61 lines · plain
1; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+simd128 | FileCheck %s2 3; Test that SIMD shifts can be lowered correctly even with shift4; values that are more complex than plain splats.5 6target triple = "wasm32-unknown-unknown"7 8; CHECK-LABEL: shl_add:9; CHECK-NEXT: .functype shl_add (v128, i32, i32) -> (v128)10; CHECK-NEXT: i32.add $push0=, $1, $211; CHECK-NEXT: i8x16.shl $push1=, $0, $pop012; CHECK-NEXT: return $pop113define <16 x i8> @shl_add(<16 x i8> %v, i8 %a, i8 %b) {14 %t1 = insertelement <16 x i8> undef, i8 %a, i32 015 %va = shufflevector <16 x i8> %t1, <16 x i8> undef, <16 x i32> zeroinitializer16 %t2 = insertelement <16 x i8> undef, i8 %b, i32 017 %vb = shufflevector <16 x i8> %t2, <16 x i8> undef, <16 x i32> zeroinitializer18 %shift = add <16 x i8> %va, %vb19 %r = shl <16 x i8> %v, %shift20 ret <16 x i8> %r21}22 23; CHECK-LABEL: shl_abs:24; CHECK-NEXT: .functype shl_abs (v128, i32) -> (v128)25; CHECK-NEXT: i8x16.splat $push0=, $126; CHECK-NEXT: i8x16.abs $push1=, $pop027; CHECK-NEXT: i8x16.extract_lane_u $push2=, $pop1, 028; CHECK-NEXT: i8x16.shl $push3=, $0, $pop229; CHECK-NEXT: return $pop330define <16 x i8> @shl_abs(<16 x i8> %v, i8 %a) {31 %t1 = insertelement <16 x i8> undef, i8 %a, i32 032 %va = shufflevector <16 x i8> %t1, <16 x i8> undef, <16 x i32> zeroinitializer33 %nva = sub <16 x i8> zeroinitializer, %va34 %c = icmp sgt <16 x i8> %va, zeroinitializer35 %shift = select <16 x i1> %c, <16 x i8> %va, <16 x i8> %nva36 %r = shl <16 x i8> %v, %shift37 ret <16 x i8> %r38}39 40; CHECK-LABEL: shl_abs_add:41; CHECK-NEXT: .functype shl_abs_add (v128, i32, i32) -> (v128)42; CHECK-NEXT: i32.add $push0=, $1, $243; CHECK-NEXT: i8x16.splat $push1=, $pop044; CHECK-NEXT: i8x16.abs $push2=, $pop145; CHECK-NEXT: i8x16.extract_lane_u $push3=, $pop2, 046; CHECK-NEXT: i8x16.shl $push4=, $0, $pop347; CHECK-NEXT: return $pop448 49define <16 x i8> @shl_abs_add(<16 x i8> %v, i8 %a, i8 %b) {50 %t1 = insertelement <16 x i8> undef, i8 %a, i32 051 %va = shufflevector <16 x i8> %t1, <16 x i8> undef, <16 x i32> zeroinitializer52 %t2 = insertelement <16 x i8> undef, i8 %b, i32 053 %vb = shufflevector <16 x i8> %t2, <16 x i8> undef, <16 x i32> zeroinitializer54 %vadd = add <16 x i8> %va, %vb55 %nvadd = sub <16 x i8> zeroinitializer, %vadd56 %c = icmp sgt <16 x i8> %vadd, zeroinitializer57 %shift = select <16 x i1> %c, <16 x i8> %vadd, <16 x i8> %nvadd58 %r = shl <16 x i8> %v, %shift59 ret <16 x i8> %r60}61