aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fix AVX bug introduced in 4ab982bc16

Browse Source
This commit is contained in:
Matt Pharr
2011-08-29 16:50:59 -07:00
parent 58e34ba4ae
commit 6de494cfdb
2 changed files with 38 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
builtins-avx-x2.ll
View File

@@ -524,10 +524,24 @@ define void @__masked_store_64(<16 x i64>* nocapture, <16 x i64>,
;; FIXME: various code elsewhere in the builtins implementations makes ;; FIXME: various code elsewhere in the builtins implementations makes
;; calls to these, basically assuming that doing so is faster than doing ;; calls to the 32/64 bit versions of these, basically assuming that doing
;; a full call to an actual masked store, which isn't likely to be the ;; so is faster than doing a full call to an actual masked store, which
;; case on AVX. So here we provide those functions but then don't actually ;; isn't likely to be the case on AVX. So here we provide those functions
;; do what the caller asked for... ;; but then don't actually do what the caller asked for...
declare void @llvm.trap()
define void @__masked_store_blend_8(<8 x i8>* nocapture, <8 x i8>,
<8 x i32>) nounwind alwaysinline {
call void @llvm.trap()
ret void
}
define void @__masked_store_blend_16(<8 x i16>* nocapture, <8 x i16>,
<8 x i32>) nounwind alwaysinline {
call void @llvm.trap()
ret void
}
define void @__masked_store_blend_32(<16 x i32>* nocapture, <16 x i32>, define void @__masked_store_blend_32(<16 x i32>* nocapture, <16 x i32>,
<16 x i32>) nounwind alwaysinline { <16 x i32>) nounwind alwaysinline {

24
builtins-avx.ll
View File

@@ -450,10 +450,26 @@ define void @__masked_store_64(<8 x i64>* nocapture, <8 x i64>,
;; FIXME: various code elsewhere in the builtins implementations makes ;; FIXME: various code elsewhere in the builtins implementations makes
;; calls to these, basically assuming that doing so is faster than doing ;; calls to the 32/64 bit versions of these, basically assuming that doing
;; a full call to an actual masked store, which isn't likely to be the ;; so is faster than doing a full call to an actual masked store, which
;; case on AVX. So here we provide those functions but then don't actually ;; isn't likely to be the case on AVX. So here we provide those functions
;; do what the caller asked for... ;; but then don't actually do what the caller asked for...
declare void @llvm.trap()
define void @__masked_store_blend_8(<8 x i8>* nocapture, <8 x i8>,
<8 x i32>) nounwind alwaysinline {
call void @llvm.trap()
ret void
}
define void @__masked_store_blend_16(<8 x i16>* nocapture, <8 x i16>,
<8 x i32>) nounwind alwaysinline {
call void @llvm.trap()
ret void
}
define void @__masked_store_blend_32(<8 x i32>* nocapture, <8 x i32>, define void @__masked_store_blend_32(<8 x i32>* nocapture, <8 x i32>,
<8 x i32>) nounwind alwaysinline { <8 x i32>) nounwind alwaysinline {