# RUN: %stinkytofu-opt --arch gfx1250 %s --BuildUseDefChainPass --PeepholeOptimizationPass --print-output # # ADD+MUL Fusion: v_add + v_mul with constants → single v_fma. # v_add v0,C1,v1 + v_mul v2,C2,v0 → v_fma v2,C2,v1,(C1*C2) # Converted from PeepholeOptimizationPassTest.*ADDMUL* #--- F32 basic: v_add v0,2.0,v1 + v_mul v2,3.0,v0 → v_fma v2,3.0,v1,6.0 # CHECK-LABEL: @add_mul_f32_basic # CHECK: v_fma_f32 # CHECK-SAME: v2 # CHECK-SAME: 3 # CHECK-SAME: v1 # CHECK-SAME: 6 # CHECK-NOT: v_add_f32 # CHECK-NOT: v_mul_f32 st.func @add_mul_f32_basic() { ^entry: v0 = "st.v_add_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(3.0, v0) { issueCycles = 1, latencyCycles = 5 } } #--- F16 basic # CHECK-LABEL: @add_mul_f16_basic # CHECK: v_fma_f16 # CHECK-SAME: 3 # CHECK-SAME: v1 # CHECK-SAME: 6 # CHECK-NOT: v_add_f16 # CHECK-NOT: v_mul_f16 st.func @add_mul_f16_basic() { ^entry: v0 = "st.v_add_f16"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f16"(3.0, v0) { issueCycles = 1, latencyCycles = 5 } } #--- Negative: ADD result has multiple uses → no fusion # CHECK-LABEL: @no_fusion_addmul_multiple_uses # CHECK: v_add_f32 # CHECK: v_mul_f32 # CHECK: v_add_f32 st.func @no_fusion_addmul_multiple_uses() { ^entry: v0 = "st.v_add_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(3.0, v0) { issueCycles = 1, latencyCycles = 5 } v3 = "st.v_add_f32"(v0, v4) { issueCycles = 1, latencyCycles = 5 } } #--- Negative: ADD has no constant → no fusion # CHECK-LABEL: @no_fusion_addmul_add_no_const # CHECK: v_add_f32 # CHECK: v_mul_f32 st.func @no_fusion_addmul_add_no_const() { ^entry: v0 = "st.v_add_f32"(v1, v2) { issueCycles = 1, latencyCycles = 5 } v3 = "st.v_mul_f32"(3.0, v0) { issueCycles = 1, latencyCycles = 5 } } #--- Negative: MUL has no constant → no fusion # CHECK-LABEL: @no_fusion_addmul_mul_no_const # CHECK: v_add_f32 # CHECK: v_mul_f32 st.func @no_fusion_addmul_mul_no_const() { ^entry: v0 = "st.v_add_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(v3, v0) { issueCycles = 1, latencyCycles = 5 } } #--- Negative constant: 3.0 * -2.0 = -6.0 # CHECK-LABEL: @add_mul_negative # CHECK: v_fma_f32 # CHECK-SAME: 3 # CHECK-SAME: v1 # CHECK-SAME: -6 # CHECK-NOT: v_add_f32 # CHECK-NOT: v_mul_f32 st.func @add_mul_negative() { ^entry: v0 = "st.v_add_f32"(-2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(3.0, v0) { issueCycles = 1, latencyCycles = 5 } } #--- Both negative: -3.0 * -2.0 = 6.0 # CHECK-LABEL: @add_mul_both_negative # CHECK: v_fma_f32 # CHECK-SAME: -3 # CHECK-SAME: v1 # CHECK-SAME: 6 # CHECK-NOT: v_add_f32 # CHECK-NOT: v_mul_f32 st.func @add_mul_both_negative() { ^entry: v0 = "st.v_add_f32"(-2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(-3.0, v0) { issueCycles = 1, latencyCycles = 5 } }