[llvm] r286233 - [VectorLegalizer] Expansion of CTLZ using CTPOP when possible

[Simon Pilgrim via llvm-commits]

> On 8 Nov 2016, at 19:50, Friedman, Eli <efriedma at codeaurora.org> wrote:
> 
> On 11/8/2016 6:10 AM, Simon Pilgrim via llvm-commits wrote:
>> Author: rksimon
>> Date: Tue Nov  8 08:10:28 2016
>> New Revision: 286233
>> 
>> URL: http://llvm.org/viewvc/llvm-project?rev=286233&view=rev
>> Log:
>> [VectorLegalizer] Expansion of CTLZ using CTPOP when possible
>> 
>> This patch avoids scalarization of CTLZ by instead expanding to use CTPOP (ref: "Hacker's Delight") when the necessary operations are available.
>> 
>> This also adds the necessary cost models for X86 SSE2 targets (the main beneficiary) to ensure vectorization only happens when its useful.
>> 
>> Differential Revision: https://reviews.llvm.org/D25910
>> 
>> Modified: llvm/trunk/test/CodeGen/X86/vec_ctbits.ll
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vec_ctbits.ll?rev=286233&r1=286232&r2=286233&view=diff
>> ==============================================================================
>> --- llvm/trunk/test/CodeGen/X86/vec_ctbits.ll (original)
>> +++ llvm/trunk/test/CodeGen/X86/vec_ctbits.ll Tue Nov  8 08:10:28 2016
>> @@ -36,17 +36,42 @@ define <2 x i64> @footz(<2 x i64> %a) no
>>  define <2 x i64> @foolz(<2 x i64> %a) nounwind {
>>  ; CHECK-LABEL: foolz:
>>  ; CHECK:       # BB#0:
>> -; CHECK-NEXT:    movd %xmm0, %rax
>> -; CHECK-NEXT:    bsrq %rax, %rax
>> -; CHECK-NEXT:    xorq $63, %rax
>> -; CHECK-NEXT:    movd %rax, %xmm1
>> -; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
>> -; CHECK-NEXT:    movd %xmm0, %rax
>> -; CHECK-NEXT:    bsrq %rax, %rax
>> -; CHECK-NEXT:    xorq $63, %rax
>> -; CHECK-NEXT:    movd %rax, %xmm0
>> -; CHECK-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
>> +; CHECK-NEXT:    movdqa %xmm0, %xmm1
>> +; CHECK-NEXT:    psrlq $1, %xmm1
>> +; CHECK-NEXT:    por %xmm0, %xmm1
>> +; CHECK-NEXT:    movdqa %xmm1, %xmm0
>> +; CHECK-NEXT:    psrlq $2, %xmm0
>> +; CHECK-NEXT:    por %xmm1, %xmm0
>> +; CHECK-NEXT:    movdqa %xmm0, %xmm1
>> +; CHECK-NEXT:    psrlq $4, %xmm1
>> +; CHECK-NEXT:    por %xmm0, %xmm1
>> +; CHECK-NEXT:    movdqa %xmm1, %xmm0
>> +; CHECK-NEXT:    psrlq $8, %xmm0
>> +; CHECK-NEXT:    por %xmm1, %xmm0
>> +; CHECK-NEXT:    movdqa %xmm0, %xmm1
>> +; CHECK-NEXT:    psrlq $16, %xmm1
>> +; CHECK-NEXT:    por %xmm0, %xmm1
>>  ; CHECK-NEXT:    movdqa %xmm1, %xmm0
>> +; CHECK-NEXT:    psrlq $32, %xmm0
>> +; CHECK-NEXT:    por %xmm1, %xmm0
>> +; CHECK-NEXT:    pcmpeqd %xmm1, %xmm1
>> +; CHECK-NEXT:    pxor %xmm0, %xmm1
>> +; CHECK-NEXT:    movdqa %xmm1, %xmm0
>> +; CHECK-NEXT:    psrlq $1, %xmm0
>> +; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
>> +; CHECK-NEXT:    psubq %xmm0, %xmm1
>> +; CHECK-NEXT:    movdqa {{.*#+}} xmm0 = [3689348814741910323,3689348814741910323]
>> +; CHECK-NEXT:    movdqa %xmm1, %xmm2
>> +; CHECK-NEXT:    pand %xmm0, %xmm2
>> +; CHECK-NEXT:    psrlq $2, %xmm1
>> +; CHECK-NEXT:    pand %xmm0, %xmm1
>> +; CHECK-NEXT:    paddq %xmm2, %xmm1
>> +; CHECK-NEXT:    movdqa %xmm1, %xmm2
>> +; CHECK-NEXT:    psrlq $4, %xmm2
>> +; CHECK-NEXT:    paddq %xmm1, %xmm2
>> +; CHECK-NEXT:    pand {{.*}}(%rip), %xmm2
>> +; CHECK-NEXT:    pxor %xmm0, %xmm0
>> +; CHECK-NEXT:    psadbw %xmm2, %xmm0
>>  ; CHECK-NEXT:    retq
> 
> This is going from 10 instructions to 35; it doesn't seem like an improvement.

The point is that this is now handled by the cost model, so we only get vectorized code when its really useful. If llc gets a request for ctlz <2 x i64> then it won’t second guess what was decided by an earlier pass.