Auto merge of #58003 - nikic:saturating-add, r=nagisa

Use LLVM intrinsics for saturating add/sub

Use the `[su](add|sub).sat` LLVM intrinsics, if we're compiling against LLVM 8, as they should optimize and codegen better than IR based on `[su](add|sub).with.overlow`.

For the fallback for LLVM < 8 I'm using the same expansion that target lowering in LLVM uses, which is not the same as Rust currently uses (in particular due to the use of selects rather than branches).

Fixes #55286.
Fixes #52203.
Fixes #44500.

r? @nagisa

2 files changed, 65 insertions, 3 deletions

diff --git a/src/librustc_codegen_llvm/context.rs b/src/librustc_codegen_llvm/context.rs
index 1d7f14b02e1..f6795588441 100644
--- a/src/librustc_codegen_llvm/context.rs
+++ b/src/librustc_codegen_llvm/context.rs
@@ -757,6 +757,30 @@ impl CodegenCx<'b, 'tcx> {
         ifn!("llvm.umul.with.overflow.i64", fn(t_i64, t_i64) -> mk_struct!{t_i64, i1});
         ifn!("llvm.umul.with.overflow.i128", fn(t_i128, t_i128) -> mk_struct!{t_i128, i1});
 
+        ifn!("llvm.sadd.sat.i8", fn(t_i8, t_i8) -> t_i8);
+        ifn!("llvm.sadd.sat.i16", fn(t_i16, t_i16) -> t_i16);
+        ifn!("llvm.sadd.sat.i32", fn(t_i32, t_i32) -> t_i32);
+        ifn!("llvm.sadd.sat.i64", fn(t_i64, t_i64) -> t_i64);
+        ifn!("llvm.sadd.sat.i128", fn(t_i128, t_i128) -> t_i128);
+
+        ifn!("llvm.uadd.sat.i8", fn(t_i8, t_i8) -> t_i8);
+        ifn!("llvm.uadd.sat.i16", fn(t_i16, t_i16) -> t_i16);
+        ifn!("llvm.uadd.sat.i32", fn(t_i32, t_i32) -> t_i32);
+        ifn!("llvm.uadd.sat.i64", fn(t_i64, t_i64) -> t_i64);
+        ifn!("llvm.uadd.sat.i128", fn(t_i128, t_i128) -> t_i128);
+
+        ifn!("llvm.ssub.sat.i8", fn(t_i8, t_i8) -> t_i8);
+        ifn!("llvm.ssub.sat.i16", fn(t_i16, t_i16) -> t_i16);
+        ifn!("llvm.ssub.sat.i32", fn(t_i32, t_i32) -> t_i32);
+        ifn!("llvm.ssub.sat.i64", fn(t_i64, t_i64) -> t_i64);
+        ifn!("llvm.ssub.sat.i128", fn(t_i128, t_i128) -> t_i128);
+
+        ifn!("llvm.usub.sat.i8", fn(t_i8, t_i8) -> t_i8);
+        ifn!("llvm.usub.sat.i16", fn(t_i16, t_i16) -> t_i16);
+        ifn!("llvm.usub.sat.i32", fn(t_i32, t_i32) -> t_i32);
+        ifn!("llvm.usub.sat.i64", fn(t_i64, t_i64) -> t_i64);
+        ifn!("llvm.usub.sat.i128", fn(t_i128, t_i128) -> t_i128);
+
         ifn!("llvm.lifetime.start", fn(t_i64,i8p) -> void);
         ifn!("llvm.lifetime.end", fn(t_i64, i8p) -> void);
 
diff --git a/src/librustc_codegen_llvm/intrinsic.rs b/src/librustc_codegen_llvm/intrinsic.rs
index 201b1684fb9..58b466dbe6f 100644
--- a/src/librustc_codegen_llvm/intrinsic.rs
+++ b/src/librustc_codegen_llvm/intrinsic.rs
@@ -14,7 +14,7 @@ use type_::Type;
 use type_of::LayoutLlvmExt;
 use rustc::ty::{self, Ty};
 use rustc::ty::layout::{self, LayoutOf, HasTyCtxt, Primitive};
-use rustc_codegen_ssa::common::TypeKind;
+use rustc_codegen_ssa::common::{IntPredicate, TypeKind};
 use rustc::hir;
 use syntax::ast::{self, FloatTy};
 use syntax::symbol::Symbol;
@@ -28,7 +28,7 @@ use rustc::session::Session;
 use syntax_pos::Span;
 
 use std::cmp::Ordering;
-use std::iter;
+use std::{iter, i128, u128};
 
 fn get_simple_intrinsic(cx: &CodegenCx<'ll, '_>, name: &str) -> Option<&'ll Value> {
     let llvm_name = match name {
@@ -342,7 +342,7 @@ impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
             "bitreverse" | "add_with_overflow" | "sub_with_overflow" |
             "mul_with_overflow" | "overflowing_add" | "overflowing_sub" | "overflowing_mul" |
             "unchecked_div" | "unchecked_rem" | "unchecked_shl" | "unchecked_shr" | "exact_div" |
-            "rotate_left" | "rotate_right" => {
+            "rotate_left" | "rotate_right" | "saturating_add" | "saturating_sub" => {
                 let ty = arg_tys[0];
                 match int_type_width_signed(ty, self) {
                     Some((width, signed)) =>
@@ -468,6 +468,44 @@ impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
                                     self.or(shift1, shift2)
                                 }
                             },
+                            "saturating_add" | "saturating_sub" => {
+                                let is_add = name == "saturating_add";
+                                let lhs = args[0].immediate();
+                                let rhs = args[1].immediate();
+                                if llvm_util::get_major_version() >= 8 {
+                                    let llvm_name = &format!("llvm.{}{}.sat.i{}",
+                                                             if signed { 's' } else { 'u' },
+                                                             if is_add { "add" } else { "sub" },
+                                                             width);
+                                    let llfn = self.get_intrinsic(llvm_name);
+                                    self.call(llfn, &[lhs, rhs], None)
+                                } else {
+                                    let llvm_name = &format!("llvm.{}{}.with.overflow.i{}",
+                                                             if signed { 's' } else { 'u' },
+                                                             if is_add { "add" } else { "sub" },
+                                                             width);
+                                    let llfn = self.get_intrinsic(llvm_name);
+                                    let pair = self.call(llfn, &[lhs, rhs], None);
+                                    let val = self.extract_value(pair, 0);
+                                    let overflow = self.extract_value(pair, 1);
+                                    let llty = self.type_ix(width);
+
+                                    let limit = if signed {
+                                        let limit_lo = self.const_uint_big(
+                                            llty, (i128::MIN >> (128 - width)) as u128);
+                                        let limit_hi = self.const_uint_big(
+                                            llty, (i128::MAX >> (128 - width)) as u128);
+                                        let neg = self.icmp(
+                                            IntPredicate::IntSLT, val, self.const_uint(llty, 0));
+                                        self.select(neg, limit_hi, limit_lo)
+                                    } else if is_add {
+                                        self.const_uint_big(llty, u128::MAX >> (128 - width))
+                                    } else {
+                                        self.const_uint(llty, 0)
+                                    };
+                                    self.select(overflow, limit, val)
+                                }
+                            },
                             _ => bug!(),
                         },
                     None => {