ARM cost model: Add cost for gather/scather

Fixes a 35% degradation compared to unvectorized code in
MiBench/automotive-susan and an equally serious regression on a private
image processing benchmark.

radar://14351991

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@186188 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMTargetTransformInfo.cpp b/lib/Target/ARM/ARMTargetTransformInfo.cpp
index 79f56a4b68cc544036bbb8120d7972f4428b228f..5cc64ded465b58f25e6c899702ccc64cd004817b 100644 (file)
--- a/lib/Target/ARM/ARMTargetTransformInfo.cpp
+++ b/lib/Target/ARM/ARMTargetTransformInfo.cpp
@@ -426,6 +426,15 @@ unsigned ARMTTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
 }
 
 unsigned ARMTTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {
+  // Address computations in vectorized code with non-consecutive addresses will
+  // likely result in more instructions compared to scalar code where the
+  // computation can more often be merged into the index mode. The resulting
+  // extra micro-ops can significantly decrease throughput.
+  unsigned NumVectorInstToHideOverhead = 10;
+
+  if (Ty->isVectorTy() && IsComplex)
+    return NumVectorInstToHideOverhead;
+
   // In many cases the address computation is not merged into the instruction
   // addressing mode.
   return 1;

diff --git a/test/Transforms/LoopVectorize/ARM/gather-cost.ll b/test/Transforms/LoopVectorize/ARM/gather-cost.ll
new file mode 100644 (file)
index 0000000..239a28f
--- /dev/null
+++ b/test/Transforms/LoopVectorize/ARM/gather-cost.ll
@@ -0,0 +1,88 @@
+; RUN: opt -loop-vectorize -mtriple=thumbv7s-apple-ios6.0.0 -S < %s | FileCheck %s
+
+target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:32:64-v128:32:128-a0:0:32-n32-S32"
+
+@kernel = global [512 x float] zeroinitializer, align 4
+@kernel2 = global [512 x float] zeroinitializer, align 4
+@kernel3 = global [512 x float] zeroinitializer, align 4
+@kernel4 = global [512 x float] zeroinitializer, align 4
+@src_data = global [1536 x float] zeroinitializer, align 4
+@r_ = global i8 0, align 4
+@g_ = global i8 0, align 4
+@b_ = global i8 0, align 4
+
+; We don't want to vectorize most loops containing gathers because they are
+; expensive. This function represents a point where vectorization starts to
+; become beneficial.
+; Make sure we are conservative and don't vectorize it.
+; CHECK-NOT: <2 x float>
+; CHECK-NOT: <4 x float>
+
+define void @_Z4testmm(i32 %size, i32 %offset) {
+entry:
+  %cmp53 = icmp eq i32 %size, 0
+  br i1 %cmp53, label %for.end, label %for.body.lr.ph
+
+for.body.lr.ph:
+  br label %for.body
+
+for.body:
+  %r.057 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add10, %for.body ]
+  %g.056 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add20, %for.body ]
+  %v.055 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
+  %b.054 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add30, %for.body ]
+  %add = add i32 %v.055, %offset
+  %mul = mul i32 %add, 3
+  %arrayidx = getelementptr inbounds [1536 x float]* @src_data, i32 0, i32 %mul
+  %0 = load float* %arrayidx, align 4
+  %arrayidx2 = getelementptr inbounds [512 x float]* @kernel, i32 0, i32 %v.055
+  %1 = load float* %arrayidx2, align 4
+  %mul3 = fmul fast float %0, %1
+  %arrayidx4 = getelementptr inbounds [512 x float]* @kernel2, i32 0, i32 %v.055
+  %2 = load float* %arrayidx4, align 4
+  %mul5 = fmul fast float %mul3, %2
+  %arrayidx6 = getelementptr inbounds [512 x float]* @kernel3, i32 0, i32 %v.055
+  %3 = load float* %arrayidx6, align 4
+  %mul7 = fmul fast float %mul5, %3
+  %arrayidx8 = getelementptr inbounds [512 x float]* @kernel4, i32 0, i32 %v.055
+  %4 = load float* %arrayidx8, align 4
+  %mul9 = fmul fast float %mul7, %4
+  %add10 = fadd fast float %r.057, %mul9
+  %arrayidx.sum = add i32 %mul, 1
+  %arrayidx11 = getelementptr inbounds [1536 x float]* @src_data, i32 0, i32 %arrayidx.sum
+  %5 = load float* %arrayidx11, align 4
+  %mul13 = fmul fast float %1, %5
+  %mul15 = fmul fast float %2, %mul13
+  %mul17 = fmul fast float %3, %mul15
+  %mul19 = fmul fast float %4, %mul17
+  %add20 = fadd fast float %g.056, %mul19
+  %arrayidx.sum52 = add i32 %mul, 2
+  %arrayidx21 = getelementptr inbounds [1536 x float]* @src_data, i32 0, i32 %arrayidx.sum52
+  %6 = load float* %arrayidx21, align 4
+  %mul23 = fmul fast float %1, %6
+  %mul25 = fmul fast float %2, %mul23
+  %mul27 = fmul fast float %3, %mul25
+  %mul29 = fmul fast float %4, %mul27
+  %add30 = fadd fast float %b.054, %mul29
+  %inc = add i32 %v.055, 1
+  %exitcond = icmp ne i32 %inc, %size
+  br i1 %exitcond, label %for.body, label %for.cond.for.end_crit_edge
+
+for.cond.for.end_crit_edge:
+  %add30.lcssa = phi float [ %add30, %for.body ]
+  %add20.lcssa = phi float [ %add20, %for.body ]
+  %add10.lcssa = phi float [ %add10, %for.body ]
+  %phitmp = fptoui float %add10.lcssa to i8
+  %phitmp60 = fptoui float %add20.lcssa to i8
+  %phitmp61 = fptoui float %add30.lcssa to i8
+  br label %for.end
+
+for.end:
+  %r.0.lcssa = phi i8 [ %phitmp, %for.cond.for.end_crit_edge ], [ 0, %entry ]
+  %g.0.lcssa = phi i8 [ %phitmp60, %for.cond.for.end_crit_edge ], [ 0, %entry ]
+  %b.0.lcssa = phi i8 [ %phitmp61, %for.cond.for.end_crit_edge ], [ 0, %entry ]
+  store i8 %r.0.lcssa, i8* @r_, align 4
+  store i8 %g.0.lcssa, i8* @g_, align 4
+  store i8 %b.0.lcssa, i8* @b_, align 4
+  ret void
+}