.0 or -1.0. */ + addi r5,r9,(L(ones)-L(anchor)) + sldi r4,r4,3 + lfdx fp0,r4,r5 + + /* Are we in the primary interval of sin or cos? */ + andi. r4,r7,0x2 + bne L(cos) + + /* Chebyshev polynomial of the form: + x+x^3*(S0+x^2*(S1+x^2*(S2+x^2*(S3+x^2*S4)))). */ + + lfd fp9,(L(S0)-L(anchor))(r9) + lfd fp10,(L(S1)-L(anchor))(r9) + lfd fp11,(L(S2)-L(anchor))(r9) + lfd fp12,(L(S3)-L(anchor))(r9) + lfd fp13,(L(S4)-L(anchor))(r9) + + fmul fp2,fp1,fp1 /* x^2 */ + fmul fp3,fp2,fp1 /* x^3 */ + + fmadd fp4,fp2,fp13,fp12 /* S3+x^2*S4 */ + fmadd fp4,fp2,fp4,fp11 /* S2+x^2*(S3+x^2*S4) */ + fmadd fp4,fp2,fp4,fp10 /* S1+x^2*(S2+x^2*(S3+x^2*S4)) */ + fmadd fp4,fp2,fp4,fp9 /* S0+x^2*(S1+x^2*(S2+x^2*(S3+x^2*S4))) */ + fmadd fp4,fp3,fp4,fp1 /* x+x^3*(S0+x^2*(S1+x^2*(S2+x^2*(S3+x^2*S4)))) */ + fmul fp4,fp4,fp0 /* Add in the sign. */ + frsp fp1,fp4 /* Round to single precision. */ + + blr + + .balign 16 +L(cos): + /* Chebyshev polynomial of the form: + 1.0+x^2*(C0+x^2*(C1+x^2*(C2+x^2*(C3+x^2*C4)))). */ + + lfd fp9,(L(C0)-L(anchor))(r9) + lfd fp10,(L(C1)-L(anchor))(r9) + lfd fp11,(L(C2)-L(anchor))(r9) + lfd fp12,(L(C3)-L(anchor))(r9) + lfd fp13,(L(C4)-L(anchor))(r9) + + fmul fp2,fp1,fp1 /* x^2 */ + lfd fp3,(L(DPone)-L(anchor))(r9) + + fmadd fp4,fp2,fp13,fp12 /* C3+x^2*C4 */ + fmadd fp4,fp2,fp4,fp11 /* C2+x^2*(C3+x^2*C4) */ + fmadd fp4,fp2,fp4,fp10 /* C1+x^2*(C2+x^2*(C3+x^2*C4)) */ + fmadd fp4,fp2,fp4,fp9 /* C0+x^2*(C1+x^2*(C2+x^2*(C3+x^2*C4))) */ + fmadd fp4,fp2,fp4,fp3 /* 1.0 + x^2*(C0+x^2*(C1+x^2*(C2+x^2*(C3+x^2*C4)))) */ + fmul fp4,fp4,fp0 /* Add in the sign. */ + frsp fp1,fp4 /* Round to single precision. */ + + blr + + .balign 16 +L(greater_or_equal_9pio4): + lis r4,INFINITY@h + ori r4,r4,INFINITY@l + cmpw r3,r4 + bge L(inf_or_nan) + + lis r4,TWO_P23@h + ori r4,r4,TWO_P23@l + cmpw r3,r4 + bge L(greater_or_equal_2p23) + + fabs fp1,fp1 /* |x| */ + + /* Calculate quotient of |x|/(PI/4). */ + lfd fp2,(L(invpio4)-L(anchor))(r9) + + lfd fp3,(L(DPone)-L(anchor))(r9) + lfd fp4,(L(DPhalf)-L(anchor))(r9) + fmul fp2,fp1,fp2 /* |x|/(PI/4) */ + friz fp2,fp2 /* n = floor(|x|/(PI/4)) */ + + /* Calculate (n + 1) / 2. */ + fadd fp2,fp2,fp3 /* n + 1 */ + fmul fp3,fp2,fp4 /* (n + 1) / 2 */ + friz fp3,fp3 + + lfd fp4,(L(pio2hi)-L(anchor))(r9) + lfd fp5,(L(pio2lo)-L(anchor))(r9) + + fmul fp6,fp4,fp3 + fadd fp6,fp6,fp1 + fmadd fp1,fp5,fp3,fp6 + + fctiduz fp2,fp2 + mfvsrd r7,v2 /* n + 1 */ + + b L(reduced) + + .balign 16 +L(inf_or_nan): + bne L(skip_errno_setting) /* Is a NAN? */ + + /* We delayed the creation of the stack frame, as well as the saving of + the link register, because only at this point, we are sure that + doing so is actually needed. */ + + stfd fp1,-8(r1) + + /* Save the link register. */ + mflr r0 + std r0,16(r1) + cfi_offset(lr, 16) + + /* Create the stack frame. */ + stdu r1,-FRAMESIZE(r1) + cfi_adjust_cfa_offset(FRAMESIZE) + + bl JUMPTARGET(__errno_location) + nop + + /* Restore the stack frame. */ + addi r1,r1,FRAMESIZE + cfi_adjust_cfa_offset(-FRAMESIZE) + /* Restore the link register. */ + ld r0,16(r1) + mtlr r0 + + lfd fp1,-8(r1) + + /* errno = EDOM */ + li r4,EDOM + stw r4,0(r3) + +L(skip_errno_setting): + fsub fp1,fp1,fp1 /* x - x */ + blr + + .balign 16 +L(greater_or_equal_2p23): + fabs fp1,fp1 + + srwi r4,r3,FLOAT_EXPONENT_SHIFT + subi r4,r4,FLOAT_EXPONENT_BIAS + + /* We reduce the input modulo pi/4, so we need 3 bits of integer + to determine where in 2*pi we are. Index into our array + accordingly. */ + addi r4,r4,INTEGER_BITS + + /* To avoid an expensive divide, for the range we care about (0 - 127) + we can transform x/28 into: + + x/28 = (x * ((0x100000000 / 28) + 1)) >> 32 + + mulhwu returns the top 32 bits of the 64 bit result, doing the + shift for us in the same instruction. The top 32 bits are undefined, + so we have to mask them. */ + + lis r6,FX_FRACTION_1_28@h + ori r6,r6,FX_FRACTION_1_28@l + mulhwu r5,r4,r6 + clrldi r5,r5,32 + + /* Get our pointer into the invpio4_table array. */ + sldi r4,r5,3 + addi r6,r9,(L(invpio4_table)-L(anchor)) + add r4,r4,r6 + + lfd fp2,0(r4) + lfd fp3,8(r4) + lfd fp4,16(r4) + lfd fp5,24(r4) + + fmul fp6,fp2,fp1 + fmul fp7,fp3,fp1 + fmul fp8,fp4,fp1 + fmul fp9,fp5,fp1 + + /* Mask off larger integer bits in highest double word that we don't + care about to avoid losing precision when combining with smaller + values. */ + fctiduz fp10,fp6 + mfvsrd r7,v10 + rldicr r7,r7,0,(63-INTEGER_BITS) + mtvsrd v10,r7 + fcfidu fp10,fp10 /* Integer bits. */ + + fsub fp6,fp6,fp10 /* highest -= integer bits */ + + /* Work out the integer component, rounded down. Use the top two + limbs for this. */ + fadd fp10,fp6,fp7 /* highest + higher */ + + fctiduz fp10,fp10 + mfvsrd r7,v10 + andi. r0,r7,1 + fcfidu fp10,fp10 + + /* Subtract integer component from highest limb. */ + fsub fp12,fp6,fp10 + + beq L(even_integer) + + /* Our integer component is odd, so we are in the -PI/4 to 0 primary + region. We need to shift our result down by PI/4, and to do this + in the mod (4/PI) space we simply subtract 1. */ + lfd fp11,(L(DPone)-L(anchor))(r9) + fsub fp12,fp12,fp11 + + /* Now add up all the limbs in order. */ + fadd fp12,fp12,fp7 + fadd fp12,fp12,fp8 + fadd fp12,fp12,fp9 + + /* And finally multiply by pi/4. */ + lfd fp13,(L(pio4)-L(anchor))(r9) + fmul fp1,fp12,fp13 + + addi r7,r7,1 + b L(reduced) + +L(even_integer): + lfd fp11,(L(DPone)-L(anchor))(r9) + + /* Now add up all the limbs in order. */ + fadd fp12,fp12,fp7 + fadd fp12,r12,fp8 + fadd fp12,r12,fp9 + + /* We need to check if the addition of all the limbs resulted in us + overflowing 1.0. */ + fcmpu 0,fp12,fp11 + bgt L(greater_than_one) + + /* And finally multiply by pi/4. */ + lfd fp13,(L(pio4)-L(anchor))(r9) + fmul fp1,fp12,fp13 + + addi r7,r7,1 + b L(reduced) + +L(greater_than_one): + /* We did overflow 1.0 when adding up all the limbs. Add 1.0 to our + integer, and subtract 1.0 from our result. Since that makes the + integer component odd, we need to subtract another 1.0 as + explained above. */ + addi r7,r7,1 + + lfd fp11,(L(DPtwo)-L(anchor))(r9) + fsub fp12,fp12,fp11 + + /* And finally multiply by pi/4. */ + lfd fp13,(L(pio4)-L(anchor))(r9) + fmul fp1,fp12,fp13 + + addi r7,r7,1 + b L(reduced) + + .balign 16 +L(less_2pn5): + lis r4,TWO_PN27@h + ori r4,r4,TWO_PN27@l + + cmpw r3,r4 + blt L(less_2pn27) + + /* A simpler Chebyshev approximation is close enough for this range: + x+x^3*(SS0+x^2*SS1). */ + + lfd fp10,(L(SS0)-L(anchor))(r9) + lfd fp11,(L(SS1)-L(anchor))(r9) + + fmul fp2,fp1,fp1 /* x^2 */ + fmul fp3,fp2,fp1 /* x^3 */ + + fmadd fp4,fp2,fp11,fp10 /* SS0+x^2*SS1 */ + fmadd fp1,fp3,fp4,fp1 /* x+x^3*(SS0+x^2*SS1) */ + + frsp fp1,fp1 /* Round to single precision. */ + + blr + + .balign 16 +L(less_2pn27): + cmpwi r3,0 + beq L(zero) + + /* Handle some special cases: + + sinf(subnormal) raises inexact/underflow + sinf(min_normalized) raises inexact/underflow + sinf(normalized) raises inexact. */ + + lfd fp2,(L(small)-L(anchor))(r9) + + fmul fp2,fp1,fp2 /* x * small */ + fsub fp1,fp1,fp2 /* x - x * small */ + + frsp fp1,fp1 + + blr + + .balign 16 +L(zero): + blr + +END (__sinf) + + .section .rodata, "a" + + .balign 8 + +L(anchor): + + /* Chebyshev constants for sin, range -PI/4 - PI/4. */ +L(S0): .8byte 0xbfc5555555551cd9 +L(S1): .8byte 0x3f81111110c2688b +L(S2): .8byte 0xbf2a019f8b4bd1f9 +L(S3): .8byte 0x3ec71d7264e6b5b4 +L(S4): .8byte 0xbe5a947e1674b58a + + /* Chebyshev constants for sin, range 2^-27 - 2^-5. */ +L(SS0): .8byte 0xbfc555555543d49d +L(SS1): .8byte 0x3f8110f475cec8c5 + + /* Chebyshev constants for cos, range -PI/4 - PI/4. */ +L(C0): .8byte 0xbfdffffffffe98ae +L(C1): .8byte 0x3fa55555545c50c7 +L(C2): .8byte 0xbf56c16b348b6874 +L(C3): .8byte 0x3efa00eb9ac43cc0 +L(C4): .8byte 0xbe923c97dd8844d7 + +L(invpio2): + .8byte 0x3fe45f306dc9c883 /* 2/PI */ + +L(invpio4): + .8byte 0x3ff45f306dc9c883 /* 4/PI */ + +L(invpio4_table): + .8byte 0x0000000000000000 + .8byte 0x3ff45f306c000000 + .8byte 0x3e3c9c882a000000 + .8byte 0x3c54fe13a8000000 + .8byte 0x3aaf47d4d0000000 + .8byte 0x38fbb81b6c000000 + .8byte 0x3714acc9e0000000 + .8byte 0x3560e4107c000000 + .8byte 0x33bca2c756000000 + .8byte 0x31fbd778ac000000 + .8byte 0x300b7246e0000000 + .8byte 0x2e5d2126e8000000 + .8byte 0x2c97003248000000 + .8byte 0x2ad77504e8000000 + .8byte 0x290921cfe0000000 + .8byte 0x274deb1cb0000000 + .8byte 0x25829a73e0000000 + .8byte 0x23fd1046be000000 + .8byte 0x2224baed10000000 + .8byte 0x20709d338e000000 + .8byte 0x1e535a2f80000000 + .8byte 0x1cef904e64000000 + .8byte 0x1b0d639830000000 + .8byte 0x1964ce7d24000000 + .8byte 0x17b908bf16000000 + +L(pio4): + .8byte 0x3fe921fb54442d18 /* PI/4 */ + +/* PI/2 as a sum of two doubles. We only use 32 bits of the upper limb + to avoid losing significant bits when multiplying with up to + (2^22)/(pi/2). */ +L(pio2hi): + .8byte 0xbff921fb54400000 + +L(pio2lo): + .8byte 0xbdd0b4611a626332 + +L(pio2_table): + .8byte 0 + .8byte 0x3ff921fb54442d18 /* 1 * PI/2 */ + .8byte 0x400921fb54442d18 /* 2 * PI/2 */ + .8byte 0x4012d97c7f3321d2 /* 3 * PI/2 */ + .8byte 0x401921fb54442d18 /* 4 * PI/2 */ + .8byte 0x401f6a7a2955385e /* 5 * PI/2 */ + .8byte 0x4022d97c7f3321d2 /* 6 * PI/2 */ + .8byte 0x4025fdbbe9bba775 /* 7 * PI/2 */ + .8byte 0x402921fb54442d18 /* 8 * PI/2 */ + .8byte 0x402c463abeccb2bb /* 9 * PI/2 */ + .8byte 0x402f6a7a2955385e /* 10 * PI/2 */ + +L(small): + .8byte 0x3cd0000000000000 /* 2^-50 */ + +L(ones): + .8byte 0x3ff0000000000000 /* +1.0 */ + .8byte 0xbff0000000000000 /* -1.0 */ + +L(DPhalf): + .8byte 0x3fe0000000000000 /* 0.5 */ + +L(DPone): + .8byte 0x3ff0000000000000 /* 1.0 */ + +L(DPtwo): + .8byte 0x4000000000000000 /* 2.0 */ + +weak_alias(__sinf, sinf) -- cgit 1.4.1