# RUN: %stinkytofu-opt --arch gfx1250 %s --BuildUseDefChainPass --PeepholeOptimizationPass --print-output # # MUL+MUL Fusion: v_mul + v_mul with constants → single v_mul with multiplied constant. # Converted from PeepholeOptimizationPassTest.*MULMUL* #--- F32 basic: 2.0 * 3.0 = 6.0 # CHECK-LABEL: @mul_mul_f32_basic # CHECK: v_mul_f32 # CHECK-SAME: v2 # CHECK-SAME: 6 # CHECK-SAME: v1 # CHECK-NOT: v_mul_f32 st.func @mul_mul_f32_basic() { ^entry: v0 = "st.v_mul_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(3.0, v0) { issueCycles = 1, latencyCycles = 5 } } #--- F32 in-place: 2.0 * 3.0 = 6.0 # CHECK-LABEL: @mul_mul_f32_inplace # CHECK: v_mul_f32 # CHECK-SAME: v0 # CHECK-SAME: 6 # CHECK-SAME: v1 # CHECK-NOT: v_mul_f32 st.func @mul_mul_f32_inplace() { ^entry: v0 = "st.v_mul_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v0 = "st.v_mul_f32"(3.0, v0) { issueCycles = 1, latencyCycles = 5 } } #--- F16 basic: 2.5 * 4.0 = 10.0 # CHECK-LABEL: @mul_mul_f16_basic # CHECK: v_mul_f16 # CHECK-SAME: v2 # CHECK-SAME: 10 # CHECK-SAME: v1 # CHECK-NOT: v_mul_f16 st.func @mul_mul_f16_basic() { ^entry: v0 = "st.v_mul_f16"(2.5, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f16"(4.0, v0) { issueCycles = 1, latencyCycles = 5 } } #--- Negative: first MUL result has multiple uses → no fusion # CHECK-LABEL: @no_fusion_mul_multiple_uses # CHECK: v_mul_f32 # CHECK: v_mul_f32 # CHECK: v_mul_f32 st.func @no_fusion_mul_multiple_uses() { ^entry: v0 = "st.v_mul_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(3.0, v0) { issueCycles = 1, latencyCycles = 5 } v3 = "st.v_mul_f32"(4.0, v0) { issueCycles = 1, latencyCycles = 5 } } #--- Negative: second MUL has no constant → no fusion # CHECK-LABEL: @no_fusion_mul_missing_const # CHECK: v_mul_f32 # CHECK: v_mul_f32 st.func @no_fusion_mul_missing_const() { ^entry: v0 = "st.v_mul_f32"(2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(v3, v0) { issueCycles = 1, latencyCycles = 5 } } #--- With zero: 5.0 * 0.0 = 0.0 # CHECK-LABEL: @mul_mul_with_zero # CHECK: v_mul_f32 # CHECK-SAME: v2 # CHECK-SAME: 0 # CHECK-SAME: v1 # CHECK-NOT: v_mul_f32 st.func @mul_mul_with_zero() { ^entry: v0 = "st.v_mul_f32"(5.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(0.0, v0) { issueCycles = 1, latencyCycles = 5 } } #--- Negative values: -2.0 * 3.0 = -6.0 # CHECK-LABEL: @mul_mul_negative # CHECK: v_mul_f32 # CHECK-SAME: v2 # CHECK-SAME: -6 # CHECK-SAME: v1 # CHECK-NOT: v_mul_f32 st.func @mul_mul_negative() { ^entry: v0 = "st.v_mul_f32"(-2.0, v1) { issueCycles = 1, latencyCycles = 5 } v2 = "st.v_mul_f32"(3.0, v0) { issueCycles = 1, latencyCycles = 5 } }