more iterjit

1 files changed, 34 insertions, 2 deletions

diff --git a/iterjit.c b/iterjit.c
index 0e069a0..e393567 100644
--- a/iterjit.c
+++ b/iterjit.c
@@ -604,7 +604,7 @@ dest_a8r8g8b8_begin (jit_dest_iter_t *	dest,
 	I_shl, dest->stride, IMM (2),
 
 	/* Line */
-	I_mov, dest->line, MEMBER(image, pixman_image_t, bits.bits),
+	I_mov, dest->line, MEMBER (image, pixman_image_t, bits.bits),
 	I_mov, tmp, dest->stride,
 	I_imul2, tmp, MEMBER (info, pixman_composite_info_t, dest_y),
 	I_add, dest->line, tmp,
@@ -663,6 +663,16 @@ dest_a8r8g8b8_process_line (jit_dest_iter_t *dest,
 
 	BEGIN_ASM (jit->fragment)
 	    I_jmp, LABEL (test),
+	END_ASM();
+	if (i == 2)
+	{
+	    /* Cache-line align the main loop */
+	    BEGIN_ASM (jit->fragment)
+		I_align, IMM (64),
+	    END_ASM ();
+	}
+	BEGIN_ASM (jit->fragment)
+	    DEFINE_LABEL (loop),
 	    I_sub, dest->w, IMM (n_pixels[i]),
 	END_ASM ();
 	
@@ -915,9 +925,23 @@ generate_kernel (jit_t *jit,
     reg_t h, composite_info;
 
     jit_switch_group (jit, "outer");
-    jit_preserve_gp (jit, rsi);
+
     composite_info = rsi;
     
+    jit_preserve_gp (jit, composite_info);
+
+    /* Preserve callee-save registers */
+    jit_preserve_gp (jit, rbx);
+    jit_preserve_gp (jit, r12);
+    jit_preserve_gp (jit, r13);
+    jit_preserve_gp (jit, r14);
+    jit_preserve_gp (jit, r15);
+
+    BEGIN_ASM (jit->fragment)
+	I_push, rbp,
+	I_mov, rbp, rsp,
+    END_ASM();
+	
     h = jit_alloc_gp (jit);
 
     BEGIN_ASM (jit->fragment)
@@ -957,8 +981,16 @@ generate_kernel (jit_t *jit,
 	mask_iter->end (mask_iter, jit);
     src_iter->end (src_iter, jit);
 
+    jit_switch_group (jit, "outer");
+    /* Restore callee-save registers */
+    jit_reload (jit, rbx);
+    jit_reload (jit, r12);
+    jit_reload (jit, r13);
+    jit_reload (jit, r14);
+    jit_reload (jit, r15);
     BEGIN_ASM (jit->fragment)
 	DEFINE_LABEL ("done"),
+	I_pop, rbp,
     END_ASM();
 }