[PATCH] math: move x86_64 fabs, fabsf to C with inline asm

[Alexander Monakov <amonakov@ispras.ru>]

[-- Attachment #1: Type: text/plain, Size: 636 bytes --]

---

Questions:

Why are there amd64-specific fabs implementations in the first place?
(Only) because GCC generated poor code for the generic C version?

Do annotations for mask manipulation in the patch help? Any way to make
them less ambiguous?


 src/math/x86_64/fabs.c  | 10 ++++++++++
 src/math/x86_64/fabs.s  |  9 ---------
 src/math/x86_64/fabsf.c | 10 ++++++++++
 src/math/x86_64/fabsf.s |  7 -------
 4 files changed, 20 insertions(+), 16 deletions(-)
 create mode 100644 src/math/x86_64/fabs.c
 delete mode 100644 src/math/x86_64/fabs.s
 create mode 100644 src/math/x86_64/fabsf.c
 delete mode 100644 src/math/x86_64/fabsf.s


[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: 0001-math-move-x86_64-fabs-fabsf-to-C-with-inline-asm.patch --]
[-- Type: text/x-patch; name="0001-math-move-x86_64-fabs-fabsf-to-C-with-inline-asm.patch", Size: 1420 bytes --]

diff --git a/src/math/x86_64/fabs.c b/src/math/x86_64/fabs.c
new file mode 100644
index 00000000..16562477
--- /dev/null
+++ b/src/math/x86_64/fabs.c
@@ -0,0 +1,10 @@
+#include <math.h>
+
+double fabs(double x)
+{
+	double t;
+	__asm__ ("pcmpeqd %0, %0" : "=x"(t));          // t = ~0
+	__asm__ ("psrlq   $1, %0" : "+x"(t));          // t >>= 1
+	__asm__ ("andps   %1, %0" : "+x"(x) : "x"(t)); // x &= t
+	return x;
+}
diff --git a/src/math/x86_64/fabs.s b/src/math/x86_64/fabs.s
deleted file mode 100644
index 5715005e..00000000
--- a/src/math/x86_64/fabs.s
+++ /dev/null
@@ -1,9 +0,0 @@
-.global fabs
-.type fabs,@function
-fabs:
-	xor %eax,%eax
-	dec %rax
-	shr %rax
-	movq %rax,%xmm1
-	andpd %xmm1,%xmm0
-	ret
diff --git a/src/math/x86_64/fabsf.c b/src/math/x86_64/fabsf.c
new file mode 100644
index 00000000..36ea7481
--- /dev/null
+++ b/src/math/x86_64/fabsf.c
@@ -0,0 +1,10 @@
+#include <math.h>
+
+float fabsf(float x)
+{
+	float t;
+	__asm__ ("pcmpeqd %0, %0" : "=x"(t));          // t = ~0
+	__asm__ ("psrld   $1, %0" : "+x"(t));          // t >>= 1
+	__asm__ ("andps   %1, %0" : "+x"(x) : "x"(t)); // x &= t
+	return x;
+}
diff --git a/src/math/x86_64/fabsf.s b/src/math/x86_64/fabsf.s
deleted file mode 100644
index 501a1f17..00000000
--- a/src/math/x86_64/fabsf.s
+++ /dev/null
@@ -1,7 +0,0 @@
-.global fabsf
-.type fabsf,@function
-fabsf:
-	mov $0x7fffffff,%eax
-	movq %rax,%xmm1
-	andps %xmm1,%xmm0
-	ret