Don't fold indexed loads into TCRETURNmi64.

We don't have enough GR64_TC registers when calling a varargs function
with 6 arguments. Since %al holds the number of vector registers used,
only %r11 is available as a scratch register.

This means that addressing modes using both base and index registers
can't be folded into TCRETURNmi64.

<rdar://problem/12282281>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@163761 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 74 insertions, 2 deletions

diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index d836c29d6b..88cfe10dfa 100644
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -204,6 +204,9 @@ namespace {
     bool SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
                     SDValue &Scale, SDValue &Index, SDValue &Disp,
                     SDValue &Segment);
+    bool SelectSingleRegAddr(SDNode *Parent, SDValue N, SDValue &Base,
+                             SDValue &Scale, SDValue &Index, SDValue &Disp,
+                             SDValue &Segment);
     bool SelectLEAAddr(SDValue N, SDValue &Base,
                        SDValue &Scale, SDValue &Index, SDValue &Disp,
                        SDValue &Segment);
@@ -1319,6 +1322,31 @@ bool X86DAGToDAGISel::SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
   return true;
 }
 
+/// SelectSingleRegAddr - Like SelectAddr, but reject any address that would
+/// require more than one allocatable register.
+///
+/// This is used for a TCRETURNmi64 instruction when used to tail call a
+/// variadic function with 6 arguments: Only %r11 is available from GR64_TC.
+/// The other scratch register, %rax, is needed to pass in the number of vector
+/// registers used in the variadic arguments.
+///
+bool X86DAGToDAGISel::SelectSingleRegAddr(SDNode *Parent, SDValue N,
+                                          SDValue &Base,
+                                          SDValue &Scale, SDValue &Index,
+                                          SDValue &Disp, SDValue &Segment) {
+  if (!SelectAddr(Parent, N, Base, Scale, Index, Disp, Segment))
+    return false;
+  // Anything %RIP relative is fine.
+  if (RegisterSDNode *Reg = dyn_cast<RegisterSDNode>(Base))
+    if (Reg->getReg() == X86::RIP)
+      return true;
+  // Check that the index register is 0.
+  if (RegisterSDNode *Reg = dyn_cast<RegisterSDNode>(Index))
+    if (Reg->getReg() == 0)
+      return true;
+  return false;
+}
+
 /// SelectScalarSSELoad - Match a scalar SSE load.  In particular, we want to
 /// match a load whose top elements are either undef or zeros.  The load flavor
 /// is derived from the type of N, which is either v4f32 or v2f64.
diff --git a/lib/Target/X86/X86InstrCompiler.td b/lib/Target/X86/X86InstrCompiler.td
index 99c2b8f955..2540455785 100644
--- a/lib/Target/X86/X86InstrCompiler.td
+++ b/lib/Target/X86/X86InstrCompiler.td
@@ -1041,7 +1041,13 @@ def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
           (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
           Requires<[In64BitMode]>;
 
-def : Pat<(X86tcret (load addr:$dst), imm:$off),
+// When calling a variadic function with 6 arguments, 7 scratch registers are
+// needed since %al holds the number of vector registers used. That leaves %r11
+// as the only remaining GR64_TC register for the addressing mode.
+//
+// The single_reg_addr pattern rejects any addressing modes that would need
+// more than one register.
+def : Pat<(X86tcret (load single_reg_addr:$dst), imm:$off),
           (TCRETURNmi64 addr:$dst, imm:$off)>,
           Requires<[In64BitMode]>;
 
diff --git a/lib/Target/X86/X86InstrInfo.td b/lib/Target/X86/X86InstrInfo.td
index aabb442f74..b91f3c0ad4 100644
--- a/lib/Target/X86/X86InstrInfo.td
+++ b/lib/Target/X86/X86InstrInfo.td
@@ -543,6 +543,10 @@ def tls64addr : ComplexPattern<i64, 5, "SelectTLSADDRAddr",
 def tls64baseaddr : ComplexPattern<i64, 5, "SelectTLSADDRAddr",
                                [tglobaltlsaddr], []>;
 
+// Same as addr, but reject addressing modes requiring more than one register.
+def single_reg_addr : ComplexPattern<iPTR, 5, "SelectSingleRegAddr", [],
+                                     [SDNPWantParent]>;
+
 //===----------------------------------------------------------------------===//
 // X86 Instruction Predicate Definitions.
 def HasCMov      : Predicate<"Subtarget->hasCMov()">;
diff --git a/test/CodeGen/X86/tailcall-64.ll b/test/CodeGen/X86/tailcall-64.ll
index 7030753415..ffbc4b8b3f 100644
--- a/test/CodeGen/X86/tailcall-64.ll
+++ b/test/CodeGen/X86/tailcall-64.ll
@@ -1,4 +1,4 @@
-; RUN: llc < %s | FileCheck %s
+; RUN: llc < %s -verify-machineinstrs | FileCheck %s
 target datalayout = "e-p:64:64:64-S128-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f16:16:16-f32:32:32-f64:64:64-f128:128:128-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
 target triple = "x86_64-apple-darwin11.4.0"
 
@@ -93,4 +93,38 @@ define { i64, i64 } @crash(i8* %this) {
   ret { i64, i64 } %mrv7
 }
 
+; <rdar://problem/12282281> Fold an indexed load into the tail call instruction.
+; Calling a varargs function with 6 arguments requires 7 registers (%al is the
+; vector count for varargs functions). This leaves %r11 as the only available
+; scratch register.
+;
+; It is not possible to fold an indexed load into TCRETURNmi64 in that case.
+;
+; typedef int (*funcptr)(void*, ...);
+; extern const funcptr funcs[];
+; int f(int n) {
+;   return funcs[n](0, 0, 0, 0, 0, 0);
+; }
+;
+; CHECK: rdar12282281
+; CHECK: jmpq *%r11 # TAILCALL
+@funcs = external constant [0 x i32 (i8*, ...)*]
+
+define i32 @rdar12282281(i32 %n) nounwind uwtable ssp {
+entry:
+  %idxprom = sext i32 %n to i64
+  %arrayidx = getelementptr inbounds [0 x i32 (i8*, ...)*]* @funcs, i64 0, i64 %idxprom
+  %0 = load i32 (i8*, ...)** %arrayidx, align 8
+  %call = tail call i32 (i8*, ...)* %0(i8* null, i32 0, i32 0, i32 0, i32 0, i32 0) nounwind
+  ret i32 %call
+}
 
+; Same thing, using a fixed offset. The load should foid.
+; CHECK: rdar12282281fixed
+; CHECK: jmpq *8(%r11) # TAILCALL
+define i32 @rdar12282281fixed() nounwind uwtable ssp {
+entry:
+  %0 = load i32 (i8*, ...)** getelementptr inbounds ([0 x i32 (i8*, ...)*]* @funcs, i64 0, i64 1), align 8
+  %call.i = tail call i32 (i8*, ...)* %0(i8* null, i32 0, i32 0, i32 0, i32 0, i32 0) nounwind
+  ret i32 %call.i
+}