# RUN: %stinkytofu-opt --arch gfx1250 %s --StinkyBuildImplicitDependencyPass --StinkyWaitCntInsertionPass --print-output
#
# StinkyWaitCntInsertionPass (gfx1250): insert s_wait_dscnt before each WMMA
# that consumes outstanding ds_load results. The pass inserts an actual
# s_wait_dscnt instruction whose immediate is the number of ds_load results
# that should still be OUTSTANDING after the wait. The inserted instruction
# also carries an `mod.swaitcnt` modifier exposing the semantic fields
# (vlcnt / vscnt / dlcnt / dscnt / kmcnt) that downstream emitters consume.
#
# Pattern (single-block):
#   ds_load_b64 v[20:21], v0        (ds_load 0)
#   ds_load_b64 v[30:31], v0        (ds_load 1)
#   ds_load_b64 v[40:41], v0        (ds_load 2)
#   ds_load_b64 v[50:51], v0        (ds_load 3)
#   v_wmma_f32_16x16x32_bf16        (uses v[20:27], v[30:37] -> needs ds_load 0,1)
#   v_wmma_f32_16x16x32_bf16        (uses v[40:47], v[50:57] -> needs ds_load 2,3)
#
# Expected waits inserted by the pass:
#   s_wait_dscnt 2   before WMMA1  (wait for ds_load 0,1 ; leave 2,3 outstanding)
#   s_wait_dscnt 0   before WMMA2  (wait for the remaining ds_load 2,3)
#
# CHECK-LABEL: @waitcnt_insertion_wmma_dsread
# CHECK: ds_load_b64
# CHECK-NEXT: ds_load_b64
# CHECK-NEXT: ds_load_b64
# CHECK-NEXT: ds_load_b64
# CHECK-NEXT: s_wait_dscnt
# CHECK-SAME: dlcnt = 2
# CHECK-NEXT: v_wmma_f32_16x16x32_bf16
# CHECK-NEXT: s_wait_dscnt
# CHECK-SAME: dlcnt = 0
# CHECK-NEXT: v_wmma_f32_16x16x32_bf16
# Only two waitcnt inserts should exist, and no tensor waitcnt should appear
# (no tensor loads are present in this block, so the tensor-wait heuristic is a no-op).
# CHECK-NOT: s_wait_tensorcnt
# CHECK-NOT: s_wait_loadcnt

st.func @waitcnt_insertion_wmma_dsread() {
^entry:
  v[20:21] = "st.ds_load_b64"(v0) { mod.ds = { na = 1, offset = 0, gds = false }, mod.memtoken = { tokens = [0] } }
  v[30:31] = "st.ds_load_b64"(v0) { mod.ds = { na = 1, offset = 16, gds = false }, mod.memtoken = { tokens = [0] } }
  v[40:41] = "st.ds_load_b64"(v0) { mod.ds = { na = 1, offset = 32, gds = false }, mod.memtoken = { tokens = [0] } }
  v[50:51] = "st.ds_load_b64"(v0) { mod.ds = { na = 1, offset = 48, gds = false }, mod.memtoken = { tokens = [0] } }
  a[10:17] = "st.v_wmma_f32_16x16x32_bf16"(v[20:27], v[30:37], a[10:17]) { mod.mfma = { reuseA = false, reuseB = false } }
  a[10:17] = "st.v_wmma_f32_16x16x32_bf16"(v[40:47], v[50:57], a[10:17]) { mod.mfma = { reuseA = false, reuseB = false } }
}