[DAG] Remove OneUse restriction on sext when folding (shl (sext (add_…

…nsw x, c1)), c2) This patch remove the restriction for folding (shl (sext (add_nsw x, c1)), c2) -> (add (shl (sext x), c2), c1 << c2), and test case from dhrystone , see this link: https://godbolt.org/z/8zfa3rnad
llvm · Oct 20, 2023 · f53dabb · f53dabb
1 parent af3ead4
commit f53dabb
Show file tree

Hide file tree

Showing 5 changed files with 24 additions and 14 deletions.
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -10054,8 +10054,7 @@ SDValue DAGCombiner::visitSHL(SDNode *N) {
   // TODO: Should we limit this with isLegalAddImmediate?
   if (N0.getOpcode() == ISD::SIGN_EXTEND &&
       N0.getOperand(0).getOpcode() == ISD::ADD &&
-      N0.getOperand(0)->getFlags().hasNoSignedWrap() && N0->hasOneUse() &&
-      N0.getOperand(0)->hasOneUse() &&
+      N0.getOperand(0)->getFlags().hasNoSignedWrap() &&
       TLI.isDesirableToCommuteWithShift(N, Level)) {
     SDValue Add = N0.getOperand(0);
     SDLoc DL(N0);

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -3370,6 +3370,17 @@ X86TargetLowering::preferredShiftLegalizationStrategy(
                                                             ExpansionFactor);
 }
 
+bool X86TargetLowering::isDesirableToCommuteWithShift(
+    const SDNode *N, CombineLevel Level) const {
+  assert((N->getOpcode() == ISD::SHL || N->getOpcode() == ISD::SRA ||
+          N->getOpcode() == ISD::SRL) &&
+         "Expected shift op");
+  SDValue N0 = N->getOperand(0).getOpcode() == ISD::SIGN_EXTEND
+                   ? N->getOperand(0)->getOperand(0)
+                   : N->getOperand(0);
+  return N0.hasOneUse();
+}
+
 bool X86TargetLowering::shouldSplatInsEltVarIndex(EVT VT) const {
   // Any legal vector type can be splatted more efficiently than
   // loading/spilling from memory.

diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -1172,6 +1172,9 @@ namespace llvm {
     preferredShiftLegalizationStrategy(SelectionDAG &DAG, SDNode *N,
                                        unsigned ExpansionFactor) const override;
 
+    bool isDesirableToCommuteWithShift(const SDNode *N,
+                                       CombineLevel Level) const override;
+
     bool shouldSplatInsEltVarIndex(EVT VT) const override;
 
     bool shouldConvertFpToSat(unsigned Op, EVT FPVT, EVT VT) const override {

diff --git a/llvm/test/CodeGen/RISCV/riscv-shifted-extend.ll b/llvm/test/CodeGen/RISCV/riscv-shifted-extend.ll
@@ -5,12 +5,10 @@
 define void @test(ptr nocapture noundef writeonly %array1, i32 noundef signext %a, i32 noundef signext %b) {
 ; RV64-LABEL: test:
 ; RV64:       # %bb.0: # %entry
-; RV64-NEXT:    addiw a3, a1, 5
-; RV64-NEXT:    slli a4, a3, 2
-; RV64-NEXT:    add a4, a0, a4
-; RV64-NEXT:    sw a2, 0(a4)
+; RV64-NEXT:    addi a3, a1, 5
 ; RV64-NEXT:    slli a1, a1, 2
 ; RV64-NEXT:    add a0, a1, a0
+; RV64-NEXT:    sw a2, 20(a0)
 ; RV64-NEXT:    sw a2, 24(a0)
 ; RV64-NEXT:    sw a3, 140(a0)
 ; RV64-NEXT:    ret
@@ -34,18 +32,16 @@ entry:
 define void @test1(ptr nocapture noundef %array1, i32 noundef signext %a, i32 noundef signext %b, i32 noundef signext %x) {
 ; RV64-LABEL: test1:
 ; RV64:       # %bb.0: # %entry
-; RV64-NEXT:    addiw a4, a1, 5
-; RV64-NEXT:    slli a5, a4, 2
-; RV64-NEXT:    add a5, a0, a5
-; RV64-NEXT:    mv a6, a4
+; RV64-NEXT:    addi a4, a1, 5
+; RV64-NEXT:    mv a5, a4
 ; RV64-NEXT:    bgtz a3, .LBB1_2
 ; RV64-NEXT:  # %bb.1: # %entry
-; RV64-NEXT:    mv a6, a2
+; RV64-NEXT:    mv a5, a2
 ; RV64-NEXT:  .LBB1_2: # %entry
-; RV64-NEXT:    sw a6, 0(a5)
 ; RV64-NEXT:    slli a1, a1, 2
 ; RV64-NEXT:    add a0, a1, a0
-; RV64-NEXT:    sw a6, 24(a0)
+; RV64-NEXT:    sw a5, 20(a0)
+; RV64-NEXT:    sw a5, 24(a0)
 ; RV64-NEXT:    sw a4, 140(a0)
 ; RV64-NEXT:    ret
 entry:

diff --git a/llvm/test/CodeGen/X86/pr65895.ll b/llvm/test/CodeGen/X86/pr65895.ll
@@ -21,10 +21,11 @@ define i32 @PR65895() {
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    jmp .LBB0_2
 ; CHECK-NEXT:  .LBB0_3: # %for.end
+; CHECK-NEXT:    movzbl %al, %ecx
 ; CHECK-NEXT:    addb $-3, %al
 ; CHECK-NEXT:    movsbl %al, %eax
 ; CHECK-NEXT:    movl %eax, d(%rip)
-; CHECK-NEXT:    leal 247(%rax,%rax,2), %eax
+; CHECK-NEXT:    leal 241(%rax,%rcx,2), %eax
 ; CHECK-NEXT:    movb $1, c(%rip)
 ; CHECK-NEXT:    movsbq %al, %rax
 ; CHECK-NEXT:    movq %rax, e(%rip)