[x86] replace div/rem with shift/mask for better shuffle combining perf

We know that shuffle masks are power-of-2 sizes, but there's no way (?) for LLVM to know that, 
so hack combineX86ShufflesRecursively() to be much faster by replacing div/rem with shift/mask.

This makes the motivating compile-time bug in PR32037 ( https://bugs.llvm.org/show_bug.cgi?id=32037 ) 
about 9% faster overall.

Differential Revision: https://reviews.llvm.org/D34174

llvm-svn: 305398

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -27970,28 +27970,45 @@ static bool combineX86ShufflesRecursively(ArrayRef<SDValue> SrcOps,
            OpMask.size() % RootMask.size() == 0) ||
           OpMask.size() == RootMask.size()) &&
          "The smaller number of elements must divide the larger.");
-  int MaskWidth = std::max<int>(OpMask.size(), RootMask.size());
-  int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
-  int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
-  assert(((RootRatio == 1 && OpRatio == 1) ||
-          (RootRatio == 1) != (OpRatio == 1)) &&
+
+  // This function can be performance-critical, so we rely on the power-of-2
+  // knowledge that we have about the mask sizes to replace div/rem ops with
+  // bit-masks and shifts.
+  assert(isPowerOf2_32(RootMask.size()) && "Non-power-of-2 shuffle mask sizes");
+  assert(isPowerOf2_32(OpMask.size()) && "Non-power-of-2 shuffle mask sizes");
+  unsigned RootMaskSizeLog2 = countTrailingZeros(RootMask.size());
+  unsigned OpMaskSizeLog2 = countTrailingZeros(OpMask.size());
+
+  unsigned MaskWidth = std::max<unsigned>(OpMask.size(), RootMask.size());
+  unsigned RootRatio = std::max<unsigned>(1, OpMask.size() >> RootMaskSizeLog2);
+  unsigned OpRatio = std::max<unsigned>(1, RootMask.size() >> OpMaskSizeLog2);
+  assert((RootRatio == 1 || OpRatio == 1) &&
          "Must not have a ratio for both incoming and op masks!");
 
-  SmallVector<int, 64> Mask((unsigned)MaskWidth, SM_SentinelUndef);
+  assert(isPowerOf2_32(MaskWidth) && "Non-power-of-2 shuffle mask sizes");
+  assert(isPowerOf2_32(RootRatio) && "Non-power-of-2 shuffle mask sizes");
+  assert(isPowerOf2_32(OpRatio) && "Non-power-of-2 shuffle mask sizes");
+  unsigned RootRatioLog2 = countTrailingZeros(RootRatio);
+  unsigned OpRatioLog2 = countTrailingZeros(OpRatio);
+
+  SmallVector<int, 64> Mask(MaskWidth, SM_SentinelUndef);
 
   // Merge this shuffle operation's mask into our accumulated mask. Note that
   // this shuffle's mask will be the first applied to the input, followed by the
   // root mask to get us all the way to the root value arrangement. The reason
   // for this order is that we are recursing up the operation chain.
-  for (int i = 0; i < MaskWidth; ++i) {
-    int RootIdx = i / RootRatio;
+  for (unsigned i = 0; i < MaskWidth; ++i) {
+    unsigned RootIdx = i >> RootRatioLog2;
     if (RootMask[RootIdx] < 0) {
       // This is a zero or undef lane, we're done.
       Mask[i] = RootMask[RootIdx];
       continue;
     }
 
-    int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
+    // TODO: Here and below, we could convert multiply to shift-left for
+    // performance because we know that our mask sizes are power-of-2.
+    unsigned RootMaskedIdx =
+        RootMask[RootIdx] * RootRatio + (i & (RootRatio - 1));
 
     // Just insert the scaled root mask value if it references an input other
     // than the SrcOp we're currently inserting.
@@ -28001,9 +28018,9 @@ static bool combineX86ShufflesRecursively(ArrayRef<SDValue> SrcOps,
       continue;
     }
 
-    RootMaskedIdx %= MaskWidth;
+    RootMaskedIdx = RootMaskedIdx & (MaskWidth - 1);
 
-    int OpIdx = RootMaskedIdx / OpRatio;
+    unsigned OpIdx = RootMaskedIdx >> OpRatioLog2;
     if (OpMask[OpIdx] < 0) {
       // The incoming lanes are zero or undef, it doesn't matter which ones we
       // are using.
@@ -28012,8 +28029,9 @@ static bool combineX86ShufflesRecursively(ArrayRef<SDValue> SrcOps,
     }
 
     // Ok, we have non-zero lanes, map them through to one of the Op's inputs.
-    int OpMaskedIdx = OpMask[OpIdx] * OpRatio + RootMaskedIdx % OpRatio;
-    OpMaskedIdx %= MaskWidth;
+    unsigned OpMaskedIdx =
+        OpMask[OpIdx] * OpRatio + (RootMaskedIdx & (OpRatio - 1));
+    OpMaskedIdx = OpMaskedIdx & (MaskWidth - 1);
 
     if (OpMask[OpIdx] < (int)OpMask.size()) {
       assert(0 <= InputIdx0 && "Unknown target shuffle input");