make simd_reduce_{mul,add}_unordered use only the 'reassoc' flag, not all fast-math flags

This commit is contained in:
Ralf Jung 2024-02-18 23:10:34 +01:00
parent 25fe3cc69d
commit 3dc631a61a
7 changed files with 30 additions and 22 deletions

View file

@ -1367,17 +1367,17 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> {
pub fn vector_reduce_fmul(&mut self, acc: &'ll Value, src: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceFMul(self.llbuilder, acc, src) }
}
pub fn vector_reduce_fadd_algebraic(&mut self, acc: &'ll Value, src: &'ll Value) -> &'ll Value {
pub fn vector_reduce_fadd_reassoc(&mut self, acc: &'ll Value, src: &'ll Value) -> &'ll Value {
unsafe {
let instr = llvm::LLVMRustBuildVectorReduceFAdd(self.llbuilder, acc, src);
llvm::LLVMRustSetAlgebraicMath(instr);
llvm::LLVMRustSetAllowReassoc(instr);
instr
}
}
pub fn vector_reduce_fmul_algebraic(&mut self, acc: &'ll Value, src: &'ll Value) -> &'ll Value {
pub fn vector_reduce_fmul_reassoc(&mut self, acc: &'ll Value, src: &'ll Value) -> &'ll Value {
unsafe {
let instr = llvm::LLVMRustBuildVectorReduceFMul(self.llbuilder, acc, src);
llvm::LLVMRustSetAlgebraicMath(instr);
llvm::LLVMRustSetAllowReassoc(instr);
instr
}
}

View file

@ -1880,14 +1880,14 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
arith_red!(
simd_reduce_add_unordered: vector_reduce_add,
vector_reduce_fadd_algebraic,
vector_reduce_fadd_reassoc,
false,
add,
0.0
);
arith_red!(
simd_reduce_mul_unordered: vector_reduce_mul,
vector_reduce_fmul_algebraic,
vector_reduce_fmul_reassoc,
false,
mul,
1.0

View file

@ -1619,6 +1619,7 @@ extern "C" {
pub fn LLVMRustSetFastMath(Instr: &Value);
pub fn LLVMRustSetAlgebraicMath(Instr: &Value);
pub fn LLVMRustSetAllowReassoc(Instr: &Value);
// Miscellaneous instructions
pub fn LLVMRustGetInstrProfIncrementIntrinsic(M: &Module) -> &Value;