diff --git a/Makefile b/Makefile
index ccee39a..b9d1d9d 100644
--- a/Makefile
+++ b/Makefile
@@ -12,7 +12,7 @@ PROJECT=shuffle3
 #	DEBUG:		Pretend we're building a debug release even though we're not. Will enable additional output messages and may interfere with some optimisations
 FEATURE_FLAGS?=
 
-COMMON_FLAGS = $(addprefix -D,$(FEATURE_FLAGS)) -Wall -pedantic $(addprefix -I,$(INCLUDE)) -fno-strict-aliasing
+COMMON_FLAGS+= $(addprefix -D,$(FEATURE_FLAGS)) -Wall -Wstrict-aliasing $(addprefix -I,$(INCLUDE)) -fno-strict-aliasing
 
 OPT_FLAGS?= -march=native -fgraphite -fopenmp -floop-parallelize-all -ftree-parallelize-loops=4 \
 	    -floop-interchange -ftree-loop-distribution -floop-strip-mine -floop-block \
@@ -21,14 +21,17 @@ OPT_FLAGS?= -march=native -fgraphite -fopenmp -floop-parallelize-all -ftree-para
 CXX_OPT_FLAGS?= $(OPT_FLAGS) -felide-constructors
 
 CFLAGS   += $(COMMON_FLAGS) --std=gnu11
-CXXFLAGS += $(COMMON_FLAGS) --std=gnu++20 -fno-exceptions
+CXXFLAGS += $(COMMON_FLAGS) --std=gnu++23 -fno-exceptions
 LDFLAGS  +=  
 
 STRIP=strip
 
-RELEASE_CFLAGS?=   -O3 -flto $(OPT_FLAGS)
-RELEASE_CXXFLAGS?= -O3 -flto $(CXX_OPT_FLAGS)
-RELEASE_LDFLAGS?=  -O3 -flto
+RELEASE_COMMON_FLAGS?= -DNOTRACE
+RELEASE_COMMON_FLAGS+= -DRELEASE 
+
+RELEASE_CFLAGS?=   -O3 -flto $(OPT_FLAGS) $(RELEASE_COMMON_FLAGS)
+RELEASE_CXXFLAGS?= -O3 -flto $(CXX_OPT_FLAGS) $(RELEASE_COMMON_FLAGS)
+RELEASE_LDFLAGS?=  -Wl,-O3 -Wl,-flto
 
 DEBUG_CFLAGS?= -O0 -g -DDEBUG
 DEBUG_CXXFLAGS?= $(DEBUG_CFLAGS)
@@ -111,14 +114,19 @@ pgo-profile: | pgo-reset pgo-generate
 	for i in {1..$(PGO_ITERATIONS)}; do \
 		dd if=/dev/urandom of=$(PGO_SET_LOC)/small bs=1024 count=1 >> /dev/null 2>&1; \
 		printf "Iteration $$i / $(PGO_ITERATIONS)\r"; \
-		./pgo-generate -s $(PGO_SET_LOC)/small >>/dev/null; \
-		./pgo-generate -u $(PGO_SET_LOC)/small >>/dev/null; \
-		./pgo-generate -h >> /dev/null; \
-		FCNT=1 ./test.sh ./pgo-generate >>/dev/null; \
-		FCNT=2 ./test.sh ./pgo-generate >>/dev/null; \
-		FCNT=3 ./test.sh ./pgo-generate >>/dev/null; \
-		FCNT=4 ./test.sh ./pgo-generate >>/dev/null; \
+		( echo ">> $$i" >&2; \
+		  echo ">> $$i"; \
+		./pgo-generate -s $(PGO_SET_LOC)/small; \
+		./pgo-generate -u $(PGO_SET_LOC)/small; \
+		./pgo-generate -h; \
+		FCNT=1 ./test.sh ./pgo-generate; \
+		FCNT=2 ./test.sh ./pgo-generate; \
+		FCNT=3 ./test.sh ./pgo-generate; \
+		FCNT=4 ./test.sh ./pgo-generate; \
+		) >>$(PGO_SET_LOC)/stdout.log 2>>$(PGO_SET_LOC)/stderr.log; \
 	done 
+	$(shell command -v bat >/dev/null && echo "bat --pager=none" || echo cat)	$(PGO_SET_LOC)/stdout.log; \
+	$(shell command -v bat >/dev/null && echo "bat --pager=none" || echo cat)	$(PGO_SET_LOC)/stderr.log >&2
 	rm -rf $(PGO_SET_LOC)
 	rm pgo-generate
 
diff --git a/include/debug.h b/include/debug.h
index 334e6d1..29fe4d9 100644
--- a/include/debug.h
+++ b/include/debug.h
@@ -1,6 +1,14 @@
 #ifndef _DEBUG_H
 #define _DEBUG_H
 
+// If `TRACE` is defined, or if `NOTRACE` is defined, use simplified debug info
+#if defined(TRACE) || !defined(NOTRACE)
+#define D_TRACE
+#define	_D__FUNC __PRETTY_FUNCTION__
+#else
+#define _D__FUNC __func__
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -8,12 +16,14 @@ extern "C" {
 struct debuginfo {
 	const char* file;
 	const char* function;
+	//const char* pfunction; //TODO: for `#if defined(TRACE)`, __PRETTY_FUNCTION__ goes here, __func__ goes above.
 	int line;
 };
 
 void _do_dprintf(struct debuginfo di, const char* fmt, ...);
 
 #ifdef __cplusplus
+#define _D_COMPTIME constexpr
 extern "C++" {
 #include <utility>
 	template<typename... Args>
@@ -24,14 +34,15 @@ extern "C++" {
 		_do_dprintf(i, fmt, std::forward<Args>(args)...);
 #endif
 	}
-#define D_dprintf(fmt, ...) _real_dprintf(__FILE__, __func__, __LINE__, fmt __VA_OPT__(,) __VA_ARGS__)
+#define D_dprintf(fmt, ...) do { if consteval { (void)0; } else { _real_dprintf(__FILE__, _D__FUNC, __LINE__, fmt __VA_OPT__(,) __VA_ARGS__); } } while(0)
 }
 #else
+#define _D_COMPTIME
 
 #ifdef DEBUG
-#define D_dprintf(fmt, ...) _do_dprintf( (struct debuginfo){.file = __FILE__, .function = __func__, .line = __LINE__}, fmt __VA_OPT__(,) __VA_ARGS__)
+#define D_dprintf(fmt, ...) _do_dprintf( (struct debuginfo){.file = __FILE__, .function = _D__FUNC, .line = __LINE__}, fmt __VA_OPT__(,) __VA_ARGS__)
 #else
-static inline void _do__nothing(const char* fmt, ...) {}
+static _D_COMPTIME inline void _do__nothing(const char* fmt, ...) {}
 #define D_dprintf(fmt, ...) _do__nothing(fmt __VA_OPT__(,) __VA_ARGS__)  //(fmt __VA_OPT__(,)  __VA_ARGS__, (void)0)
 #endif
 #endif
@@ -39,5 +50,6 @@ static inline void _do__nothing(const char* fmt, ...) {}
 #ifdef __cplusplus
 }
 #endif
+#undef _D_COMPTIME
 
 #endif /* _DEBUG_H */
diff --git a/include/uuid.hpp b/include/uuid.hpp
index 5331ccc..56dc06c 100644
--- a/include/uuid.hpp
+++ b/include/uuid.hpp
@@ -1,5 +1,6 @@
 #pragma once
 
+#include <array>
 #include <random>
 #include <string>
 
diff --git a/src/debug.c b/src/debug.c
index c830c4b..d2bcc2f 100644
--- a/src/debug.c
+++ b/src/debug.c
@@ -1,17 +1,85 @@
 #include <stdlib.h>
 #include <stdio.h>
 #include <stdarg.h>
+#include <time.h>
+#include <unistd.h>
+#include <string.h>
 
 #include <debug.h>
 
 #include <colours.h>
 
+enum jt_comp {
+	JT_COMP_UNIX,
+	JT_COMP_DATE,
+	JT_COMP_TIME,
+	JT_COMP_DATETIME,
+};
+
+#define CLONE(p) ({  __auto_type $__p = (p); \
+			__auto_type $__r = __builtin_choose_expr(__builtin_classify_type($__p) == 5, *$__p, $__p); \
+			/* ((__typeof__($__r) []) { $__r }); }) */ \
+			__typeof__($__r) $__rp[1]; \
+			$__rp[0] = $__r; \
+			&$__rp[0]; })
+
+__attribute__((returns_nonnull, nonnull(1)))
+static inline
+const char* jtime(const struct tm* restrict lt, enum jt_comp kind)
+{
+	_Thread_local static char datebuf[1
+			+ 4 + 2 + 4 + 1 // `dd/mm/yyyy`
+			+ 6 + 3 // ` hh:MM:ss`
+	];
+	_Static_assert(sizeof(datebuf) != sizeof(char*), "Unexpected decay");
+
+	//memset(datebuf, '0', sizeof(datebuf)-1);
+	if(!kind)
+		snprintf(datebuf, sizeof(datebuf), "%lu", mktime(CLONE(lt)));
+	else {
+		size_t w = 0, sz = sizeof(datebuf);
+		if(kind & JT_COMP_DATE)
+		w += snprintf(datebuf, sz, "%02u/%02u/%4.4u"
+			, (unsigned)lt->tm_mday
+			, (unsigned)lt->tm_mon
+			, (unsigned)(1900 + lt->tm_year));
+
+		if(kind & JT_COMP_TIME)
+		w += snprintf(datebuf+w, sz - w, " %02u:%02u:%02u"
+			, (unsigned)lt->tm_hour
+			, (unsigned)lt->tm_min
+			, (unsigned)lt->tm_sec);
+	}
+	return datebuf;
+}
+
 void _do_dprintf(struct debuginfo info, const char* fmt, ...)
 {
 #ifdef DEBUG
 	va_list li;
+	
+	time_t utime;
+	struct tm ltime;
+	time(&utime);
+	localtime_r(&utime, &ltime);
+
 	va_start(li, fmt);
-	fprintf(stderr, "["FGRN("dbg")" " FYEL("@") BOLD("%s") "->"  FWHT("%s") ":" FYEL("%d") "]: ", info.file, info.function,info.line);
+
+#ifdef D_TRACE
+#define FMT_FUNC "%s"
+#else
+#define	FMT_FUNC "%s()"
+#endif
+
+#define TTY_OUT "["FGRN("dbg")" (%s %s) " FYEL("@") BOLD("%s") "->"  FWHT(FMT_FUNC) ":" FYEL("%d") "]: "
+#define NCOL(x) x
+#define FIL_OUT "["NCOL("dbg")" (%s %s) " NCOL("@") NCOL("%s") "->"  NCOL(FMT_FUNC) ":" NCOL("%d") "]: "
+
+	fprintf(stderr, isatty(fileno(stderr)) ? (TTY_OUT) : (FIL_OUT)
+			, jtime(&ltime, JT_COMP_DATETIME), ltime.tm_zone
+			, info.file
+			, info.function
+			, info.line);
 	vfprintf(stderr, fmt, li);
 	fprintf(stderr, "\n");
 	va_end(li);
diff --git a/src/work.cpp b/src/work.cpp
index 1df52b9..06efa40 100644
--- a/src/work.cpp
+++ b/src/work.cpp
@@ -1,5 +1,6 @@
 #include <tuple>
 #include <functional>
+#include <bit>
 #include <cfloat>
 
 #include <shuffle3.h>
@@ -62,12 +63,12 @@ namespace work
 
 			auto [long_l, long_h] = minmax_t(map.as_span().reinterpret<std::int64_t>());
 			D_dprintf("MMX res (u64): %ld -- %ld", long_l, long_h);
-			rng::xoroshiro128plus	xorng(*(const std::uint64_t*)&long_l, *(const std::uint64_t*)&long_h);
+			rng::xoroshiro128plus	xorng(std::bit_cast<std::uint64_t>(long_l), std::bit_cast<std::uint64_t>(long_h));
 			rng::unshuffle(xorng, map.as_span().reinterpret<std::int64_t>());
 		} else {
 			auto [long_l, long_h] = minmax_t(map.as_span().reinterpret<std::int64_t>());
 			D_dprintf("MMX res (u64): %ld -- %ld", long_l, long_h);
-			rng::xoroshiro128plus	xorng(*(const std::uint64_t*)&long_l, *(const std::uint64_t*)&long_h);
+			rng::xoroshiro128plus	xorng(std::bit_cast<std::uint64_t>(long_l), std::bit_cast<std::uint64_t>(long_h));
 			rng::shuffle(xorng, map.as_span().reinterpret<std::int64_t>());
 
 			auto [float_l, float_h] = minmax_t(map.as_span().reinterpret<float>(), [](float f) -> bool { return !( (f!=f) || f < -FLT_MAX || f > FLT_MAX); });