diff options
Diffstat (limited to 'arch/x86/math-emu/reg_u_sub.S')
| -rw-r--r-- | arch/x86/math-emu/reg_u_sub.S | 274 |
1 files changed, 274 insertions, 0 deletions
diff --git a/arch/x86/math-emu/reg_u_sub.S b/arch/x86/math-emu/reg_u_sub.S new file mode 100644 index 000000000..f05dea7de --- /dev/null +++ b/arch/x86/math-emu/reg_u_sub.S @@ -0,0 +1,274 @@ +/* SPDX-License-Identifier: GPL-2.0 */ + .file "reg_u_sub.S" +/*---------------------------------------------------------------------------+ + | reg_u_sub.S | + | | + | Core floating point subtraction routine. | + | | + | Copyright (C) 1992,1993,1995,1997 | + | W. Metzenthen, 22 Parker St, Ormond, Vic 3163, Australia | + | E-mail billm@suburbia.net | + | | + | Call from C as: | + | int FPU_u_sub(FPU_REG *arg1, FPU_REG *arg2, FPU_REG *answ, | + | int control_w) | + | Return value is the tag of the answer, or-ed with FPU_Exception if | + | one was raised, or -1 on internal error. | + | | + +---------------------------------------------------------------------------*/ + +/* + | Kernel subtraction routine FPU_u_sub(reg *arg1, reg *arg2, reg *answ). + | Takes two valid reg f.p. numbers (TAG_Valid), which are + | treated as unsigned numbers, + | and returns their difference as a TAG_Valid or TAG_Zero f.p. + | number. + | The first number (arg1) must be the larger. + | The returned number is normalized. + | Basic checks are performed if PARANOID is defined. + */ + +#include "exception.h" +#include "fpu_emu.h" +#include "control_w.h" + +.text +ENTRY(FPU_u_sub) + pushl %ebp + movl %esp,%ebp + pushl %esi + pushl %edi + pushl %ebx + + movl PARAM1,%esi /* source 1 */ + movl PARAM2,%edi /* source 2 */ + + movl PARAM6,%ecx + subl PARAM7,%ecx /* exp1 - exp2 */ + +#ifdef PARANOID + /* source 2 is always smaller than source 1 */ + js L_bugged_1 + + testl $0x80000000,SIGH(%edi) /* The args are assumed to be be normalized */ + je L_bugged_2 + + testl $0x80000000,SIGH(%esi) + je L_bugged_2 +#endif /* PARANOID */ + +/*--------------------------------------+ + | Form a register holding the | + | smaller number | + +--------------------------------------*/ + movl SIGH(%edi),%eax /* register ms word */ + movl SIGL(%edi),%ebx /* register ls word */ + + movl PARAM3,%edi /* destination */ + movl PARAM6,%edx + movw %dx,EXP(%edi) /* Copy exponent to destination */ + + xorl %edx,%edx /* register extension */ + +/*--------------------------------------+ + | Shift the temporary register | + | right the required number of | + | places. | + +--------------------------------------*/ + + cmpw $32,%cx /* shrd only works for 0..31 bits */ + jnc L_more_than_31 + +/* less than 32 bits */ + shrd %cl,%ebx,%edx + shrd %cl,%eax,%ebx + shr %cl,%eax + jmp L_shift_done + +L_more_than_31: + cmpw $64,%cx + jnc L_more_than_63 + + subb $32,%cl + jz L_exactly_32 + + shrd %cl,%eax,%edx + shr %cl,%eax + orl %ebx,%ebx + jz L_more_31_no_low /* none of the lowest bits is set */ + + orl $1,%edx /* record the fact in the extension */ + +L_more_31_no_low: + movl %eax,%ebx + xorl %eax,%eax + jmp L_shift_done + +L_exactly_32: + movl %ebx,%edx + movl %eax,%ebx + xorl %eax,%eax + jmp L_shift_done + +L_more_than_63: + cmpw $65,%cx + jnc L_more_than_64 + + /* Shift right by 64 bits */ + movl %eax,%edx + orl %ebx,%ebx + jz L_more_63_no_low + + orl $1,%edx + jmp L_more_63_no_low + +L_more_than_64: + jne L_more_than_65 + + /* Shift right by 65 bits */ + /* Carry is clear if we get here */ + movl %eax,%edx + rcrl %edx + jnc L_shift_65_nc + + orl $1,%edx + jmp L_more_63_no_low + +L_shift_65_nc: + orl %ebx,%ebx + jz L_more_63_no_low + + orl $1,%edx + jmp L_more_63_no_low + +L_more_than_65: + movl $1,%edx /* The shifted nr always at least one '1' */ + +L_more_63_no_low: + xorl %ebx,%ebx + xorl %eax,%eax + +L_shift_done: +L_subtr: +/*------------------------------+ + | Do the subtraction | + +------------------------------*/ + xorl %ecx,%ecx + subl %edx,%ecx + movl %ecx,%edx + movl SIGL(%esi),%ecx + sbbl %ebx,%ecx + movl %ecx,%ebx + movl SIGH(%esi),%ecx + sbbl %eax,%ecx + movl %ecx,%eax + +#ifdef PARANOID + /* We can never get a borrow */ + jc L_bugged +#endif /* PARANOID */ + +/*--------------------------------------+ + | Normalize the result | + +--------------------------------------*/ + testl $0x80000000,%eax + jnz L_round /* no shifting needed */ + + orl %eax,%eax + jnz L_shift_1 /* shift left 1 - 31 bits */ + + orl %ebx,%ebx + jnz L_shift_32 /* shift left 32 - 63 bits */ + +/* + * A rare case, the only one which is non-zero if we got here + * is: 1000000 .... 0000 + * -0111111 .... 1111 1 + * -------------------- + * 0000000 .... 0000 1 + */ + + cmpl $0x80000000,%edx + jnz L_must_be_zero + + /* Shift left 64 bits */ + subw $64,EXP(%edi) + xchg %edx,%eax + jmp fpu_reg_round + +L_must_be_zero: +#ifdef PARANOID + orl %edx,%edx + jnz L_bugged_3 +#endif /* PARANOID */ + + /* The result is zero */ + movw $0,EXP(%edi) /* exponent */ + movl $0,SIGL(%edi) + movl $0,SIGH(%edi) + movl TAG_Zero,%eax + jmp L_exit + +L_shift_32: + movl %ebx,%eax + movl %edx,%ebx + movl $0,%edx + subw $32,EXP(%edi) /* Can get underflow here */ + +/* We need to shift left by 1 - 31 bits */ +L_shift_1: + bsrl %eax,%ecx /* get the required shift in %ecx */ + subl $31,%ecx + negl %ecx + shld %cl,%ebx,%eax + shld %cl,%edx,%ebx + shl %cl,%edx + subw %cx,EXP(%edi) /* Can get underflow here */ + +L_round: + jmp fpu_reg_round /* Round the result */ + + +#ifdef PARANOID +L_bugged_1: + pushl EX_INTERNAL|0x206 + call EXCEPTION + pop %ebx + jmp L_error_exit + +L_bugged_2: + pushl EX_INTERNAL|0x209 + call EXCEPTION + pop %ebx + jmp L_error_exit + +L_bugged_3: + pushl EX_INTERNAL|0x210 + call EXCEPTION + pop %ebx + jmp L_error_exit + +L_bugged_4: + pushl EX_INTERNAL|0x211 + call EXCEPTION + pop %ebx + jmp L_error_exit + +L_bugged: + pushl EX_INTERNAL|0x212 + call EXCEPTION + pop %ebx + jmp L_error_exit + +L_error_exit: + movl $-1,%eax + +#endif /* PARANOID */ + +L_exit: + popl %ebx + popl %edi + popl %esi + leave + ret +ENDPROC(FPU_u_sub) |