aarch64: Add bitrev,clz,cls,ctz for i128 values

This commit is contained in:
Afonso Bordado
2021-06-03 16:31:45 +01:00
parent 09fec151eb
commit 9fc89d2316
6 changed files with 329 additions and 63 deletions

View File

@@ -52,6 +52,19 @@ block0(v0: i64):
; nextln: ldp fp, lr, [sp], #16
; nextln: ret
function %a(i128) -> i128 {
block0(v0: i128):
v1 = bitrev v0
return v1
}
; check: stp fp, lr, [sp, #-16]!
; nextln: mov fp, sp
; nextln: rbit x2, x0
; nextln: rbit x0, x1
; nextln: mov x1, x2
; nextln: ldp fp, lr, [sp], #16
; nextln: ret
function %b(i8) -> i8 {
block0(v0: i8):
@@ -103,6 +116,22 @@ block0(v0: i64):
; nextln: ldp fp, lr, [sp], #16
; nextln: ret
function %b(i128) -> i128 {
block0(v0: i128):
v1 = clz v0
return v1
}
; check: stp fp, lr, [sp, #-16]!
; nextln: mov fp, sp
; nextln: clz x1, x1
; nextln: clz x0, x0
; nextln: lsr x2, x1, #6
; nextln: madd x0, x0, x2, x1
; nextln: movz x1, #0
; nextln: ldp fp, lr, [sp], #16
; nextln: ret
function %c(i8) -> i8 {
block0(v0: i8):
v1 = cls v0
@@ -153,6 +182,26 @@ block0(v0: i64):
; nextln: ldp fp, lr, [sp], #16
; nextln: ret
function %c(i128) -> i128 {
block0(v0: i128):
v1 = cls v0
return v1
}
; check: stp fp, lr, [sp, #-16]!
; nextln: mov fp, sp
; nextln: cls x2, x0
; nextln: cls x3, x1
; nextln: eon x0, x1, x0
; nextln: lsr x0, x0, #63
; nextln: madd x0, x2, x0, x0
; nextln: subs xzr, x3, #63
; nextln: csel x0, x0, xzr, eq
; nextln: add x0, x0, x3
; nextln: movz x1, #0
; nextln: ldp fp, lr, [sp], #16
; nextln: ret
function %d(i8) -> i8 {
block0(v0: i8):
v1 = ctz v0
@@ -207,6 +256,24 @@ block0(v0: i64):
; nextln: ldp fp, lr, [sp], #16
; nextln: ret
function %d(i128) -> i128 {
block0(v0: i128):
v1 = ctz v0
return v1
}
; check: stp fp, lr, [sp, #-16]!
; nextln: mov fp, sp
; nextln: rbit x0, x0
; nextln: rbit x1, x1
; nextln: clz x0, x0
; nextln: clz x1, x1
; nextln: lsr x2, x0, #6
; nextln: madd x0, x1, x2, x0
; nextln: movz x1, #0
; nextln: ldp fp, lr, [sp], #16
; nextln: ret
function %d(i128) -> i128 {
block0(v0: i128):
v1 = popcnt v0

View File

@@ -0,0 +1,24 @@
test run
target aarch64
; TODO: Move this test into i128-bitops-count.clif when x86_64 supports it
function %cls_i128(i64, i64) -> i64 {
block0(v0: i64, v1: i64):
v2 = iconcat v0, v1
v3 = cls v2
v4, v5 = isplit v3
v6 = iadd v4, v5
return v6
}
; run: %cls_i128(0x00000000_00000000, 0x00000000_00000000) == 127
; run: %cls_i128(0xFFFFFFFF_FFFFFFFF, 0x00000000_00000000) == 63
; run: %cls_i128(0x00000000_00000000, 0xFFFFFFFF_FFFFFFFF) == 63
; run: %cls_i128(0xFFFFFFFF_FFFFFFFF, 0xFFFFFFFF_FFFFFFFF) == 127
; run: %cls_i128(0xFFFFFFFF_FFFFFFFF, 0x7FFFFFFF_FFFFFFFF) == 0
; run: %cls_i128(0xFFFFFFFF_FFFFFFFF, 0x3FFFFFFF_FFFFFFFF) == 1
; run: %cls_i128(0x7FFFFFFF_FFFFFFFF, 0xFFFFFFFF_FFFFFFFF) == 63
; run: %cls_i128(0x80000000_00000000, 0xC0000000_00000000) == 1
; run: %cls_i128(0x00000000_00000000, 0xC0000000_00000000) == 1
; run: %cls_i128(0x80000000_00000000, 0x80000000_00000000) == 0

View File

@@ -1,30 +1,49 @@
test run
target aarch64
; target s390x TODO: Not yet implemented on s390x
target x86_64 machinst
function %ctz(i64, i64) -> i8 {
function %ctz_i128(i64, i64) -> i64 {
block0(v0: i64, v1: i64):
v2 = iconcat v0, v1
v3 = ctz.i128 v2
v4 = ireduce.i8 v3
return v4
}
; run: %ctz(0x00000000_00000000, 0x00000001_00000000) == 96
; run: %ctz(0x00000000_00010000, 0x00000001_00000000) == 16
; run: %ctz(0x00000000_00010000, 0x00000000_00000000) == 16
; run: %ctz(0x00000000_00000000, 0x00000000_00000000) == 128
function %clz(i64, i64) -> i8 {
v3 = ctz v2
v4, v5 = isplit v3
v6 = iadd v4, v5
return v6
}
; run: %ctz_i128(0x00000000_00000000, 0x00000000_00000000) == 128
; run: %ctz_i128(0xFFFFFFFF_FFFFFFFF, 0x00000000_00000000) == 0
; run: %ctz_i128(0x00000000_00000000, 0xFFFFFFFF_FFFFFFFF) == 64
; run: %ctz_i128(0xFFFFFFFF_FFFFFFFF, 0xFFFFFFFF_FFFFFFFF) == 0
; run: %ctz_i128(0xFFFFFFFF_00000000, 0xF0000000_00000000) == 32
; run: %ctz_i128(0xF0000000_00000000, 0xFF000000_00000000) == 60
; run: %ctz_i128(0x00000001_00000000, 0x00000000_00000000) == 32
; run: %ctz_i128(0x00000000_00000000, 0x00000001_00000000) == 96
; run: %ctz_i128(0x00000000_00010000, 0x00000001_00000000) == 16
; run: %ctz_i128(0x00000000_00010000, 0x00000000_00000000) == 16
function %clz_i128(i64, i64) -> i64 {
block0(v0: i64, v1: i64):
v2 = iconcat v0, v1
v3 = clz.i128 v2
v4 = ireduce.i8 v3
return v4
v3 = clz v2
v4, v5 = isplit v3
v6 = iadd v4, v5
return v6
}
; run: %clz(0x00000000_00000000, 0x00000001_00000000) == 31
; run: %clz(0x00000000_00010000, 0x00000001_00000000) == 31
; run: %clz(0x00000000_00010000, 0x00000000_00000000) == 111
; run: %clz(0x00000000_00000000, 0x00000000_00000000) == 128
; run: %clz_i128(0x00000000_00000000, 0x00000000_00000000) == 128
; run: %clz_i128(0xFFFFFFFF_FFFFFFFF, 0x00000000_00000000) == 64
; run: %clz_i128(0x00000000_00000000, 0xFFFFFFFF_FFFFFFFF) == 0
; run: %clz_i128(0xFFFFFFFF_FFFFFFFF, 0xFFFFFFFF_FFFFFFFF) == 0
; run: %clz_i128(0xFFFFFFFF_FFFFFFFF, 0x40000000_00000000) == 1
; run: %clz_i128(0xFFFFFFFF_FFFFFFFF, 0x20000000_00000000) == 2
; run: %clz_i128(0x00000000_00000000, 0x00000000_80000000) == 32
; run: %clz_i128(0x00000000_00000000, 0x00000001_00000000) == 31
; run: %clz_i128(0x00000000_00010000, 0x00000001_00000000) == 31
; run: %clz_i128(0x00000000_00010000, 0x00000000_00000000) == 111
function %popcnt_i128(i64, i64) -> i64 {
block0(v0: i64, v1: i64):
@@ -36,9 +55,9 @@ block0(v0: i64, v1: i64):
v6 = iadd v4, v5
return v6
}
; run: %popcnt_i128(0, 0) == 0
; run: %popcnt_i128(-1, 0) == 64
; run: %popcnt_i128(0, -1) == 64
; run: %popcnt_i128(-1, -1) == 128
; run: %popcnt_i128(0x00000000_00000000, 0x00000000_00000000) == 0
; run: %popcnt_i128(0xFFFFFFFF_FFFFFFFF, 0x00000000_00000000) == 64
; run: %popcnt_i128(0x00000000_00000000, 0xFFFFFFFF_FFFFFFFF) == 64
; run: %popcnt_i128(0xFFFFFFFF_FFFFFFFF, 0xFFFFFFFF_FFFFFFFF) == 128
; run: %popcnt_i128(0x55555555_55555555, 0x55555555_55555555) == 64
; run: %popcnt_i128(0xC0FFEEEE_DECAFFFF, 0xDECAFFFF_C0FFEEEE) == 96

View File

@@ -133,3 +133,21 @@ return v7, v8
; run: %bxor_not_i128(0x01234567_89ABCDEF, 0xFEDCBA98_76543210, 0xFEDCBA98_76543210, 0x01234567_89ABCDEF) == [0, 0]
; run: %bxor_not_i128(0x8FA50A64_8FA50A64, 0x9440A07D_9440A07D, 0xB0A51B75_B0A51B75, 0xB575A07D_B575A07D) == [0xC0FFEEEE_C0FFEEEE, 0xDECAFFFF_DECAFFFF]
function %bitrev_i128(i64, i64) -> i64, i64 {
block0(v0: i64, v1: i64):
v2 = iconcat v0, v1
v3 = bitrev v2
v4, v5 = isplit v3
return v4, v5
}
; run: %bitrev_i128(0, 0) == [0, 0]
; run: %bitrev_i128(-1, -1) == [-1, -1]
; run: %bitrev_i128(-1, 0) == [0, -1]
; run: %bitrev_i128(0, -1) == [-1, 0]
; run: %bitrev_i128(0x00000000_00000000, 0x80000000_00000000) == [1, 0]
; run: %bitrev_i128(0x01234567_89ABCDEF, 0xFEDCBA98_76543210) == [0x084C2A6E_195D3B7F, 0xF7B3D591_E6A2C480]
; run: %bitrev_i128(0xC0FFEEEE_DECAFFFF, 0xDECAFFFF_C0FFEEEE) == [0x7777FF03_FFFF537B, 0xFFFF537B_7777FF03]