diff options
Diffstat (limited to 'sysdeps/libm-i387/s_cbrt.S')
| -rw-r--r-- | sysdeps/libm-i387/s_cbrt.S | 78 |
1 files changed, 47 insertions, 31 deletions
diff --git a/sysdeps/libm-i387/s_cbrt.S b/sysdeps/libm-i387/s_cbrt.S index 4b9a2b11c6..3f6a0174f2 100644 --- a/sysdeps/libm-i387/s_cbrt.S +++ b/sysdeps/libm-i387/s_cbrt.S @@ -28,34 +28,36 @@ #endif .align ALIGNARG(4) - ASM_TYPE_DIRECTIVE(f1,@object) -f1: .double 0.354895765043919860 - ASM_SIZE_DIRECTIVE(f1) - ASM_TYPE_DIRECTIVE(f2,@object) -f2: .double 1.50819193781584896 - ASM_SIZE_DIRECTIVE(f2) - ASM_TYPE_DIRECTIVE(f3,@object) -f3: .double -2.11499494167371287 - ASM_SIZE_DIRECTIVE(f3) - ASM_TYPE_DIRECTIVE(f4,@object) -f4: .double 2.44693122563534430 - ASM_SIZE_DIRECTIVE(f4) - ASM_TYPE_DIRECTIVE(f5,@object) -f5: .double -1.83469277483613086 - ASM_SIZE_DIRECTIVE(f5) - ASM_TYPE_DIRECTIVE(f6,@object) -f6: .double 0.784932344976639262 - ASM_SIZE_DIRECTIVE(f6) ASM_TYPE_DIRECTIVE(f7,@object) f7: .double -0.145263899385486377 ASM_SIZE_DIRECTIVE(f7) + ASM_TYPE_DIRECTIVE(f6,@object) +f6: .double 0.784932344976639262 + ASM_SIZE_DIRECTIVE(f6) + ASM_TYPE_DIRECTIVE(f5,@object) +f5: .double -1.83469277483613086 + ASM_SIZE_DIRECTIVE(f5) + ASM_TYPE_DIRECTIVE(f4,@object) +f4: .double 2.44693122563534430 + ASM_SIZE_DIRECTIVE(f4) + ASM_TYPE_DIRECTIVE(f3,@object) +f3: .double -2.11499494167371287 + ASM_SIZE_DIRECTIVE(f3) + ASM_TYPE_DIRECTIVE(f2,@object) +f2: .double 1.50819193781584896 + ASM_SIZE_DIRECTIVE(f2) + ASM_TYPE_DIRECTIVE(f1,@object) +f1: .double 0.354895765043919860 + 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 +factor: .double ONE_SQR_CBRT2 + .double ONE_CBRT2 .double 1.0 .double CBRT2 .double SQR_CBRT2 @@ -67,10 +69,10 @@ two54: .byte 0, 0, 0, 0, 0, 0, 0x50, 0x43 #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 @@ -102,8 +104,13 @@ ENTRY(__cbrt) #endif fmull MO(two54) movl $-54, %ecx +#ifdef PIC + fstpl 8(%esp) + movl 12(%esp), %eax +#else fstpl 4(%esp) movl 8(%esp), %eax +#endif movl %eax, %edx andl $0x7fffffff, %eax @@ -123,11 +130,16 @@ ENTRY(__cbrt) #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 */ @@ -161,20 +173,24 @@ ENTRY(__cbrt) 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 12(%esp), %eax popl %ebx +#else + movl 8(%esp), %eax #endif - testl $0x80000000, 8(%esp) - jz 4f + testl %eax, %eax + fstp %st(1) + jns 4f fchs 4: ret |