# RUN: %stinkytofu-opt --arch gfx1250 %s --BuildUseDefChainPass --PeepholeOptimizationPass --print-output
#
# Move Propagation: v_mov_b32 + consumer → consumer with propagated source.
# Single-use MOV is eliminated; consumer uses the original source register.
# Converted from PeepholeOptimizationPassTest.*MovPropagation*

#--- F32 add: v_mov v1,v0 + v_add v2,2.0,v1 → v_add v2,2.0,v0
# CHECK-LABEL: @mov_prop_f32_add
# CHECK: v_add_f32
# CHECK-SAME: v2
# CHECK-SAME: 2
# CHECK-SAME: v0
# CHECK-NOT: v_mov_b32

st.func @mov_prop_f32_add() {
^entry:
  v1 = "st.v_mov_b32"(v0) { issueCycles = 1, latencyCycles = 5 }
  v2 = "st.v_add_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 }
}

#--- F16 add
# CHECK-LABEL: @mov_prop_f16_add
# CHECK: v_add_f16
# CHECK-SAME: v2
# CHECK-SAME: 2
# CHECK-SAME: v0
# CHECK-NOT: v_mov_b32

st.func @mov_prop_f16_add() {
^entry:
  v1 = "st.v_mov_b32"(v0) { issueCycles = 1, latencyCycles = 5 }
  v2 = "st.v_add_f16"(2.0, v1) { issueCycles = 1, latencyCycles = 5 }
}

#--- F32 mul
# CHECK-LABEL: @mov_prop_f32_mul
# CHECK: v_mul_f32
# CHECK-SAME: v2
# CHECK-SAME: 2
# CHECK-SAME: v0
# CHECK-NOT: v_mov_b32

st.func @mov_prop_f32_mul() {
^entry:
  v1 = "st.v_mov_b32"(v0) { issueCycles = 1, latencyCycles = 5 }
  v2 = "st.v_mul_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 }
}

#--- F16 mul
# CHECK-LABEL: @mov_prop_f16_mul
# CHECK: v_mul_f16
# CHECK-SAME: v2
# CHECK-SAME: 2
# CHECK-SAME: v0
# CHECK-NOT: v_mov_b32

st.func @mov_prop_f16_mul() {
^entry:
  v1 = "st.v_mov_b32"(v0) { issueCycles = 1, latencyCycles = 5 }
  v2 = "st.v_mul_f16"(2.0, v1) { issueCycles = 1, latencyCycles = 5 }
}

#--- Negative: MOV result has multiple uses → no propagation
# CHECK-LABEL: @no_mov_prop_multiple_uses
# CHECK: v_mov_b32
# CHECK: v_add_f32
# CHECK: v_mul_f32

st.func @no_mov_prop_multiple_uses() {
^entry:
  v1 = "st.v_mov_b32"(v0) { issueCycles = 1, latencyCycles = 5 }
  v2 = "st.v_add_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 }
  v3 = "st.v_mul_f32"(3.0, v1) { issueCycles = 1, latencyCycles = 5 }
}