X86 cost model: Adjust cost for custom lowered vector multiplies

This matters for example in following matrix multiply:

int **mmult(int rows, int cols, int **m1, int **m2, int **m3) {
  int i, j, k, val;
  for (i=0; i<rows; i++) {
    for (j=0; j<cols; j++) {
      val = 0;
      for (k=0; k<cols; k++) {
        val += m1[i][k] * m2[k][j];
      }
      m3[i][j] = val;
    }
  }
  return(m3);
}

Taken from the test-suite benchmark Shootout.

We estimate the cost of the multiply to be 2 while we generate 9 instructions
for it and end up being quite a bit slower than the scalar version (48% on my
machine).

Also, properly differentiate between avx1 and avx2. On avx-1 we still split the
vector into 2 128bits and handle the subvector muls like above with 9
instructions.
Only on avx-2 will we have a cost of 9 for v4i64.

I changed the test case in test/Transforms/LoopVectorize/X86/avx1.ll to use an
add instead of a mul because with a mul we now no longer vectorize. I did
verify that the mul would be indeed more expensive when vectorized with 3
kernels:

for (i ...)
   r += a[i] * 3;
for (i ...)
  m1[i] = m1[i] * 3; // This matches the test case in avx1.ll
and a matrix multiply.

In each case the vectorized version was considerably slower.

radar://13304919

llvm-svn: 176403

lib/Target/X86/X86TargetTransformInfo.cpp

@@ -176,18 +176,42 @@ unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty) const {
     { ISD::MUL,     MVT::v8i32,    4 },
     { ISD::SUB,     MVT::v8i32,    4 },
     { ISD::ADD,     MVT::v8i32,    4 },
-    { ISD::MUL,     MVT::v4i64,    4 },
     { ISD::SUB,     MVT::v4i64,    4 },
     { ISD::ADD,     MVT::v4i64,    4 },
-    };
+    // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
+    // are lowered as a series of long multiplies(3), shifts(4) and adds(2)
+    // Because we believe v4i64 to be a legal type, we must also include the
+    // split factor of two in the cost table. Therefore, the cost here is 18
+    // instead of 9.
+    { ISD::MUL,     MVT::v4i64,    18 },
+  };
 
   // Look for AVX1 lowering tricks.
-  if (ST->hasAVX()) {
-    int Idx = CostTableLookup<MVT>(AVX1CostTable, array_lengthof(AVX1CostTable), ISD,
-                          LT.second);
+  if (ST->hasAVX() && !ST->hasAVX2()) {
+    int Idx = CostTableLookup<MVT>(AVX1CostTable, array_lengthof(AVX1CostTable),
+                                   ISD, LT.second);
     if (Idx != -1)
       return LT.first * AVX1CostTable[Idx].Cost;
   }
+
+  // Custom lowering of vectors.
+  static const CostTblEntry<MVT> CustomLowered[] = {
+    // A v2i64/v4i64 and multiply is custom lowered as a series of long
+    // multiplies(3), shifts(4) and adds(2).
+    { ISD::MUL,     MVT::v2i64,    9 },
+    { ISD::MUL,     MVT::v4i64,    9 },
+  };
+  int Idx = CostTableLookup<MVT>(CustomLowered, array_lengthof(CustomLowered),
+                                 ISD, LT.second);
+  if (Idx != -1)
+    return LT.first * CustomLowered[Idx].Cost;
+
+  // Special lowering of v4i32 mul on sse2, sse3: Lower v4i32 mul as 2x shuffle,
+  // 2x pmuludq, 2x shuffle.
+  if (ISD == ISD::MUL && LT.second == MVT::v4i32 && ST->hasSSE2() &&
+      !ST->hasSSE41())
+    return 6;
+
   // Fallback to the default implementation.
   return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty);
 }

test/Analysis/CostModel/X86/arith.ll

@@ -1,4 +1,6 @@
 ; RUN: opt < %s  -cost-model -analyze -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s
+; RUN: opt < %s  -cost-model -analyze -mtriple=x86_64-apple-macosx10.8.0 -mcpu=core2 | FileCheck %s --check-prefix=SSE3
+; RUN: opt < %s  -cost-model -analyze -mtriple=x86_64-apple-macosx10.8.0 -mcpu=core-avx2 | FileCheck %s --check-prefix=AVX2
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.8.0"
@@ -32,7 +34,37 @@ define i32 @xor(i32 %arg) {
   ret i32 undef
 }
 
+; CHECK: mul
+define void @mul() {
+  ; A <2 x i32> gets expanded to a <2 x i64> vector.
+  ; A <2 x i64> vector multiply is implemented using
+  ; 3 PMULUDQ and 2 PADDS and 4 shifts.
+  ;CHECK: cost of 9 {{.*}} mul
+  %A0 = mul <2 x i32> undef, undef
+  ;CHECK: cost of 9 {{.*}} mul
+  %A1 = mul <2 x i64> undef, undef
+  ;CHECK: cost of 18 {{.*}} mul
+  %A2 = mul <4 x i64> undef, undef
+  ret void
+}
 
+; SSE3: sse3mull
+define void @sse3mull() {
+  ; SSE3: cost of 6 {{.*}} mul
+  %A0 = mul <4 x i32> undef, undef
+  ret void
+  ; SSE3: avx2mull
+}
+
+; AVX2: avx2mull
+define void @avx2mull() {
+  ; AVX2: cost of 9 {{.*}} mul
+  %A0 = mul <4 x i64> undef, undef
+  ret void
+  ; AVX2: fmul
+}
+
+; CHECK: fmul
 define i32 @fmul(i32 %arg) {
   ;CHECK: cost of 1 {{.*}} fmul
   %A = fmul <4 x float> undef, undef

test/Transforms/LoopVectorize/X86/avx1.ll

@@ -27,7 +27,7 @@ define i32 @read_mod_write_single_ptr(float* nocapture %a, i32 %n) nounwind uwta
 
 
 ;CHECK: @read_mod_i64
-;CHECK: load <4 x i64>
+;CHECK: load <2 x i64>
 ;CHECK: ret i32
 define i32 @read_mod_i64(i64* nocapture %a, i32 %n) nounwind uwtable ssp {
   %1 = icmp sgt i32 %n, 0
@@ -37,7 +37,7 @@ define i32 @read_mod_i64(i64* nocapture %a, i32 %n) nounwind uwtable ssp {
   %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %0 ]
   %2 = getelementptr inbounds i64* %a, i64 %indvars.iv
   %3 = load i64* %2, align 4
-  %4 = mul i64 %3, 3
+  %4 = add i64 %3, 3
   store i64 %4, i64* %2, align 4
   %indvars.iv.next = add i64 %indvars.iv, 1
   %lftr.wideiv = trunc i64 %indvars.iv.next to i32