[InstCombine] Add transforms for icmp uPred (trunc x),(truncOrZext(y)) -> icmp uPred x,y
#71309
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-llvm-transforms Author: None (goldsteinn) Changes
Full diff: https://github.com/llvm/llvm-project/pull/71309.diff 3 Files Affected:
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index f06657c8cd7633d..ee749e1b0537334 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -1545,6 +1545,53 @@ Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp,
return nullptr;
}
+/// Fold icmp (trunc X), (trunc Y).
+/// Fold icmp (trunc X), (zext Y).
+static Instruction *foldICmpTruncWithTruncOrExt(ICmpInst &Cmp,
+ InstCombinerImpl &IC,
+ const SimplifyQuery &Q) {
+ if (!Cmp.isEquality() && !Cmp.isUnsigned())
+ return nullptr;
+
+ Value *X, *Y;
+ ICmpInst::Predicate Pred;
+ bool YIsZext = false;
+ // Try to match icmp (trunc X), (trunc Y)
+ if (match(&Cmp, m_ICmp(Pred, m_Trunc(m_Value(X)), m_Trunc(m_Value(Y))))) {
+ if (X->getType() != Y->getType() &&
+ (!Cmp.getOperand(0)->hasOneUse() || !Cmp.getOperand(1)->hasOneUse()))
+ return nullptr;
+ }
+ // Try to match icmp (trunc X), (zext Y)
+ else if (match(&Cmp, m_c_ICmp(Pred, m_Trunc(m_Value(X)),
+ m_OneUse(m_ZExt(m_Value(Y)))))) {
+
+ YIsZext = true;
+ } else {
+ return nullptr;
+ }
+
+ Type *TruncTy = Cmp.getOperand(0)->getType();
+ Type *BaseTy = X->getType();
+
+ unsigned TruncBits = TruncTy->getScalarSizeInBits();
+
+ // Check if the trunc is unneeded.
+ KnownBits KnownX = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
+ if (KnownX.Zero.countl_one() < KnownX.getBitWidth() - TruncBits)
+ return nullptr;
+
+ if (!YIsZext) {
+ // If Y is also a trunc, make sure it is unneeded.
+ KnownBits KnownY = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
+ if (KnownY.Zero.countl_one() < KnownY.getBitWidth() - TruncBits)
+ return nullptr;
+ }
+
+ Value *NewY = IC.Builder.CreateZExtOrTrunc(Y, BaseTy);
+ return IC.replaceInstUsesWith(Cmp, IC.Builder.CreateICmp(Pred, X, NewY));
+}
+
/// Fold icmp (xor X, Y), C.
Instruction *InstCombinerImpl::foldICmpXorConstant(ICmpInst &Cmp,
BinaryOperator *Xor,
@@ -6907,6 +6954,9 @@ Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
if (Instruction *Res = foldICmpUsingKnownBits(I))
return Res;
+ if (Instruction *Res = foldICmpTruncWithTruncOrExt(I, *this, Q))
+ return Res;
+
// Test if the ICmpInst instruction is used exclusively by a select as
// part of a minimum or maximum operation. If so, refrain from doing
// any other folding. This helps out other analyses which understand
diff --git a/llvm/test/Transforms/InstCombine/eq-of-parts.ll b/llvm/test/Transforms/InstCombine/eq-of-parts.ll
index 57b15ae3b96e66e..217e37b85933949 100644
--- a/llvm/test/Transforms/InstCombine/eq-of-parts.ll
+++ b/llvm/test/Transforms/InstCombine/eq-of-parts.ll
@@ -584,17 +584,8 @@ define i1 @eq_21_not_adjacent(i32 %x, i32 %y) {
define i1 @eq_shift_in_zeros(i32 %x, i32 %y) {
; CHECK-LABEL: @eq_shift_in_zeros(
-; CHECK-NEXT: [[X_321:%.*]] = lshr i32 [[X:%.*]], 8
-; CHECK-NEXT: [[X_1:%.*]] = trunc i32 [[X_321]] to i8
-; CHECK-NEXT: [[X_32:%.*]] = lshr i32 [[X]], 16
-; CHECK-NEXT: [[X_2:%.*]] = trunc i32 [[X_32]] to i24
-; CHECK-NEXT: [[Y_321:%.*]] = lshr i32 [[Y:%.*]], 8
-; CHECK-NEXT: [[Y_1:%.*]] = trunc i32 [[Y_321]] to i8
-; CHECK-NEXT: [[Y_32:%.*]] = lshr i32 [[Y]], 16
-; CHECK-NEXT: [[Y_2:%.*]] = trunc i32 [[Y_32]] to i24
-; CHECK-NEXT: [[C_1:%.*]] = icmp eq i8 [[X_1]], [[Y_1]]
-; CHECK-NEXT: [[C_2:%.*]] = icmp eq i24 [[X_2]], [[Y_2]]
-; CHECK-NEXT: [[C_210:%.*]] = and i1 [[C_2]], [[C_1]]
+; CHECK-NEXT: [[C_210_UNSHIFTED:%.*]] = xor i32 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[C_210:%.*]] = icmp ult i32 [[C_210_UNSHIFTED]], 256
; CHECK-NEXT: ret i1 [[C_210]]
;
%x.321 = lshr i32 %x, 8
@@ -1249,17 +1240,8 @@ define i1 @ne_21_not_adjacent(i32 %x, i32 %y) {
define i1 @ne_shift_in_zeros(i32 %x, i32 %y) {
; CHECK-LABEL: @ne_shift_in_zeros(
-; CHECK-NEXT: [[X_321:%.*]] = lshr i32 [[X:%.*]], 8
-; CHECK-NEXT: [[X_1:%.*]] = trunc i32 [[X_321]] to i8
-; CHECK-NEXT: [[X_32:%.*]] = lshr i32 [[X]], 16
-; CHECK-NEXT: [[X_2:%.*]] = trunc i32 [[X_32]] to i24
-; CHECK-NEXT: [[Y_321:%.*]] = lshr i32 [[Y:%.*]], 8
-; CHECK-NEXT: [[Y_1:%.*]] = trunc i32 [[Y_321]] to i8
-; CHECK-NEXT: [[Y_32:%.*]] = lshr i32 [[Y]], 16
-; CHECK-NEXT: [[Y_2:%.*]] = trunc i32 [[Y_32]] to i24
-; CHECK-NEXT: [[C_1:%.*]] = icmp ne i8 [[X_1]], [[Y_1]]
-; CHECK-NEXT: [[C_2:%.*]] = icmp ne i24 [[X_2]], [[Y_2]]
-; CHECK-NEXT: [[C_210:%.*]] = or i1 [[C_2]], [[C_1]]
+; CHECK-NEXT: [[C_210_UNSHIFTED:%.*]] = xor i32 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[C_210:%.*]] = icmp ugt i32 [[C_210_UNSHIFTED]], 255
; CHECK-NEXT: ret i1 [[C_210]]
;
%x.321 = lshr i32 %x, 8
diff --git a/llvm/test/Transforms/InstCombine/icmp-of-trunc-ext.ll b/llvm/test/Transforms/InstCombine/icmp-of-trunc-ext.ll
new file mode 100644
index 000000000000000..0673f013790d5c8
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/icmp-of-trunc-ext.ll
@@ -0,0 +1,172 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -passes=instcombine < %s | FileCheck %s
+
+declare void @llvm.assume(i1)
+declare void @use(i16)
+define i1 @icmp_trunc_x_trunc_y(i32 %x, i32 %y) {
+; CHECK-LABEL: @icmp_trunc_x_trunc_y(
+; CHECK-NEXT: [[X_LB_ONLY:%.*]] = icmp ult i32 [[X:%.*]], 65536
+; CHECK-NEXT: [[Y_LB_ONLY:%.*]] = icmp ult i32 [[Y:%.*]], 65536
+; CHECK-NEXT: call void @llvm.assume(i1 [[X_LB_ONLY]])
+; CHECK-NEXT: call void @llvm.assume(i1 [[Y_LB_ONLY]])
+; CHECK-NEXT: [[R:%.*]] = icmp eq i32 [[X]], [[Y]]
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %x_lb_only = icmp ult i32 %x, 65536
+ %y_lb_only = icmp ult i32 %y, 65536
+ call void @llvm.assume(i1 %x_lb_only)
+ call void @llvm.assume(i1 %y_lb_only)
+ %x16 = trunc i32 %x to i16
+ %y16 = trunc i32 %y to i16
+ %r = icmp eq i16 %x16, %y16
+ ret i1 %r
+}
+
+define i1 @icmp_trunc_x_trunc_y_2(i32 %x, i64 %y) {
+; CHECK-LABEL: @icmp_trunc_x_trunc_y_2(
+; CHECK-NEXT: [[X_LB_ONLY:%.*]] = icmp ult i32 [[X:%.*]], 12345
+; CHECK-NEXT: [[Y_LB_ONLY:%.*]] = icmp ult i64 [[Y:%.*]], 65536
+; CHECK-NEXT: call void @llvm.assume(i1 [[X_LB_ONLY]])
+; CHECK-NEXT: call void @llvm.assume(i1 [[Y_LB_ONLY]])
+; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[X]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp ugt i64 [[TMP1]], [[Y]]
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %x_lb_only = icmp ult i32 %x, 12345
+ %y_lb_only = icmp ult i64 %y, 65536
+ call void @llvm.assume(i1 %x_lb_only)
+ call void @llvm.assume(i1 %y_lb_only)
+ %x16 = trunc i32 %x to i16
+ %y16 = trunc i64 %y to i16
+ %r = icmp ult i16 %y16, %x16
+ ret i1 %r
+}
+
+define i1 @icmp_trunc_x_trunc_y_3(i64 %x, i32 %y) {
+; CHECK-LABEL: @icmp_trunc_x_trunc_y_3(
+; CHECK-NEXT: [[X_LB_ONLY:%.*]] = icmp ult i64 [[X:%.*]], 123
+; CHECK-NEXT: [[Y_LB_ONLY:%.*]] = icmp ult i32 [[Y:%.*]], 256
+; CHECK-NEXT: call void @llvm.assume(i1 [[X_LB_ONLY]])
+; CHECK-NEXT: call void @llvm.assume(i1 [[Y_LB_ONLY]])
+; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[X]] to i32
+; CHECK-NEXT: [[R:%.*]] = icmp uge i32 [[TMP1]], [[Y]]
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %x_lb_only = icmp ult i64 %x, 123
+ %y_lb_only = icmp ult i32 %y, 256
+ call void @llvm.assume(i1 %x_lb_only)
+ call void @llvm.assume(i1 %y_lb_only)
+ %xi8 = trunc i64 %x to i8
+ %yi8 = trunc i32 %y to i8
+ %r = icmp ule i8 %yi8, %xi8
+ ret i1 %r
+}
+
+define i1 @icmp_trunc_x_trunc_y_fail_maybe_dirty_upper(i32 %x, i32 %y) {
+; CHECK-LABEL: @icmp_trunc_x_trunc_y_fail_maybe_dirty_upper(
+; CHECK-NEXT: [[X_LB_ONLY:%.*]] = icmp ult i32 [[X:%.*]], 65536
+; CHECK-NEXT: [[Y_LB_ONLY:%.*]] = icmp ult i32 [[Y:%.*]], 65537
+; CHECK-NEXT: call void @llvm.assume(i1 [[X_LB_ONLY]])
+; CHECK-NEXT: call void @llvm.assume(i1 [[Y_LB_ONLY]])
+; CHECK-NEXT: [[X16:%.*]] = trunc i32 [[X]] to i16
+; CHECK-NEXT: [[Y16:%.*]] = trunc i32 [[Y]] to i16
+; CHECK-NEXT: [[R:%.*]] = icmp ne i16 [[X16]], [[Y16]]
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %x_lb_only = icmp ult i32 %x, 65536
+ %y_lb_only = icmp ult i32 %y, 65537
+ call void @llvm.assume(i1 %x_lb_only)
+ call void @llvm.assume(i1 %y_lb_only)
+ %x16 = trunc i32 %x to i16
+ %y16 = trunc i32 %y to i16
+ %r = icmp ne i16 %x16, %y16
+ ret i1 %r
+}
+
+define i1 @icmp_trunc_x_trunc_y_fail_maybe_dirty_upper_2(i32 %x, i32 %y) {
+; CHECK-LABEL: @icmp_trunc_x_trunc_y_fail_maybe_dirty_upper_2(
+; CHECK-NEXT: [[X_LB_ONLY:%.*]] = icmp slt i32 [[X:%.*]], 65536
+; CHECK-NEXT: [[Y_LB_ONLY:%.*]] = icmp ult i32 [[Y:%.*]], 65536
+; CHECK-NEXT: call void @llvm.assume(i1 [[X_LB_ONLY]])
+; CHECK-NEXT: call void @llvm.assume(i1 [[Y_LB_ONLY]])
+; CHECK-NEXT: [[X16:%.*]] = trunc i32 [[X]] to i16
+; CHECK-NEXT: [[Y16:%.*]] = trunc i32 [[Y]] to i16
+; CHECK-NEXT: [[R:%.*]] = icmp ne i16 [[X16]], [[Y16]]
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %x_lb_only = icmp slt i32 %x, 65536
+ %y_lb_only = icmp ult i32 %y, 65536
+ call void @llvm.assume(i1 %x_lb_only)
+ call void @llvm.assume(i1 %y_lb_only)
+ %x16 = trunc i32 %x to i16
+ %y16 = trunc i32 %y to i16
+ %r = icmp ne i16 %x16, %y16
+ ret i1 %r
+}
+
+define i1 @icmp_trunc_x_zext_y(i32 %x, i8 %y) {
+; CHECK-LABEL: @icmp_trunc_x_zext_y(
+; CHECK-NEXT: [[X_LB_ONLY:%.*]] = icmp ult i32 [[X:%.*]], 65536
+; CHECK-NEXT: call void @llvm.assume(i1 [[X_LB_ONLY]])
+; CHECK-NEXT: [[TMP1:%.*]] = zext i8 [[Y:%.*]] to i32
+; CHECK-NEXT: [[R:%.*]] = icmp ult i32 [[TMP1]], [[X]]
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %x_lb_only = icmp ult i32 %x, 65536
+ call void @llvm.assume(i1 %x_lb_only)
+ %x16 = trunc i32 %x to i16
+ %y16 = zext i8 %y to i16
+ %r = icmp ugt i16 %x16, %y16
+ ret i1 %r
+}
+
+define i1 @icmp_trunc_x_zext_y_2(i64 %x, i8 %y) {
+; CHECK-LABEL: @icmp_trunc_x_zext_y_2(
+; CHECK-NEXT: [[X_LB_ONLY:%.*]] = icmp ult i64 [[X:%.*]], 65536
+; CHECK-NEXT: call void @llvm.assume(i1 [[X_LB_ONLY]])
+; CHECK-NEXT: [[TMP1:%.*]] = zext i8 [[Y:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp uge i64 [[TMP1]], [[X]]
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %x_lb_only = icmp ult i64 %x, 65536
+ call void @llvm.assume(i1 %x_lb_only)
+ %x16 = trunc i64 %x to i16
+ %y16 = zext i8 %y to i16
+ %r = icmp uge i16 %y16, %x16
+ ret i1 %r
+}
+
+define i1 @icmp_trunc_x_zext_y_3(i8 %x, i64 %y) {
+; CHECK-LABEL: @icmp_trunc_x_zext_y_3(
+; CHECK-NEXT: [[Y_LB_ONLY:%.*]] = icmp ult i64 [[Y:%.*]], 65536
+; CHECK-NEXT: call void @llvm.assume(i1 [[Y_LB_ONLY]])
+; CHECK-NEXT: [[TMP1:%.*]] = zext i8 [[X:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp ne i64 [[TMP1]], [[Y]]
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %y_lb_only = icmp ult i64 %y, 65536
+ call void @llvm.assume(i1 %y_lb_only)
+ %x16 = zext i8 %x to i16
+ %y16 = trunc i64 %y to i16
+ %r = icmp ne i16 %y16, %x16
+ ret i1 %r
+}
+
+define i1 @icmp_trunc_x_zext_y_fail_multiuse(i32 %x, i8 %y) {
+; CHECK-LABEL: @icmp_trunc_x_zext_y_fail_multiuse(
+; CHECK-NEXT: [[X_LB_ONLY:%.*]] = icmp ult i32 [[X:%.*]], 65536
+; CHECK-NEXT: call void @llvm.assume(i1 [[X_LB_ONLY]])
+; CHECK-NEXT: [[X16:%.*]] = trunc i32 [[X]] to i16
+; CHECK-NEXT: [[Y16:%.*]] = zext i8 [[Y:%.*]] to i16
+; CHECK-NEXT: call void @use(i16 [[Y16]])
+; CHECK-NEXT: [[R:%.*]] = icmp ule i16 [[X16]], [[Y16]]
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %x_lb_only = icmp ult i32 %x, 65536
+ call void @llvm.assume(i1 %x_lb_only)
+ %x16 = trunc i32 %x to i16
+ %y16 = zext i8 %y to i16
+ call void @use(i16 %y16)
+ %r = icmp ule i16 %x16, %y16
+ ret i1 %r
+}
|
dtcxzyw
changed the title
icmp uPred (trunc x),(truncOrZext(y)) -> icmp uPred x,y
|
Should this be guarded by a shouldChangeType() heuristic? Say you're doing |
dtcxzyw
reviewed
Nov 5, 2023
nikic
reviewed
Nov 5, 2023
goldsteinn
force-pushed
the
goldsteinn/trunc
branch
from
2636ebe
to
eab767a
Compare
dtcxzyw
reviewed
Nov 8, 2023
nikic
reviewed
Nov 14, 2023
goldsteinn
force-pushed
the
goldsteinn/trunc
branch
from
eab767a
to
2343dfc
Compare
…y)))`->`(icmp uPred x,y)` Three transforms (all commutative): https://alive2.llvm.org/ce/z/Bc-nh4
goldsteinn
force-pushed
the
goldsteinn/trunc
branch
from
2343dfc
to
ac60c78
Compare
Seems like minimal affect: https://llvm-compile-time-tracker.com/compare.php?from=2b71f91b06ad4f5a0c54725b06283fd731620b92&to=2343dfccca46d158de240fca8e0588ae4cc46203&stat=instructions:u Slight regression in ThinLTO. Larger improvement in stage2 + O3. |
sr-tream
pushed a commit
to sr-tream/llvm-project
that referenced
this pull request
Nov 20, 2023
…y)))`->`(icmp uPred x,y)` Three transforms (all commutative): https://alive2.llvm.org/ce/z/Bc-nh4 Closes llvm#71309
zahiraam
pushed a commit
to zahiraam/llvm-project
that referenced
this pull request
Nov 20, 2023
…y)))`->`(icmp uPred x,y)` Three transforms (all commutative): https://alive2.llvm.org/ce/z/Bc-nh4 Closes llvm#71309
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
(icmp eq/ne trunc(x), truncOrZext(y)); NFC(icmp uPred (trunc x),(truncOrZext(y)))->(icmp uPred x,y)