LLVM miscompiles passing/returning `half` on several backends by using lossy conversions

[beetrees]

Consider the following IR:

define half @to_half(i16 %bits) {
    %f = bitcast i16 %bits to half
    ret half %f
}

define i16 @from_half(half %f) {
    %bits = bitcast half %f to i16
    ret i16 %bits
}

As the only operation involved is a bitcast (in particular, there are no floating point type conversions in the LLVM IR), the values returned from to_half and from_half should be bit-for-bit identical to the value passed to them as their only argument (just a different type). However, several targets pass/return half as a float. On these targets, LLVM will use the default float conversion builtins (such as __gnu_h2f_ieee and __gnu_f2h_ieee) to convert between half and float. The issue is that these builtins silence signalling NaNs which changes the NaN payload, meaning that the roundtrip of half -> float -> half is not lossless and causes the generated ASM to violate the semantics of LLVM IR. This miscompilation is similar to #66803.

By inspecting the assembly LLVM emits, I've discovered that at least the following backends appear to be affected (in brackets are a specific target triple that I've checked):

• C-SKY (csky-unknown-linux-gnuabiv2 with -mcpu=ck860fv -mattr=+hard-float)
• Hexagon (hexagon-unknown-linux-musl): Fixed by [HEXAGON] Add support to lower "FREEZE a half(f16)" instruction on Hexagon and fix the isel-buildvector-v2f16.ll assertion #130977
• LoongArch (loongarch64-unknown-linux-gnu with -mattr=+f): Fixed by [loongarch][DAG][FREEZE] Fix crash when FREEZE a half(f16) type on loongarch #107791
• MIPS (mips64el-unknown-linux-gnuabi64): Fixed by [MIPS] Use softPromoteHalf legalization for fp16 rather than PromoteFloat #110199
• PowerPC (powerpc64le-unknown-linux-gnu)
• SPARC (sparc64-unknown-linux-gnu)
• WASM (wasm32-unknown-wasi): Already reported in wasm32 queitens half signalling NaNs in some situations when passing/returning them #96438

As none of these target's ABI specifications (that I've been able to find) specify how half should be passed (nor does Clang support _Float16 on any of these targets), and given that these targets are a subset of those affected by #97975, I'm filing this as a single issue as the ABI has probably been selected as an automatic default by LLVM rather than a deliberate choice by the backends. Ultimately there are two possible solutions: either fix LLVM to codegen lossless conversions between half and float when needed for the ABI (one way to do this would be with a new pair of builtins that don't silence signalling NaNs), or change the ABIs to pass/return half without converting it to float (probably using the same ABI as i16, but some targets might have better options).

Related to #97975.

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

Author: None (beetrees)

Consider the following IR: ```llvm define half @to_half(i16 %bits) { %f = bitcast i16 %bits to half ret half %f }

define i16 @from_half(half %f) {
%bits = bitcast half %f to i16
ret i16 %bits
}

As the only operation involved is a bitcast (in particular, there are no floating point type conversions in the LLVM IR), the values returned from `to_half` and `from_half` should be bit-for-bit identical to the value passed to them as their only argument (just a different type). However, several targets pass/return `half` as a `float`. On these targets, LLVM will use the default float conversion builtins (such as `__gnu_h2f_ieee` and `__gnu_f2h_ieee`) to convert between `half` and `float`. The issue is that these builtins silence signalling NaNs which changes the NaN payload, meaning that the roundtrip of `half` -&gt; `float` -&gt; `half` is not lossless and causes the generated ASM to violate the semantics of LLVM IR. This miscompilation is similar to #<!-- -->66803.

By inspecting the assembly LLVM emits, I've discovered that at least the following backends appear to be affected (in brackets are a specific target triple that I've checked):

- Hexagon (`hexagon-unknown-linux-musl`)
- MIPS (`mips64el-unknown-linux-gnuabi64`)
- PowerPC (`powerpc64le-unknown-linux-gnu`)
- SPARC (`sparc64-unknown-linux-gnu`)
- WASM (`wasm32-unknown-wasi`): Already reported in #<!-- -->96438

As none of these target's ABI specifications (that I've been able to find) specify how `half` should be passed (nor does Clang support `_Float16` on any of these targets), and given that these targets are a subset of those affected by #<!-- -->97975, I'm filing this as a single issue as the ABI has probably been selected as an automatic default by LLVM rather than a deliberate choice by the backends. Ultimately there are two possible solutions: either fix LLVM to codegen lossless conversions between `half` and `float` when needed for the ABI (one way to do this would be with a new pair of builtins that don't silence signalling NaNs), or change the ABIs to pass/return `half` without converting it to `float` (probably using the same ABI as `i16`, but some targets might have better options).
</details>

[llvmbot]

@llvm/issue-subscribers-backend-mips

Author: None (beetrees)

Consider the following IR: ```llvm define half @to_half(i16 %bits) { %f = bitcast i16 %bits to half ret half %f }

define i16 @from_half(half %f) {
%bits = bitcast half %f to i16
ret i16 %bits
}

As the only operation involved is a bitcast (in particular, there are no floating point type conversions in the LLVM IR), the values returned from `to_half` and `from_half` should be bit-for-bit identical to the value passed to them as their only argument (just a different type). However, several targets pass/return `half` as a `float`. On these targets, LLVM will use the default float conversion builtins (such as `__gnu_h2f_ieee` and `__gnu_f2h_ieee`) to convert between `half` and `float`. The issue is that these builtins silence signalling NaNs which changes the NaN payload, meaning that the roundtrip of `half` -&gt; `float` -&gt; `half` is not lossless and causes the generated ASM to violate the semantics of LLVM IR. This miscompilation is similar to #<!-- -->66803.

By inspecting the assembly LLVM emits, I've discovered that at least the following backends appear to be affected (in brackets are a specific target triple that I've checked):

- Hexagon (`hexagon-unknown-linux-musl`)
- MIPS (`mips64el-unknown-linux-gnuabi64`)
- PowerPC (`powerpc64le-unknown-linux-gnu`)
- SPARC (`sparc64-unknown-linux-gnu`)
- WASM (`wasm32-unknown-wasi`): Already reported in #<!-- -->96438

As none of these target's ABI specifications (that I've been able to find) specify how `half` should be passed (nor does Clang support `_Float16` on any of these targets), and given that these targets are a subset of those affected by #<!-- -->97975, I'm filing this as a single issue as the ABI has probably been selected as an automatic default by LLVM rather than a deliberate choice by the backends. Ultimately there are two possible solutions: either fix LLVM to codegen lossless conversions between `half` and `float` when needed for the ABI (one way to do this would be with a new pair of builtins that don't silence signalling NaNs), or change the ABIs to pass/return `half` without converting it to `float` (probably using the same ABI as `i16`, but some targets might have better options).
</details>

[llvmbot]

@llvm/issue-subscribers-backend-hexagon

Author: None (beetrees)

Consider the following IR: ```llvm define half @to_half(i16 %bits) { %f = bitcast i16 %bits to half ret half %f }

define i16 @from_half(half %f) {
%bits = bitcast half %f to i16
ret i16 %bits
}

As the only operation involved is a bitcast (in particular, there are no floating point type conversions in the LLVM IR), the values returned from `to_half` and `from_half` should be bit-for-bit identical to the value passed to them as their only argument (just a different type). However, several targets pass/return `half` as a `float`. On these targets, LLVM will use the default float conversion builtins (such as `__gnu_h2f_ieee` and `__gnu_f2h_ieee`) to convert between `half` and `float`. The issue is that these builtins silence signalling NaNs which changes the NaN payload, meaning that the roundtrip of `half` -&gt; `float` -&gt; `half` is not lossless and causes the generated ASM to violate the semantics of LLVM IR. This miscompilation is similar to #<!-- -->66803.

By inspecting the assembly LLVM emits, I've discovered that at least the following backends appear to be affected (in brackets are a specific target triple that I've checked):

- Hexagon (`hexagon-unknown-linux-musl`)
- MIPS (`mips64el-unknown-linux-gnuabi64`)
- PowerPC (`powerpc64le-unknown-linux-gnu`)
- SPARC (`sparc64-unknown-linux-gnu`)
- WASM (`wasm32-unknown-wasi`): Already reported in #<!-- -->96438

As none of these target's ABI specifications (that I've been able to find) specify how `half` should be passed (nor does Clang support `_Float16` on any of these targets), and given that these targets are a subset of those affected by #<!-- -->97975, I'm filing this as a single issue as the ABI has probably been selected as an automatic default by LLVM rather than a deliberate choice by the backends. Ultimately there are two possible solutions: either fix LLVM to codegen lossless conversions between `half` and `float` when needed for the ABI (one way to do this would be with a new pair of builtins that don't silence signalling NaNs), or change the ABIs to pass/return `half` without converting it to `float` (probably using the same ABI as `i16`, but some targets might have better options).
</details>

[llvmbot]

@llvm/issue-subscribers-backend-webassembly

Author: None (beetrees)

Consider the following IR: ```llvm define half @to_half(i16 %bits) { %f = bitcast i16 %bits to half ret half %f }

define i16 @from_half(half %f) {
%bits = bitcast half %f to i16
ret i16 %bits
}

As the only operation involved is a bitcast (in particular, there are no floating point type conversions in the LLVM IR), the values returned from `to_half` and `from_half` should be bit-for-bit identical to the value passed to them as their only argument (just a different type). However, several targets pass/return `half` as a `float`. On these targets, LLVM will use the default float conversion builtins (such as `__gnu_h2f_ieee` and `__gnu_f2h_ieee`) to convert between `half` and `float`. The issue is that these builtins silence signalling NaNs which changes the NaN payload, meaning that the roundtrip of `half` -&gt; `float` -&gt; `half` is not lossless and causes the generated ASM to violate the semantics of LLVM IR. This miscompilation is similar to #<!-- -->66803.

By inspecting the assembly LLVM emits, I've discovered that at least the following backends appear to be affected (in brackets are a specific target triple that I've checked):

- Hexagon (`hexagon-unknown-linux-musl`)
- MIPS (`mips64el-unknown-linux-gnuabi64`)
- PowerPC (`powerpc64le-unknown-linux-gnu`)
- SPARC (`sparc64-unknown-linux-gnu`)
- WASM (`wasm32-unknown-wasi`): Already reported in #<!-- -->96438

As none of these target's ABI specifications (that I've been able to find) specify how `half` should be passed (nor does Clang support `_Float16` on any of these targets), and given that these targets are a subset of those affected by #<!-- -->97975, I'm filing this as a single issue as the ABI has probably been selected as an automatic default by LLVM rather than a deliberate choice by the backends. Ultimately there are two possible solutions: either fix LLVM to codegen lossless conversions between `half` and `float` when needed for the ABI (one way to do this would be with a new pair of builtins that don't silence signalling NaNs), or change the ABIs to pass/return `half` without converting it to `float` (probably using the same ABI as `i16`, but some targets might have better options).
</details>

[vchuravy]

FYI: On PPC atleast you need to add zext I think. (On PPC i16 is being passed as i32 and that can cause issue)

define half @to_half(zext i16 %bits) {
    %f = bitcast i16 %bits to half
    ret half %f
}