Allow truncations when matching arguments for intrinsic overloads. This eliminates the need for explicit scalar extractions from vectors for arguments that are scalar by nature. This encompasses any vectors passed for scalars, allowing the truncation, but emitting a warning the same as is done for other assignments of vectors to scalars.

This maintains splats as the preferred transformations and promotes perfect matches to be preferred over that. This has the effect of removing the need to carefully order intrinsics to ensure that the right variant gets matched first before another one incorrectly takes its place with a faulty cast.

Allowing truncations causes a problems with a small subset of intrinsics that have explicit overloads for various matrix,vector, scalar combinations. Namely the mul overloads. These could be simplified to accept a new range of template types except the dimensions need to be matched in unconventional ways.

For these, the notion of uncastable or "ONLY" variants of the template/layout types are introduced. These are indicated with a trailing "!" after the parameter typename in gen_intrin_main, which directs them to an array that contains a NOCAST enum that, when encountered, will skip the attempts to splat or truncate.

Fixes microsoft#7079

Load operations previously split out the addr into separate components in individual locations where they were required. This was error prone as it had a lot of places that were missing. This changes the contents of resload and atomic helpers to contain the broken down coordinates instead of the individual Value. This catches a number of new locations, but it requires the intrinsic overloads to be defined as vec1s to handle that case and it still involves a lot of fiddly places for the code that don't produce truncation warnings when longer vectors are used. Not recommended, but functional.

Fixes microsoft#7079

Allow truncations when matching arguments for intrinsic overloads. This eliminates the need for explicit scalar extractions from vectors for arguments that are scalar by nature. This encompasses any vectors passed for scalars, allowing the truncation, but emitting a warning the same as is done for other assignments of vectors to scalars.

This maintains splats as the preferred transformations and promotes perfect matches to be preferred over that. This has the effect of removing the need to carefully order intrinsics to ensure that the right variant gets matched first before another one incorrectly takes its place with a faulty cast.

Allowing truncations causes a problems with a small subset of intrinsics that have explicit overloads for various matrix,vector, scalar combinations. Namely the mul overloads. These could be simplified to accept a new range of template types except the dimensions need to be matched in unconventional ways.

For these, the notion of uncastable or "ONLY" variants of the template/layout types are introduced. These are indicated with a trailing "!" after the parameter typename in gen_intrin_main, which directs them to an array that contains a NOCAST enum that, when encountered, will skip the attempts to splat or truncate.

Fixes microsoft#7079

It was producing a validation error previously because they are represented as u32 and i1 is not valid for output resources. This preserves the cmp conversions for values read from the resource, but satisfies the validation requirement. Performs some incidental consolidation of types that are represented as u32.

Related to microsoft#7079

It was producing a validation error previously because they are represented as u32 and i1 is not valid for output resources. This preserves the cmp conversions for values read from the resource, but satisfies the validation requirement. Performs some incidental consolidation of types that are represented as u32.

Related to microsoft#7079

Allow truncations when matching arguments for intrinsic overloads. This eliminates the need for explicit scalar extractions from vectors for arguments that are scalar by nature. This encompasses any vectors passed for scalars, allowing the truncation, but emitting a warning the same as is done for other assignments of vectors to scalars.

This maintains splats as the preferred transformations and promotes perfect matches to be preferred over that. This has the effect of removing the need to carefully order intrinsics to ensure that the right variant gets matched first before another one incorrectly takes its place with a faulty cast.

Allowing truncations causes a problems with a small subset of intrinsics that have explicit overloads for various matrix,vector, scalar combinations. Namely the mul overloads. These could be simplified to accept a new range of template types except the dimensions need to be matched in unconventional ways.

For these, the notion of uncastable or "ONLY" variants of the template/layout types are introduced. These are indicated with a trailing "!" after the parameter typename in gen_intrin_main, which directs them to an array that contains a NOCAST enum that, when encountered, will skip the attempts to splat or truncate.

Fixes microsoft#7079

Allow truncations when matching arguments for intrinsic overloads. This eliminates the need for explicit scalar extractions from vectors for arguments that are scalar by nature. This encompasses any vectors passed for scalars, allowing the truncation, but emitting a warning the same as is done for other assignments of vectors to scalars.

This maintains splats as the preferred transformations and promotes perfect matches to be preferred over that. This has the effect of removing the need to carefully order intrinsics to ensure that the right variant gets matched first before another one incorrectly takes its place with a faulty cast.

Allowing truncations causes a problems with a small subset of intrinsics that have explicit overloads for various matrix,vector, scalar combinations. Namely the mul overloads. These could be simplified to accept a new range of template types except the dimensions need to be matched in unconventional ways.

For these, the notion of uncastable or "ONLY" variants of the template/layout types are introduced. These are indicated with a trailing "!" after the parameter typename in gen_intrin_main, which directs them to an array that contains a NOCAST enum that, when encountered, will skip the attempts to splat or truncate.

Fixes microsoft#7079

It was producing a validation error previously because they are represented as u32 and i1 is not valid for output resources. This preserves the cmp conversions for values read from the resource, but satisfies the validation requirement. Performs some incidental consolidation of types that are represented as u32.

Related to microsoft#7079

Allow truncations when matching arguments for intrinsic overloads. This eliminates the need for explicit scalar extractions from vectors for arguments that are scalar by nature. This encompasses any vectors passed for scalars, allowing the truncation, but emitting a warning the same as is done for other assignments of vectors to scalars.

This maintains splats as the preferred transformations and promotes perfect matches to be preferred over that. This has the effect of removing the need to carefully order intrinsics to ensure that the right variant gets matched first before another one incorrectly takes its place with a faulty cast.

Allowing truncations causes a problems with a small subset of intrinsics that have explicit overloads for various matrix,vector, scalar combinations. Namely the mul overloads. These could be simplified to accept a new range of template types except the dimensions need to be matched in unconventional ways.

For these, the notion of uncastable or "ONLY" variants of the template/layout types are introduced. These are indicated with a trailing "!" after the parameter typename in gen_intrin_main, which directs them to an array that contains a NOCAST enum that, when encountered, will skip the attempts to splat or truncate.

Fixes microsoft#7079

It was producing a validation error previously because they are represented as u32 and i1 is not valid for output resources. This preserves the cmp conversions for values read from the resource, but satisfies the validation requirement. Performs some incidental consolidation of types that are represented as u32.

Related to microsoft#7079

Allow truncations when matching arguments for intrinsic overloads. This eliminates the need for explicit scalar extractions from vectors for arguments that are scalar by nature. This encompasses any vectors passed for scalars, allowing the truncation, but emitting a warning the same as is done for other assignments of vectors to scalars.

This maintains splats as the preferred transformations and promotes perfect matches to be preferred over that. This has the effect of removing the need to carefully order intrinsics to ensure that the right variant gets matched first before another one incorrectly takes its place with a faulty cast.

Allowing truncations causes a problems with a small subset of intrinsics that have explicit overloads for various matrix,vector, scalar combinations. Namely the mul overloads. These could be simplified to accept a new range of template types except the dimensions need to be matched in unconventional ways.

For these, the notion of uncastable or "ONLY" variants of the template/layout types are introduced. These are indicated with a trailing "!" after the parameter typename in gen_intrin_main, which directs them to an array that contains a NOCAST enum that, when encountered, will skip the attempts to splat or truncate.

Fixes microsoft#7079

It was producing a validation error previously because they are represented as u32 and i1 is not valid for output resources. This preserves the cmp conversions for values read from the resource, but satisfies the validation requirement. Performs some incidental consolidation of types that are represented as u32.

Related to microsoft#7079

Allow truncations when matching arguments for intrinsic overloads. This eliminates the need for explicit scalar extractions from vectors for arguments that are scalar by nature. This encompasses any vectors passed for scalars, allowing the truncation, but emitting a warning the same as is done for other assignments of vectors to scalars.

This maintains splats as the preferred transformations and promotes perfect matches to be preferred over that. This has the effect of removing the need to carefully order intrinsics to ensure that the right variant gets matched first before another one incorrectly takes its place with a faulty cast.

Allowing truncations causes a problems with a small subset of intrinsics that have explicit overloads for various matrix,vector, scalar combinations. Namely the mul overloads. These could be simplified to accept a new range of template types except the dimensions need to be matched in unconventional ways.

For these, the notion of uncastable or "ONLY" variants of the template/layout types are introduced. These are indicated with a trailing "!" after the parameter typename in gen_intrin_main, which directs them to an array that contains a NOCAST enum that, when encountered, will skip the attempts to splat or truncate.

Fixes microsoft#7079

It was producing a validation error previously because they are represented as u32 and i1 is not valid for output resources. This preserves the cmp conversions for values read from the resource, but satisfies the validation requirement. Performs some incidental consolidation of types that are represented as u32.

Related to microsoft#7079

remove dead array handing code

The array type checks were for a type extracted from a vector. Arrays cannot be elements of vectors. This is leftover from when that was not the case. The code shows as never executed in coverage

It was producing a validation error previously because they are
represented as u32 and i1 is not valid for output resources. This
preserves the cmp conversions for values read from the resource, but
satisfies the validation requirement. Performs some incidental
consolidation of types that are represented as u32.

Includes a couple incidental NFC fixes as well:

Moves lit filecheck tests into the correct directory. I got CodeGenHLSL
and CodeGenDXIL mixed up.

Remove dead array handing code. The array type checks were for a type
extracted from a vector. Arrays cannot be elements of vectors. This is
leftover from when the type came from more sources. The code shows as
never executed in coverage.

Related to #7079

It was producing a validation error previously because they are
represented as u32 and i1 is not valid for output resources. This
preserves the cmp conversions for values read from the resource, but
satisfies the validation requirement. Performs some incidental
consolidation of types that are represented as u32.

Includes a couple incidental NFC fixes as well:

Moves lit filecheck tests into the correct directory. I got CodeGenHLSL
and CodeGenDXIL mixed up.

Remove dead array handing code. The array type checks were for a type
extracted from a vector. Arrays cannot be elements of vectors. This is
leftover from when the type came from more sources. The code shows as
never executed in coverage.

Related to microsoft#7079