Stream: git-wasmtime

Topic: wasmtime / issue #3075 Cranelift: Wrong results in x86_64...

Wasmtime GitHub notifications bot (Jul 10 2021 at 11:43):

afonso360 labeled issue #3075:

Hey, In #3074 we introduce a number of shift tests, some of those found issues with the shift instructions that all backends have.

.clif Test Case

test run
target x86_64 machinst
target s390x

function %ishl_i16_i64(i16, i64) -> i16 {
block0(v0: i16, v1: i64):
    v2 = ishl.i16 v0, v1
    return v2
}
; run: %ishl_i16_i64(0x0000, 0) == 0x0000
; run: %ishl_i16_i64(0x0000, 1) == 0x0000
; run: %ishl_i16_i64(0x000f, 0) == 0x000f
; run: %ishl_i16_i64(0x000f, 4) == 0x00f0
; This fails for the values below
; run: %ishl_i16_i64(0x0004, 16) == 0x0004
; run: %ishl_i16_i64(0x0004, 17) == 0x0008
; run: %ishl_i16_i64(0x0004, 18) == 0x0010

Note, this is a sample CLIF, we found errors for all shifts with destination types of i16 and i8. A full test suite is introduced in #3074, in the files:

• cranelift/filetests/filetests/runtests/i128-shifts-small-types.clif
• cranelift/filetests/filetests/runtests/shifts-small-types.clif

Expected Results and Actual Results

For shift amounts larger than the size of the type, a wrapping behaviour should occur, such that i32 << 32 is equivalent to performing a i32 << 0. And this is what happens for i32 and i64 types, however for i16 and i8 this does not happen
and the result is always 0.

Versions and Environment

Cranelift version or commit: main
Architecture: x86_64 and s390x

Wasmtime GitHub notifications bot (Jul 10 2021 at 11:43):

afonso360 labeled issue #3075:

Hey, In #3074 we introduce a number of shift tests, some of those found issues with the shift instructions that all backends have.

.clif Test Case

test run
target x86_64 machinst
target s390x

function %ishl_i16_i64(i16, i64) -> i16 {
block0(v0: i16, v1: i64):
    v2 = ishl.i16 v0, v1
    return v2
}
; run: %ishl_i16_i64(0x0000, 0) == 0x0000
; run: %ishl_i16_i64(0x0000, 1) == 0x0000
; run: %ishl_i16_i64(0x000f, 0) == 0x000f
; run: %ishl_i16_i64(0x000f, 4) == 0x00f0
; This fails for the values below
; run: %ishl_i16_i64(0x0004, 16) == 0x0004
; run: %ishl_i16_i64(0x0004, 17) == 0x0008
; run: %ishl_i16_i64(0x0004, 18) == 0x0010

Note, this is a sample CLIF, we found errors for all shifts with destination types of i16 and i8. A full test suite is introduced in #3074, in the files:

• cranelift/filetests/filetests/runtests/i128-shifts-small-types.clif
• cranelift/filetests/filetests/runtests/shifts-small-types.clif

Expected Results and Actual Results

For shift amounts larger than the size of the type, a wrapping behaviour should occur, such that i32 << 32 is equivalent to performing a i32 << 0. And this is what happens for i32 and i64 types, however for i16 and i8 this does not happen
and the result is always 0.

Versions and Environment

Cranelift version or commit: main
Architecture: x86_64 and s390x

Wasmtime GitHub notifications bot (Jul 10 2021 at 11:43):

afonso360 opened issue #3075:

Hey, In #3074 we introduce a number of shift tests, some of those found issues with the shift instructions that all backends have.

.clif Test Case

test run
target x86_64 machinst
target s390x

function %ishl_i16_i64(i16, i64) -> i16 {
block0(v0: i16, v1: i64):
    v2 = ishl.i16 v0, v1
    return v2
}
; run: %ishl_i16_i64(0x0000, 0) == 0x0000
; run: %ishl_i16_i64(0x0000, 1) == 0x0000
; run: %ishl_i16_i64(0x000f, 0) == 0x000f
; run: %ishl_i16_i64(0x000f, 4) == 0x00f0
; This fails for the values below
; run: %ishl_i16_i64(0x0004, 16) == 0x0004
; run: %ishl_i16_i64(0x0004, 17) == 0x0008
; run: %ishl_i16_i64(0x0004, 18) == 0x0010

Note, this is a sample CLIF, we found errors for all shifts with destination types of i16 and i8. A full test suite is introduced in #3074, in the files:

• cranelift/filetests/filetests/runtests/i128-shifts-small-types.clif
• cranelift/filetests/filetests/runtests/shifts-small-types.clif

Expected Results and Actual Results

For shift amounts larger than the size of the type, a wrapping behaviour should occur, such that i32 << 32 is equivalent to performing a i32 << 0. And this is what happens for i32 and i64 types, however for i16 and i8 this does not happen
and the result is always 0.

Versions and Environment

Cranelift version or commit: main
Architecture: x86_64 and s390x

Wasmtime GitHub notifications bot (May 04 2022 at 22:59):

cfallin labeled issue #3075:

Hey, In #3074 we introduce a number of shift tests, some of those found issues with the shift instructions that all backends have.

.clif Test Case

test run
target x86_64 machinst
target s390x

function %ishl_i16_i64(i16, i64) -> i16 {
block0(v0: i16, v1: i64):
    v2 = ishl.i16 v0, v1
    return v2
}
; run: %ishl_i16_i64(0x0000, 0) == 0x0000
; run: %ishl_i16_i64(0x0000, 1) == 0x0000
; run: %ishl_i16_i64(0x000f, 0) == 0x000f
; run: %ishl_i16_i64(0x000f, 4) == 0x00f0
; This fails for the values below
; run: %ishl_i16_i64(0x0004, 16) == 0x0004
; run: %ishl_i16_i64(0x0004, 17) == 0x0008
; run: %ishl_i16_i64(0x0004, 18) == 0x0010

Note, this is a sample CLIF, we found errors for all shifts with destination types of i16 and i8. A full test suite is introduced in #3074, in the files:

• cranelift/filetests/filetests/runtests/i128-shifts-small-types.clif
• cranelift/filetests/filetests/runtests/shifts-small-types.clif

Expected Results and Actual Results

For shift amounts larger than the size of the type, a wrapping behaviour should occur, such that i32 << 32 is equivalent to performing a i32 << 0. And this is what happens for i32 and i64 types, however for i16 and i8 this does not happen
and the result is always 0.

Versions and Environment

Cranelift version or commit: main
Architecture: x86_64 and s390x

Wasmtime GitHub notifications bot (May 04 2022 at 22:59):

cfallin labeled issue #3075:

Hey, In #3074 we introduce a number of shift tests, some of those found issues with the shift instructions that all backends have.

.clif Test Case

test run
target x86_64 machinst
target s390x

function %ishl_i16_i64(i16, i64) -> i16 {
block0(v0: i16, v1: i64):
    v2 = ishl.i16 v0, v1
    return v2
}
; run: %ishl_i16_i64(0x0000, 0) == 0x0000
; run: %ishl_i16_i64(0x0000, 1) == 0x0000
; run: %ishl_i16_i64(0x000f, 0) == 0x000f
; run: %ishl_i16_i64(0x000f, 4) == 0x00f0
; This fails for the values below
; run: %ishl_i16_i64(0x0004, 16) == 0x0004
; run: %ishl_i16_i64(0x0004, 17) == 0x0008
; run: %ishl_i16_i64(0x0004, 18) == 0x0010

Note, this is a sample CLIF, we found errors for all shifts with destination types of i16 and i8. A full test suite is introduced in #3074, in the files:

• cranelift/filetests/filetests/runtests/i128-shifts-small-types.clif
• cranelift/filetests/filetests/runtests/shifts-small-types.clif

Expected Results and Actual Results

For shift amounts larger than the size of the type, a wrapping behaviour should occur, such that i32 << 32 is equivalent to performing a i32 << 0. And this is what happens for i32 and i64 types, however for i16 and i8 this does not happen
and the result is always 0.

Versions and Environment

Cranelift version or commit: main
Architecture: x86_64 and s390x

Wasmtime GitHub notifications bot (Sep 02 2022 at 15:56):

afonso360 commented on issue #3075:

This was fixed in #4752 for x86 and #3702 for s390x.

Wasmtime GitHub notifications bot (Sep 02 2022 at 15:56):

afonso360 closed issue #3075:

Hey, In #3074 we introduce a number of shift tests, some of those found issues with the shift instructions that all backends have.

.clif Test Case

test run
target x86_64 machinst
target s390x

function %ishl_i16_i64(i16, i64) -> i16 {
block0(v0: i16, v1: i64):
    v2 = ishl.i16 v0, v1
    return v2
}
; run: %ishl_i16_i64(0x0000, 0) == 0x0000
; run: %ishl_i16_i64(0x0000, 1) == 0x0000
; run: %ishl_i16_i64(0x000f, 0) == 0x000f
; run: %ishl_i16_i64(0x000f, 4) == 0x00f0
; This fails for the values below
; run: %ishl_i16_i64(0x0004, 16) == 0x0004
; run: %ishl_i16_i64(0x0004, 17) == 0x0008
; run: %ishl_i16_i64(0x0004, 18) == 0x0010

Note, this is a sample CLIF, we found errors for all shifts with destination types of i16 and i8. A full test suite is introduced in #3074, in the files:

• cranelift/filetests/filetests/runtests/i128-shifts-small-types.clif
• cranelift/filetests/filetests/runtests/shifts-small-types.clif

Expected Results and Actual Results

For shift amounts larger than the size of the type, a wrapping behaviour should occur, such that i32 << 32 is equivalent to performing a i32 << 0. And this is what happens for i32 and i64 types, however for i16 and i8 this does not happen
and the result is always 0.

Versions and Environment

Cranelift version or commit: main
Architecture: x86_64 and s390x

Last updated: Nov 22 2024 at 16:03 UTC