https://github.com/jasilvanus created
https://github.com/llvm/llvm-project/pull/75448
Vectors are always bit-packed and don't respect the elements' alignment
requirements. This is different from arrays. This means offsets of vector GEPs
need to be computed differently than offsets of array GEPs.
https://github.com/jasilvanus updated
https://github.com/llvm/llvm-project/pull/75448
>From 2c367fba42b716d803ee088af45c1b57fe4bcbcd Mon Sep 17 00:00:00 2001
From: Jannik Silvanus
Date: Thu, 14 Dec 2023 09:24:51 +0100
Subject: [PATCH 1/3] [InstCombine] Precommit test exhibiting miscompile
Inst
https://github.com/jasilvanus updated
https://github.com/llvm/llvm-project/pull/75448
>From 2c367fba42b716d803ee088af45c1b57fe4bcbcd Mon Sep 17 00:00:00 2001
From: Jannik Silvanus
Date: Thu, 14 Dec 2023 09:24:51 +0100
Subject: [PATCH 1/3] [InstCombine] Precommit test exhibiting miscompile
Inst
jasilvanus wrote:
> Can this solve #68566 too?
I don't think it solves it, as it only fixes offset computations within GEPs
and doesn't teach code in general about the correct vector layout. However, it
is reducing the amount of code assuming the wrong layout.
There seem to be some clang test
jasilvanus wrote:
> There seem to be some clang test failures though that I'm currently looking
> into.
Solved: The clang failures were caused by me having `dxv` in `$PATH`, which
implicitly enabled some extra `dxv`-based tests which also fail without this
patch for me.
Clang was invoking `
https://github.com/jasilvanus updated
https://github.com/llvm/llvm-project/pull/75448
>From 2c367fba42b716d803ee088af45c1b57fe4bcbcd Mon Sep 17 00:00:00 2001
From: Jannik Silvanus
Date: Thu, 14 Dec 2023 09:24:51 +0100
Subject: [PATCH 1/4] [InstCombine] Precommit test exhibiting miscompile
Inst
@@ -108,7 +143,23 @@ class generic_gep_type_iterator {
// that.
bool isStruct() const { return isa(CurTy); }
- bool isSequential() const { return isa(CurTy); }
+ bool isVector() const { return isa(CurTy); }
+ bool isSequential() const { return !isStruct(); }
+
+ // For
https://github.com/jasilvanus updated
https://github.com/llvm/llvm-project/pull/75448
>From 2c367fba42b716d803ee088af45c1b57fe4bcbcd Mon Sep 17 00:00:00 2001
From: Jannik Silvanus
Date: Thu, 14 Dec 2023 09:24:51 +0100
Subject: [PATCH 1/5] [InstCombine] Precommit test exhibiting miscompile
Inst
@@ -111,6 +111,20 @@ define void @test_evaluate_gep_as_ptrs_array(ptr
addrspace(2) %B) {
ret void
}
+define void @test_overaligned_vec(i8 %B) {
+; This should be turned into a constexpr instead of being an
instruction
+; CHECK-LABEL: @test_overaligned_vec(
+; TODO:
jasilvanus wrote:
> Alternative would be to forbid GEP indexing into vectors entirely.
I agree that it would be better if there just were no vector GEPs, but that
breaks importing older modules, and that cannot be easily auto-upgraded
(convert to byte-GEPs?). Even worse, DXIL uses different ve
jasilvanus wrote:
I'm now on vacation for two weeks, returning Jan 4th.
I'll merge the PR then (if it is approved by then).
https://github.com/llvm/llvm-project/pull/75448
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/c
https://github.com/jasilvanus closed
https://github.com/llvm/llvm-project/pull/75448
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
jasilvanus wrote:
Interesting, here it is online: https://alive2.llvm.org/ce/z/Bm5gP2
The complaint goes away if `half` is naturally aligned (`f16:16`).
Looks like an Alive2 bug with respect to vector bit layout?
https://github.com/llvm/llvm-project/pull/75448
___
Author: Jannik Silvanus
Date: 2023-06-21T17:18:36+02:00
New Revision: 9741ac5b3b3e7942e3332e37af829e58e5e2bb82
URL:
https://github.com/llvm/llvm-project/commit/9741ac5b3b3e7942e3332e37af829e58e5e2bb82
DIFF:
https://github.com/llvm/llvm-project/commit/9741ac5b3b3e7942e3332e37af829e58e5e2bb82.dif
https://github.com/jasilvanus created
https://github.com/llvm/llvm-project/pull/89228
When serializing a formatting style to YAML, we were emitting a comment `#
BasedOnStyle:
@@ -807,12 +807,18 @@ template <> struct MappingTraits {
FormatStyle PredefinedStyle;
if (getPredefinedStyle(StyleName, Style.Language, &PredefinedStyle) &&
Style == PredefinedStyle) {
- IO.mapOptional("# BasedOnStyle", StyleName);
+
jasilvanus wrote:
> This doesn't feel correct to me...
I'd also be fine with just removing the comment.
https://github.com/llvm/llvm-project/pull/89228
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/list
@@ -807,12 +807,18 @@ template <> struct MappingTraits {
FormatStyle PredefinedStyle;
if (getPredefinedStyle(StyleName, Style.Language, &PredefinedStyle) &&
Style == PredefinedStyle) {
- IO.mapOptional("# BasedOnStyle", StyleName);
-
https://github.com/jasilvanus updated
https://github.com/llvm/llvm-project/pull/89228
>From 5d4a3b0f922b0c28960f610c695c92da7d3538c1 Mon Sep 17 00:00:00 2001
From: Jannik Silvanus
Date: Thu, 18 Apr 2024 14:56:47 +0200
Subject: [PATCH 1/2] [clang-format] Remove YAML hack to emit a BasedOnStyle
@@ -807,12 +807,18 @@ template <> struct MappingTraits {
FormatStyle PredefinedStyle;
if (getPredefinedStyle(StyleName, Style.Language, &PredefinedStyle) &&
Style == PredefinedStyle) {
- IO.mapOptional("# BasedOnStyle", StyleName);
+
jasilvanus wrote:
I've removed the dummy field and the comment now. This should prevent confusion.
I don't know whether anyone actually uses the emitted comment, there is no
explanation,
and tests pass without it.
I suggest that if we later learn that there are valid uses that require the
comm
@@ -807,12 +807,18 @@ template <> struct MappingTraits {
FormatStyle PredefinedStyle;
if (getPredefinedStyle(StyleName, Style.Language, &PredefinedStyle) &&
Style == PredefinedStyle) {
- IO.mapOptional("# BasedOnStyle", StyleName);
-
https://github.com/jasilvanus closed
https://github.com/llvm/llvm-project/pull/89228
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
23 matches
Mail list logo
