motiejus/zig
1
Fork 0
Code Releases Activity

stage1: Let LLVM legalize the overflowing ops on vectors

Browse Source 
LLVM is smart enough to produce tight overflow checking sequences.

On x86_64 a simple v4u32 addition is turned from:

```
  c5 e9 d4 d4                  	vpaddq	%xmm4, %xmm2, %xmm2
  c5 f1 d4 cb                  	vpaddq	%xmm3, %xmm1, %xmm1
  c5 f8 28 e9                  	vmovaps	%xmm1, %xmm5
  c4 e3 55 18 ea 01            	vinsertf128	$1, %xmm2, %ymm5, %ymm5
  c5 f0 c6 da 88               	vshufps	$136, %xmm2, %xmm1, %xmm3
  c5 c8 57 f6                  	vxorps	%xmm6, %xmm6, %xmm6
  c4 e3 55 0c ee aa            	vblendps	$170, %ymm6, %ymm5, %ymm5
  c4 e3 7d 19 ec 01            	vextractf128	$1, %ymm5, %xmm4
  c4 e2 69 29 d4               	vpcmpeqq	%xmm4, %xmm2, %xmm2
  c5 e9 ef d0                  	vpxor	%xmm0, %xmm2, %xmm2
  c4 e2 71 29 cd               	vpcmpeqq	%xmm5, %xmm1, %xmm1
  c5 f1 ef c0                  	vpxor	%xmm0, %xmm1, %xmm0
  c5 f9 6b c2                  	vpackssdw	%xmm2, %xmm0, %xmm0
  c5 f8 50 c0                  	vmovmskps	%xmm0, %eax
  84 c0                        	testb	%al, %al
```

into:

```
  c5 f9 fe c9                  	vpaddd	%xmm1, %xmm0, %xmm1
  c4 e2 71 3f c0               	vpmaxud	%xmm0, %xmm1, %xmm0
  c5 f1 76 c0                  	vpcmpeqd	%xmm0, %xmm1, %xmm0
  c5 f8 50 c0                  	vmovmskps	%xmm0, %eax
  34 0f                        	xorb	$15, %al
  0f 95 c1                     	setne	%cl
  f6 c1 01                     	testb	$1, %cl
```
This commit is contained in:
LemonBoy
2020-10-17 13:11:10 +02:00
committed by Andrew Kelley
parent 8f1408578c
commit 1ac512df70
1 changed files with 9 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
src/stage1/codegen.cpp
View File

@@ -1593,35 +1593,16 @@ static const BuildBinOpFunc unsigned_op[3] = { LLVMBuildNUWAdd, LLVMBuildNUWSub,
static LLVMValueRef gen_overflow_op(CodeGen *g, ZigType *operand_type, AddSubMul op,
LLVMValueRef val1, LLVMValueRef val2)
{
LLVMValueRef overflow_bit;
LLVMValueRef result;
LLVMValueRef fn_val = get_int_overflow_fn(g, operand_type, op);
LLVMValueRef params[] = {
val1,
val2,
};
LLVMValueRef result_struct = LLVMBuildCall(g->builder, fn_val, params, 2, "");
LLVMValueRef result = LLVMBuildExtractValue(g->builder, result_struct, 0, "");
LLVMValueRef overflow_bit = LLVMBuildExtractValue(g->builder, result_struct, 1, "");
if (operand_type->id == ZigTypeIdVector) {
ZigType *int_type = operand_type->data.vector.elem_type;
assert(int_type->id == ZigTypeIdInt);
LLVMTypeRef one_more_bit_int = LLVMIntType(int_type->data.integral.bit_count + 1);
LLVMTypeRef one_more_bit_int_vector = LLVMVectorType(one_more_bit_int, operand_type->data.vector.len);
const auto buildExtFn = int_type->data.integral.is_signed ? LLVMBuildSExt : LLVMBuildZExt;
LLVMValueRef extended1 = buildExtFn(g->builder, val1, one_more_bit_int_vector, "");
LLVMValueRef extended2 = buildExtFn(g->builder, val2, one_more_bit_int_vector, "");
LLVMValueRef extended_result = wrap_op[op](g->builder, extended1, extended2, "");
result = LLVMBuildTrunc(g->builder, extended_result, get_llvm_type(g, operand_type), "");
LLVMValueRef re_extended_result = buildExtFn(g->builder, result, one_more_bit_int_vector, "");
LLVMValueRef overflow_vector = LLVMBuildICmp(g->builder, LLVMIntNE, extended_result, re_extended_result, "");
LLVMTypeRef bitcast_int_type = LLVMIntType(operand_type->data.vector.len);
LLVMValueRef bitcasted_overflow = LLVMBuildBitCast(g->builder, overflow_vector, bitcast_int_type, "");
LLVMValueRef zero = LLVMConstNull(bitcast_int_type);
overflow_bit = LLVMBuildICmp(g->builder, LLVMIntNE, bitcasted_overflow, zero, "");
} else {
LLVMValueRef fn_val = get_int_overflow_fn(g, operand_type, op);
LLVMValueRef params[] = {
val1,
val2,
};
LLVMValueRef result_struct = LLVMBuildCall(g->builder, fn_val, params, 2, "");
result = LLVMBuildExtractValue(g->builder, result_struct, 0, "");
overflow_bit = LLVMBuildExtractValue(g->builder, result_struct, 1, "");
overflow_bit = ZigLLVMBuildOrReduce(g->builder, overflow_bit);
}
LLVMBasicBlockRef fail_block = LLVMAppendBasicBlock(g->cur_fn_val, "OverflowFail");