From 1fb01691547e85be253a4f1543be3920de651e27 Mon Sep 17 00:00:00 2001
From: nsz <nsz@port70.net>
Date: Wed, 8 Aug 2012 20:18:16 +0200
Subject: [PATCH] math: fix exp.s on i386 and x86_64 so the exception flags are
 correct

exp(inf), exp(-inf), exp(nan) used to raise wrong flags
---
 src/math/i386/exp.s    | 41 +++++++++++++++++++----------------------
 src/math/x86_64/expl.s | 38 ++++++++++++++++++--------------------
 2 files changed, 37 insertions(+), 42 deletions(-)

diff --git a/src/math/i386/exp.s b/src/math/i386/exp.s
index ca0de1d4..c7f5ad0f 100644
--- a/src/math/i386/exp.s
+++ b/src/math/i386/exp.s
@@ -69,18 +69,18 @@ exp:
 exp2:
 	fldl 4(%esp)
 1:	pushl $0x467ff000
-	flds (%esp)
+	flds (%esp)       # 16380
 	xorl %eax,%eax
 	pushl $0x80000000
 	push %eax
 	fld %st(1)
 	fabs
-	fucom %st(1)
+	fucomp %st(1)
 	fnstsw
 	fstp %st(0)
-	fstp %st(0)
 	sahf
-	ja 2f
+	ja 3f             # |x| > 16380
+	jp 2f             # x is nan (avoid invalid except in fistp)
 	fld %st(0)
 	fistpl 8(%esp)
 	fildl 8(%esp)
@@ -90,30 +90,27 @@ exp2:
 	add %eax,8(%esp)
 	f2xm1
 	fld1
-	faddp
-	fldt (%esp)
+	faddp             # 2^(x-rint(x))
+	fldt (%esp)       # 2^rint(x)
 	fmulp
 	fstp %st(1)
-	add $12,%esp
+2:	add $12,%esp
 	ret
 
-2:	fld %st(0)
+3:	fld %st(0)
 	fstpt (%esp)
-	mov 9(%esp),%ah
-	and $0x7f,%ah
-	cmp $0x7f,%ah
-	jne 1f
-	decb 9(%esp)
-	fstp %st(0)
-	fldt (%esp)
-1:	fld %st(0)
-	frndint
-	fxch %st(1)
-	fsub %st(1)
-	f2xm1
 	fld1
-	faddp
-	fscale
+	mov 8(%esp),%ax
+	and $0x7fff,%ax
+	cmp $0x7fff,%ax
+	je 1f             # x = +-inf
+	fld %st(1)
+	frndint
+	fxch %st(2)
+	fsub %st(2)       # st(0)=x-rint(x), st(1)=1, st(2)=rint(x)
+	f2xm1
+	faddp             # 2^(x-rint(x))
+1:	fscale
 	fstp %st(1)
 	add $12,%esp
 	ret
diff --git a/src/math/x86_64/expl.s b/src/math/x86_64/expl.s
index 64c1c78e..740bc779 100644
--- a/src/math/x86_64/expl.s
+++ b/src/math/x86_64/expl.s
@@ -40,7 +40,7 @@ exp2l:
 	mov %eax,-20(%rsp)
 	xor %eax,%eax
 	mov %eax,-24(%rsp)
-	flds -16(%rsp)
+	flds -16(%rsp)    # 16380
 	fld %st(1)
 	fabs
 	fucom %st(1)
@@ -48,7 +48,8 @@ exp2l:
 	fstp %st(0)
 	fstp %st(0)
 	sahf
-	ja 2f
+	ja 3f             # |x| > 16380
+	jp 2f             # x is nan (avoid invalid except in fistp)
 	fld %st(0)
 	fistpl -16(%rsp)
 	fildl -16(%rsp)
@@ -58,28 +59,25 @@ exp2l:
 	add %eax,-16(%rsp)
 	f2xm1
 	fld1
-	faddp
-	fldt -24(%rsp)
+	faddp             # 2^(x-rint(x))
+	fldt -24(%rsp)    # 2^rint(x)
 	fmulp
-	fstp %st(1)
+2:	fstp %st(1)
 	ret
 
-2:	fld %st(0)
+3:	fld %st(0)
 	fstpt -24(%rsp)
-	mov -15(%rsp),%ah
-	and $0x7f,%ah
-	cmp $0x7f,%ah
-	jne 1f
-	decb -15(%rsp)
-	fstp %st(0)
-	fldt -24(%rsp)
-1:	fld %st(0)
-	frndint
-	fxch %st(1)
-	fsub %st(1)
-	f2xm1
 	fld1
-	faddp
-	fscale
+	mov -15(%rsp),%ax
+	and $0x7fff,%ax
+	cmp $0x7fff,%ax
+	je 1f             # x = +-inf
+	fld %st(1)
+	frndint
+	fxch %st(2)
+	fsub %st(2)       # st(0)=x-rint(x), st(1)=1, st(2)=rint(x)
+	f2xm1
+	faddp             # 2^(x-rint(x))
+1:	fscale
 	fstp %st(1)
 	ret