Merge pull request #1431 from rust-lang/even_more_simd_intrinsics

Implement another batch of vendor intrinsics

1 files changed, 347 insertions, 21 deletions

diff --git a/src/intrinsics/llvm_x86.rs b/src/intrinsics/llvm_x86.rs
index 78a0a347cfc..07b95b7933d 100644
--- a/src/intrinsics/llvm_x86.rs
+++ b/src/intrinsics/llvm_x86.rs
@@ -22,6 +22,11 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             // Spin loop hint
         }
 
+        "llvm.x86.avx.vzeroupper" => {
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_zeroupper&ig_expand=7218
+            // Do nothing. It is a perf hint anyway.
+        }
+
         // Used by is_x86_feature_detected!();
         "llvm.x86.xgetbv" => {
             intrinsic_args!(fx, args => (xcr_no); intrinsic);
@@ -69,6 +74,103 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             ret.write_cvalue(fx, val);
         }
 
+        "llvm.x86.avx2.gather.d.d"
+        | "llvm.x86.avx2.gather.d.q"
+        | "llvm.x86.avx2.gather.d.ps"
+        | "llvm.x86.avx2.gather.d.pd"
+        | "llvm.x86.avx2.gather.d.d.256"
+        | "llvm.x86.avx2.gather.d.q.256"
+        | "llvm.x86.avx2.gather.d.ps.256"
+        | "llvm.x86.avx2.gather.d.pd.256"
+        | "llvm.x86.avx2.gather.q.d"
+        | "llvm.x86.avx2.gather.q.q"
+        | "llvm.x86.avx2.gather.q.ps"
+        | "llvm.x86.avx2.gather.q.pd"
+        | "llvm.x86.avx2.gather.q.d.256"
+        | "llvm.x86.avx2.gather.q.q.256"
+        | "llvm.x86.avx2.gather.q.ps.256"
+        | "llvm.x86.avx2.gather.q.pd.256" => {
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_i64gather_pd&ig_expand=3818
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_mask_i64gather_pd&ig_expand=3819
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_i64gather_pd&ig_expand=3821
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_mask_i64gather_pd&ig_expand=3822
+            // ...
+
+            intrinsic_args!(fx, args => (src, ptr, index, mask, scale); intrinsic);
+
+            let (src_lane_count, src_lane_ty) = src.layout().ty.simd_size_and_type(fx.tcx);
+            let (index_lane_count, index_lane_ty) = index.layout().ty.simd_size_and_type(fx.tcx);
+            let (mask_lane_count, mask_lane_ty) = mask.layout().ty.simd_size_and_type(fx.tcx);
+            let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
+            assert_eq!(src_lane_ty, ret_lane_ty);
+            assert!(index_lane_ty.is_integral());
+            assert_eq!(src_lane_count, mask_lane_count);
+            assert_eq!(src_lane_count, ret_lane_count);
+
+            let lane_clif_ty = fx.clif_type(ret_lane_ty).unwrap();
+            let index_lane_clif_ty = fx.clif_type(index_lane_ty).unwrap();
+            let mask_lane_clif_ty = fx.clif_type(mask_lane_ty).unwrap();
+            let ret_lane_layout = fx.layout_of(ret_lane_ty);
+
+            let ptr = ptr.load_scalar(fx);
+            let scale = scale.load_scalar(fx);
+            let scale = fx.bcx.ins().uextend(types::I64, scale);
+            for lane_idx in 0..std::cmp::min(src_lane_count, index_lane_count) {
+                let src_lane = src.value_lane(fx, lane_idx).load_scalar(fx);
+                let index_lane = index.value_lane(fx, lane_idx).load_scalar(fx);
+                let mask_lane = mask.value_lane(fx, lane_idx).load_scalar(fx);
+                let mask_lane =
+                    fx.bcx.ins().bitcast(mask_lane_clif_ty.as_int(), MemFlags::new(), mask_lane);
+
+                let if_enabled = fx.bcx.create_block();
+                let if_disabled = fx.bcx.create_block();
+                let next = fx.bcx.create_block();
+                let res_lane = fx.bcx.append_block_param(next, lane_clif_ty);
+
+                let mask_lane = match mask_lane_clif_ty {
+                    types::I32 | types::F32 => {
+                        fx.bcx.ins().band_imm(mask_lane, 0x8000_0000u64 as i64)
+                    }
+                    types::I64 | types::F64 => {
+                        fx.bcx.ins().band_imm(mask_lane, 0x8000_0000_0000_0000u64 as i64)
+                    }
+                    _ => unreachable!(),
+                };
+                fx.bcx.ins().brif(mask_lane, if_enabled, &[], if_disabled, &[]);
+                fx.bcx.seal_block(if_enabled);
+                fx.bcx.seal_block(if_disabled);
+
+                fx.bcx.switch_to_block(if_enabled);
+                let index_lane = if index_lane_clif_ty != types::I64 {
+                    fx.bcx.ins().sextend(types::I64, index_lane)
+                } else {
+                    index_lane
+                };
+                let offset = fx.bcx.ins().imul(index_lane, scale);
+                let lane_ptr = fx.bcx.ins().iadd(ptr, offset);
+                let res = fx.bcx.ins().load(lane_clif_ty, MemFlags::trusted(), lane_ptr, 0);
+                fx.bcx.ins().jump(next, &[res]);
+
+                fx.bcx.switch_to_block(if_disabled);
+                fx.bcx.ins().jump(next, &[src_lane]);
+
+                fx.bcx.seal_block(next);
+                fx.bcx.switch_to_block(next);
+
+                fx.bcx.ins().nop();
+
+                ret.place_lane(fx, lane_idx)
+                    .write_cvalue(fx, CValue::by_val(res_lane, ret_lane_layout));
+            }
+
+            for lane_idx in std::cmp::min(src_lane_count, index_lane_count)..ret_lane_count {
+                let zero_lane = fx.bcx.ins().iconst(mask_lane_clif_ty.as_int(), 0);
+                let zero_lane = fx.bcx.ins().bitcast(mask_lane_clif_ty, MemFlags::new(), zero_lane);
+                ret.place_lane(fx, lane_idx)
+                    .write_cvalue(fx, CValue::by_val(zero_lane, ret_lane_layout));
+            }
+        }
+
         "llvm.x86.sse.cmp.ps" | "llvm.x86.sse2.cmp.pd" => {
             let (x, y, kind) = match args {
                 [x, y, kind] => (x, y, kind),
@@ -273,16 +375,31 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             );
         }
         "llvm.x86.ssse3.pabs.b.128" | "llvm.x86.ssse3.pabs.w.128" | "llvm.x86.ssse3.pabs.d.128" => {
-            let a = match args {
-                [a] => a,
-                _ => bug!("wrong number of args for intrinsic {intrinsic}"),
-            };
-            let a = codegen_operand(fx, a);
+            intrinsic_args!(fx, args => (a); intrinsic);
 
             simd_for_each_lane(fx, a, ret, &|fx, _lane_ty, _res_lane_ty, lane| {
                 fx.bcx.ins().iabs(lane)
             });
         }
+        "llvm.x86.sse2.cvttps2dq" => {
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvttps_epi32&ig_expand=2429
+            intrinsic_args!(fx, args => (a); intrinsic);
+            let a = a.load_scalar(fx);
+
+            // Using inline asm instead of fcvt_to_sint_sat as unrepresentable values are turned
+            // into 0x80000000 for which Cranelift doesn't have a native instruction.
+            codegen_inline_asm_inner(
+                fx,
+                &[InlineAsmTemplatePiece::String(format!("cvttps2dq xmm0, xmm0"))],
+                &[CInlineAsmOperand::InOut {
+                    reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm0)),
+                    _late: true,
+                    in_value: a,
+                    out_place: Some(ret),
+                }],
+                InlineAsmOptions::NOSTACK | InlineAsmOptions::PURE | InlineAsmOptions::NOMEM,
+            );
+        }
         "llvm.x86.addcarry.32" | "llvm.x86.addcarry.64" => {
             intrinsic_args!(fx, args => (c_in, a, b); intrinsic);
             let c_in = c_in.load_scalar(fx);
@@ -439,12 +556,12 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             let ret_lane_layout = fx.layout_of(fx.tcx.types.i32);
             for out_lane_idx in 0..lane_count / 2 {
                 let a_lane0 = a.value_lane(fx, out_lane_idx * 2).load_scalar(fx);
-                let a_lane0 = fx.bcx.ins().uextend(types::I32, a_lane0);
+                let a_lane0 = fx.bcx.ins().sextend(types::I32, a_lane0);
                 let b_lane0 = b.value_lane(fx, out_lane_idx * 2).load_scalar(fx);
                 let b_lane0 = fx.bcx.ins().sextend(types::I32, b_lane0);
 
                 let a_lane1 = a.value_lane(fx, out_lane_idx * 2 + 1).load_scalar(fx);
-                let a_lane1 = fx.bcx.ins().uextend(types::I32, a_lane1);
+                let a_lane1 = fx.bcx.ins().sextend(types::I32, a_lane1);
                 let b_lane1 = b.value_lane(fx, out_lane_idx * 2 + 1).load_scalar(fx);
                 let b_lane1 = fx.bcx.ins().sextend(types::I32, b_lane1);
 
@@ -599,14 +716,14 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             assert_eq!(ret_lane_ty, fx.tcx.types.i16);
             assert_eq!(lane_count * 2, ret_lane_count);
 
-            let min_i16 = fx.bcx.ins().iconst(types::I32, i64::from(i16::MIN as u16));
-            let max_i16 = fx.bcx.ins().iconst(types::I32, i64::from(i16::MAX as u16));
+            let min_i16 = fx.bcx.ins().iconst(types::I32, i32::from(i16::MIN) as u32 as i64);
+            let max_i16 = fx.bcx.ins().iconst(types::I32, i32::from(i16::MAX) as u32 as i64);
             let ret_lane_layout = fx.layout_of(fx.tcx.types.i16);
 
             for idx in 0..lane_count {
                 let lane = a.value_lane(fx, idx).load_scalar(fx);
                 let sat = fx.bcx.ins().smax(lane, min_i16);
-                let sat = fx.bcx.ins().umin(sat, max_i16);
+                let sat = fx.bcx.ins().smin(sat, max_i16);
                 let res = fx.bcx.ins().ireduce(types::I16, sat);
 
                 let res_lane = CValue::by_val(res, ret_lane_layout);
@@ -616,7 +733,7 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             for idx in 0..lane_count {
                 let lane = b.value_lane(fx, idx).load_scalar(fx);
                 let sat = fx.bcx.ins().smax(lane, min_i16);
-                let sat = fx.bcx.ins().umin(sat, max_i16);
+                let sat = fx.bcx.ins().smin(sat, max_i16);
                 let res = fx.bcx.ins().ireduce(types::I16, sat);
 
                 let res_lane = CValue::by_val(res, ret_lane_layout);
@@ -643,8 +760,8 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
 
             for idx in 0..lane_count {
                 let lane = a.value_lane(fx, idx).load_scalar(fx);
-                let sat = fx.bcx.ins().umax(lane, min_u16);
-                let sat = fx.bcx.ins().umin(sat, max_u16);
+                let sat = fx.bcx.ins().smax(lane, min_u16);
+                let sat = fx.bcx.ins().smin(sat, max_u16);
                 let res = fx.bcx.ins().ireduce(types::I16, sat);
 
                 let res_lane = CValue::by_val(res, ret_lane_layout);
@@ -653,8 +770,8 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
 
             for idx in 0..lane_count {
                 let lane = b.value_lane(fx, idx).load_scalar(fx);
-                let sat = fx.bcx.ins().umax(lane, min_u16);
-                let sat = fx.bcx.ins().umin(sat, max_u16);
+                let sat = fx.bcx.ins().smax(lane, min_u16);
+                let sat = fx.bcx.ins().smin(sat, max_u16);
                 let res = fx.bcx.ins().ireduce(types::I16, sat);
 
                 let res_lane = CValue::by_val(res, ret_lane_layout);
@@ -675,14 +792,14 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             assert_eq!(ret_lane_ty, fx.tcx.types.i16);
             assert_eq!(lane_count * 2, ret_lane_count);
 
-            let min_i16 = fx.bcx.ins().iconst(types::I32, i64::from(i16::MIN as u16));
-            let max_i16 = fx.bcx.ins().iconst(types::I32, i64::from(i16::MAX as u16));
+            let min_i16 = fx.bcx.ins().iconst(types::I32, i32::from(i16::MIN) as u32 as i64);
+            let max_i16 = fx.bcx.ins().iconst(types::I32, i32::from(i16::MAX) as u32 as i64);
             let ret_lane_layout = fx.layout_of(fx.tcx.types.i16);
 
             for idx in 0..lane_count / 2 {
                 let lane = a.value_lane(fx, idx).load_scalar(fx);
                 let sat = fx.bcx.ins().smax(lane, min_i16);
-                let sat = fx.bcx.ins().umin(sat, max_i16);
+                let sat = fx.bcx.ins().smin(sat, max_i16);
                 let res = fx.bcx.ins().ireduce(types::I16, sat);
 
                 let res_lane = CValue::by_val(res, ret_lane_layout);
@@ -692,7 +809,7 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             for idx in 0..lane_count / 2 {
                 let lane = b.value_lane(fx, idx).load_scalar(fx);
                 let sat = fx.bcx.ins().smax(lane, min_i16);
-                let sat = fx.bcx.ins().umin(sat, max_i16);
+                let sat = fx.bcx.ins().smin(sat, max_i16);
                 let res = fx.bcx.ins().ireduce(types::I16, sat);
 
                 let res_lane = CValue::by_val(res, ret_lane_layout);
@@ -702,7 +819,7 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             for idx in 0..lane_count / 2 {
                 let lane = a.value_lane(fx, idx).load_scalar(fx);
                 let sat = fx.bcx.ins().smax(lane, min_i16);
-                let sat = fx.bcx.ins().umin(sat, max_i16);
+                let sat = fx.bcx.ins().smin(sat, max_i16);
                 let res = fx.bcx.ins().ireduce(types::I16, sat);
 
                 let res_lane = CValue::by_val(res, ret_lane_layout);
@@ -712,7 +829,7 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             for idx in 0..lane_count / 2 {
                 let lane = b.value_lane(fx, idx).load_scalar(fx);
                 let sat = fx.bcx.ins().smax(lane, min_i16);
-                let sat = fx.bcx.ins().umin(sat, max_i16);
+                let sat = fx.bcx.ins().smin(sat, max_i16);
                 let res = fx.bcx.ins().ireduce(types::I16, sat);
 
                 let res_lane = CValue::by_val(res, ret_lane_layout);
@@ -720,6 +837,215 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
             }
         }
 
+        "llvm.x86.fma.vfmaddsub.ps"
+        | "llvm.x86.fma.vfmaddsub.pd"
+        | "llvm.x86.fma.vfmaddsub.ps.256"
+        | "llvm.x86.fma.vfmaddsub.pd.256" => {
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fmaddsub_ps&ig_expand=3205
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fmaddsub_pd&ig_expand=3181
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fmaddsub_ps&ig_expand=3209
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fmaddsub_pd&ig_expand=3185
+            intrinsic_args!(fx, args => (a, b, c); intrinsic);
+
+            assert_eq!(a.layout(), b.layout());
+            assert_eq!(a.layout(), c.layout());
+            let layout = a.layout();
+
+            let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
+            let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
+            assert!(lane_ty.is_floating_point());
+            assert!(ret_lane_ty.is_floating_point());
+            assert_eq!(lane_count, ret_lane_count);
+            let ret_lane_layout = fx.layout_of(ret_lane_ty);
+
+            for idx in 0..lane_count {
+                let a_lane = a.value_lane(fx, idx).load_scalar(fx);
+                let b_lane = b.value_lane(fx, idx).load_scalar(fx);
+                let c_lane = c.value_lane(fx, idx).load_scalar(fx);
+
+                let mul = fx.bcx.ins().fmul(a_lane, b_lane);
+                let res = if idx & 1 == 0 {
+                    fx.bcx.ins().fsub(mul, c_lane)
+                } else {
+                    fx.bcx.ins().fadd(mul, c_lane)
+                };
+
+                let res_lane = CValue::by_val(res, ret_lane_layout);
+                ret.place_lane(fx, idx).write_cvalue(fx, res_lane);
+            }
+        }
+
+        "llvm.x86.fma.vfmsubadd.ps"
+        | "llvm.x86.fma.vfmsubadd.pd"
+        | "llvm.x86.fma.vfmsubadd.ps.256"
+        | "llvm.x86.fma.vfmsubadd.pd.256" => {
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fmsubadd_ps&ig_expand=3325
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fmsubadd_pd&ig_expand=3301
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fmsubadd_ps&ig_expand=3329
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fmsubadd_pd&ig_expand=3305
+            intrinsic_args!(fx, args => (a, b, c); intrinsic);
+
+            assert_eq!(a.layout(), b.layout());
+            assert_eq!(a.layout(), c.layout());
+            let layout = a.layout();
+
+            let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
+            let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
+            assert!(lane_ty.is_floating_point());
+            assert!(ret_lane_ty.is_floating_point());
+            assert_eq!(lane_count, ret_lane_count);
+            let ret_lane_layout = fx.layout_of(ret_lane_ty);
+
+            for idx in 0..lane_count {
+                let a_lane = a.value_lane(fx, idx).load_scalar(fx);
+                let b_lane = b.value_lane(fx, idx).load_scalar(fx);
+                let c_lane = c.value_lane(fx, idx).load_scalar(fx);
+
+                let mul = fx.bcx.ins().fmul(a_lane, b_lane);
+                let res = if idx & 1 == 0 {
+                    fx.bcx.ins().fadd(mul, c_lane)
+                } else {
+                    fx.bcx.ins().fsub(mul, c_lane)
+                };
+
+                let res_lane = CValue::by_val(res, ret_lane_layout);
+                ret.place_lane(fx, idx).write_cvalue(fx, res_lane);
+            }
+        }
+
+        "llvm.x86.fma.vfnmadd.ps"
+        | "llvm.x86.fma.vfnmadd.pd"
+        | "llvm.x86.fma.vfnmadd.ps.256"
+        | "llvm.x86.fma.vfnmadd.pd.256" => {
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fnmadd_ps&ig_expand=3391
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fnmadd_pd&ig_expand=3367
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fnmadd_ps&ig_expand=3395
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fnmadd_pd&ig_expand=3371
+            intrinsic_args!(fx, args => (a, b, c); intrinsic);
+
+            assert_eq!(a.layout(), b.layout());
+            assert_eq!(a.layout(), c.layout());
+            let layout = a.layout();
+
+            let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
+            let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
+            assert!(lane_ty.is_floating_point());
+            assert!(ret_lane_ty.is_floating_point());
+            assert_eq!(lane_count, ret_lane_count);
+            let ret_lane_layout = fx.layout_of(ret_lane_ty);
+
+            for idx in 0..lane_count {
+                let a_lane = a.value_lane(fx, idx).load_scalar(fx);
+                let b_lane = b.value_lane(fx, idx).load_scalar(fx);
+                let c_lane = c.value_lane(fx, idx).load_scalar(fx);
+
+                let mul = fx.bcx.ins().fmul(a_lane, b_lane);
+                let neg_mul = fx.bcx.ins().fneg(mul);
+                let res = fx.bcx.ins().fadd(neg_mul, c_lane);
+
+                let res_lane = CValue::by_val(res, ret_lane_layout);
+                ret.place_lane(fx, idx).write_cvalue(fx, res_lane);
+            }
+        }
+
+        "llvm.x86.sse42.pcmpestri128" => {
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cmpestri&ig_expand=939
+            intrinsic_args!(fx, args => (a, la, b, lb, _imm8); intrinsic);
+
+            let a = a.load_scalar(fx);
+            let la = la.load_scalar(fx);
+            let b = b.load_scalar(fx);
+            let lb = lb.load_scalar(fx);
+
+            let imm8 = if let Some(imm8) = crate::constant::mir_operand_get_const_val(fx, &args[4])
+            {
+                imm8
+            } else {
+                fx.tcx.sess.span_fatal(span, "Index argument for `_mm_cmpestri` is not a constant");
+            };
+
+            let imm8 = imm8.try_to_u8().unwrap_or_else(|_| panic!("kind not scalar: {:?}", imm8));
+
+            codegen_inline_asm_inner(
+                fx,
+                &[InlineAsmTemplatePiece::String(format!("pcmpestri xmm0, xmm1, {imm8}"))],
+                &[
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm0)),
+                        value: a,
+                    },
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm1)),
+                        value: b,
+                    },
+                    // Implicit argument to the pcmpestri intrinsic
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::ax)),
+                        value: la,
+                    },
+                    // Implicit argument to the pcmpestri intrinsic
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::dx)),
+                        value: lb,
+                    },
+                    // Implicit result of the pcmpestri intrinsic
+                    CInlineAsmOperand::Out {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::cx)),
+                        late: true,
+                        place: Some(ret),
+                    },
+                ],
+                InlineAsmOptions::NOSTACK | InlineAsmOptions::PURE | InlineAsmOptions::NOMEM,
+            );
+        }
+
+        "llvm.x86.sse42.pcmpestrm128" => {
+            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cmpestrm&ig_expand=940
+            intrinsic_args!(fx, args => (a, la, b, lb, _imm8); intrinsic);
+
+            let a = a.load_scalar(fx);
+            let la = la.load_scalar(fx);
+            let b = b.load_scalar(fx);
+            let lb = lb.load_scalar(fx);
+
+            let imm8 = if let Some(imm8) = crate::constant::mir_operand_get_const_val(fx, &args[4])
+            {
+                imm8
+            } else {
+                fx.tcx.sess.span_fatal(span, "Index argument for `_mm_cmpestrm` is not a constant");
+            };
+
+            let imm8 = imm8.try_to_u8().unwrap_or_else(|_| panic!("kind not scalar: {:?}", imm8));
+
+            codegen_inline_asm_inner(
+                fx,
+                &[InlineAsmTemplatePiece::String(format!("pcmpestrm xmm0, xmm1, {imm8}"))],
+                &[
+                    CInlineAsmOperand::InOut {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm0)),
+                        _late: true,
+                        in_value: a,
+                        out_place: Some(ret),
+                    },
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm1)),
+                        value: b,
+                    },
+                    // Implicit argument to the pcmpestri intrinsic
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::ax)),
+                        value: la,
+                    },
+                    // Implicit argument to the pcmpestri intrinsic
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::dx)),
+                        value: lb,
+                    },
+                ],
+                InlineAsmOptions::NOSTACK | InlineAsmOptions::PURE | InlineAsmOptions::NOMEM,
+            );
+        }
+
         "llvm.x86.pclmulqdq" => {
             // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_clmulepi64_si128&ig_expand=772
             intrinsic_args!(fx, args => (a, b, _imm8); intrinsic);