Skip to content

A lightweight WebAssembly runtime that is fast, secure, and standards-compliant

Cranelift: fix alignment on cmpps/cmppd. (#14641)

Chris Fallinec1ef6c3992e
diff --git a/cranelift/assembler-x64/meta/src/instructions/cmp.rs b/cranelift/assembler-x64/meta/src/instructions/cmp.rs
index 7a65fc9159..b178ce1d1a 100644
--- a/cranelift/assembler-x64/meta/src/instructions/cmp.rs
+++ b/cranelift/assembler-x64/meta/src/instructions/cmp.rs
@@ -48,8 +48,8 @@ pub fn list() -> Vec<Inst> {
// Floating-point comparisons.
inst("cmpss", fmt("A", [rw(xmm1), r(xmm_m32), r(imm8)]), rex([0xF3, 0x0F, 0xC2]).r().ib(), (_64b | compat) & sse).custom(Display),
inst("cmpsd", fmt("A", [rw(xmm1), r(xmm_m64), r(imm8)]), rex([0xF2, 0x0F, 0xC2]).r().ib(), (_64b | compat) & sse2).custom(Display),
- inst("cmpps", fmt("A", [rw(xmm1), r(xmm_m128), r(imm8)]), rex([0x0F, 0xC2]).r().ib(), (_64b | compat) & sse).custom(Display),
- inst("cmppd", fmt("A", [rw(xmm1), r(xmm_m128), r(imm8)]), rex([0x66, 0x0F, 0xC2]).r().ib(), (_64b | compat) & sse2).custom(Display),
+ inst("cmpps", fmt("A", [rw(xmm1), r(align(xmm_m128)), r(imm8)]), rex([0x0F, 0xC2]).r().ib(), (_64b | compat) & sse).custom(Display),
+ inst("cmppd", fmt("A", [rw(xmm1), r(align(xmm_m128)), r(imm8)]), rex([0x66, 0x0F, 0xC2]).r().ib(), (_64b | compat) & sse2).custom(Display),
inst("vcmpss", fmt("B", [w(xmm1), r(xmm2), r(xmm_m32), r(imm8)]), vex(LIG)._f3()._0f().op(0xC2).r().ib(), (_64b | compat) & avx).custom(Display),
inst("vcmpsd", fmt("B", [w(xmm1), r(xmm2), r(xmm_m64), r(imm8)]), vex(LIG)._f2()._0f().op(0xC2).r().ib(), (_64b | compat) & avx).custom(Display),
inst("vcmpps", fmt("B", [w(xmm1), r(xmm2), r(xmm_m128), r(imm8)]), vex(L128)._0f().op(0xC2).r().ib(), (_64b | compat) & avx).custom(Display),
diff --git a/cranelift/filetests/filetests/isa/x64/simd-fcmp-unaligned.clif b/cranelift/filetests/filetests/isa/x64/simd-fcmp-unaligned.clif
new file mode 100644
index 0000000000..d71bf52c48
--- /dev/null
+++ b/cranelift/filetests/filetests/isa/x64/simd-fcmp-unaligned.clif
@@ -0,0 +1,85 @@
+test compile precise-output
+target x86_64
+
+function %fcmp_eq_f32x4_load(f32x4, i64) -> i32x4 {
+block0(v0: f32x4, v1: i64):
+ v2 = load.f32x4 v1+1
+ v3 = fcmp eq v0, v2
+ return v3
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; movdqu 1(%rdi), %xmm3
+; cmpeqps %xmm3, %xmm0
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; movdqu 1(%rdi), %xmm3 ; trap: heap_oob
+; cmpeqps %xmm3, %xmm0
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+
+function %fcmp_lt_f64x2_load(f64x2, i64) -> i64x2 {
+block0(v0: f64x2, v1: i64):
+ v2 = load.f64x2 v1+1
+ v3 = fcmp lt v0, v2
+ return v3
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; movdqu 1(%rdi), %xmm3
+; cmpltpd %xmm3, %xmm0
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; movdqu 1(%rdi), %xmm3 ; trap: heap_oob
+; cmpltpd %xmm3, %xmm0
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+
+function %fcmp_eq_f32x4_load_aligned(f32x4, i64) -> i32x4 {
+block0(v0: f32x4, v1: i64):
+ v2 = load.f32x4 aligned v1
+ v3 = fcmp eq v0, v2
+ return v3
+}
+
+; VCode:
+; pushq %rbp
+; movq %rsp, %rbp
+; block0:
+; cmpeqps (%rdi), %xmm0
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+;
+; Disassembled:
+; block0: ; offset 0x0
+; pushq %rbp
+; movq %rsp, %rbp
+; block1: ; offset 0x4
+; cmpeqps (%rdi), %xmm0 ; trap: heap_oob
+; movq %rbp, %rsp
+; popq %rbp
+; retq
+
diff --git a/cranelift/filetests/filetests/runtests/simd-fcmp-unaligned.clif b/cranelift/filetests/filetests/runtests/simd-fcmp-unaligned.clif
new file mode 100644
index 0000000000..8784c72dc3
--- /dev/null
+++ b/cranelift/filetests/filetests/runtests/simd-fcmp-unaligned.clif
@@ -0,0 +1,34 @@
+test interpret
+test run
+target aarch64
+target s390x
+target x86_64
+target x86_64 has_sse3 has_ssse3 has_sse41
+target x86_64 has_sse3 has_ssse3 has_sse41 has_avx
+set enable_multi_ret_implicit_sret
+target riscv64 has_v
+target riscv64 has_v has_c has_zcb
+target pulley64
+target pulley64be
+
+function %fcmp_eq_unaligned_f32x4(f32x4) -> i32x4 {
+ ss0 = explicit_slot 48, align = 16
+block0(v0: f32x4):
+ v1 = stack_addr.i64 ss0
+ store.f32x4 v0, v1+1
+ v2 = load.f32x4 v1+1
+ v3 = fcmp eq v0, v2
+ return v3
+}
+; run: %fcmp_eq_unaligned_f32x4([0x1.0 0x2.0 0x3.0 0x4.0]) == [-1 -1 -1 -1]
+
+function %fcmp_lt_unaligned_f64x2(f64x2) -> i64x2 {
+ ss0 = explicit_slot 48, align = 16
+block0(v0: f64x2):
+ v1 = stack_addr.i64 ss0
+ store.f64x2 v0, v1+1
+ v2 = load.f64x2 v1+1
+ v3 = fcmp lt v0, v2
+ return v3
+}
+; run: %fcmp_lt_unaligned_f64x2([0x1.0 0x2.0]) == [0 0]