summaryrefslogtreecommitdiffstats
path: root/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp
diff options
context:
space:
mode:
authorSanjoy Das <sanjoy@playingwithpointers.com>2017-12-04 19:21:58 +0000
committerSanjoy Das <sanjoy@playingwithpointers.com>2017-12-04 19:21:58 +0000
commitaa92cae14e998670780065f92ae9ed1431d69ff1 (patch)
tree602685dd6d055e7070d5512130067e1ae413b959 /llvm/lib/Transforms/Utils/BypassSlowDivision.cpp
parentfeb8343281cf723081c189d3e661209490f92ac8 (diff)
downloadbcm5719-llvm-aa92cae14e998670780065f92ae9ed1431d69ff1.tar.gz
bcm5719-llvm-aa92cae14e998670780065f92ae9ed1431d69ff1.zip
[BypassSlowDivision] Improve our handling of divisions by constants
(This reapplies r314253. r314253 was reverted on r314482 because of a correctness regression on P100, but that regression was identified to be something else.) Summary: Don't bail out on constant divisors for divisions that can be narrowed without introducing control flow . This gives us a 32 bit multiply instead of an emulated 64 bit multiply in the generated PTX assembly. Reviewers: jlebar Subscribers: jholewinski, mcrosier, llvm-commits Differential Revision: https://reviews.llvm.org/D38265 llvm-svn: 319677
Diffstat (limited to 'llvm/lib/Transforms/Utils/BypassSlowDivision.cpp')
-rw-r--r--llvm/lib/Transforms/Utils/BypassSlowDivision.cpp20
1 files changed, 13 insertions, 7 deletions
diff --git a/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp b/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp
index e9c14c93a9a..f711b192f60 100644
--- a/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp
+++ b/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp
@@ -352,11 +352,6 @@ Optional<QuotRemPair> FastDivInsertionTask::insertFastDivAndRem() {
Value *Dividend = SlowDivOrRem->getOperand(0);
Value *Divisor = SlowDivOrRem->getOperand(1);
- if (isa<ConstantInt>(Divisor)) {
- // Keep division by a constant for DAGCombiner.
- return None;
- }
-
VisitedSetTy SetL;
ValueRange DividendRange = getValueRange(Dividend, SetL);
if (DividendRange == VALRNG_LIKELY_LONG)
@@ -372,7 +367,9 @@ Optional<QuotRemPair> FastDivInsertionTask::insertFastDivAndRem() {
if (DividendShort && DivisorShort) {
// If both operands are known to be short then just replace the long
- // division with a short one in-place.
+ // division with a short one in-place. Since we're not introducing control
+ // flow in this case, narrowing the division is always a win, even if the
+ // divisor is a constant (and will later get replaced by a multiplication).
IRBuilder<> Builder(SlowDivOrRem);
Value *TruncDividend = Builder.CreateTrunc(Dividend, BypassType);
@@ -382,7 +379,16 @@ Optional<QuotRemPair> FastDivInsertionTask::insertFastDivAndRem() {
Value *ExtDiv = Builder.CreateZExt(TruncDiv, getSlowType());
Value *ExtRem = Builder.CreateZExt(TruncRem, getSlowType());
return QuotRemPair(ExtDiv, ExtRem);
- } else if (DividendShort && !isSignedOp()) {
+ }
+
+ if (isa<ConstantInt>(Divisor)) {
+ // If the divisor is not a constant, DAGCombiner will convert it to a
+ // multiplication by a magic constant. It isn't clear if it is worth
+ // introducing control flow to get a narrower multiply.
+ return None;
+ }
+
+ if (DividendShort && !isSignedOp()) {
// If the division is unsigned and Dividend is known to be short, then
// either
// 1) Divisor is less or equal to Dividend, and the result can be computed
OpenPOWER on IntegriCloud