Explorar el Código

Merge pull request #5124 from Barinzaya/core-simd-bmi-intrinsics

BMI/BMI2 Intrinsics
gingerBill hace 4 meses
padre
commit
e074518983
Se han modificado 2 ficheros con 125 adiciones y 0 borrados
  1. 79 0
      core/simd/x86/bmi.odin
  2. 46 0
      core/simd/x86/bmi2.odin

+ 79 - 0
core/simd/x86/bmi.odin

@@ -0,0 +1,79 @@
+#+build i386, amd64
+package simd_x86
+
+import "base:intrinsics"
+
+@(require_results, enable_target_feature="bmi")
+_andn_u32 :: #force_inline proc "c" (a, b: u32) -> u32 {
+	return a &~ b
+}
+@(require_results, enable_target_feature="bmi")
+_andn_u64 :: #force_inline proc "c" (a, b: u64) -> u64 {
+	return a &~ b
+}
+
+@(require_results, enable_target_feature="bmi")
+_bextr_u32 :: #force_inline proc "c" (a, start, len: u32) -> u32 {
+	return bextr_u32(a, (start & 0xff) | (len << 8))
+}
+@(require_results, enable_target_feature="bmi")
+_bextr_u64 :: #force_inline proc "c" (a: u64, start, len: u32) -> u64 {
+	return bextr_u64(a, cast(u64)((start & 0xff) | (len << 8)))
+}
+
+@(require_results, enable_target_feature="bmi")
+_bextr2_u32 :: #force_inline proc "c" (a, control: u32) -> u32 {
+	return bextr_u32(a, control)
+}
+@(require_results, enable_target_feature="bmi")
+_bextr2_u64 :: #force_inline proc "c" (a, control: u64) -> u64 {
+	return bextr_u64(a, control)
+}
+
+@(require_results, enable_target_feature="bmi")
+_blsi_u32 :: #force_inline proc "c" (a: u32) -> u32 {
+	return a & -a
+}
+@(require_results, enable_target_feature="bmi")
+_blsi_u64 :: #force_inline proc "c" (a: u64) -> u64 {
+	return a & -a
+}
+
+@(require_results, enable_target_feature="bmi")
+_blsmsk_u32 :: #force_inline proc "c" (a: u32) -> u32 {
+	return a ~ (a-1)
+}
+@(require_results, enable_target_feature="bmi")
+_blsmsk_u64 :: #force_inline proc "c" (a: u64) -> u64 {
+	return a ~ (a-1)
+}
+
+@(require_results, enable_target_feature="bmi")
+_blsr_u32 :: #force_inline proc "c" (a: u32) -> u32 {
+	return a & (a-1)
+}
+@(require_results, enable_target_feature="bmi")
+_blsr_u64 :: #force_inline proc "c" (a: u64) -> u64 {
+	return a & (a-1)
+}
+
+@(require_results, enable_target_feature = "bmi")
+_tzcnt_u16 :: #force_inline proc "c" (a: u16) -> u16 {
+	return intrinsics.count_trailing_zeros(a)
+}
+@(require_results, enable_target_feature = "bmi")
+_tzcnt_u32 :: #force_inline proc "c" (a: u32) -> u32 {
+	return intrinsics.count_trailing_zeros(a)
+}
+@(require_results, enable_target_feature = "bmi")
+_tzcnt_u64 :: #force_inline proc "c" (a: u64) -> u64 {
+	return intrinsics.count_trailing_zeros(a)
+}
+
+@(private, default_calling_convention = "none")
+foreign _ {
+	@(link_name = "llvm.x86.bmi.bextr.32")
+	bextr_u32 :: proc(a, control: u32) -> u32 ---
+	@(link_name = "llvm.x86.bmi.bextr.64")
+	bextr_u64 :: proc(a, control: u64) -> u64 ---
+}

+ 46 - 0
core/simd/x86/bmi2.odin

@@ -0,0 +1,46 @@
+#+build i386, amd64
+package simd_x86
+
+@(require_results, enable_target_feature = "bmi2")
+_bzhi_u32 :: #force_inline proc "c" (a, index: u32) -> u32 {
+	return bzhi_u32(a, index)
+}
+@(require_results, enable_target_feature = "bmi2")
+_bzhi_u64 :: #force_inline proc "c" (a, index: u64) -> u64 {
+	return bzhi_u64(a, index)
+}
+
+@(require_results, enable_target_feature = "bmi2")
+_pdep_u32 :: #force_inline proc "c" (a, mask: u32) -> u32 {
+	return pdep_u32(a, mask)
+}
+@(require_results, enable_target_feature = "bmi2")
+_pdep_u64 :: #force_inline proc "c" (a, mask: u64) -> u64 {
+	return pdep_u64(a, mask)
+}
+
+@(require_results, enable_target_feature = "bmi2")
+_pext_u32 :: #force_inline proc "c" (a, mask: u32) -> u32 {
+	return pext_u32(a, mask)
+}
+@(require_results, enable_target_feature = "bmi2")
+_pext_u64 :: #force_inline proc "c" (a, mask: u64) -> u64 {
+	return pext_u64(a, mask)
+}
+
+
+@(private, default_calling_convention = "none")
+foreign _ {
+	@(link_name = "llvm.x86.bmi.bzhi.32")
+	bzhi_u32 :: proc(a, index: u32) -> u32 ---
+	@(link_name = "llvm.x86.bmi.bzhi.64")
+	bzhi_u64 :: proc(a, index: u64) -> u64 ---
+	@(link_name = "llvm.x86.bmi.pdep.32")
+	pdep_u32 :: proc(a, mask: u32) -> u32 ---
+	@(link_name = "llvm.x86.bmi.pdep.64")
+	pdep_u64 :: proc(a, mask: u64) -> u64 ---
+	@(link_name = "llvm.x86.bmi.pext.32")
+	pext_u32 :: proc(a, mask: u32) -> u32 ---
+	@(link_name = "llvm.x86.bmi.pext.64")
+	pext_u64 :: proc(a, mask: u64) -> u64 ---
+}