1c4880a2d3
If we use multiply it would be with 0x0101 which is 1 more than a power of 2. On some targets we would expand this to shl+add. By avoiding the multiply earlier, we can generate better code. Note, PowerPC doesn't do the shl+add expansion of multiply so one of the tests increased in instruction count. Limiting to scalars because it almost always increased the number of instructions in vector tests. Reviewed By: RKSimon Differential Revision: https://reviews.llvm.org/D125638
158 lines
5.1 KiB
LLVM
158 lines
5.1 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-popcnt,+sse2 | FileCheck %s --check-prefix=NOPOPCNT
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+popcnt,+sse2 | FileCheck %s --check-prefix=POPCNT
|
|
|
|
define i1 @noncanonical_parity(<16 x i1> %x) {
|
|
; NOPOPCNT-LABEL: noncanonical_parity:
|
|
; NOPOPCNT: # %bb.0:
|
|
; NOPOPCNT-NEXT: psllw $7, %xmm0
|
|
; NOPOPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; NOPOPCNT-NEXT: xorb %ah, %al
|
|
; NOPOPCNT-NEXT: setnp %al
|
|
; NOPOPCNT-NEXT: retq
|
|
;
|
|
; POPCNT-LABEL: noncanonical_parity:
|
|
; POPCNT: # %bb.0:
|
|
; POPCNT-NEXT: psllw $7, %xmm0
|
|
; POPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; POPCNT-NEXT: popcntl %eax, %eax
|
|
; POPCNT-NEXT: andl $1, %eax
|
|
; POPCNT-NEXT: # kill: def $al killed $al killed $eax
|
|
; POPCNT-NEXT: retq
|
|
%r = call i1 @llvm.vector.reduce.xor.v16i1(<16 x i1> %x)
|
|
ret i1 %r
|
|
}
|
|
define i1 @canonical_parity(<16 x i1> %x) {
|
|
; NOPOPCNT-LABEL: canonical_parity:
|
|
; NOPOPCNT: # %bb.0:
|
|
; NOPOPCNT-NEXT: psllw $7, %xmm0
|
|
; NOPOPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; NOPOPCNT-NEXT: xorb %ah, %al
|
|
; NOPOPCNT-NEXT: setnp %al
|
|
; NOPOPCNT-NEXT: retq
|
|
;
|
|
; POPCNT-LABEL: canonical_parity:
|
|
; POPCNT: # %bb.0:
|
|
; POPCNT-NEXT: psllw $7, %xmm0
|
|
; POPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; POPCNT-NEXT: popcntl %eax, %eax
|
|
; POPCNT-NEXT: andl $1, %eax
|
|
; POPCNT-NEXT: # kill: def $al killed $al killed $eax
|
|
; POPCNT-NEXT: retq
|
|
%i1 = bitcast <16 x i1> %x to i16
|
|
%i2 = call i16 @llvm.ctpop.i16(i16 %i1)
|
|
%i3 = and i16 %i2, 1
|
|
%i4 = icmp ne i16 %i3, 0
|
|
ret i1 %i4
|
|
}
|
|
define i1 @canonical_parity_noncanonical_pred(<16 x i1> %x) {
|
|
; NOPOPCNT-LABEL: canonical_parity_noncanonical_pred:
|
|
; NOPOPCNT: # %bb.0:
|
|
; NOPOPCNT-NEXT: psllw $7, %xmm0
|
|
; NOPOPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; NOPOPCNT-NEXT: movl %eax, %ecx
|
|
; NOPOPCNT-NEXT: shrl %ecx
|
|
; NOPOPCNT-NEXT: andl $21845, %ecx # imm = 0x5555
|
|
; NOPOPCNT-NEXT: subl %ecx, %eax
|
|
; NOPOPCNT-NEXT: movl %eax, %ecx
|
|
; NOPOPCNT-NEXT: andl $13107, %ecx # imm = 0x3333
|
|
; NOPOPCNT-NEXT: shrl $2, %eax
|
|
; NOPOPCNT-NEXT: andl $13107, %eax # imm = 0x3333
|
|
; NOPOPCNT-NEXT: addl %ecx, %eax
|
|
; NOPOPCNT-NEXT: movl %eax, %ecx
|
|
; NOPOPCNT-NEXT: shrl $4, %ecx
|
|
; NOPOPCNT-NEXT: addl %eax, %ecx
|
|
; NOPOPCNT-NEXT: andl $3855, %ecx # imm = 0xF0F
|
|
; NOPOPCNT-NEXT: movl %ecx, %eax
|
|
; NOPOPCNT-NEXT: shrl $8, %eax
|
|
; NOPOPCNT-NEXT: addl %ecx, %eax
|
|
; NOPOPCNT-NEXT: # kill: def $al killed $al killed $eax
|
|
; NOPOPCNT-NEXT: retq
|
|
;
|
|
; POPCNT-LABEL: canonical_parity_noncanonical_pred:
|
|
; POPCNT: # %bb.0:
|
|
; POPCNT-NEXT: psllw $7, %xmm0
|
|
; POPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; POPCNT-NEXT: popcntl %eax, %eax
|
|
; POPCNT-NEXT: # kill: def $al killed $al killed $eax
|
|
; POPCNT-NEXT: retq
|
|
%i1 = bitcast <16 x i1> %x to i16
|
|
%i2 = call i16 @llvm.ctpop.i16(i16 %i1)
|
|
%i3 = and i16 %i2, 1
|
|
%i4 = icmp eq i16 %i3, 1
|
|
ret i1 %i4
|
|
}
|
|
|
|
define i1 @noncanonical_nonparity(<16 x i1> %x) {
|
|
; NOPOPCNT-LABEL: noncanonical_nonparity:
|
|
; NOPOPCNT: # %bb.0:
|
|
; NOPOPCNT-NEXT: psllw $7, %xmm0
|
|
; NOPOPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; NOPOPCNT-NEXT: xorb %ah, %al
|
|
; NOPOPCNT-NEXT: setp %al
|
|
; NOPOPCNT-NEXT: retq
|
|
;
|
|
; POPCNT-LABEL: noncanonical_nonparity:
|
|
; POPCNT: # %bb.0:
|
|
; POPCNT-NEXT: psllw $7, %xmm0
|
|
; POPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; POPCNT-NEXT: popcntl %eax, %eax
|
|
; POPCNT-NEXT: andl $1, %eax
|
|
; POPCNT-NEXT: xorb $1, %al
|
|
; POPCNT-NEXT: # kill: def $al killed $al killed $eax
|
|
; POPCNT-NEXT: retq
|
|
%r.inv = call i1 @llvm.vector.reduce.xor.v16i1(<16 x i1> %x)
|
|
%r = xor i1 %r.inv, -1
|
|
ret i1 %r
|
|
}
|
|
define i1 @canonical_nonparity(<16 x i1> %x) {
|
|
; NOPOPCNT-LABEL: canonical_nonparity:
|
|
; NOPOPCNT: # %bb.0:
|
|
; NOPOPCNT-NEXT: psllw $7, %xmm0
|
|
; NOPOPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; NOPOPCNT-NEXT: xorb %ah, %al
|
|
; NOPOPCNT-NEXT: setp %al
|
|
; NOPOPCNT-NEXT: retq
|
|
;
|
|
; POPCNT-LABEL: canonical_nonparity:
|
|
; POPCNT: # %bb.0:
|
|
; POPCNT-NEXT: psllw $7, %xmm0
|
|
; POPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; POPCNT-NEXT: popcntl %eax, %eax
|
|
; POPCNT-NEXT: testb $1, %al
|
|
; POPCNT-NEXT: sete %al
|
|
; POPCNT-NEXT: retq
|
|
%i1 = bitcast <16 x i1> %x to i16
|
|
%i2 = call i16 @llvm.ctpop.i16(i16 %i1)
|
|
%i3 = and i16 %i2, 1
|
|
%i4 = icmp eq i16 %i3, 0
|
|
ret i1 %i4
|
|
}
|
|
define i1 @canonical_nonparity_noncanonical_pred(<16 x i1> %x) {
|
|
; NOPOPCNT-LABEL: canonical_nonparity_noncanonical_pred:
|
|
; NOPOPCNT: # %bb.0:
|
|
; NOPOPCNT-NEXT: psllw $7, %xmm0
|
|
; NOPOPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; NOPOPCNT-NEXT: xorb %ah, %al
|
|
; NOPOPCNT-NEXT: setp %al
|
|
; NOPOPCNT-NEXT: retq
|
|
;
|
|
; POPCNT-LABEL: canonical_nonparity_noncanonical_pred:
|
|
; POPCNT: # %bb.0:
|
|
; POPCNT-NEXT: psllw $7, %xmm0
|
|
; POPCNT-NEXT: pmovmskb %xmm0, %eax
|
|
; POPCNT-NEXT: popcntl %eax, %eax
|
|
; POPCNT-NEXT: andl $1, %eax
|
|
; POPCNT-NEXT: xorb $1, %al
|
|
; POPCNT-NEXT: # kill: def $al killed $al killed $eax
|
|
; POPCNT-NEXT: retq
|
|
%i1 = bitcast <16 x i1> %x to i16
|
|
%i2 = call i16 @llvm.ctpop.i16(i16 %i1)
|
|
%i3 = and i16 %i2, 1
|
|
%i4 = icmp ne i16 %i3, 1
|
|
ret i1 %i4
|
|
}
|
|
|
|
declare i1 @llvm.vector.reduce.xor.v16i1(<16 x i1>)
|
|
declare i16 @llvm.ctpop.i16(i16)
|