[LAYOUTS] Generalise HoistLayoutConversion to work with arbitrary layouts and chains of ops

[lezcano]

We generalise HoistLayoutConversion to lift a given convert_layout dot_operand
above any chain of operations that do not require data movement. We
could totally generalise this in the future to lift it over other ops. We do
this as a first step to keep the code somewhat similar to the previous
one.

Regarding the previous limitations of canHoistDotOpEncV2 I did a bit of archeology:

• The "don't hoist past select" was added in this issue Floating Point Exception in F16xF16 Matmul with Pipelining on A100 #2857. I run the repro and with the recent layout fixes, it now passes.
• The TruncOps being skipped comes from [BACKEND] Refactor RemoveLayoutConversion pass #2181. I think this is related with the hack that was removed in [BACKEND] Get rid of unpack/pack I32  #5044, so now it should work
• Same same for the UIToFpOp, this is now supported after [BACKEND] Get rid of unpack/pack I32  #5044
• Mixed dtype hack is not necessary either as now everything works as expected with the convert_layout rework.

We also add proper support for isPure for elementwise_inline_asm ops

On the location of the code, we just leave it in RemoveLayoutConversion.cpp to
take advantage of the rather generic implementation of rewriteSlice. We could totally
move this pass outside of remove-layout-conversion, as it's probably enough to run
it once. This code will go through further changes in the near future, so we'll assess this
then.

[lezcano]

This is all codemovement + adding this test that was proposed but not merged in #5349 (comment)
as we now hoist everything as expected

[ThomasRaoux]

LGTM. One comment is I'm not sure if we need the speculatively part but that's kind of a detail

[Mogball]

Should ConstantOp just be put inside noDataMovement?

[lezcano]

Probably yeah, but leaving it as-is for now because it currently works and we are probably going to refactor this pass in the near future so whatever.

[lezcano]

I had to add a fix for MemDescTransOp to behave like TransOp when it comes to comparing LLs and legacy layouts as equal if they are structurally equal. Thank you @Mogball who helped me with the MLIR rough parts in #5747, which makes this transition a breeze (inheritance is so nice when it works)

[lezcano]

WTF, the A100 tests failed:

8 workers [16973 items]
<unknown>:0: error: 'tt.trans' op inferred type(s) 'tensor<16x128xbf16, #ttg.linear<{register = [[1, 0], [0, 8], [8, 0], [0, [32](https://github.com/triton-lang/triton/actions/runs/13071386381/job/36473639009?pr=5673#step:11:33)], [0, 64]], lane = [[2, 0], [4, 0], [0, 1], [0, 2], [0, 4]], warp = [[0, 0], [0, 16]], block = []}>>' are incompatible with return type(s) of operation 'tensor<128x16xbf16, #ttg.dot_op<{opIdx = 0, parent = #ttg.nvidia_mma<{versionMajor = 2, versionMinor = 0, warpsPerCTA = [2, 2], instrShape = [16, 8]}>, kWidth = 2}>>'
<unknown>:0: error: 'tt.trans' op failed to infer returned types
<unknown>:0: error: 'tt.trans' op inferred type(s) 'tensor<16x64xbf16, #ttg.linear<{register = [[1, 0], [0, 8], [8, 0], [0, 32]], lane = [[2, 0], [4, 0], [0, 1], [0, 2], [0, 4]], warp = [[0, 0], [0, 16]], block = []}>>' are incompatible with return type(s) of operation 'tensor<64x16xbf16, #ttg.dot_op<{opIdx = 0, parent = #ttg.nvidia_mma<{versionMajor = 2, versionMinor = 0, warpsPerCTA = [2, 2], instrShape = [16, 8]}>, kWidth = 2}>>'
<unknown>:0: error: 'tt.trans' op failed to infer returned types
<unknown>:0: error: 'tt.trans' op inferred type(s) 'tensor<16x128xbf16, #ttg.linear<{register = [[1, 0], [0, 8], [8, 0], [0, 32], [0, 64]], lane = [[2, 0], [4, 0], [0, 1], [0, 2], [0, 4]], warp = [[0, 0], [0, 16]], block = []}>>' are incompatible with return type(s) of operation 'tensor<128x16xbf16, #ttg.dot_op<{opIdx = 0, parent = #ttg.nvidia_mma<{versionMajor = 2, versionMinor = 0, warpsPerCTA = [2, 2], instrShape = [16, 8]}>, kWidth = 2}>>'
<unknown>:0: error: 'tt.trans' op failed to infer returned types
<unknown>:0: error: 'tt.trans' op inferred type(s) 'tensor<16x128xbf16, #ttg.linear<{register = [[1, 0], [0, 8], [8, 0], [0, 32], [0, 64]], lane = [[2, 0], [4, 0], [0, 1], [0, 2], [0, 4]], warp = [[0, 0], [0, 16]], block = []}>>' are incompatible with return type(s) of operation 'tensor<128x16xbf16, #ttg.dot_op<{opIdx = 0, parent = #ttg.nvidia_mma<{versionMajor = 2, versionMinor = 0, warpsPerCTA = [2, 2], instrShape = [16, 8]}>, kWidth = 2}>>'
<unknown>:0: error: 'tt.trans' op failed to infer returned types
<unknown>:0: error: 'tt.trans' op inferred type(s) 'tensor<16x64xbf16, #ttg.linear<{register = [[1, 0], [0, 8], [8, 0], [0, 32]], lane = [[2, 0], [4, 0], [0, 1], [0, 2], [0, 4]], warp = [[0, 0], [0, 16]], block = []}>>' are incompatible with return type(s) of operation 'tensor<64x16xbf16, #ttg.dot_op<{opIdx = 0, parent = #ttg.nvidia_mma<{versionMajor = 2, versionMinor = 0, warpsPerCTA = [2, 2], instrShape = [16, 8]}>, kWidth = 2}>>'
<unknown>:0: error: 'tt.trans' op failed to infer returned types
<unknown>:0: error: 'tt.trans' op inferred type(s) 'tensor<16x128xbf16, #ttg.linear<{register = [[1, 0], [0, 8], [8, 0], [0, 32], [0, 64]], lane = [[2, 0], [4, 0], [0, 1], [0, 2], [0, 4]], warp = [[0, 0], [0, 16]], block = []}>>' are incompatible with return type(s) of operation 'tensor<128x16xbf16, #ttg.dot_op<{opIdx = 0, parent = #ttg.nvidia_mma<{versionMajor = 2, versionMinor = 0, warpsPerCTA = [2, 2], instrShape = [16, 8]}>, kWidth = 2}>>'
<unknown>:0: error: 'tt.trans' op failed to infer returned types
<unknown>:0: error: operation scheduled before its operands

but CI marked it as green?

[lezcano]

Reopened at #5788