Add pattern to canonicalize for loop bounds

[bondhugula]

• add pattern to canonicalize affine.for loop bounds (using canonicalizeMapAndOperands)
• rename AffineForLoopBoundFolder -> AffineForOpBoundFolder for consistency

Signed-off-by: Uday Bondhugula uday@polymagelabs.com

[bondhugula]

Fixed merge conflicts. @nicolasvasilache @River707

[joker-eph]

I see failures on the examples tests, for example:

mlir/examples/Linalg/Linalg3/Example.cpp:65:12: error: CHECK: expected string not found in input
 // CHECK: affine.for %{{.*}} = 0 to (d0) -> (d0)(%[[N]]) {
           ^
<stdin>:10:2: note: scanning from here
 affine.for %arg3 = 0 to %1 {
 ^
<stdin>:10:2: note: with "N" equal to "1"
 affine.for %arg3 = 0 to %1 {
 ^

[bondhugula]

I see failures on the examples tests, for example:

mlir/examples/Linalg/Linalg3/Example.cpp:65:12: error: CHECK: expected string not found in input
 // CHECK: affine.for %{{.*}} = 0 to (d0) -> (d0)(%[[N]]) {
           ^
<stdin>:10:2: note: scanning from here
 affine.for %arg3 = 0 to %1 {
 ^
<stdin>:10:2: note: with "N" equal to "1"
 affine.for %arg3 = 0 to %1 {
 ^

How do I reproduce these? I already have the default cmake conf (per README.md) with LLVM_BUILD_EXAMPLES=ON, and check-mlir passes for me.

[joker-eph]

How do I reproduce these? I already have the default cmake conf (per README.md) with LLVM_BUILD_EXAMPLES=ON, and check-mlir passes for me.

That's why I kicked the "Kokoro" builds earlier to see if this is an internal only issue, and these builds seem to be passing. I'll involve @nicolasvasilache to debug this! Sorry for the delay this is causing!

[bondhugula]

How do I reproduce these? I already have the default cmake conf (per README.md) with LLVM_BUILD_EXAMPLES=ON, and check-mlir passes for me.

That's why I kicked the "Kokoro" builds earlier to see if this is an internal only issue, and these builds seem to be passing. I'll involve @nicolasvasilache to debug this! Sorry for the delay this is causing!

This one is pretty easy to fix. As a result of canonicalization, (d0) -> (d0) here turns into ()[s0] -> (s0), and (s0) -> (s0)(%[[N]]) will just be printed as %[[N]]. The check line will have to be updated to:

// CHECK: affine.for %{{.*}} = 0 to %[[N]] {

But I wanted a way to have these missing tests run - so that I could catch these and any more on my own.

[nicolasvasilache]

Took a quick look, mlir/examples/Linalg/Linalg3/Example.cpp and mlir/examples/Linalg/Linalg4/Example.cpp should be very easy to fix "with a blindfold". They all seem to be instances of map + parameter ub that should just become the parameter:

 // CHECK: affine.for %{{.*}} = 0 to (d0) -> (d0)(%[[N]]) {
           ^
<stdin>:93:2: note: scanning from here
 affine.for %arg3 = 0 to %1 {
 ^
<stdin>:93:2: note: with "N" equal to "1"
 affine.for %arg3 = 0 to %1 {

and

 // CHECK: affine.for %{{.*}} = 0 to (d0) -> (d0)(%[[M]]) step 8 {
           ^
<stdin>:32:2: note: scanning from here
 affine.for %arg3 = 0 to %0 step 8 {
 ^
<stdin>:32:2: note: with "M" equal to "0"
 affine.for %arg3 = 0 to %0 step 8 {
 ^

Maybe take a quick crack at them in this CL? Otherwise I can pick it up Monday when I know how we proceed with making changes to a CL from another user.

[bondhugula]

Fixing it isn't the issue! I was asking how I could run these tests in the first place? (Please see above - they aren't running as part of chrck-mlir.) would like to avoid a situation where some more test cases to update are found after a few days.

[jpienaar]

Fixing it isn't the issue! I was asking how I could run these tests in the first place? (Please see above - they aren't running as part of chrck-mlir.) would like to avoid a situation where some more test cases to update are found after a few days.

It seems there is a pattern missing that groups these into a convenient cmake target, but you can run them individually. E.,g,

ninja linalg-example-3 ./bin/linalg-example-3 | ./bin/FileCheck ../mlir/examples/Linalg/Linalg3/Example.cpp

[bondhugula]

Fixing it isn't the issue! I was asking how I could run these tests in the first place? (Please see above - they aren't running as part of chrck-mlir.) would like to avoid a situation where some more test cases to update are found after a few days.

It seems there is a pattern missing that groups these into a convenient cmake target, but you can run them individually. E.,g,

ninja linalg-example-3 ./bin/linalg-example-3 | ./bin/FileCheck ../mlir/examples/Linalg/Linalg3/Example.cpp

Thanks very much! This is exactly what I was looking for! Fixed now.

[joker-eph]

We still see the other Linalg example failing apparently:

/build/work/a961fabe6974af72c61c07bcbfdfb8f5a2ba/google3/runfiles/google3/third_party/llvm/llvm/projects/google_mlir/examples/Linalg/Linalg4/Example.cpp:71:12: error: CHECK: expected string not found in input
 // CHECK: affine.for %{{.*}} = 0 to (d0) -> (d0)(%[[M]]) step 8 {
           ^
<stdin>:7:2: note: scanning from here
 affine.for %arg3 = 0 to %0 step 8 {
 ^
<stdin>:7:2: note: with "M" equal to "0"
 affine.for %arg3 = 0 to %0 step 8 {
 ^
/build/work/a961fabe6974af72c61c07bcbfdfb8f5a2ba/google3/runfiles/google3/third_party/llvm/llvm/projects/google_mlir/examples/Linalg/Linalg4/Example.cpp:101:12: error: CHECK: expected string not found in input
 // CHECK: affine.for %{{.*}} = 0 to (d0) -> (d0)(%[[M]]) step 8 {
           ^
<stdin>:32:2: note: scanning from here
 affine.for %arg3 = 0 to %0 step 8 {
 ^
<stdin>:32:2: note: with "M" equal to "0"
 affine.for %arg3 = 0 to %0 step 8 {
 ^
/build/work/a961fabe6974af72c61c07bcbfdfb8f5a2ba/google3/runfiles/google3/third_party/llvm/llvm/projects/google_mlir/examples/Linalg/Linalg4/Example.cpp:135:12: error: CHECK: expected string not found in input
 // CHECK: affine.for %{{.*}} = 0 to (d0) -> (d0)(%[[M]]) step 8 {
           ^
<stdin>:54:2: note: scanning from here
 affine.for %arg3 = 0 to %0 step 8 {
 ^
<stdin>:54:2: note: with "M" equal to "0"
 affine.for %arg3 = 0 to %0 step 8 {
 ^

[bondhugula]

We still see the other Linalg example failing apparently:

@joker-eph Fixed now.