diff --git a/src/panfrost/bifrost/valhall/va_lower_split_64bit.c b/src/panfrost/bifrost/valhall/va_lower_split_64bit.c
index bc315221f86..e193ab32109 100644
--- a/src/panfrost/bifrost/valhall/va_lower_split_64bit.c
+++ b/src/panfrost/bifrost/valhall/va_lower_split_64bit.c
@@ -34,16 +34,6 @@
 static void
 lower_split_src(bi_context *ctx, bi_instr *I, unsigned s)
 {
-   /* Fix up for mismatch between IR and Valhall BLEND */
-   if (I->op == BI_OPCODE_BLEND) {
-      /* Only the blend descriptor needs lowering */
-      if (s != 1)
-         return;
-
-      /* But it is in a different place! */
-      s = 2;
-   }
-
    /* Skip sources that are already split properly */
    bi_index offset_fau = I->src[s];
    offset_fau.offset++;
diff --git a/src/panfrost/bifrost/valhall/valhall.h b/src/panfrost/bifrost/valhall/valhall.h
index a634a1487eb..df14129ed66 100644
--- a/src/panfrost/bifrost/valhall/valhall.h
+++ b/src/panfrost/bifrost/valhall/valhall.h
@@ -112,12 +112,17 @@ valhall_opcodes[BI_NUM_OPCODES];
 /* Bifrost specifies the source of bitwise operations as (A, B, shift), but
  * Valhall specifies (A, shift, B). We follow Bifrost conventions in the
  * compiler, so normalize.
+ *
+ * Bifrost specifies BLEND as staging + (coverage, blend descriptor), but
+ * Valhall specifies staging + (blend descriptor, coverage). Given we put
+ * staging sources first, this works out to the same swap as bitwise ops.
  */
 
 static inline bool
 va_swap_12(enum bi_opcode op)
 {
    switch (op) {
+   case BI_OPCODE_BLEND:
    case BI_OPCODE_LSHIFT_AND_I32:
    case BI_OPCODE_LSHIFT_AND_V2I16:
    case BI_OPCODE_LSHIFT_AND_V4I8: