Add llvm.x86.vcvtps2ph.128 (#1613)

* Add `llvm.x86.vcvtps2ph.128`

* `cargo fmt`

* Test `_mm_cvtps_ph`

Author: opstic

example/std_example.rs

@@ -259,6 +259,9 @@ unsafe fn test_simd() {
         test_mm_cvttps_epi32();
         test_mm_cvtsi128_si64();
 
+        #[cfg(not(jit))]
+        test_mm_cvtps_ph();
+
         test_mm_extract_epi8();
         test_mm_insert_epi16();
         test_mm_shuffle_epi8();
@@ -558,6 +561,21 @@ unsafe fn test_mm_cvttps_epi32() {
     }
 }
 
+#[cfg(target_arch = "x86_64")]
+#[target_feature(enable = "f16c")]
+#[cfg(not(jit))]
+unsafe fn test_mm_cvtps_ph() {
+    const F16_ONE: i16 = 0x3c00;
+    const F16_TWO: i16 = 0x4000;
+    const F16_THREE: i16 = 0x4200;
+    const F16_FOUR: i16 = 0x4400;
+
+    let a = _mm_set_ps(1.0, 2.0, 3.0, 4.0);
+    let r = _mm_cvtps_ph::<_MM_FROUND_CUR_DIRECTION>(a);
+    let e = _mm_set_epi16(0, 0, 0, 0, F16_ONE, F16_TWO, F16_THREE, F16_FOUR);
+    assert_eq_m128i(r, e);
+}
+
 fn test_checked_mul() {
     let u: Option<u8> = u8::from_str_radix("1000", 10).ok();
     assert_eq!(u, None);

src/intrinsics/llvm_x86.rs

@@ -1313,6 +1313,35 @@ pub(super) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             ret.write_cvalue_transmute(fx, res);
         }
 
+        "llvm.x86.vcvtps2ph.128" => {
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvtps_ph
+            intrinsic_args!(fx, args => (a, _imm8); intrinsic);
+            let a = a.load_scalar(fx);
+
+            let imm8 =
+                if let Some(imm8) = crate::constant::mir_operand_get_const_val(fx, &args[1].node) {
+                    imm8
+                } else {
+                    fx.tcx
+                        .dcx()
+                        .span_fatal(span, "Index argument for `_mm_cvtps_ph` is not a constant");
+                };
+
+            let imm8 = imm8.to_u32();
+
+            codegen_inline_asm_inner(
+                fx,
+                &[InlineAsmTemplatePiece::String(format!("vcvtps2ph xmm0, xmm0, {imm8}").into())],
+                &[CInlineAsmOperand::InOut {
+                    reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm0)),
+                    _late: true,
+                    in_value: a,
+                    out_place: Some(ret),
+                }],
+                InlineAsmOptions::NOSTACK | InlineAsmOptions::PURE | InlineAsmOptions::NOMEM,
+            );
+        }
+
         _ => {
             fx.tcx
                 .dcx()