Initial import

third_party/compiler_rt/nexgen32e/floatundixf.S

@@ -0,0 +1,75 @@
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+
+#include "third_party/compiler_rt/assembly.h"
+
+// long double __floatundixf(du_int a);
+
+#ifdef __x86_64__
+
+	.section .yoink
+	nop	huge_compiler_rt_license
+	.previous
+
+CONST_SECTION
+
+	.balign 16
+twop64:
+	.quad 0x43f0000000000000
+
+#define REL_ADDR(_a)	(_a)(%rip)
+
+	.text
+
+	.balign 4
+DEFINE_COMPILERRT_FUNCTION(__floatundixf)
+	movq	%rdi,	 -8(%rsp)
+	fildq	-8(%rsp)
+	test	%rdi,		%rdi
+	js		1f
+	ret
+1:	faddl	REL_ADDR(twop64)
+	ret
+END_COMPILERRT_FUNCTION(__floatundixf)
+
+#endif // __x86_64__
+
+
+/* Branch-free implementation is ever so slightly slower, but more beautiful.
+   It is likely superior for inlining, so I kept it around for future reference.
+
+#ifdef __x86_64__
+
+CONST_SECTION
+
+	.balign 4
+twop52:
+	.quad 0x4330000000000000
+twop84_plus_twop52_neg:
+	.quad 0xc530000000100000
+twop84:
+	.quad 0x4530000000000000
+
+#define REL_ADDR(_a)	(_a)(%rip)
+
+.text
+.balign 4
+DEFINE_COMPILERRT_FUNCTION(__floatundixf)
+	movl	%edi,				%esi			// low 32 bits of input
+	shrq	$32,				%rdi			// hi 32 bits of input
+	orq		REL_ADDR(twop84),	%rdi			// 2^84 + hi (as a double)
+	orq		REL_ADDR(twop52),	%rsi			// 2^52 + lo (as a double)
+	movq	%rdi,			 -8(%rsp)
+	movq	%rsi,			-16(%rsp)
+	fldl	REL_ADDR(twop84_plus_twop52_neg)	
+	faddl	-8(%rsp)	// hi - 2^52 (as double extended, no rounding occurs)
+	faddl	-16(%rsp)	// hi + lo (as double extended)
+	ret
+END_COMPILERRT_FUNCTION(__floatundixf)
+
+#endif // __x86_64__
+
+*/
+
+NO_EXEC_STACK_DIRECTIVE
+