From c89cfbd4ddfa4e01fea1eb87aba9cdcd0e31d3a8 Mon Sep 17 00:00:00 2001
From: David Green <david.green@arm.com>
Date: Tue, 1 Feb 2022 20:18:40 +0000
Subject: [PATCH] Revert "[DAG] Extend SearchForAndLoads with any_extend
 handling"

This reverts commit 100763a88fe97b22cd5e3f69d203669aac3ed48f as it was
making incorrect assumptions about implicit zero_extends.
---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |  41 ++----
 .../CodeGen/AArch64/combine-andintoload.ll    | 120 ++++++++++--------
 llvm/test/CodeGen/X86/pr35763.ll              |   8 +-
 3 files changed, 84 insertions(+), 85 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 0ed2bd9d6d62..041d7e5b4a4a 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -5531,8 +5531,6 @@ bool DAGCombiner::SearchForAndLoads(SDNode *N,
 
     // Some constants may need fixing up later if they are too large.
     if (auto *C = dyn_cast<ConstantSDNode>(Op)) {
-      if (Mask->getValueType(0) != C->getValueType(0))
-        return false;
       if ((N->getOpcode() == ISD::OR || N->getOpcode() == ISD::XOR) &&
           (Mask->getAPIntValue() & C->getAPIntValue()) != C->getAPIntValue())
         NodesWithConsts.insert(N);
@@ -5566,9 +5564,9 @@ bool DAGCombiner::SearchForAndLoads(SDNode *N,
     case ISD::AssertZext: {
       unsigned ActiveBits = Mask->getAPIntValue().countTrailingOnes();
       EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
-      EVT VT = Op.getOpcode() == ISD::AssertZext
-                   ? cast<VTSDNode>(Op.getOperand(1))->getVT()
-                   : Op.getOperand(0).getValueType();
+      EVT VT = Op.getOpcode() == ISD::AssertZext ?
+        cast<VTSDNode>(Op.getOperand(1))->getVT() :
+        Op.getOperand(0).getValueType();
 
       // We can accept extending nodes if the mask is wider or an equal
       // width to the original type.
@@ -5576,15 +5574,6 @@ bool DAGCombiner::SearchForAndLoads(SDNode *N,
         continue;
       break;
     }
-    case ISD::ANY_EXTEND: {
-      unsigned ActiveBits = Mask->getAPIntValue().countTrailingOnes();
-      EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
-      EVT VT = Op.getOperand(0).getValueType();
-      if (ExtVT.bitsGE(VT))
-        break;
-      // Fallthrough to searching for nodes from the operands of the extend.
-      LLVM_FALLTHROUGH;
-    }
     case ISD::OR:
     case ISD::XOR:
     case ISD::AND:
@@ -5644,14 +5633,12 @@ bool DAGCombiner::BackwardsPropagateMask(SDNode *N) {
     // masking.
     if (FixupNode) {
       LLVM_DEBUG(dbgs() << "First, need to fix up: "; FixupNode->dump());
-      SDValue MaskOpT = DAG.getZExtOrTrunc(MaskOp, SDLoc(FixupNode),
-                                           FixupNode->getValueType(0));
-      SDValue And =
-          DAG.getNode(ISD::AND, SDLoc(FixupNode), FixupNode->getValueType(0),
-                      SDValue(FixupNode, 0), MaskOpT);
+      SDValue And = DAG.getNode(ISD::AND, SDLoc(FixupNode),
+                                FixupNode->getValueType(0),
+                                SDValue(FixupNode, 0), MaskOp);
       DAG.ReplaceAllUsesOfValueWith(SDValue(FixupNode, 0), And);
       if (And.getOpcode() == ISD ::AND)
-        DAG.UpdateNodeOperands(And.getNode(), SDValue(FixupNode, 0), MaskOpT);
+        DAG.UpdateNodeOperands(And.getNode(), SDValue(FixupNode, 0), MaskOp);
     }
 
     // Narrow any constants that need it.
@@ -5660,12 +5647,10 @@ bool DAGCombiner::BackwardsPropagateMask(SDNode *N) {
       SDValue Op1 = LogicN->getOperand(1);
 
       if (isa<ConstantSDNode>(Op0))
-        std::swap(Op0, Op1);
+          std::swap(Op0, Op1);
 
-      SDValue MaskOpT =
-          DAG.getZExtOrTrunc(MaskOp, SDLoc(Op1), Op1.getValueType());
-      SDValue And =
-          DAG.getNode(ISD::AND, SDLoc(Op1), Op1.getValueType(), Op1, MaskOpT);
+      SDValue And = DAG.getNode(ISD::AND, SDLoc(Op1), Op1.getValueType(),
+                                Op1, MaskOp);
 
       DAG.UpdateNodeOperands(LogicN, Op0, And);
     }
@@ -5673,14 +5658,12 @@ bool DAGCombiner::BackwardsPropagateMask(SDNode *N) {
     // Create narrow loads.
     for (auto *Load : Loads) {
       LLVM_DEBUG(dbgs() << "Propagate AND back to: "; Load->dump());
-      SDValue MaskOpT =
-          DAG.getZExtOrTrunc(MaskOp, SDLoc(Load), Load->getValueType(0));
       SDValue And = DAG.getNode(ISD::AND, SDLoc(Load), Load->getValueType(0),
-                                SDValue(Load, 0), MaskOpT);
+                                SDValue(Load, 0), MaskOp);
       DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), And);
       if (And.getOpcode() == ISD ::AND)
         And = SDValue(
-            DAG.UpdateNodeOperands(And.getNode(), SDValue(Load, 0), MaskOpT), 0);
+            DAG.UpdateNodeOperands(And.getNode(), SDValue(Load, 0), MaskOp), 0);
       SDValue NewLoad = reduceLoadWidth(And.getNode());
       assert(NewLoad &&
              "Shouldn't be masking the load if it can't be narrowed");
diff --git a/llvm/test/CodeGen/AArch64/combine-andintoload.ll b/llvm/test/CodeGen/AArch64/combine-andintoload.ll
index d7a69fdd1193..693d318b272a 100644
--- a/llvm/test/CodeGen/AArch64/combine-andintoload.ll
+++ b/llvm/test/CodeGen/AArch64/combine-andintoload.ll
@@ -5,14 +5,16 @@
 define i64 @load32_and16_and(i32* %p, i64 %y) {
 ; CHECK-LABEL: load32_and16_and:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrh w8, [x0]
-; CHECK-NEXT:    and w0, w1, w8
+; CHECK-NEXT:    ldr w8, [x0]
+; CHECK-NEXT:    and w8, w1, w8
+; CHECK-NEXT:    and x0, x8, #0xffff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: load32_and16_and:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrh w8, [x0, #2]
-; CHECKBE-NEXT:    and w0, w1, w8
+; CHECKBE-NEXT:    ldr w8, [x0]
+; CHECKBE-NEXT:    and w8, w1, w8
+; CHECKBE-NEXT:    and x0, x8, #0xffff
 ; CHECKBE-NEXT:    ret
   %x = load i32, i32* %p, align 4
   %xz = zext i32 %x to i64
@@ -24,14 +26,16 @@ define i64 @load32_and16_and(i32* %p, i64 %y) {
 define i64 @load32_and16_andr(i32* %p, i64 %y) {
 ; CHECK-LABEL: load32_and16_andr:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrh w8, [x0]
-; CHECK-NEXT:    and w0, w1, w8
+; CHECK-NEXT:    ldr w8, [x0]
+; CHECK-NEXT:    and w8, w1, w8
+; CHECK-NEXT:    and x0, x8, #0xffff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: load32_and16_andr:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrh w8, [x0, #2]
-; CHECKBE-NEXT:    and w0, w1, w8
+; CHECKBE-NEXT:    ldr w8, [x0]
+; CHECKBE-NEXT:    and w8, w1, w8
+; CHECKBE-NEXT:    and x0, x8, #0xffff
 ; CHECKBE-NEXT:    ret
   %x = load i32, i32* %p, align 4
   %xz = zext i32 %x to i64
@@ -43,14 +47,16 @@ define i64 @load32_and16_andr(i32* %p, i64 %y) {
 define i64 @load32_and16_and_sext(i32* %p, i64 %y) {
 ; CHECK-LABEL: load32_and16_and_sext:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrh w8, [x0]
-; CHECK-NEXT:    and w0, w1, w8
+; CHECK-NEXT:    ldr w8, [x0]
+; CHECK-NEXT:    and w8, w1, w8
+; CHECK-NEXT:    and x0, x8, #0xffff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: load32_and16_and_sext:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrh w8, [x0, #2]
-; CHECKBE-NEXT:    and w0, w1, w8
+; CHECKBE-NEXT:    ldr w8, [x0]
+; CHECKBE-NEXT:    and w8, w1, w8
+; CHECKBE-NEXT:    and x0, x8, #0xffff
 ; CHECKBE-NEXT:    ret
   %x = load i32, i32* %p, align 4
   %xz = sext i32 %x to i64
@@ -62,16 +68,16 @@ define i64 @load32_and16_and_sext(i32* %p, i64 %y) {
 define i64 @load32_and16_or(i32* %p, i64 %y) {
 ; CHECK-LABEL: load32_and16_or:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrh w8, [x0]
-; CHECK-NEXT:    and w9, w1, #0xffff
-; CHECK-NEXT:    orr w0, w9, w8
+; CHECK-NEXT:    ldr w8, [x0]
+; CHECK-NEXT:    orr w8, w1, w8
+; CHECK-NEXT:    and x0, x8, #0xffff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: load32_and16_or:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrh w8, [x0, #2]
-; CHECKBE-NEXT:    and w9, w1, #0xffff
-; CHECKBE-NEXT:    orr w0, w9, w8
+; CHECKBE-NEXT:    ldr w8, [x0]
+; CHECKBE-NEXT:    orr w8, w1, w8
+; CHECKBE-NEXT:    and x0, x8, #0xffff
 ; CHECKBE-NEXT:    ret
   %x = load i32, i32* %p, align 4
   %xz = zext i32 %x to i64
@@ -164,14 +170,16 @@ define i64 @load16_and16(i16* %p, i64 %y) {
 define i64 @load16_and8(i16* %p, i64 %y) {
 ; CHECK-LABEL: load16_and8:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrb w8, [x0]
-; CHECK-NEXT:    and w0, w1, w8
+; CHECK-NEXT:    ldrh w8, [x0]
+; CHECK-NEXT:    and w8, w1, w8
+; CHECK-NEXT:    and x0, x8, #0xff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: load16_and8:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrb w8, [x0, #1]
-; CHECKBE-NEXT:    and w0, w1, w8
+; CHECKBE-NEXT:    ldrh w8, [x0]
+; CHECKBE-NEXT:    and w8, w1, w8
+; CHECKBE-NEXT:    and x0, x8, #0xff
 ; CHECKBE-NEXT:    ret
   %x = load i16, i16* %p, align 4
   %xz = zext i16 %x to i64
@@ -224,13 +232,15 @@ define i64 @load8_and16_zext(i8* %p, i8 %y) {
 ; CHECK-LABEL: load8_and16_zext:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    ldrb w8, [x0]
-; CHECK-NEXT:    and w0, w1, w8
+; CHECK-NEXT:    and w8, w1, w8
+; CHECK-NEXT:    and x0, x8, #0xff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: load8_and16_zext:
 ; CHECKBE:       // %bb.0:
 ; CHECKBE-NEXT:    ldrb w8, [x0]
-; CHECKBE-NEXT:    and w0, w1, w8
+; CHECKBE-NEXT:    and w8, w1, w8
+; CHECKBE-NEXT:    and x0, x8, #0xff
 ; CHECKBE-NEXT:    ret
   %x = load i8, i8* %p, align 4
   %xz = zext i8 %x to i64
@@ -286,14 +296,16 @@ define i64 @load8_and16_or(i8* %p, i64 %y) {
 define i64 @load16_and8_manyext(i16* %p, i32 %y) {
 ; CHECK-LABEL: load16_and8_manyext:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrb w8, [x0]
-; CHECK-NEXT:    and w0, w1, w8
+; CHECK-NEXT:    ldrh w8, [x0]
+; CHECK-NEXT:    and w8, w1, w8
+; CHECK-NEXT:    and x0, x8, #0xff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: load16_and8_manyext:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrb w8, [x0, #1]
-; CHECKBE-NEXT:    and w0, w1, w8
+; CHECKBE-NEXT:    ldrh w8, [x0]
+; CHECKBE-NEXT:    and w8, w1, w8
+; CHECKBE-NEXT:    and x0, x8, #0xff
 ; CHECKBE-NEXT:    ret
   %x = load i16, i16* %p, align 4
   %xz = zext i16 %x to i32
@@ -306,16 +318,18 @@ define i64 @load16_and8_manyext(i16* %p, i32 %y) {
 define i64 @multiple_load(i16* %p, i32* %q) {
 ; CHECK-LABEL: multiple_load:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrb w8, [x0]
-; CHECK-NEXT:    ldrb w9, [x1]
-; CHECK-NEXT:    and w0, w9, w8
+; CHECK-NEXT:    ldrh w8, [x0]
+; CHECK-NEXT:    ldr w9, [x1]
+; CHECK-NEXT:    and w8, w9, w8
+; CHECK-NEXT:    and x0, x8, #0xff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: multiple_load:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrb w8, [x0, #1]
-; CHECKBE-NEXT:    ldrb w9, [x1, #3]
-; CHECKBE-NEXT:    and w0, w9, w8
+; CHECKBE-NEXT:    ldrh w8, [x0]
+; CHECKBE-NEXT:    ldr w9, [x1]
+; CHECKBE-NEXT:    and w8, w9, w8
+; CHECKBE-NEXT:    and x0, x8, #0xff
 ; CHECKBE-NEXT:    ret
   %x = load i16, i16* %p, align 4
   %xz = zext i16 %x to i64
@@ -329,16 +343,18 @@ define i64 @multiple_load(i16* %p, i32* %q) {
 define i64 @multiple_load_or(i16* %p, i32* %q) {
 ; CHECK-LABEL: multiple_load_or:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrb w8, [x0]
-; CHECK-NEXT:    ldrb w9, [x1]
-; CHECK-NEXT:    orr w0, w9, w8
+; CHECK-NEXT:    ldrh w8, [x0]
+; CHECK-NEXT:    ldr w9, [x1]
+; CHECK-NEXT:    orr w8, w9, w8
+; CHECK-NEXT:    and x0, x8, #0xff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: multiple_load_or:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrb w8, [x0, #1]
-; CHECKBE-NEXT:    ldrb w9, [x1, #3]
-; CHECKBE-NEXT:    orr w0, w9, w8
+; CHECKBE-NEXT:    ldrh w8, [x0]
+; CHECKBE-NEXT:    ldr w9, [x1]
+; CHECKBE-NEXT:    orr w8, w9, w8
+; CHECKBE-NEXT:    and x0, x8, #0xff
 ; CHECKBE-NEXT:    ret
   %x = load i16, i16* %p, align 4
   %xz = zext i16 %x to i64
@@ -352,16 +368,16 @@ define i64 @multiple_load_or(i16* %p, i32* %q) {
 define i64 @load32_and16_zexty(i32* %p, i32 %y) {
 ; CHECK-LABEL: load32_and16_zexty:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrh w8, [x0]
-; CHECK-NEXT:    and w9, w1, #0xffff
-; CHECK-NEXT:    orr w0, w9, w8
+; CHECK-NEXT:    ldr w8, [x0]
+; CHECK-NEXT:    orr w8, w1, w8
+; CHECK-NEXT:    and x0, x8, #0xffff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: load32_and16_zexty:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrh w8, [x0, #2]
-; CHECKBE-NEXT:    and w9, w1, #0xffff
-; CHECKBE-NEXT:    orr w0, w9, w8
+; CHECKBE-NEXT:    ldr w8, [x0]
+; CHECKBE-NEXT:    orr w8, w1, w8
+; CHECKBE-NEXT:    and x0, x8, #0xffff
 ; CHECKBE-NEXT:    ret
   %x = load i32, i32* %p, align 4
   %xz = zext i32 %x to i64
@@ -374,16 +390,16 @@ define i64 @load32_and16_zexty(i32* %p, i32 %y) {
 define i64 @load32_and16_sexty(i32* %p, i32 %y) {
 ; CHECK-LABEL: load32_and16_sexty:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldrh w8, [x0]
-; CHECK-NEXT:    and w9, w1, #0xffff
-; CHECK-NEXT:    orr w0, w9, w8
+; CHECK-NEXT:    ldr w8, [x0]
+; CHECK-NEXT:    orr w8, w1, w8
+; CHECK-NEXT:    and x0, x8, #0xffff
 ; CHECK-NEXT:    ret
 ;
 ; CHECKBE-LABEL: load32_and16_sexty:
 ; CHECKBE:       // %bb.0:
-; CHECKBE-NEXT:    ldrh w8, [x0, #2]
-; CHECKBE-NEXT:    and w9, w1, #0xffff
-; CHECKBE-NEXT:    orr w0, w9, w8
+; CHECKBE-NEXT:    ldr w8, [x0]
+; CHECKBE-NEXT:    orr w8, w1, w8
+; CHECKBE-NEXT:    and x0, x8, #0xffff
 ; CHECKBE-NEXT:    ret
   %x = load i32, i32* %p, align 4
   %xz = zext i32 %x to i64
diff --git a/llvm/test/CodeGen/X86/pr35763.ll b/llvm/test/CodeGen/X86/pr35763.ll
index 53a0a0284d11..8b3e91dc577a 100644
--- a/llvm/test/CodeGen/X86/pr35763.ll
+++ b/llvm/test/CodeGen/X86/pr35763.ll
@@ -10,10 +10,10 @@
 define dso_local void @PR35763() {
 ; CHECK-LABEL: PR35763:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    movzwl z(%rip), %eax
-; CHECK-NEXT:    movzwl z+2(%rip), %ecx
-; CHECK-NEXT:    orl %eax, %ecx
-; CHECK-NEXT:    movq %rcx, tf_3_var_136(%rip)
+; CHECK-NEXT:    movl z(%rip), %eax
+; CHECK-NEXT:    orl z+2(%rip), %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    movq %rax, tf_3_var_136(%rip)
 ; CHECK-NEXT:    movl z+6(%rip), %eax
 ; CHECK-NEXT:    movzbl z+10(%rip), %ecx
 ; CHECK-NEXT:    shlq $32, %rcx
-- 
2.35.1.723.g4982287a31-goog