[llvm-dev] Loop invariant not being optimized

[Nema, Ashutosh via llvm-dev]

If I understood it correctly, __declspec(noalias) is not the same as specifying restrict on each parameter.

It means in the mentioned example a, b & c don't modify or reference any global state, but they are free to alias one another.

You could specify restrict on each one to indicate that they do not alias each other.

For more details refer: https://msdn.microsoft.com/en-us/library/k649tyc7.aspx

Regards,
Ashutosh

From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of Phil Tomson via llvm-dev
Sent: Friday, November 18, 2016 12:23 AM
To: LLVM Developers Mailing List <llvm-dev at lists.llvm.org>
Subject: [llvm-dev] Loop invariant not being optimized

I've got an example where I think that there should be some loop-invariant optimization happening, but it's not.  Here's the C code:

#define DIM 8
#define UNROLL_DIM DIM
typedef double InArray[DIM][DIM];

__declspec(noalias) void f1( InArray c, const InArray a, const InArray b )
{

#pragma clang loop unroll_count(UNROLL_DIM)
    for( int i=0;i<DIM;i++)
#pragma clang loop unroll_count(UNROLL_DIM)
        for( int j=0;j<DIM;j++)
#pragma clang loop  unroll_count(UNROLL_DIM)
            for( int k=0;k<DIM;k++) {
                c[i][k] = c[i][k] + a[i][j]*b[j][k];
            }
}
The "a[i][j]" there is invariant in that inner loop. I've unrolled the loops with the unroll pragma to make the assembly easier to read, here's what I see (LVM 3.9, compiling with: clang -fms-compatibility -funroll-loops -O3   -c fma.c -o fma.o )


0000000000000000 <f1>:
       0: 29580c0000000000  load  r3,r0,0x0,64
       8: 2958100200000000  load  r4,r1,0x0,64 #r4 <- a[0][0]
      10: 2958140400000000  load  r5,r2,0x0,64
      18: c0580c0805018000  fmaf  r3,r4,r5,r3,64
      20: 79b80c0000000000  store r3,r0,0x0,64
      28: 2958100000000008  load  r4,r0,0x8,64
      30: 2958140200000000  load  r5,r1,0x0,64 #r5 <- a[0][0]
      38: 2958180400000008  load  r6,r2,0x8,64
      40: c058100a06020000  fmaf  r4,r5,r6,r4,64
      48: 79b8100000000008  store r4,r0,0x8,64
      50: 2958140000000010  load  r5,r0,0x10,64
      58: 2958180200000000  load  r6,r1,0x0,64 #r6 <- a[0][0]
      60: 29581c0400000010  load  r7,r2,0x10,64
      68: c058140c07028000  fmaf  r5,r6,r7,r5,64
      70: 79b8140000000010  store r5,r0,0x10,64
      78: 2958180000000018  load  r6,r0,0x18,64
      80: 29581c0200000000  load  r7,r1,0x0,64 #r7 <- a[0][0]
      88: 2958200400000018  load  r8,r2,0x18,64
      90: c058180e08030000  fmaf  r6,r7,r8,r6,64
...
(fmaf semantics are: fmaf r1,r2,r3,r4, SIZE    r1 <- r2*r3+r4 )
(load semantics are: load r1,r2,imm, SIZE     r1<- mem[r2+imm] )

All three of the addresses are loaded in every loop. Only two need to be reloaded in the inner loop. I added the 'noalias' declspec in the C code above thinking that it would indicate that the pointers going into the function are not aliased and that that would allow the optimization, but it didn't make any difference.
Of course it's easy to rewrite the example code to avoid this extra load/inner loop, but I would have thought this would be a fairly straighforward optimization for the optimizer. Am I missing something?
Phil

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161118/df4577e5/attachment.html>