From af4637aff659381ce914da5f7368c4424d12764b Mon Sep 17 00:00:00 2001
From: Andrew Brown <andrew.brown@intel.com>
Date: Fri, 25 Oct 2019 11:18:26 -0700
Subject: [PATCH] Add x86 SIMD legalizations for icmp less-than

---
 .../codegen/meta/src/isa/x86/legalize.rs      | 30 +++++++++-----
 .../isa/x86/simd-comparison-run.clif          | 40 +++++++++++++++++++
 2 files changed, 59 insertions(+), 11 deletions(-)

diff --git a/cranelift/codegen/meta/src/isa/x86/legalize.rs b/cranelift/codegen/meta/src/isa/x86/legalize.rs
index e30ea592a9..f45edd513b 100644
--- a/cranelift/codegen/meta/src/isa/x86/legalize.rs
+++ b/cranelift/codegen/meta/src/isa/x86/legalize.rs
@@ -509,34 +509,42 @@ pub(crate) fn define(shared: &mut SharedDefinitions, x86_instructions: &Instruct
         );
     }
 
-    // SIMD icmp ugt
+    // SIMD icmp greater-/less-than
+    let sgt = Literal::enumerator_for(&imm.intcc, "sgt");
     let ugt = Literal::enumerator_for(&imm.intcc, "ugt");
+    let sge = Literal::enumerator_for(&imm.intcc, "sge");
+    let uge = Literal::enumerator_for(&imm.intcc, "uge");
+    let slt = Literal::enumerator_for(&imm.intcc, "slt");
+    let ult = Literal::enumerator_for(&imm.intcc, "ult");
+    let sle = Literal::enumerator_for(&imm.intcc, "sle");
+    let ule = Literal::enumerator_for(&imm.intcc, "ule");
     for ty in &[I8, I16, I32] {
+        // greater-than
         let icmp_ = icmp.bind(vector(*ty, sse_vector_size));
         narrow.legalize(
             def!(c = icmp_(ugt, a, b)),
             vec![def!(x = x86_pmaxu(a, b)), def!(c = icmp(eq, a, x))],
         );
-    }
-
-    // SIMD icmp sge
-    let sge = Literal::enumerator_for(&imm.intcc, "sge");
-    for ty in &[I8, I16, I32] {
         let icmp_ = icmp.bind(vector(*ty, sse_vector_size));
         narrow.legalize(
             def!(c = icmp_(sge, a, b)),
             vec![def!(x = x86_pmins(a, b)), def!(c = icmp(eq, x, b))],
         );
-    }
-
-    // SIMD icmp uge
-    let uge = Literal::enumerator_for(&imm.intcc, "uge");
-    for ty in &[I8, I16, I32] {
         let icmp_ = icmp.bind(vector(*ty, sse_vector_size));
         narrow.legalize(
             def!(c = icmp_(uge, a, b)),
             vec![def!(x = x86_pminu(a, b)), def!(c = icmp(eq, x, b))],
         );
+
+        // less-than
+        let icmp_ = icmp.bind(vector(*ty, sse_vector_size));
+        narrow.legalize(def!(c = icmp_(slt, a, b)), vec![def!(c = icmp(sgt, b, a))]);
+        let icmp_ = icmp.bind(vector(*ty, sse_vector_size));
+        narrow.legalize(def!(c = icmp_(ult, a, b)), vec![def!(c = icmp(ugt, b, a))]);
+        let icmp_ = icmp.bind(vector(*ty, sse_vector_size));
+        narrow.legalize(def!(c = icmp_(sle, a, b)), vec![def!(c = icmp(sge, b, a))]);
+        let icmp_ = icmp.bind(vector(*ty, sse_vector_size));
+        narrow.legalize(def!(c = icmp_(ule, a, b)), vec![def!(c = icmp(uge, b, a))]);
     }
 
     narrow.custom_legalize(shuffle, "convert_shuffle");
diff --git a/cranelift/filetests/filetests/isa/x86/simd-comparison-run.clif b/cranelift/filetests/filetests/isa/x86/simd-comparison-run.clif
index 0cffbc6708..5d96585be0 100644
--- a/cranelift/filetests/filetests/isa/x86/simd-comparison-run.clif
+++ b/cranelift/filetests/filetests/isa/x86/simd-comparison-run.clif
@@ -137,3 +137,43 @@ ebb0:
     return v8
 }
 ; run
+
+function %icmp_slt_i32x4() -> b1 {
+ebb0:
+    v0 = vconst.i32x4 [-1 1 1 1]
+    v1 = vconst.i32x4 [1 2 3 4]
+    v2 = icmp slt v0, v1
+    v8 = vall_true v2
+    return v8
+}
+; run
+
+function %icmp_ult_i32x4() -> b1 {
+ebb0:
+    v0 = vconst.i32x4 [1 1 1 1]
+    v1 = vconst.i32x4 [-1 2 3 4] ; -1 = 0xffff... will be greater than 1 when unsigned
+    v2 = icmp ult v0, v1
+    v8 = vall_true v2
+    return v8
+}
+; run
+
+function %icmp_sle_i16x8() -> b1 {
+ebb0:
+    v0 = vconst.i16x8 [-1 -1 0 0 0 0 0 0]
+    v1 = vconst.i16x8 [-1  0 0 0 0 0 0 0]
+    v2 = icmp sle v0, v1
+    v8 = vall_true v2
+    return v8
+}
+; run
+
+function %icmp_ule_i16x8() -> b1 {
+ebb0:
+    v0 = vconst.i16x8 [-1  0 0 0 0 0 0 0]
+    v1 = vconst.i16x8 [-1 -1 0 0 0 0 0 0]
+    v2 = icmp ule v0, v1
+    v8 = vall_true v2
+    return v8
+}
+; run