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

Fixed a number of issues related to memory alignment; a number of places

Browse Source 
were expecting vector-width-aligned pointers where in point of fact,
there's no guarantee that they would have been in general.

Removed the aligned memory allocation routines from some of the examples;
they're no longer needed.

No perf. difference on Core2/Core i5 CPUs; older CPUs may see some
regressions.

Still need to update the documentation for this change and finish reviewing
alignment issues in Load/Store instructions generated by .cpp files.
This commit is contained in:
Matt Pharr
2011-06-23 18:18:33 -07:00
parent d340dcbfcc
commit b84167dddd
11 changed files with 45 additions and 112 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
stdlib.m4
View File

@@ -452,7 +452,7 @@ define internal <$1 x i32> @__load_uint16([0 x i32] *, i32 %offset) nounwind alw
%ptr16 = bitcast [0 x i32] *%0 to i16 *
%ptr = getelementptr i16 * %ptr16, i32 %offset
%ptr64 = bitcast i16 * %ptr to i`'eval(16*$1) *
%val = load i`'eval(16*$1) * %ptr64, align 1
%val = load i`'eval(16*$1) * %ptr64, align 2
%vval = bitcast i`'eval(16*$1) %val to <$1 x i16>
; unsigned, so use zero-extent...
@@ -479,7 +479,7 @@ define internal void @__store_uint8([0 x i32] *, i32 %offset, <$1 x i32> %val32,
%oldmasked = and i`'eval(8*$1) %old, %notmask
%newmasked = and i`'eval(8*$1) %val64, %mask64
%final = or i`'eval(8*$1) %oldmasked, %newmasked
store i`'eval(8*$1) %final, i`'eval(8*$1) * %ptr64
store i`'eval(8*$1) %final, i`'eval(8*$1) * %ptr64, align 1
ret void
}
@@ -498,11 +498,11 @@ define internal void @__store_uint16([0 x i32] *, i32 %offset, <$1 x i32> %val32
%ptr64 = bitcast i16 * %ptr to i`'eval(16*$1) *
;; as above, use mask to do blending with logical ops...
%old = load i`'eval(16*$1) * %ptr64, align 1
%old = load i`'eval(16*$1) * %ptr64, align 2
%oldmasked = and i`'eval(16*$1) %old, %notmask
%newmasked = and i`'eval(16*$1) %val64, %mask64
%final = or i`'eval(16*$1) %oldmasked, %newmasked
store i`'eval(16*$1) %final, i`'eval(16*$1) * %ptr64
store i`'eval(16*$1) %final, i`'eval(16*$1) * %ptr64, align 2
ret void
}
@@ -544,7 +544,7 @@ all_on:
;; vector load
%vecptr = bitcast i32 *%startptr to <$1 x i32> *
%vec_load = load <$1 x i32> *%vecptr, align 4
store <$1 x i32> %vec_load, <$1 x i32> * %val_ptr
store <$1 x i32> %vec_load, <$1 x i32> * %val_ptr, align 4
ret i32 $1
not_all_on: