diff options
Diffstat (limited to 'sysdeps/libm-i387/s_cbrtf.S')
| -rw-r--r-- | sysdeps/libm-i387/s_cbrtf.S | 50 |
1 files changed, 31 insertions, 19 deletions
diff --git a/sysdeps/libm-i387/s_cbrtf.S b/sysdeps/libm-i387/s_cbrtf.S index 6978da2d40..a14e04ed2f 100644 --- a/sysdeps/libm-i387/s_cbrtf.S +++ b/sysdeps/libm-i387/s_cbrtf.S @@ -28,22 +28,25 @@ #endif .align ALIGNARG(4) - ASM_TYPE_DIRECTIVE(f1,@object) -f1: .double 0.492659620528969547 - ASM_SIZE_DIRECTIVE(f1) - ASM_TYPE_DIRECTIVE(f2,@object) -f2: .double 0.697570460207922770 - ASM_SIZE_DIRECTIVE(f2) ASM_TYPE_DIRECTIVE(f3,@object) f3: .double 0.191502161678719066 ASM_SIZE_DIRECTIVE(f3) + ASM_TYPE_DIRECTIVE(f2,@object) +f2: .double 0.697570460207922770 + ASM_SIZE_DIRECTIVE(f2) + ASM_TYPE_DIRECTIVE(f1,@object) +f1: .double 0.492659620528969547 + ASM_SIZE_DIRECTIVE(f1) -#define CBRT2 1.2599210498948731648 -#define SQR_CBRT2 1.5874010519681994748 +#define CBRT2 1.2599210498948731648 +#define ONE_CBRT2 0.793700525984099737355196796584 +#define SQR_CBRT2 1.5874010519681994748 +#define ONE_SQR_CBRT2 0.629960524947436582364439673883 ASM_TYPE_DIRECTIVE(factor,@object) -factor: .double 1.0 / SQR_CBRT2 - .double 1.0 / CBRT2 + .align ALIGNARG(4) +factor: .double ONE_SQR_CBRT2 + .double ONE_CBRT2 .double 1.0 .double CBRT2 .double SQR_CBRT2 @@ -55,10 +58,10 @@ two25: .byte 0, 0, 0, 0x4c #ifdef PIC #define MO(op) op##@GOTOFF(%ebx) -#define MOX(op,x,f) op##@GOTOFF(%ebx,x,f) +#define MOX(op,x) op##@GOTOFF(%ebx,x,1) #else #define MO(op) op -#define MOX(op,x,f) op(,x,f) +#define MOX(op,x) op(x) #endif .text @@ -114,11 +117,16 @@ ENTRY(__cbrtf) #endif fabs - /* The following code has two track: + /* The following code has two tracks: a) compute the normalized cbrt value b) compute xe/3 and xe%3 The right track computes the value for b) and this is done - in an optimized way by avoiding division. */ + in an optimized way by avoiding division. + + But why two tracks at all? Very easy: efficiency. Some FP + instruction can overlap with a certain amount of integer (and + FP) instructions. So we get (except for the imull) all + instructions for free. */ fld %st(0) /* xm : xm */ fmull MO(f3) /* f3*xm : xm */ @@ -142,20 +150,24 @@ ENTRY(__cbrtf) fadd %st(0) /* 2*t2 : t2+2*xm : u : xm */ subl %edx, %ecx faddp %st, %st(3) /* t2+2*xm : u : 2*t2+xm */ + shll $3, %ecx fmulp /* u*(t2+2*xm) : 2*t2+xm */ fdivp %st, %st(1) /* u*(t2+2*xm)/(2*t2+xm) */ - fmull MOX(16+factor,%ecx,8) /* u*(t2+2*xm)/(2*t2+xm)*FACT */ + fmull MOX(16+factor,%ecx) /* u*(t2+2*xm)/(2*t2+xm)*FACT */ pushl %eax fildl (%esp) /* xe/3 : u*(t2+2*xm)/(2*t2+xm)*FACT */ fxch /* u*(t2+2*xm)/(2*t2+xm)*FACT : xe/3 */ - popl %eax fscale /* u*(t2+2*xm)/(2*t2+xm)*FACT*2^xe/3 */ - fstp %st(1) + popl %edx #ifdef PIC + movl 8(%esp), %eax popl %ebx +#else + movl 4(%esp), %eax #endif - testl $0x80000000, 4(%esp) - jz 4f + testl %eax, %eax + fstp %st(1) + jns 4f fchs 4: ret |