From: Sanjay Patel <spatel@rotateright.com>
Date: Sat, 13 Aug 2016 18:42:14 +0000 (+0000)
Subject: [x86] add tests to show missed 64-bit immediate merging
X-Git-Tag: llvmorg-4.0.0-rc1~12509
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=08c876673ef186627353655c2cf66c019b30f417;p=platform%2Fupstream%2Fllvm.git

[x86] add tests to show missed 64-bit immediate merging

Tests are slightly modified versions of those written by
Sunita Marathe in D23391.

llvm-svn: 278599
---

diff --git a/llvm/test/CodeGen/X86/immediate_merging64.ll b/llvm/test/CodeGen/X86/immediate_merging64.ll
new file mode 100644
index 0000000..b0d6666
--- /dev/null
+++ b/llvm/test/CodeGen/X86/immediate_merging64.ll
@@ -0,0 +1,54 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s
+
+; Check that multiple instances of 64-bit constants encodable as
+; 32-bit immediates are merged for code size savings.
+
+@g1 = common global i64 0, align 8
+@g2 = common global i64 0, align 8
+@g3 = common global i64 0, align 8
+@g4 = common global i64 0, align 8
+
+; Immediates with multiple users should not be pulled into instructions when
+; optimizing for code size.
+define void @imm_multiple_users(i64 %l1, i64 %l2, i64 %l3, i64 %l4) optsize {
+; CHECK-LABEL: imm_multiple_users:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    movq $-1, {{.*}}(%rip)
+; CHECK-NEXT:    cmpq $-1, %rdx
+; CHECK-NEXT:    cmovneq %rsi, %rdi
+; CHECK-NEXT:    movq %rdi, {{.*}}(%rip)
+; CHECK-NEXT:    movq $-1, %rax
+; CHECK-NEXT:    # kill: %CL<def> %CL<kill> %RCX<kill>
+; CHECK-NEXT:    shlq %cl, %rax
+; CHECK-NEXT:    movq %rax, {{.*}}(%rip)
+; CHECK-NEXT:    movq $0, {{.*}}(%rip)
+; CHECK-NEXT:    retq
+;
+  store i64 -1, i64* @g1, align 8
+  %cmp = icmp eq i64 %l3, -1
+  %sel = select i1 %cmp, i64 %l1, i64 %l2
+  store i64 %sel, i64* @g2, align 8
+  %and = and i64 %l4, 63
+  %shl = shl i64 -1, %and
+  store i64 %shl, i64* @g3, align 8
+  store i64 0, i64* @g4, align 8
+  ret void
+}
+
+declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1)
+
+; Inlined memsets requiring multiple same-sized stores should be lowered using
+; the register, rather than immediate, form of stores when optimizing for
+; code size.
+define void @memset_zero(i8* noalias nocapture %D) optsize {
+; CHECK-LABEL: memset_zero:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    movq $0, 7(%rdi)
+; CHECK-NEXT:    movq $0, (%rdi)
+; CHECK-NEXT:    retq
+;
+  tail call void @llvm.memset.p0i8.i64(i8* %D, i8 0, i64 15, i32 1, i1 false)
+  ret void
+}
+