[WebAssembly] [Codegen] Add patterns for relaxed dot

[badumbatish]

The pattern I added for relaxed dot similar to normal dot @ #151775.

For relaxed dot add, i noticed that in the proposal the portion of dot implementation is similar to relaxed dot, so I think we can add a pattern where after we do relaxed dot and do extadd pairwise, we can do relaxed dot add.

One current obstacles is I don't think there is any pattern to singly create a extadd pairwise from other instructions so the relaxed dot add pattern would not cover a wide range of instructions.

related to #55932

[llvmbot]

@llvm/pr-subscribers-backend-webassembly

Author: Jasmine Tang (badumbatish)

Changes

The pattern I added for relaxed dot similar to normal dot @ #151775.

For relaxed dot add, i noticed that in the proposal the portion of dot implementation is similar to relaxed dot, so I think we can add a pattern where after we do relaxed dot and do extmul pairwise, we can do relaxed dot add.

One current obstacles is I don't think there is any pattern to singly create a extmul pairwise from other instructions so the relaxed dot add pattern would not cover a wide range of instructions.

related to #55932

---

Full diff: https://github.com/llvm/llvm-project/pull/163266.diff

2 Files Affected:

• (modified) llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td (+27)
• (added) llvm/test/CodeGen/WebAssembly/simd-relaxed-dot.ll (+104)

diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
index 49af78bce68c3..34b8167ee9c07 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
@@ -1711,6 +1711,26 @@ defm RELAXED_DOT :
             "i16x8.relaxed_dot_i8x16_i7x16_s\t$dst, $lhs, $rhs",
             "i16x8.relaxed_dot_i8x16_i7x16_s", 0x112>;
 
+def : Pat<
+  (v8i16 (add
+    (wasm_shuffle
+      (v8i16 (extmul_low_s v16i8:$lhs, v16i8:$rhs)),
+      (v8i16 (extmul_high_s v16i8:$lhs, v16i8:$rhs)),
+      (i32 0), (i32 1), (i32 4), (i32 5),
+      (i32 8), (i32 9), (i32 12), (i32 13),
+      (i32 16), (i32 17), (i32 20), (i32 21),
+      (i32 24), (i32 25), (i32 28), (i32 29)),
+    (wasm_shuffle
+      (v8i16 (extmul_low_s v16i8:$lhs, v16i8:$rhs)),
+      (v8i16 (extmul_high_s v16i8:$lhs, v16i8:$rhs)),
+      (i32 2), (i32 3), (i32 6), (i32 7),
+      (i32 10), (i32 11), (i32 14), (i32 15),
+      (i32 18), (i32 19), (i32 22), (i32 23),
+      (i32 26), (i32 27), (i32 30), (i32 31)))
+  ),
+  (v8i16 (RELAXED_DOT v16i8:$lhs, v16i8:$rhs))
+>;
+
 defm RELAXED_DOT_ADD :
   RELAXED_I<(outs V128:$dst), (ins V128:$lhs, V128:$rhs, V128:$acc),
             (outs), (ins),
@@ -1719,6 +1739,13 @@ defm RELAXED_DOT_ADD :
             "i32x4.relaxed_dot_i8x16_i7x16_add_s\t$dst, $lhs, $rhs, $acc",
             "i32x4.relaxed_dot_i8x16_i7x16_add_s", 0x113>;
 
+def : Pat<
+  (v4i32 (add
+    (v4i32 (int_wasm_extadd_pairwise_signed
+      (v8i16 (int_wasm_relaxed_dot_i8x16_i7x16_signed v16i8:$lhs, v16i8:$rhs)))),
+    (v4i32 V128:$acc))),
+  (v4i32 (RELAXED_DOT_ADD v16i8:$lhs, v16i8:$rhs, (v4i32 V128:$acc)))
+    >;
 //===----------------------------------------------------------------------===//
 // Relaxed BFloat16 dot product
 //===----------------------------------------------------------------------===//
diff --git a/llvm/test/CodeGen/WebAssembly/simd-relaxed-dot.ll b/llvm/test/CodeGen/WebAssembly/simd-relaxed-dot.ll
new file mode 100644
index 0000000000000..9716cbe077080
--- /dev/null
+++ b/llvm/test/CodeGen/WebAssembly/simd-relaxed-dot.ll
@@ -0,0 +1,104 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers  -mattr=+fp16,+simd128,+relaxed-simd | FileCheck %s
+
+target triple = "wasm32"
+; relaxed_dot stands for relaxed_dot_i8x16_i7x16_s, as in td
+; relaxed_dot_add stands for i32x4.relaxed_dot_i8x16_i7x16_add_s, as in td
+
+define <8 x i16> @relaxed_dot_sext_1(<16 x i8> %a, <16 x i8> %b) {
+; CHECK-LABEL: relaxed_dot_sext_1:
+; CHECK:         .functype relaxed_dot_sext_1 (v128, v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    i16x8.relaxed_dot_i8x16_i7x16_s $push0=, $0, $1
+; CHECK-NEXT:    return $pop0
+  %sext1 = sext <16 x i8> %a to <16 x i16>
+  %sext2 = sext <16 x i8> %b to <16 x i16>
+  %mul = mul <16 x i16> %sext1, %sext2
+  %shuffle1 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>
+  %shuffle2 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
+  %res = add <8 x i16> %shuffle1, %shuffle2
+  ret <8 x i16> %res
+}
+
+
+define <8 x i16> @relaxed_dot_sext_2(<16 x i8> %a, <16 x i8> %b) {
+; CHECK-LABEL: relaxed_dot_sext_2:
+; CHECK:         .functype relaxed_dot_sext_2 (v128, v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    i16x8.relaxed_dot_i8x16_i7x16_s $push0=, $0, $1
+; CHECK-NEXT:    return $pop0
+  %sext1 = sext <16 x i8> %a to <16 x i16>
+  %sext2 = sext <16 x i8> %b to <16 x i16>
+  %mul = mul <16 x i16> %sext1, %sext2
+  %shuffle1 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>
+  %shuffle2 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
+  %res = add <8 x i16> %shuffle2, %shuffle1
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @relaxed_dot_sext_self(<16 x i8> %v) {
+; CHECK-LABEL: relaxed_dot_sext_self:
+; CHECK:         .functype relaxed_dot_sext_self (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    i16x8.relaxed_dot_i8x16_i7x16_s $push0=, $0, $0
+; CHECK-NEXT:    return $pop0
+  %sext = sext <16 x i8> %v to <16 x i16>
+  %mul = mul <16 x i16> %sext, %sext
+  %shuffle1 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>
+  %shuffle2 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
+  %res = add <8 x i16> %shuffle1, %shuffle2
+  ret <8 x i16> %res
+}
+
+define <4 x i32> @relaxed_dot_add_from_relaxed_dot(<16 x i8> %a, <16 x i8> %b, <4 x i32> %c) {
+; CHECK-LABEL: relaxed_dot_add_from_relaxed_dot:
+; CHECK:         .functype relaxed_dot_add_from_relaxed_dot (v128, v128, v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    i32x4.relaxed_dot_i8x16_i7x16_add_s $push0=, $0, $1, $2
+; CHECK-NEXT:    return $pop0
+  %relaxed_dot_call = call <8 x i16> @llvm.wasm.relaxed.dot.i8x16.i7x16.signed(<16 x i8> %a, <16 x i8> %b)
+  %sext = call <4 x i32> @llvm.wasm.extadd.pairwise.signed.v4i32(<8 x i16> %relaxed_dot_call)
+  %res = add <4 x i32> %sext, %c
+  ret <4 x i32> %res
+}
+
+; INFO: Negative test
+define <8 x i16> @relaxed_dot_zext(<16 x i8> %a, <16 x i8> %b) {
+; CHECK-LABEL: relaxed_dot_zext:
+; CHECK:         .functype relaxed_dot_zext (v128, v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    i16x8.extmul_low_i8x16_u $push6=, $0, $1
+; CHECK-NEXT:    local.tee $push5=, $2=, $pop6
+; CHECK-NEXT:    i16x8.extmul_high_i8x16_u $push4=, $0, $1
+; CHECK-NEXT:    local.tee $push3=, $1=, $pop4
+; CHECK-NEXT:    i8x16.shuffle $push1=, $pop5, $pop3, 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29
+; CHECK-NEXT:    i8x16.shuffle $push0=, $2, $1, 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31
+; CHECK-NEXT:    i16x8.add $push2=, $pop1, $pop0
+; CHECK-NEXT:    return $pop2
+  %zext1 = zext <16 x i8> %a to <16 x i16>
+  %zext2 = zext <16 x i8> %b to <16 x i16>
+  %mul = mul <16 x i16> %zext1, %zext2
+  %shuffle1 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>
+  %shuffle2 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
+  %res = add <8 x i16> %shuffle1, %shuffle2
+  ret <8 x i16> %res
+
+}
+
+; INFO: Negative test
+define <8 x i16> @relaxed_dot_wrong_shuffle(<16 x i8> %a, <16 x i8> %b) {
+; CHECK-LABEL: relaxed_dot_wrong_shuffle:
+; CHECK:         .functype relaxed_dot_wrong_shuffle (v128, v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    i16x8.extmul_low_i8x16_s $push1=, $0, $1
+; CHECK-NEXT:    i16x8.extmul_high_i8x16_s $push0=, $0, $1
+; CHECK-NEXT:    i16x8.add $push2=, $pop1, $pop0
+; CHECK-NEXT:    return $pop2
+  %sext1 = sext <16 x i8> %a to <16 x i16>
+  %sext2 = sext <16 x i8> %b to <16 x i16>
+  %mul = mul <16 x i16> %sext1, %sext2
+  %shuffle1 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %shuffle2 = shufflevector <16 x i16> %mul, <16 x i16> poison, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %res = add <8 x i16> %shuffle1, %shuffle2
+  ret <8 x i16> %res
+}

[sparker-arm]

I guess we can also support patterns where rhs is zext?

EDIT: Maybe this is a bad idea, as this is likely to cause undesirable behaviour.

[sparker-arm]

I've also just noticed that the current implementation of these instructions are hard-coded for i7/i8 inputs but they need to be parameterized to support i15/i16.

[badumbatish]

I've also just noticed that the current implementation of these instructions are hard-coded for i7/i8 inputs but they need to be parameterized to support i15/i16.

i've added parameterization to the pattern, let me know if it works

[sparker-arm]

Thanks. And sorry, I should have deleted that comment after my rethink. If we're only supporting sext i8, then we don't need the parameterization, plus the instruction definition still only uses i8x16_i7x16, and I don't think we'll support i16/i15.

Sorry I got confused, these instructions are weird! We'll go with your original implementation.