[llvm] b865eea - [InstCombine] eliminate sext and/or trunc if value has enough signbits
Sanjay Patel via llvm-commits
llvm-commits at lists.llvm.org
Thu Jun 3 10:59:41 PDT 2021
Author: Sanjay Patel
Date: 2021-06-03T13:58:19-04:00
New Revision: b865eead76577b031c6fae5e2490e7be0073f201
URL: https://github.com/llvm/llvm-project/commit/b865eead76577b031c6fae5e2490e7be0073f201
DIFF: https://github.com/llvm/llvm-project/commit/b865eead76577b031c6fae5e2490e7be0073f201.diff
LOG: [InstCombine] eliminate sext and/or trunc if value has enough signbits
If we have enough signbits in a source value, we can skip an
intermediate cast for a trunc+sext pair:
https://alive2.llvm.org/ce/z/A_mQt-
This is the original problem shown in:
https://llvm.org/PR49543
There's a test that shows we transformed what used to be
a pair of shifts, so that suggests we could add another
ComputeNumSignBits fold starting from a shift.
There does not appear to be any change in compile-time
from the extra analysis:
https://llvm-compile-time-tracker.com/compare.php?from=3d2c9069dcafd0cbb641841aa3dd6e851fb7d760&to=b9513cdf2419704c7bb0c3a02a9ca06aae13d902&stat=instructions
Differential Revision: https://reviews.llvm.org/D103617
Added:
Modified:
llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
llvm/test/Transforms/InstCombine/sext-of-trunc-nsw.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
index ca5f02d20025a..f658dc35deea7 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -1526,13 +1526,20 @@ Instruction *InstCombinerImpl::visitSExt(SExtInst &CI) {
ShAmt);
}
- // If the input is a trunc from the destination type, then turn sext(trunc(x))
- // into shifts.
Value *X;
- if (match(Src, m_OneUse(m_Trunc(m_Value(X)))) && X->getType() == DestTy) {
- // sext (trunc X) --> ashr (shl X, C), C
- Constant *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
- return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShAmt), ShAmt);
+ if (match(Src, m_Trunc(m_Value(X)))) {
+ // If the input has more sign bits than bits truncated, then convert
+ // directly to final type.
+ unsigned XBitSize = X->getType()->getScalarSizeInBits();
+ if (ComputeNumSignBits(X, 0, &CI) > XBitSize - SrcBitSize)
+ return CastInst::CreateIntegerCast(X, DestTy, /* isSigned */ true);
+
+ // If input is a trunc from the destination type, then convert into shifts.
+ if (Src->hasOneUse() && X->getType() == DestTy) {
+ // sext (trunc X) --> ashr (shl X, C), C
+ Constant *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
+ return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShAmt), ShAmt);
+ }
}
if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
diff --git a/llvm/test/Transforms/InstCombine/sext-of-trunc-nsw.ll b/llvm/test/Transforms/InstCombine/sext-of-trunc-nsw.ll
index 8aa244971cf41..c2211deb6aa23 100644
--- a/llvm/test/Transforms/InstCombine/sext-of-trunc-nsw.ll
+++ b/llvm/test/Transforms/InstCombine/sext-of-trunc-nsw.ll
@@ -13,8 +13,7 @@ define i16 @t0(i8 %x) {
; CHECK-LABEL: @t0(
; CHECK-NEXT: [[A:%.*]] = ashr i8 [[X:%.*]], 5
; CHECK-NEXT: call void @use8(i8 [[A]])
-; CHECK-NEXT: [[B:%.*]] = trunc i8 [[A]] to i4
-; CHECK-NEXT: [[C:%.*]] = sext i4 [[B]] to i16
+; CHECK-NEXT: [[C:%.*]] = sext i8 [[A]] to i16
; CHECK-NEXT: ret i16 [[C]]
;
%a = ashr i8 %x, 5
@@ -28,8 +27,7 @@ define i16 @t1(i8 %x) {
; CHECK-LABEL: @t1(
; CHECK-NEXT: [[A:%.*]] = ashr i8 [[X:%.*]], 4
; CHECK-NEXT: call void @use8(i8 [[A]])
-; CHECK-NEXT: [[B:%.*]] = trunc i8 [[A]] to i4
-; CHECK-NEXT: [[C:%.*]] = sext i4 [[B]] to i16
+; CHECK-NEXT: [[C:%.*]] = sext i8 [[A]] to i16
; CHECK-NEXT: ret i16 [[C]]
;
%a = ashr i8 %x, 4
@@ -59,8 +57,7 @@ define <2 x i16> @t3_vec(<2 x i8> %x) {
; CHECK-LABEL: @t3_vec(
; CHECK-NEXT: [[A:%.*]] = ashr <2 x i8> [[X:%.*]], <i8 4, i8 4>
; CHECK-NEXT: call void @usevec(<2 x i8> [[A]])
-; CHECK-NEXT: [[B:%.*]] = trunc <2 x i8> [[A]] to <2 x i4>
-; CHECK-NEXT: [[C:%.*]] = sext <2 x i4> [[B]] to <2 x i16>
+; CHECK-NEXT: [[C:%.*]] = sext <2 x i8> [[A]] to <2 x i16>
; CHECK-NEXT: ret <2 x i16> [[C]]
;
%a = ashr <2 x i8> %x, <i8 4, i8 4>
@@ -91,7 +88,7 @@ define i16 @t5_extrause(i8 %x) {
; CHECK-NEXT: call void @use8(i8 [[A]])
; CHECK-NEXT: [[B:%.*]] = trunc i8 [[A]] to i4
; CHECK-NEXT: call void @use4(i4 [[B]])
-; CHECK-NEXT: [[C:%.*]] = sext i4 [[B]] to i16
+; CHECK-NEXT: [[C:%.*]] = sext i8 [[A]] to i16
; CHECK-NEXT: ret i16 [[C]]
;
%a = ashr i8 %x, 5
@@ -106,8 +103,7 @@ define i64 @narrow_source_matching_signbits(i32 %x) {
; CHECK-LABEL: @narrow_source_matching_signbits(
; CHECK-NEXT: [[M:%.*]] = and i32 [[X:%.*]], 7
; CHECK-NEXT: [[A:%.*]] = shl nsw i32 -1, [[M]]
-; CHECK-NEXT: [[B:%.*]] = trunc i32 [[A]] to i8
-; CHECK-NEXT: [[C:%.*]] = sext i8 [[B]] to i64
+; CHECK-NEXT: [[C:%.*]] = sext i32 [[A]] to i64
; CHECK-NEXT: ret i64 [[C]]
;
%m = and i32 %x, 7
@@ -117,6 +113,8 @@ define i64 @narrow_source_matching_signbits(i32 %x) {
ret i64 %c
}
+; negative test - not enough sign-bits
+
define i64 @narrow_source_not_matching_signbits(i32 %x) {
; CHECK-LABEL: @narrow_source_not_matching_signbits(
; CHECK-NEXT: [[M:%.*]] = and i32 [[X:%.*]], 8
@@ -136,8 +134,7 @@ define i24 @wide_source_matching_signbits(i32 %x) {
; CHECK-LABEL: @wide_source_matching_signbits(
; CHECK-NEXT: [[M:%.*]] = and i32 [[X:%.*]], 7
; CHECK-NEXT: [[A:%.*]] = shl nsw i32 -1, [[M]]
-; CHECK-NEXT: [[B:%.*]] = trunc i32 [[A]] to i8
-; CHECK-NEXT: [[C:%.*]] = sext i8 [[B]] to i24
+; CHECK-NEXT: [[C:%.*]] = trunc i32 [[A]] to i24
; CHECK-NEXT: ret i24 [[C]]
;
%m = and i32 %x, 7
@@ -147,6 +144,8 @@ define i24 @wide_source_matching_signbits(i32 %x) {
ret i24 %c
}
+; negative test - not enough sign-bits
+
define i24 @wide_source_not_matching_signbits(i32 %x) {
; CHECK-LABEL: @wide_source_not_matching_signbits(
; CHECK-NEXT: [[M2:%.*]] = and i32 [[X:%.*]], 8
@@ -165,9 +164,8 @@ define i24 @wide_source_not_matching_signbits(i32 %x) {
define i32 @same_source_matching_signbits(i32 %x) {
; CHECK-LABEL: @same_source_matching_signbits(
; CHECK-NEXT: [[M:%.*]] = and i32 [[X:%.*]], 7
-; CHECK-NEXT: [[TMP1:%.*]] = shl i32 -16777216, [[M]]
-; CHECK-NEXT: [[C:%.*]] = ashr exact i32 [[TMP1]], 24
-; CHECK-NEXT: ret i32 [[C]]
+; CHECK-NEXT: [[A:%.*]] = shl nsw i32 -1, [[M]]
+; CHECK-NEXT: ret i32 [[A]]
;
%m = and i32 %x, 7
%a = shl nsw i32 -1, %m
@@ -176,6 +174,8 @@ define i32 @same_source_matching_signbits(i32 %x) {
ret i32 %c
}
+; negative test - not enough sign-bits
+
define i32 @same_source_not_matching_signbits(i32 %x) {
; CHECK-LABEL: @same_source_not_matching_signbits(
; CHECK-NEXT: [[M2:%.*]] = and i32 [[X:%.*]], 8
@@ -196,8 +196,7 @@ define i32 @same_source_matching_signbits_extra_use(i32 %x) {
; CHECK-NEXT: [[A:%.*]] = shl nsw i32 -1, [[M]]
; CHECK-NEXT: [[B:%.*]] = trunc i32 [[A]] to i8
; CHECK-NEXT: call void @use8(i8 [[B]])
-; CHECK-NEXT: [[C:%.*]] = sext i8 [[B]] to i32
-; CHECK-NEXT: ret i32 [[C]]
+; CHECK-NEXT: ret i32 [[A]]
;
%m = and i32 %x, 7
%a = shl nsw i32 -1, %m
More information about the llvm-commits
mailing list