<html><head><meta http-equiv="Content-Type" content="text/html charset=utf-8"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class=""><br class=""><div><blockquote type="cite" class=""><div class="">On 25 Sep 2017, at 19:08, Friedman, Eli via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" class="">llvm-dev@lists.llvm.org</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><div class="moz-cite-prefix" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255);">On 9/25/2017 9:14 AM, Björn Pettersson A wrote:<br class=""></div><blockquote type="cite" cite="mid:HE1PR0701MB25723C2FC2B41565BBA52F69B07A0@HE1PR0701MB2572.eurprd07.prod.outlook.com" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255);" class=""><div class="WordSection1" style="page: WordSection1;"><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">(Not sure if this exactly maps to “truncating store”, but I think it at least touches some of the subjects discussed in this thread)<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""> </o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">Our out-of-tree-target need several patches to get things working correctly for us.<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">We have introduced i24 and i40 types in ValueTypes/MachineValueTypes (in addition to the normal pow-of-2 types). And we have vectors of those (v2i40, v4i40).<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">And the byte size in our target is 16 bits.<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""> </o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">When storing an i40 we need to store it as three 16-bit bytes, i.e. 48 bits.<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">When storing a v4i40 vector it will be stored as 4x48 bits.<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""> </o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">One thing that we have had to patch is the getStoreSize() method in ValueTypes/MachineValueTypes where we assume that vectors are bitpacked when the element size is smaller than the byte size (“BitsPerByte”):<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""> </o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span>/// Return the number of bytes overwritten by a store of the specified value<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span>/// type.<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span>unsigned getStoreSize() const {<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">- return (getSizeInBits() + 7) / 8;<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">+ // We assume that vectors with elements smaller than the byte size are<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">+ // bitpacked. And that elements larger than the byte size should be padded<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">+ // (e.g. i40 type for Phoenix is stored using 3 bytes (48 bits)).<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">+ bool PadElementsToByteSize =<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">+ isVector() && getScalarSizeInBits() >= BitsPerByte;<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">+ if (PadElementsToByteSize)<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">+ return getVectorNumElements() * getScalarType().getStoreSize();<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">+ return (getSizeInBits() + (BitsPerByte-1)) / BitsPerByte;<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span>}<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""> </o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">The patch seems to work for in-tree-target tests as well as our out-of-tree target.<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">If it is a correct assumption for all targets is beyond my knowledge. Maybe only i1 vectors should be bitpacked?<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""> </o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">Anyway, I think the bitpacked cases is very special (we do not use it for our target…).<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">AFAIK bitcast is defined as writing to memory followed by a load using a different type. And I think that doing several scalar operations should give the same result as when using vectors. So bitcast of bitpacked vectors should probably be avoided?</span></div></div></blockquote><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255);" class=""><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255); float: none; display: inline !important;" class="">Yes, store+load is the right definition of bitcast. And in fact, the backend will lower a bitcast to a store+load to a stack temporary in cases where there isn't some other lowering specified.</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255);" class=""><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255);" class=""><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255); float: none; display: inline !important;" class="">The end result is probably going to be pretty inefficient unless your target has a special instruction to handle it (x86 has pmovmskb for i1 vector bitcasts, but otherwise you probably end up with some terrible lowering involving a lot of shifts).</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255);" class=""></div></blockquote><div><br class=""></div><div>We still struggle with this in many cases - llvm/test/CodeGen/X86/vector-compare-results.ll has some pretty shocking cases that haven’t been addressed.</div><div><br class=""></div><div>Weren’t the Embescom chaps working on better support for targets with base types other than 8 bits?</div><br class=""><blockquote type="cite" class=""><blockquote type="cite" cite="mid:HE1PR0701MB25723C2FC2B41565BBA52F69B07A0@HE1PR0701MB2572.eurprd07.prod.outlook.com" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255);" class=""><div class="WordSection1" style="page: WordSection1;"><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span style="font-size: 11pt;" class="">This also reminded me of the following test case that is in trunk: test/CodeGen/X86/pr20011.ll</span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""> </o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">%destTy = type { i2, i2 }<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">define void @crash(i64 %x0, i64 %y0, %destTy* nocapture %dest) nounwind {<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">; X64-LABEL: crash:<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">; X64: # BB#0:<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">; X64-NEXT: andl $3, %esi<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">; X64-NEXT: movb %sil, (%rdx)<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">; X64-NEXT: andl $3, %edi<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">; X64-NEXT: movb %dil, (%rdx)<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">; X64-NEXT: retq<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span>%x1 = trunc i64 %x0 to i2<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span>%y1 = trunc i64 %y0 to i2<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span>%1 = bitcast %destTy* %dest to <2 x i2>*<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span></span><span class="">%2 = insertelement <2 x i2> undef, i2 %x1, i32 0<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span class=""> <span class="Apple-converted-space"> </span></span><span lang="EN-US" class="">%3 = insertelement <2 x i2> %2, i2 %y1, i32 1<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span>store <2 x i2> %3, <2 x i2>* %1, align 1<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""> <span class="Apple-converted-space"> </span>ret void<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">}<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""> </o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">As you can see by the “X64” checks the behavior is quite weird.<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">Both movb instructions writes to the same address. So the result of the store <2 x i2> will be the same as when only storing one of the elements.<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">Is this really expected?<o:p class=""></o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class=""><o:p class=""> </o:p></span></div><div style="margin: 0cm 0cm 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif;" class=""><span lang="EN-US" class="">We have emailed Simon Pilgrim who added the test case to show that we no longer crash on this test case (see<span class="Apple-converted-space"> </span><a href="https://bugs.llvm.org/show_bug.cgi?id=20011" moz-do-not-send="true" style="color: purple; text-decoration: underline;" class="">https://bugs.llvm.org/show_bug.cgi?id=20011</a>). But even if the compiler doesn’t crash, the behavior seems wrong to me.<span class="Apple-converted-space"> </span><br class=""></span></div></div></blockquote><p style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: rgb(255, 255, 255);" class="">Yes, the behavior here is wrong. DAGTypeLegalizer::SplitVecOp_STORE/DAGTypeLegalizer::SplitVecRes_LOAD/etc. assume the element size is a multiple of 8. I'm sure this has been discussed before, but I guess nobody ever wrote a patch to fix it...?</p></blockquote>Sorry I might have missed that email. I ended up creating PR31265 as a meta because there are far too many cases like this, where we don’t correctly pack illegal vector types, PR1784 goes into more details on this. My particular interest was in bool vectors, especially after AVX512 went with the bitpacked data representations for mask registers. </div></body></html>