[DAGCombiner] Support bswap as a part of load combine patterns

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D29397

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@294201 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 02221d700bfd6936e0bb76ceca2de2a82e04c467..97ec5d6b2d7f8ec5dab024417f951d72feb3c222 100644 (file)
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -4463,6 +4463,9 @@ const Optional<ByteProvider> calculateByteProvider(SDValue Op, unsigned Index,
                ? ByteProvider::getConstantZero()
                : calculateByteProvider(NarrowOp, Index, Depth + 1);
   }
+  case ISD::BSWAP:
+    return calculateByteProvider(Op->getOperand(0), ByteWidth - Index - 1,
+                                 Depth + 1);
   case ISD::LOAD: {
     auto L = cast<LoadSDNode>(Op.getNode());
 

diff --git a/test/CodeGen/AArch64/load-combine-big-endian.ll b/test/CodeGen/AArch64/load-combine-big-endian.ll
index 692a57c4471ce20e4947c24ae151f7adb80483ab..9fc852b71745b0f4fbc383f051441f312eaae0e1 100644 (file)
--- a/test/CodeGen/AArch64/load-combine-big-endian.ll
+++ b/test/CodeGen/AArch64/load-combine-big-endian.ll
@@ -331,3 +331,25 @@ define i32 @load_i32_by_i8_neg_offset_bswap(i32* %arg) {
   %tmp18 = or i32 %tmp13, %tmp17
   ret i32 %tmp18
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p; // p is 4 byte aligned
+; (i32) bswap(p[0]) | (i32) bswap(p[1] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK: ldr   w8, [x0]
+; CHECK-NEXT: rev w0, w8
+; CHECK-NEXT: ret
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp5, 16
+  %tmp7 = or i32 %tmp6, %tmp2
+  ret i32 %tmp7
+}

diff --git a/test/CodeGen/AArch64/load-combine.ll b/test/CodeGen/AArch64/load-combine.ll
index 86461e2a1d73ee26f757e91c0f5c4d4b4c2bd9af..baa9628d4e322d46c8c19a618340ab53c617dd9f 100644 (file)
--- a/test/CodeGen/AArch64/load-combine.ll
+++ b/test/CodeGen/AArch64/load-combine.ll
@@ -318,3 +318,26 @@ define i32 @load_i32_by_i8_neg_offset_bswap(i32* %arg) {
   %tmp18 = or i32 %tmp13, %tmp17
   ret i32 %tmp18
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p; // p is 4 byte aligned
+; (i32) bswap(p[1]) | (i32) bswap(p[0] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK: ldr    w8, [x0]
+; CHECK-NEXT: rev w0, w8
+; CHECK-NEXT: ret
+
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp2, 16
+  %tmp7 = or i32 %tmp6, %tmp5
+  ret i32 %tmp7
+}

diff --git a/test/CodeGen/ARM/load-combine-big-endian.ll b/test/CodeGen/ARM/load-combine-big-endian.ll
index 795e69fc4f589a66e4f12da36258f25544197e10..b5a405df9679888b54d04f6c2658844db9a2df41 100644 (file)
--- a/test/CodeGen/ARM/load-combine-big-endian.ll
+++ b/test/CodeGen/ARM/load-combine-big-endian.ll
@@ -449,3 +449,37 @@ define i32 @load_i32_by_i8_neg_offset_bswap(i32* %arg) {
   %tmp18 = or i32 %tmp13, %tmp17
   ret i32 %tmp18
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p; // p is 4 byte aligned
+; (i32) bswap(p[0]) | (i32) bswap(p[1] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK: ldr r0, [r0]
+; CHECK-NEXT: mov r1, #65280
+; CHECK-NEXT: mov r2, #16711680
+; CHECK-NEXT: and r1, r1, r0, lsr #8
+; CHECK-NEXT: and r2, r2, r0, lsl #8
+; CHECK-NEXT: orr r1, r1, r0, lsr #24
+; CHECK-NEXT: orr r0, r2, r0, lsl #24
+; CHECK-NEXT: orr r0, r0, r1
+; CHECK-NEXT: mov pc, lr
+
+; CHECK-ARMv6-LABEL: load_i32_by_bswap_i16:
+; CHECK-ARMv6: ldr  r0, [r0]
+; CHECK-ARMv6-NEXT: rev r0, r0
+; CHECK-ARMv6-NEXT: bx  lr
+
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp5, 16
+  %tmp7 = or i32 %tmp6, %tmp2
+  ret i32 %tmp7
+}

diff --git a/test/CodeGen/ARM/load-combine.ll b/test/CodeGen/ARM/load-combine.ll
index bafa13894c69495d0ce23a67f9a13b33c621056f..eb476b673b186d362b4c1a3bdce7ebc2f98d91c5 100644 (file)
--- a/test/CodeGen/ARM/load-combine.ll
+++ b/test/CodeGen/ARM/load-combine.ll
@@ -407,3 +407,37 @@ define i32 @load_i32_by_i8_neg_offset_bswap(i32* %arg) {
   %tmp18 = or i32 %tmp13, %tmp17
   ret i32 %tmp18
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p; // p is 4 byte aligned
+; (i32) bswap(p[1]) | (i32) bswap(p[0] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK: ldr  r0, [r0]
+; CHECK-NEXT: mov r1, #65280
+; CHECK-NEXT: mov r2, #16711680
+; CHECK-NEXT: and r1, r1, r0, lsr #8
+; CHECK-NEXT: and r2, r2, r0, lsl #8
+; CHECK-NEXT: orr r1, r1, r0, lsr #24
+; CHECK-NEXT: orr r0, r2, r0, lsl #24
+; CHECK-NEXT: orr r0, r0, r1
+; CHECK-NEXT: mov pc, lr
+
+; CHECK-ARMv6-LABEL: load_i32_by_bswap_i16:
+; CHECK-ARMv6: ldr r0, [r0]
+; CHECK-ARMv6-NEXT: rev r0, r0
+; CHECK-ARMv6-NEXT: bx  lr
+
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp2, 16
+  %tmp7 = or i32 %tmp6, %tmp5
+  ret i32 %tmp7
+}

diff --git a/test/CodeGen/X86/load-combine.ll b/test/CodeGen/X86/load-combine.ll
index 8942ca09c144dbbeb9cda4f1c8e4ffbaceff7fdc..7c88165305786826e1863b639d2b9b0bd078ad1a 100644 (file)
--- a/test/CodeGen/X86/load-combine.ll
+++ b/test/CodeGen/X86/load-combine.ll
@@ -869,3 +869,35 @@ entry:
   store i64 %conv75, i64* %dst, align 8
   ret void
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p;
+; (i32) bswap(p[1]) | (i32) bswap(p[0] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; CHECK-NEXT:    movl (%eax), %eax
+; CHECK-NEXT:    bswapl %eax
+; CHECK-NEXT:    retl
+;
+; CHECK64-LABEL: load_i32_by_bswap_i16:
+; CHECK64:       # BB#0:
+; CHECK64-NEXT:    movl (%rdi), %eax
+; CHECK64-NEXT:    bswapl %eax
+; CHECK64-NEXT:    retq
+
+
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp2, 16
+  %tmp7 = or i32 %tmp6, %tmp5
+  ret i32 %tmp7
+}