Index: llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp =================================================================== --- llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -3842,9 +3842,16 @@ EVT ExtVT; if (isAndLoadExtLoad(Mask, Load, Load->getValueType(0), ExtVT) && isLegalNarrowLoad(Load, ISD::ZEXTLOAD, ExtVT)) { - // Only add this load if we can make it more narrow. - if (ExtVT.bitsLT(Load->getMemoryVT())) + + // ZEXTLOAD is already small enough. + if (Load->getExtensionType() == ISD::ZEXTLOAD && + ExtVT.bitsGE(Load->getMemoryVT())) + continue; + + // Use LE to convert equal sized loads to zext. + if (ExtVT.bitsLE(Load->getMemoryVT())) Loads.insert(Load); + continue; } return false; @@ -3899,11 +3906,13 @@ if (Loads.size() == 0) return false; + DEBUG(dbgs() << "Backwards propagate AND: "; N->dump()); SDValue MaskOp = N->getOperand(1); // If it exists, fixup the single node we allow in the tree that needs // masking. if (FixupNode) { + DEBUG(dbgs() << "First, need to fix up: "; FixupNode->dump()); SDValue And = DAG.getNode(ISD::AND, SDLoc(FixupNode), FixupNode->getValueType(0), SDValue(FixupNode, 0), MaskOp); @@ -3922,6 +3931,7 @@ // Create narrow loads. for (auto *Load : Loads) { + DEBUG(dbgs() << "Propagate AND back to: "; Load->dump()); SDValue And = DAG.getNode(ISD::AND, SDLoc(Load), Load->getValueType(0), SDValue(Load, 0), MaskOp); DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), And); Index: llvm/trunk/test/CodeGen/ARM/and-load-combine.ll =================================================================== --- llvm/trunk/test/CodeGen/ARM/and-load-combine.ll +++ llvm/trunk/test/CodeGen/ARM/and-load-combine.ll @@ -852,8 +852,7 @@ ; ARM: @ %bb.0: @ %entry ; ARM-NEXT: ldrb r0, [r0] ; ARM-NEXT: uxtb r2, r2 -; ARM-NEXT: and r0, r0, r1 -; ARM-NEXT: uxtb r1, r0 +; ARM-NEXT: and r1, r0, r1 ; ARM-NEXT: mov r0, #0 ; ARM-NEXT: cmp r1, r2 ; ARM-NEXT: movweq r0, #1 @@ -863,8 +862,7 @@ ; ARMEB: @ %bb.0: @ %entry ; ARMEB-NEXT: ldrb r0, [r0] ; ARMEB-NEXT: uxtb r2, r2 -; ARMEB-NEXT: and r0, r0, r1 -; ARMEB-NEXT: uxtb r1, r0 +; ARMEB-NEXT: and r1, r0, r1 ; ARMEB-NEXT: mov r0, #0 ; ARMEB-NEXT: cmp r1, r2 ; ARMEB-NEXT: movweq r0, #1 @@ -872,9 +870,8 @@ ; ; THUMB1-LABEL: test6: ; THUMB1: @ %bb.0: @ %entry -; THUMB1-NEXT: ldrb r0, [r0] -; THUMB1-NEXT: ands r0, r1 -; THUMB1-NEXT: uxtb r3, r0 +; THUMB1-NEXT: ldrb r3, [r0] +; THUMB1-NEXT: ands r3, r1 ; THUMB1-NEXT: uxtb r2, r2 ; THUMB1-NEXT: movs r0, #1 ; THUMB1-NEXT: movs r1, #0 @@ -889,8 +886,7 @@ ; THUMB2: @ %bb.0: @ %entry ; THUMB2-NEXT: ldrb r0, [r0] ; THUMB2-NEXT: uxtb r2, r2 -; THUMB2-NEXT: ands r0, r1 -; THUMB2-NEXT: uxtb r1, r0 +; THUMB2-NEXT: ands r1, r0 ; THUMB2-NEXT: movs r0, #0 ; THUMB2-NEXT: cmp r1, r2 ; THUMB2-NEXT: it eq Index: llvm/trunk/test/CodeGen/X86/pr37563.ll =================================================================== --- llvm/trunk/test/CodeGen/X86/pr37563.ll +++ llvm/trunk/test/CodeGen/X86/pr37563.ll @@ -10,9 +10,10 @@ define void @PR35763() { ; CHECK-LABEL: PR35763: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: movzwl z+{{.*}}(%rip), %eax -; CHECK-NEXT: orl {{.*}}(%rip), %eax -; CHECK-NEXT: movq %rax, {{.*}}(%rip) +; CHECK-NEXT: movzwl {{.*}}(%rip), %eax +; CHECK-NEXT: movzwl z+{{.*}}(%rip), %ecx +; CHECK-NEXT: orl %eax, %ecx +; CHECK-NEXT: movq %rcx, {{.*}}(%rip) ; CHECK-NEXT: movl z+{{.*}}(%rip), %eax ; CHECK-NEXT: movzbl z+{{.*}}(%rip), %ecx ; CHECK-NEXT: shlq $32, %rcx