From da26b84bd0cf996badc6c1c876f5a363158ca8d4 Mon Sep 17 00:00:00 2001
From: Thomas Lively <tlively@google.com>
Date: Thu, 23 Aug 2018 19:23:13 +0000
Subject: [PATCH] [WebAssembly] Prioritize splats over v128.consts

Summary:
Splats are fewer bytes than v128.consts, so use them when either could
apply.

Reviewers: aheejin, dschuff

Subscribers: sbc100, jgravelle-google, sunfish, llvm-commits

Differential Revision: https://reviews.llvm.org/D51179

llvm-svn: 340569
---
 .../lib/Target/WebAssembly/WebAssemblyInstrSIMD.td |  2 ++
 llvm/test/CodeGen/WebAssembly/simd.ll              | 35 ++++++++++++++++++++++
 2 files changed, 37 insertions(+)

diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
index 07d3dbc..2ea5c8d 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
@@ -85,6 +85,8 @@ def splat16 : PatFrag<(ops node:$x), (build_vector
                         node:$x, node:$x, node:$x, node:$x)>;
 multiclass Splat<ValueType vec_t, string vec, WebAssemblyRegClass reg_t,
                  PatFrag splat_pat, bits<32> simdop> {
+  // Prefer splats over v128.const for const splats (65 is lowest that works)
+  let AddedComplexity = 65 in
   defm SPLAT_#vec_t : SIMD_I<(outs V128:$dst), (ins reg_t:$x), (outs), (ins),
                              [(set (vec_t V128:$dst), (splat_pat reg_t:$x))],
                              vec#".splat\t$dst, $x", vec#".splat", simdop>;
diff --git a/llvm/test/CodeGen/WebAssembly/simd.ll b/llvm/test/CodeGen/WebAssembly/simd.ll
index 4d33761..0cb3630 100644
--- a/llvm/test/CodeGen/WebAssembly/simd.ll
+++ b/llvm/test/CodeGen/WebAssembly/simd.ll
@@ -37,6 +37,13 @@ define <16 x i8> @splat_v16i8(i8 %x) {
   ret <16 x i8> %res
 }
 
+; CHECK-LABEL: const_splat_v16i8
+; SIMD128; i8x16.splat
+define <16 x i8> @const_splat_v16i8() {
+  ret <16 x i8> <i8 42, i8 42, i8 42, i8 42, i8 42, i8 42, i8 42, i8 42,
+                 i8 42, i8 42, i8 42, i8 42, i8 42, i8 42, i8 42, i8 42>
+}
+
 ; CHECK-LABEL: extract_v16i8_s:
 ; NO-SIMD128-NOT: i8x16
 ; SIMD128: .param v128{{$}}
@@ -155,6 +162,12 @@ define <8 x i16> @splat_v8i16(i16 %x) {
   ret <8 x i16> %res
 }
 
+; CHECK-LABEL: const_splat_v8i16
+; SIMD128; i16x8.splat
+define <8 x i16> @const_splat_v8i16() {
+  ret <8 x i16> <i16 42, i16 42, i16 42, i16 42, i16 42, i16 42, i16 42, i16 42>
+}
+
 ; CHECK-LABEL: extract_v8i16_s:
 ; NO-SIMD128-NOT: i16x8
 ; SIMD128: .param v128{{$}}
@@ -254,6 +267,12 @@ define <4 x i32> @splat_v4i32(i32 %x) {
   ret <4 x i32> %res
 }
 
+; CHECK-LABEL: const_splat_v4i32
+; SIMD128; i32x4.splat
+define <4 x i32> @const_splat_v4i32() {
+  ret <4 x i32> <i32 42, i32 42, i32 42, i32 42>
+}
+
 ; CHECK-LABEL: extract_v4i32:
 ; NO-SIMD128-NOT: i32x4
 ; SIMD128: .param v128{{$}}
@@ -345,6 +364,10 @@ define <2 x i64> @replace_v2i64(<2 x i64> %v, i64 %x) {
   ret <2 x i64> %res
 }
 
+define <2 x i64> @const_splat_v2i64() {
+  ret <2 x i64> <i64 42, i64 42>
+}
+
 ; CHECK-LABEL: build_v2i64:
 ; NO-SIMD128-NOT: i64x2
 ; SIMD128-VM-NOT: i64x2
@@ -388,6 +411,12 @@ define <4 x float> @splat_v4f32(float %x) {
   ret <4 x float> %res
 }
 
+; CHECK-LABEL: const_splat_v4f32
+; SIMD128; f32x4.splat
+define <4 x float> @const_splat_v4f32() {
+  ret <4 x float> <float 42., float 42., float 42., float 42.>
+}
+
 ; CHECK-LABEL: extract_v4f32:
 ; NO-SIMD128-NOT: f32x4
 ; SIMD128: .param v128{{$}}
@@ -454,6 +483,12 @@ define <2 x double> @splat_v2f64(double %x) {
   ret <2 x double> %res
 }
 
+; CHECK-LABEL: const_splat_v2f64:
+; SIMD128; f64x2.splat
+define <2 x double> @const_splat_v2f64() {
+  ret <2 x double> <double 42., double 42.>
+}
+
 ; CHECK-LABEL: extract_v2f64:
 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-VM-NOT: f64x2
-- 
2.7.4