94 lines · plain
1; RUN: opt -mtriple=armv8m.main-none-none-eabi -mattr=+dsp < %s -arm-parallel-dsp -S | FileCheck %s2 3; CHECK-LABEL: @test14; CHECK: %mac1{{\.}}026 = phi i32 [ [[V8:%[0-9]+]], %for.body ], [ 0, %for.body.preheader ]5; CHECK: [[V5:%[0-9]+]] = load i32, ptr %arrayidx, align 26; CHECK: [[V7:%[0-9]+]] = load i32, ptr %arrayidx3, align 27; CHECK: [[V8]] = call i32 @llvm.arm.smlad(i32 [[V7]], i32 [[V5]], i32 %mac1{{\.}}026)8 9define dso_local i32 @test1(i32 %arg, ptr nocapture readnone %arg1, ptr nocapture readonly %arg2, ptr nocapture readonly %arg3) {10entry:11 %cmp24 = icmp sgt i32 %arg, 012 br i1 %cmp24, label %for.body.preheader, label %for.cond.cleanup13 14for.body.preheader:15 %.pre = load i16, ptr %arg3, align 216 %.pre27 = load i16, ptr %arg2, align 217 br label %for.body18 19for.cond.cleanup:20 %mac1.0.lcssa = phi i32 [ 0, %entry ], [ %add11, %for.body ]21 ret i32 %mac1.0.lcssa22 23for.body:24 %mac1.026 = phi i32 [ %add11, %for.body ], [ 0, %for.body.preheader ]25 %i.025 = phi i32 [ %add, %for.body ], [ 0, %for.body.preheader ]26 %arrayidx = getelementptr inbounds i16, ptr %arg3, i32 %i.02527 %0 = load i16, ptr %arrayidx, align 228 %add = add nuw nsw i32 %i.025, 129 %arrayidx1 = getelementptr inbounds i16, ptr %arg3, i32 %add30 %1 = load i16, ptr %arrayidx1, align 231 %arrayidx3 = getelementptr inbounds i16, ptr %arg2, i32 %i.02532 %2 = load i16, ptr %arrayidx3, align 233 %conv = sext i16 %2 to i3234 %conv4 = sext i16 %0 to i3235 %mul = mul nsw i32 %conv, %conv436 %arrayidx6 = getelementptr inbounds i16, ptr %arg2, i32 %add37 %3 = load i16, ptr %arrayidx6, align 238 %conv7 = sext i16 %3 to i3239 %conv8 = sext i16 %1 to i3240 %mul9 = mul nsw i32 %conv7, %conv841 %add10 = add i32 %mul, %mac1.02642 43; And here the Add is the LHS, the Mul the RHS44 %add11 = add i32 %add10, %mul945 46 %exitcond = icmp ne i32 %add, %arg47 br i1 %exitcond, label %for.body, label %for.cond.cleanup48}49 50; Here we have i8 loads, which we do want to support, but don't handle yet.51;52; CHECK-LABEL: @test253; CHECK-NOT: call i32 @llvm.arm.smlad54;55define dso_local i32 @test2(i32 %arg, ptr nocapture readnone %arg1, ptr nocapture readonly %arg2, ptr nocapture readonly %arg3) {56entry:57 %cmp24 = icmp sgt i32 %arg, 058 br i1 %cmp24, label %for.body.preheader, label %for.cond.cleanup59 60for.body.preheader:61 %.pre = load i8, ptr %arg3, align 262 %.pre27 = load i8, ptr %arg2, align 263 br label %for.body64 65for.cond.cleanup:66 %mac1.0.lcssa = phi i32 [ 0, %entry ], [ %add11, %for.body ]67 ret i32 %mac1.0.lcssa68 69for.body:70 %mac1.026 = phi i32 [ %add11, %for.body ], [ 0, %for.body.preheader ]71 %i.025 = phi i32 [ %add, %for.body ], [ 0, %for.body.preheader ]72 %arrayidx = getelementptr inbounds i8, ptr %arg3, i32 %i.02573 %0 = load i8, ptr %arrayidx, align 274 %add = add nuw nsw i32 %i.025, 175 %arrayidx1 = getelementptr inbounds i8, ptr %arg3, i32 %add76 %1 = load i8, ptr %arrayidx1, align 277 %arrayidx3 = getelementptr inbounds i8, ptr %arg2, i32 %i.02578 %2 = load i8, ptr %arrayidx3, align 279 %conv = sext i8 %2 to i3280 %conv4 = sext i8 %0 to i3281 %mul = mul nsw i32 %conv, %conv482 %arrayidx6 = getelementptr inbounds i8, ptr %arg2, i32 %add83 %3 = load i8, ptr %arrayidx6, align 284 %conv7 = sext i8 %3 to i3285 %conv8 = sext i8 %1 to i3286 %mul9 = mul nsw i32 %conv7, %conv887 %add10 = add i32 %mul, %mac1.02688 %add11 = add i32 %add10, %mul989 %exitcond = icmp ne i32 %add, %arg90 br i1 %exitcond, label %for.body, label %for.cond.cleanup91}92 93 94