bjoernager/rust
bjoernager/rust
1
Fork 0
Code Issues Pull requests Activity
rust/tests/assembly/simd-intrinsic-mask-load.rs

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

84 lines
2.7 KiB
Rust
Raw Permalink Normal View History

tests: use minicore more minicore makes it much easier to add new language items to all of the existing `no_core` tests.
2025-02-24 09:26:54 +00:00
//@ add-core-stubs
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
//@ revisions: x86-avx2 x86-avx512
//@ [x86-avx2] compile-flags: --target=x86_64-unknown-linux-gnu -C llvm-args=-x86-asm-syntax=intel
//@ [x86-avx2] compile-flags: -C target-feature=+avx2
//@ [x86-avx2] needs-llvm-components: x86
//@ [x86-avx512] compile-flags: --target=x86_64-unknown-linux-gnu -C llvm-args=-x86-asm-syntax=intel
//@ [x86-avx512] compile-flags: -C target-feature=+avx512f,+avx512vl,+avx512bw,+avx512dq
//@ [x86-avx512] needs-llvm-components: x86
//@ assembly-output: emit-asm
tests/assembly: use -Copt-level=3 instead of -O
2025-02-08 18:56:57 -08:00
//@ compile-flags: --crate-type=lib -Copt-level=3 -C panic=abort
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
#![feature(no_core, lang_items, repr_simd, intrinsics)]
#![no_core]
#![allow(non_camel_case_types)]
tests: use minicore more minicore makes it much easier to add new language items to all of the existing `no_core` tests.
2025-02-24 09:26:54 +00:00
extern crate minicore;
use minicore::*;
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
#[repr(simd)]
pub struct i8x16([i8; 16]);
#[repr(simd)]
pub struct m8x16([i8; 16]);
#[repr(simd)]
pub struct f32x8([f32; 8]);
#[repr(simd)]
pub struct m32x8([i32; 8]);
#[repr(simd)]
pub struct f64x4([f64; 4]);
#[repr(simd)]
pub struct m64x4([i64; 4]);
update/bless tests
2025-04-06 15:12:24 +02:00
#[rustc_intrinsic]
unsafe fn simd_masked_load<M, P, T>(mask: M, pointer: P, values: T) -> T;
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
// CHECK-LABEL: load_i8x16
#[no_mangle]
pub unsafe extern "C" fn load_i8x16(mask: m8x16, pointer: *const i8) -> i8x16 {
// Since avx2 supports no masked loads for bytes, the code tests each individual bit
// and jumps to code that inserts individual bytes.
Consistently use the most significant bit of vector masks This improves the codegen for vector `select`, `gather`, `scatter` and boolean reduction intrinsics and fixes rust-lang/portable-simd#316. The current behavior of most mask operations during llvm codegen is to truncate the mask vector to <N x i1>, telling llvm to use the least significat bit. The exception is the `simd_bitmask` intrinsics, which already used the most signifiant bit. Since sse/avx instructions are defined to use the most significant bit, truncating means that llvm has to insert a left shift to move the bit into the most significant position, before the mask can actually be used. Similarly on aarch64, mask operations like blend work bit by bit, repeating the least significant bit across the whole lane involves shifting it into the sign position and then comparing against zero. By shifting before truncating to <N x i1>, we tell llvm that we only consider the most significant bit, removing the need for additional shift instructions in the assembly.
2023-01-04 23:55:40 +01:00
// x86-avx2-NOT: vpsllw
// x86-avx2-DAG: vpmovmskb eax
// x86-avx2-DAG: vpxor
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
// x86-avx2-NEXT: test al, 1
// x86-avx2-NEXT: jne
// x86-avx2-NEXT: test al, 2
// x86-avx2-NEXT: jne
// x86-avx2-DAG: movzx [[REG:[a-z]+]], byte ptr [rdi]
// x86-avx2-NEXT: vmovd xmm0, [[REG]]
// x86-avx2-DAG: vpinsrb xmm0, xmm0, byte ptr [rdi + 1], 1
//
Consistently use the most significant bit of vector masks This improves the codegen for vector `select`, `gather`, `scatter` and boolean reduction intrinsics and fixes rust-lang/portable-simd#316. The current behavior of most mask operations during llvm codegen is to truncate the mask vector to <N x i1>, telling llvm to use the least significat bit. The exception is the `simd_bitmask` intrinsics, which already used the most signifiant bit. Since sse/avx instructions are defined to use the most significant bit, truncating means that llvm has to insert a left shift to move the bit into the most significant position, before the mask can actually be used. Similarly on aarch64, mask operations like blend work bit by bit, repeating the least significant bit across the whole lane involves shifting it into the sign position and then comparing against zero. By shifting before truncating to <N x i1>, we tell llvm that we only consider the most significant bit, removing the need for additional shift instructions in the assembly.
2023-01-04 23:55:40 +01:00
// x86-avx512-NOT: vpsllw
// x86-avx512: vpmovb2m k1, xmm0
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
// x86-avx512-NEXT: vmovdqu8 xmm0 {k1} {z}, xmmword ptr [rdi]
simd_masked_load(mask, pointer, i8x16([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]))
}
// CHECK-LABEL: load_f32x8
#[no_mangle]
pub unsafe extern "C" fn load_f32x8(mask: m32x8, pointer: *const f32) -> f32x8 {
Consistently use the most significant bit of vector masks This improves the codegen for vector `select`, `gather`, `scatter` and boolean reduction intrinsics and fixes rust-lang/portable-simd#316. The current behavior of most mask operations during llvm codegen is to truncate the mask vector to <N x i1>, telling llvm to use the least significat bit. The exception is the `simd_bitmask` intrinsics, which already used the most signifiant bit. Since sse/avx instructions are defined to use the most significant bit, truncating means that llvm has to insert a left shift to move the bit into the most significant position, before the mask can actually be used. Similarly on aarch64, mask operations like blend work bit by bit, repeating the least significant bit across the whole lane involves shifting it into the sign position and then comparing against zero. By shifting before truncating to <N x i1>, we tell llvm that we only consider the most significant bit, removing the need for additional shift instructions in the assembly.
2023-01-04 23:55:40 +01:00
// x86-avx2-NOT: vpslld
// x86-avx2: vmaskmovps ymm0, ymm0, ymmword ptr [rdi]
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
//
Consistently use the most significant bit of vector masks This improves the codegen for vector `select`, `gather`, `scatter` and boolean reduction intrinsics and fixes rust-lang/portable-simd#316. The current behavior of most mask operations during llvm codegen is to truncate the mask vector to <N x i1>, telling llvm to use the least significat bit. The exception is the `simd_bitmask` intrinsics, which already used the most signifiant bit. Since sse/avx instructions are defined to use the most significant bit, truncating means that llvm has to insert a left shift to move the bit into the most significant position, before the mask can actually be used. Similarly on aarch64, mask operations like blend work bit by bit, repeating the least significant bit across the whole lane involves shifting it into the sign position and then comparing against zero. By shifting before truncating to <N x i1>, we tell llvm that we only consider the most significant bit, removing the need for additional shift instructions in the assembly.
2023-01-04 23:55:40 +01:00
// x86-avx512-NOT: vpslld
// x86-avx512: vpmovd2m k1, ymm0
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
// x86-avx512-NEXT: vmovups ymm0 {k1} {z}, ymmword ptr [rdi]
simd_masked_load(mask, pointer, f32x8([0_f32, 0_f32, 0_f32, 0_f32, 0_f32, 0_f32, 0_f32, 0_f32]))
}
// CHECK-LABEL: load_f64x4
#[no_mangle]
pub unsafe extern "C" fn load_f64x4(mask: m64x4, pointer: *const f64) -> f64x4 {
Consistently use the most significant bit of vector masks This improves the codegen for vector `select`, `gather`, `scatter` and boolean reduction intrinsics and fixes rust-lang/portable-simd#316. The current behavior of most mask operations during llvm codegen is to truncate the mask vector to <N x i1>, telling llvm to use the least significat bit. The exception is the `simd_bitmask` intrinsics, which already used the most signifiant bit. Since sse/avx instructions are defined to use the most significant bit, truncating means that llvm has to insert a left shift to move the bit into the most significant position, before the mask can actually be used. Similarly on aarch64, mask operations like blend work bit by bit, repeating the least significant bit across the whole lane involves shifting it into the sign position and then comparing against zero. By shifting before truncating to <N x i1>, we tell llvm that we only consider the most significant bit, removing the need for additional shift instructions in the assembly.
2023-01-04 23:55:40 +01:00
// x86-avx2-NOT: vpsllq
// x86-avx2: vmaskmovpd ymm0, ymm0, ymmword ptr [rdi]
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
//
Consistently use the most significant bit of vector masks This improves the codegen for vector `select`, `gather`, `scatter` and boolean reduction intrinsics and fixes rust-lang/portable-simd#316. The current behavior of most mask operations during llvm codegen is to truncate the mask vector to <N x i1>, telling llvm to use the least significat bit. The exception is the `simd_bitmask` intrinsics, which already used the most signifiant bit. Since sse/avx instructions are defined to use the most significant bit, truncating means that llvm has to insert a left shift to move the bit into the most significant position, before the mask can actually be used. Similarly on aarch64, mask operations like blend work bit by bit, repeating the least significant bit across the whole lane involves shifting it into the sign position and then comparing against zero. By shifting before truncating to <N x i1>, we tell llvm that we only consider the most significant bit, removing the need for additional shift instructions in the assembly.
2023-01-04 23:55:40 +01:00
// x86-avx512-NOT: vpsllq
// x86-avx512: vpmovq2m k1, ymm0
Add tests for the generated assembly of mask related simd instructions. The tests show that the code generation currently uses the least significant bits of <iX x N> vector masks when converting to <i1 xN>. This leads to an additional left shift operation in the assembly for x86, since mask operations on x86 operate based on the most significant bit. On aarch64 the left shift is followed by a comparison against zero, which repeats the sign bit across the whole lane. The exception, which does not introduce an unneeded shift, is simd_bitmask, because the code generation already shifts before truncating. By using the "C" calling convention the tests should be stable regarding changes in register allocation, but it is possible that future llvm updates will require updating some of the checks. This additional instruction would be removed by the fix in #104693, which uses the most significant bit for all mask operations.
2024-03-02 10:59:11 +01:00
simd_masked_load(mask, pointer, f64x4([0_f64, 0_f64, 0_f64, 0_f64]))
}
Reference in a new issue Copy permalink