https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113517
Bug ID: 113517 Summary: vector SLP cost model should be improved Product: gcc Version: 14.0 Status: UNCONFIRMED Keywords: missed-optimization Severity: enhancement Priority: P3 Component: target Assignee: unassigned at gcc dot gnu.org Reporter: pinskia at gcc dot gnu.org Target Milestone: --- Target: aarch64 Take: ``` void f(short *a, signed int *b) { int sum = 0; sum += a[0]; sum += a[1]; sum += a[2]; sum += a[3]; *b = sum; } ``` Right now by default this produces: ``` ldrsh w3, [x0] ldrsh w4, [x0, 2] ldrsh w2, [x0, 4] add w3, w3, w4 ldrsh w0, [x0, 6] add w2, w2, w3 add w0, w0, w2 str w0, [x1] ``` But disabling the cost model we get: ``` ldr d31, [x0] saddlv s31, v31.4h str s31, [x1] ret ``` (note this is better code generation than what LLVM produces as that uses sshll/addv ). For most cores, doing a float (vector) load and one vector instruction and one vector store is better than doing 4 scalar loads and 3 scalar instructions and one scalar store. This is true on even ThunderX 1 and Cortex-A57.