[SVE] Workaround for ICE with GCC-10 & SVE (#391)

With this patch, `-fno-tree-vrp` compiler option is added if the SVE code is compiled with GCC.
shibatch · Jan 12, 2021 · ea29e62 · ea29e62
1 parent e7b4784
commit ea29e62
Show file tree

Hide file tree

Showing 14 changed files with 65 additions and 42 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -75,14 +75,14 @@ command_arguments(MKMASKED_PARAMS_GNUABI_SVE_sp sve s -4)
 
 #
 
-set(COSTOVERRIDE_AVX512F 10.0)
-set(COSTOVERRIDE_AVX512FNOFMA 10.0)
-set(COSTOVERRIDE_AVX2 2.0)
-set(COSTOVERRIDE_AVX 2.0)
-set(COSTOVERRIDE_NEON32 2.0)
-set(COSTOVERRIDE_NEON32VFPV4 2.0)
-set(COSTOVERRIDE_SVE 10.0)
-set(COSTOVERRIDE_SVENOFMA 10.0)
+set(COSTOVERRIDE_AVX512F 10)
+set(COSTOVERRIDE_AVX512FNOFMA 10)
+set(COSTOVERRIDE_AVX2 2)
+set(COSTOVERRIDE_AVX 2)
+set(COSTOVERRIDE_NEON32 2)
+set(COSTOVERRIDE_NEON32VFPV4 2)
+set(COSTOVERRIDE_SVE 10)
+set(COSTOVERRIDE_SVENOFMA 10)
 
 #
 

diff --git a/Configure.cmake b/Configure.cmake
@@ -283,6 +283,10 @@ else()
   set(DFT_C_FLAGS "${FLAGS_WALL} ${FLAGS_NOSTRICTALIASING} ${FLAGS_FASTMATH} ${FLAGS_OTHERS}")
 endif()
 
+if(CMAKE_C_COMPILER_ID MATCHES "GNU")
+  set(FLAGS_ENABLE_SVE "${FLAGS_ENABLE_SVE};-fno-tree-vrp")
+endif()
+
 if (CMAKE_SYSTEM_PROCESSOR MATCHES "^i.86$" AND CMAKE_C_COMPILER_ID MATCHES "GNU")
   set(SLEEF_C_FLAGS "${SLEEF_C_FLAGS} -msse2 -mfpmath=sse")
   set(DFT_C_FLAGS "${DFT_C_FLAGS} -msse2 -mfpmath=sse -m128bit-long-double")

diff --git a/Jenkinsfile b/Jenkinsfile
@@ -53,7 +53,7 @@ pipeline {
 			 rm -rf build
  			 mkdir build
 			 cd build
-			 cmake -GNinja -DCMAKE_INSTALL_PREFIX=../install -DSLEEF_SHOW_CONFIG=1 -DBUILD_SHARED_LIBS=FALSE -DENFORCE_TESTER3=TRUE -DFORCE_AAVPCS=On -DENABLE_GNUABI=On -DBUILD_QUAD=TRUE -DBUILD_DFT=TRUE -DDISABLE_SVE=TRUE -DBUILD_INLINE_HEADERS=TRUE -DENABLE_CUDA=TRUE -DENFORCE_CUDA=TRUE ..
+			 cmake -GNinja -DCMAKE_INSTALL_PREFIX=../install -DSLEEF_SHOW_CONFIG=1 -DBUILD_SHARED_LIBS=FALSE -DENFORCE_TESTER3=TRUE -DFORCE_AAVPCS=On -DENABLE_GNUABI=On -DBUILD_QUAD=TRUE -DBUILD_DFT=TRUE -DBUILD_INLINE_HEADERS=TRUE -DENABLE_CUDA=TRUE -DENFORCE_CUDA=TRUE -DENFORCE_SVE=TRUE -DEMULATOR=qemu-aarch64 ..
 			 ninja
 			 export OMP_WAIT_POLICY=passive
 		         export CTEST_OUTPUT_ON_FAILURE=TRUE

diff --git a/src/arch/helperpurec_scalar.h b/src/arch/helperpurec_scalar.h
@@ -56,13 +56,17 @@
 
 #if defined(__AVX2__) || defined(__aarch64__) || defined(__arm__) || defined(__powerpc64__) || defined(__zarch__) || CONFIG == 3
 #ifndef FP_FAST_FMA
+//@#ifndef FP_FAST_FMA
 #define FP_FAST_FMA
 //@#define FP_FAST_FMA
 #endif
+//@#endif
 #ifndef FP_FAST_FMAF
+//@#ifndef FP_FAST_FMAF
 #define FP_FAST_FMAF
 //@#define FP_FAST_FMAF
 #endif
+//@#endif
 #endif
 
 #if (!defined(FP_FAST_FMA) || !defined(FP_FAST_FMAF)) && !defined(SLEEF_GENHEADER)
@@ -119,6 +123,8 @@ typedef struct {
 
 #if defined(ENABLEFLOAT128) && CONFIG != 3
 typedef __float128 vargquad;
+#elif defined(__SIZEOF_LONG_DOUBLE__) && defined(__aarch64__) && CONFIG != 3
+typedef long double vargquad;
 #else
 typedef vquad vargquad;
 #endif

diff --git a/src/common/misc.h b/src/common/misc.h
@@ -179,6 +179,10 @@ typedef struct {
 #define Sleef_quad_DEFINED
 #if defined(ENABLEFLOAT128)
 typedef __float128 Sleef_quad;
+#define SLEEF_QUAD_C(x) (x ## Q)
+#elif defined(__SIZEOF_LONG_DOUBLE__) && defined(__aarch64__)
+typedef long double Sleef_quad;
+#define SLEEF_QUAD_C(x) (x ## L)
 #else
 typedef struct { uint64_t x, y; } Sleef_quad;
 #endif

diff --git a/src/libm/CMakeLists.txt b/src/libm/CMakeLists.txt
@@ -549,7 +549,7 @@ if(BUILD_INLINE_HEADERS)
             > ${PROJECT_BINARY_DIR}/include/sleefinline_${SIMDLC}.h                                        # > /build/include/sleefinline_sse2.h
 
 	  MAIN_DEPENDENCY ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimddp.c ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimdsp.c ${HEADER_${SIMD}}
-	  DEPENDS ${HEADER_${SIMD}} addSuffix
+	  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimddp.c ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimdsp.c ${HEADER_${SIMD}} addSuffix
 	  VERBATIM
 	  )
 
@@ -630,7 +630,7 @@ if(BUILD_INLINE_HEADERS)
       > ${PROJECT_BINARY_DIR}/include/sleefinline_${SIMDLC}.h                                          # > /build/include/sleefinline_cuda.h
 
       MAIN_DEPENDENCY ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimddp.c ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimdsp.c ${HEADER_${SIMD}}
-      DEPENDS ${HEADER_${SIMD}} addSuffix
+      DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimddp.c ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimdsp.c ${HEADER_${SIMD}} addSuffix
       VERBATIM
       )
 

diff --git a/src/libm/sleeflibm_header.h.org.in b/src/libm/sleeflibm_header.h.org.in
@@ -132,9 +132,9 @@ typedef struct {
 #if defined(__SIZEOF_FLOAT128__) || (defined(__linux__) && defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))) || (defined(__PPC64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 8)
 typedef __float128 Sleef_quad;
 #define SLEEF_QUAD_C(x) (x ## Q)
-//#elif defined(__SIZEOF_LONG_DOUBLE__) && defined(__aarch64__)
-//typedef long double Sleef_quad;
-//#define SLEEF_QUAD_C(x) (x ## L)
+#elif defined(__SIZEOF_LONG_DOUBLE__) && defined(__aarch64__)
+typedef long double Sleef_quad;
+#define SLEEF_QUAD_C(x) (x ## L)
 #else
 typedef struct { uint64_t x, y; } Sleef_quad;
 #endif

diff --git a/src/quad-tester/CMakeLists.txt b/src/quad-tester/CMakeLists.txt
@@ -71,7 +71,7 @@ function(add_test_iut IUT C)
       set_tests_properties(${IUT} PROPERTIES COST ${C})
     else()
       add_test(NAME ${IUT}
-	COMMAND ${QTESTER} ${EMULATOR} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${IUT}
+	COMMAND ${QTESTER} "--qemu" ${EMULATOR} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${IUT}
 	WORKING_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
       set_tests_properties(${IUT} PROPERTIES COST ${C})
     endif()
@@ -97,7 +97,8 @@ macro(test_extension SIMD)
     add_dependencies(${TARGET_IUT${SIMD}} sleefquad ${TARGET_LIBSLEEF})
     set_target_properties(${TARGET_IUT${SIMD}} PROPERTIES C_STANDARD 99)
     if (DEFINED COSTOVERRIDE_${SIMD})
-      add_test_iut(${TARGET_IUT${SIMD}} ${COSTOVERRIDE_${SIMD}})
+      math(EXPR C "${COSTOVERRIDE_${SIMD}} + 1")
+      add_test_iut(${TARGET_IUT${SIMD}} ${C})
     else()
       add_test_iut(${TARGET_IUT${SIMD}} 0.5)
     endif()
@@ -125,7 +126,8 @@ macro(test_extension SIMD)
 	add_dependencies(${IUTINAME} ${TARGET_QINLINE_HEADERS})
 	set_target_properties(${IUTINAME} PROPERTIES C_STANDARD 99)
 	if (DEFINED COSTOVERRIDE_${SIMD})
-	  add_test_iut(${IUTINAME} ${COSTOVERRIDE_${SIMD}})
+	  math(EXPR C "${COSTOVERRIDE_${SIMD}} + 1")
+	  add_test_iut(${IUTINAME} ${C})
 	else()
 	  add_test_iut(${IUTINAME} 0.5)
 	endif()

diff --git a/src/quad-tester/qtester.c b/src/quad-tester/qtester.c
@@ -1499,7 +1499,7 @@ void do_test(int options) {
 }
 
 int main(int argc, char **argv) {
-  char *argv2[argc+2], *commandSde = NULL;
+  char *argv2[argc+2], *commandSde = NULL, *commandQEmu = NULL;
   int i, a2s, options;
 
   // BUGFIX: this flush is to prevent incorrect syncing with the
@@ -1511,6 +1511,9 @@ int main(int argc, char **argv) {
     if (a2s+1 < argc && strcmp(argv[a2s], "--sde") == 0) {
       commandSde = argv[a2s+1];
       a2s++;
+    } else if (a2s+1 < argc && strcmp(argv[a2s], "--qemu") == 0) {
+      commandQEmu = argv[a2s+1];
+      a2s++;
     } else {
       break;
     }
@@ -1532,14 +1535,20 @@ int main(int argc, char **argv) {
     if (readln(ctop[0], str, 255) < 1 ||
 	sscanf(str, "%d", &options) != 1 ||
 	(options & 1) == 0) {
-      if (commandSde != NULL) {
+      if (commandSde != NULL || commandQEmu != NULL) {
 	close(ctop[0]);
 	close(ptoc[1]);
 
-	argv2[0] = commandSde;
-	argv2[1] = "--";
-	for(i=a2s;i<argc;i++) argv2[i-a2s+2] = argv[i];
-	argv2[argc-a2s+2] = NULL;
+	if (commandSde) {
+	  argv2[0] = commandSde;
+	  argv2[1] = "--";
+	  for(i=a2s;i<argc;i++) argv2[i-a2s+2] = argv[i];
+	  argv2[argc-a2s+2] = NULL;
+	} else {
+	  argv2[0] = commandQEmu;
+	  for(i=a2s;i<argc;i++) argv2[i-a2s+1] = argv[i];
+	  argv2[argc-a2s+1] = NULL;
+	}
 
 	startChild(argv2[0], argv2);
 
@@ -1550,7 +1559,7 @@ int main(int argc, char **argv) {
 	  return 0;
 	}
 
-	fprintf(stderr, "*** Using SDE\n");
+	fprintf(stderr, "*** Using emulator\n");
       } else {
 	int status;
 	waitpid(pid, &status, 0);
@@ -1565,8 +1574,6 @@ int main(int argc, char **argv) {
     }
   }
 
-  fprintf(stderr, "\n\n*** qtester : now testing %s\n", argv2[0]);
-
   fpctop = fdopen(ctop[0], "r");
 
   do_test(options);

diff --git a/src/quad/CMakeLists.txt b/src/quad/CMakeLists.txt
@@ -235,7 +235,7 @@ if(BUILD_INLINE_HEADERS)
             > ${PROJECT_BINARY_DIR}/include/sleefquadinline_${SIMDLC}.h                                     # > /build/include/sleefquadinline_sse2.h
 
 	  MAIN_DEPENDENCY ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimdqp.c ${HEADER_${SIMD}}
-	  DEPENDS ${HEADER_${SIMD}} addSuffix
+	  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimdqp.c ${HEADER_${SIMD}} addSuffix
 	  VERBATIM
 	  )
 
@@ -307,7 +307,7 @@ if(BUILD_INLINE_HEADERS)
       > ${PROJECT_BINARY_DIR}/include/sleefquadinline_${SIMDLC}.h                                       # > /build/include/sleefquadinline_cuda.h
 
       MAIN_DEPENDENCY ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimdqp.c ${HEADER_${SIMD}}
-      DEPENDS ${HEADER_${SIMD}} addSuffix
+      DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/sleefsimdqp.c ${HEADER_${SIMD}} addSuffix
       VERBATIM
       )
 

diff --git a/src/quad/sleefquadinline_cuda_header.h.org b/src/quad/sleefquadinline_cuda_header.h.org
@@ -12,9 +12,9 @@
 #if defined(__SIZEOF_FLOAT128__) || (defined(__linux__) && defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))) || (defined(__PPC64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 8)
 typedef __float128 Sleef_quad;
 #define SLEEF_QUAD_C(x) (x ## Q)
-//#elif defined(__SIZEOF_LONG_DOUBLE__) && defined(__aarch64__)
-//typedef long double Sleef_quad;
-//#define SLEEF_QUAD_C(x) (x ## L)
+#elif defined(__SIZEOF_LONG_DOUBLE__) && defined(__aarch64__)
+typedef long double Sleef_quad;
+#define SLEEF_QUAD_C(x) (x ## L)
 #else
 typedef struct { uint64_t x, y; } Sleef_quad;
 #endif

diff --git a/src/quad/sleefquadinline_header.h.org b/src/quad/sleefquadinline_header.h.org
@@ -25,9 +25,9 @@
 #if defined(__SIZEOF_FLOAT128__) || (defined(__linux__) && defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))) || (defined(__PPC64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 8)
 typedef __float128 Sleef_quad;
 #define SLEEF_QUAD_C(x) (x ## Q)
-//#elif defined(__SIZEOF_LONG_DOUBLE__) && defined(__aarch64__)
-//typedef long double Sleef_quad;
-//#define SLEEF_QUAD_C(x) (x ## L)
+#elif defined(__SIZEOF_LONG_DOUBLE__) && defined(__aarch64__)
+typedef long double Sleef_quad;
+#define SLEEF_QUAD_C(x) (x ## L)
 #else
 typedef struct { uint64_t x, y; } Sleef_quad;
 #endif

diff --git a/src/quad/sleefsimdqp.c b/src/quad/sleefsimdqp.c
@@ -261,15 +261,15 @@ static INLINE CONST VECTOR_CC vdouble2 twoprod_vd2_vd_vd(vdouble x, vdouble y) {
 }
 
 static INLINE CONST VECTOR_CC vdouble3 scale_vd3_vd3_vd(vdouble3 d, vdouble s) {
-  return (vdouble3) { vmul_vd_vd_vd(vd3getx_vd_vd3(d), s), vmul_vd_vd_vd(vd3gety_vd_vd3(d), s), vmul_vd_vd_vd(vd3getz_vd_vd3(d), s) };
+  return cast_vd3_vd_vd_vd(vmul_vd_vd_vd(vd3getx_vd_vd3(d), s), vmul_vd_vd_vd(vd3gety_vd_vd3(d), s), vmul_vd_vd_vd(vd3getz_vd_vd3(d), s));
 }
 
 static INLINE CONST VECTOR_CC vdouble3 scale_vd3_vd3_d(vdouble3 d, double s) { return scale_vd3_vd3_vd(d, vcast_vd_d(s)); }
 
 static INLINE CONST VECTOR_CC vdouble3 quickrenormalize_vd3_vd3(vdouble3 td) {
   vdouble2 u = quicktwosum_vd2_vd_vd(vd3getx_vd_vd3(td), vd3gety_vd_vd3(td));
   vdouble2 v = quicktwosum_vd2_vd_vd(vd2gety_vd_vd2(u), vd3getz_vd_vd3(td));
-  return (vdouble3) { vd2getx_vd_vd2(u), vd2getx_vd_vd2(v), vd2gety_vd_vd2(v) };
+  return cast_vd3_vd_vd_vd(vd2getx_vd_vd2(u), vd2getx_vd_vd2(v), vd2gety_vd_vd2(v));
 }
 
 static INLINE CONST VECTOR_CC vdouble3 normalize_vd3_vd3(vdouble3 td) {
@@ -405,31 +405,31 @@ static INLINE CONST VECTOR_CC vdouble3 mul_vd3_vd2_vd2(vdouble2 x, vdouble2 y) {
 }
 
 static INLINE CONST VECTOR_CC vdouble3 div2_vd3_vd3_vd3(vdouble3 n, vdouble3 q) {
-  vdouble2 d = ddrec_vd2_vd2((vdouble2) {vd3getx_vd_vd3(q), vd3gety_vd_vd3(q)});
+  vdouble2 d = ddrec_vd2_vd2(vcast_vd2_vd_vd(vd3getx_vd_vd3(q), vd3gety_vd_vd3(q)));
   return mul2_vd3_vd3_vd3(n, add_vd3_vd2_vd3(d, mul_vd3_vd2_vd3(ddscale_vd2_vd2_d(d, -1),
 								add_vd3_vd_vd3(vcast_vd_d(-1), mul_vd3_vd2_vd3(d, q)))));
 }
 
 static INLINE CONST VECTOR_CC vdouble3 div_vd3_vd3_vd3(vdouble3 n, vdouble3 q) {
-  vdouble2 d = ddrec_vd2_vd2((vdouble2) {vd3getx_vd_vd3(q), vd3gety_vd_vd3(q)});
+  vdouble2 d = ddrec_vd2_vd2(vcast_vd2_vd_vd(vd3getx_vd_vd3(q), vd3gety_vd_vd3(q)));
   return mul_vd3_vd3_vd3(n, add_vd3_vd2_vd3(d, mul_vd3_vd2_vd3(ddscale_vd2_vd2_d(d, -1),
 							       add_vd3_vd_vd3(vcast_vd_d(-1), mul_vd3_vd2_vd3(d, q)))));
 }
 
 static INLINE CONST VECTOR_CC vdouble3 rec_vd3_vd3(vdouble3 q) {
-  vdouble2 d = ddrec_vd2_vd2((vdouble2) {vd3getx_vd_vd3(q), vd3gety_vd_vd3(q)});
+  vdouble2 d = ddrec_vd2_vd2(vcast_vd2_vd_vd(vd3getx_vd_vd3(q), vd3gety_vd_vd3(q)));
   return add2_vd3_vd2_vd3(d, mul_vd3_vd2_vd3(ddscale_vd2_vd2_d(d, -1),
 					     add_vd3_vd_vd3(vcast_vd_d(-1), mul_vd3_vd2_vd3(d, q))));
 }
 
 static INLINE CONST VECTOR_CC vdouble3 rec_vd3_vd2(vdouble2 q) {
-  vdouble2 d = ddrec_vd2_vd2((vdouble2) {vd2getx_vd_vd2(q), vd2gety_vd_vd2(q)});
+  vdouble2 d = ddrec_vd2_vd2(vcast_vd2_vd_vd(vd2getx_vd_vd2(q), vd2gety_vd_vd2(q)));
   return add2_vd3_vd2_vd3(d, mul_vd3_vd2_vd3(ddscale_vd2_vd2_d(d, -1),
 					     add_vd3_vd_vd3(vcast_vd_d(-1), mul_vd3_vd2_vd2(d, q))));
 }
 
 static INLINE CONST VECTOR_CC vdouble3 sqrt_vd3_vd3(vdouble3 d) {
-  vdouble2 t = ddsqrt_vd2_vd2((vdouble2) {vd3getx_vd_vd3(d), vd3gety_vd_vd3(d)});
+  vdouble2 t = ddsqrt_vd2_vd2(vcast_vd2_vd_vd(vd3getx_vd_vd3(d), vd3gety_vd_vd3(d)));
   vdouble3 r = mul2_vd3_vd3_vd3(add2_vd3_vd3_vd3(d, mul_vd3_vd2_vd2(t, t)), rec_vd3_vd2(t));
   r = sel_vd3_vo_vd3_vd3(veq_vo_vd_vd(vd3getx_vd_vd3(d), vcast_vd_d(0)), cast_vd3_d_d_d(0, 0, 0), scale_vd3_vd3_d(r, 0.5));
   return r;

diff --git a/travis/before_script.arm64-gcc-sve.sh b/travis/before_script.arm64-gcc-sve.sh
@@ -5,4 +5,4 @@ cd sleef.build
 export PATH=/opt/local/bin:$PATH
 export LD_LIBRARY_PATH=/opt/local/lib:$LD_LIBRARY_PATH
 export CC=gcc-10
-cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../install -DSLEEF_SHOW_CONFIG=1 -DEMULATOR=qemu-aarch64 -DENFORCE_TESTER3=TRUE -DBUILD_INLINE_HEADERS=TRUE -DBUILD_QUAD=FALSE -DBUILD_DFT=TRUE -DENFORCE_SVE=TRUE ..
+cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../install -DSLEEF_SHOW_CONFIG=1 -DEMULATOR=qemu-aarch64 -DENFORCE_TESTER3=TRUE -DBUILD_INLINE_HEADERS=TRUE -DBUILD_QUAD=TRUE -DBUILD_DFT=TRUE -DENFORCE_SVE=TRUE ..