[block-freq] Add BlockFrequency::scale that returns a remainder from the division and make the private scale in BlockFrequency more performant.
This change is the first in a series of changes improving LLVM's Block
Frequency propogation implementation to not lose probability mass in
branchy code when propogating block frequency information from a basic
block to its successors. This patch is a simple infrastructure
improvement that does not actually modify the block frequency
algorithm. The specific changes are:
1. Changes the division algorithm used when scaling block frequencies by
branch probabilities to a short division algorithm. This gives us the
remainder for free as well as provides a nice speed boost. When I
benched the old routine and the new routine on a Sandy Bridge iMac with
disabled turbo mode performing 8192 iterations on an array of length
32768, I saw ~600% increase in speed in mean/median performance.
2. Exposes a scale method that returns a remainder. This is important so
we can ensure that when we scale a block frequency by some branch
probability BP = N/D, the remainder from the division by D can be
retrieved and propagated to other children to ensure no probability mass
is lost (more to come on this).
llvm-svn: 194950
diff --git a/llvm/lib/Support/BlockFrequency.cpp b/llvm/lib/Support/BlockFrequency.cpp
index 5e45e46..00efe90 100644
--- a/llvm/lib/Support/BlockFrequency.cpp
+++ b/llvm/lib/Support/BlockFrequency.cpp
@@ -19,52 +19,69 @@
using namespace llvm;
/// Multiply FREQ by N and store result in W array.
-static void mult96bit(uint64_t freq, uint32_t N, uint64_t W[2]) {
+static void mult96bit(uint64_t freq, uint32_t N, uint32_t W[3]) {
uint64_t u0 = freq & UINT32_MAX;
uint64_t u1 = freq >> 32;
- // Represent 96-bit value as w[2]:w[1]:w[0];
- uint32_t w[3] = { 0, 0, 0 };
-
+ // Represent 96-bit value as W[2]:W[1]:W[0];
uint64_t t = u0 * N;
uint64_t k = t >> 32;
- w[0] = t;
+ W[0] = t;
t = u1 * N + k;
- w[1] = t;
- w[2] = t >> 32;
-
- // W[1] - higher bits.
- // W[0] - lower bits.
- W[0] = w[0] + ((uint64_t) w[1] << 32);
- W[1] = w[2];
+ W[1] = t;
+ W[2] = t >> 32;
}
+/// Divide 96-bit value stored in W[2]:W[1]:W[0] by D. Since our word size is a
+/// 32 bit unsigned integer, we can use a short division algorithm.
+static uint64_t divrem96bit(uint32_t W[3], uint32_t D, uint32_t *Rout) {
+ // We assume that W[2] is non-zero since if W[2] is not then the user should
+ // just use hardware division.
+ assert(W[2] && "This routine assumes that W[2] is non-zero since if W[2] is "
+ "zero, the caller should just use 64/32 hardware.");
+ uint32_t Q[3] = { 0, 0, 0 };
-/// Divide 96-bit value stored in W array by D.
-/// Return 64-bit quotient, saturated to UINT64_MAX on overflow.
-static uint64_t div96bit(uint64_t W[2], uint32_t D) {
- uint64_t y = W[0];
- uint64_t x = W[1];
- unsigned i;
-
- assert(x != 0 && "This is really a 64-bit division");
-
- // This long division algorithm automatically saturates on overflow.
- for (i = 0; i < 64 && x; ++i) {
- uint32_t t = -((x >> 31) & 1); // Splat bit 31 to bits 0-31.
- x = (x << 1) | (y >> 63);
- y = y << 1;
- if ((x | t) >= D) {
- x -= D;
- ++y;
+ // The generalized short division algorithm sets i to m + n - 1, where n is
+ // the number of words in the divisior and m is the number of words by which
+ // the divident exceeds the divisor (i.e. m + n == the length of the dividend
+ // in words). Due to our assumption that W[2] is non-zero, we know that the
+ // dividend is of length 3 implying since n is 1 that m = 2. Thus we set i to
+ // m + n - 1 = 2 + 1 - 1 = 2.
+ uint32_t R = 0;
+ for (int i = 2; i >= 0; --i) {
+ uint64_t PartialD = uint64_t(R) << 32 | W[i];
+ if (PartialD == 0) {
+ Q[i] = 0;
+ R = 0;
+ } else if (PartialD < D) {
+ Q[i] = 0;
+ R = uint32_t(PartialD);
+ } else if (PartialD == D) {
+ Q[i] = 1;
+ R = 0;
+ } else {
+ Q[i] = uint32_t(PartialD / D);
+ R = uint32_t(PartialD - (Q[i] * D));
}
}
- return y << (64 - i);
+ // If Q[2] is non-zero, then we overflowed.
+ uint64_t Result;
+ if (Q[2]) {
+ Result = UINT64_MAX;
+ R = D;
+ } else {
+ // Form the final uint64_t result, avoiding endianness issues.
+ Result = uint64_t(Q[0]) | (uint64_t(Q[1]) << 32);
+ }
+
+ if (Rout)
+ *Rout = R;
+
+ return Result;
}
-
-void BlockFrequency::scale(uint32_t N, uint32_t D) {
+uint32_t BlockFrequency::scale(uint32_t N, uint32_t D) {
assert(D != 0 && "Division by zero");
// Calculate Frequency * N.
@@ -75,15 +92,16 @@
// If the product fits in 64 bits, just use built-in division.
if (MulHi <= UINT32_MAX && MulRes >= MulLo) {
Frequency = MulRes / D;
- return;
+ return MulRes % D;
}
// Product overflowed, use 96-bit operations.
- // 96-bit value represented as W[1]:W[0].
- uint64_t W[2];
+ // 96-bit value represented as W[2]:W[1]:W[0].
+ uint32_t W[3];
+ uint32_t R;
mult96bit(Frequency, N, W);
- Frequency = div96bit(W, D);
- return;
+ Frequency = divrem96bit(W, D, &R);
+ return R;
}
BlockFrequency &BlockFrequency::operator*=(const BranchProbability &Prob) {
@@ -127,6 +145,10 @@
return Freq;
}
+uint32_t BlockFrequency::scale(const BranchProbability &Prob) {
+ return scale(Prob.getNumerator(), Prob.getDenominator());
+}
+
void BlockFrequency::print(raw_ostream &OS) const {
// Convert fixed-point number to decimal.
OS << Frequency / getEntryFrequency() << ".";