https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106346
Bug ID: 106346 Summary: Potential regression on vectorization of left shift with constants Product: gcc Version: 13.0 Status: UNCONFIRMED Keywords: missed-optimization Severity: normal Priority: P3 Component: tree-optimization Assignee: unassigned at gcc dot gnu.org Reporter: manolis.tsamis at vrull dot eu Target Milestone: --- Target: aarch64 Created attachment 53317 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53317&action=edit Does not vectorize on GCC > 10.3 The following test case: void foo (uint32_t dst[8], uint8_t src1[8], uint8_t src2[8]) { uint16_t diff_e0 = src1[0] - src2[0]; uint16_t diff_e1 = src1[1] - src2[1]; uint16_t diff_e2 = src1[2] - src2[2]; uint16_t diff_e3 = src1[3] - src2[3]; uint16_t diff_e4 = src1[4] - src2[4]; uint16_t diff_e5 = src1[5] - src2[5]; uint16_t diff_e6 = src1[6] - src2[6]; uint16_t diff_e7 = src1[7] - src2[7]; uint32_t a0 = diff_e0 << 1; uint32_t a1 = diff_e1 << 3; uint32_t a2 = diff_e2 << 4; uint32_t a3 = diff_e3 << 2; uint32_t a4 = diff_e4 << 12; uint32_t a5 = diff_e5 << 11; uint32_t a6 = diff_e6 << 9; uint32_t a7 = diff_e7 << 3; dst[0] = a0; dst[1] = a1; dst[2] = a2; dst[3] = a3; dst[4] = a4; dst[5] = a5; dst[6] = a6; dst[7] = a7; } Compiles at -O3 to nice vectorized code by loading the constants from memory in GCC 10.3: ldr d0, [x1] adrp x3, .LC0 ldr d1, [x2] adrp x1, .LC1 ldr q3, [x3, #:lo12:.LC0] usubl v0.8h, v0.8b, v1.8b ldr q2, [x1, #:lo12:.LC1] uxtl v1.4s, v0.4h uxtl2 v0.4s, v0.8h sshl v1.4s, v1.4s, v3.4s sshl v0.4s, v0.4s, v2.4s stp q1, q0, [x0] ret But this has regressed in later releases, with GCC still loading the constants from memory but also emitting a lot of scalar code before that. For example GCC 13 produces: adrp x3, .LC0 ldrb w6, [x1, 4] fmov d0, x6 ldrb w7, [x1] ldr q5, [x3, #:lo12:.LC0] fmov d1, x7 ldrb w3, [x1, 5] ldrb w4, [x1, 1] ldrb w8, [x2, 4] ldrb w5, [x2, 5] ins v0.h[1], w3 ldrb w6, [x2] fmov d2, x8 ldrb w3, [x2, 1] fmov d3, x6 ins v2.h[1], w5 ins v1.h[1], w4 ldrb w9, [x1, 2] ins v3.h[1], w3 ldrb w8, [x1, 6] ldrb w7, [x2, 2] ldrb w6, [x2, 6] ins v1.h[2], w9 ins v0.h[2], w8 ldrb w5, [x1, 3] ins v3.h[2], w7 ldrb w4, [x1, 7] ins v2.h[2], w6 ldrb w1, [x2, 7] ldrb w3, [x2, 3] ins v1.h[3], w5 ins v0.h[3], w4 ins v2.h[3], w1 ins v3.h[3], w3 adrp x1, .LC1 ldr q4, [x1, #:lo12:.LC1] sub v1.4h, v1.4h, v3.4h sub v0.4h, v0.4h, v2.4h uxtl v1.4s, v1.4h uxtl v0.4s, v0.4h sshl v1.4s, v1.4s, v5.4s sshl v0.4s, v0.4s, v4.4s stp q1, q0, [x0] ret Interestingly, this happens only with left shift and not with right shift. GCC 10.3 vs trunk comparison: https://godbolt.org/z/xWbfGdfen