A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
fix: support `fmin` and `fmax` in interpreter on vector types (#12457)
Tom Schollenberger23c0f09d0ba1
diff --git a/cranelift/filetests/filetests/runtests/fmax-vector.clif b/cranelift/filetests/filetests/runtests/fmax-vector.clifnew file mode 100644index 0000000000..b8e26dc989--- /dev/null+++ b/cranelift/filetests/filetests/runtests/fmax-vector.clif@@ -0,0 +1,51 @@+test interpret+test run+target aarch64+target s390x+target x86_64+target x86_64 has_avx+set enable_multi_ret_implicit_sret+target riscv64 has_v+target riscv64 has_v has_c has_zcb+target pulley32+target pulley32be+target pulley64+target pulley64be++function %fmax_f32x4(f32x4, f32x4) -> f32x4 {+block0(v0: f32x4, v1: f32x4):+ v2 = fmax v0, v1+ return v2+}+; run: %fmax_f32x4([0x0.0 0x1.0 0x2.0 0x3.0], [0x0.0 0x1.0 0x2.0 0x3.0]) == [0x0.0 0x1.0 0x2.0 0x3.0]+; run: %fmax_f32x4([0x1.0p10 0x1.0p11 0x1.0p12 0x1.0p13], [0x1.0p10 0x1.0p10 0x1.0p10 0x1.0p10]) == [0x1.0p10 0x1.0p11 0x1.0p12 0x1.0p13]+; run: %fmax_f32x4([0x1.0 0x2.0 0x3.0 0x4.0], [0x0.5 0x3.0 0x2.5 0x5.0]) == [0x1.0 0x3.0 0x3.0 0x5.0]+; run: %fmax_f32x4([-0x0.0 -0x0.0 -0x0.0 -0x0.0], [0x0.0 0x0.0 0x0.0 0x0.0]) == [0x0.0 0x0.0 0x0.0 0x0.0]+; run: %fmax_f32x4([+Inf +Inf +Inf +Inf], [0x0.0 0x0.0 0x0.0 0x0.0]) == [+Inf +Inf +Inf +Inf]+; run: %fmax_f32x4([-Inf -Inf -Inf -Inf], [0x0.0 0x0.0 0x0.0 0x0.0]) == [0x0.0 0x0.0 0x0.0 0x0.0]+; run: %fmax_f32x4([+Inf -Inf +Inf -Inf], [0x0.0 0x0.0 0x0.0 0x0.0]) == [+Inf 0x0.0 +Inf 0x0.0]++; run: %fmax_f32x4([0x1.000002p-23 0x1.000000p-23 0x1.000002p-23 0x1.000000p-23], [0x1.000000p-23 0x1.000002p-23 0x1.000000p-23 0x1.000002p-23]) == [0x1.000002p-23 0x1.000002p-23 0x1.000002p-23 0x1.000002p-23]+; run: %fmax_f32x4([0x1.fffffcp127 0x1.fffffep127 0x1.fffffcp127 0x1.fffffep127], [0x1.fffffep127 0x1.fffffcp127 0x1.fffffep127 0x1.fffffcp127]) == [0x1.fffffep127 0x1.fffffep127 0x1.fffffep127 0x1.fffffep127]+; run: %fmax_f32x4([0x1.000000p-126 0x1.000000p-126 0x1.000000p-126 0x1.000000p-126], [0x1.000000p-126 0x1.000000p-126 0x1.000000p-126 0x1.000000p-126]) == [0x1.000000p-126 0x1.000000p-126 0x1.000000p-126 0x1.000000p-126]++; run: %fmax_f32x4([0x0.800002p-126 0x0.800000p-126 -0x0.800002p-126 -0x0.800000p-126], [-0x0.800002p-126 -0x0.800000p-126 0x0.800002p-126 0x0.800000p-126]) == [0x0.800002p-126 0x0.800000p-126 0x0.800002p-126 0x0.800000p-126]++function %fmax_f64x2(f64x2, f64x2) -> f64x2 {+block0(v0: f64x2, v1: f64x2):+ v2 = fmax v0, v1+ return v2+}+; run: %fmax_f64x2([0x0.0 0x1.0], [0x0.0 0x1.0]) == [0x0.0 0x1.0]+; run: %fmax_f64x2([0x1.0p10 0x1.0p11], [0x1.0p10 0x1.0p10]) == [0x1.0p10 0x1.0p11]+; run: %fmax_f64x2([0x1.0 0x2.0], [0x0.5 0x3.0]) == [0x1.0 0x3.0]+; run: %fmax_f64x2([-0x0.0 -0x0.0], [0x0.0 0x0.0]) == [0x0.0 0x0.0]+; run: %fmax_f64x2([+Inf +Inf], [0x0.0 0x0.0]) == [+Inf +Inf]+; run: %fmax_f64x2([-Inf -Inf], [0x0.0 0x0.0]) == [0x0.0 0x0.0]+; run: %fmax_f64x2([+Inf -Inf], [0x0.0 0x0.0]) == [+Inf 0x0.0]++; run: %fmax_f64x2([0x1.0000000000002p-52 0x1.0000000000000p-52], [0x1.0000000000000p-52 0x1.0000000000002p-52]) == [0x1.0000000000002p-52 0x1.0000000000002p-52]+; run: %fmax_f64x2([0x1.ffffffffffffcp1023 0x1.fffffffffffffp1023], [0x1.fffffffffffffp1023 0x1.ffffffffffffcp1023]) == [0x1.fffffffffffffp1023 0x1.fffffffffffffp1023]+; run: %fmax_f64x2([0x1.0000000000000p-1022 0x1.0000000000000p-1022], [0x1.0000000000000p-1022 0x1.0000000000000p-1022]) == [0x1.0000000000000p-1022 0x1.0000000000000p-1022]++; run: %fmax_f64x2([0x0.8000000000002p-1022 0x0.8000000000000p-1022], [-0x0.8000000000002p-1022 -0x0.8000000000000p-1022]) == [0x0.8000000000002p-1022 0x0.8000000000000p-1022]diff --git a/cranelift/filetests/filetests/runtests/fmin-vector.clif b/cranelift/filetests/filetests/runtests/fmin-vector.clifnew file mode 100644index 0000000000..8a57f02f67--- /dev/null+++ b/cranelift/filetests/filetests/runtests/fmin-vector.clif@@ -0,0 +1,51 @@+test interpret+test run+target aarch64+target s390x+target x86_64+target x86_64 has_avx+set enable_multi_ret_implicit_sret+target riscv64 has_v+target riscv64 has_v has_c has_zcb+target pulley32+target pulley32be+target pulley64+target pulley64be++function %fmin_f32x4(f32x4, f32x4) -> f32x4 {+block0(v0: f32x4, v1: f32x4):+ v2 = fmin v0, v1+ return v2+}+; run: %fmin_f32x4([0x0.0 0x1.0 0x2.0 0x3.0], [0x0.0 0x1.0 0x2.0 0x3.0]) == [0x0.0 0x1.0 0x2.0 0x3.0]+; run: %fmin_f32x4([0x1.0p10 0x1.0p11 0x1.0p12 0x1.0p13], [0x1.0p10 0x1.0p10 0x1.0p10 0x1.0p10]) == [0x1.0p10 0x1.0p10 0x1.0p10 0x1.0p10]+; run: %fmin_f32x4([0x1.0 0x2.0 0x3.0 0x4.0], [0x0.5 0x3.0 0x2.5 0x5.0]) == [0x0.5 0x2.0 0x2.5 0x4.0]+; run: %fmin_f32x4([-0x0.0 -0x0.0 -0x0.0 -0x0.0], [0x0.0 0x0.0 0x0.0 0x0.0]) == [-0x0.0 -0x0.0 -0x0.0 -0x0.0]+; run: %fmin_f32x4([+Inf +Inf +Inf +Inf], [0x0.0 0x0.0 0x0.0 0x0.0]) == [0x0.0 0x0.0 0x0.0 0x0.0]+; run: %fmin_f32x4([-Inf -Inf -Inf -Inf], [0x0.0 0x0.0 0x0.0 0x0.0]) == [-Inf -Inf -Inf -Inf]+; run: %fmin_f32x4([+Inf -Inf +Inf -Inf], [0x0.0 0x0.0 0x0.0 0x0.0]) == [0x0.0 -Inf 0x0.0 -Inf]++; run: %fmin_f32x4([0x1.000002p-23 0x1.000000p-23 0x1.000002p-23 0x1.000000p-23], [0x1.000000p-23 0x1.000002p-23 0x1.000000p-23 0x1.000002p-23]) == [0x1.000000p-23 0x1.000000p-23 0x1.000000p-23 0x1.000000p-23]+; run: %fmin_f32x4([0x1.fffffcp127 0x1.fffffep127 0x1.fffffcp127 0x1.fffffep127], [0x1.fffffep127 0x1.fffffcp127 0x1.fffffep127 0x1.fffffcp127]) == [0x1.fffffcp127 0x1.fffffcp127 0x1.fffffcp127 0x1.fffffcp127]+; run: %fmin_f32x4([0x1.000000p-126 0x1.000000p-126 0x1.000000p-126 0x1.000000p-126], [0x1.000000p-126 0x1.000000p-126 0x1.000000p-126 0x1.000000p-126]) == [0x1.000000p-126 0x1.000000p-126 0x1.000000p-126 0x1.000000p-126]++; run: %fmin_f32x4([0x0.800002p-126 0x0.800000p-126 -0x0.800002p-126 -0x0.800000p-126], [-0x0.800002p-126 -0x0.800000p-126 0x0.800002p-126 0x0.800000p-126]) == [-0x0.800002p-126 -0x0.800000p-126 -0x0.800002p-126 -0x0.800000p-126]++function %fmin_f64x2(f64x2, f64x2) -> f64x2 {+block0(v0: f64x2, v1: f64x2):+ v2 = fmin v0, v1+ return v2+}+; run: %fmin_f64x2([0x0.0 0x1.0], [0x0.0 0x1.0]) == [0x0.0 0x1.0]+; run: %fmin_f64x2([0x1.0p10 0x1.0p11], [0x1.0p10 0x1.0p10]) == [0x1.0p10 0x1.0p10]+; run: %fmin_f64x2([0x1.0 0x2.0], [0x0.5 0x3.0]) == [0x0.5 0x2.0]+; run: %fmin_f64x2([-0x0.0 -0x0.0], [0x0.0 0x0.0]) == [-0x0.0 -0x0.0]+; run: %fmin_f64x2([+Inf +Inf], [0x0.0 0x0.0]) == [0x0.0 0x0.0]+; run: %fmin_f64x2([-Inf -Inf], [0x0.0 0x0.0]) == [-Inf -Inf]+; run: %fmin_f64x2([+Inf -Inf], [0x0.0 0x0.0]) == [0x0.0 -Inf]++; run: %fmin_f64x2([0x1.0000000000002p-52 0x1.0000000000000p-52], [0x1.0000000000000p-52 0x1.0000000000002p-52]) == [0x1.0000000000000p-52 0x1.0000000000000p-52]+; run: %fmin_f64x2([0x1.ffffffffffffcp1023 0x1.fffffffffffffp1023], [0x1.fffffffffffffp1023 0x1.ffffffffffffcp1023]) == [0x1.ffffffffffffcp1023 0x1.ffffffffffffcp1023]+; run: %fmin_f64x2([0x1.0000000000000p-1022 0x1.0000000000000p-1022], [0x1.0000000000000p-1022 0x1.0000000000000p-1022]) == [0x1.0000000000000p-1022 0x1.0000000000000p-1022]++; run: %fmin_f64x2([0x0.8000000000002p-1022 0x0.8000000000000p-1022], [-0x0.8000000000002p-1022 -0x0.8000000000000p-1022]) == [-0x0.8000000000002p-1022 -0x0.8000000000000p-1022]diff --git a/cranelift/interpreter/src/step.rs b/cranelift/interpreter/src/step.rsindex 8a7fad2323..7910325c74 100644--- a/cranelift/interpreter/src/step.rs+++ b/cranelift/interpreter/src/step.rs@@ -886,20 +886,60 @@ whereOpcode::Fneg => unary(DataValueExt::neg, arg(0))?,Opcode::Fabs => unary(DataValueExt::abs, arg(0))?,Opcode::Fcopysign => binary(DataValueExt::copysign, arg(0), arg(1))?,- Opcode::Fmin => assign(match (arg(0), arg(1)) {- (a, _) if a.is_nan()? => a,- (_, b) if b.is_nan()? => b,- (a, b) if a.is_zero()? && b.is_zero()? && a.is_negative()? => a,- (a, b) if a.is_zero()? && b.is_zero()? && b.is_negative()? => b,- (a, b) => a.smin(b)?,- }),- Opcode::Fmax => assign(match (arg(0), arg(1)) {- (a, _) if a.is_nan()? => a,- (_, b) if b.is_nan()? => b,- (a, b) if a.is_zero()? && b.is_zero()? && a.is_negative()? => b,- (a, b) if a.is_zero()? && b.is_zero()? && b.is_negative()? => a,- (a, b) => a.smax(b)?,- }),+ Opcode::Fmin => {+ let scalar_min = |a: DataValue, b: DataValue| -> ValueResult<DataValue> {+ Ok(match (a, b) {+ (a, _) if a.is_nan()? => a,+ (_, b) if b.is_nan()? => b,+ (a, b) if a.is_zero()? && b.is_zero()? && a.is_negative()? => a,+ (a, b) if a.is_zero()? && b.is_zero()? && b.is_negative()? => b,+ (a, b) => a.smin(b)?,+ })+ };++ if ctrl_ty.is_vector() {+ let arg0 = extractlanes(&arg(0), ctrl_ty)?;+ let arg1 = extractlanes(&arg(1), ctrl_ty)?;++ assign(vectorizelanes(+ &(arg0+ .into_iter()+ .zip(arg1.into_iter())+ .map(|(a, b)| scalar_min(a, b))+ .collect::<ValueResult<SimdVec<DataValue>>>()?),+ ctrl_ty,+ )?)+ } else {+ assign(scalar_min(arg(0), arg(1))?)+ }+ }+ Opcode::Fmax => {+ let scalar_max = |a: DataValue, b: DataValue| -> ValueResult<DataValue> {+ Ok(match (a, b) {+ (a, _) if a.is_nan()? => a,+ (_, b) if b.is_nan()? => b,+ (a, b) if a.is_zero()? && b.is_zero()? && a.is_negative()? => b,+ (a, b) if a.is_zero()? && b.is_zero()? && b.is_negative()? => a,+ (a, b) => a.smax(b)?,+ })+ };++ if ctrl_ty.is_vector() {+ let arg0 = extractlanes(&arg(0), ctrl_ty)?;+ let arg1 = extractlanes(&arg(1), ctrl_ty)?;++ assign(vectorizelanes(+ &(arg0+ .into_iter()+ .zip(arg1.into_iter())+ .map(|(a, b)| scalar_max(a, b))+ .collect::<ValueResult<SimdVec<DataValue>>>()?),+ ctrl_ty,+ )?)+ } else {+ assign(scalar_max(arg(0), arg(1))?)+ }+ }Opcode::Ceil => unary(DataValueExt::ceil, arg(0))?,Opcode::Floor => unary(DataValueExt::floor, arg(0))?,Opcode::Trunc => unary(DataValueExt::trunc, arg(0))?,