[llvm] r304163 - [X86] Don't fold away the memory operand of an xchg.

[Benjamin Kramer via llvm-commits]

Author: d0k
Date: Mon May 29 11:25:20 2017
New Revision: 304163

URL: http://llvm.org/viewvc/llvm-project?rev=304163&view=rev
Log:
[X86] Don't fold away the memory operand of an xchg.

xchg with a mem operand has different locking semantics. If we unfold it
into a xchg r,r we will loose the implicit lock. Likewise we never want
to fold a register xchg into a memory one as it would be a lot slower.

This triggers during LLVM selfhost.

Added:
    llvm/trunk/test/CodeGen/X86/xchg-nofold.ll
Modified:
    llvm/trunk/utils/TableGen/X86FoldTablesEmitter.cpp

Added: llvm/trunk/test/CodeGen/X86/xchg-nofold.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/xchg-nofold.ll?rev=304163&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/xchg-nofold.ll (added)
+++ llvm/trunk/test/CodeGen/X86/xchg-nofold.ll Mon May 29 11:25:20 2017
@@ -0,0 +1,37 @@
+; RUN: llc -mtriple=x86_64-linux-gnu < %s | FileCheck %s
+
+%"struct.std::atomic" = type { %"struct.std::atomic_bool" }
+%"struct.std::atomic_bool" = type { %"struct.std::__atomic_base" }
+%"struct.std::__atomic_base" = type { i8 }
+
+; CHECK-LABEL: _Z3fooRSt6atomicIbEb
+define zeroext i1 @_Z3fooRSt6atomicIbEb(%"struct.std::atomic"* nocapture dereferenceable(1) %a, i1 returned zeroext %b) nounwind {
+entry:
+  %frombool.i.i = zext i1 %b to i8
+  %_M_i.i.i = getelementptr inbounds %"struct.std::atomic", %"struct.std::atomic"* %a, i64 0, i32 0, i32 0, i32 0
+  %0 = ptrtoint i8* %_M_i.i.i to i64
+  %1 = lshr i64 %0, 3
+  %2 = add i64 %1, 2147450880
+  %3 = inttoptr i64 %2 to i8*
+  %4 = load i8, i8* %3
+  %5 = icmp ne i8 %4, 0
+  br i1 %5, label %6, label %11
+
+; <label>:6:                                      ; preds = %entry
+  %7 = and i64 %0, 7
+  %8 = trunc i64 %7 to i8
+  %9 = icmp sge i8 %8, %4
+  br i1 %9, label %10, label %11
+
+; <label>:10:                                     ; preds = %6
+  call void @__asan_report_store1(i64 %0)
+  call void asm sideeffect "", ""()
+  unreachable
+
+; <label>:11:                                     ; preds = %6, %entry
+  store atomic i8 %frombool.i.i, i8* %_M_i.i.i seq_cst, align 1
+; CHECK: xchgb	%{{.*}}, (%{{.*}})
+  ret i1 %b
+}
+
+declare void @__asan_report_store1(i64)

Modified: llvm/trunk/utils/TableGen/X86FoldTablesEmitter.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/utils/TableGen/X86FoldTablesEmitter.cpp?rev=304163&r1=304162&r2=304163&view=diff
==============================================================================
--- llvm/trunk/utils/TableGen/X86FoldTablesEmitter.cpp (original)
+++ llvm/trunk/utils/TableGen/X86FoldTablesEmitter.cpp Mon May 29 11:25:20 2017
@@ -77,7 +77,14 @@ const char *const NoFoldSet[] = {
     "TCRETURNri64",
     "TCRETURNmi64", // Special dealing (in X86InstrCompiler.td under
     "TCRETURNri",   // "tailcall stuff" section).
-    "TCRETURNmi"
+    "TCRETURNmi",
+
+    // Never fold XCHG, the register and memory forms have different locking
+    // semantics.
+    "XCHG8rr",  "XCHG8rm",
+    "XCHG16rr", "XCHG16rm",
+    "XCHG32rr", "XCHG32rm",
+    "XCHG64rr", "XCHG64rm",
 
     // Different calculations of the folded operand between
     // memory and register forms (folding is illegal).