From b9d2bb67d65619f1815fd56aa49563800d447165 Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Sat, 3 May 2014 11:34:07 -0700
Subject: [PATCH] dsp/neon.h: coalesce intrinsics-related defines

Change-Id: Ifadd41a5bbf7f99eeb6d75d2b67daa25e0544946
---
 src/dsp/dec_neon.c      | 12 +-----------
 src/dsp/enc_neon.c      | 14 +++-----------
 src/dsp/lossless_neon.c |  5 +----
 src/dsp/neon.h          | 11 +++++++++++
 4 files changed, 16 insertions(+), 26 deletions(-)

diff --git a/src/dsp/dec_neon.c b/src/dsp/dec_neon.c
index 7216ea12..7b189ed8 100644
--- a/src/dsp/dec_neon.c
+++ b/src/dsp/dec_neon.c
@@ -16,19 +16,9 @@
 
 #if defined(WEBP_USE_NEON)
 
-// #define USE_INTRINSICS   // use intrinsics when possible
-
 #include "./neon.h"
-
 #include "../dec/vp8i.h"
 
-// if using intrinsics, this flag avoids some functions that make gcc-4.6.3
-// crash ("internal compiler error: in immed_double_const, at emit-rtl.").
-// (probably similar to gcc.gnu.org/bugzilla/show_bug.cgi?id=48183)
-#if !LOCAL_GCC_PREREQ(4,8)
-#define WORK_AROUND_GCC
-#endif
-
 //------------------------------------------------------------------------------
 // NxM Loading functions
 
@@ -84,7 +74,7 @@ static WEBP_INLINE void Load4x16(const uint8_t* const src, int stride,
   *q1 = vcombine_u8(row0.val[3], row8.val[3]);
 }
 
-#else
+#else  // WORK_AROUND_GCC
 
 #define LOADQ_LANE_32b(VALUE, LANE) do {                             \
   (VALUE) = vld1q_lane_u32((const uint32_t*)src, (VALUE), (LANE));   \
diff --git a/src/dsp/enc_neon.c b/src/dsp/enc_neon.c
index 7e74e84e..210f10a8 100644
--- a/src/dsp/enc_neon.c
+++ b/src/dsp/enc_neon.c
@@ -15,8 +15,6 @@
 
 #if defined(WEBP_USE_NEON)
 
-// #define USE_INTRINSICS   // use intrinsics when possible
-
 #include <assert.h>
 
 #include "./neon.h"
@@ -969,13 +967,7 @@ static int SSE4x4(const uint8_t* a, const uint8_t* b) {
 
 //------------------------------------------------------------------------------
 
-// Compilation with gcc-4.6.x is problematic for now. Disable this function
-// in this case.
-#if !LOCAL_GCC_PREREQ(4,8)
-#define SKIP_QUANTIZE
-#endif
-
-#if !defined(SKIP_QUANTIZE)
+#if !defined(WORK_AROUND_GCC)
 
 static int16x8_t Quantize(int16_t* const in,
                           const VP8Matrix* const mtx, int offset) {
@@ -1033,7 +1025,7 @@ static int QuantizeBlock(int16_t in[16], int16_t out[16],
   return 0;
 }
 
-#endif   // SKIP_QUANTIZE
+#endif   // !WORK_AROUND_GCC
 
 #endif   // WEBP_USE_NEON
 
@@ -1056,7 +1048,7 @@ void VP8EncDspInitNEON(void) {
   VP8SSE16x8 = SSE16x8;
   VP8SSE8x8 = SSE8x8;
   VP8SSE4x4 = SSE4x4;
-#if !defined(SKIP_QUANTIZE)
+#if !defined(WORK_AROUND_GCC)
   VP8EncQuantizeBlock = QuantizeBlock;
 #endif
 #endif   // WEBP_USE_NEON
diff --git a/src/dsp/lossless_neon.c b/src/dsp/lossless_neon.c
index d0ae21c9..5597d9d2 100644
--- a/src/dsp/lossless_neon.c
+++ b/src/dsp/lossless_neon.c
@@ -18,10 +18,7 @@
 #include <arm_neon.h>
 
 #include "./lossless.h"
-
-// Right now, some intrinsics function seem slower, so we disable it.
-// Uncomment the following for testing:
-// #define USE_INTRINSICS
+#include "./neon.h"
 
 //------------------------------------------------------------------------------
 // Colorspace conversion functions
diff --git a/src/dsp/neon.h b/src/dsp/neon.h
index 96c089ec..2b42f9e5 100644
--- a/src/dsp/neon.h
+++ b/src/dsp/neon.h
@@ -14,6 +14,17 @@
 
 #include <arm_neon.h>
 
+#include "./dsp.h"
+
+// #define USE_INTRINSICS   // use intrinsics when possible
+
+// if using intrinsics, this flag avoids some functions that make gcc-4.6.3
+// crash ("internal compiler error: in immed_double_const, at emit-rtl.").
+// (probably similar to gcc.gnu.org/bugzilla/show_bug.cgi?id=48183)
+#if !LOCAL_GCC_PREREQ(4,8)
+#define WORK_AROUND_GCC
+#endif
+
 static WEBP_INLINE int32x4x4_t Transpose4x4(const int32x4x4_t rows) {
   uint64x2x2_t row01, row23;