A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
x64: Remove an add128 lower rule that turns out just hurts (#14056)
Alex Crichton0bcb8223a70e
diff --git a/cranelift/codegen/src/isa/x64/lower.isle b/cranelift/codegen/src/isa/x64/lower.isleindex 7dcec4ee0d..8739c6fb6c 100644--- a/cranelift/codegen/src/isa/x64/lower.isle+++ b/cranelift/codegen/src/isa/x64/lower.isle@@ -134,23 +134,6 @@(iadd128 (value_regs_get_gpr x 0) (value_regs_get_gpr x 1)y (gpr_mem_imm_64_imm 0))))-;; Specialized lowering rule for `iadd` of two 64-bit unsigned integers, meaning-;; that we can skip the `adc` and instead use `setb`. This is in some sense a-;; way of modeling `uadd_overflow`.-(rule 4 (lower (iadd $I128 (uextend _ x) (uextend _ y @ (value_type $I64))))- (let (- (x Gpr (extend_to_gpr x $I64 (ExtendKind.Zero)))- (ret ValueRegs (with_flags (x64_add_with_flags_paired $I64 x y)- (x64_setcc_paired (CC.B))))- )- ;; FIXME: this `movzx` ideally would happen before the `add` itself to- ;; zero out the destination register with `xor %dst,%dst` and then- ;; the `setb` would just write to the lower bytes. That would probably- ;; require modeling this as a pseudo-inst which isn't quite worth it- ;; at this time.- (value_regs (value_regs_get ret 0)- (x64_movzx (ExtMode.BQ) (value_regs_get ret 1)))))-;; Helper for lowering 128-bit addition with the 64-bit halves of the lhs/rhs;; already split. The first two arguments are lo/hi for the lhs and the second;; two are lo/hi for the rhs.diff --git a/cranelift/filetests/filetests/isa/x64/i128.clif b/cranelift/filetests/filetests/isa/x64/i128.clifindex 3b601d4d24..d607d884bf 100644--- a/cranelift/filetests/filetests/isa/x64/i128.clif+++ b/cranelift/filetests/filetests/isa/x64/i128.clif@@ -1111,21 +1111,22 @@ block2(v8: i128):; pushq %rbp; movq %rsp, %rbp; block0:+; movq %rdx, %rcx; uninit %rax; xorq %rax, %rax-; testb %dl, %dl+; uninit %rdx+; xorq %rdx, %rdx+; testb %cl, %cl; jnz label2; j label1; block1:; addq $0x2, %rax-; setb %r11b-; movzbq %r11b, %rdx+; adcq $0x0, %rdx; movq %rbp, %rsp; popq %rbp; retq; block2:; addq $0x1, %rax-; setb %cl-; movzbq %cl, %rdx+; adcq $0x0, %rdx; movq %rbp, %rsp; popq %rbp; retq@@ -1135,20 +1136,20 @@ block2(v8: i128):; pushq %rbp; movq %rsp, %rbp; block1: ; offset 0x4+; movq %rdx, %rcx; xorq %rax, %rax-; testb %dl, %dl-; jne 0x20-; block2: ; offset 0xf+; xorq %rdx, %rdx+; testb %cl, %cl+; jne 0x22+; block2: ; offset 0x15; addq $2, %rax-; setb %r11b-; movzbq %r11b, %rdx+; adcq $0, %rdx; movq %rbp, %rsp; popq %rbp; retq-; block3: ; offset 0x20+; block3: ; offset 0x22; addq $1, %rax-; setb %cl-; movzbq %cl, %rdx+; adcq $0, %rdx; movq %rbp, %rsp; popq %rbp; retq@@ -2055,10 +2056,11 @@ block0(v0: i64, v1: i64):; pushq %rbp; movq %rsp, %rbp; block0:+; uninit %rdx+; xorq %rdx, %rdx; movq %rsi, %rax; addq (%rdi), %rax-; setb %sil-; movzbq %sil, %rdx+; adcq $0x0, %rdx; movq %rbp, %rsp; popq %rbp; retq@@ -2068,10 +2070,10 @@ block0(v0: i64, v1: i64):; pushq %rbp; movq %rsp, %rbp; block1: ; offset 0x4+; xorq %rdx, %rdx; movq %rsi, %rax; addq (%rdi), %rax ; trap: heap_oob-; setb %sil-; movzbq %sil, %rdx+; adcq $0, %rdx; movq %rbp, %rsp; popq %rbp; retq@@ -2204,10 +2206,11 @@ block0(v0: i64, v1: i64):; pushq %rbp; movq %rsp, %rbp; block0:+; uninit %rdx+; xorq %rdx, %rdx; movq %rdi, %rax; addq %rsi, %rax-; setb %sil-; movzbq %sil, %rdx+; adcq $0x0, %rdx; movq %rbp, %rsp; popq %rbp; retq@@ -2217,10 +2220,10 @@ block0(v0: i64, v1: i64):; pushq %rbp; movq %rsp, %rbp; block1: ; offset 0x4+; xorq %rdx, %rdx; movq %rdi, %rax; addq %rsi, %rax-; setb %sil-; movzbq %sil, %rdx+; adcq $0, %rdx; movq %rbp, %rsp; popq %rbp; retq