# RUN: %stinkytofu-opt --arch gfx1250 %s --StinkyBuildImplicitDependencyPass --StinkyDAGSchedulerPass --print-output
#
# Full LDS prefetch flow through the DAG scheduler (gfx1250):
#
#   buffer_load  ->  ds_store  ->  s_barrier(signal/wait)  ->  ds_load  ->  v_wmma
#
# StinkyBuildImplicitDependencyPass ties the ds_store / barrier / ds_load on
# each memory token together via LDS pseudo registers. Given that def-use
# constraint, StinkyDAGSchedulerPass MUST preserve the producer -> barrier ->
# consumer ordering for each token group. Within the pre-barrier region the
# scheduler is free to interleave the independent chains (token 0 vs token 1)
# and order buffer_loads; within the post-barrier region the ds_loads may be
# interleaved as they feed a common WMMA.
#
# Input (single block):
#   buffer_load_b128 v[50:53]           (global, no memtoken, feeds ds_store(t=0))
#   buffer_load_b128 v[54:57]           (feeds ds_store(t=0))
#   buffer_load_b128 v[58:61]           (feeds ds_store(t=1))
#   buffer_load_b128 v[62:65]           (feeds ds_store(t=1))
#   ds_store_b128    (v10, v[50:53])    tokens=[0]
#   ds_store_b128    (v10, v[54:57])    tokens=[0]
#   ds_store_b128    (v11, v[58:61])    tokens=[1]
#   ds_store_b128    (v11, v[62:65])    tokens=[1]
#   s_barrier_signal (-1)               tokens=[0, 1]
#   s_barrier_wait   (-1)               tokens=[0, 1]
#   ds_load_b128     v[20:23]           tokens=[0]   (operand A first half)
#   ds_load_b128     v[24:27]           tokens=[0]   (operand A second half)
#   ds_load_b128     v[30:33]           tokens=[1]   (operand B first half)
#   ds_load_b128     v[34:37]           tokens=[1]   (operand B second half)
#   v_wmma_f32_16x16x16_bf16  v[20:27] x v[30:37]
#
# Required ordering after BuildImplicitDep + DAGScheduler:
#   - at least one buffer_load appears before any ds_store (value dependency)
#   - every ds_store appears before both barrier instructions
#   - the barrier_signal appears before the barrier_wait
#   - every ds_load appears after the barrier_wait  (LDS token edge)
#   - the WMMA is the last memory-sensitive instruction in the block
#
# CHECK-LABEL: @dag_wmma_prefetch_flow
#
# ---- Pre-barrier stage: at least one buffer_load must precede the ds_stores ----
# CHECK:      buffer_load_b128
#
# ---- Pre-barrier stage: ds_stores (all of them) must precede the barrier ----
# CHECK:      ds_store_b128
#
# ---- LDS synchronization barrier ----
# CHECK:      s_barrier_signal
# CHECK:      s_barrier_wait
#
# ---- Post-barrier stage: every ds_load appears after the barrier ----
# CHECK:      ds_load_b128
#
# ---- WMMA is the terminator of the memory-producer/consumer chain ----
# CHECK:      v_wmma_f32_16x16x16_bf16
#
# ---- Nothing memory-related may appear after the WMMA in this block ----
# CHECK-NOT: buffer_load_b128
# CHECK-NOT: ds_store_b128
# CHECK-NOT: ds_load_b128
# CHECK-NOT: s_barrier_signal
# CHECK-NOT: s_barrier_wait

st.func @dag_wmma_prefetch_flow() {
^entry:
  v[50:53] = "st.buffer_load_b128"(v8, s[0:3], 0) { issueCycles = 12, latencyCycles = 116, mod.mubuf = { offen = true, offset12 = 0, glc = false, slc = false, nt = false, lds = false } }
  v[54:57] = "st.buffer_load_b128"(v8, s[0:3], s0) { issueCycles = 12, latencyCycles = 116, mod.mubuf = { offen = true, offset12 = 0, glc = false, slc = false, nt = false, lds = false } }
  v[58:61] = "st.buffer_load_b128"(v9, s[0:3], 0) { issueCycles = 12, latencyCycles = 116, mod.mubuf = { offen = true, offset12 = 0, glc = false, slc = false, nt = false, lds = false } }
  v[62:65] = "st.buffer_load_b128"(v9, s[0:3], s0) { issueCycles = 12, latencyCycles = 116, mod.mubuf = { offen = true, offset12 = 0, glc = false, slc = false, nt = false, lds = false } }
  "st.ds_store_b128"(v10, v[50:53]) { issueCycles = 1, latencyCycles = 1, mod.ds = { na = 1, offset = 0, gds = false }, mod.memtoken = { tokens = [0] } }
  "st.ds_store_b128"(v10, v[54:57]) { issueCycles = 1, latencyCycles = 1, mod.ds = { na = 1, offset = 32, gds = false }, mod.memtoken = { tokens = [0] } }
  "st.ds_store_b128"(v11, v[58:61]) { issueCycles = 1, latencyCycles = 1, mod.ds = { na = 1, offset = 0, gds = false }, mod.memtoken = { tokens = [1] } }
  "st.ds_store_b128"(v11, v[62:65]) { issueCycles = 1, latencyCycles = 1, mod.ds = { na = 1, offset = 32, gds = false }, mod.memtoken = { tokens = [1] } }
  "st.s_barrier_signal"(-1) { issueCycles = 1, latencyCycles = 1, mod.memtoken = { tokens = [0, 1] } }
  "st.s_barrier_wait"(-1) { issueCycles = 1, latencyCycles = 1, mod.memtoken = { tokens = [0, 1] } }
  v[20:23] = "st.ds_load_b128"(v12) { issueCycles = 1, latencyCycles = 56, mod.ds = { na = 1, offset = 0, gds = false }, mod.memtoken = { tokens = [0] } }
  v[24:27] = "st.ds_load_b128"(v12) { issueCycles = 1, latencyCycles = 56, mod.ds = { na = 1, offset = 32, gds = false }, mod.memtoken = { tokens = [0] } }
  v[30:33] = "st.ds_load_b128"(v13) { issueCycles = 1, latencyCycles = 56, mod.ds = { na = 1, offset = 0, gds = false }, mod.memtoken = { tokens = [1] } }
  v[34:37] = "st.ds_load_b128"(v13) { issueCycles = 1, latencyCycles = 56, mod.ds = { na = 1, offset = 32, gds = false }, mod.memtoken = { tokens = [1] } }
  v[0:7] = "st.v_wmma_f32_16x16x16_bf16"(v[20:27], v[30:37], v[0:7]) { mod.mfma = { reuseA = false, reuseB = false } }
}