diff options
Diffstat (limited to 'sysdeps/powerpc/powerpc32/power4/fpu/s_llround.S')
| -rw-r--r-- | sysdeps/powerpc/powerpc32/power4/fpu/s_llround.S | 71 |
1 files changed, 45 insertions, 26 deletions
diff --git a/sysdeps/powerpc/powerpc32/power4/fpu/s_llround.S b/sysdeps/powerpc/powerpc32/power4/fpu/s_llround.S index 952d2aa6a5..4b1691efd3 100644 --- a/sysdeps/powerpc/powerpc32/power4/fpu/s_llround.S +++ b/sysdeps/powerpc/powerpc32/power4/fpu/s_llround.S @@ -1,5 +1,5 @@ /* llround function. PowerPC32 on PowerPC64 version. - Copyright (C) 2004, 2006 Free Software Foundation, Inc. + Copyright (C) 2004, 2006, 2007 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 @@ -20,15 +20,15 @@ #include <sysdep.h> #include <math_ldbl_opt.h> - .section .rodata.cst8,"aM",@progbits,8 - .align 2 -.LC0: /* 0.0 */ + .section .rodata.cst12,"aM",@progbits,12 + .align 3 + .LC0: /* 0x1.0000000000000p+52 == 2^52 */ + .long 0x43300000 .long 0x00000000 -.LC1: /* 0.5 */ - .long 0x3f000000 + .long 0x3f000000 /* Use this for 0.5 */ .section ".text" - + /* long [r3] lround (float x [fp1]) IEEE 1003.1 lround function. IEEE specifies "round to the nearest integer value, rounding halfway cases away from zero, regardless of @@ -37,7 +37,15 @@ tie, choose the one that is even (least significant bit o).". So we can't use the PowerPC "round to Nearest" mode. Instead we set "round toward Zero" mode and round by adding +-0.5 before rounding - to the integer value. */ + to the integer value. + + It is necessary to detect when x is (+-)0x1.fffffffffffffp-2 + because adding +-0.5 in this case will cause an erroneous shift, + carry and round. We simply return 0 if 0.5 > x > -0.5. Likewise + if x is and odd number between +-(2^52 and 2^53-1) a shift and + carry will erroneously round if biased with +-0.5. Therefore if x + is greater/less than +-2^52 we don't need to bias the number with + +-0.5. */ ENTRY (__llround) stwu r1,-16(r1) @@ -57,30 +65,41 @@ ENTRY (__llround) # endif mtlr r11 cfi_same_value (lr) - lfs fp12,0(r9) - lfs fp10,.LC1-.LC0(r9) + lfd fp9,0(r9) + lfs fp10,8(r9) #else - lis r9,.LC0@ha - lis r10,.LC1@ha - lfs fp12,.LC0@l(r9) - lfs fp10,.LC1@l(r10) + lis r9,.LC0@ha + lfd fp9,.LC0@l(r9) /* Load 2^52 into fpr9. */ + lfs fp10,.LC0@l+8(r9) /* Load 0.5 into fpr10. */ #endif - fcmpu cr6,fp1,fp12 /* if (x > 0.0) */ - ble- cr6,.L4 - fadd fp1,fp1,fp10 /* x+= 0.5; */ -.L9: - fctidz fp2,fp1 /* Convert To Integer DW round toward 0. */ - stfd fp2,8(r1) - nop /* Ensure the following load is in a different dispatch */ - nop /* group to avoid pipe stall on POWER4&5. */ + fabs fp2,fp1 /* Get the absolute value of x. */ + fsub fp12,fp10,fp10 /* Compute 0.0 into fpr12. */ + fcmpu cr6,fp2,fp10 /* if |x| < 0.5 */ + fcmpu cr4,fp2,fp9 /* if |x| >= 2^52 */ + fcmpu cr3,fp1,fp12 /* x is negative? x < 0.0 */ + blt- cr6,.Lretzero /* 0.5 > x < -0.5 so just return 0. */ + bge- cr4,.Lnobias /* 2^52 > x < -2^52 just convert with no bias. */ + fadd fp3,fp2,fp10 /* |x|+=0.5 bias to prepare to round. */ + bge cr3,.Lconvert /* x is positive so don't negate x. */ + fnabs fp3,fp3 /* -(|x|+=0.5) */ +.Lconvert: + fctidz fp4,fp3 /* Convert to Integer double word round toward 0. */ + stfd fp4,8(r1) + nop + nop nop - lwz r4,12(r1) + lwz r4,12(r1) /* Load return as integer. */ lwz r3,8(r1) +.Lout: addi r1,r1,16 blr -.L4: - fsub fp1,fp1,fp10 /* x-= 0.5; */ - b .L9 +.Lretzero: /* 0.5 > x > -0.5 */ + li r3,0 /* return 0. */ + li r4,0 + b .Lout +.Lnobias: + fmr fp3,fp1 + b .Lconvert END (__llround) weak_alias (__llround, llround) |