diff options
| author | Jørgen P. Tjernø <[email protected]> | 2013-12-03 10:34:47 -0800 |
|---|---|---|
| committer | Jørgen P. Tjernø <[email protected]> | 2013-12-03 10:34:47 -0800 |
| commit | 30064a19770809fbb25f5f19a11389812c7f22a9 (patch) | |
| tree | 3f8b0d442e3dae7b429168148fe118e2f843d91b | |
| parent | Fix some accidental new-line stomping. (diff) | |
| download | source-sdk-2013-30064a19770809fbb25f5f19a11389812c7f22a9.tar.xz source-sdk-2013-30064a19770809fbb25f5f19a11389812c7f22a9.zip | |
Specify clobbered registers in POSIX SSE functions
We weren't specifying clobbered registers for our SSE implementations of
VectorNormalize and InvRSquared. Supposedly GCC and clang does not inspect the
assembly to figure out which registers are getting clobbered, so this could
potentially fix some bugs.
This fixes GH-195.
| -rw-r--r-- | sp/src/mathlib/sse.cpp | 6 |
1 files changed, 4 insertions, 2 deletions
diff --git a/sp/src/mathlib/sse.cpp b/sp/src/mathlib/sse.cpp index a2445c79..2260550b 100644 --- a/sp/src/mathlib/sse.cpp +++ b/sp/src/mathlib/sse.cpp @@ -259,6 +259,7 @@ float FASTCALL _SSE_VectorNormalize (Vector& vec) "movaps %%xmm4, %1 \n\t" : "=m" (radius), "=m" (result) : "m" (*v) + : "xmm1", "xmm2", "xmm3", "xmm4" ); #else #error "Not Implemented" @@ -322,8 +323,9 @@ float _SSE_InvRSquared(const float* v) "maxss %%xmm5, %%xmm1 \n\t" "rcpss %%xmm1, %%xmm0 \n\t" "movss %%xmm0, %0 \n\t" - : "=m" (inv_r2) - : "m" (*v), "0" (inv_r2) + : "+m" (inv_r2) + : "m" (*v) + : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" ); #else #error "Not Implemented" |