Merge branch 'mcount-nofp'

Fix crash of clang compiler when building uftrace source for x86-64.
The -mno-sse2 option can cause an internal compiler error when used
with floating-point types.  Clang folks suggested -mgeneral-regs-only
option, so let's use it instead.

Also change libmcount not to generate instructions that use floating-
point type in the common code (under utils/ directory).

Fixed: #1877

Signed-off-by: Namhyung Kim <[email protected]>

Makefile

@@ -89,7 +89,7 @@ BENCH_CFLAGS       = -D_GNU_SOURCE -g -pg $(CFLAGS_$@) $(CFLAGS_bench)
 TRACEEVENT_CFLAGS  = $(COMMON_CFLAGS) $(CFLAGS_$@) $(CFLAGS_traceevent)
 LIB_CFLAGS         = $(COMMON_CFLAGS) $(CFLAGS_$@) $(CFLAGS_lib)
 LIB_CFLAGS        += -fPIC -fvisibility=hidden -fno-omit-frame-pointer
-LIB_CFLAGS        += -fno-builtin -fno-tree-vectorize
+LIB_CFLAGS        += -fno-builtin -fno-tree-vectorize -DLIBMCOUNT
 TEST_CFLAGS        = $(COMMON_CFLAGS) -DUNIT_TEST
 PYTHON_CFLAGS      = $(COMMON_CFLAGS) -fPIC
 

check-deps/Makefile

@@ -15,12 +15,13 @@ CHECK_LIST += have_libunwind
 CHECK_LIST += have_libcapstone
 CHECK_LIST += cc_has_minline_all_stringops
 CHECK_LIST += have_libtraceevent
+CHECK_LIST += cc_has_mgeneral_regs_only
 
 #
 # This is needed for checking build dependency
 #
 
-CHECK_CFLAGS  = $(CFLAGS)  $(CFLAGS_$@) -Werror
+CHECK_CFLAGS  = $(CFLAGS)  $(CFLAGS_$@) -O2 -Werror
 CHECK_LDFLAGS = $(LDFLAGS) $(LDFLAGS_$@)
 
 # libpython3 provides an embed version of pkg-config file since python3.8
@@ -37,7 +38,7 @@ endif
 CFLAGS_cc_has_mfentry = -mfentry
 LDFLAGS_cxa_demangle = -lstdc++
 LDFLAGS_have_libelf = -lelf
-CFLAGS_cc_has_mno_sse2 = -O2 -mno-sse -mno-sse2
+CFLAGS_cc_has_mno_sse2 = -mno-sse -mno-sse2
 LDFLAGS_have_libpython2.7 = -lpython2.7
 CFLAGS_have_libpython3 = $(shell pkg-config python3$(EMBED) --cflags 2> /dev/null)
 LDFLAGS_have_libpython3 = $(shell pkg-config python3$(EMBED) --libs 2> /dev/null)
@@ -54,6 +55,7 @@ LDFLAGS_have_libunwind = $(shell pkg-config --libs   libunwind 2> /dev/null)
 CFLAGS_cc_has_minline_all_stringops = -minline-all-stringops
 CFLAGS_have_libtraceevent = $(shell pkg-config --cflags libtraceevent 2> /dev/null)
 LDFLAGS_have_libtraceevent = $(shell pkg-config --libs  libtraceevent 2> /dev/null)
+CFLAGS_cc_has_mgeneral_regs_only = -mgeneral-regs-only
 
 check-build: $(CHECK_LIST)
 

check-deps/Makefile.check

@@ -11,7 +11,9 @@ ifneq ($(wildcard $(objdir)/check-deps/cxa_demangle),)
   COMMON_LDFLAGS += -lstdc++
 endif
 
-ifneq ($(wildcard $(objdir)/check-deps/cc_has_mno_sse2),)
+ifneq ($(wildcard $(objdir)/check-deps/cc_has_mgeneral_regs_only),)
+  LIB_CFLAGS += -mgeneral-regs-only
+else ifneq ($(wildcard $(objdir)/check-deps/cc_has_mno_sse2),)
   LIB_CFLAGS += -mno-sse -mno-sse2
 endif
 

check-deps/__cc_has_mgeneral_regs_only.c

@@ -0,0 +1,7 @@
+#include <stdio.h>
+
+int main(void)
+{
+	printf("%f", 123.45);
+	return 0;
+}

utils/debug.c

@@ -275,6 +275,7 @@ void __pr_color(char code, const char *fmt, ...)
 	color(ec, outfp);
 }
 
+#ifndef LIBMCOUNT
 static void __print_time_unit(int64_t delta_nsec, bool needs_sign)
 {
 	uint64_t delta = llabs(delta_nsec);
@@ -441,3 +442,4 @@ void print_diff_count(uint64_t base, uint64_t pair)
 	else
 		pr_out("%9s", "+0");
 }
+#endif

utils/script-luajit.c

@@ -97,7 +97,7 @@ static void setup_argument_context(bool is_retval, struct script_context *sc_ctx
 		unsigned short slen;
 		char ch_str[2];
 		char *str;
-		double dval;
+		double dval __maybe_unused;
 
 		/* skip unwanted arguments or retval */
 		if (is_retval != (spec->idx == RETVAL_IDX))
@@ -143,6 +143,7 @@ static void setup_argument_context(bool is_retval, struct script_context *sc_ctx
 			break;
 		case ARG_FMT_FLOAT:
 			memcpy(val.v, data, spec->size);
+#ifndef LIBMCOUNT
 			switch (spec->size) {
 			case 4:
 				dval = val.f;
@@ -161,6 +162,7 @@ static void setup_argument_context(bool is_retval, struct script_context *sc_ctx
 			dllua_pushinteger(L, ++count);
 			dllua_pushnumber(L, dval);
 			dllua_settable(L, -3);
+#endif
 			data += ALIGN(spec->size, 4);
 			break;
 

utils/script-python.c

@@ -349,7 +349,7 @@ static void insert_tuple_string(PyObject *tuple, int idx, char *v)
 	python_insert_tuple(tuple, 's', idx, val);
 }
 
-static void insert_tuple_double(PyObject *tuple, int idx, double v)
+static void __maybe_unused insert_tuple_double(PyObject *tuple, int idx, double v)
 {
 	union python_val val = {
 		.f = v,
@@ -429,7 +429,7 @@ static void setup_argument_context(PyObject **pDict, bool is_retval, struct scri
 		unsigned short slen;
 		char ch_str[2];
 		char *str;
-		double dval;
+		double dval __maybe_unused;
 
 		/* skip unwanted arguments or retval */
 		if (is_retval != (spec->idx == RETVAL_IDX))
@@ -468,6 +468,7 @@ static void setup_argument_context(PyObject **pDict, bool is_retval, struct scri
 
 		case ARG_FMT_FLOAT:
 			memcpy(val.v, data, spec->size);
+#ifndef LIBMCOUNT
 			switch (spec->size) {
 			case 4:
 				dval = val.f;
@@ -484,6 +485,7 @@ static void setup_argument_context(PyObject **pDict, bool is_retval, struct scri
 				break;
 			}
 			insert_tuple_double(args, count++, dval);
+#endif
 			data += ALIGN(spec->size, 4);
 			break;
 

utils/script.h

@@ -52,9 +52,12 @@ union script_arg_val {
 	int i;
 	long l;
 	long long L;
+/* libmcount should not access floating-point types. */
+#ifndef LIBMCOUNT
 	float f;
 	double d;
 	long double D;
+#endif
 	unsigned char v[16];
 };