[llvm] [SDAG] fix miss opt: shl nuw + zext adds unnecessary masking (PR #172046)

via llvm-commits llvm-commits at lists.llvm.org
Fri Dec 12 22:24:24 PST 2025


https://github.com/actinks updated https://github.com/llvm/llvm-project/pull/172046

>From a80b806e05cf32ce5cb0426d08d0d439a14d6432 Mon Sep 17 00:00:00 2001
From: actink <actink at 163.com>
Date: Fri, 12 Dec 2025 23:08:25 +0800
Subject: [PATCH 1/2] [SDAG] fix miss opt: shl nuw + zext adds unnecessary
 masking

close: #171750
---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |  6 ++-
 .../X86/fold-int-pow2-with-fmul-or-fdiv.ll    |  4 --
 llvm/test/CodeGen/X86/known-never-zero.ll     |  4 --
 llvm/test/CodeGen/X86/known-pow2.ll           |  3 +-
 llvm/test/CodeGen/X86/pr171750.ll             | 47 +++++++++++++++++++
 llvm/test/CodeGen/X86/pr89877.ll              |  2 -
 6 files changed, 52 insertions(+), 14 deletions(-)
 create mode 100644 llvm/test/CodeGen/X86/pr171750.ll

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 6a99d4e29b64f..a0212774d8f17 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -1673,8 +1673,10 @@ SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
 
     SDLoc DL(Op);
     SDValue N1 = Op.getOperand(1);
-    SDValue RV =
-        DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, N0, N1));
+    SDValue POp = DAG.getNode(Opc, DL, PVT, N0, N1);
+    if (Opc == ISD::SHL && Op->getFlags().hasNoUnsignedWrap())
+      POp = DAG.getNode(ISD::AssertZext, DL, PVT, POp, DAG.getValueType(VT));
+    SDValue RV = DAG.getNode(ISD::TRUNCATE, DL, VT, POp);
 
     if (Replace)
       ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
diff --git a/llvm/test/CodeGen/X86/fold-int-pow2-with-fmul-or-fdiv.ll b/llvm/test/CodeGen/X86/fold-int-pow2-with-fmul-or-fdiv.ll
index b655bda68f906..dea6b47a5961a 100644
--- a/llvm/test/CodeGen/X86/fold-int-pow2-with-fmul-or-fdiv.ll
+++ b/llvm/test/CodeGen/X86/fold-int-pow2-with-fmul-or-fdiv.ll
@@ -1529,7 +1529,6 @@ define half @fdiv_pow_shl_cnt_fail_out_of_bound2(i16 %cnt) nounwind {
 ; CHECK-SSE-NEXT:    movl $1, %eax
 ; CHECK-SSE-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; CHECK-SSE-NEXT:    shll %cl, %eax
-; CHECK-SSE-NEXT:    movzwl %ax, %eax
 ; CHECK-SSE-NEXT:    cvtsi2ss %eax, %xmm0
 ; CHECK-SSE-NEXT:    callq __truncsfhf2 at PLT
 ; CHECK-SSE-NEXT:    callq __extendhfsf2 at PLT
@@ -1547,7 +1546,6 @@ define half @fdiv_pow_shl_cnt_fail_out_of_bound2(i16 %cnt) nounwind {
 ; CHECK-AVX2-NEXT:    movl $1, %eax
 ; CHECK-AVX2-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; CHECK-AVX2-NEXT:    shll %cl, %eax
-; CHECK-AVX2-NEXT:    movzwl %ax, %eax
 ; CHECK-AVX2-NEXT:    vcvtsi2ss %eax, %xmm15, %xmm0
 ; CHECK-AVX2-NEXT:    callq __truncsfhf2 at PLT
 ; CHECK-AVX2-NEXT:    callq __extendhfsf2 at PLT
@@ -1563,7 +1561,6 @@ define half @fdiv_pow_shl_cnt_fail_out_of_bound2(i16 %cnt) nounwind {
 ; CHECK-ONLY-AVX512F-NEXT:    movl $1, %eax
 ; CHECK-ONLY-AVX512F-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; CHECK-ONLY-AVX512F-NEXT:    shll %cl, %eax
-; CHECK-ONLY-AVX512F-NEXT:    movzwl %ax, %eax
 ; CHECK-ONLY-AVX512F-NEXT:    vcvtsi2ss %eax, %xmm15, %xmm0
 ; CHECK-ONLY-AVX512F-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
 ; CHECK-ONLY-AVX512F-NEXT:    vcvtph2ps %xmm0, %xmm0
@@ -1576,7 +1573,6 @@ define half @fdiv_pow_shl_cnt_fail_out_of_bound2(i16 %cnt) nounwind {
 ; CHECK-SKX:       # %bb.0:
 ; CHECK-SKX-NEXT:    movl $1, %eax
 ; CHECK-SKX-NEXT:    shlxl %edi, %eax, %eax
-; CHECK-SKX-NEXT:    movzwl %ax, %eax
 ; CHECK-SKX-NEXT:    vcvtsi2ss %eax, %xmm15, %xmm0
 ; CHECK-SKX-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
 ; CHECK-SKX-NEXT:    vcvtph2ps %xmm0, %xmm0
diff --git a/llvm/test/CodeGen/X86/known-never-zero.ll b/llvm/test/CodeGen/X86/known-never-zero.ll
index 63336ffa7c6c8..758ebc6177f14 100644
--- a/llvm/test/CodeGen/X86/known-never-zero.ll
+++ b/llvm/test/CodeGen/X86/known-never-zero.ll
@@ -1318,7 +1318,6 @@ define i32 @zext_known_nonzero(i16 %xx) {
 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    movl $256, %eax # imm = 0x100
 ; X86-NEXT:    shll %cl, %eax
-; X86-NEXT:    movzwl %ax, %eax
 ; X86-NEXT:    rep bsfl %eax, %eax
 ; X86-NEXT:    retl
 ;
@@ -1328,7 +1327,6 @@ define i32 @zext_known_nonzero(i16 %xx) {
 ; X64-NEXT:    movl $256, %eax # imm = 0x100
 ; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; X64-NEXT:    shll %cl, %eax
-; X64-NEXT:    movzwl %ax, %eax
 ; X64-NEXT:    rep bsfl %eax, %eax
 ; X64-NEXT:    retq
   %x = shl nuw nsw i16 256, %xx
@@ -1363,7 +1361,6 @@ define i32 @sext_known_nonzero(i16 %xx) {
 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    movl $256, %eax # imm = 0x100
 ; X86-NEXT:    shll %cl, %eax
-; X86-NEXT:    movzwl %ax, %eax
 ; X86-NEXT:    rep bsfl %eax, %eax
 ; X86-NEXT:    retl
 ;
@@ -1373,7 +1370,6 @@ define i32 @sext_known_nonzero(i16 %xx) {
 ; X64-NEXT:    movl $256, %eax # imm = 0x100
 ; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; X64-NEXT:    shll %cl, %eax
-; X64-NEXT:    movzwl %ax, %eax
 ; X64-NEXT:    rep bsfl %eax, %eax
 ; X64-NEXT:    retq
   %x = shl nuw nsw i16 256, %xx
diff --git a/llvm/test/CodeGen/X86/known-pow2.ll b/llvm/test/CodeGen/X86/known-pow2.ll
index 019bca7e53b4c..09ceaf8eca3ac 100644
--- a/llvm/test/CodeGen/X86/known-pow2.ll
+++ b/llvm/test/CodeGen/X86/known-pow2.ll
@@ -847,8 +847,7 @@ define i1 @pow2_though_zext(i32 %x, i16 %y) {
 ; CHECK-NEXT:    movl $4, %eax
 ; CHECK-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; CHECK-NEXT:    shll %cl, %eax
-; CHECK-NEXT:    andl %edi, %eax
-; CHECK-NEXT:    testl $65535, %eax # imm = 0xFFFF
+; CHECK-NEXT:    testl %eax, %edi
 ; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %dd = shl nuw nsw i16 4, %y
diff --git a/llvm/test/CodeGen/X86/pr171750.ll b/llvm/test/CodeGen/X86/pr171750.ll
new file mode 100644
index 0000000000000..b70536ee215d3
--- /dev/null
+++ b/llvm/test/CodeGen/X86/pr171750.ll
@@ -0,0 +1,47 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=X86
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=X64
+
+define i32 @_Z1ft(i16 zeroext %0) {
+; X86-LABEL: _Z1ft:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    shll $3, %eax
+; X86-NEXT:    retl
+;
+; X64-LABEL: _Z1ft:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    # kill: def $edi killed $edi def $rdi
+; X64-NEXT:    leal (,%rdi,8), %eax
+; X64-NEXT:    retq
+entry:
+  %3 = shl nuw i16 %0, 3
+  %4 = zext i16 %3 to i32
+  ret i32 %4
+}
+
+define i32 @_Z1gt(i16 zeroext %x) {
+; X86-LABEL: _Z1gt:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    leal (,%eax,8), %ecx
+; X86-NEXT:    shll $16, %eax
+; X86-NEXT:    orl %ecx, %eax
+; X86-NEXT:    retl
+;
+; X64-LABEL: _Z1gt:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    # kill: def $edi killed $edi def $rdi
+; X64-NEXT:    leal (,%rdi,8), %eax
+; X64-NEXT:    shll $16, %edi
+; X64-NEXT:    orl %edi, %eax
+; X64-NEXT:    retq
+entry:
+  %conv = zext nneg i16 %x to i32
+  %shl = shl nuw i16 %x, 3
+  %conv3 = zext i16 %shl to i32
+  %shl5 = shl nuw nsw i32 %conv, 16
+  %or = or disjoint i32 %shl5, %conv3
+  ret i32 %or
+}
+
diff --git a/llvm/test/CodeGen/X86/pr89877.ll b/llvm/test/CodeGen/X86/pr89877.ll
index a40ad8f941278..eeb10edb1002f 100644
--- a/llvm/test/CodeGen/X86/pr89877.ll
+++ b/llvm/test/CodeGen/X86/pr89877.ll
@@ -86,7 +86,6 @@ define i32 @sext_known_nonzero_nuw_nsw(i16 %xx) {
 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    movl $256, %eax # imm = 0x100
 ; X86-NEXT:    shll %cl, %eax
-; X86-NEXT:    movzwl %ax, %eax
 ; X86-NEXT:    rep bsfl %eax, %eax
 ; X86-NEXT:    retl
 ;
@@ -96,7 +95,6 @@ define i32 @sext_known_nonzero_nuw_nsw(i16 %xx) {
 ; X64-NEXT:    movl $256, %eax # imm = 0x100
 ; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; X64-NEXT:    shll %cl, %eax
-; X64-NEXT:    movzwl %ax, %eax
 ; X64-NEXT:    rep bsfl %eax, %eax
 ; X64-NEXT:    retq
   %x = shl nuw nsw i16 256, %xx

>From 38affdb9eb26308d8769627530903830fa1ccd0b Mon Sep 17 00:00:00 2001
From: actink <actink at 163.com>
Date: Sat, 13 Dec 2025 11:49:43 +0800
Subject: [PATCH 2/2] fixup shl's promote, add trunc flags

---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 20 +++++++++++++------
 .../lib/CodeGen/SelectionDAG/SelectionDAG.cpp |  3 +++
 llvm/test/CodeGen/X86/known-never-zero.ll     |  2 ++
 llvm/test/CodeGen/X86/known-pow2.ll           |  3 ++-
 .../CodeGen/X86/{pr171750.ll => pr172046.ll}  | 14 +++++++------
 llvm/test/CodeGen/X86/pr89877.ll              |  2 --
 6 files changed, 29 insertions(+), 15 deletions(-)
 rename llvm/test/CodeGen/X86/{pr171750.ll => pr172046.ll} (72%)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index a0212774d8f17..666b52be42b72 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -1659,24 +1659,32 @@ SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
 
     LLVM_DEBUG(dbgs() << "\nPromoting "; Op.dump(&DAG));
 
+    SDNodeFlags Flags;
     bool Replace = false;
     SDValue N0 = Op.getOperand(0);
     if (Opc == ISD::SRA)
       N0 = SExtPromoteOperand(N0, PVT);
     else if (Opc == ISD::SRL)
       N0 = ZExtPromoteOperand(N0, PVT);
-    else
-      N0 = PromoteOperand(N0, PVT, Replace);
+    else {
+      if (Op->getFlags().hasNoUnsignedWrap()) {
+        Flags = SDNodeFlags::NoUnsignedWrap;
+        N0 = ZExtPromoteOperand(N0, PVT);
+      } else if (Op->getFlags().hasNoSignedWrap()) {
+        Flags = SDNodeFlags::NoSignedWrap;
+        N0 = SExtPromoteOperand(N0, PVT);
+      } else
+        N0 = PromoteOperand(N0, PVT, Replace);
+    }
 
     if (!N0.getNode())
       return SDValue();
 
     SDLoc DL(Op);
     SDValue N1 = Op.getOperand(1);
-    SDValue POp = DAG.getNode(Opc, DL, PVT, N0, N1);
-    if (Opc == ISD::SHL && Op->getFlags().hasNoUnsignedWrap())
-      POp = DAG.getNode(ISD::AssertZext, DL, PVT, POp, DAG.getValueType(VT));
-    SDValue RV = DAG.getNode(ISD::TRUNCATE, DL, VT, POp);
+
+    SDValue RV = DAG.getNode(ISD::TRUNCATE, DL, VT,
+                             DAG.getNode(Opc, DL, PVT, N0, N1), Flags);
 
     if (Replace)
       ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index cbe3236eb80d9..7795a004d658d 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -1572,6 +1572,9 @@ SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, const SDLoc &DL, EVT VT) {
   assert(VT.bitsLE(OpVT) && "Not extending!");
   if (OpVT == VT)
     return Op;
+  if (Op.getOpcode() == ISD::AssertZext &&
+      cast<VTSDNode>(Op.getOperand(1))->getVT().bitsLE(VT))
+    return Op;
   APInt Imm = APInt::getLowBitsSet(OpVT.getScalarSizeInBits(),
                                    VT.getScalarSizeInBits());
   return getNode(ISD::AND, DL, OpVT, Op, getConstant(Imm, DL, OpVT));
diff --git a/llvm/test/CodeGen/X86/known-never-zero.ll b/llvm/test/CodeGen/X86/known-never-zero.ll
index 758ebc6177f14..b9353a2cddd1a 100644
--- a/llvm/test/CodeGen/X86/known-never-zero.ll
+++ b/llvm/test/CodeGen/X86/known-never-zero.ll
@@ -1318,6 +1318,7 @@ define i32 @zext_known_nonzero(i16 %xx) {
 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    movl $256, %eax # imm = 0x100
 ; X86-NEXT:    shll %cl, %eax
+; X86-NEXT:    movzwl %ax, %eax
 ; X86-NEXT:    rep bsfl %eax, %eax
 ; X86-NEXT:    retl
 ;
@@ -1327,6 +1328,7 @@ define i32 @zext_known_nonzero(i16 %xx) {
 ; X64-NEXT:    movl $256, %eax # imm = 0x100
 ; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; X64-NEXT:    shll %cl, %eax
+; X64-NEXT:    movzwl %ax, %eax
 ; X64-NEXT:    rep bsfl %eax, %eax
 ; X64-NEXT:    retq
   %x = shl nuw nsw i16 256, %xx
diff --git a/llvm/test/CodeGen/X86/known-pow2.ll b/llvm/test/CodeGen/X86/known-pow2.ll
index 09ceaf8eca3ac..019bca7e53b4c 100644
--- a/llvm/test/CodeGen/X86/known-pow2.ll
+++ b/llvm/test/CodeGen/X86/known-pow2.ll
@@ -847,7 +847,8 @@ define i1 @pow2_though_zext(i32 %x, i16 %y) {
 ; CHECK-NEXT:    movl $4, %eax
 ; CHECK-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; CHECK-NEXT:    shll %cl, %eax
-; CHECK-NEXT:    testl %eax, %edi
+; CHECK-NEXT:    andl %edi, %eax
+; CHECK-NEXT:    testl $65535, %eax # imm = 0xFFFF
 ; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %dd = shl nuw nsw i16 4, %y
diff --git a/llvm/test/CodeGen/X86/pr171750.ll b/llvm/test/CodeGen/X86/pr172046.ll
similarity index 72%
rename from llvm/test/CodeGen/X86/pr171750.ll
rename to llvm/test/CodeGen/X86/pr172046.ll
index b70536ee215d3..df8c82adb3a59 100644
--- a/llvm/test/CodeGen/X86/pr171750.ll
+++ b/llvm/test/CodeGen/X86/pr172046.ll
@@ -1,12 +1,13 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
-; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=X86
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=X64
+; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s --check-prefixes=X86
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefixes=X64
 
 define i32 @_Z1ft(i16 zeroext %0) {
 ; X86-LABEL: _Z1ft:
 ; X86:       # %bb.0: # %entry
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    shll $3, %eax
+; X86-NEXT:    movzwl %ax, %eax
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: _Z1ft:
@@ -24,17 +25,18 @@ define i32 @_Z1gt(i16 zeroext %x) {
 ; X86-LABEL: _Z1gt:
 ; X86:       # %bb.0: # %entry
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    leal (,%eax,8), %ecx
+; X86-NEXT:    movl %eax, %ecx
+; X86-NEXT:    andl $8191, %ecx # imm = 0x1FFF
 ; X86-NEXT:    shll $16, %eax
-; X86-NEXT:    orl %ecx, %eax
+; X86-NEXT:    leal (%eax,%ecx,8), %eax
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: _Z1gt:
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    # kill: def $edi killed $edi def $rdi
-; X64-NEXT:    leal (,%rdi,8), %eax
+; X64-NEXT:    movl %edi, %eax
 ; X64-NEXT:    shll $16, %edi
-; X64-NEXT:    orl %edi, %eax
+; X64-NEXT:    leal (%rdi,%rax,8), %eax
 ; X64-NEXT:    retq
 entry:
   %conv = zext nneg i16 %x to i32
diff --git a/llvm/test/CodeGen/X86/pr89877.ll b/llvm/test/CodeGen/X86/pr89877.ll
index eeb10edb1002f..ed85e7fe9fb60 100644
--- a/llvm/test/CodeGen/X86/pr89877.ll
+++ b/llvm/test/CodeGen/X86/pr89877.ll
@@ -61,7 +61,6 @@ define i32 @sext_known_nonzero_nsw(i16 %xx) {
 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    movl $256, %eax # imm = 0x100
 ; X86-NEXT:    shll %cl, %eax
-; X86-NEXT:    movzwl %ax, %eax
 ; X86-NEXT:    rep bsfl %eax, %eax
 ; X86-NEXT:    retl
 ;
@@ -71,7 +70,6 @@ define i32 @sext_known_nonzero_nsw(i16 %xx) {
 ; X64-NEXT:    movl $256, %eax # imm = 0x100
 ; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; X64-NEXT:    shll %cl, %eax
-; X64-NEXT:    movzwl %ax, %eax
 ; X64-NEXT:    rep bsfl %eax, %eax
 ; X64-NEXT:    retq
   %x = shl nsw i16 256, %xx



More information about the llvm-commits mailing list