diff options
Diffstat (limited to 'sysdeps/libm-i387/e_powf.S')
| -rw-r--r-- | sysdeps/libm-i387/e_powf.S | 230 |
1 files changed, 211 insertions, 19 deletions
diff --git a/sysdeps/libm-i387/e_powf.S b/sysdeps/libm-i387/e_powf.S index 9d6bc510b6..54af93c961 100644 --- a/sysdeps/libm-i387/e_powf.S +++ b/sysdeps/libm-i387/e_powf.S @@ -1,5 +1,5 @@ /* ix87 specific implementation of pow function. - Copyright (C) 1996 Free Software Foundation, Inc. + Copyright (C) 1996, 1997 Free Software Foundation, Inc. This file is part of the GNU C Library. Contributed by Ulrich Drepper <drepper@cygnus.com>, 1996. @@ -27,67 +27,107 @@ #endif .align ALIGNARG(4) + ASM_TYPE_DIRECTIVE(infinity,@object) +inf_zero: +infinity: + .byte 0, 0, 0, 0, 0, 0, 0xf0, 0x7f + ASM_SIZE_DIRECTIVE(infinity) + ASM_TYPE_DIRECTIVE(zero,@object) +zero: .double 0.0 + ASM_SIZE_DIRECTIVE(zero) + ASM_TYPE_DIRECTIVE(minf_mzero,@object) +minf_mzero: +minfinity: + .byte 0, 0, 0, 0, 0, 0, 0xf0, 0xff +mzero: + .byte 0, 0, 0, 0, 0, 0, 0, 0x80 + ASM_SIZE_DIRECTIVE(minf_mzero) ASM_TYPE_DIRECTIVE(one,@object) one: .double 1.0 ASM_SIZE_DIRECTIVE(one) ASM_TYPE_DIRECTIVE(limit,@object) limit: .double 0.29 ASM_SIZE_DIRECTIVE(limit) + ASM_TYPE_DIRECTIVE(nan,@object) +nan: .byte 0, 0, 0, 0, 0, 0, 0xff, 0x7f + ASM_SIZE_DIRECTIVE(nan) #ifdef PIC #define MO(op) op##@GOTOFF(%ecx) +#define MOX(op,x,f) op##@GOTOFF(%ecx,x,f) #else #define MO(op) op +#define MOX(op,x,f) op(,x,f) #endif .text ENTRY(__ieee754_powf) - flds 4(%esp) // x - flds 8(%esp) // y : x + flds 8(%esp) // y + fxam + fnstsw + movb %ah, %dl + andb $0x45, %ah + cmpb $0x40, %ah // is y == 0 ? + je 11f + + cmpb $0x05, %ah // is y == ±inf ? + je 12f + + cmpb $0x01, %ah // is y == NaN ? + je 30f #ifdef PIC call 1f 1: popl %ecx addl $_GLOBAL_OFFSET_TABLE_+[.-1b], %ecx #endif - subl $8,%esp + + flds 4(%esp) // x : y + + subl $4, %esp + + fxam + fnstsw + movb %ah, %dh + andb $0x45, %ah + cmpb $0x40, %ah + je 20f // x is ±0 + + cmpb $0x05, %ah + je 15f // x is ±inf + + fxch // y : x /* First see whether `y' is a natural number. In this case we can use a more precise algorithm. */ fld %st // y : y : x - fistpll (%esp) // y : x - fildll (%esp) // int(y) : y : x + fistpl (%esp) // y : x + fildl (%esp) // int(y) : y : x fucomp %st(1) // y : x fnstsw sahf jne 2f /* OK, we have an integer value for y. */ - ftst // y : x - fstp %st(0) // x - fnstsw - sahf - popl %eax popl %edx - jnc 4f // y >= 0, jump + orl $0, %edx + fstp %st(0) // x + jns 4f // y >= 0, jump fdivrl MO(one) // 1/x (now referred to as x) - negl %eax - adcl $0, %edx negl %edx 4: fldl MO(one) // 1 : x fxch -6: shrdl $1, %edx, %eax +6: shrl $1, %edx jnc 5f fxch fmul %st(1) // x : ST*x fxch 5: fmul %st(0), %st // x*x : ST*x - movl %eax, %ecx - orl %edx, %ecx + testl %edx, %edx jnz 6b fstp %st(0) // ST*x - ret +30: ret .align ALIGNARG(4) 2: /* y is a real number. */ @@ -114,7 +154,159 @@ ENTRY(__ieee754_powf) f2xm1 // 2^fract(y*log2(x))-1 : int(y*log2(x)) faddl MO(one) // 2^fract(y*log2(x)) : int(y*log2(x)) fscale // 2^fract(y*log2(x))*2^int(y*log2(x)) : int(y*log2(x)) - addl $8, %esp + addl $4, %esp fstp %st(1) // 2^fract(y*log2(x))*2^int(y*log2(x)) ret + + + // pow(x,±0) = 1 + .align ALIGNARG(4) +11: fstp %st(0) // pop y + fldl MO(one) + ret + + // y == ±inf + .align ALIGNARG(4) +12: fstp %st(0) // pop y + flds 4(%esp) // x + fabs + fcompl MO(one) // < 1, == 1, or > 1 + fnstsw + andb $0x45, %ah + cmpb $0x45, %ah + je 13f // jump if x is NaN + + cmpb $0x40, %ah + je 14f // jump if |x| == 1 + + shlb $1, %ah + xorb %ah, %dl + andl $2, %edx + fldl MOX(inf_zero, %edx, 4) + ret + + .align ALIGNARG(4) +14: fldl MO(nan) + faddl MO(zero) // raise invalid exception + ret + + .align ALIGNARG(4) +13: flds 4(%esp) // load x == NaN + ret + + .align ALIGNARG(4) + // x is ±inf +15: fstp %st(0) // y + testb $2, %dh + jz 16f // jump if x == +inf + + // We must find out whether y is an odd integer. + fld %st // y : y + fistpl (%esp) // y + fildl (%esp) // int(y) : y + fucompp // <empty> + fnstsw + sahf + jne 17f + + // OK, the value is an integer, but is the number of bits small + // enough so that all are coming from the mantissa? + popl %edx + testb $1, %dl + jz 18f // jump if not odd + movl %edx, %eax + orl %edx, %edx + jns 155f + negl %eax +155: cmpl $0x01000000, %eax + ja 18f // does not fit in mantissa bits + // It's an odd integer. + shrl $31, %edx + fldl MOX(minf_mzero, %edx, 8) + ret + + .align ALIGNARG(4) +16: fcompl MO(zero) + addl $4, %esp + fnstsw + shrl $5, %eax + andl $8, %eax + fldl MOX(inf_zero, %eax, 1) + ret + + .align ALIGNARG(4) +17: shll $30, %edx // sign bit for y in right position + addl $4, %esp +18: shrl $31, %edx + fldl MOX(inf_zero, %edx, 8) + ret + + .align ALIGNARG(4) + // x is ±0 +20: fstp %st(0) // y + testb $2, %dl + jz 21f // y > 0 + + // x is ±0 and y is < 0. We must find out whether y is an odd integer. + testb $2, %dh + jz 25f + + fld %st // y : y + fistpl (%esp) // y + fildl (%esp) // int(y) : y + fucompp // <empty> + fnstsw + sahf + jne 26f + + // OK, the value is an integer, but is the number of bits small + // enough so that all are coming from the mantissa? + popl %edx + testb $1, %dl + jz 27f // jump if not odd + cmpl $0xff000000, %edx + jbe 27f // does not fit in mantissa bits + // It's an odd integer. + // Raise divide-by-zero exception and get minus infinity value. + fldl MO(one) + fdivl MO(zero) + fchs + ret + +25: fstp %st(0) +26: popl %eax +27: // Raise divide-by-zero exception and get infinity value. + fldl MO(one) + fdivl MO(zero) + ret + + .align ALIGNARG(4) + // x is ±0 and y is > 0. We must find out whether y is an odd integer. +21: testb $2, %dh + jz 22f + + fld %st // y : y + fistpl (%esp) // y + fildl (%esp) // int(y) : y + fucompp // <empty> + fnstsw + sahf + jne 23f + + // OK, the value is an integer, but is the number of bits small + // enough so that all are coming from the mantissa? + popl %edx + testb $1, %dl + jz 24f // jump if not odd + cmpl $0xff000000, %edx + jbe 24f // does not fit in mantissa bits + // It's an odd integer. + fldl MO(mzero) + ret + +22: fstp %st(0) +23: popl %eax +24: fldl MO(zero) + ret + END(__ieee754_powf) |