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

identified issue with __movmsk.

Browse Source
This commit is contained in:
Evghenii
2014-01-23 10:45:10 +01:00
parent 2e7609156a
commit e87e332d2f
2 changed files with 14 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
builtins/target-nvptx.ll
View File

@@ -722,9 +722,16 @@ svml_stubs(double,d,WIDTH)
define i64 @__movmsk(<1 x i1>) nounwind readnone alwaysinline {
%v = extractelement <1 x i1> %0, i32 0
%call = call i32 @__ballot_nvptx(i1 zeroext %v)
%v64 = zext i32 %call to i64
ret i64 %v64
;; if 0
;; this one fails with ./tests/popcnt-4.ispc and others ...
;; %v0 = call i32 @__ballot_nvptx(i1 %v)
;; %v64 = zext i32 %v0 to i64
;; else
;; this one just copies mask
%v64 = zext i1 %v to i64
;; endif
ret i64 %v64
}
define i1 @__any(<1 x i1>) nounwind readnone alwaysinline {

42
builtins/util-nvptx.m4
View File

@@ -2693,44 +2693,10 @@ declare i8* @malloc(i32)
declare i32 @posix_memalign(i8**, i32, i32)
declare void @free(i8 *)
define noalias i8 * @__new_uniform_32rt(i64 %size) {
%ptr = alloca i8*
%conv = trunc i64 %size to i32
%alignment = load i32* @memory_alignment
%call1 = call i32 @posix_memalign(i8** %ptr, i32 %alignment, i32 %conv)
%ptr_val = load i8** %ptr
ret i8* %ptr_val
}
define <WIDTH x i64> @__new_varying32_32rt(<WIDTH x i32> %size, <WIDTH x MASK> %mask) {
%ret = alloca <WIDTH x i64>
store <WIDTH x i64> zeroinitializer, <WIDTH x i64> * %ret
%ret64 = bitcast <WIDTH x i64> * %ret to i64 *
%alignment = load i32* @memory_alignment
per_lane(WIDTH, <WIDTH x MASK> %mask, `
%sz_LANE_ID = extractelement <WIDTH x i32> %size, i32 LANE
%store_LANE_ID = getelementptr i64 * %ret64, i32 LANE
%ptr_LANE_ID = bitcast i64* %store_LANE_ID to i8**
%call_LANE_ID = call i32 @posix_memalign(i8** %ptr_LANE_ID, i32 %alignment, i32 %sz_LANE_ID)')
%r = load <WIDTH x i64> * %ret
ret <WIDTH x i64> %r
}
define void @__delete_uniform_32rt(i8 * %ptr) {
call void @free(i8 * %ptr)
ret void
}
define void @__delete_varying_32rt(<WIDTH x i64> %ptr, <WIDTH x MASK> %mask) {
per_lane(WIDTH, <WIDTH x MASK> %mask, `
%iptr_LANE_ID = extractelement <WIDTH x i64> %ptr, i32 LANE
%ptr_LANE_ID = inttoptr i64 %iptr_LANE_ID to i8 *
call void @free(i8 * %ptr_LANE_ID)
')
ret void
}
declare noalias i8 * @__new_uniform_32rt(i64 %size);
declare <WIDTH x i64> @__new_varying32_32rt(<WIDTH x i32> %size, <WIDTH x MASK> %mask);
declare void @__delete_uniform_32rt(i8 * %ptr);
declare void @__delete_varying_32rt(<WIDTH x i64> %ptr, <WIDTH x MASK> %mask);
',
RUNTIME, `64',