https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114544
--- Comment #5 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Roger Sayle <[email protected]>: https://gcc.gnu.org/g:6b1ffdc831718b1e7639d25c37ea1cdef48e1b53 commit r17-4566-g6b1ffdc831718b1e7639d25c37ea1cdef48e1b53 Author: Roger Sayle <[email protected]> Date: Tue Sep 22 15:57:40 2026 +0100 x86_64: Improved 128-bit TImode moves from SSE to GPRs with SSE2. This patch improves TImode moves on x86_64 between SSE and general registers with SSE2. This is a partial solution for PR target/114544. Consider the following simple function: typedef __int128 v1ti __attribute__((vector_size(16))); __int128 foo(v1ti x) { return x[0]; } Currently, with default -O2 (-msse2) the 128-bit value is moved via the stack: foo: movaps %xmm0, -24(%rsp) // 5 bytes movq -24(%rsp), %rax // 5 bytes movq -16(%rsp), %rdx // 5 bytes ret Though with -O2 -msse4.1, the transfers can be done directly. foo: movq %xmm0, %rax pextrq $1, %xmm0, %rdx ret With this patch, we make use of SSE2's shuffle functionality to avoid the transfers to and from memory: foo: movq %xmm0, %rax // 5 bytes shufpd $1, %xmm0, %xmm0 // 5 bytes movq %xmm0, %rdx // 5 bytes ret Technically, to avoid using a second register we swap/rotate the high and low lanes between movq instructions, and (not shown above) also swap/rotate the lanes back again (so the source register remains unchanged if the value is still live). This second (restoring) shufpd is eliminated by later RTL passes if it's discovered to be dead code. To make this possible, the Yd constraint that enables allows *movti_internal to match on SSE4_1, is tweaked to form a new YD constraint that allows *movti_internal to match on SSE2. 2026-09-22 Roger Sayle <[email protected]> Hongtao Liu <[email protected]> gcc/ChangeLog PR target/114544 * config/i386/constraints.md (YD): New constraint, like Yd but allows SSE_REGS on TARGET_SSE2. * config/i386/i386.cc (inline_secondary_memory_needed): Update that TImode moves to GENERAL_REGS don't need to use memory. * config/i386/i386.md (*movti_internal): Change Yd constraint to YD. (define_split): On TARGET_64BIT && TARGET_SSE2 && !TARGET_SSE4_1, split *movti_internal into movq+shufpd+movq+shufpd. gcc/testsuite/ChangeLog * gcc.target/i386/pr126293-15.c: Update test case. * gcc.target/i386/pr126293-5.c: Likewise.
