ockl: Don't use wave32 ballot builtin

Wave32 and wave64 paths cannot really co-exist in the same function
or callgraph. They need to be treated as a hard ABI incompatibility.
We cannot handle the wave32 operation on wave64, but we can and do
handle the wave64 operation on wave32. Given the current linking scheme,
the most expedient fix for this not working is to pretend wave32 does
not exist and just use the wave64 ballot. The optimizer will fold
the 64-bit ballot intrinsic to the 32-bit one when it sees a 32-bit
target.

This was reported broken in
llvm#89332

Change-Id: If897f631066e257e18d8adf574cd17d3f9278ca7

Author: arsenm

amd/device-libs/ockl/src/dm.cl

@@ -294,9 +294,14 @@ first(__global void * v)
     return __builtin_astype(w2, __global void *);
 }
 
-REQUIRES_WAVE64
+// Read val from one active lane whose predicate is one.
+// If no lanes have the predicate set, return none
+// This is like first, except that first may not have its predicate set
 static uint
-elect_uint_wave64(int pred, uint val, uint none) {
+elect_uint(int pred, uint val, uint none)
+{
+  // Pretend wave32 doesn't exist. The wave64 ballot works, and the high half
+  // will fold out as 0.
     uint ret = none;
 
     ulong mask = __builtin_amdgcn_ballot_w64(pred != 0);
@@ -308,49 +313,12 @@ elect_uint_wave64(int pred, uint val, uint none) {
     return ret;
 }
 
-REQUIRES_WAVE32
-static uint
-elect_uint_wave32(int pred, uint val, uint none) {
-    uint ret = none;
-    uint mask = __builtin_amdgcn_ballot_w32(pred != 0);
-    if (mask != 0U) {
-        uint l = __ockl_ctz_u32(mask);
-        ret = __builtin_amdgcn_ds_bpermute(l << 2, val);
-    }
-
-    return ret;
-}
-
-// Read val from one active lane whose predicate is one.
-// If no lanes have the predicate set, return none
-// This is like first, except that first may not have its predicate set
-static uint
-elect_uint(int pred, uint val, uint none)
-{
-    return __oclc_wavefrontsize64 ?  elect_uint_wave64(pred, val, none) : elect_uint_wave32(pred, val, none);
-}
-
-REQUIRES_WAVE64
-static uint
-votes_wave64(bool b)
-{
-    ulong mask = __builtin_amdgcn_ballot_w64(b);
-    return __builtin_popcountl(mask);
-}
-
-REQUIRES_WAVE32
-static uint
-votes_wave32(bool b)
-{
-    uint mask = __builtin_amdgcn_ballot_w32(b);
-    return __builtin_popcount(mask);
-}
-
 // Count the number of nonzero arguments across the wave
 static uint
 votes(bool b)
 {
-    return __oclc_wavefrontsize64 ?  votes_wave64(b) : votes_wave32(b);
+    ulong mask = __builtin_amdgcn_ballot_w64(b);
+    return __builtin_popcountl(mask);
 }
 
 // The kind of the smallest block that can hold sz bytes

amd/device-libs/ockl/src/wfaas.cl

@@ -22,60 +22,25 @@ static int optimizationBarrierHack(int in_val)
     return out_val;
 }
 
-REQUIRES_WAVE32
-static bool wfany_impl_w32(int e) {
-    return __builtin_amdgcn_ballot_w32(e) != 0;
-}
-
-REQUIRES_WAVE64
-static bool wfany_impl_w64(int e) {
-    return __builtin_amdgcn_ballot_w64(e) != 0;
-}
-
 ATTR bool
 OCKL_MANGLE_I32(wfany)(int e)
 {
     e = optimizationBarrierHack(e);
-    return __oclc_wavefrontsize64 ?
-        wfany_impl_w64(e) : wfany_impl_w32(e);
-}
-
-REQUIRES_WAVE32
-static bool wfall_impl_w32(int e) {
-    return __builtin_amdgcn_ballot_w32(e) == __builtin_amdgcn_read_exec_lo();
-}
-
-REQUIRES_WAVE64
-static bool wfall_impl_w64(int e) {
-    return __builtin_amdgcn_ballot_w64(e) == __builtin_amdgcn_read_exec();
+    return __builtin_amdgcn_ballot_w64(e) != 0;
 }
 
 ATTR bool
 OCKL_MANGLE_I32(wfall)(int e)
 {
     e = optimizationBarrierHack(e);
-    return __oclc_wavefrontsize64 ?
-        wfall_impl_w64(e) : wfall_impl_w32(e);
-}
-
-
-REQUIRES_WAVE32
-static bool wfsame_impl_w32(int e) {
-    uint u = __builtin_amdgcn_ballot_w32(e);
-    return (u == 0) | (u == __builtin_amdgcn_read_exec_lo());
-}
-
-REQUIRES_WAVE64
-static bool wfsame_impl_w64(int e) {
-    ulong u = __builtin_amdgcn_ballot_w64(e);
-    return (u == 0UL) | (u == __builtin_amdgcn_read_exec());
+    return __builtin_amdgcn_ballot_w64(e) == __builtin_amdgcn_read_exec();
 }
 
 ATTR bool
 OCKL_MANGLE_I32(wfsame)(int e)
 {
     e = optimizationBarrierHack(e);
-    return __oclc_wavefrontsize64 ?
-        wfsame_impl_w64(e) : wfsame_impl_w32(e);
+    ulong u = __builtin_amdgcn_ballot_w64(e);
+    return (u == 0UL) | (u == __builtin_amdgcn_read_exec());
 }