[llvm] 9594c0d - [InstCombine] Move sext(trunc(...)) combines before EvaluateInDifferentType (#227321)

via llvm-commits llvm-commits at lists.llvm.org
Wed Oct 7 15:56:30 PDT 2026


Author: Luke Lau
Date: 2026-10-07T23:56:14+01:00
New Revision: 9594c0d77e42ceeeeb1d75c87ed76f9954067b22

URL: https://github.com/llvm/llvm-project/commit/9594c0d77e42ceeeeb1d75c87ed76f9954067b22
DIFF: https://github.com/llvm/llvm-project/commit/9594c0d77e42ceeeeb1d75c87ed76f9954067b22.diff

LOG: [InstCombine] Move sext(trunc(...)) combines before EvaluateInDifferentType (#227321)

EvaluateInDifferentType loses some of the sign bit information which the
sext(trunc(...)) combines can use to remove intermediate casts, so run
the latter earlier. This helps towards fixing a codegen pattern in x264

Tests written with opus 5.5

Added: 
    llvm/test/Transforms/InstCombine/AArch64/lshr-trunc-sext-to-ashr-sext.ll
    llvm/test/Transforms/InstCombine/RISCV/lshr-trunc-sext-to-ashr-sext.ll
    llvm/test/Transforms/InstCombine/X86/lshr-trunc-sext-to-ashr-sext.ll

Modified: 
    llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
index 0275bf52a13f25..f3741e9b16b94e 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -1945,37 +1945,8 @@ Instruction *InstCombinerImpl::visitSExt(SExtInst &Sext) {
     return CI;
   }
 
-  // Try to extend the entire expression tree to the wide destination type.
-  bool ShouldExtendExpression = true;
-  Value *TruncSrc = nullptr;
-  // It is not desirable to extend expression in the trunc + sext pattern when
-  // destination type is narrower than original (pre-trunc) type.
-  if (match(Src, m_Trunc(m_Value(TruncSrc))))
-    if (TruncSrc->getType()->getScalarSizeInBits() > DestBitSize)
-      ShouldExtendExpression = false;
-  if (ShouldExtendExpression && shouldChangeType(SrcTy, DestTy) &&
-      TypeEvaluationHelper::canEvaluateSExtd(Src, DestTy)) {
-    // Okay, we can transform this!  Insert the new expression now.
-    LLVM_DEBUG(
-        dbgs() << "ICE: EvaluateInDifferentType converting expression type"
-                  " to avoid sign extend: "
-               << Sext << '\n');
-    Value *Res = EvaluateInDifferentType(Src, DestTy, true);
-    assert(Res->getType() == DestTy);
-
-    // If the high bits are already filled with sign bit, just replace this
-    // cast with the result.
-    if (ComputeNumSignBits(Res, &Sext) > DestBitSize - SrcBitSize)
-      return replaceInstUsesWith(Sext, Res);
-
-    // We need to emit a shl + ashr to do the sign extend.
-    Value *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
-    return BinaryOperator::CreateAShr(Builder.CreateShl(Res, ShAmt, "sext"),
-                                      ShAmt);
-  }
-
-  Value *X = TruncSrc;
-  if (X) {
+  Value *X;
+  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();
@@ -2012,6 +1983,35 @@ Instruction *InstCombinerImpl::visitSExt(SExtInst &Sext) {
     }
   }
 
+  // Try to extend the entire expression tree to the wide destination type.
+  bool ShouldExtendExpression = true;
+  Value *TruncSrc = nullptr;
+  // It is not desirable to extend expression in the trunc + sext pattern when
+  // destination type is narrower than original (pre-trunc) type.
+  if (match(Src, m_Trunc(m_Value(TruncSrc))))
+    if (TruncSrc->getType()->getScalarSizeInBits() > DestBitSize)
+      ShouldExtendExpression = false;
+  if (ShouldExtendExpression && shouldChangeType(SrcTy, DestTy) &&
+      TypeEvaluationHelper::canEvaluateSExtd(Src, DestTy)) {
+    // Okay, we can transform this!  Insert the new expression now.
+    LLVM_DEBUG(
+        dbgs() << "ICE: EvaluateInDifferentType converting expression type"
+                  " to avoid sign extend: "
+               << Sext << '\n');
+    Value *Res = EvaluateInDifferentType(Src, DestTy, true);
+    assert(Res->getType() == DestTy);
+
+    // If the high bits are already filled with sign bit, just replace this
+    // cast with the result.
+    if (ComputeNumSignBits(Res, &Sext) > DestBitSize - SrcBitSize)
+      return replaceInstUsesWith(Sext, Res);
+
+    // We need to emit a shl + ashr to do the sign extend.
+    Value *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
+    return BinaryOperator::CreateAShr(Builder.CreateShl(Res, ShAmt, "sext"),
+                                      ShAmt);
+  }
+
   if (auto *Cmp = dyn_cast<ICmpInst>(Src))
     return transformSExtICmp(Cmp, Sext);
 

diff  --git a/llvm/test/Transforms/InstCombine/AArch64/lshr-trunc-sext-to-ashr-sext.ll b/llvm/test/Transforms/InstCombine/AArch64/lshr-trunc-sext-to-ashr-sext.ll
new file mode 100644
index 00000000000000..6451486688a123
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/AArch64/lshr-trunc-sext-to-ashr-sext.ll
@@ -0,0 +1,32 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt < %s -p instcombine -mtriple aarch64 -S | FileCheck %s
+
+define i64 @t0(i32 %x) {
+; CHECK-LABEL: define i64 @t0(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = ashr i32 [[X]], 16
+; CHECK-NEXT:    [[C:%.*]] = sext i32 [[TMP1]] to i64
+; CHECK-NEXT:    ret i64 [[C]]
+;
+  %a = lshr i32 %x, 16
+  %b = trunc i32 %a to i16
+  %c = sext i16 %b to i64
+  ret i64 %c
+}
+
+define i32 @t1(i16 %x) {
+; CHECK-LABEL: define i32 @t1(
+; CHECK-SAME: i16 [[X:%.*]]) {
+; CHECK-NEXT:    [[S:%.*]] = sext i16 [[X]] to i32
+; CHECK-NEXT:    [[Y:%.*]] = shl nsw i32 [[S]], 10
+; CHECK-NEXT:    [[TMP1:%.*]] = add nsw i32 [[Y]], 32768
+; CHECK-NEXT:    [[C:%.*]] = ashr i32 [[TMP1]], 16
+; CHECK-NEXT:    ret i32 [[C]]
+;
+  %s = sext i16 %x to i32
+  %y = add i32 %s, 32
+  %a = ashr i32 %y, 6
+  %b = trunc i32 %a to i16
+  %c = sext i16 %b to i32
+  ret i32 %c
+}

diff  --git a/llvm/test/Transforms/InstCombine/RISCV/lshr-trunc-sext-to-ashr-sext.ll b/llvm/test/Transforms/InstCombine/RISCV/lshr-trunc-sext-to-ashr-sext.ll
new file mode 100644
index 00000000000000..09d35c2cfa2c3b
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/RISCV/lshr-trunc-sext-to-ashr-sext.ll
@@ -0,0 +1,32 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt < %s -p instcombine -mtriple riscv64 -S | FileCheck %s
+
+define i64 @t0(i32 %x) {
+; CHECK-LABEL: define i64 @t0(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = ashr i32 [[X]], 16
+; CHECK-NEXT:    [[C:%.*]] = sext i32 [[TMP1]] to i64
+; CHECK-NEXT:    ret i64 [[C]]
+;
+  %a = lshr i32 %x, 16
+  %b = trunc i32 %a to i16
+  %c = sext i16 %b to i64
+  ret i64 %c
+}
+
+define i32 @t1(i16 %x) {
+; CHECK-LABEL: define i32 @t1(
+; CHECK-SAME: i16 [[X:%.*]]) {
+; CHECK-NEXT:    [[S:%.*]] = sext i16 [[X]] to i32
+; CHECK-NEXT:    [[Y:%.*]] = shl nsw i32 [[S]], 10
+; CHECK-NEXT:    [[TMP1:%.*]] = add nsw i32 [[Y]], 32768
+; CHECK-NEXT:    [[C:%.*]] = ashr i32 [[TMP1]], 16
+; CHECK-NEXT:    ret i32 [[C]]
+;
+  %s = sext i16 %x to i32
+  %y = add i32 %s, 32
+  %a = ashr i32 %y, 6
+  %b = trunc i32 %a to i16
+  %c = sext i16 %b to i32
+  ret i32 %c
+}

diff  --git a/llvm/test/Transforms/InstCombine/X86/lshr-trunc-sext-to-ashr-sext.ll b/llvm/test/Transforms/InstCombine/X86/lshr-trunc-sext-to-ashr-sext.ll
new file mode 100644
index 00000000000000..852b6bc0daefea
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/X86/lshr-trunc-sext-to-ashr-sext.ll
@@ -0,0 +1,32 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt < %s -p instcombine -mtriple x86_64 -S | FileCheck %s
+
+define i64 @t0(i32 %x) {
+; CHECK-LABEL: define i64 @t0(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = ashr i32 [[X]], 16
+; CHECK-NEXT:    [[C:%.*]] = sext i32 [[TMP1]] to i64
+; CHECK-NEXT:    ret i64 [[C]]
+;
+  %a = lshr i32 %x, 16
+  %b = trunc i32 %a to i16
+  %c = sext i16 %b to i64
+  ret i64 %c
+}
+
+define i32 @t1(i16 %x) {
+; CHECK-LABEL: define i32 @t1(
+; CHECK-SAME: i16 [[X:%.*]]) {
+; CHECK-NEXT:    [[S:%.*]] = sext i16 [[X]] to i32
+; CHECK-NEXT:    [[Y:%.*]] = shl nsw i32 [[S]], 10
+; CHECK-NEXT:    [[TMP1:%.*]] = add nsw i32 [[Y]], 32768
+; CHECK-NEXT:    [[C:%.*]] = ashr i32 [[TMP1]], 16
+; CHECK-NEXT:    ret i32 [[C]]
+;
+  %s = sext i16 %x to i32
+  %y = add i32 %s, 32
+  %a = ashr i32 %y, 6
+  %b = trunc i32 %a to i16
+  %c = sext i16 %b to i32
+  ret i32 %c
+}


        


More information about the llvm-commits mailing list