[X86] Improve transform for add-like nodes to add

We previously did this only in tablegen, after we would have already
dropped `disjoint` flag from `or`.

Author: goldsteinn

llvm/lib/Target/X86/X86ISelDAGToDAG.cpp

@@ -5295,6 +5295,20 @@ void X86DAGToDAGISel::Select(SDNode *Node) {
     if (tryVPTERNLOG(Node))
       return;
 
+    // Convert addlike to add before final selection. Do this before we drop
+    // flags like `disjoint`.
+    // NB: Conversion to add is preferable so we use `lea` in codegen.
+    if (NVT.isScalarInteger() &&
+        (Opcode == ISD::OR ||
+         (NVT == MVT::i8 || NVT == MVT::i16 || NVT == MVT::i32)) &&
+        CurDAG->isADDLike(SDValue(Node, 0))) {
+      SDValue AsAdd = CurDAG->getNode(ISD::ADD, SDLoc(Node), NVT,
+                                      Node->getOperand(0), Node->getOperand(1));
+      ReplaceUses(SDValue(Node, 0), AsAdd);
+      CurDAG->RemoveDeadNode(Node);
+      Node = AsAdd.getNode();
+      Opcode = ISD::ADD;
+    }
     [[fallthrough]];
   case ISD::ADD:
     if (Opcode == ISD::ADD && matchBitExtract(Node))

llvm/lib/Target/X86/X86InstrCompiler.td

@@ -1560,21 +1560,38 @@ let Predicates = [HasNDD] in {
 }
 
 // Depositing value to 8/16 bit subreg:
-def : Pat<(or (and GR64:$dst, -256), 
+def : Pat<(or (and GR64:$dst, -256),
               (i64 (zextloadi8 addr:$src))),
-          (INSERT_SUBREG (i64 (COPY $dst)), (MOV8rm  i8mem:$src), sub_8bit)>; 
+          (INSERT_SUBREG (i64 (COPY $dst)), (MOV8rm  i8mem:$src), sub_8bit)>;
 
-def : Pat<(or (and GR32:$dst, -256), 
+def : Pat<(or (and GR32:$dst, -256),
               (i32 (zextloadi8 addr:$src))),
-          (INSERT_SUBREG (i32 (COPY $dst)), (MOV8rm  i8mem:$src), sub_8bit)>; 
+          (INSERT_SUBREG (i32 (COPY $dst)), (MOV8rm  i8mem:$src), sub_8bit)>;
 
-def : Pat<(or (and GR64:$dst, -65536), 
+def : Pat<(or (and GR64:$dst, -65536),
               (i64 (zextloadi16 addr:$src))),
           (INSERT_SUBREG (i64 (COPY $dst)), (MOV16rm  i16mem:$src), sub_16bit)>;
 
-def : Pat<(or (and GR32:$dst, -65536), 
+def : Pat<(or (and GR32:$dst, -65536),
               (i32 (zextloadi16 addr:$src))),
-          (INSERT_SUBREG (i32 (COPY $dst)), (MOV16rm  i16mem:$src), sub_16bit)>; 
+          (INSERT_SUBREG (i32 (COPY $dst)), (MOV16rm  i16mem:$src), sub_16bit)>;
+
+// We convert or -> add when the or is disjoint so need to handle for add as well.
+def : Pat<(add (and GR64:$dst, -256),
+               (i64 (zextloadi8 addr:$src))),
+          (INSERT_SUBREG (i64 (COPY $dst)), (MOV8rm  i8mem:$src), sub_8bit)>;
+
+def : Pat<(add (and GR32:$dst, -256),
+               (i32 (zextloadi8 addr:$src))),
+          (INSERT_SUBREG (i32 (COPY $dst)), (MOV8rm  i8mem:$src), sub_8bit)>;
+
+def : Pat<(add (and GR64:$dst, -65536),
+               (i64 (zextloadi16 addr:$src))),
+          (INSERT_SUBREG (i64 (COPY $dst)), (MOV16rm  i16mem:$src), sub_16bit)>;
+
+def : Pat<(add (and GR32:$dst, -65536),
+               (i32 (zextloadi16 addr:$src))),
+          (INSERT_SUBREG (i32 (COPY $dst)), (MOV16rm  i16mem:$src), sub_16bit)>;
 
 // To avoid needing to materialize an immediate in a register, use a 32-bit and
 // with implicit zero-extension instead of a 64-bit and if the immediate has at

llvm/test/CodeGen/X86/2009-05-23-dagcombine-shifts.ll

@@ -1,3 +1,4 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
 ; RUN: llc < %s | FileCheck %s
 
 ; Check that the shr(shl X, 56), 48) is not mistakenly turned into
@@ -16,11 +17,13 @@ target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f3
 target triple = "x86_64-unknown-linux-gnu"
 
 define i64 @foo(i64 %b) nounwind readnone {
-entry:
 ; CHECK-LABEL: foo:
-; CHECK: movsbq %dil, %rax
-; CHECK: shlq $8, %rax
-; CHECK: orq $1, %rax
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movsbq %dil, %rax
+; CHECK-NEXT:    shlq $8, %rax
+; CHECK-NEXT:    incq %rax
+; CHECK-NEXT:    retq
+entry:
 	%shl = shl i64 %b, 56		; <i64> [#uses=1]
 	%shr = ashr i64 %shl, 48		; <i64> [#uses=1]
 	%add5 = or i64 %shr, 1		; <i64> [#uses=1]

llvm/test/CodeGen/X86/3addr-or.ll

@@ -24,7 +24,7 @@ define i64 @test2(i8 %A, i8 %B) nounwind {
 ; CHECK-NEXT:    andl $48, %edi
 ; CHECK-NEXT:    movzbl %sil, %eax
 ; CHECK-NEXT:    shrl $4, %eax
-; CHECK-NEXT:    orl %edi, %eax
+; CHECK-NEXT:    addl %edi, %eax
 ; CHECK-NEXT:    retq
   %C = zext i8 %A to i64
   %D = shl i64 %C, 4
@@ -42,7 +42,7 @@ define void @test3(i32 %x, ptr %P) nounwind readnone ssp {
 ; CHECK-LABEL: test3:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    shll $5, %edi
-; CHECK-NEXT:    orl $3, %edi
+; CHECK-NEXT:    addl $3, %edi
 ; CHECK-NEXT:    movl %edi, (%rsi)
 ; CHECK-NEXT:    retq
   %t0 = shl i32 %x, 5
@@ -71,7 +71,7 @@ define void @test5(i32 %a, i32 %b, ptr nocapture %P) nounwind ssp {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    andl $6, %edi
 ; CHECK-NEXT:    andl $16, %esi
-; CHECK-NEXT:    orl %edi, %esi
+; CHECK-NEXT:    addl %edi, %esi
 ; CHECK-NEXT:    movl %esi, (%rdx)
 ; CHECK-NEXT:    retq
   %and = and i32 %a, 6

llvm/test/CodeGen/X86/addcarry2.ll

@@ -138,7 +138,7 @@ define void @adc_load_store_32_127(ptr inreg %x, ptr inreg %x2, i32 inreg %y) no
 ; X64-NEXT:    movl (%rdi), %eax # encoding: [0x8b,0x07]
 ; X64-NEXT:    shlq $32, %rax # encoding: [0x48,0xc1,0xe0,0x20]
 ; X64-NEXT:    movl %edx, %ecx # encoding: [0x89,0xd1]
-; X64-NEXT:    orq %rax, %rcx # encoding: [0x48,0x09,0xc1]
+; X64-NEXT:    addq %rax, %rcx # encoding: [0x48,0x01,0xc1]
 ; X64-NEXT:    movabsq $545460846593, %rax # encoding: [0x48,0xb8,0x01,0x00,0x00,0x00,0x7f,0x00,0x00,0x00]
 ; X64-NEXT:    # imm = 0x7F00000001
 ; X64-NEXT:    xorl %edx, %edx # encoding: [0x31,0xd2]
@@ -178,7 +178,7 @@ define void @adc_load_store_32_128(ptr inreg %x, ptr inreg %x2, i32 inreg %y) no
 ; X64-NEXT:    movl (%rdi), %eax # encoding: [0x8b,0x07]
 ; X64-NEXT:    shlq $32, %rax # encoding: [0x48,0xc1,0xe0,0x20]
 ; X64-NEXT:    movl %edx, %ecx # encoding: [0x89,0xd1]
-; X64-NEXT:    orq %rax, %rcx # encoding: [0x48,0x09,0xc1]
+; X64-NEXT:    addq %rax, %rcx # encoding: [0x48,0x01,0xc1]
 ; X64-NEXT:    movabsq $549755813889, %rax # encoding: [0x48,0xb8,0x01,0x00,0x00,0x00,0x80,0x00,0x00,0x00]
 ; X64-NEXT:    # imm = 0x8000000001
 ; X64-NEXT:    xorl %edx, %edx # encoding: [0x31,0xd2]

llvm/test/CodeGen/X86/and-or-fold.ll

@@ -45,7 +45,7 @@ define i32 @test1(i32 %x, i16 %y) {
 ; DARWIN-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
 ; DARWIN-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; DARWIN-NEXT:    shll $16, %eax
-; DARWIN-NEXT:    orl %ecx, %eax
+; DARWIN-NEXT:    addl %ecx, %eax
 ; DARWIN-NEXT:    andl $16711807, %eax ## imm = 0xFF007F
 ; DARWIN-NEXT:    retl
 ;
@@ -54,7 +54,7 @@ define i32 @test1(i32 %x, i16 %y) {
 ; DARWIN-OPT-NEXT:    andl $127, %esi
 ; DARWIN-OPT-NEXT:    movzbl %dil, %eax
 ; DARWIN-OPT-NEXT:    shll $16, %eax
-; DARWIN-OPT-NEXT:    orl %esi, %eax
+; DARWIN-OPT-NEXT:    addl %esi, %eax
 ; DARWIN-OPT-NEXT:    retq
   %tmp1 = zext i16 %y to i32
   %tmp2 = and i32 %tmp1, 127

llvm/test/CodeGen/X86/andimm8.ll

@@ -29,15 +29,15 @@ define void @foo(i64 %zed, ptr %x) nounwind {
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx # encoding: [0x8b,0x4c,0x24,0x04]
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx # encoding: [0x8b,0x54,0x24,0x08]
 ; X86-NEXT:    andl $-4, %ecx # encoding: [0x83,0xe1,0xfc]
-; X86-NEXT:    orl $2, %ecx # encoding: [0x83,0xc9,0x02]
+; X86-NEXT:    addl $2, %ecx # encoding: [0x83,0xc1,0x02]
 ; X86-NEXT:    movl %edx, 4(%eax) # encoding: [0x89,0x50,0x04]
 ; X86-NEXT:    movl %ecx, (%eax) # encoding: [0x89,0x08]
 ; X86-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-LABEL: foo:
 ; X64:       # %bb.0:
 ; X64-NEXT:    andq $-4, %rdi # encoding: [0x48,0x83,0xe7,0xfc]
-; X64-NEXT:    orq $2, %rdi # encoding: [0x48,0x83,0xcf,0x02]
+; X64-NEXT:    addq $2, %rdi # encoding: [0x48,0x83,0xc7,0x02]
 ; X64-NEXT:    movq %rdi, (%rsi) # encoding: [0x48,0x89,0x3e]
 ; X64-NEXT:    retq # encoding: [0xc3]
   %t1 = and i64 %zed, -4

llvm/test/CodeGen/X86/atomic-unordered.ll

@@ -2359,7 +2359,7 @@ define i16 @load_combine(ptr %p) {
 ; CHECK-O3-NEXT:    movzbl (%rdi), %ecx
 ; CHECK-O3-NEXT:    movzbl 1(%rdi), %eax
 ; CHECK-O3-NEXT:    shll $8, %eax
-; CHECK-O3-NEXT:    orl %ecx, %eax
+; CHECK-O3-NEXT:    addl %ecx, %eax
 ; CHECK-O3-NEXT:    # kill: def $ax killed $ax killed $eax
 ; CHECK-O3-NEXT:    retq
   %v1 = load atomic i8, ptr %p unordered, align 2