diff --git a/src/codegen_new/codegen_backend_arm64_ops.c b/src/codegen_new/codegen_backend_arm64_ops.c index 915cae93d..21f0df34d 100644 --- a/src/codegen_new/codegen_backend_arm64_ops.c +++ b/src/codegen_new/codegen_backend_arm64_ops.c @@ -180,6 +180,7 @@ # define OPCODE_SQSUB_V8B (0x0e202c00) # define OPCODE_SQSUB_V4H (0x0e602c00) # define OPCODE_SQXTN_V8B_8H (0x0e214800) +# define OPCODE_SQXTUN_V8B_8H (0x7e212800) # define OPCODE_SQXTN_V4H_4S (0x0e614800) # define OPCODE_SHL_VD (0x0f005400) # define OPCODE_SHL_VQ (0x4f005400) @@ -1225,6 +1226,12 @@ host_arm64_SQXTN_V8B_8H(codeblock_t *block, int dst_reg, int src_reg) { codegen_addlong(block, OPCODE_SQXTN_V8B_8H | Rd(dst_reg) | Rn(src_reg)); } + +void host_arm64_SQXTUN_V8B_8H(codeblock_t *block, int dst_reg, int src_reg) +{ + codegen_addlong(block, OPCODE_SQXTUN_V8B_8H | Rd(dst_reg) | Rn(src_reg)); +} + void host_arm64_SQXTN_V4H_4S(codeblock_t *block, int dst_reg, int src_reg) { diff --git a/src/codegen_new/codegen_backend_arm64_ops.h b/src/codegen_new/codegen_backend_arm64_ops.h index df751b4aa..084bbb404 100644 --- a/src/codegen_new/codegen_backend_arm64_ops.h +++ b/src/codegen_new/codegen_backend_arm64_ops.h @@ -184,6 +184,7 @@ void host_arm64_SQSUB_V8B(codeblock_t *block, int dst_reg, int src_n_reg, int sr void host_arm64_SQSUB_V4H(codeblock_t *block, int dst_reg, int src_n_reg, int src_m_reg); void host_arm64_SQXTN_V8B_8H(codeblock_t *block, int dst_reg, int src_reg); +void host_arm64_SQXTUN_V8B_8H(codeblock_t *block, int dst_reg, int src_reg); void host_arm64_SQXTN_V4H_4S(codeblock_t *block, int dst_reg, int src_reg); void host_arm64_SHL_V4H(codeblock_t *block, int dst_reg, int src_reg, int shift); diff --git a/src/codegen_new/codegen_backend_arm64_uops.c b/src/codegen_new/codegen_backend_arm64_uops.c index 2bb6281ff..f6038f71f 100644 --- a/src/codegen_new/codegen_backend_arm64_uops.c +++ b/src/codegen_new/codegen_backend_arm64_uops.c @@ -1480,8 +1480,8 @@ codegen_PACKUSWB(codeblock_t *block, uop_t *uop) int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real), src_size_b = IREG_GET_SIZE(uop->src_reg_b_real); if (REG_IS_Q(dest_size) && REG_IS_Q(src_size_b) && uop->dest_reg_a_real == uop->src_reg_a_real) { - host_arm64_UQXTN_V8B_8H(block, REG_V_TEMP, src_reg_b); - host_arm64_UQXTN_V8B_8H(block, dest_reg, dest_reg); + host_arm64_SQXTUN_V8B_8H(block, REG_V_TEMP, src_reg_b); + host_arm64_SQXTUN_V8B_8H(block, dest_reg, dest_reg); host_arm64_ZIP1_V2S(block, dest_reg, dest_reg, REG_V_TEMP); } else fatal("PACKUSWB %02x %02x %02x\n", uop->dest_reg_a_real, uop->src_reg_a_real, uop->src_reg_b_real);