From 7bba7d0efc69ff3da5a064a5d77513dd5aca0f15 Mon Sep 17 00:00:00 2001
From: Jakob Stoklund Olesen <stoklund@2pi.dk>
Date: Thu, 13 Sep 2012 18:31:27 +0000
Subject: [PATCH] Fix the TCRETURNmi64 bug differently.

Add a PatFrag to match X86tcret using 6 fixed registers or less. This
avoids folding loads into TCRETURNmi64 using 7 or more volatile
registers.

<rdar://problem/12282281>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@163819 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86InstrCompiler.td | 23 ++++++++++++++++--
 test/CodeGen/X86/tailcall-64.ll    | 39 ++++++++++++++++++++++++++++++
 2 files changed, 60 insertions(+), 2 deletions(-)

diff --git a/lib/Target/X86/X86InstrCompiler.td b/lib/Target/X86/X86InstrCompiler.td
index 99c2b8f955e..75ceeb9b182 100644
--- a/lib/Target/X86/X86InstrCompiler.td
+++ b/lib/Target/X86/X86InstrCompiler.td
@@ -1017,7 +1017,24 @@ def : Pat<(X86call (i64 tglobaladdr:$dst)),
 def : Pat<(X86call (i64 texternalsym:$dst)),
           (CALL64pcrel32 texternalsym:$dst)>;
 
-// tailcall stuff
+// Tailcall stuff. The TCRETURN instructions execute after the epilog, so they
+// can never use callee-saved registers. That is the purpose of the GR64_TC
+// register classes.
+//
+// The only volatile register that is never used by the calling convention is
+// %r11. This happens when calling a vararg function with 6 arguments.
+//
+// Match an X86tcret that uses less than 7 volatile registers.
+def X86tcret_6regs : PatFrag<(ops node:$ptr, node:$off),
+                             (X86tcret node:$ptr, node:$off), [{
+  // X86tcret args: (*chain, ptr, imm, regs..., glue)
+  unsigned NumRegs = 0;
+  for (unsigned i = 3, e = N->getNumOperands(); i != e; ++i)
+    if (isa<RegisterSDNode>(N->getOperand(i)) && ++NumRegs > 6)
+      return false;
+  return true;
+}]>;
+
 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
           (TCRETURNri ptr_rc_tailcall:$dst, imm:$off)>,
           Requires<[In32BitMode]>;
@@ -1041,7 +1058,9 @@ 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),
+// Don't fold loads into X86tcret requiring more than 6 regs.
+// There wouldn't be enough scratch registers for base+index.
+def : Pat<(X86tcret_6regs (load addr:$dst), imm:$off),
           (TCRETURNmi64 addr:$dst, imm:$off)>,
           Requires<[In64BitMode]>;
 
diff --git a/test/CodeGen/X86/tailcall-64.ll b/test/CodeGen/X86/tailcall-64.ll
index 70307534156..df7af69a996 100644
--- a/test/CodeGen/X86/tailcall-64.ll
+++ b/test/CodeGen/X86/tailcall-64.ll
@@ -93,4 +93,43 @@ define { i64, i64 } @crash(i8* %this) {
   ret { i64, i64 } %mrv7
 }
 
+; Check that we can fold an indexed load into a tail call instruction.
+; CHECK: fold_indexed_load
+; CHECK: leaq (%rsi,%rsi,4), %[[RAX:r..]]
+; CHECK: movq _func_table@GOTPCREL(%rip), %[[RCX:r..]]
+; CHECK: jmpq *16(%[[RCX]],%[[RAX]],8)  # TAILCALL
+%struct.funcs = type { i32 (i8*, i32*, i32)*, i32 (i8*)*, i32 (i8*)*, i32 (i8*, i32)*, i32 }
+@func_table = external global [0 x %struct.funcs]
+define void @fold_indexed_load(i8* %mbstr, i64 %idxprom) nounwind uwtable ssp {
+entry:
+  %dsplen = getelementptr inbounds [0 x %struct.funcs]* @func_table, i64 0, i64 %idxprom, i32 2
+  %x1 = load i32 (i8*)** %dsplen, align 8
+  %call = tail call i32 %x1(i8* %mbstr) nounwind
+  ret void
+}
 
+; <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
+}
-- 
2.34.1