afonso360 opened issue #3248:
Hey, @dheaton-arm reported an interesting implementation detail from the aarch64 backend on the
cls
instruction. However, further testing shows that this instruction is not correctly implemented in any backend fori16
andi8
types.
.clif
Test Casetest run target aarch64 target x86_64 machinst target s390x function %cls_i16(i16) -> i16 { block0(v0: i16): v1 = cls.i16 v0 return v1 } ; run: %cls_i16(0x0000) == 15 ; run: %cls_i16(0xFFFF) == 15 ; run: %cls_i16(0x8000) == 1 ; run: %cls_i16(0xC000) == 2 ; run: %cls_i16(0x2000) == 1 ; run: %cls_i16(0x1000) == 2 function %cls_i8(i8) -> i8 { block0(v0: i8): v1 = cls.i8 v0 return v1 } ; run: %cls_i8(0x00) == 7 ; run: %cls_i8(0xFF) == 7 ; run: %cls_i8(0x80) == 1 ; run: %cls_i8(0xC0) == 2 ; run: %cls_i8(0x20) == 1 ; run: %cls_i8(0x10) == 2
Steps to Reproduce
clif-util test ./the-above.clif
Expected Results
All three backends should pass the run tests.
Actual Results
x86_64:
cls
is not implemented, for either type.
aarch64: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns31
)
s390x: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns4096
)Versions and Environment
Cranelift version or commit: main
Operating system: Linux / qemu emulator for aarch64 and s390x
Architecture: x86_64 / aarch64 / s390x
afonso360 labeled issue #3248:
Hey, @dheaton-arm reported an interesting implementation detail from the aarch64 backend on the
cls
instruction. However, further testing shows that this instruction is not correctly implemented in any backend fori16
andi8
types.
.clif
Test Casetest run target aarch64 target x86_64 machinst target s390x function %cls_i16(i16) -> i16 { block0(v0: i16): v1 = cls.i16 v0 return v1 } ; run: %cls_i16(0x0000) == 15 ; run: %cls_i16(0xFFFF) == 15 ; run: %cls_i16(0x8000) == 1 ; run: %cls_i16(0xC000) == 2 ; run: %cls_i16(0x2000) == 1 ; run: %cls_i16(0x1000) == 2 function %cls_i8(i8) -> i8 { block0(v0: i8): v1 = cls.i8 v0 return v1 } ; run: %cls_i8(0x00) == 7 ; run: %cls_i8(0xFF) == 7 ; run: %cls_i8(0x80) == 1 ; run: %cls_i8(0xC0) == 2 ; run: %cls_i8(0x20) == 1 ; run: %cls_i8(0x10) == 2
Steps to Reproduce
clif-util test ./the-above.clif
Expected Results
All three backends should pass the run tests.
Actual Results
x86_64:
cls
is not implemented, for either type.
aarch64: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns31
)
s390x: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns4096
)Versions and Environment
Cranelift version or commit: main
Operating system: Linux / qemu emulator for aarch64 and s390x
Architecture: x86_64 / aarch64 / s390x
afonso360 labeled issue #3248:
Hey, @dheaton-arm reported an interesting implementation detail from the aarch64 backend on the
cls
instruction. However, further testing shows that this instruction is not correctly implemented in any backend fori16
andi8
types.
.clif
Test Casetest run target aarch64 target x86_64 machinst target s390x function %cls_i16(i16) -> i16 { block0(v0: i16): v1 = cls.i16 v0 return v1 } ; run: %cls_i16(0x0000) == 15 ; run: %cls_i16(0xFFFF) == 15 ; run: %cls_i16(0x8000) == 1 ; run: %cls_i16(0xC000) == 2 ; run: %cls_i16(0x2000) == 1 ; run: %cls_i16(0x1000) == 2 function %cls_i8(i8) -> i8 { block0(v0: i8): v1 = cls.i8 v0 return v1 } ; run: %cls_i8(0x00) == 7 ; run: %cls_i8(0xFF) == 7 ; run: %cls_i8(0x80) == 1 ; run: %cls_i8(0xC0) == 2 ; run: %cls_i8(0x20) == 1 ; run: %cls_i8(0x10) == 2
Steps to Reproduce
clif-util test ./the-above.clif
Expected Results
All three backends should pass the run tests.
Actual Results
x86_64:
cls
is not implemented, for either type.
aarch64: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns31
)
s390x: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns4096
)Versions and Environment
Cranelift version or commit: main
Operating system: Linux / qemu emulator for aarch64 and s390x
Architecture: x86_64 / aarch64 / s390x
afonso360 commented on issue #3248:
In fact, we have issues with this instruction in
i32
andi64
types.Testcase:
function %cls_i64(i64) -> i64 { block0(v0: i64): v1 = cls.i64 v0 return v1 } ; run: %cls_i64(0) == 63 function %cls_i32(i32) -> i32 { block0(v0: i32): v1 = cls.i32 v0 return v1 } ; run: %cls_i32(0) == 31
aarch64: Passes these test cases
x86_64: Does not implement these
x390x: Fails with a wrong value (Failed test: run: %cls_i64(0) == 63, actual: 4611686018427387904
)
afonso360 edited issue #3248:
Hey, @dheaton-arm reported an interesting implementation detail from the aarch64 backend on the
cls
instruction. However, further testing shows that this instruction is not correctly implemented in any backend fori16
andi8
types.
.clif
Test Casetest run target aarch64 target x86_64 machinst target s390x function %cls_i16(i16) -> i16 { block0(v0: i16): v1 = cls.i16 v0 return v1 } ; run: %cls_i16(0x0000) == 15 ; run: %cls_i16(0xFFFF) == 15 ; run: %cls_i16(0x8000) == 0 ; run: %cls_i16(0xC000) == 1 ; run: %cls_i16(0x4000) == 0 ; run: %cls_i16(0x2000) == 1 function %cls_i8(i8) -> i8 { block0(v0: i8): v1 = cls.i8 v0 return v1 } ; run: %cls_i8(0x00) == 7 ; run: %cls_i8(0xFF) == 7 ; run: %cls_i8(0x80) == 0 ; run: %cls_i8(0xC0) == 1 ; run: %cls_i8(0x40) == 0 ; run: %cls_i8(0x20) == 1
Steps to Reproduce
clif-util test ./the-above.clif
Expected Results
All three backends should pass the run tests.
Actual Results
x86_64:
cls
is not implemented, for either type.
aarch64: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns31
)
s390x: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns4096
)Versions and Environment
Cranelift version or commit: main
Operating system: Linux / qemu emulator for aarch64 and s390x
Architecture: x86_64 / aarch64 / s390x
sparker-arm commented on issue #3248:
At least for AArch64, we can't use a 1-1 mapping between CLIF and machine instructions when using a type smaller than a naturally supported register width (32 and 64-bits). I think we need to expand this instruction in the case that the target doesn't directly support the type, something like:
cls (x) -> sub (cls (swiden (x)), bitwidth(x))
akirilov-arm labeled issue #3248:
Hey, @dheaton-arm reported an interesting implementation detail from the aarch64 backend on the
cls
instruction. However, further testing shows that this instruction is not correctly implemented in any backend fori16
andi8
types.
.clif
Test Casetest run target aarch64 target x86_64 machinst target s390x function %cls_i16(i16) -> i16 { block0(v0: i16): v1 = cls.i16 v0 return v1 } ; run: %cls_i16(0x0000) == 15 ; run: %cls_i16(0xFFFF) == 15 ; run: %cls_i16(0x8000) == 0 ; run: %cls_i16(0xC000) == 1 ; run: %cls_i16(0x4000) == 0 ; run: %cls_i16(0x2000) == 1 function %cls_i8(i8) -> i8 { block0(v0: i8): v1 = cls.i8 v0 return v1 } ; run: %cls_i8(0x00) == 7 ; run: %cls_i8(0xFF) == 7 ; run: %cls_i8(0x80) == 0 ; run: %cls_i8(0xC0) == 1 ; run: %cls_i8(0x40) == 0 ; run: %cls_i8(0x20) == 1
Steps to Reproduce
clif-util test ./the-above.clif
Expected Results
All three backends should pass the run tests.
Actual Results
x86_64:
cls
is not implemented, for either type.
aarch64: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns31
)
s390x: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns4096
)Versions and Environment
Cranelift version or commit: main
Operating system: Linux / qemu emulator for aarch64 and s390x
Architecture: x86_64 / aarch64 / s390x
akirilov-arm labeled issue #3248:
Hey, @dheaton-arm reported an interesting implementation detail from the aarch64 backend on the
cls
instruction. However, further testing shows that this instruction is not correctly implemented in any backend fori16
andi8
types.
.clif
Test Casetest run target aarch64 target x86_64 machinst target s390x function %cls_i16(i16) -> i16 { block0(v0: i16): v1 = cls.i16 v0 return v1 } ; run: %cls_i16(0x0000) == 15 ; run: %cls_i16(0xFFFF) == 15 ; run: %cls_i16(0x8000) == 0 ; run: %cls_i16(0xC000) == 1 ; run: %cls_i16(0x4000) == 0 ; run: %cls_i16(0x2000) == 1 function %cls_i8(i8) -> i8 { block0(v0: i8): v1 = cls.i8 v0 return v1 } ; run: %cls_i8(0x00) == 7 ; run: %cls_i8(0xFF) == 7 ; run: %cls_i8(0x80) == 0 ; run: %cls_i8(0xC0) == 1 ; run: %cls_i8(0x40) == 0 ; run: %cls_i8(0x20) == 1
Steps to Reproduce
clif-util test ./the-above.clif
Expected Results
All three backends should pass the run tests.
Actual Results
x86_64:
cls
is not implemented, for either type.
aarch64: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns31
)
s390x: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns4096
)Versions and Environment
Cranelift version or commit: main
Operating system: Linux / qemu emulator for aarch64 and s390x
Architecture: x86_64 / aarch64 / s390x
fitzgen closed issue #3248:
Hey, @dheaton-arm reported an interesting implementation detail from the aarch64 backend on the
cls
instruction. However, further testing shows that this instruction is not correctly implemented in any backend fori16
andi8
types.
.clif
Test Casetest run target aarch64 target x86_64 machinst target s390x function %cls_i16(i16) -> i16 { block0(v0: i16): v1 = cls.i16 v0 return v1 } ; run: %cls_i16(0x0000) == 15 ; run: %cls_i16(0xFFFF) == 15 ; run: %cls_i16(0x8000) == 0 ; run: %cls_i16(0xC000) == 1 ; run: %cls_i16(0x4000) == 0 ; run: %cls_i16(0x2000) == 1 function %cls_i8(i8) -> i8 { block0(v0: i8): v1 = cls.i8 v0 return v1 } ; run: %cls_i8(0x00) == 7 ; run: %cls_i8(0xFF) == 7 ; run: %cls_i8(0x80) == 0 ; run: %cls_i8(0xC0) == 1 ; run: %cls_i8(0x40) == 0 ; run: %cls_i8(0x20) == 1
Steps to Reproduce
clif-util test ./the-above.clif
Expected Results
All three backends should pass the run tests.
Actual Results
x86_64:
cls
is not implemented, for either type.
aarch64: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns31
)
s390x: Returns the wrong results (i.e.run: %cls_i16(0x0000) == 15
returns4096
)Versions and Environment
Cranelift version or commit: main
Operating system: Linux / qemu emulator for aarch64 and s390x
Architecture: x86_64 / aarch64 / s390x
Last updated: Nov 22 2024 at 16:03 UTC