Implement 128bit shl and shr binops

src/abi.rs

@@ -270,6 +270,7 @@ impl<'a, 'tcx: 'a, B: Backend + 'a> FunctionCx<'a, 'tcx, B> {
             .module
             .declare_func_in_func(func_id, &mut self.bcx.func);
         let call_inst = self.bcx.ins().call(func_ref, args);
+        self.add_comment(call_inst, format!("easy_call {}", name));
         if output_ty.is_none() {
             return None;
         }

src/codegen_i128.rs

@@ -19,7 +19,13 @@ pub fn maybe_codegen<'a, 'tcx>(
     let rhs_val = rhs.load_scalar(fx);
 
     match bin_op {
-        BinOp::Add | BinOp::Sub | BinOp::BitAnd | BinOp::BitOr | BinOp::BitXor => return None,
+        BinOp::BitAnd | BinOp::BitOr | BinOp::BitXor => {
+            assert!(!checked);
+            return None;
+        }
+        BinOp::Add | BinOp::Sub => {
+            return None; // FIXME implement checked versions
+        }
         BinOp::Offset => unreachable!("offset should only be used on pointers, not 128bit ints"),
         BinOp::Mul => {
             let res = if checked {
@@ -99,9 +105,79 @@ pub fn maybe_codegen<'a, 'tcx>(
             return Some(res);
         }
         BinOp::Shl | BinOp::Shr => {
-            // FIXME implement it
-            let out_layout = fx.layout_of(out_ty);
-            return Some(crate::trap::trap_unreachable_ret_value(fx, out_layout, format!("unimplemented 128bit binop {:?}", bin_op)));
+            let is_overflow = if checked {
+                // rhs >= 128
+
+                // FIXME support non 128bit rhs
+                /*let (rhs_lsb, rhs_msb) = fx.bcx.ins().isplit(rhs_val);
+                let rhs_msb_gt_0 = fx.bcx.ins().icmp_imm(IntCC::NotEqual, rhs_msb, 0);
+                let rhs_lsb_ge_128 = fx.bcx.ins().icmp_imm(IntCC::SignedGreaterThan, rhs_lsb, 127);
+                let is_overflow = fx.bcx.ins().bor(rhs_msb_gt_0, rhs_lsb_ge_128);*/
+                let is_overflow = fx.bcx.ins().bconst(types::B1, false);
+
+                Some(fx.bcx.ins().bint(types::I8, is_overflow))
+            } else {
+                None
+            };
+
+            // Optimize `val >> 64`, because compiler_builtins uses it to deconstruct an 128bit
+            // integer into its lsb and msb.
+            // https://github.com/rust-lang-nursery/compiler-builtins/blob/79a6a1603d5672cbb9187ff41ff4d9b5048ac1cb/src/int/mod.rs#L217
+            if let Some(64) = resolve_value_imm(fx.bcx.func, rhs_val) {
+                let (lhs_lsb, lhs_msb) = fx.bcx.ins().isplit(lhs_val);
+                let all_zeros = fx.bcx.ins().iconst(types::I64, 0);
+                let val = match (bin_op, is_signed) {
+                    (BinOp::Shr, false) => {
+                        let val = fx.bcx.ins().iconcat(lhs_msb, all_zeros);
+                        Some(CValue::by_val(val, fx.layout_of(fx.tcx.types.u128)))
+                    }
+                    (BinOp::Shr, true) => {
+                        let sign = fx.bcx.ins().icmp_imm(IntCC::SignedLessThan, lhs_msb, 0);
+                        let all_ones = fx.bcx.ins().iconst(types::I64, u64::max_value() as i64);
+                        let all_sign_bits = fx.bcx.ins().select(sign, all_zeros, all_ones);
+
+                        let val = fx.bcx.ins().iconcat(lhs_msb, all_sign_bits);
+                        Some(CValue::by_val(val, fx.layout_of(fx.tcx.types.i128)))
+                    }
+                    (BinOp::Shl, _) => {
+                        let val = fx.bcx.ins().iconcat(all_zeros, lhs_lsb);
+                        Some(CValue::by_val(val, fx.layout_of(out_ty)))
+                    }
+                    _ => None
+                };
+                if let Some(val) = val {
+                    if let Some(is_overflow) = is_overflow {
+                        let val = val.load_scalar(fx);
+                        return Some(CValue::by_val_pair(val, is_overflow, fx.layout_of(out_ty)))
+                    } else {
+                        return Some(val);
+                    }
+                }
+            }
+
+            let truncated_rhs = clif_intcast(fx, rhs_val, types::I32, false);
+            let truncated_rhs = CValue::by_val(truncated_rhs, fx.layout_of(fx.tcx.types.u32));
+            let val = match (bin_op, is_signed) {
+                (BinOp::Shl, false) => {
+                    fx.easy_call("__ashlti3", &[lhs, truncated_rhs], fx.tcx.types.u128)
+                }
+                (BinOp::Shl, true) => {
+                    fx.easy_call("__ashlti3", &[lhs, truncated_rhs], fx.tcx.types.i128)
+                }
+                (BinOp::Shr, false) => {
+                    fx.easy_call("__lshrti3", &[lhs, truncated_rhs], fx.tcx.types.u128)
+                }
+                (BinOp::Shr, true) => {
+                    fx.easy_call("__ashrti3", &[lhs, truncated_rhs], fx.tcx.types.i128)
+                }
+                (_, _) => unreachable!(),
+            };
+            if let Some(is_overflow) = is_overflow {
+                let val = val.load_scalar(fx);
+                Some(CValue::by_val_pair(val, is_overflow, fx.layout_of(out_ty)))
+            } else {
+                Some(val)
+            }
         }
     }
 }

src/common.rs

@@ -1,5 +1,6 @@
 use rustc_target::spec::{HasTargetSpec, Target};
 
+use cranelift::codegen::ir::{Opcode, InstructionData, ValueDef};
 use cranelift_module::Module;
 
 use crate::prelude::*;
@@ -62,7 +63,7 @@ pub fn codegen_select(bcx: &mut FunctionBuilder, cond: Value, lhs: Value, rhs: V
     let rhs_ty = bcx.func.dfg.value_type(rhs);
     assert_eq!(lhs_ty, rhs_ty);
     if lhs_ty == types::I8 || lhs_ty == types::I16 {
-        // FIXME workaround for missing enocding for select.i8
+        // FIXME workaround for missing encoding for select.i8
         let lhs = bcx.ins().uextend(types::I32, lhs);
         let rhs = bcx.ins().uextend(types::I32, rhs);
         let res = bcx.ins().select(cond, lhs, rhs);
@@ -118,6 +119,49 @@ pub fn clif_intcast<'a, 'tcx: 'a>(
     }
 }
 
+fn resolve_normal_value_imm(func: &Function, val: Value) -> Option<i64> {
+    if let ValueDef::Result(inst, 0 /*param*/) = func.dfg.value_def(val) {
+        if let InstructionData::UnaryImm {
+            opcode: Opcode::Iconst,
+            imm,
+        } = func.dfg[inst] {
+            Some(imm.into())
+        } else {
+            None
+        }
+    } else {
+        None
+    }
+}
+
+fn resolve_128bit_value_imm(func: &Function, val: Value) -> Option<u128> {
+    let (lsb, msb) = if let ValueDef::Result(inst, 0 /*param*/) = func.dfg.value_def(val) {
+        if let InstructionData::Binary {
+            opcode: Opcode::Iconcat,
+            args: [lsb, msb],
+        } = func.dfg[inst] {
+            (lsb, msb)
+        } else {
+            return None;
+        }
+    } else {
+        return None;
+    };
+
+    let lsb = resolve_normal_value_imm(func, lsb)? as u64 as u128;
+    let msb = resolve_normal_value_imm(func, msb)? as u64 as u128;
+
+    Some(msb << 64 | lsb)
+}
+
+pub fn resolve_value_imm(func: &Function, val: Value) -> Option<u128> {
+    if func.dfg.value_type(val) == types::I128 {
+        resolve_128bit_value_imm(func, val)
+    } else {
+        resolve_normal_value_imm(func, val).map(|imm| imm as u64 as u128)
+    }
+}
+
 pub struct FunctionCx<'a, 'tcx: 'a, B: Backend> {
     // FIXME use a reference to `CodegenCx` instead of `tcx`, `module` and `constants` and `caches`
     pub tcx: TyCtxt<'tcx>,