From 643eece25370447aeacf4b772f0d48af803fcc9a Mon Sep 17 00:00:00 2001
From: Avril <flanchan@cumallover.me>
Date: Wed, 9 Dec 2020 17:47:51 +0000
Subject: [PATCH] day9: fix alignment

day9: pgo
---
 .gitignore    |  3 +++
 day9/Makefile | 36 +++++++++++++++++++++++++++++++-----
 day9/day9.c   |  2 --
 day9/map.c    | 21 ++++++++++++++++++++-
 4 files changed, 54 insertions(+), 8 deletions(-)

diff --git a/.gitignore b/.gitignore
index 7186bb3..37efd27 100644
--- a/.gitignore
+++ b/.gitignore
@@ -4,3 +4,6 @@ part2*
 
 *.lock
 target/
+
+*.o
+*.gcda
diff --git a/day9/Makefile b/day9/Makefile
index d35f96e..6b996ca 100644
--- a/day9/Makefile
+++ b/day9/Makefile
@@ -5,7 +5,7 @@ OPT_FLAGS?= -march=native -flto \
 	 -floop-interchange -ftree-loop-distribution -floop-strip-mine -floop-block \
 	 -fno-stack-check -fno-strict-aliasing 
 
-COMMON_FLAGS?=-pipe -O3 -Wall -pedantic -Wextra -Wstrict-aliasing
+COMMON_FLAGS?=-pipe -O3 -Wall -pedantic -Wextra -Wstrict-aliasing -Wcast-align
 
 C_OPT_FLAGS=$(OPT_FLAGS)
 CXX_OPT_FLAGS=$(OPT_FLAGS) -felide-constructors -fno-exceptions
@@ -19,9 +19,19 @@ LDFLAGS?=$(LD_OPT_FLAGS)
 
 CFLAGS+=-I$(INCLUDE)
 
+STRIP=strip
+
+# Number of iterations to profile program
+PROF_ITERATIONS?=40
+# Set concurrent profiling runs to be number of *physical* cpu cores
+PROF_CONCURRENT?=$(shell grep ^cpu\\scores /proc/cpuinfo | uniq |  awk '{print $4}') 
+
 .PHONY: all
 all: part1 part2
 
+.PHONY: test
+test: part1-test part2-test
+
 inpu%.h: inpu%
 	@rm -f $@
 	while read line; do \
@@ -30,16 +40,32 @@ inpu%.h: inpu%
 	
 part1: $(SRC) | input.h
 	$(CC) $(SRC) $(CFLAGS) -o $@ $(LDFLAGS)
-	strip $@
+	$(STRIP) $@
 
 part2: $(SRC) | input.h
-	$(CC) $(SRC) -DPART2 $(CFLAGS) -o $@ $(LDFLAGS)
-	strip $@
+	rm -f *.gcda
+	$(CC) $(SRC) -DPART2 $(CFLAGS) -fprofile-generate -o $@ $(LDFLAGS)
+	for i in {0..$(PROF_ITERATIONS)}; do \
+		printf "\rIteration $$i / $(PROF_ITERATIONS)"; \
+		for j in {0..$(PROF_CONCURRENT)}; do \
+			./$@ >> /dev/null 2>&1 & \
+		done; \
+		wait; \
+	done
+	@echo " DONE"
+	rm $@
+	$(CC) $(SRC) -DPART2 $(CFLAGS) -fprofile-use -fprofile-correction -o $@ -lgcov $(LDFLAGS) -fprofile-use -fprofile-correction
+	$(STRIP) $@
 
 part1-test: $(SRC) | input-test.h
 	$(CC) $(SRC) -DTEST $(CFLAGS) -o $@ $(LDFLAGS)
-	strip $@
+	$(STRIP) $@
+
+part2-test: $(SRC) | input-test.h
+	$(CC) $(SRC) -DTEST -DPART2 $(CFLAGS) -o $@ $(LDFLAGS)
+	$(STRIP) $@
 
 clean:
 	rm -f part{1,2}{,-test}
 	rm -f input{,-test}.h
+	rm -f *.gcda
diff --git a/day9/day9.c b/day9/day9.c
index e7239bd..6df48ad 100644
--- a/day9/day9.c
+++ b/day9/day9.c
@@ -24,8 +24,6 @@ const u64 input[] = {
 #endif
 };
 
-#define box(t) (calloc(sizeof(t), 1))
-
 struct pair {
 	bool valid;
 	u64 num;
diff --git a/day9/map.c b/day9/map.c
index 4674858..225380d 100644
--- a/day9/map.c
+++ b/day9/map.c
@@ -3,8 +3,27 @@
 #include <string.h>
 
 #include "map.h"
-
+/*
+static inline void* align_ptr(void* ptr, size_t al)
+{
+	unsigned char* buffer = ptr;
+	return buffer + al - ((intptr_t)buffer) % al;
+}
+*/
+#ifdef _EXPR_EXT			
+#define box(t) ({ t* restrict p = aligned_alloc(_Alignof(t), sizeof(t)); \
+			*p = (t){0}; \
+		p; })
+#elif !defined(IGNORE_ALIGMENT)
+inline extern void* _zero_ptr(void* ptr, size_t n)
+{
+	memset(ptr, 0, n);
+	return ptr;
+}
+#define box(t) _zero_ptr(aligned_alloc(_Alignof(t), sizeof(t)), sizeof(t))
+#else
 #define box(t) calloc(sizeof(t), 1)
+#endif
 
 struct page {
 	struct entry {