From 50fe56949ff6864d0f1cc6aa92a348188aaf7554 Mon Sep 17 00:00:00 2001
From: flysand7 <thebumboni@gmail.com>
Date: Sat, 5 Aug 2023 22:20:20 +1100
Subject: [PATCH] Fix SSE by aligning the stack

---
 build.py                    | 5 ++---
 src/linux/entry.c           | 2 +-
 src/loader/loader-entry.asm | 5 +++++
 3 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/build.py b/build.py
index 031ae15..2dcbb54 100755
--- a/build.py
+++ b/build.py
@@ -72,7 +72,7 @@ dependencies.append(args.compiler)
 includes = ['src/include']
 cc = args.compiler
 cc_defines = []
-cc_flags = ['-nostdlib', '-fno-stack-protector', '-mno-sse']
+cc_flags = ['-nostdlib', '-fno-stack-protector']
 crt_file = 'crt.lib'
 lib_file = 'cia.lib'
 dl_file = 'ld-cia.so'
@@ -272,7 +272,7 @@ dl_lib = f'lib/{dl_file}'
 dl_flags = [
     '-shared',
     '-nostdlib',
-    '-mno-sse',
+    # '-mno-sse',
     '-ffreestanding',
     '-fno-stack-protector',
     '-Wl,-e,_dlstart',
@@ -292,7 +292,6 @@ if target == 'linux':
 
 print_step(f'Compiling {crt_file}\n')
 if target == 'linux':
-    # assemble('src/linux/crt-entry.asm', 'bin/crt-entry.o')
     compile(['src/linux/crt-ctors.c'], 'bin/crt-ctors.o', '-c')
     archive(['bin/crt-ctors.o'], crt_lib)
 elif target == 'windows':
diff --git a/src/linux/entry.c b/src/linux/entry.c
index fe63f79..b9dac28 100644
--- a/src/linux/entry.c
+++ b/src/linux/entry.c
@@ -14,7 +14,7 @@ void __stack_chk_fail(void) {
 extern int main(int argc, char **argv, char **envp);
 static void _fileapi_init();
 
-void _start() {
+void _cia_start() {
     // Get the envp
     // char **envp = argv + (argc + 1);
     // init(argc, argv, envp);
diff --git a/src/loader/loader-entry.asm b/src/loader/loader-entry.asm
index 24cd490..9759162 100644
--- a/src/loader/loader-entry.asm
+++ b/src/loader/loader-entry.asm
@@ -11,3 +11,8 @@ _dlstart:
     mov rdi, rsp
     lea rsi, _DYNAMIC
     and rsp, -16
+    ; `call` pushes 8-byte value onto the stack
+    ; by pushing an 8-bit value ourselves we can make
+    ; sure the stack is aligned after rbp push in prologue
+    sub rsp, 8
+