nv50: comment on a so-far unseen bug

diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c
index af5ca84c8e79e63e96cfa9b79333d63affe27ecf..39597c548143322933cc47c5053ad59be8dfe6da 100644 (file)
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -22,6 +22,9 @@
  *     - Maybe even relax restrictions a bit, can't do P_RESULT + P_IMMD,
  *       but can emit to P_TEMP first - then MOV later. NVIDIA does this
  *
+ * In ops such as ADD it's possible to construct a bad opcode in the !is_long()
+ * case, if the emit_src() causes the inst to suddenly become long.
+ *
  * Verify half-insns work where expected - and force disable them where they
  * don't work - MUL has it forcibly disabled atm as it fixes POW..
  *