A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
cranelift: lower i128 umin/umax/smin/smax on x86_64 (#14374)
Demetrios Chiuratto Agourakiscf0673fc4e3a
diff --git a/cranelift/codegen/src/isa/x64/lower.isle b/cranelift/codegen/src/isa/x64/lower.isleindex 617616445f..09adefd5ae 100644--- a/cranelift/codegen/src/isa/x64/lower.isle+++ b/cranelift/codegen/src/isa/x64/lower.isle@@ -1799,6 +1799,15 @@(rule -1 (lower (smax (fits_in_64 ty) x y))(cmp_and_choose ty (CC.NL) x y))+(rule 3 (lower (umin $I128 x y))+ (lower_select (emit_cmp (IntCC.UnsignedLessThan) x y) x y))+(rule 3 (lower (umax $I128 x y))+ (lower_select (emit_cmp (IntCC.UnsignedGreaterThan) x y) x y))+(rule 3 (lower (smin $I128 x y))+ (lower_select (emit_cmp (IntCC.SignedLessThan) x y) x y))+(rule 3 (lower (smax $I128 x y))+ (lower_select (emit_cmp (IntCC.SignedGreaterThan) x y) x y))+;; SSE helpers for determining if single-instruction lowerings are available.(decl pure has_pmins (Type) bool)diff --git a/cranelift/filetests/filetests/isa/x64/iminmax-i128.clif b/cranelift/filetests/filetests/isa/x64/iminmax-i128.clifnew file mode 100644index 0000000000..851b38bc02--- /dev/null+++ b/cranelift/filetests/filetests/isa/x64/iminmax-i128.clif@@ -0,0 +1,206 @@+test compile precise-output+set enable_llvm_abi_extensions+set opt_level=speed_and_size+target x86_64++function %select_ugt_to_umin() -> i128 system_v {+block0:+ v0 = iconst.i64 0+ v1 = uextend.i128 v0+ v2 = iconst.i64 0+ v3 = iconcat v2, v2+ v4 = isub v3, v1+ v5 = icmp ugt v4, v3+ v6 = select v5, v3, v4+ return v6+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; uninit %rax+; xorq %rax, %rax+; uninit %rdx+; xorq %rdx, %rdx+; movq %rax, %rdi+; subq $0x0, %rdi+; movq %rdx, %r9+; sbbq $0x0, %r9+; cmpq %rax, %rdi+; movq %r9, %r8+; sbbq %rdx, %r8+; cmovbq %rdi, %rax+; cmovbq %r9, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; xorq %rax, %rax+; xorq %rdx, %rdx+; movq %rax, %rdi+; subq $0, %rdi+; movq %rdx, %r9+; sbbq $0, %r9+; cmpq %rax, %rdi+; movq %r9, %r8+; sbbq %rdx, %r8+; cmovbq %rdi, %rax+; cmovbq %r9, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq++function %umin_i128(i128, i128) -> i128 system_v {+block0(v0: i128, v1: i128):+ v2 = umin v0, v1+ return v2+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; cmpq %rdx, %rdi+; movq %rsi, %r8+; sbbq %rcx, %r8+; movq %rdx, %rax+; cmovbq %rdi, %rax+; movq %rcx, %rdx+; cmovbq %rsi, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; cmpq %rdx, %rdi+; movq %rsi, %r8+; sbbq %rcx, %r8+; movq %rdx, %rax+; cmovbq %rdi, %rax+; movq %rcx, %rdx+; cmovbq %rsi, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq++function %umax_i128(i128, i128) -> i128 system_v {+block0(v0: i128, v1: i128):+ v2 = umax v0, v1+ return v2+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; cmpq %rdi, %rdx+; movq %rcx, %r8+; sbbq %rsi, %r8+; movq %rdx, %rax+; cmovbq %rdi, %rax+; movq %rcx, %rdx+; cmovbq %rsi, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; cmpq %rdi, %rdx+; movq %rcx, %r8+; sbbq %rsi, %r8+; movq %rdx, %rax+; cmovbq %rdi, %rax+; movq %rcx, %rdx+; cmovbq %rsi, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq++function %smin_i128(i128, i128) -> i128 system_v {+block0(v0: i128, v1: i128):+ v2 = smin v0, v1+ return v2+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; cmpq %rdx, %rdi+; movq %rsi, %r8+; sbbq %rcx, %r8+; movq %rdx, %rax+; cmovlq %rdi, %rax+; movq %rcx, %rdx+; cmovlq %rsi, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; cmpq %rdx, %rdi+; movq %rsi, %r8+; sbbq %rcx, %r8+; movq %rdx, %rax+; cmovlq %rdi, %rax+; movq %rcx, %rdx+; cmovlq %rsi, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq++function %smax_i128(i128, i128) -> i128 system_v {+block0(v0: i128, v1: i128):+ v2 = smax v0, v1+ return v2+}++; VCode:+; pushq %rbp+; movq %rsp, %rbp+; block0:+; cmpq %rdi, %rdx+; movq %rcx, %r8+; sbbq %rsi, %r8+; movq %rdx, %rax+; cmovlq %rdi, %rax+; movq %rcx, %rdx+; cmovlq %rsi, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq+;+; Disassembled:+; block0: ; offset 0x0+; pushq %rbp+; movq %rsp, %rbp+; block1: ; offset 0x4+; cmpq %rdi, %rdx+; movq %rcx, %r8+; sbbq %rsi, %r8+; movq %rdx, %rax+; cmovlq %rdi, %rax+; movq %rcx, %rdx+; cmovlq %rsi, %rdx+; movq %rbp, %rsp+; popq %rbp+; retq+diff --git a/cranelift/filetests/filetests/runtests/i128-min-max.clif b/cranelift/filetests/filetests/runtests/i128-min-max.clifindex ea130e5478..29ba1be5af 100644--- a/cranelift/filetests/filetests/runtests/i128-min-max.clif+++ b/cranelift/filetests/filetests/runtests/i128-min-max.clif@@ -5,6 +5,9 @@ target riscv64target riscv64 has_c has_zcbset enable_multi_ret_implicit_srettarget s390x+;; x86_64 now lowers i128 min/max (#13790).+set enable_llvm_abi_extensions+target x86_64function %smin_i128(i128, i128) -> i128 {diff --git a/cranelift/filetests/filetests/runtests/issue-13790-umin-i128.clif b/cranelift/filetests/filetests/runtests/issue-13790-umin-i128.clifnew file mode 100644index 0000000000..6a5bd5c0d6--- /dev/null+++ b/cranelift/filetests/filetests/runtests/issue-13790-umin-i128.clif@@ -0,0 +1,24 @@+test interpret+test run+set enable_llvm_abi_extensions+set opt_level=speed_and_size+target aarch64+target riscv64+target riscv64 has_c has_zcb+set enable_multi_ret_implicit_sret+target s390x+target x86_64++function %select_ugt_to_umin(i128, i128) -> i128 {+block0(v0: i128, v1: i128):+ v2 = icmp ugt v0, v1+ v3 = select v2, v1, v0+ return v3+}+; run: %select_ugt_to_umin(0, 0) == 0+; run: %select_ugt_to_umin(1, 0) == 0+; run: %select_ugt_to_umin(0, 1) == 0+; run: %select_ugt_to_umin(5, 9) == 5+; run: %select_ugt_to_umin(9, 5) == 5+; run: %select_ugt_to_umin(0x10000000000000000, 1) == 1+; run: %select_ugt_to_umin(1, 0x10000000000000000) == 1