cranelift: Align Scalar and SIMD shift semantics (#4520)

* cranelift: Reorganize test suite

Group some SIMD operations by instruction.

* cranelift: Deduplicate some shift tests

Also, new tests with the mod behaviour

* aarch64: Lower shifts with mod behaviour

* x64: Lower shifts with mod behaviour

* wasmtime: Don't mask SIMD shifts
This commit is contained in:
Afonso Bordado
2022-07-27 18:54:00 +01:00
committed by GitHub
parent e121c209fc
commit 0508932174
15 changed files with 314 additions and 423 deletions

View File

@@ -13,3 +13,33 @@ block0(v0: i32x4, v1: i32x4, v2: i32x4):
; run: %bitselect_i32x4(0x11111111111111111111111111111111, 0x11111111111111111111111111111111, 0x00000000000000000000000000000000) == 0x11111111111111111111111111111111
; run: %bitselect_i32x4(0x01010011000011110000000011111111, 0x11111111111111111111111111111111, 0x00000000000000000000000000000000) == 0x01010011000011110000000011111111
; run: %bitselect_i32x4(0x00000000000000001111111111111111, 0x00000000000000000000000000000000, 0x11111111111111111111111111111111) == 0x11111111111111110000000000000000
function %bitselect_i8x16(i8x16, i8x16, i8x16) -> i8x16 {
block0(v0: i8x16, v1: i8x16, v2: i8x16):
v3 = bitselect v0, v1, v2
return v3
}
; Remember that bitselect accepts: 1) the selector vector, 2) the "if true" vector, and 3) the "if false" vector.
; run: %bitselect_i8x16([0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 255], [127 0 0 0 0 0 0 0 0 0 0 0 0 0 0 42], [42 0 0 0 0 0 0 0 0 0 0 0 0 0 0 127]) == [42 0 0 0 0 0 0 0 0 0 0 0 0 0 0 42]
function %bitselect_i8x16() -> b1 {
block0:
v0 = vconst.i8x16 [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 255] ; the selector vector
v1 = vconst.i8x16 [127 0 0 0 0 0 0 0 0 0 0 0 0 0 0 42] ; for each 1-bit in v0 the bit of v1 is selected
v2 = vconst.i8x16 [42 0 0 0 0 0 0 0 0 0 0 0 0 0 0 127] ; for each 0-bit in v0 the bit of v2 is selected
v3 = bitselect v0, v1, v2
v4 = extractlane v3, 0
v5 = icmp_imm eq v4, 42
v6 = extractlane v3, 1
v7 = icmp_imm eq v6, 0
v8 = extractlane v3, 15
v9 = icmp_imm eq v8, 42
v10 = band v5, v7
v11 = band v10, v9
return v11
}
; run

View File

@@ -1,216 +0,0 @@
test run
set enable_simd
target aarch64
; target s390x FIXME: s390x implements modulo semantics for shift counts
target x86_64 skylake
; TODO: once available, replace all lane extraction with `icmp + all_ones`
function %ishl_i32x4() -> b1 {
block0:
v0 = iconst.i32 1
v1 = vconst.i32x4 [1 2 4 8]
v2 = ishl v1, v0
v3 = extractlane v2, 0
v4 = icmp_imm eq v3, 2
v5 = extractlane v2, 3
v6 = icmp_imm eq v5, 16
v7 = band v4, v6
return v7
}
; run
function %ishl_too_large_i16x8() -> b1 {
block0:
v0 = iconst.i32 17 ; note that this will shift off the end of each lane
v1 = vconst.i16x8 [1 2 4 8 16 32 64 128]
v2 = ishl v1, v0
v3 = extractlane v2, 0
v4 = icmp_imm eq v3, 0
v5 = extractlane v2, 3
v6 = icmp_imm eq v5, 0
v7 = band v4, v6
return v7
}
; run
function %ushr_i8x16() -> b1 {
block0:
v0 = iconst.i32 1
v1 = vconst.i8x16 [0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15]
v2 = ushr v1, v0
v3 = vconst.i8x16 [0 0 1 1 2 2 3 3 4 4 5 5 6 6 7 7]
v4 = icmp eq v2, v3
v5 = vall_true v4
return v5
}
; run
function %sshr_i8x16() -> b1 {
block0:
v0 = iconst.i32 1
v1 = vconst.i8x16 [0 0xff 2 0xfd 4 0xfb 6 0xf9 8 0xf7 10 0xf5 12 0xf3 14 0xf1]
v2 = sshr v1, v0
v3 = vconst.i8x16 [0 0xff 1 0xfe 2 0xfd 3 0xfc 4 0xfb 5 0xfa 6 0xf9 7 0xf8]
v4 = icmp eq v2, v3
v5 = vall_true v4
return v5
}
; run
function %ishl_i8x16() -> b1 {
block0:
v0 = iconst.i32 1
v1 = vconst.i8x16 [0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15]
v2 = ishl v1, v0
v3 = vconst.i8x16 [0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30]
v4 = icmp eq v2, v3
v5 = vall_true v4
return v5
}
; run
function %ushr_i64x2() -> b1 {
block0:
v0 = iconst.i32 1
v1 = vconst.i64x2 [1 2]
v2 = ushr v1, v0
v3 = extractlane v2, 0
v4 = icmp_imm eq v3, 0
v5 = extractlane v2, 1
v6 = icmp_imm eq v5, 1
v7 = band v4, v6
return v7
}
; run
function %ushr_too_large_i32x4() -> b1 {
block0:
v0 = iconst.i32 33 ; note that this will shift off the end of each lane
v1 = vconst.i32x4 [1 2 4 8]
v2 = ushr v1, v0
v3 = extractlane v2, 0
v4 = icmp_imm eq v3, 0
v5 = extractlane v2, 3
v6 = icmp_imm eq v5, 0
v7 = band v4, v6
return v7
}
; run
function %sshr_i16x8() -> b1 {
block0:
v0 = iconst.i32 1
v1 = vconst.i16x8 [-1 2 4 8 -16 32 64 128]
v2 = sshr v1, v0
v3 = extractlane v2, 0
v4 = icmp_imm eq v3, 0xffff ; because of the shifted-in sign-bit, this remains 0xffff == -1
v5 = extractlane v2, 4
v6 = icmp_imm eq v5, 0xfff8 ; -16 has been shifted to -8 == 0xfff8
v7 = band v4, v6
return v7
}
; run
function %sshr_too_large_i32x4() -> b1 {
block0:
v0 = iconst.i32 33 ; note that this will shift off the end of each lane
v1 = vconst.i32x4 [1 2 4 -8]
v2 = sshr v1, v0
v3 = extractlane v2, 0
v4 = icmp_imm eq v3, 0
v5 = extractlane v2, 3
v6 = icmp_imm eq v5, 0xffff_ffff ; shifting in the sign-bit repeatedly fills the result with 1s
v7 = band v4, v6
return v7
}
; run
function %sshr_i64x2(i64x2, i32) -> i64x2 {
block0(v0:i64x2, v1:i32):
v2 = sshr v0, v1
return v2
}
; run: %sshr_i64x2([1 -1], 0) == [1 -1]
; run: %sshr_i64x2([1 -1], 1) == [0 -1] ; note the -1 shift result
; run: %sshr_i64x2([2 -2], 1) == [1 -1]
; run: %sshr_i64x2([0x80000000_00000000 0x7FFFFFFF_FFFFFFFF], 63) == [0xFFFFFFFF_FFFFFFFF 0]
function %bitselect_i8x16() -> b1 {
block0:
v0 = vconst.i8x16 [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 255] ; the selector vector
v1 = vconst.i8x16 [127 0 0 0 0 0 0 0 0 0 0 0 0 0 0 42] ; for each 1-bit in v0 the bit of v1 is selected
v2 = vconst.i8x16 [42 0 0 0 0 0 0 0 0 0 0 0 0 0 0 127] ; for each 0-bit in v0 the bit of v2 is selected
v3 = bitselect v0, v1, v2
v4 = extractlane v3, 0
v5 = icmp_imm eq v4, 42
v6 = extractlane v3, 1
v7 = icmp_imm eq v6, 0
v8 = extractlane v3, 15
v9 = icmp_imm eq v8, 42
v10 = band v5, v7
v11 = band v10, v9
return v11
}
; run
function %sshr_imm_i32x4() -> b1 {
block0:
v1 = vconst.i32x4 [1 2 4 -8]
v2 = sshr_imm v1, 1
v3 = vconst.i32x4 [0 1 2 -4]
v4 = icmp eq v2, v3
v5 = vall_true v4
return v5
}
; run
function %sshr_imm_i16x8() -> b1 {
block0:
v1 = vconst.i16x8 [1 2 4 -8 0 0 0 0]
v2 = ushr_imm v1, 1
v3 = vconst.i16x8 [0 1 2 32764 0 0 0 0] ; -4 with MSB unset == 32764
v4 = icmp eq v2, v3
v5 = vall_true v4
return v5
}
; run
function %ishl_imm_i64x2() -> b1 {
block0:
v1 = vconst.i64x2 [1 0]
v2 = ishl_imm v1, 1
v3 = vconst.i64x2 [2 0]
v4 = icmp eq v2, v3
v5 = vall_true v4
return v5
}
; run

View File

@@ -1,132 +0,0 @@
test run
target aarch64
; target s390x FIXME: s390x implements modulo semantics for shift counts
set enable_simd
target x86_64 skylake
function %bitselect_i8x16(i8x16, i8x16, i8x16) -> i8x16 {
block0(v0: i8x16, v1: i8x16, v2: i8x16):
v3 = bitselect v0, v1, v2
return v3
}
; Remember that bitselect accepts: 1) the selector vector, 2) the "if true" vector, and 3) the "if false" vector.
; run: %bitselect_i8x16([0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 255], [127 0 0 0 0 0 0 0 0 0 0 0 0 0 0 42], [42 0 0 0 0 0 0 0 0 0 0 0 0 0 0 127]) == [42 0 0 0 0 0 0 0 0 0 0 0 0 0 0 42]
function %vselect_i32x4(i32x4, i32x4) -> i32x4 {
block0(v1: i32x4, v2: i32x4):
; `make_trampoline` still does not know how to convert boolean vector types
; so we load the value directly here.
v0 = vconst.b32x4 [true true false false]
v3 = vselect v0, v1, v2
return v3
}
; Remember that vselect accepts: 1) the selector vector, 2) the "if true" vector, and 3) the "if false" vector.
; run: %vselect_i8x16([1 2 -1 -1], [-1 -1 3 4]) == [1 2 3 4]
; shift left
function %ishl_i8x16(i8x16, i32) -> i8x16 {
block0(v0: i8x16, v1: i32):
v2 = ishl v0, v1
return v2
}
; run: %ishl_i8x16([0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15], 4) == [0x00 0x10 0x20 0x30 0x40 0x50 0x60 0x70 0x80 0x90 0xa0 0xb0 0xc0 0xd0 0xe0 0xf0]
function %ishl_i16x8(i16x8, i32) -> i16x8 {
block0(v0: i16x8, v1: i32):
v2 = ishl v0, v1
return v2
}
; run: %ishl_i16x8([1 2 4 8 16 32 64 128], 17) == [0 0 0 0 0 0 0 0]
function %ishl_i32x4(i32x4, i32) -> i32x4 {
block0(v0: i32x4, v1: i32):
v2 = ishl v0, v1
return v2
}
; run: %ishl_i32x4([1 2 4 8], 1) == [2 4 8 16]
function %ishl_imm_i64x2(i64x2) -> i64x2 {
block0(v0: i64x2):
v2 = ishl_imm v0, 1
return v2
}
; run: %ishl_imm_i64x2([1 0]) == [2 0]
; shift right (logical)
function %ushr_i8x16(i8x16, i32) -> i8x16 {
block0(v0: i8x16, v1: i32):
v2 = ushr v0, v1
return v2
}
; run: %ushr_i8x16([0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15], 1) == [0 0 1 1 2 2 3 3 4 4 5 5 6 6 7 7]
function %ushr_i32x4(i32x4, i32) -> i32x4 {
block0(v0: i32x4, v1: i32):
v2 = ushr v0, v1
return v2
}
; run: %ushr_i32x4([1 2 4 8], 33) == [0 0 0 0]
function %ushr_i64x2(i64x2, i32) -> i64x2 {
block0(v0: i64x2, v1: i32):
v2 = ushr v0, v1
return v2
}
; run: %ushr_i64x2([1 2], 1) == [0 1]
; shift right (arithmetic)
function %sshr_i8x16(i8x16, i32) -> i8x16 {
block0(v0: i8x16, v1: i32):
v2 = sshr v0, v1
return v2
}
; run: %sshr_i8x16([0 0xff 2 0xfd 4 0xfb 6 0xf9 8 0xf7 10 0xf5 12 0xf3 14 0xf1], 1) == [0 0xff 1 0xfe 2 0xfd 3 0xfc 4 0xfb 5 0xfa 6 0xf9 7 0xf8]
function %sshr_i16x8(i16x8, i32) -> i16x8 {
block0(v0: i16x8, v1: i32):
v2 = sshr v0, v1
return v2
}
; note: because of the shifted-in sign-bit, lane 0 remains -1 == 0xffff, whereas lane 4 has been shifted to -8 == 0xfff8
; run: %ushr_i16x8([-1 2 4 8 -16 32 64 128], 1) == [-1 1 2 4 -8 16 32 64]
function %sshr_i32x4(i32x4, i32) -> i32x4 {
block0(v0: i32x4, v1: i32):
v2 = sshr v0, v1
return v2
}
; note: shifting in the sign-bit repeatedly in lane 3 fills the result with 1s (-1 == 0xffff_ffff)
; run: %ushr_i32x4([1 2 4 -8], 33) == [0 0 0 0xffff_ffff]
function %sshr_i64x2(i64x2, i32) -> i64x2 {
block0(v0:i64x2, v1:i32):
v2 = sshr v0, v1
return v2
}
; run: %sshr_i64x2([1 -1], 0) == [1 -1]
; run: %sshr_i64x2([1 -1], 1) == [0 -1] ; note the -1 shift result
; run: %sshr_i64x2([2 -2], 1) == [1 -1]
; run: %sshr_i64x2([0x80000000_00000000 0x7FFFFFFF_FFFFFFFF], 63) == [0xFFFFFFFF_FFFFFFFF 0]
function %sshr_imm_i32x4(i32x4) -> i32x4 {
block0(v0: i32x4):
v1 = sshr_imm v0, 1
return v1
}
; run: %sshr_imm_i32x4([1 2 4 -8]) == [0 1 2 -4]
function %sshr_imm_i16x8(i16x8) -> i16x8 {
block0(v0: i16x8):
v1 = sshr_imm v0, 1
return v1
}
; run: %sshr_imm_i16x8([1 2 4 -8 0 0 0 0]) == [0 1 2 -4 0 0 0 0]

View File

@@ -0,0 +1,46 @@
test run
set enable_simd
target aarch64
target s390x
target x86_64 skylake
function %ishl_i8x16(i8x16, i32) -> i8x16 {
block0(v0: i8x16, v1: i32):
v2 = ishl v0, v1
return v2
}
; run: %ishl_i8x16([0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15], 4) == [0x00 0x10 0x20 0x30 0x40 0x50 0x60 0x70 0x80 0x90 0xa0 0xb0 0xc0 0xd0 0xe0 0xf0]
; run: %ishl_i8x16([0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15], 12) == [0x00 0x10 0x20 0x30 0x40 0x50 0x60 0x70 0x80 0x90 0xa0 0xb0 0xc0 0xd0 0xe0 0xf0]
function %ishl_i16x8(i16x8, i32) -> i16x8 {
block0(v0: i16x8, v1: i32):
v2 = ishl v0, v1
return v2
}
; run: %ishl_i16x8([1 2 4 8 16 32 64 128], 1) == [2 4 8 16 32 64 128 256]
; run: %ishl_i16x8([1 2 4 8 16 32 64 128], 17) == [2 4 8 16 32 64 128 256]
function %ishl_i32x4(i32x4, i32) -> i32x4 {
block0(v0: i32x4, v1: i32):
v2 = ishl v0, v1
return v2
}
; run: %ishl_i32x4([1 2 4 8], 1) == [2 4 8 16]
; run: %ishl_i32x4([1 2 4 8], 33) == [2 4 8 16]
function %ishl_i64x2(i64x2, i32) -> i64x2 {
block0(v0: i64x2, v1: i32):
v2 = ishl v0, v1
return v2
}
; run: %ishl_i64x2([1 2], 1) == [2 4]
; run: %ishl_i64x2([1 2], 65) == [2 4]
function %ishl_imm_i64x2(i64x2) -> i64x2 {
block0(v0: i64x2):
v2 = ishl_imm v0, 1
return v2
}
; run: %ishl_imm_i64x2([1 0]) == [2 0]

View File

@@ -0,0 +1,58 @@
test run
set enable_simd
target aarch64
target s390x
target x86_64 skylake
function %sshr_i8x16(i8x16, i32) -> i8x16 {
block0(v0: i8x16, v1: i32):
v2 = sshr v0, v1
return v2
}
; run: %sshr_i8x16([0 0xff 2 0xfd 4 0xfb 6 0xf9 8 0xf7 10 0xf5 12 0xf3 14 0xf1], 1) == [0 0xff 1 0xfe 2 0xfd 3 0xfc 4 0xfb 5 0xfa 6 0xf9 7 0xf8]
; run: %sshr_i8x16([0 0xff 2 0xfd 4 0xfb 6 0xf9 8 0xf7 10 0xf5 12 0xf3 14 0xf1], 9) == [0 0xff 1 0xfe 2 0xfd 3 0xfc 4 0xfb 5 0xfa 6 0xf9 7 0xf8]
function %sshr_i16x8(i16x8, i32) -> i16x8 {
block0(v0: i16x8, v1: i32):
v2 = sshr v0, v1
return v2
}
; note: because of the shifted-in sign-bit, lane 0 remains -1 == 0xffff, whereas lane 4 has been shifted to -8 == 0xfff8
; run: %sshr_i16x8([-1 2 4 8 -16 32 64 128], 1) == [-1 1 2 4 -8 16 32 64]
; run: %sshr_i16x8([-1 2 4 8 -16 32 64 128], 17) == [-1 1 2 4 -8 16 32 64]
function %sshr_i32x4(i32x4, i32) -> i32x4 {
block0(v0: i32x4, v1: i32):
v2 = sshr v0, v1
return v2
}
; run: %sshr_i32x4([1 2 4 -8], 1) == [0 1 2 -4]
; run: %sshr_i32x4([1 2 4 -8], 33) == [0 1 2 -4]
function %sshr_i64x2(i64x2, i32) -> i64x2 {
block0(v0:i64x2, v1:i32):
v2 = sshr v0, v1
return v2
}
; run: %sshr_i64x2([1 -1], 0) == [1 -1]
; run: %sshr_i64x2([1 -1], 1) == [0 -1] ; note the -1 shift result
; run: %sshr_i64x2([2 -2], 1) == [1 -1]
; run: %sshr_i64x2([0x80000000_00000000 0x7FFFFFFF_FFFFFFFF], 63) == [0xFFFFFFFF_FFFFFFFF 0]
; run: %sshr_i64x2([2 -2], 65) == [1 -1]
function %sshr_imm_i32x4(i32x4) -> i32x4 {
block0(v0: i32x4):
v1 = sshr_imm v0, 1
return v1
}
; run: %sshr_imm_i32x4([1 2 4 -8]) == [0 1 2 -4]
function %sshr_imm_i16x8(i16x8) -> i16x8 {
block0(v0: i16x8):
v1 = sshr_imm v0, 1
return v1
}
; run: %sshr_imm_i16x8([1 2 4 -8 0 0 0 0]) == [0 1 2 -4 0 0 0 0]

View File

@@ -0,0 +1,52 @@
test run
set enable_simd
target aarch64
target s390x
target x86_64 skylake
function %ushr_i8x16(i8x16, i32) -> i8x16 {
block0(v0: i8x16, v1: i32):
v2 = ushr v0, v1
return v2
}
; run: %ushr_i8x16([0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15], 1) == [0 0 1 1 2 2 3 3 4 4 5 5 6 6 7 7]
; run: %ushr_i8x16([0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15], 9) == [0 0 1 1 2 2 3 3 4 4 5 5 6 6 7 7]
function %ushr_i16x8(i16x8, i32) -> i16x8 {
block0(v0: i16x8, v1: i32):
v2 = ushr v0, v1
return v2
}
; run: %ushr_i16x8([0 1 2 3 4 5 6 7], 1) == [0 0 1 1 2 2 3 3]
; run: %ushr_i16x8([0 1 2 3 4 5 6 7], 17) == [0 0 1 1 2 2 3 3]
; run: %ushr_i16x8([1 2 4 -8 0 0 0 0], 1) == [0 1 2 32764 0 0 0 0]
function %ushr_i32x4(i32x4, i32) -> i32x4 {
block0(v0: i32x4, v1: i32):
v2 = ushr v0, v1
return v2
}
; run: %ushr_i32x4([1 2 4 8], 1) == [0 1 2 4]
; run: %ushr_i32x4([1 2 4 8], 33) == [0 1 2 4]
function %ushr_i64x2(i64x2, i32) -> i64x2 {
block0(v0: i64x2, v1: i32):
v2 = ushr v0, v1
return v2
}
; run: %ushr_i64x2([1 2], 1) == [0 1]
; run: %ushr_i64x2([1 2], 65) == [0 1]
function %sshr_imm_i16x8() -> b1 {
block0:
v1 = vconst.i16x8 [1 2 4 -8 0 0 0 0]
v2 = ushr_imm v1, 1
v3 = vconst.i16x8 [0 1 2 32764 0 0 0 0] ; -4 with MSB unset == 32764
v4 = icmp eq v2, v3
v5 = vall_true v4
return v5
}
; run

View File

@@ -72,3 +72,15 @@ block0(v0: b64x2, v1: i64x2, v2: i64x2):
return v3
}
; run: %vselect_p_i64x2([true false], [1 2], [100000000000 200000000000]) == [1 200000000000]
function %vselect_i32x4(i32x4, i32x4) -> i32x4 {
block0(v1: i32x4, v2: i32x4):
; `make_trampoline` still does not know how to convert boolean vector types
; so we load the value directly here.
v0 = vconst.b32x4 [true true false false]
v3 = vselect v0, v1, v2
return v3
}
; Remember that vselect accepts: 1) the selector vector, 2) the "if true" vector, and 3) the "if false" vector.
; run: %vselect_i32x4([1 2 -1 -1], [-1 -1 3 4]) == [1 2 3 4]