summaryrefslogtreecommitdiffstats
path: root/llvm
diff options
context:
space:
mode:
authorSanjoy Das <sanjoy@playingwithpointers.com>2017-09-29 00:54:16 +0000
committerSanjoy Das <sanjoy@playingwithpointers.com>2017-09-29 00:54:16 +0000
commit0ac5ba5ade654f4eb9506e7f1d8ceaafd90ee1c1 (patch)
tree7f5e0ce68843db340f343bdcd04599341ea5080a /llvm
parentf51e78017d8b334c019ca8b90d9d1f00886a416a (diff)
downloadbcm5719-llvm-0ac5ba5ade654f4eb9506e7f1d8ceaafd90ee1c1.tar.gz
bcm5719-llvm-0ac5ba5ade654f4eb9506e7f1d8ceaafd90ee1c1.zip
Revert "[BypassSlowDivision] Improve our handling of divisions by constants"
This reverts commit r314253. It causes a miscompile on P100 in an internal benchmark. Reverting while I investigate. llvm-svn: 314482
Diffstat (limited to 'llvm')
-rw-r--r--llvm/lib/Transforms/Utils/BypassSlowDivision.cpp20
-rw-r--r--llvm/test/Transforms/CodeGenPrepare/NVPTX/bypass-slow-div.ll77
2 files changed, 7 insertions, 90 deletions
diff --git a/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp b/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp
index 4aed897d641..d6c31f282e8 100644
--- a/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp
+++ b/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp
@@ -339,6 +339,11 @@ 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)
@@ -354,9 +359,7 @@ 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. 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).
+ // division with a short one in-place.
IRBuilder<> Builder(SlowDivOrRem);
Value *TruncDividend = Builder.CreateTrunc(Dividend, BypassType);
@@ -366,16 +369,7 @@ Optional<QuotRemPair> FastDivInsertionTask::insertFastDivAndRem() {
Value *ExtDiv = Builder.CreateZExt(TruncDiv, getSlowType());
Value *ExtRem = Builder.CreateZExt(TruncRem, getSlowType());
return QuotRemPair(ExtDiv, ExtRem);
- }
-
- 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()) {
+ } else 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
diff --git a/llvm/test/Transforms/CodeGenPrepare/NVPTX/bypass-slow-div.ll b/llvm/test/Transforms/CodeGenPrepare/NVPTX/bypass-slow-div.ll
index 4d824e450ff..4846d52f4d2 100644
--- a/llvm/test/Transforms/CodeGenPrepare/NVPTX/bypass-slow-div.ll
+++ b/llvm/test/Transforms/CodeGenPrepare/NVPTX/bypass-slow-div.ll
@@ -27,80 +27,3 @@ define void @rem_only(i64 %a, i64 %b, i64* %retptr) {
store i64 %d, i64* %retptr
ret void
}
-
-; CHECK-LABEL: @udiv_by_constant(
-define i64 @udiv_by_constant(i32 %a) {
-; CHECK-NEXT: [[A_ZEXT:%.*]] = zext i32 [[A:%.*]] to i64
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[A_ZEXT]] to i32
-; CHECK-NEXT: [[TMP2:%.*]] = udiv i32 [[TMP1]], 50
-; CHECK-NEXT: [[TMP3:%.*]] = zext i32 [[TMP2]] to i64
-; CHECK-NEXT: ret i64 [[TMP3]]
-
- %a.zext = zext i32 %a to i64
- %wide.div = udiv i64 %a.zext, 50
- ret i64 %wide.div
-}
-
-; CHECK-LABEL: @urem_by_constant(
-define i64 @urem_by_constant(i32 %a) {
-; CHECK-NEXT: [[A_ZEXT:%.*]] = zext i32 [[A:%.*]] to i64
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[A_ZEXT]] to i32
-; CHECK-NEXT: [[TMP2:%.*]] = urem i32 [[TMP1]], 50
-; CHECK-NEXT: [[TMP3:%.*]] = zext i32 [[TMP2]] to i64
-; CHECK-NEXT: ret i64 [[TMP3]]
-
- %a.zext = zext i32 %a to i64
- %wide.div = urem i64 %a.zext, 50
- ret i64 %wide.div
-}
-
-; Negative test: instead of emitting a runtime check on %a, we prefer to let the
-; DAGCombiner transform this division by constant into a multiplication (with a
-; "magic constant").
-;
-; CHECK-LABEL: @udiv_by_constant_negative_0(
-define i64 @udiv_by_constant_negative_0(i64 %a) {
-; CHECK-NEXT: [[WIDE_DIV:%.*]] = udiv i64 [[A:%.*]], 50
-; CHECK-NEXT: ret i64 [[WIDE_DIV]]
-
- %wide.div = udiv i64 %a, 50
- ret i64 %wide.div
-}
-
-; Negative test: while we know the dividend is short, the divisor isn't. This
-; test is here for completeness, but instcombine will optimize this to return 0.
-;
-; CHECK-LABEL: @udiv_by_constant_negative_1(
-define i64 @udiv_by_constant_negative_1(i32 %a) {
-; CHECK-NEXT: [[A_ZEXT:%.*]] = zext i32 [[A:%.*]] to i64
-; CHECK-NEXT: [[WIDE_DIV:%.*]] = udiv i64 [[A_ZEXT]], 8589934592
-; CHECK-NEXT: ret i64 [[WIDE_DIV]]
-
- %a.zext = zext i32 %a to i64
- %wide.div = udiv i64 %a.zext, 8589934592 ;; == 1 << 33
- ret i64 %wide.div
-}
-
-; URem version of udiv_by_constant_negative_0
-;
-; CHECK-LABEL: @urem_by_constant_negative_0(
-define i64 @urem_by_constant_negative_0(i64 %a) {
-; CHECK-NEXT: [[WIDE_DIV:%.*]] = urem i64 [[A:%.*]], 50
-; CHECK-NEXT: ret i64 [[WIDE_DIV]]
-
- %wide.div = urem i64 %a, 50
- ret i64 %wide.div
-}
-
-; URem version of udiv_by_constant_negative_1
-;
-; CHECK-LABEL: @urem_by_constant_negative_1(
-define i64 @urem_by_constant_negative_1(i32 %a) {
-; CHECK-NEXT: [[A_ZEXT:%.*]] = zext i32 [[A:%.*]] to i64
-; CHECK-NEXT: [[WIDE_DIV:%.*]] = urem i64 [[A_ZEXT]], 8589934592
-; CHECK-NEXT: ret i64 [[WIDE_DIV]]
-
- %a.zext = zext i32 %a to i64
- %wide.div = urem i64 %a.zext, 8589934592 ;; == 1 << 33
- ret i64 %wide.div
-}
OpenPOWER on IntegriCloud