Auto merge of #28136 - huonw:simd, r=alexcrichton

I believe everything that doesn't take a constant integer up to SSE4.2
should now be correct (I don't have any reason to believe that those
that do take constant integers are wrong; they're just more complicated
and I just haven't tested them in detail).

(Also, takes out two unused code paths from trans.)

Author: bors

src/etc/platform-intrinsics/x86/sse2.json

@@ -19,8 +19,8 @@
             "intrinsic": "_madd_epi16",
             "width": [128],
             "llvm": "pmadd.wd",
-            "ret": "s16",
-            "args": ["0", "0"]
+            "ret": "s32",
+            "args": ["s16", "s16"]
         },
         {
             "intrinsic": "_max_{0.data_type}",
@@ -68,11 +68,11 @@
             "intrinsic": "_mul_epu32",
             "width": [128],
             "llvm": "pmulu.dq",
-            "ret": "s64",
+            "ret": "u64",
             "args": ["0dn", "0dn"]
         },
         {
-            "intrinsic": "_mulhi_ep{0.kind}16",
+            "intrinsic": "_mulhi_{0.data_type}",
             "width": [128],
             "llvm": "pmulh{0.kind_short}.w",
             "ret": "i16",

src/etc/platform-intrinsics/x86/sse41.json

@@ -39,7 +39,7 @@
         {
             "intrinsic": "_mul_epi32",
             "width": [128],
-            "llvm": "muldq",
+            "llvm": "pmuldq",
             "ret": "s64",
             "args": ["s32", "s32"]
         },
@@ -58,9 +58,9 @@
             "args": ["u64", "u64"]
         },
         {
-            "intrinsic": "_testncz_si128",
+            "intrinsic": "_testnzc_si128",
             "width": [128],
-            "llvm": "ptest.nzc",
+            "llvm": "ptestnzc",
             "ret": "S32",
             "args": ["u64", "u64"]
         },

src/etc/platform-intrinsics/x86/ssse3.json

@@ -4,7 +4,7 @@
         {
             "intrinsic": "_abs_{0.data_type}",
             "width": [128],
-            "llvm": "pabs.{0.data_type_short}",
+            "llvm": "pabs.{0.data_type_short}.128",
             "ret": "s(8-32)",
             "args": ["0"]
         },
@@ -41,7 +41,7 @@
             "width": [128],
             "llvm": "pmadd.ub.sw.128",
             "ret": "s16",
-            "args": ["s8", "s8"]
+            "args": ["u8", "s8"]
         },
         {
             "intrinsic": "_mulhrs_epi16",
@@ -61,7 +61,7 @@
             "intrinsic": "_sign_{0.data_type}",
             "width": [128],
             "llvm": "psign.{0.data_type_short}.128",
-            "ret": "s(8-16)",
+            "ret": "s(8-32)",
             "args": ["0", "0"]
         }
     ]

src/librustc_platform_intrinsics/x86.rs

@@ -82,7 +82,7 @@ pub fn find<'tcx>(_tcx: &ty::ctxt<'tcx>, name: &str) -> Option<Intrinsic> {
         },
         "_madd_epi16" => Intrinsic {
             inputs: vec![v(i(16), 8), v(i(16), 8)],
-            output: v(i(16), 8),
+            output: v(i(32), 4),
             definition: Named("llvm.x86.sse2.pmadd.wd")
         },
         "_max_epi16" => Intrinsic {
@@ -126,11 +126,11 @@ pub fn find<'tcx>(_tcx: &ty::ctxt<'tcx>, name: &str) -> Option<Intrinsic> {
             definition: Named("llvm.x86.sse2.pmovmskb.128")
         },
         "_mul_epu32" => Intrinsic {
-            inputs: vec![v(i(32), 4), v(i(32), 4)],
-            output: v(i(64), 2),
+            inputs: vec![v(u(32), 4), v(u(32), 4)],
+            output: v(u(64), 2),
             definition: Named("llvm.x86.sse2.pmulu.dq")
         },
-        "_mulhi_eps16" => Intrinsic {
+        "_mulhi_epi16" => Intrinsic {
             inputs: vec![v(i(16), 8), v(i(16), 8)],
             output: v(i(16), 8),
             definition: Named("llvm.x86.sse2.pmulh.w")
@@ -218,17 +218,17 @@ pub fn find<'tcx>(_tcx: &ty::ctxt<'tcx>, name: &str) -> Option<Intrinsic> {
         "_abs_epi8" => Intrinsic {
             inputs: vec![v(i(8), 16)],
             output: v(i(8), 16),
-            definition: Named("llvm.x86.ssse3.pabs.b")
+            definition: Named("llvm.x86.ssse3.pabs.b.128")
         },
         "_abs_epi16" => Intrinsic {
             inputs: vec![v(i(16), 8)],
             output: v(i(16), 8),
-            definition: Named("llvm.x86.ssse3.pabs.w")
+            definition: Named("llvm.x86.ssse3.pabs.w.128")
         },
         "_abs_epi32" => Intrinsic {
             inputs: vec![v(i(32), 4)],
             output: v(i(32), 4),
-            definition: Named("llvm.x86.ssse3.pabs.d")
+            definition: Named("llvm.x86.ssse3.pabs.d.128")
         },
         "_hadd_epi16" => Intrinsic {
             inputs: vec![v(i(16), 8), v(i(16), 8)],
@@ -261,7 +261,7 @@ pub fn find<'tcx>(_tcx: &ty::ctxt<'tcx>, name: &str) -> Option<Intrinsic> {
             definition: Named("llvm.x86.ssse3.phsub.sw.128")
         },
         "_maddubs_epi16" => Intrinsic {
-            inputs: vec![v(i(8), 16), v(i(8), 16)],
+            inputs: vec![v(u(8), 16), v(i(8), 16)],
             output: v(i(16), 8),
             definition: Named("llvm.x86.ssse3.pmadd.ub.sw.128")
         },
@@ -285,6 +285,11 @@ pub fn find<'tcx>(_tcx: &ty::ctxt<'tcx>, name: &str) -> Option<Intrinsic> {
             output: v(i(16), 8),
             definition: Named("llvm.x86.ssse3.psign.w.128")
         },
+        "_sign_epi32" => Intrinsic {
+            inputs: vec![v(i(32), 4), v(i(32), 4)],
+            output: v(i(32), 4),
+            definition: Named("llvm.x86.ssse3.psign.d.128")
+        },
         "_dp_ps" => Intrinsic {
             inputs: vec![v(f(32), 4), v(f(32), 4), i_(32, 8)],
             output: v(f(32), 4),
@@ -348,7 +353,7 @@ pub fn find<'tcx>(_tcx: &ty::ctxt<'tcx>, name: &str) -> Option<Intrinsic> {
         "_mul_epi32" => Intrinsic {
             inputs: vec![v(i(32), 4), v(i(32), 4)],
             output: v(i(64), 2),
-            definition: Named("llvm.x86.sse41.muldq")
+            definition: Named("llvm.x86.sse41.pmuldq")
         },
         "_packus_epi32" => Intrinsic {
             inputs: vec![v(i(32), 4), v(i(32), 4)],
@@ -360,10 +365,10 @@ pub fn find<'tcx>(_tcx: &ty::ctxt<'tcx>, name: &str) -> Option<Intrinsic> {
             output: i(32),
             definition: Named("llvm.x86.sse41.ptestc")
         },
-        "_testncz_si128" => Intrinsic {
+        "_testnzc_si128" => Intrinsic {
             inputs: vec![v(u(64), 2), v(u(64), 2)],
             output: i(32),
-            definition: Named("llvm.x86.sse41.ptest.nzc")
+            definition: Named("llvm.x86.sse41.ptestnzc")
         },
         "_testz_si128" => Intrinsic {
             inputs: vec![v(u(64), 2), v(u(64), 2)],

src/librustc_trans/trans/consts.rs

@@ -501,14 +501,9 @@ fn const_expr_unadjusted<'a, 'tcx>(cx: &CrateContext<'a, 'tcx>,
             debug!("const_expr_unadjusted: te1={}, ty={:?}",
                    cx.tn().val_to_string(te1),
                    ty);
-            let is_simd = ty.is_simd();
-            let intype = if is_simd {
-                ty.simd_type(cx.tcx())
-            } else {
-                ty
-            };
-            let is_float = intype.is_fp();
-            let signed = intype.is_signed();
+            assert!(!ty.is_simd());
+            let is_float = ty.is_fp();
+            let signed = ty.is_signed();
 
             let (te2, _) = const_expr(cx, &**e2, param_substs, fn_args);
 
@@ -552,14 +547,7 @@ fn const_expr_unadjusted<'a, 'tcx>(cx: &CrateContext<'a, 'tcx>,
                         ConstFCmp(cmp, te1, te2)
                     } else {
                         let cmp = base::bin_op_to_icmp_predicate(cx, b.node, signed);
-                        let bool_val = ConstICmp(cmp, te1, te2);
-                        if is_simd {
-                            // LLVM outputs an `< size x i1 >`, so we need to perform
-                            // a sign extension to get the correctly sized type.
-                            llvm::LLVMConstIntCast(bool_val, val_ty(te1).to_ref(), True)
-                        } else {
-                            bool_val
-                        }
+                        ConstICmp(cmp, te1, te2)
                     }
                 },
             } } // unsafe { match b.node {

src/librustc_trans/trans/expr.rs

@@ -1693,14 +1693,9 @@ fn trans_eager_binop<'blk, 'tcx>(bcx: Block<'blk, 'tcx>,
     let _icx = push_ctxt("trans_eager_binop");
 
     let tcx = bcx.tcx();
-    let is_simd = lhs_t.is_simd();
-    let intype = if is_simd {
-        lhs_t.simd_type(tcx)
-    } else {
-        lhs_t
-    };
-    let is_float = intype.is_fp();
-    let is_signed = intype.is_signed();
+    assert!(!lhs_t.is_simd());
+    let is_float = lhs_t.is_fp();
+    let is_signed = lhs_t.is_signed();
     let info = expr_info(binop_expr);
 
     let binop_debug_loc = binop_expr.debug_loc();
@@ -1710,8 +1705,6 @@ fn trans_eager_binop<'blk, 'tcx>(bcx: Block<'blk, 'tcx>,
       ast::BiAdd => {
         if is_float {
             FAdd(bcx, lhs, rhs, binop_debug_loc)
-        } else if is_simd {
-            Add(bcx, lhs, rhs, binop_debug_loc)
         } else {
             let (newbcx, res) = with_overflow_check(
                 bcx, OverflowOp::Add, info, lhs_t, lhs, rhs, binop_debug_loc);
@@ -1722,8 +1715,6 @@ fn trans_eager_binop<'blk, 'tcx>(bcx: Block<'blk, 'tcx>,
       ast::BiSub => {
         if is_float {
             FSub(bcx, lhs, rhs, binop_debug_loc)
-        } else if is_simd {
-            Sub(bcx, lhs, rhs, binop_debug_loc)
         } else {
             let (newbcx, res) = with_overflow_check(
                 bcx, OverflowOp::Sub, info, lhs_t, lhs, rhs, binop_debug_loc);
@@ -1734,8 +1725,6 @@ fn trans_eager_binop<'blk, 'tcx>(bcx: Block<'blk, 'tcx>,
       ast::BiMul => {
         if is_float {
             FMul(bcx, lhs, rhs, binop_debug_loc)
-        } else if is_simd {
-            Mul(bcx, lhs, rhs, binop_debug_loc)
         } else {
             let (newbcx, res) = with_overflow_check(
                 bcx, OverflowOp::Mul, info, lhs_t, lhs, rhs, binop_debug_loc);
@@ -1828,11 +1817,7 @@ fn trans_eager_binop<'blk, 'tcx>(bcx: Block<'blk, 'tcx>,
           res
       }
       ast::BiEq | ast::BiNe | ast::BiLt | ast::BiGe | ast::BiLe | ast::BiGt => {
-        if is_simd {
-            base::compare_simd_types(bcx, lhs, rhs, intype, val_ty(lhs), op.node, binop_debug_loc)
-        } else {
-            base::compare_scalar_types(bcx, lhs, rhs, intype, op.node, binop_debug_loc)
-        }
+          base::compare_scalar_types(bcx, lhs, rhs, lhs_t, op.node, binop_debug_loc)
       }
       _ => {
         bcx.tcx().sess.span_bug(binop_expr.span, "unexpected binop");
@@ -2533,14 +2518,7 @@ fn build_unchecked_rshift<'blk, 'tcx>(bcx: Block<'blk, 'tcx>,
     let rhs = base::cast_shift_expr_rhs(bcx, ast::BinOp_::BiShr, lhs, rhs);
     // #1877, #10183: Ensure that input is always valid
     let rhs = shift_mask_rhs(bcx, rhs, binop_debug_loc);
-    let tcx = bcx.tcx();
-    let is_simd = lhs_t.is_simd();
-    let intype = if is_simd {
-        lhs_t.simd_type(tcx)
-    } else {
-        lhs_t
-    };
-    let is_signed = intype.is_signed();
+    let is_signed = lhs_t.is_signed();
     if is_signed {
         AShr(bcx, lhs, rhs, binop_debug_loc)
     } else {