[RegAllocFast] Handle implicit definitions conservatively.

Prior to this commit, physical registers defined implicitly were considered free
right after their definition, i.e.. like dead definitions. Therefore, their uses
had to immediately follow their definitions, otherwise the related register may
be reused to allocate a virtual register.

This commit fixes this assumption by keeping implicit definitions alive until
they are actually used. The downside is that if the implicit definition was dead
(and not marked at such), we block an otherwise available register. This is
however conservatively correct and makes the fast register allocator much more
robust in particular regarding the scheduling of the instructions.

Fixes PR21700.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223317 91177308-0d34-0410-b5e6-96231b3b80d8

test/CodeGen/X86/crash-O0.ll

@@ -1,4 +1,4 @@
-; RUN: llc -O0 -relocation-model=pic -disable-fp-elim < %s
+; RUN: llc -O0 -relocation-model=pic -disable-fp-elim < %s | FileCheck %s
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
 target triple = "x86_64-apple-darwin10"
 
@@ -29,3 +29,23 @@ entry:
 "41":                                             ; preds = %"39"
   unreachable
 }
+
+; When using fast isel, sdiv is lowered into a sequence of CQO + DIV64.
+; CQO defines implicitly AX and DIV64 uses it implicitly too.
+; When an instruction gets between those two, RegAllocFast was reusing
+; AX for the vreg defined in between and the compiler crashed.
+;
+; An instruction gets between CQO and DIV64 because the load is folded
+; into the division but it requires a sign extension.
+; PR21700
+; CHECK-LABEL: addressModeWith32bitIndex:
+; CHECK: cqto
+; CHECK-NEXT: movslq
+; CHECK-NEXT: idivq
+; CHECK: retq
+define i64 @addressModeWith32bitIndex(i32 %V) {
+  %gep = getelementptr i64* null, i32 %V
+  %load = load i64* %gep
+  %sdiv = sdiv i64 0, %load
+  ret i64 %sdiv
+}