[llvm] [DAGCombiner][X86] Correctly clean up high bits in `combinei64TruncSrlAdd` (PR #128353)

Yingwei Zheng via llvm-commits llvm-commits at lists.llvm.org
Sat Feb 22 04:14:10 PST 2025


https://github.com/dtcxzyw created https://github.com/llvm/llvm-project/pull/128353

A counterexample for original implementation: https://alive2.llvm.org/ce/z/YowPZY
We should keep low `64 - shamt` bits instead of `shamt - 32`.
Proof: https://alive2.llvm.org/ce/z/z_jdHD



>From 270fc0ed253fc2d72e99038ff338f57e19233c44 Mon Sep 17 00:00:00 2001
From: Yingwei Zheng <dtcxzyw2333 at gmail.com>
Date: Sat, 22 Feb 2025 19:59:20 +0800
Subject: [PATCH 1/2] [DAGCombiner][X86] Add pre-commit tests. NFC.

---
 .../CodeGen/X86/combine-i64-trunc-srl-add.ll    | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

diff --git a/llvm/test/CodeGen/X86/combine-i64-trunc-srl-add.ll b/llvm/test/CodeGen/X86/combine-i64-trunc-srl-add.ll
index 41e1a24b239a6..3401f6ce78fd5 100644
--- a/llvm/test/CodeGen/X86/combine-i64-trunc-srl-add.ll
+++ b/llvm/test/CodeGen/X86/combine-i64-trunc-srl-add.ll
@@ -151,3 +151,20 @@ for.body:
 exit:
   ret i32 0
 }
+
+define i64 @pr128309(i64 %x) {
+; X64-LABEL: pr128309:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    movl %edi, %eax
+; X64-NEXT:    andl $18114, %eax # imm = 0x46C2
+; X64-NEXT:    addl $-65530, %eax # imm = 0xFFFF0006
+; X64-NEXT:    andl %edi, %eax
+; X64-NEXT:    retq
+entry:
+  %shl = shl i64 %x, 48
+  %and = and i64 %shl, 5098637728136822784
+  %add = add i64 %and, 1688849860263936
+  %lshr = lshr i64 %add, 48
+  %res = and i64 %lshr, %x
+  ret i64 %res
+}

>From 6c4f7bf4d66132cbc36ee602b1a987f276f677a3 Mon Sep 17 00:00:00 2001
From: Yingwei Zheng <dtcxzyw2333 at gmail.com>
Date: Sat, 22 Feb 2025 20:10:17 +0800
Subject: [PATCH 2/2] [DAGCombiner][X86] Correctly clean up high bits in
 `combinei64TruncSrlAdd`

---
 llvm/lib/Target/X86/X86ISelLowering.cpp       | 11 ++++------
 .../CodeGen/X86/combine-i64-trunc-srl-add.ll  | 20 +++++++++++--------
 2 files changed, 16 insertions(+), 15 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 95fd7f9b94282..0bd0dbeac2087 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -53712,23 +53712,20 @@ static SDValue combinei64TruncSrlAdd(SDValue N, EVT VT, SelectionDAG &DAG,
                                  m_ConstInt(SrlConst)))))
     return SDValue();
 
-  if (SrlConst.ule(32) || AddConst.lshr(SrlConst).shl(SrlConst) != AddConst)
+  if (SrlConst.ule(32) || AddConst.countr_zero() < SrlConst.getZExtValue())
     return SDValue();
 
   SDValue AddLHSSrl =
       DAG.getNode(ISD::SRL, DL, MVT::i64, AddLhs, N.getOperand(1));
   SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, VT, AddLHSSrl);
 
-  APInt NewAddConstVal =
-      (~((~AddConst).lshr(SrlConst))).trunc(VT.getSizeInBits());
+  APInt NewAddConstVal = AddConst.lshr(SrlConst).trunc(VT.getSizeInBits());
   SDValue NewAddConst = DAG.getConstant(NewAddConstVal, DL, VT);
   SDValue NewAddNode = DAG.getNode(ISD::ADD, DL, VT, Trunc, NewAddConst);
 
-  APInt CleanupSizeConstVal = (SrlConst - 32).zextOrTrunc(VT.getSizeInBits());
   EVT CleanUpVT =
-      EVT::getIntegerVT(*DAG.getContext(), CleanupSizeConstVal.getZExtValue());
-  SDValue CleanUp = DAG.getAnyExtOrTrunc(NewAddNode, DL, CleanUpVT);
-  return DAG.getAnyExtOrTrunc(CleanUp, DL, VT);
+      EVT::getIntegerVT(*DAG.getContext(), 64 - SrlConst.getZExtValue());
+  return DAG.getZeroExtendInReg(NewAddNode, DL, CleanUpVT);
 }
 
 /// Attempt to pre-truncate inputs to arithmetic ops if it will simplify
diff --git a/llvm/test/CodeGen/X86/combine-i64-trunc-srl-add.ll b/llvm/test/CodeGen/X86/combine-i64-trunc-srl-add.ll
index 3401f6ce78fd5..14992ca5bf488 100644
--- a/llvm/test/CodeGen/X86/combine-i64-trunc-srl-add.ll
+++ b/llvm/test/CodeGen/X86/combine-i64-trunc-srl-add.ll
@@ -7,8 +7,9 @@ define i1 @test_ult_trunc_add(i64 %x) {
 ; X64-LABEL: test_ult_trunc_add:
 ; X64:       # %bb.0:
 ; X64-NEXT:    shrq $48, %rdi
-; X64-NEXT:    addl $-65522, %edi # imm = 0xFFFF000E
-; X64-NEXT:    cmpl $3, %edi
+; X64-NEXT:    addl $14, %edi
+; X64-NEXT:    movzwl %di, %eax
+; X64-NEXT:    cmpl $3, %eax
 ; X64-NEXT:    setb %al
 ; X64-NEXT:    retq
   %add = add i64 %x, 3940649673949184
@@ -22,8 +23,9 @@ define i1 @test_ult_add(i64 %x) {
 ; X64-LABEL: test_ult_add:
 ; X64:       # %bb.0:
 ; X64-NEXT:    shrq $48, %rdi
-; X64-NEXT:    addl $-65522, %edi # imm = 0xFFFF000E
-; X64-NEXT:    cmpl $3, %edi
+; X64-NEXT:    addl $14, %edi
+; X64-NEXT:    movzwl %di, %eax
+; X64-NEXT:    cmpl $3, %eax
 ; X64-NEXT:    setb %al
 ; X64-NEXT:    retq
   %add = add i64 3940649673949184, %x
@@ -35,8 +37,9 @@ define i1 @test_ugt_trunc_add(i64 %x) {
 ; X64-LABEL: test_ugt_trunc_add:
 ; X64:       # %bb.0:
 ; X64-NEXT:    shrq $48, %rdi
-; X64-NEXT:    addl $-65522, %edi # imm = 0xFFFF000E
-; X64-NEXT:    cmpl $4, %edi
+; X64-NEXT:    addl $14, %edi
+; X64-NEXT:    movzwl %di, %eax
+; X64-NEXT:    cmpl $4, %eax
 ; X64-NEXT:    setae %al
 ; X64-NEXT:    retq
   %add = add i64 %x, 3940649673949184
@@ -116,7 +119,8 @@ define i32 @test_trunc_add(i64 %x) {
 ; X64-LABEL: test_trunc_add:
 ; X64:       # %bb.0:
 ; X64-NEXT:    shrq $48, %rdi
-; X64-NEXT:    leal -65522(%rdi), %eax
+; X64-NEXT:    addl $14, %edi
+; X64-NEXT:    movzwl %di, %eax
 ; X64-NEXT:    retq
   %add = add i64 %x, 3940649673949184
   %shr = lshr i64 %add, 48
@@ -157,7 +161,7 @@ define i64 @pr128309(i64 %x) {
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    movl %edi, %eax
 ; X64-NEXT:    andl $18114, %eax # imm = 0x46C2
-; X64-NEXT:    addl $-65530, %eax # imm = 0xFFFF0006
+; X64-NEXT:    addl $6, %eax
 ; X64-NEXT:    andl %edi, %eax
 ; X64-NEXT:    retq
 entry:



More information about the llvm-commits mailing list