diff --git a/src/asahi/compiler/agx_optimizer.c b/src/asahi/compiler/agx_optimizer.c
index 1df093b72249..ff2ac82e865c 100644
--- a/src/asahi/compiler/agx_optimizer.c
+++ b/src/asahi/compiler/agx_optimizer.c
@@ -616,6 +616,15 @@ agx_optimizer_fmov_rev(agx_context *ctx, agx_instr *I, agx_instr *use)
        (I->src[0].abs || I->src[0].neg || I->src[1].abs || I->src[1].neg))
       return false;
 
+   /* M5 (AGX3C): the native perspective multiply only encodes a word
+    * destination (agx3_persp_mul_encode). Folding the f32-to-f16 move that
+    * narrows a mediump varying into it made the binary packer refuse the
+    * shader and the draw was skipped. Keep the explicit move. */
+   if (ctx->key && ctx->key->dev.encoding == AGX_ISA_ENCODING_AGX3C &&
+       I->op == AGX_OPCODE_PERSP_MUL_AGX3 &&
+       I->dest[0].size == AGX_SIZE_32 && use->dest[0].size == AGX_SIZE_16)
+      return false;
+
    /* saturate(saturate(x)) = saturate(x) */
    I->saturate |= use->saturate;
    I->dest[0] = use->dest[0];
diff --git a/src/asahi/compiler/test/test-optimizer.cpp b/src/asahi/compiler/test/test-optimizer.cpp
index 0bd63a55091b..2317b594ae89 100644
--- a/src/asahi/compiler/test/test-optimizer.cpp
+++ b/src/asahi/compiler/test/test-optimizer.cpp
@@ -204,6 +204,31 @@ TEST_F(Optimizer, G17PPreservesHalfWidenBeforeFltnorm)
    ASSERT_SHADER_EQUAL(actual->shader, expected->shader);
 }
 
+/*
+ * The native projective f32mul (PERSP_MUL) has no 16-bit destination form. A
+ * mediump varying narrows its product with an f32 to f16 move; folding that
+ * move into the multiply made the shader unencodable on M5 (AGX3C), so the
+ * move stays a separate instruction.
+ */
+TEST_F(Optimizer, G17CKeepsPerspMulNarrowingMove)
+{
+   agx_builder *actual = agx_test_builder(mem_ctx);
+   agx_builder *expected = agx_test_builder(mem_ctx);
+   struct agx_shader_key key = {};
+   key.dev.encoding = AGX_ISA_ENCODING_AGX3C;
+   key.dev.chip = AGX_CHIP_G17C;
+   for (agx_builder *b : {actual, expected}) {
+      b->shader->key = &key;
+      agx_index product = agx_temp(b->shader, AGX_SIZE_32);
+      agx_persp_mul_agx3_to(b, product, wx, wy, 1);
+      agx_index narrow = agx_temp(b->shader, AGX_SIZE_16);
+      agx_fmov_to(b, narrow, product);
+      agx_unit_test(b, narrow);
+   }
+   agx_optimize_and_dce(actual->shader);
+   ASSERT_SHADER_EQUAL(actual->shader, expected->shader);
+}
+
 /*
  * AGX3 has no select form without a register compare operand, so a select of
  * two immediates -- here fused out of a compare of two constants -- has to