1222656Sed// This file is dual licensed under the MIT and the University of Illinois Open 2222656Sed// Source Licenses. See LICENSE.TXT for details. 3214152Sed 4214152Sed#include "../assembly.h" 5214152Sed 6214152Sed// float __floatundisf(du_int a); 7214152Sed 8214152Sed// Note that there is a hardware instruction, fildll, that does most of what 9214152Sed// this function needs to do. However, because of our ia32 ABI, it will take 10214152Sed// a write-small read-large stall, so the software implementation here is 11214152Sed// actually several cycles faster. 12214152Sed 13214152Sed// This is a branch-free implementation. A branchy implementation might be 14214152Sed// faster for the common case if you know something a priori about the input 15214152Sed// distribution. 16214152Sed 17214152Sed/* branch-free x87 implementation - one cycle slower than without x87. 18214152Sed 19214152Sed#ifdef __i386__ 20214152Sed 21214152Sed.const 22214152Sed.align 3 23214152Sed 24214152Sed .quad 0x43f0000000000000 25214152Sedtwop64: .quad 0x0000000000000000 26214152Sed 27214152Sed#define TWOp64 twop64-0b(%ecx,%eax,8) 28214152Sed 29214152Sed.text 30214152Sed.align 4 31214152SedDEFINE_COMPILERRT_FUNCTION(__floatundisf) 32214152Sed movl 8(%esp), %eax 33214152Sed movd 8(%esp), %xmm1 34214152Sed movd 4(%esp), %xmm0 35214152Sed punpckldq %xmm1, %xmm0 36214152Sed calll 0f 37214152Sed0: popl %ecx 38214152Sed sarl $31, %eax 39214152Sed movq %xmm0, 4(%esp) 40214152Sed fildll 4(%esp) 41214152Sed faddl TWOp64 42214152Sed fstps 4(%esp) 43214152Sed flds 4(%esp) 44214152Sed ret 45214152Sed 46214152Sed#endif // __i386__ 47214152Sed 48214152Sed*/ 49214152Sed 50214152Sed/* branch-free, x87-free implementation - faster at the expense of code size */ 51214152Sed 52214152Sed#ifdef __i386__ 53214152Sed 54214152Sed#ifndef __ELF__ 55214152Sed.const 56214152Sed.align 3 57214152Sed#else 58214152Sed.align 8 59214152Sed#endif 60214152Sedtwop52: .quad 0x4330000000000000 61214152Sed .quad 0x0000000000000fff 62214152Sedsticky: .quad 0x0000000000000000 63214152Sed .long 0x00000012 64214152Sedtwelve: .long 0x00000000 65214152Sed 66214152Sed#define TWOp52 twop52-0b(%ecx) 67214152Sed#define STICKY sticky-0b(%ecx,%eax,8) 68214152Sed 69214152Sed.text 70214152Sed.align 4 71214152SedDEFINE_COMPILERRT_FUNCTION(__floatundisf) 72214152Sed movl 8(%esp), %eax 73214152Sed movd 8(%esp), %xmm1 74214152Sed movd 4(%esp), %xmm0 75214152Sed punpckldq %xmm1, %xmm0 76214152Sed 77214152Sed calll 0f 78214152Sed0: popl %ecx 79214152Sed shrl %eax // high 31 bits of input as sint32 80214152Sed addl $0x7ff80000, %eax 81214152Sed sarl $31, %eax // (big input) ? -1 : 0 82214152Sed movsd STICKY, %xmm1 // (big input) ? 0xfff : 0 83214152Sed movl $12, %edx 84214152Sed andl %eax, %edx // (big input) ? 12 : 0 85214152Sed movd %edx, %xmm3 86214152Sed andpd %xmm0, %xmm1 // (big input) ? input & 0xfff : 0 87214152Sed movsd TWOp52, %xmm2 // 0x1.0p52 88214152Sed psrlq %xmm3, %xmm0 // (big input) ? input >> 12 : input 89214152Sed orpd %xmm2, %xmm1 // 0x1.0p52 + ((big input) ? input & 0xfff : input) 90214152Sed orpd %xmm1, %xmm0 // 0x1.0p52 + ((big input) ? (input >> 12 | input & 0xfff) : input) 91214152Sed subsd %xmm2, %xmm0 // (double)((big input) ? (input >> 12 | input & 0xfff) : input) 92214152Sed cvtsd2ss %xmm0, %xmm0 // (float)((big input) ? (input >> 12 | input & 0xfff) : input) 93214152Sed pslld $23, %xmm3 94214152Sed paddd %xmm3, %xmm0 // (float)input 95214152Sed movd %xmm0, 4(%esp) 96214152Sed flds 4(%esp) 97214152Sed ret 98214152Sed 99214152Sed#endif // __i386__ 100