Merge pull request #3115 from bjorn3/fminmax_pseudo_scalar

Implement fmin_pseudo and fmax_pseudo for scalars

Author: cfallin

cranelift/codegen/src/isa/x64/lower.rs

@@ -4413,6 +4413,10 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
             let ty = ty.unwrap();
             ctx.emit(Inst::gen_move(dst, rhs, ty));
             let sse_opcode = match (ty, op) {
+                (types::F32, Opcode::FminPseudo) => SseOpcode::Minss,
+                (types::F32, Opcode::FmaxPseudo) => SseOpcode::Maxss,
+                (types::F64, Opcode::FminPseudo) => SseOpcode::Minsd,
+                (types::F64, Opcode::FmaxPseudo) => SseOpcode::Maxsd,
                 (types::F32X4, Opcode::FminPseudo) => SseOpcode::Minps,
                 (types::F32X4, Opcode::FmaxPseudo) => SseOpcode::Maxps,
                 (types::F64X2, Opcode::FminPseudo) => SseOpcode::Minpd,

cranelift/codegen/src/preopt.serialized

@@ -0,0 +1,36 @@
+test run
+; target s390x TODO: Not yet implemented on s390x
+set enable_simd
+target aarch64
+target x86_64 machinst skylake
+
+function %fmin_pseudo_f32x4(f32x4, f32x4) -> f32x4 {
+block0(v0:f32x4, v1:f32x4):
+    v2 = fmin_pseudo v0, v1
+    return v2
+}
+; run: %fmin_pseudo_f32x4([0x1.0 NaN 0x0.1 -0x0.0], [0x2.0 0x2.0 NaN 0x0.0]) == [0x1.0 NaN 0x0.1 -0x0.0]
+
+function %fmax_pseudo_f32x4(f32x4, f32x4) -> f32x4 {
+block0(v0:f32x4, v1:f32x4):
+    v2 = fmax_pseudo v0, v1
+    return v2
+}
+; run: %fmax_pseudo_f32x4([0x1.0 NaN 0x0.1 -0x0.0], [0x2.0 0x2.0 NaN 0x0.0]) == [0x2.0 NaN 0x0.1 -0x0.0]
+
+function %fmin_pseudo_f64x2(f64x2, f64x2) -> f64x2 {
+block0(v0:f64x2, v1:f64x2):
+    v2 = fmin_pseudo v0, v1
+    return v2
+}
+; run: %fmin_pseudo_f64x2([0x1.0 NaN], [0x2.0 0x2.0]) == [0x1.0 NaN]
+; run: %fmin_pseudo_f64x2([0x0.1 -0x0.0], [NaN 0x0.0]) == [0x0.1 -0x0.0]
+
+function %fmax_pseudo_f64x2(f64x2, f64x2) -> f64x2 {
+block0(v0:f64x2, v1:f64x2):
+    v2 = fmax_pseudo v0, v1
+    return v2
+}
+; run: %fmax_pseudo_f64x2([0x1.0 NaN], [0x2.0 0x2.0]) == [0x2.0 NaN]
+; run: %fmax_pseudo_f64x2([0x0.1 -0x0.0], [NaN 0x0.0]) == [0x0.1 -0x0.0]
+

cranelift/filetests/filetests/runtests/fmin-max-pseudo-vector.clif

@@ -0,0 +1,50 @@
+test run
+; target s390x TODO: Not yet implemented on s390x
+; target aarch64 TODO: Not yet implemented on aarch64
+set enable_simd
+target x86_64 machinst skylake
+
+function %fmin_pseudo_f32(f32, f32) -> f32 {
+block0(v0:f32, v1:f32):
+    v2 = fmin_pseudo v0, v1
+    return v2
+}
+; run: %fmin_pseudo_f32(0x1.0, 0x2.0) == 0x1.0
+; run: %fmin_pseudo_f32(NaN, 0x2.0) == NaN
+; run: %fmin_pseudo_f32(0x0.1, NaN) == 0x0.1
+; run: %fmin_pseudo_f32(0x0.0, -0x0.0) == 0x0.0
+; run: %fmin_pseudo_f32(-0x0.0, 0x0.0) == -0x0.0
+
+function %fmax_pseudo_f32(f32, f32) -> f32 {
+block0(v0:f32, v1:f32):
+    v2 = fmax_pseudo v0, v1
+    return v2
+}
+; run: %fmax_pseudo_f32(0x1.0, 0x2.0) == 0x2.0
+; run: %fmax_pseudo_f32(NaN, 0x2.0) == NaN
+; run: %fmax_pseudo_f32(0x0.1, NaN) == 0x0.1
+; run: %fmax_pseudo_f32(0x0.0, 0x0.0) == 0x0.0
+; run: %fmax_pseudo_f32(-0x0.0, 0x0.0) == -0x0.0
+
+function %fmin_pseudo_f64(f64, f64) -> f64 {
+block0(v0:f64, v1:f64):
+    v2 = fmin_pseudo v0, v1
+    return v2
+}
+; run: %fmin_pseudo_f64(0x1.0, 0x2.0) == 0x1.0
+; run: %fmin_pseudo_f64(NaN, 0x2.0) == NaN
+; run: %fmin_pseudo_f64(0x0.1, NaN) == 0x0.1
+; run: %fmin_pseudo_f64(0x0.0, -0x0.0) == 0x0.0
+; run: %fmin_pseudo_f64(-0x0.0, 0x0.0) == -0x0.0
+
+function %fmax_pseudo_f64(f64, f64) -> f64 {
+block0(v0:f64, v1:f64):
+    v2 = fmax_pseudo v0, v1
+    return v2
+}
+; run: %fmax_pseudo_f64(0x1.0, 0x2.0) == 0x2.0
+; run: %fmax_pseudo_f64(NaN, 0x2.0) == NaN
+; run: %fmax_pseudo_f64(0x0.1, NaN) == 0x0.1
+; run: %fmax_pseudo_f64(0x0.0, 0x0.0) == 0x0.0
+; run: %fmax_pseudo_f64(-0x0.0, 0x0.0) == -0x0.0
+