From fd7135a399c064399ba176e97f3386bd41b8a279 Mon Sep 17 00:00:00 2001
From: Thomas Lively <tlively@google.com>
Date: Sat, 31 Aug 2019 00:12:29 +0000
Subject: [PATCH] [WebAssembly] Add SIMD QFMA/QFMS

Summary:
Adds clang builtins and LLVM intrinsics for these experimental
instructions. They are not implemented in engines yet, but that is ok
because the user must opt into using them by calling the builtins.

Reviewers: aheejin, dschuff

Reviewed By: aheejin

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

Tags: #clang, #llvm

Differential Revision: https://reviews.llvm.org/D67020

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@370556 91177308-0d34-0410-b5e6-96231b3b80d8
---
 include/llvm/IR/IntrinsicsWebAssembly.td      |  8 +++
 .../WebAssembly/WebAssemblyInstrSIMD.td       | 21 ++++++++
 test/CodeGen/WebAssembly/simd-intrinsics.ll   | 50 ++++++++++++++++++-
 test/MC/WebAssembly/simd-encodings.s          | 12 +++++
 4 files changed, 90 insertions(+), 1 deletion(-)

diff --git a/include/llvm/IR/IntrinsicsWebAssembly.td b/include/llvm/IR/IntrinsicsWebAssembly.td
index 4750c315611..73d190b0b55 100644
--- a/include/llvm/IR/IntrinsicsWebAssembly.td
+++ b/include/llvm/IR/IntrinsicsWebAssembly.td
@@ -109,6 +109,14 @@ def int_wasm_alltrue :
   Intrinsic<[llvm_i32_ty],
             [llvm_anyvector_ty],
             [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_qfma :
+  Intrinsic<[llvm_anyvector_ty],
+            [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_qfms :
+  Intrinsic<[llvm_anyvector_ty],
+            [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
+            [IntrNoMem, IntrSpeculatable]>;
 
 //===----------------------------------------------------------------------===//
 // Bulk memory intrinsics
diff --git a/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td b/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
index dd8930f079b..ff031675959 100644
--- a/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
+++ b/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
@@ -732,3 +732,24 @@ foreach t2 = !foldl(
   )
 ) in
 def : Pat<(t1 (bitconvert (t2 V128:$v))), (t1 V128:$v)>;
+
+//===----------------------------------------------------------------------===//
+// Quasi-Fused Multiply- Add and Subtract (QFMA/QFMS)
+//===----------------------------------------------------------------------===//
+multiclass SIMDQFM<ValueType vec_t, string vec, bits<32> baseInst> {
+  defm QFMA_#vec_t :
+    SIMD_I<(outs V128:$dst), (ins V128:$a, V128:$b, V128:$c),
+           (outs), (ins),
+           [(set (vec_t V128:$dst),
+             (int_wasm_qfma (vec_t V128:$a), (vec_t V128:$b), (vec_t V128:$c)))],
+           vec#".qfma\t$dst, $a, $b, $c", vec#".qfma", baseInst>;
+  defm QFMS_#vec_t :
+    SIMD_I<(outs V128:$dst), (ins V128:$a, V128:$b, V128:$c),
+           (outs), (ins),
+           [(set (vec_t V128:$dst),
+             (int_wasm_qfms (vec_t V128:$a), (vec_t V128:$b), (vec_t V128:$c)))],
+           vec#".qfms\t$dst, $a, $b, $c", vec#".qfms", !add(baseInst, 1)>;
+}
+
+defm "" : SIMDQFM<v4f32, "f32x4", 0x98>;
+defm "" : SIMDQFM<v2f64, "f64x2", 0xa3>;
diff --git a/test/CodeGen/WebAssembly/simd-intrinsics.ll b/test/CodeGen/WebAssembly/simd-intrinsics.ll
index 53c98d2722b..2077cf8c70e 100644
--- a/test/CodeGen/WebAssembly/simd-intrinsics.ll
+++ b/test/CodeGen/WebAssembly/simd-intrinsics.ll
@@ -290,11 +290,35 @@ define <2 x i64> @trunc_sat_u_v2i64(<2 x double> %x) {
 declare <4 x float> @llvm.wasm.bitselect.v4f32(<4 x float>, <4 x float>, <4 x float>)
 define <4 x float> @bitselect_v4f32(<4 x float> %v1, <4 x float> %v2, <4 x float> %c) {
   %a = call <4 x float> @llvm.wasm.bitselect.v4f32(
-     <4 x float> %v1, <4 x float> %v2, <4 x float> %c
+    <4 x float> %v1, <4 x float> %v2, <4 x float> %c
   )
   ret <4 x float> %a
 }
 
+; CHECK-LABEL: qfma_v4f32:
+; SIMD128-NEXT: .functype qfma_v4f32 (v128, v128, v128) -> (v128){{$}}
+; SIMD128-NEXT: f32x4.qfma $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x float> @llvm.wasm.qfma.v4f32(<4 x float>, <4 x float>, <4 x float>)
+define <4 x float> @qfma_v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c) {
+  %v = call <4 x float> @llvm.wasm.qfma.v4f32(
+    <4 x float> %a, <4 x float> %b, <4 x float> %c
+  )
+  ret <4 x float> %v
+}
+
+; CHECK-LABEL: qfms_v4f32:
+; SIMD128-NEXT: .functype qfms_v4f32 (v128, v128, v128) -> (v128){{$}}
+; SIMD128-NEXT: f32x4.qfms $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x float> @llvm.wasm.qfms.v4f32(<4 x float>, <4 x float>, <4 x float>)
+define <4 x float> @qfms_v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c) {
+  %v = call <4 x float> @llvm.wasm.qfms.v4f32(
+    <4 x float> %a, <4 x float> %b, <4 x float> %c
+  )
+  ret <4 x float> %v
+}
+
 ; ==============================================================================
 ; 2 x f64
 ; ==============================================================================
@@ -309,3 +333,27 @@ define <2 x double> @bitselect_v2f64(<2 x double> %v1, <2 x double> %v2, <2 x do
   )
   ret <2 x double> %a
 }
+
+; CHECK-LABEL: qfma_v2f64:
+; SIMD128-NEXT: .functype qfma_v2f64 (v128, v128, v128) -> (v128){{$}}
+; SIMD128-NEXT: f64x2.qfma $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <2 x double> @llvm.wasm.qfma.v2f64(<2 x double>, <2 x double>, <2 x double>)
+define <2 x double> @qfma_v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c) {
+  %v = call <2 x double> @llvm.wasm.qfma.v2f64(
+    <2 x double> %a, <2 x double> %b, <2 x double> %c
+  )
+  ret <2 x double> %v
+}
+
+; CHECK-LABEL: qfms_v2f64:
+; SIMD128-NEXT: .functype qfms_v2f64 (v128, v128, v128) -> (v128){{$}}
+; SIMD128-NEXT: f64x2.qfms $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <2 x double> @llvm.wasm.qfms.v2f64(<2 x double>, <2 x double>, <2 x double>)
+define <2 x double> @qfms_v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c) {
+  %v = call <2 x double> @llvm.wasm.qfms.v2f64(
+    <2 x double> %a, <2 x double> %b, <2 x double> %c
+  )
+  ret <2 x double> %v
+}
diff --git a/test/MC/WebAssembly/simd-encodings.s b/test/MC/WebAssembly/simd-encodings.s
index b2a39e7cce9..491b4844d7f 100644
--- a/test/MC/WebAssembly/simd-encodings.s
+++ b/test/MC/WebAssembly/simd-encodings.s
@@ -382,6 +382,12 @@ main:
     # CHECK: f32x4.sqrt # encoding: [0xfd,0x97,0x01]
     f32x4.sqrt
 
+    # CHECK: f32x4.qfma # encoding: [0xfd,0x98,0x01]
+    f32x4.qfma
+
+    # CHECK: f32x4.qfms # encoding: [0xfd,0x99,0x01]
+    f32x4.qfms
+
     # CHECK: f32x4.add # encoding: [0xfd,0x9a,0x01]
     f32x4.add
 
@@ -409,6 +415,12 @@ main:
     # CHECK: f64x2.sqrt # encoding: [0xfd,0xa2,0x01]
     f64x2.sqrt
 
+    # CHECK: f64x2.qfma # encoding: [0xfd,0xa3,0x01]
+    f64x2.qfma
+
+    # CHECK: f64x2.qfms # encoding: [0xfd,0xa4,0x01]
+    f64x2.qfms
+
     # CHECK: f64x2.add # encoding: [0xfd,0xa5,0x01]
     f64x2.add
 
-- 
2.50.1