[LLVMbugs] [Bug 16009] New: Inefficient code for load + mask

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=16009

            Bug ID: 16009
           Summary: Inefficient code for load + mask
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: silvas at purdue.edu
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

Both of these examples generate inefficient code. At least a part of the issue
seems to be that we fail to reduce an i32 load masked with 0xFF to a byte load
with zext.

/****** Case 1 ******/

extern void use(int);
void test_pointless_moving_and_masking(int *p) {
  use(p[0] & 0xFF);
}

IR:
define void @test_pointless_moving_and_masking(i32* nocapture %p) #0 {
  %1 = load i32* %p, align 4, !tbaa !0
  %2 = and i32 %1, 255
  tail call void @use(i32 %2) #2
  ret void
}

Our code is shockingly bad:
0000000000000000 <test_pointless_moving_and_masking>:
   0:   b8 ff 00 00 00          mov    eax,0xff
   5:   23 07                   and    eax,DWORD PTR [rdi]
   7:   89 c7                   mov    edi,eax
   9:   e9 00 00 00 00          jmp    e
<test_pointless_moving_and_masking+0xe>

GCC (4.7.3) produces far superior code:
0000000000000000 <test_pointless_moving_and_masking>:
   0:   0f b6 3f                movzx  edi,BYTE PTR [rdi]
   3:   e9 00 00 00 00          jmp    8
<test_pointless_moving_and_masking+0x8>


/****** Case 2 ******/

extern int *arr[0xFF];
extern void use_intp(int *);

void test_pointless_zext(int *p) {
  use_intp(arr[p[0] & 0xFF]);
}

IR:
define void @test_pointless_zext(i32* nocapture %p) #0 {
  %1 = load i32* %p, align 4, !tbaa !0
  %2 = and i32 %1, 255
  %3 = zext i32 %2 to i64
  %4 = getelementptr inbounds [255 x i32*]* @arr, i64 0, i64 %3
  %5 = load i32** %4, align 8, !tbaa !3
  tail call void @use_intp(i32* %5) #2
  ret void
}

0000000000000010 <test_pointless_zext>:
  10:   8b 07                   mov    eax,DWORD PTR [rdi]
  12:   0f b6 c0                movzx  eax,al
  15:   48 8b 3c c5 00 00 00 00 mov    rdi,QWORD PTR [rax*8+0x0]
  1d:   e9 00 00 00 00          jmp    22 <test_pointless_zext+0x12>


GCC (4.7.3) directly does a movzx with memory operand rather than a dword load
+ movzx:
0000000000000010 <test_pointless_zext>:
  10:   0f b6 07                movzx  eax,BYTE PTR [rdi]
  13:   48 8b 3c c5 00 00 00 00 mov    rdi,QWORD PTR [rax*8+0x0]
  1b:   e9 00 00 00 00          jmp    20 <test_pointless_zext+0x10>



Tested with clang is built from trunk on Apr 20, 2013. (this is all with -O2
and -O3)

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20130515/f3ce852b/attachment.html>