From ac928830e6afec8c547a58ce39ac40472f87114f Mon Sep 17 00:00:00 2001
From: Ilya Kurdyukov <jpegqs@gmail.com>
Date: Tue, 8 Jun 2021 10:13:27 +0700
Subject: [PATCH] fftw-3.3.8 e2k support

---
 kernel/cycle.h           | 19 +++++++++++++++++++
 simd-support/avx.c       | 11 +++++++++++
 simd-support/avx2.c      | 10 ++++++++++
 simd-support/simd-avx.h  |  2 +-
 simd-support/simd-avx2.h |  2 +-
 simd-support/simd-sse2.h |  2 +-
 simd-support/sse2.c      |  2 +-
 7 files changed, 44 insertions(+), 4 deletions(-)

diff --git a/kernel/cycle.h b/kernel/cycle.h
index fe3dd50..b20fb28 100644
--- a/kernel/cycle.h
+++ b/kernel/cycle.h
@@ -260,6 +260,25 @@ INLINE_ELAPSED(__inline)
 #define TIME_MIN 5000.0
 #endif
 
+/*----------------------------------------------------------------*/
+/*
+ * E2K cycle counter
+ */
+
+#ifdef __e2k__
+#include <stdint.h>
+#include <x86intrin.h>
+typedef uint64_t ticks;
+#define getticks __rdtsc
+
+static __attribute__((__always_inline__)) inline double elapsed(ticks t1, ticks t0)
+{
+     return t1 - t0;
+}
+
+#define HAVE_TICK_COUNTER
+#endif
+
 /*----------------------------------------------------------------*/
 /*
  * IA64 cycle counter
diff --git a/simd-support/avx.c b/simd-support/avx.c
index 7e19be0..c29db77 100644
--- a/simd-support/avx.c
+++ b/simd-support/avx.c
@@ -23,6 +23,15 @@
 
 #if HAVE_AVX
 
+#ifdef __e2k__
+
+int X(have_simd_avx)(void)
+{
+     return 1;
+}
+
+#else
+
 #if defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64)
 #    include "amd64-cpuid.h"
 #else
@@ -52,3 +61,5 @@ int X(have_simd_avx)(void)
 
 #endif
 
+#endif
+
diff --git a/simd-support/avx2.c b/simd-support/avx2.c
index e240645..e9079c4 100644
--- a/simd-support/avx2.c
+++ b/simd-support/avx2.c
@@ -23,6 +23,15 @@
 
 #if HAVE_AVX2
 
+#ifdef __e2k__
+
+int X(have_simd_avx2_128)(void)
+{
+     return 1;
+}
+
+#else
+
 #if defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64)
 #    include "amd64-cpuid.h"
 #else
@@ -53,6 +62,7 @@ int X(have_simd_avx2_128)(void)
     }
     return res;
 }
+#endif
 
 int X(have_simd_avx2)(void)
 {
diff --git a/simd-support/simd-avx.h b/simd-support/simd-avx.h
index 7ca6363..9ec9bca 100644
--- a/simd-support/simd-avx.h
+++ b/simd-support/simd-avx.h
@@ -192,7 +192,7 @@ static inline __m128d VMOVAPD_LD(const R *x)
 
 	Force the use of vmovapd via asm until compilers stabilize.
      */
-#if defined(__GNUC__)
+#if defined(__GNUC__) && !defined(__e2k__)
      __m128d var;
      __asm__("vmovapd %1, %0\n" : "=x"(var) : "m"(x[0]));
      return var;
diff --git a/simd-support/simd-avx2.h b/simd-support/simd-avx2.h
index 5217026..7897153 100644
--- a/simd-support/simd-avx2.h
+++ b/simd-support/simd-avx2.h
@@ -196,7 +196,7 @@ static inline __m128d VMOVAPD_LD(const R *x)
 
 	Force the use of vmovapd via asm until compilers stabilize.
      */
-#if defined(__GNUC__)
+#if defined(__GNUC__) && !defined(__e2k__)
      __m128d var;
      __asm__("vmovapd %1, %0\n" : "=x"(var) : "m"(x[0]));
      return var;
diff --git a/simd-support/simd-sse2.h b/simd-support/simd-sse2.h
index 44b11b7..f7781db 100644
--- a/simd-support/simd-sse2.h
+++ b/simd-support/simd-sse2.h
@@ -140,7 +140,7 @@ static inline V LD(const R *x, INT ivs, const R *aligned_like)
 {
      V var;
      (void)aligned_like; /* UNUSED */
-#  ifdef __GNUC__
+#  if defined(__GNUC__) && !defined(__e2k__)
      /* We use inline asm because gcc-3.x generates slow code for
 	_mm_loadh_pi().  gcc-3.x insists upon having an existing variable for
 	VAL, which is however never used.  Thus, it generates code to move
diff --git a/simd-support/sse2.c b/simd-support/sse2.c
index c52c852..971f291 100644
--- a/simd-support/sse2.c
+++ b/simd-support/sse2.c
@@ -29,7 +29,7 @@
 
 #if HAVE_SSE2
 
-# if defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64)
+# if defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || defined(__e2k__)
 
   int X(have_simd_sse2)(void)
   {
-- 
2.17.1