From aa8b9993c23f9915f1ba694502333f67a627e8d0 Mon Sep 17 00:00:00 2001
From: Jessica Paquette <jpaquette@apple.com>
Date: Fri, 26 Jul 2019 23:28:53 +0000
Subject: [PATCH] [AArch64][GlobalISel] Select @llvm.aarch64.stlxr for 32-bit
 pointers

Add partial instruction selection for intrinsics like this:

```
declare i32 @llvm.aarch64.stlxr(i64, i32*)
```

(This only handles the case where a G_ZEXT is feeding the intrinsic.)

Also make sure that the added store instruction actually has the memory op from
the original G_STORE.

Update select-stlxr-intrin.mir and arm64-ldxr-stxr.ll.

Differential Revision: https://reviews.llvm.org/D65355

llvm-svn: 367163
---
 .../AArch64/AArch64InstructionSelector.cpp    | 24 +++++++++++++--
 .../GlobalISel/select-stlxr-intrin.mir        | 30 ++++++++++++++++++-
 llvm/test/CodeGen/AArch64/arm64-ldxr-stxr.ll  |  5 ++++
 3 files changed, 55 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp b/llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp
index 2a45ee662ef7..4f2b4dc7f468 100644
--- a/llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp
+++ b/llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp
@@ -3892,7 +3892,9 @@ static unsigned findIntrinsicID(MachineInstr &I) {
 /// intrinsic.
 static unsigned getStlxrOpcode(unsigned NumBytesToStore) {
   switch (NumBytesToStore) {
-  // TODO: 1, 2, and 4 byte stores.
+  // TODO: 1 and 2 byte stores
+  case 4:
+    return AArch64::STLXRW;
   case 8:
     return AArch64::STLXRX;
   default:
@@ -3946,8 +3948,24 @@ bool AArch64InstructionSelector::selectIntrinsicWithSideEffects(
     unsigned Opc = getStlxrOpcode(NumBytesToStore);
     if (!Opc)
       return false;
-
-    auto StoreMI = MIRBuilder.buildInstr(Opc, {StatReg}, {SrcReg, PtrReg});
+    unsigned NumBitsToStore = NumBytesToStore * 8;
+    if (NumBitsToStore != 64) {
+      // The intrinsic always has a 64-bit source, but we might actually want
+      // a differently-sized source for the instruction. Try to get it.
+      // TODO: For 1 and 2-byte stores, this will have a G_AND. For now, let's
+      // just handle 4-byte stores.
+      // TODO: If we don't find a G_ZEXT, we'll have to truncate the value down
+      // to the right size for the STLXR.
+      MachineInstr *Zext = getOpcodeDef(TargetOpcode::G_ZEXT, SrcReg, MRI);
+      if (!Zext)
+        return false;
+      SrcReg = Zext->getOperand(1).getReg();
+      // We should get an appropriately-sized register here.
+      if (RBI.getSizeInBits(SrcReg, MRI, TRI) != NumBitsToStore)
+        return false;
+    }
+    auto StoreMI = MIRBuilder.buildInstr(Opc, {StatReg}, {SrcReg, PtrReg})
+                       .addMemOperand(*I.memoperands_begin());
     constrainSelectedInstRegOperands(*StoreMI, TII, TRI, RBI);
   }
 
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/select-stlxr-intrin.mir b/llvm/test/CodeGen/AArch64/GlobalISel/select-stlxr-intrin.mir
index dd775a20b38e..d60015d5d98e 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/select-stlxr-intrin.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/select-stlxr-intrin.mir
@@ -5,6 +5,10 @@
   define i32 @test_store_release_i64(i32 %a, i64* %addr) {
     ret i32 %a
   }
+
+  define i32 @test_store_release_i32(i32 %a, i64* %addr) {
+    ret i32 %a
+  }
 ...
 ---
 name:            test_store_release_i64
@@ -20,7 +24,7 @@ body:             |
     ; CHECK: liveins: $w0, $x1, $x2
     ; CHECK: [[COPY:%[0-9]+]]:gpr64 = COPY $x1
     ; CHECK: [[COPY1:%[0-9]+]]:gpr64sp = COPY $x2
-    ; CHECK: early-clobber %2:gpr32 = STLXRX [[COPY]], [[COPY1]]
+    ; CHECK: early-clobber %2:gpr32 = STLXRX [[COPY]], [[COPY1]] :: (volatile store 8 into %ir.addr)
     ; CHECK: $w0 = COPY %2
     ; CHECK: RET_ReallyLR implicit $w0
     %1:gpr(s64) = COPY $x1
@@ -30,3 +34,27 @@ body:             |
     RET_ReallyLR implicit $w0
 
 ...
+---
+name:            test_store_release_i32
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $w0, $w1, $x2
+    ; CHECK-LABEL: name: test_store_release_i32
+    ; CHECK: liveins: $w0, $w1, $x2
+    ; CHECK: [[COPY:%[0-9]+]]:gpr32 = COPY $w1
+    ; CHECK: [[COPY1:%[0-9]+]]:gpr64sp = COPY $x2
+    ; CHECK: early-clobber %3:gpr32 = STLXRW [[COPY]], [[COPY1]] :: (volatile store 4 into %ir.addr)
+    ; CHECK: $w0 = COPY %3
+    ; CHECK: RET_ReallyLR implicit $w0
+    %1:gpr(s32) = COPY $w1
+    %2:gpr(p0) = COPY $x2
+    %3:gpr(s64) = G_ZEXT %1(s32)
+    %4:gpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.aarch64.stlxr), %3(s64), %2(p0) :: (volatile store 4 into %ir.addr)
+    $w0 = COPY %4(s32)
+    RET_ReallyLR implicit $w0
+
+...
diff --git a/llvm/test/CodeGen/AArch64/arm64-ldxr-stxr.ll b/llvm/test/CodeGen/AArch64/arm64-ldxr-stxr.ll
index 4f87121c6ae4..fffcbe9fbf8a 100644
--- a/llvm/test/CodeGen/AArch64/arm64-ldxr-stxr.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-ldxr-stxr.ll
@@ -248,11 +248,16 @@ define i32 @test_store_release_i16(i32, i16 %val, i16* %addr) {
   ret i32 %res
 }
 
+; FALLBACK-NOT: remark:{{.*}}test_store_release_i32
 define i32 @test_store_release_i32(i32, i32 %val, i32* %addr) {
 ; CHECK-LABEL: test_store_release_i32:
 ; CHECK-NOT: uxtw
 ; CHECK-NOT: and
 ; CHECK: stlxr w0, w1, [x2]
+; GISEL-LABEL: test_store_release_i32:
+; GISEL-NOT: uxtw
+; GISEL-NOT: and
+; GISEL: stlxr w0, w1, [x2]
   %extval = zext i32 %val to i64
   %res = call i32 @llvm.aarch64.stlxr.p0i32(i64 %extval, i32* %addr)
   ret i32 %res