From e5dacc55ad0122eb4b8913955cf6339fa3ff2d6e Mon Sep 17 00:00:00 2001
From: Dan Gohman <gohman@apple.com>
Date: Mon, 11 Jan 2010 17:58:34 +0000
Subject: [PATCH] Use a 32-bit and with implicit zero-extension instead of a
 64-bit and if it has an immediate with at least 32 bits of leading zeros, to
 avoid needing to materialize that immediate in a register first.

FileCheckize, tidy, and extend a testcase to cover this case.

This fixes rdar://7527390.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@93160 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86Instr64bit.td     | 11 ++++++++
 test/CodeGen/X86/x86-64-and-mask.ll | 43 +++++++++++++++++++++++++++--
 2 files changed, 51 insertions(+), 3 deletions(-)

diff --git a/lib/Target/X86/X86Instr64bit.td b/lib/Target/X86/X86Instr64bit.td
index 1835c8feb62..bc36be21955 100644
--- a/lib/Target/X86/X86Instr64bit.td
+++ b/lib/Target/X86/X86Instr64bit.td
@@ -1966,6 +1966,17 @@ def : Pat<(add GR64:$src1, 0x0000000080000000),
 def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
           (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
 
+// Use a 32-bit and with implicit zero-extension instead of a 64-bit and if it
+// has an immediate with at least 32 bits of leading zeros, to avoid needing to
+// materialize that immediate in a register first.
+def : Pat<(and GR64:$src, i64immZExt32:$imm),
+          (SUBREG_TO_REG
+            (i64 0),
+            (AND32ri
+              (EXTRACT_SUBREG GR64:$src, x86_subreg_32bit),
+              imm:$imm),
+            x86_subreg_32bit)>;
+
 // r & (2^32-1) ==> movz
 def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
           (MOVZX64rr32 (EXTRACT_SUBREG GR64:$src, x86_subreg_32bit))>;
diff --git a/test/CodeGen/X86/x86-64-and-mask.ll b/test/CodeGen/X86/x86-64-and-mask.ll
index 3c738911126..2465f23a768 100644
--- a/test/CodeGen/X86/x86-64-and-mask.ll
+++ b/test/CodeGen/X86/x86-64-and-mask.ll
@@ -1,12 +1,49 @@
-; RUN: llc < %s | grep {movl.*%edi, %eax}
-; This should be a single mov, not a load of immediate + andq.
+; RUN: llc < %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"
 target triple = "x86_64-apple-darwin8"
 
-define i64 @test(i64 %x) nounwind  {
+; This should be a single mov, not a load of immediate + andq.
+; CHECK: test:
+; CHECK: movl %edi, %eax
+
+define i64 @test(i64 %x) nounwind {
 entry:
 	%tmp123 = and i64 %x, 4294967295		; <i64> [#uses=1]
 	ret i64 %tmp123
 }
 
+; This copy can't be coalesced away because it needs the implicit zero-extend.
+; CHECK: bbb:
+; CHECK: movl %edi, %edi
+
+define void @bbb(i64 %x) nounwind {
+  %t = and i64 %x, 4294967295
+  call void @foo(i64 %t)
+  ret void
+}
+
+; This should use a 32-bit and with implicit zero-extension, not a 64-bit and
+; with a separate mov to materialize the mask.
+; rdar://7527390
+; CHECK: ccc:
+; CHECK: andl $-1048593, %edi
+
+declare void @foo(i64 %x) nounwind
+
+define void @ccc(i64 %x) nounwind {
+  %t = and i64 %x, 4293918703
+  call void @foo(i64 %t)
+  ret void
+}
+
+; This requires a mov and a 64-bit and.
+; CHECK: ddd:
+; CHECK: movabsq $4294967296, %rax
+; CHECK: andq %rax, %rdi
+
+define void @ddd(i64 %x) nounwind {
+  %t = and i64 %x, 4294967296
+  call void @foo(i64 %t)
+  ret void
+}
-- 
2.34.1