[WebAssembly] Optimize ISel for SIMD Boolean reductions

[tlively]

Summary

Converting the result *.{all,any}_true to a bool at the source level
generates LLVM IR that compares the result to 0. This check is
redundant since these instructions already return either 0 or 1 and
therefore conform to the BooleanContents setting for WebAssembly. This
CL adds patterns to detect and remove such redundant operations on the
result of Boolean reductions.

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

 defm NEG : SIMDUnaryInt<ivneg, "neg", 81>;
 
 // Any lane true: any_true
 defm ANYTRUE : SIMDReduce<int_wasm_anytrue, "any_true", 82>;
 
 // All lanes true: all_true
 defm ALLTRUE : SIMDReduce<int_wasm_alltrue, "all_true", 83>;
 
+// Reductions already return 0 or 1, so and 1, setne 0, and seteq 1
+// can be folded out
+foreach reduction =
+  [["int_wasm_anytrue", "ANYTRUE"], ["int_wasm_alltrue", "ALLTRUE"]] in
+foreach ty = [v16i8, v8i16, v4i32, v2i64] in {
+def : Pat<(i32 (and
+            (i32 (!cast<Intrinsic>(reduction[0]) (ty V128:$x))),
+            (i32 1)
+          )),
+          (i32 (!cast<NI>(reduction[1]#"_"#ty) (ty V128:$x)))>;
+def : Pat<(i32 (setne
+            (i32 (!cast<Intrinsic>(reduction[0]) (ty V128:$x))),
+            (i32 0)
+          )),
+          (i32 (!cast<NI>(reduction[1]#"_"#ty) (ty V128:$x)))>;
+def : Pat<(i32 (seteq
+            (i32 (!cast<Intrinsic>(reduction[0]) (ty V128:$x))),
+            (i32 1)
+          )),
+          (i32 (!cast<NI>(reduction[1]#"_"#ty) (ty V128:$x)))>;
+}
+
 //===----------------------------------------------------------------------===//
 // Bit shifts
 //===----------------------------------------------------------------------===//
 
 multiclass SIMDShift<ValueType vec_t, string vec, SDNode node, dag shift_vec,
                      string name, bits<32> simdop> {
   defm _#vec_t : SIMD_I<(outs V128:$dst), (ins V128:$vec, I32:$x),
                         (outs), (ins),

llvm/test/CodeGen/WebAssembly/simd-reductions.ll

+; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+unimplemented-simd128 | FileCheck %s --check-prefixes CHECK,SIMD128
+
+; Tests that redundant masking and conversions are folded out
+; following SIMD reduction instructions.
+
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32-unknown-unknown"
+
+; ==============================================================================
+; 16 x i8
+; ==============================================================================
+declare i32 @llvm.wasm.anytrue.v16i8(<16 x i8>)
+declare i32 @llvm.wasm.alltrue.v16i8(<16 x i8>)
+
+; CHECK-LABEL: any_v16i8_trunc:
+; SIMD128-NEXT: .functype any_v16i8_trunc (v128) -> (i32){{$}}
+; SIMD128-NEXT: i8x16.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v16i8_trunc(<16 x i8> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v16i8(<16 x i8> %x)
+  %b = trunc i32 %a to i1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: any_v16i8_ne:
+; SIMD128-NEXT: .functype any_v16i8_ne (v128) -> (i32){{$}}
+; SIMD128-NEXT: i8x16.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v16i8_ne(<16 x i8> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v16i8(<16 x i8> %x)
+  %b = icmp ne i32 %a, 0
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: any_v16i8_eq:
+; SIMD128-NEXT: .functype any_v16i8_eq (v128) -> (i32){{$}}
+; SIMD128-NEXT: i8x16.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v16i8_eq(<16 x i8> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v16i8(<16 x i8> %x)
+  %b = icmp eq i32 %a, 1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v16i8_trunc:
+; SIMD128-NEXT: .functype all_v16i8_trunc (v128) -> (i32){{$}}
+; SIMD128-NEXT: i8x16.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v16i8_trunc(<16 x i8> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v16i8(<16 x i8> %x)
+  %b = trunc i32 %a to i1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v16i8_ne:
+; SIMD128-NEXT: .functype all_v16i8_ne (v128) -> (i32){{$}}
+; SIMD128-NEXT: i8x16.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v16i8_ne(<16 x i8> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v16i8(<16 x i8> %x)
+  %b = icmp ne i32 %a, 0
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v16i8_eq:
+; SIMD128-NEXT: .functype all_v16i8_eq (v128) -> (i32){{$}}
+; SIMD128-NEXT: i8x16.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v16i8_eq(<16 x i8> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v16i8(<16 x i8> %x)
+  %b = icmp eq i32 %a, 1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; ==============================================================================
+; 8 x i16
+; ==============================================================================
+declare i32 @llvm.wasm.anytrue.v8i16(<8 x i16>)
+declare i32 @llvm.wasm.alltrue.v8i16(<8 x i16>)
+
+; CHECK-LABEL: any_v8i16_trunc:
+; SIMD128-NEXT: .functype any_v8i16_trunc (v128) -> (i32){{$}}
+; SIMD128-NEXT: i16x8.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v8i16_trunc(<8 x i16> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v8i16(<8 x i16> %x)
+  %b = trunc i32 %a to i1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: any_v8i16_ne:
+; SIMD128-NEXT: .functype any_v8i16_ne (v128) -> (i32){{$}}
+; SIMD128-NEXT: i16x8.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v8i16_ne(<8 x i16> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v8i16(<8 x i16> %x)
+  %b = icmp ne i32 %a, 0
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: any_v8i16_eq:
+; SIMD128-NEXT: .functype any_v8i16_eq (v128) -> (i32){{$}}
+; SIMD128-NEXT: i16x8.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v8i16_eq(<8 x i16> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v8i16(<8 x i16> %x)
+  %b = icmp eq i32 %a, 1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v8i16_trunc:
+; SIMD128-NEXT: .functype all_v8i16_trunc (v128) -> (i32){{$}}
+; SIMD128-NEXT: i16x8.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v8i16_trunc(<8 x i16> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v8i16(<8 x i16> %x)
+  %b = trunc i32 %a to i1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v8i16_ne:
+; SIMD128-NEXT: .functype all_v8i16_ne (v128) -> (i32){{$}}
+; SIMD128-NEXT: i16x8.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v8i16_ne(<8 x i16> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v8i16(<8 x i16> %x)
+  %b = icmp ne i32 %a, 0
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v8i16_eq:
+; SIMD128-NEXT: .functype all_v8i16_eq (v128) -> (i32){{$}}
+; SIMD128-NEXT: i16x8.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v8i16_eq(<8 x i16> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v8i16(<8 x i16> %x)
+  %b = icmp eq i32 %a, 1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; ==============================================================================
+; 4 x i32
+; ==============================================================================
+declare i32 @llvm.wasm.anytrue.v4i32(<4 x i32>)
+declare i32 @llvm.wasm.alltrue.v4i32(<4 x i32>)
+
+; CHECK-LABEL: any_v4i32_trunc:
+; SIMD128-NEXT: .functype any_v4i32_trunc (v128) -> (i32){{$}}
+; SIMD128-NEXT: i32x4.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v4i32_trunc(<4 x i32> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v4i32(<4 x i32> %x)
+  %b = trunc i32 %a to i1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: any_v4i32_ne:
+; SIMD128-NEXT: .functype any_v4i32_ne (v128) -> (i32){{$}}
+; SIMD128-NEXT: i32x4.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v4i32_ne(<4 x i32> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v4i32(<4 x i32> %x)
+  %b = icmp ne i32 %a, 0
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: any_v4i32_eq:
+; SIMD128-NEXT: .functype any_v4i32_eq (v128) -> (i32){{$}}
+; SIMD128-NEXT: i32x4.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v4i32_eq(<4 x i32> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v4i32(<4 x i32> %x)
+  %b = icmp eq i32 %a, 1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v4i32_trunc:
+; SIMD128-NEXT: .functype all_v4i32_trunc (v128) -> (i32){{$}}
+; SIMD128-NEXT: i32x4.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v4i32_trunc(<4 x i32> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v4i32(<4 x i32> %x)
+  %b = trunc i32 %a to i1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v4i32_ne:
+; SIMD128-NEXT: .functype all_v4i32_ne (v128) -> (i32){{$}}
+; SIMD128-NEXT: i32x4.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v4i32_ne(<4 x i32> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v4i32(<4 x i32> %x)
+  %b = icmp ne i32 %a, 0
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v4i32_eq:
+; SIMD128-NEXT: .functype all_v4i32_eq (v128) -> (i32){{$}}
+; SIMD128-NEXT: i32x4.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v4i32_eq(<4 x i32> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v4i32(<4 x i32> %x)
+  %b = icmp eq i32 %a, 1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; ==============================================================================
+; 2 x i64
+; ==============================================================================
+declare i32 @llvm.wasm.anytrue.v2i64(<2 x i64>)
+declare i32 @llvm.wasm.alltrue.v2i64(<2 x i64>)
+
+; CHECK-LABEL: any_v2i64_trunc:
+; SIMD128-NEXT: .functype any_v2i64_trunc (v128) -> (i32){{$}}
+; SIMD128-NEXT: i64x2.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v2i64_trunc(<2 x i64> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v2i64(<2 x i64> %x)
+  %b = trunc i32 %a to i1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: any_v2i64_ne:
+; SIMD128-NEXT: .functype any_v2i64_ne (v128) -> (i32){{$}}
+; SIMD128-NEXT: i64x2.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v2i64_ne(<2 x i64> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v2i64(<2 x i64> %x)
+  %b = icmp ne i32 %a, 0
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: any_v2i64_eq:
+; SIMD128-NEXT: .functype any_v2i64_eq (v128) -> (i32){{$}}
+; SIMD128-NEXT: i64x2.any_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @any_v2i64_eq(<2 x i64> %x) {
+  %a = call i32 @llvm.wasm.anytrue.v2i64(<2 x i64> %x)
+  %b = icmp eq i32 %a, 1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v2i64_trunc:
+; SIMD128-NEXT: .functype all_v2i64_trunc (v128) -> (i32){{$}}
+; SIMD128-NEXT: i64x2.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v2i64_trunc(<2 x i64> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v2i64(<2 x i64> %x)
+  %b = trunc i32 %a to i1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v2i64_ne:
+; SIMD128-NEXT: .functype all_v2i64_ne (v128) -> (i32){{$}}
+; SIMD128-NEXT: i64x2.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v2i64_ne(<2 x i64> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v2i64(<2 x i64> %x)
+  %b = icmp ne i32 %a, 0
+  %c = zext i1 %b to i32
+  ret i32 %c
+}
+
+; CHECK-LABEL: all_v2i64_eq:
+; SIMD128-NEXT: .functype all_v2i64_eq (v128) -> (i32){{$}}
+; SIMD128-NEXT: i64x2.all_true $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+define i32 @all_v2i64_eq(<2 x i64> %x) {
+  %a = call i32 @llvm.wasm.alltrue.v2i64(<2 x i64> %x)
+  %b = icmp eq i32 %a, 1
+  %c = zext i1 %b to i32
+  ret i32 %c
+}