libclc: llvm assembly implementation of clz

Untested... currently crashes in the same manner as add_sat.

5 files changed, 151 insertions, 11 deletions

diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
index 5bb3a11..4aeb85c 100644
--- a/generic/lib/SOURCES
+++ b/generic/lib/SOURCES
@@ -9,6 +9,8 @@ integer/add_sat.cl
 integer/add_sat.ll
 integer/add_sat_impl.ll
 integer/clz.cl
+integer/clz.ll
+integer/clz_impl.ll
 integer/rotate.cl
 integer/sub_sat.cl
 integer/sub_sat.ll
diff --git a/generic/lib/integer/clz.cl b/generic/lib/integer/clz.cl
index 93994a4..83ef2dd 100644
--- a/generic/lib/integer/clz.cl
+++ b/generic/lib/integer/clz.cl
@@ -1,4 +1,52 @@
 #include <clc/clc.h>
 
-#define BODY <clz.inc>
-#include <clc/integer/gentype.inc>
+// From clz.ll
+_CLC_DECL char   __clc_clz_s8(char);
+_CLC_DECL uchar  __clc_clz_u8(uchar);
+_CLC_DECL short  __clc_clz_s16(short);
+_CLC_DECL ushort __clc_clz_u16(ushort);
+_CLC_DECL int    __clc_clz_s32(int);
+_CLC_DECL uint   __clc_clz_u32(uint);
+_CLC_DECL long   __clc_clz_s64(long);
+_CLC_DECL ulong  __clc_clz_u64(ulong);
+
+_CLC_OVERLOAD _CLC_DEF char clz(char x) {
+  return __clc_clz_s8(x);
+}
+
+_CLC_OVERLOAD _CLC_DEF uchar clz(uchar x) {
+  return __clc_clz_u8(x);
+}
+
+_CLC_OVERLOAD _CLC_DEF short clz(short x) {
+  return __clc_clz_s16(x);
+}
+
+_CLC_OVERLOAD _CLC_DEF ushort clz(ushort x) {
+  return __clc_clz_u16(x);
+}
+
+_CLC_OVERLOAD _CLC_DEF int clz(int x) {
+  return __clc_clz_s32(x);
+}
+
+_CLC_OVERLOAD _CLC_DEF uint clz(uint x) {
+  return __clc_clz_u32(x);
+}
+
+_CLC_OVERLOAD _CLC_DEF long clz(long x) {
+  return __clc_clz_s64(x);
+}
+
+_CLC_OVERLOAD _CLC_DEF ulong clz(ulong x) {
+  return __clc_clz_u64(x);
+}
+
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, char, clz, char)
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, uchar, clz, uchar)
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, short, clz, short)
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, ushort, clz, ushort)
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, int, clz, int)
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, uint, clz, uint)
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, long, clz, long)
+_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, ulong, clz, ulong)
diff --git a/generic/lib/integer/clz.inc b/generic/lib/integer/clz.inc
deleted file mode 100644
index 2fa1ad9..0000000
--- a/generic/lib/integer/clz.inc
+++ /dev/null
@@ -1,9 +0,0 @@
-_CLC_OVERLOAD _CLC_DEF GENTYPE clz(GENTYPE x) {
-#if (GENTYPE == 'int') || (GENTYPE == 'uint')
-    //Doesn't seem to actually work.
-    return __builtin_clz(x);
-#else
-    //BAh!
-    return 0;
-#endif
-}
diff --git a/generic/lib/integer/clz.ll b/generic/lib/integer/clz.ll
new file mode 100644
index 0000000..23dfc74
--- /dev/null
+++ b/generic/lib/integer/clz.ll
@@ -0,0 +1,55 @@
+declare i8 @__clc_clz_impl_s8(i8 %x)
+
+define i8 @__clc_clz_s8(i8 %x) nounwind readnone alwaysinline {
+  %call = call i8 @__clc_clz_impl_s8(i8 %x)
+  ret i8 %call
+}
+
+declare i8 @__clc_clz_impl_u8(i8 %x)
+
+define i8 @__clc_clz_u8(i8 %x) nounwind readnone alwaysinline {
+  %call = call i8 @__clc_clz_impl_u8(i8 %x)
+  ret i8 %call
+}
+
+declare i16 @__clc_clz_impl_s16(i16 %x)
+
+define i16 @__clc_clz_s16(i16 %x) nounwind readnone alwaysinline {
+  %call = call i16 @__clc_clz_impl_s16(i16 %x)
+  ret i16 %call
+}
+
+declare i16 @__clc_clz_impl_u16(i16 %x)
+
+define i16 @__clc_clz_u16(i16 %x) nounwind readnone alwaysinline {
+  %call = call i16 @__clc_clz_impl_u16(i16 %x)
+  ret i16 %call
+}
+
+declare i32 @__clc_clz_impl_s32(i32 %x)
+
+define i32 @__clc_clz_s32(i32 %x) nounwind readnone alwaysinline {
+  %call = call i32 @__clc_clz_impl_s32(i32 %x)
+  ret i32 %call
+}
+
+declare i32 @__clc_clz_impl_u32(i32 %x)
+
+define i32 @__clc_clz_u32(i32 %x) nounwind readnone alwaysinline {
+  %call = call i32 @__clc_clz_impl_u32(i32 %x)
+  ret i32 %call
+}
+
+declare i64 @__clc_clz_impl_s64(i64 %x)
+
+define i64 @__clc_clz_s64(i64 %x) nounwind readnone alwaysinline {
+  %call = call i64 @__clc_clz_impl_s64(i64 %x)
+  ret i64 %call
+}
+
+declare i64 @__clc_clz_impl_u64(i64 %x)
+
+define i64 @__clc_clz_u64(i64 %x) nounwind readnone alwaysinline {
+  %call = call i64 @__clc_clz_impl_u64(i64 %x)
+  ret i64 %call
+}
diff --git a/generic/lib/integer/clz_impl.ll b/generic/lib/integer/clz_impl.ll
new file mode 100644
index 0000000..b5c3d98
--- /dev/null
+++ b/generic/lib/integer/clz_impl.ll
@@ -0,0 +1,44 @@
+declare i8 @llvm.ctlz.i8(i8, i1)
+declare i16 @llvm.ctlz.i16(i16, i1)
+declare i32 @llvm.ctlz.i32(i32, i1)
+declare i64 @llvm.ctlz.i64(i64, i1)
+
+define i8 @__clc_clz_impl_s8(i8 %x) nounwind readnone alwaysinline {
+  %call = call i8 @llvm.ctlz.i8(i8 %x, i1 0)
+  ret i8 %call
+}
+
+define i8 @__clc_clz_impl_u8(i8 %x) nounwind readnone alwaysinline {
+  %call = call i8 @llvm.ctlz.i8(i8 %x, i1 0)
+  ret i8 %call
+}
+
+define i16 @__clc_clz_impl_s16(i16 %x) nounwind readnone alwaysinline {
+  %call = call i16 @llvm.ctlz.i16(i16 %x, i1 0)
+  ret i16 %call
+}
+
+define i16 @__clc_clz_impl_u16(i16 %x) nounwind readnone alwaysinline {
+  %call = call i16 @llvm.ctlz.i16(i16 %x, i1 0)
+  ret i16 %call
+}
+
+define i32 @__clc_clz_impl_s32(i32 %x) nounwind readnone alwaysinline {
+  %call = call i32 @llvm.ctlz.i32(i32 %x, i1 0)
+  ret i32 %call
+}
+
+define i32 @__clc_clz_impl_u32(i32 %x) nounwind readnone alwaysinline {
+  %call = call i32 @llvm.ctlz.i32(i32 %x, i1 0)
+  ret i32 %call
+}
+
+define i64 @__clc_clz_impl_s64(i64 %x) nounwind readnone alwaysinline {
+  %call = call i64 @llvm.ctlz.i64(i64 %x, i1 0)
+  ret i64 %call
+}
+
+define i64 @__clc_clz_impl_u64(i64 %x) nounwind readnone alwaysinline {
+  %call = call i64 @llvm.ctlz.i64(i64 %x, i1 0)
+  ret i64 %call
+}