<table border="1" cellspacing="0" cellpadding="8">
<tr>
<th>Issue</th>
<td>
<a href=https://github.com/llvm/llvm-project/issues/54171>54171</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>
[X86] Failure to recognise when icmp_sgt (pcmpgt) can be used instead of icmp_ugt
</td>
</tr>
<tr>
<th>Labels</th>
<td>
backend:X86
</td>
</tr>
<tr>
<th>Assignees</th>
<td>
</td>
</tr>
<tr>
<th>Reporter</th>
<td>
RKSimon
</td>
</tr>
</table>
<pre>
https://simd.godbolt.org/z/zfsohdWhe
```
#include <x86intrin.h>
void foo(__m256i &mask0, __m256i &mask1, size_t i) {
size_t i8 = (i / 8) * 8;
if ((i - i8) > 0) {
mask0 = _mm256_cmpgt_epi32(_mm256_set1_epi32((int)(i - i8)), _mm256_setr_epi32(0,0,1,1,2,2,3,3));
mask1 = _mm256_cmpgt_epi32(_mm256_set1_epi32((int)(i - i8)), _mm256_setr_epi32(4,4,5,5,6,6,7,7));
}
}
define void @foo(<4 x i64>* %mask0, <4 x i64>* %mask1, i64 %i) {
entry:
%sub = and i64 %i, 7
%cmp.not = icmp eq i64 %sub, 0
br i1 %cmp.not, label %if.end, label %if.then
if.then: ; preds = %entry
%conv = trunc i64 %sub to i32
%vecinit.i.i = insertelement <8 x i32> undef, i32 %conv, i64 0
%vecinit7.i.i = shufflevector <8 x i32> %vecinit.i.i, <8 x i32> poison, <8 x i32> zeroinitializer
%cmp.i = icmp ugt <8 x i32> %vecinit7.i.i, <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
%sext.i = sext <8 x i1> %cmp.i to <8 x i32>
%0 = bitcast <4 x i64>* %mask0 to <8 x i32>*
store <8 x i32> %sext.i, <8 x i32>* %0, align 32, !tbaa !3
%cmp.i18 = icmp ugt <8 x i32> %vecinit7.i.i, <i32 4, i32 4, i32 5, i32 5, i32 6, i32 6, i32 7, i32 7>
%sext.i19 = sext <8 x i1> %cmp.i18 to <8 x i32>
%1 = bitcast <4 x i64>* %mask1 to <8 x i32>*
store <8 x i32> %sext.i19, <8 x i32>* %1, align 32, !tbaa !3
br label %if.end
if.end: ; preds = %if.then, %entry
ret void
}
foo:
andq $7, %rdx
je .LBB0_2
vmovd %edx, %xmm0
vpbroadcastd %xmm0, %ymm0
vpmaxud .LCPI0_0(%rip), %ymm0, %ymm1
vpmaxud .LCPI0_1(%rip), %ymm0, %ymm2
vpcmpeqd %ymm1, %ymm0, %ymm1
vpcmpeqd %ymm2, %ymm0, %ymm0
vmovdqa %ymm1, (%rdi)
vmovdqa %ymm0, (%rsi)
.LBB0_2: # %if.end
vzeroupper
retq
```
IR has simplified this to a icmp_ugt as it recognised that the comparisons are always positive, but we fail to realise that we can lower to VPCMPGT in the backend for the same reason - resulting in the unnecessary VPCMPEQ+VPMAXU icmp_ugt pattern.
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJy1V1uTmzYU_jX4RRMGhMH4wQ9r724n02Zmm7Zp3jwChK0EJFYSXm9-fY4kwDZ4L81MGYSPpHP9jnQkZ6J4Xu21bpQX3Xj4Hl7F6sLfiSITlfaF3MHQD9NKJfbFv3vqBbdecNN9k6B7XRdHjOdVW1DkRZtjmjCuJeP-3ovuHMdBsAKVQng43W5rHCcMeTipifoeeHiDRmOhGVPsB91qxDy8RN5i7fQgePqJFIzdgkhq5O5RahnxDRDRund2yUrDYZk-gIjlie5QcKEV-OC13lid29q4s83rZqe3tGERNn67QUV1OIwZvVyDrnMD9t2gE78c-E2wpoVdw12LXLOi0cSp8P91ag4TpsVdS7q2cO3CKYO_t7jtPBwI-y1oyThFNtXePHDZhvUwR0fEkrlZC5AdD8dD2l-atPmHUdO_zD-FhfVslmznDTCoNrP4EF6cyWzQ4owHQPO50JaPQQfRx54XxA130HNnErHwTMZMViSjlVVc-pQX4yG9p_wciH4osnC97wGEUSNpobpFHbtIz0IQ_GDntGx5fuY90gKZRJ5YDzRnnGmf-cxFzBWVmla0BqUG9NSADiKwEVoOabNwR7g306MfTHUuBqVq35ZlRWFcCzlSeulDl-mz-UYwJfh0_AeVwkgxUsEel6MEslP62t04jpGLnW4TVNBHNxDhhMATIjoRfQ1zq40edY8AkIMXYeeEcxRScuHemQJXYTKmc6L0izsgmKrAw6JXgDidxu9cm6DaabXRA647jiIbpYdDnRFifqMx0mH6a1jPe9QGIp4QyYRYnIgrWIfLN9AGZ1_BO3wX3uEv4x0uX0I8fA_iUG7G1eWykpiR_1JIppWkL0fWg8u6Iqm29fpqPTcF_FRp3QNF9tEawfNFp1AWx0ueb9T9-n-s18EWX04eanEorIKYgqBTcazrYMTWZFKQwuSscPYck-N_vsJfk2NbgM3Nw8dgG9hjMJas6c69Xmggw1flwzflx2E1sBTpYzEkoTPyLtPXRPFV0XHUBsxHcmnNOl6YY_M15uCMWZ2Y-5y9veLgzjdes4MlU8fbpjlVcPfAanu8en_8-BntiYKLXd1UrGS0QHrPlNmSxNagralBwMA06MjFjjNleYiGD0W5qBsizZGiEIGdSqon8qzglFFwlByoiTRrNXqiqCSsMmolhX2pqFMB4znhqBJPVJrJLw-bTw-__Q3HptWekfw7hAi3V2n7itTUKABzcLmSVLWVZnzXs7ec05wqReSz03T3p4fXXx4-3Xz95xRMQ7Smkvsu_lmxiopltCQzzXRFV168_pomXnyL7sHhFkKyPneRoyfYzU6VMrUZp429FJqLkgkkAycMPnDsa0oKJMrB7qyV1ejav2N632Y-YAidqjr0Px8aKb7B4Q5dplRLFRDxPFyEs_0K4wLjNMvjBY1xnsYRzdK0JMlimQQ4yMqZLWnKxOFh3AEIFk1QGENcM7bCAcZBFEQhhk_q5_kiyZI0zpIoxekyhQskrSF43_hi_o7M5Mq6lbU7BZOQPq1Ok0QpqLPUQmf0k1bvhVx9_v0vVgs-sxGsrPs_ASKof4g">