fix-arm64

Gbp-Pq: Name fix-arm64.patch

diff --git a/client/vector/analyzeFuncs_vector.cpp b/client/vector/analyzeFuncs_vector.cpp
index d80e3045a03314a1d6951fd8f8318c0683fa0ba7..db74892aaf45d7c4d38cbccc297679253530be16 100644 (file)
--- a/client/vector/analyzeFuncs_vector.cpp
+++ b/client/vector/analyzeFuncs_vector.cpp
@@ -313,9 +313,7 @@ void SetCapabilities(void) {
 #ifdef USE_NEON
   if (strstr(app_init_data.host_info.p_features," neon ")) CPUCaps |= BA_NEON;
 #endif
-#if  defined(__aarch64__) 
-  CPUCaps |= BA_VFPV4 | BA_VFPV4D16 | BA_VFPV3 | BA_VFPV3D16 | BA_VFP | BA_NEON;
-#endif
+
 #endif // ANDROID
 #endif // __arm__
 #ifdef USE_MANUAL_CALLSTACK
@@ -353,7 +351,7 @@ GPStb GetPowerSpectrumFuncs[]={
      v_avxGetPowerSpectrum, BA_AVX, "v_avxGetPowerSpectrum", 
 #  endif
 #endif
-#if (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)) || defined(__aarch64__)
+#if defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)
      vfp_GetPowerSpectrum, BA_VFP, "vfp_GetPowerSpectrum",
 #ifdef USE_NEON
      neon_GetPowerSpectrum, BA_NEON, "neon_GetPowerSpectrum",
@@ -409,7 +407,7 @@ CDtb ChirpDataFuncs[]={
      avx_ChirpData_d, BA_AVX, "avx_ChirpData_d", 
 #  endif
 #endif
-#if (defined(__arm__) && defined(ANDROID) && defined(__VFP_FP__) && !defined(__SOFTFP__)) || defined(__aarch64__)
+#if defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)
      vfp_ChirpData, BA_VFP, "vfp_ChirpData",
 #ifdef USE_NEON
      neon_ChirpData, BA_NEON, "neon_ChirpData",
@@ -459,7 +457,7 @@ TPtb TransposeFuncs[]={
      v_avxTranspose8x8ntw_b, BA_AVX, "v_avxTranspose8x8ntw_b", 
 #  endif
 #endif
-#if (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)) || defined(__arch64__)
+#if defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)
     v_pfTranspose2, BA_VFP, "v_pfTranspose2",      
     v_pfTranspose4, BA_VFP, "v_pfTranspose4",      
     v_pfTranspose8, BA_VFP, "v_pfTranspose8",      
@@ -489,7 +487,7 @@ FolSub FoldSubs[] = {
   &AVXfold_a,    BA_AVX,
   &AVXfold_c,    BA_AVX,
 #endif
-#if (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)) || defined(__aarch64__)
+#if defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)
   &vfpFoldMain,  BA_VFP,
 #ifdef USE_NEON
   &neonFoldMain, BA_NEON,
@@ -531,7 +529,7 @@ bool TestBoincSignalHandling() {
         __asm__ ("movl %cr4,%eax");
 #elif defined(__ppc__) || defined(__sparc__) 
         __asm__ (".long %0": : "i" (0xfeedface) );
-#elif defined(__arm__) || defined (__aarch64__)
+#elif defined(__arm__)
         __asm__ (".word 0xf7f000f0" );
 #endif
 #elif defined(_MSC_VER) && ( defined(_M_IX86) || defined(_M_X64) )

diff --git a/client/vector/analyzeFuncs_vector.h b/client/vector/analyzeFuncs_vector.h
index 49d1dd71ff52998815b9ebb3dc899fade1eed020..c068d2ab0e0926a8914e1bbb0768447793f0c827 100644 (file)
--- a/client/vector/analyzeFuncs_vector.h
+++ b/client/vector/analyzeFuncs_vector.h
@@ -236,7 +236,7 @@ extern int sse3_ChirpData_ak8(
 
 extern int v_vTranspose(int i, int j, float *in, float *out);
 #if defined(__i386__) || defined(__x86_64__) || defined(USE_SSE) || \
-    (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)) || defined(__aarch64__)
+    (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__))
 extern int v_pfTranspose2(int i, int j, float *in, float *out);
 extern int v_pfTranspose4(int i, int j, float *in, float *out);
 extern int v_pfTranspose8(int i, int j, float *in, float *out);
@@ -249,7 +249,7 @@ extern int v_vTranspose4x8ntw(int i, int j, float *in, float *out);
 extern int v_vTranspose4x16ntw(int i, int j, float *in, float *out);
 extern int v_vpfTranspose8x4ntw(int i, int j, float *in, float *out);
 #endif
-#if (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)) || defined(__aarch64__)
+#if (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__))
 extern int v_vfpTranspose2(int i, int j, float *in, float *out);
 #endif
 #ifdef USE_FFTWF
@@ -361,7 +361,7 @@ extern int v_avxGetPowerSpectrum(
 );
 #endif
 
-#if (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)) || defined(__aarch64__)
+#if defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)
 #ifdef USE_NEON
 extern int neon_GetPowerSpectrum(
     sah_complex * cx_FreqData,

diff --git a/client/vector/analyzeFuncs_vfp_aux.cpp b/client/vector/analyzeFuncs_vfp_aux.cpp
index a0fbb53b3235b59c89181027dafd7ed9a90ba884..88bc9eee00c8ea6852a7174567a52f4901b4ee5c 100644 (file)
--- a/client/vector/analyzeFuncs_vfp_aux.cpp
+++ b/client/vector/analyzeFuncs_vfp_aux.cpp
@@ -33,7 +33,7 @@
 #include <vector>
 #include <cmath>
 
-#if (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)) || defined(__aarch64__)
+#if (defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__))
 
 
 #define INVALID_CHIRP 2e+20
@@ -199,7 +199,7 @@ static inline int vfp_subTranspose2(int x, int y, float *in, float *out) {
 #ifdef USE_MANUAL_CALLSTACK
     call_stack.enter("vfp_subTranspose2()");
 #endif
-#ifndef __aarch64__
+#if defined(__arm__) && defined(__VFP_FP__) && !defined(__SOFTFP__)
     __asm__ __volatile__ (
         "fldmias  %2, {s4, s5}\n"
         "fldmias  %3, {s6, s7}\n"
@@ -211,8 +211,18 @@ static inline int vfp_subTranspose2(int x, int y, float *in, float *out) {
     :  
     : "r" (out), "r" (out+y), "r" (in), "r" (in+x)
     : "s3", "s4", "s5", "s6", "s7" );
-#else
-    v_pfsubTranspose<2>(in,out,x,y);
+#elif defined(__aarch64__)
+    __asm__ __volatile__ (
+        "ldp x4, x5, [%2], #16\n"
+        "ldp x6, x7, [%3], #16\n"
+        "mov x3,x6\n"
+        "mov x6,x5\n"
+        "mov x5,x3\n"
+        "stp x4, x5, [%0], #16\n"
+        "stp x6, x7, [%1], #16\n"
+        :
+        : "x" (out), "x" (out+y), "x" (in), "x" (in+x)
+        : "x3", "x4", "x5", "x6", "x7" );
 #endif
 #ifdef USE_MANUAL_CALLSTACK
     call_stack.exit();

diff --git a/client/vector/fp_arm.h b/client/vector/fp_arm.h
index e85ea35ac202da0ef36848e47f3fe3af46846aa6..9a7ba952945b343b1056e15c6e98cba6ad57df7e 100644 (file)
--- a/client/vector/fp_arm.h
+++ b/client/vector/fp_arm.h
@@ -36,7 +36,7 @@
 // work so modified without permission of the authors of those packages.
 //
 
-#if (defined(__arm__) || defined(__aarch64__))
+#if defined(__arm__) 
 #include <signal.h>
 #include <setjmp.h>
 #include "s_util.h"
@@ -125,11 +125,7 @@ inline static unsigned int restorefp() {
 
 inline void pld(void *arg1,const int arg2=0) {
     __asm__ __volatile__ (
-#ifndef __aarch64__
       "pld [%0,%1]\n"
-#else
-      "prfm PLDL1STRM,[%0,%1]\n"
-#endif
       : 
       : "r" (arg1), "Jr" (arg2)
     );