110 lines · plain
1// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.2// See https://llvm.org/LICENSE.txt for license information.3// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception4 5#include "../assembly.h"6 7// float __floatundisf(du_int a);8 9// Note that there is a hardware instruction, fildll, that does most of what10// this function needs to do. However, because of our ia32 ABI, it will take11// a write-small read-large stall, so the software implementation here is12// actually several cycles faster.13 14// This is a branch-free implementation. A branchy implementation might be15// faster for the common case if you know something a priori about the input16// distribution.17 18/* branch-free x87 implementation - one cycle slower than without x87.19 20#ifdef __i386__21 22CONST_SECTION23.balign 324 25 .quad 0x43f000000000000026twop64: .quad 0x000000000000000027 28#define TWOp64 twop64-0b(%ecx,%eax,8)29 30.text31.balign 432DEFINE_COMPILERRT_FUNCTION(__floatundisf)33 movl 8(%esp), %eax34 movd 8(%esp), %xmm135 movd 4(%esp), %xmm036 punpckldq %xmm1, %xmm037 calll 0f380: popl %ecx39 sarl $31, %eax40 movq %xmm0, 4(%esp)41 fildll 4(%esp)42 faddl TWOp6443 fstps 4(%esp)44 flds 4(%esp)45 ret46END_COMPILERRT_FUNCTION(__floatundisf)47 48#endif // __i386__49 50*/51 52// branch-free, x87-free implementation - faster at the expense of code size53 54#ifdef __i386__55 56CONST_SECTION57 58 .balign 1659twop52:60 .quad 0x433000000000000061 .quad 0x0000000000000fff62 63 .balign 1664sticky:65 .quad 0x000000000000000066 .long 0x0000001267 68 .balign 1669twelve:70 .long 0x0000000071 72#define TWOp52 twop52-0b(%ecx)73#define STICKY sticky-0b(%ecx,%eax,8)74 75.text76.balign 477DEFINE_COMPILERRT_FUNCTION(__floatundisf)78 movl 8(%esp), %eax79 movd 8(%esp), %xmm180 movd 4(%esp), %xmm081 punpckldq %xmm1, %xmm082 83 calll 0f840: popl %ecx85 shrl %eax // high 31 bits of input as sint3286 addl $0x7ff80000, %eax87 sarl $31, %eax // (big input) ? -1 : 088 movsd STICKY, %xmm1 // (big input) ? 0xfff : 089 movl $12, %edx90 andl %eax, %edx // (big input) ? 12 : 091 movd %edx, %xmm392 andpd %xmm0, %xmm1 // (big input) ? input & 0xfff : 093 movsd TWOp52, %xmm2 // 0x1.0p5294 psrlq %xmm3, %xmm0 // (big input) ? input >> 12 : input95 orpd %xmm2, %xmm1 // 0x1.0p52 + ((big input) ? input & 0xfff : input)96 orpd %xmm1, %xmm0 // 0x1.0p52 + ((big input) ? (input >> 12 | input & 0xfff) : input)97 subsd %xmm2, %xmm0 // (double)((big input) ? (input >> 12 | input & 0xfff) : input)98 cvtsd2ss %xmm0, %xmm0 // (float)((big input) ? (input >> 12 | input & 0xfff) : input)99 pslld $23, %xmm3100 paddd %xmm3, %xmm0 // (float)input101 movd %xmm0, 4(%esp)102 flds 4(%esp)103 ret104END_COMPILERRT_FUNCTION(__floatundisf)105 106#endif // __i386__107 108NO_EXEC_STACK_DIRECTIVE109 110