# RUN: %stinkytofu-opt --arch gfx1250 %s --BuildUseDefChainPass --PeepholeOptimizationPass --print-output # # MUL+FMA Fusion: v_mul + v_fma with constants → single v_fma with folded constant. # v_mul v0,C1,v1 + v_fma v2,C2,v0,v3 → v_fma v2,(C1*C2),v1,v3 # Converted from PeepholeOptimizationPassTest.*MULFMA* #--- F32 basic: 2.0 * 3.0 = 6.0 # CHECK-LABEL: @mul_fma_f32_basic # CHECK: v_fma_f32 # CHECK-SAME: v2 # CHECK-SAME: 6 # CHECK-SAME: v1 # CHECK-SAME: v3 # CHECK-NOT: v_mul_f32 st.func @mul_fma_f32_basic() { ^entry: v0 = "st.v_mul_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_fma_f32"(3.0, v0, v3) { issueCycles = 1, latencyCycles = 5 } } #--- F32 commutative: MUL result in first position of FMA multiply # CHECK-LABEL: @mul_fma_f32_commutative # CHECK: v_fma_f32 # CHECK-SAME: 6 # CHECK-NOT: v_mul_f32 st.func @mul_fma_f32_commutative() { ^entry: v0 = "st.v_mul_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_fma_f32"(v0, 3.0, v3) { issueCycles = 1, latencyCycles = 5 } } #--- F16 basic: 2.0 * 3.0 = 6.0 # CHECK-LABEL: @mul_fma_f16_basic # CHECK: v_fma_f16 # CHECK-SAME: 6 # CHECK-SAME: v1 # CHECK-SAME: v3 # CHECK-NOT: v_mul_f16 st.func @mul_fma_f16_basic() { ^entry: v0 = "st.v_mul_f16"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_fma_f16"(3.0, v0, v3) { issueCycles = 1, latencyCycles = 5 } } #--- Negative: MUL result has multiple uses → no fusion # CHECK-LABEL: @no_fusion_mulfma_multiple_uses # CHECK: v_mul_f32 # CHECK: v_fma_f32 # CHECK: v_add_f32 st.func @no_fusion_mulfma_multiple_uses() { ^entry: v0 = "st.v_mul_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_fma_f32"(3.0, v0, v3) { issueCycles = 1, latencyCycles = 5 } v4 = "st.v_add_f32"(v0, v5) { issueCycles = 1, latencyCycles = 5 } } #--- Negative: MUL has no constant → no fusion # CHECK-LABEL: @no_fusion_mulfma_missing_const # CHECK: v_mul_f32 # CHECK: v_fma_f32 st.func @no_fusion_mulfma_missing_const() { ^entry: v0 = "st.v_mul_f32"(v1, v2) { issueCycles = 1, latencyCycles = 5 } v3 = "st.v_fma_f32"(3.0, v0, v4) { issueCycles = 1, latencyCycles = 5 } } #--- Negative values: -2.0 * 3.0 = -6.0 # CHECK-LABEL: @mul_fma_negative # CHECK: v_fma_f32 # CHECK-SAME: -6 # CHECK-NOT: v_mul_f32 st.func @mul_fma_negative() { ^entry: v0 = "st.v_mul_f32"(-2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_fma_f32"(3.0, v0, v3) { issueCycles = 1, latencyCycles = 5 } }