floatundisf.S revision 214152
1109998Smarkm// This file is distributed under the University of Illinois Open Source 2109998Smarkm// License. See LICENSE.TXT for details. 3109998Smarkm 4109998Smarkm#include "../assembly.h" 5109998Smarkm 6109998Smarkm// float __floatundisf(du_int a); 7109998Smarkm 8109998Smarkm// Note that there is a hardware instruction, fildll, that does most of what 9109998Smarkm// this function needs to do. However, because of our ia32 ABI, it will take 10109998Smarkm// a write-small read-large stall, so the software implementation here is 11109998Smarkm// actually several cycles faster. 12109998Smarkm 13109998Smarkm// This is a branch-free implementation. A branchy implementation might be 14109998Smarkm// faster for the common case if you know something a priori about the input 15109998Smarkm// distribution. 16109998Smarkm 17109998Smarkm/* branch-free x87 implementation - one cycle slower than without x87. 18109998Smarkm 19109998Smarkm#ifdef __i386__ 20109998Smarkm 21109998Smarkm.const 22109998Smarkm.align 3 23109998Smarkm 24109998Smarkm .quad 0x43f0000000000000 25109998Smarkmtwop64: .quad 0x0000000000000000 26109998Smarkm 27109998Smarkm#define TWOp64 twop64-0b(%ecx,%eax,8) 28109998Smarkm 29109998Smarkm.text 30109998Smarkm.align 4 31109998SmarkmDEFINE_COMPILERRT_FUNCTION(__floatundisf) 32109998Smarkm movl 8(%esp), %eax 33109998Smarkm movd 8(%esp), %xmm1 34109998Smarkm movd 4(%esp), %xmm0 35109998Smarkm punpckldq %xmm1, %xmm0 36109998Smarkm calll 0f 37109998Smarkm0: popl %ecx 38109998Smarkm sarl $31, %eax 39109998Smarkm movq %xmm0, 4(%esp) 40109998Smarkm fildll 4(%esp) 41109998Smarkm faddl TWOp64 42109998Smarkm fstps 4(%esp) 43109998Smarkm flds 4(%esp) 44109998Smarkm ret 45109998Smarkm 46109998Smarkm#endif // __i386__ 47109998Smarkm 48109998Smarkm*/ 49109998Smarkm 50109998Smarkm/* branch-free, x87-free implementation - faster at the expense of code size */ 51109998Smarkm 52109998Smarkm#ifdef __i386__ 53109998Smarkm 54109998Smarkm#ifndef __ELF__ 55160814Ssimon.const 56109998Smarkm.align 3 57111147Snectar#else 58109998Smarkm.align 8 59111147Snectar#endif 60111147Snectartwop52: .quad 0x4330000000000000 61109998Smarkm .quad 0x0000000000000fff 62109998Smarkmsticky: .quad 0x0000000000000000 63109998Smarkm .long 0x00000012 64109998Smarkmtwelve: .long 0x00000000 65109998Smarkm 66109998Smarkm#define TWOp52 twop52-0b(%ecx) 67109998Smarkm#define STICKY sticky-0b(%ecx,%eax,8) 68109998Smarkm 69109998Smarkm.text 70109998Smarkm.align 4 71109998SmarkmDEFINE_COMPILERRT_FUNCTION(__floatundisf) 72109998Smarkm movl 8(%esp), %eax 73109998Smarkm movd 8(%esp), %xmm1 74109998Smarkm movd 4(%esp), %xmm0 75109998Smarkm punpckldq %xmm1, %xmm0 76109998Smarkm 77109998Smarkm calll 0f 78109998Smarkm0: popl %ecx 79109998Smarkm shrl %eax // high 31 bits of input as sint32 80109998Smarkm addl $0x7ff80000, %eax 81109998Smarkm sarl $31, %eax // (big input) ? -1 : 0 82109998Smarkm movsd STICKY, %xmm1 // (big input) ? 0xfff : 0 83109998Smarkm movl $12, %edx 84109998Smarkm andl %eax, %edx // (big input) ? 12 : 0 85109998Smarkm movd %edx, %xmm3 86109998Smarkm andpd %xmm0, %xmm1 // (big input) ? input & 0xfff : 0 87109998Smarkm movsd TWOp52, %xmm2 // 0x1.0p52 88109998Smarkm psrlq %xmm3, %xmm0 // (big input) ? input >> 12 : input 89109998Smarkm orpd %xmm2, %xmm1 // 0x1.0p52 + ((big input) ? input & 0xfff : input) 90109998Smarkm orpd %xmm1, %xmm0 // 0x1.0p52 + ((big input) ? (input >> 12 | input & 0xfff) : input) 91109998Smarkm subsd %xmm2, %xmm0 // (double)((big input) ? (input >> 12 | input & 0xfff) : input) 92109998Smarkm cvtsd2ss %xmm0, %xmm0 // (float)((big input) ? (input >> 12 | input & 0xfff) : input) 93109998Smarkm pslld $23, %xmm3 94109998Smarkm paddd %xmm3, %xmm0 // (float)input 95109998Smarkm movd %xmm0, 4(%esp) 96109998Smarkm flds 4(%esp) 97109998Smarkm ret 98109998Smarkm 99109998Smarkm#endif // __i386__ 100109998Smarkm