[Tinycc-devel] [PATCH] Eliminate call to memset() on x86-64 when zeroing an array

[Yao Zi]

See thread "win32: -Wl,-nostdlib: undefined symbol 'memset'"

This patch has been tested on x86-64 Alpine (musl) and all tests are passed.
I do not have a Windows platform, so many thanks if someone could help me test
it.

---
 tccgen.c     |  5 +++++
 x86_64-gen.c | 38 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 43 insertions(+)

diff --git a/tccgen.c b/tccgen.c
index fd0680f..da07675 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -7373,6 +7373,10 @@ static void init_putz(init_params *p, unsigned long c, 
int size)
     if (p->sec) {
         /* nothing to do because globals are already set to zero */
     } else {
+#ifdef TCC_TARGET_NATIVE_PUT_ZERO
+        vseti(VT_LOCAL, c);
+        gen_put_zero(size);
+#else
         vpush_helper_func(TOK_memset);
         vseti(VT_LOCAL, c);
 #ifdef TCC_TARGET_ARM
@@ -7383,6 +7387,7 @@ static void init_putz(init_params *p, unsigned long c, 
int size)
         vpushs(size);
 #endif
         gfunc_call(3);
+#endif
     }
 }
 
diff --git a/x86_64-gen.c b/x86_64-gen.c
index e04df10..fae6f54 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -109,7 +109,9 @@ enum {
 #define PROMOTE_RET
 
 #define TCC_TARGET_NATIVE_STRUCT_COPY
+#define TCC_TARGET_NATIVE_PUT_ZERO
 ST_FUNC void gen_struct_copy(int size);
+ST_FUNC void gen_put_zero(int size);
 
 /******************************************************/
 #else /* ! TARGET_DEFS_ONLY */
@@ -2325,6 +2327,42 @@ ST_FUNC void gen_struct_copy(int size)
     vpop();
 }
 
+ST_FUNC void gen_put_zero(int size)
+{
+    int n = size / PTR_SIZE;
+
+#ifdef TCC_TARGET_PE
+    o(0x57); /* push rdi */
+#endif
+
+    vpushi(0);
+    gv2(RC_RDI, RC_RAX);
+
+    if (n <= 4) {
+        for (; n; n--)
+            o(0xab48);
+    } else {
+        vpushi(n);
+        gv(RC_RCX);
+        o(0xab48f3);
+       vpop();
+    }
+
+    if (size & 0x04)
+        o(0xab);
+    if (size & 0x02)
+        o(0xab66);
+    if (size & 0x01)
+        o(0xaa);
+
+#ifdef TCC_TARGET_PE
+   o(0x5f); /* pop rdi */
+#endif
+
+   vpop();
+   vpop();
+}
+
 /* end of x86-64 code generator */
 /*************************************************************/
 #endif /* ! TARGET_DEFS_ONLY */
-- 
2.42.0